Three Essays on Gender and Cultural Economics

Aix-Marseille Université - École Doctorale de Sciences Économiques et de Gestion d’Aix-Marseille N°372 Faculté d’Économie et de Gestion

Aix-Marseille School of Economics

Numéro attribué par la bibliothèque

a a a a a a a a a a

Thèse pour le Doctorat ès Sciences Économiques Présentée et soutenue publiquement par

Jordan LOPER

le 28 août 2020 en vue de l’obtention du grade de docteur d’Aix-Marseille Université

Three Essays on Gender and Cultural Economics

Jury :

Jean-Marie BALAND University of Namur, Rapporteur et Président du Jury Nicolas BERMAN Aix-Marseille University, Examinateur

Nathan NUNN Harvard University, Examinateur

Christian SCHLUTER Aix-Marseille University, Co-directeur de thèse

Alessandra VOENA University of Chicago, Rapporteur

Roberta ZIPARO Aix-Marseille University, Co-directrice de thèse ii L’université d’Aix-Marseille n’entend ni approuver, ni désapprouver les opinions particulières du candidat : Ces opinions doivent être considérées comme propres à leur auteur.

iii iv A mes parents, à ma famille, à Eva, à mes amis, à toutes les personnes que j’ai croisées sur mon chemin...

v vi Remerciements

Quelle aventure que de se lancer dans une thèse !... A l’image de la vie, le parcours est vallonné, semé d’embûches, et les sommets à grimper son multiples... Chaque sommet gravi est pourtant l’occasion d’un enrichissement, d’une prise de vue assez exceptionnelle sur la vie, mais aussi sur la compréhension du monde... Un tel chemin ne se parcourt pas seul, et je veux ici remercier toutes les personnes qui m’ont apporté, de près comme de loin, l’énergie et les ressources nécessaires jusqu’à l’ascension du dernier sommet.

En premier lieu, je veux exprimer ma gratitude inﬁnie à Roberta Ziparo, formidable co-directrice de thèse durant ces 4 années. Tu as cru en mon envie ainsi qu’en mon projet dès le tout premier jour, et tu as su faire conﬁance en l’avenir pour cheminer avec moi durant 4 années, alors que nous ne nous

étions jamais rencontrés auparavant et que tu débutais ta carrière académique. Tu as également fait de moi le chercheur que je suis aujourd’hui, m’apprenant les rouages du métier, et alimentant sans cesse ma curiosité. Tu as par ailleurs toujours su faire preuve d’ouverture d’esprit, et tu as toujours priviliégié l’esprit d’analyse sur l’esprit de jugement. Enﬁn, tu as toujours fait preuve d’une grande humanité: la plus grande des qualités à mes yeux. Pour tout cela, est plus encore, je t’exprime ma gratitude inﬁnie. J’espère que nous continuerons de cheminer ensemble...

Je voudrais aussi exprimer ma gratitude à mes autres co-directeurs. A Olivier Bargain, qui à

également cru en mon projet et accepté de me superviser dès le premier jour. Bien que vous ayez passé le relai à Christian Schluter à la fin de ma première année de thèse suite à votre changement d’université, vous êtes par la suite resté un fidèle co-auteur de grande exigence. J’ai beaucoup appris sur le métier à vos côtés, et j’espère que la route sera belle pour nos projets en cours... Christian, je voudrais également vous remercier d’avoir accepté de prendre la suite d’Olivier. Votre regard aiguisé m’a été précieux pour améliorer la qualité de mon travail. Je vous remercie enfin pour votre gentillesse tout au long de mon parcours...

Je remercie chaleureusement Nathan Nunn, pour avoir accepté de faire partie de mon jury de thèse, mais aussi pour avoir accepté de m’inviter en visiting au Département d’Economie de Harvard University. Le petit “Cévenol” que je suis a eu l’impression de vivre dans un ﬁlm le temps de ce visiting. Changement de pays, changement de culture, changement de perspective... Un dépaysement total... Une autre illustration que la vie nous réserve parfois quelques bonnes surprises... Je mesure la chance d’avoir pu interagir avec l’un de mes mentors dans la profession. Nos discussions ont grandement contribué à améliorer la qualité de mon travail et m’ont permis d’élargir le spectre de mes réﬂexions. Je remercie Harvard University de m’avoir accueilli, et également toutes les personnes que j’ai rencontrées lors de ce visiting, et en premier lieu les chercheurs, les doctorants et l’équipe administrative pour leur accueil et pour les discussions de qualité. J’ai une pensée toute particulière pour Dr. Lewis Patsavos, qui m’a ouvert les portes de sa maison durant les 3 mois de mon visiting: merci pour votre grande chaleur humaine, ainsi que de m’avoir transmis avec passion l’histoire de

Cambridge, Massachusetts.

Plus généralement, je remercie Jean-Marie Baland, Nicolas Berman, Nathan Nunn et Alessandra

Voena d’avoir accepté de faire partie de jury de thèse. Je sais vos calendriers très remplis, et je mesure la chance de pourvoir bénéﬁcier de vos encouragements, vos suggestions, vos commentaires ou encore vos critiques. Bien que “challenging”, je vois ces dernières comme une occasion unique de progresser toujours plus...

Ma gratitude également à Aix-Marseille Université et à Aix-Marseille School of Economics

(AMSE), ainsi qu’à toutes les personnes qui font de cette formidable institution une grande famille.

En premier lieu, un immense merci aux Directeurs du Programme Doctoral: Marc Sangnier puis

Nicolas Berman et Sebastian Bervoets qui ont pris sa suite. Vous accomplissez un travail formidable, et votre support auprès des doctorants est très précieux. Vous contribuez par ailleurs à forger une vraie identité au Programme Doctoral de l’AMSE: un grand merci à vous pour cela. Un grand merci

également à l’ensemble de la communauté des chercheurs de l’AMSE. La grande accessibilité des

Professeurs “seniors” est une grande richesse pour les doctorants ! Je remercie justement également les doctorants de l’AMSE, mes compagnons, mes partenaires d’aventure. Apprendre le métier de chercheur s’accompagne de “très hauts” comme “de très bas”, et la solidarité qui existe entre nous est un atout plus que précieux. Je sais combien la recherche peut isoler, et je veux vous remercier pour les discussions, les retours d’expérience, mais aussi les moments d’amitié entre nous. J’ai évidemment une pensée toute particulière pour les doctorants du Château Lafarge (dit le “Château)”, au premier rang desquels Laurène, Rémi, Mathilde, Morgan, Fatemeh, Marie-Christine, François, Fabien, Nan- deeta et Rosnel. Merci à vous pour votre présence, votre simplicité et votre générosité qui m’ont

viii permis de tenir le cap au quotidien. Enﬁn, je veux exprimer ma plus grande gratitude à toutes les personnes faisant partie de l’équipe administrative, et en tout particulier à Bernadette Vouriot, Anne-

Sophie Therond et Mathilde Martelli. Votre humanité et votre générosité sont l’ADN de l’AMSE.

Vous êtes le socle sur lequel repose l’institution et vous faites du laboratoire une grande famille: encore merci à vous !

Je remercie aussi la Toulouse School of Economics (TSE), mon école d’origine, de m’avoir fourni des bases solides nécessaires pour devenir un bon chercheur; mais également de m’avoir offert un contrat d’Attaché Temporaire d’Enseignement et de Recherche (ATER), qui m’a permis de bénéﬁcier d’une quatrième année pour terminer ma thèse dans la sérénité tout en me lançant dans l’aventure du “Job Market” pour trouver un poste après la thèse. Je vous remercie de m’avoir accueilli comme l’un de vos propres doctorants, m’offrant un bureau dans des locaux ﬂambants neufs, avec une vue imprenable sur la Garonne et Toulouse, la “Ville Rose” si chère à mon cœur...

Enﬁn, et probablement le plus important, je voudrais exprimer ma gratitude inﬁnie à toutes les personnes qui sont chères à mon cœur, et qui donnent tout le sel à mon passage sur Terre. Ma famille en premier lieu: un immense merci en particulier à mes parents et à mon frère pour leur patience et leur soutien inconditionnel. Vous si loin de mon travail et pourtant si près de mon cœur. Vous aussi la raison qui me fait persévérer vers la recherche du mieux. J’ai commis des erreurs, je continue d’en commettre et j’en commettrai, mais j’espère continuer de progresser, de m’améliorer... Merci pour votre amour inconditionnel.

Je voudrais aussi remercier Eva, du fond de mon cœur. Tu auras été ma partenaire de vie, la femme que j’aurais aimée, ma conﬁdente pendant ces 4 années de thèse. Mon être porte ta trace marquée au fer rouge... Je te remercie pour tout ce que tu m’as apporté, pour m’avoir aimé malgré mes failles.

Je te remercie de m’avoir laissé partager un bout de chemin avec toi... Envole-toi maintenant, petit colibri...

Enﬁn, je voudrais remercier mes amis, ma “Dream Team” en particulier (Alexandre, Hugo,

Maxime et Xavier, par ordre alphabétique): je vous vois comme différentes facettes d’un même dia- mant. Merci d’être dans ma vie. Je voudrais aussi, autant que mes bras me le permettent, enlacer toutes les personnes qui comptent dans ma vie: la liste est trop longue pour être énumérée ici, mais rien ne pourrait contredire ce qui nous unit. Vous êtes un exhausteur de vie et je ne suis rien d’autre qu’une mosaïque de vos empreintes. Merci d’être dans ma vie.

ix x Summary

Matrilineality in Sub-Saharan Africa” and explores the long-term effect of matrilineality on contemporaneous female HIV in Sub-Saharan Africa. Using within Sub-Saharan African countries variation in ethnic groups’ ancestral kinship organizations, I find that females originating from ancestrally matrilineal ethnic groups are today more likely to be infected by HIV. This finding is robust to the inclusion of subnational fixed effects, as well as a large set of cultural, historical, geographical, and environmental factors. I find consistent results using a number of alternative estimation strategies, including a geographic regression discontinuity design at ethnic boundaries and an instrumental variable strategy. Matrilineal females’ riskier sexual and contraceptive behaviours constitute the main explanatory mechanisms. Building an epidemiological model, I simulate how these differences in sexual and contraceptive behaviours translate into different gender-specific HIV rates dynamics. The

2nd chapter is entitled “Traditional Norms, Access to Divorce and Women’s Empowerment: Evi- dence from Indonesia” and explores how cultural norms and development policies interact in shaping women’s empowerment and well-being in developing countries. My co-authors and I examine this question in the context of legal reforms and their differentiated impact on divorce and empowerment across traditional modes of post-marital cohabitation. We theoretically establish how women originating from matrilocal ethnic groups should respond to the reform compared to those from patrilocal ethnicities. We conﬁrm the model predictions using a panel difference-in-difference approach: the former divorce more and, when in stable marriages, experience a signiﬁcant improvement in well- being and empowerment. This conclusion calls for better tailored policies that can transcend cultural contexts and overcome the adherence to informal laws. The 3rd paper is entitled “Women’s Em- powerment and Husband’s Migration: Evidence from Indonesia” and examines the link between the distribution of power in marriage and the decision to split-migrate (one spouse migrates alone) in

Indonesia. My co-authors and I exploit a national policy experiment that has exogenously increased women’s bargaining power among ethnic groups of matrilocal tradition relative to patrilocal groups.

We ﬁnd that the propensity of matrilocal husbands to split-migrate, relative to patrilocal husbands, increases by 2-3.4 percentage points, i.e. a rise of 41-76%, following the reform. We suggest that empowered women may have ex-ante gained control over outcomes that are costlier to monitor for husbands when they migrate. Hence, empowerment restores some efﬁciency in migration decisions by reducing the anticipated information asymmetry and the moral hazard associated with migration.

Consistently, we show that households with empowered women are more able to cushion shocks due to natural disasters and, among all households experiencing split-migration, matrilocal women are better off than their patrilocal counterparts. We provide a theoretical framework that rationalizes the intra-household mechanisms lying behind these results.

Keywords: Gender, cultural persistence, kinship systems, HIV, sexual behaviour, female empowerment, migration

JEL classiﬁcation: D13, D91, I15, J1, K38, N37, O15, R23, Z13

xii Résumé

Afrique Sub-Saharienne”. En utilisant la variation intra-pays dans les types de structures familiales des groupes ethniques, je trouve que les femmes originaires de groupes ethniques ancestralement ma- trilinéaires sont aujourd’hui plus susceptibles d’être séropositives. Ce résultat est robuste à l’inclusion d’effets ﬁxes infranationaux et d’un large ensemble de contrôles culturels, historiques, géographiques et environnementaux. Je trouve un résultat similaire en utilisant des stratégies d’estimation différentes, telle qu’une approche de discontinuité géographique de la régression, ou une stratégie de variable instrumentale. Le principal mécanisme explicatif est le suivant : les femmes matrilinéaires adoptent des comportements sexuels et de contraception plus à risque. Finalement, je construis un modèle

épidémiologique et conduit une simulation permettant de comprendre comment les différences de comportements sexuels et de contraception trouvées entre les femmes matrilinéaires et patrilinéaires se traduisent en différentes dynamiques de transmission du VIH. Le 2e chapitre est intitulé “Normes

Traditionnelles, Accès au Divorce et Emancipation des Femmes : Etude de Cas de l’Indonésie” et explore la manière dont les normes traditionnelles interagissent avec les politiques de développe- ment visant à soutenir l’émancipation et le bien-être des femmes dans les pays en développement.

Mes co-auteurs et moi-même examinons l’impact différencié de réformes légales sur le divorce et l’émancipation des femmes, en fonction de la norme traditionnelle de cohabitation post-mariage (ma- trilocalité et patrilocalité). Nous montrons théoriquement comment les femmes originaires de groupes ethniques matrilocaux répondent aux réformes, relativement aux femmes originaires de groupes ethniques patrilocaux. Nous conﬁrmons ensuite empiriquement les prédictions du modèle avec une approche en double différences : nous trouvons que suite aux réformes les femmes matrilocales divorcent plus et, lorsqu’elles restent mariées, bénéﬁcient d’une amélioration de leur bien-être et de leur

xiii pouvoir de décision au sein du ménage. Cette conclusion souligne le besoin de penser des politiques qui soient plus adaptées aux différents contextes culturels. Le 3e chapitre s’intitule “Emancipation des Femmes et Migration du Mari : Etude de Cas de l’Indonésie” et examine le lien entre la distribution du pouvoir de décision entre les époux et la décision de migrer seul du mari. Mes co-auteurs et moi-même utilisons une série de réformes qui ont augmenté de façon exogène le pouvoir de déci- sion des femmes matrilocales, relativement aux femmes patrilocales. Nous trouvons que suite à ces réformes la propension à migrer seul des époux matrilocaux a augmenté de 2-3,4 points de pourcentage relativement aux époux patrilocaux, une augmentation de 41-76%. Ce résultat suggère que les femmes émancipées détiennent ex-ante le pouvoir de décision sur des aspects de la vie de famille qui sont difﬁciles à surveiller par le mari lorsqu’il migre seul. Ainsi, l’émancipation des femmes restaure de l’efﬁcacité dans la décision de migration du mari, en réduisant l’asymétrie d’information anticipée et l’aléa moral associés à la migration. Nous montrons aussi que les ménages où les femmes sont plus

émancipées sont davantage capables d’absorber les chocs liés aux catastrophes naturelles en envoyant un mari en migration. Nous trouvons également que parmi les ménages ayant un mari migrant, les ménages matrilocaux se retrouvent en meilleure posture que les ménages patrilocaux. Pour conclure, nous élaborons un cadre théorique aﬁn de rationaliser les mécanismes intra-ménages se trouvant der- rière ces résultats.

Mots-Clés: Genre, persistence culturelle, systèmes de parenté, VIH, comportements sexuels, éman- cipation des femmes, migration

Classiﬁcation JEL: D13, D91, I15, J1, K38, N37, O15, R23, Z13

xiv xv xvi Contents

General Introduction1

References...... 3

1 Ancestral Matrilineality and Female HIV in Sub-Saharan Africa7

1.1 Introduction...... 9

1.2 Context and Conceptual Framework...... 17

1.2.1 Ancestral Matrilineality in Sub-Saharan Africa...... 17

1.2.2 Female HIV in Sub-Saharan Africa...... 18

1.2.3 Conceptual Framework...... 20

1.2.3.1 Gender Differences in Sexual Strategies...... 20

1.2.3.2 Women’s Empowerment and Contraceptive Use...... 24

1.3 Data and Empirical Strategy...... 25

1.3.1 Contemporary Data...... 25

1.3.2 Historical Data...... 26

1.3.3 OLS Empirical Strategy...... 27

1.4 The Long-Term Effect of Ancestral Matrilineality on Female HIV...... 31

1.4.1 Main Results - OLS Estimates...... 31

1.4.2 Robustness Checks...... 33

1.4.2.1 Selection Analysis...... 33

1.4.2.2 Other Health Outcomes as a Falsiﬁcation Test...... 35

1.4.2.3 Robustness to Alternative Channels...... 35

1.4.2.4 Assessing Selection on Unobservables...... 42

1.4.3 Alternative Identiﬁcation Strategies...... 44

1.4.3.1 Accounting for Unobservables: Geographic RD Estimates..... 44

xvii 1.4.3.2 Instrumenting for Ancestral Matrilineality...... 48

1.4.3.3 Nearest Neighbor Matching...... 50

1.5 Mechanisms...... 52

1.5.1 Female Sexual Autonomy...... 52

1.5.2 Sexual Behaviour...... 56

1.5.3 Contraceptive Behaviour...... 58

1.5.4 Discarded Mechanisms...... 62

1.6 An Epidemiological Approach...... 64

1.6.1 The Model...... 64

1.6.2 The Effect of Matrilineality: Comparative Statics...... 67

1.6.3 Simulation...... 68

1.7 Conclusions...... 71

References...... 73

Appendices 81

1.A Data Description...... 81

1.A.1 List of the countries included in the analysis...... 81

1.A.2 Description of the main controls...... 82

1.A.2.1 Individual-level Data...... 82

1.A.2.2 Ethnicity-level Data...... 82

1.A.2.3 Village-level Data...... 83

1.B Additional Figures...... 86

1.C Additional Tables...... 93

2 Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from In-

donesia 111

2.1 Introduction...... 113

2.2 Background on Social Norms and Legal Reforms...... 115

2.2.1 Traditional Norms and Female Empowerment: Global Evidence...... 115

2.2.2 Traditional Residence Norms: the Indonesian Context...... 119

2.2.3 National Access-to-Justice Strategy: a Natural Experiment...... 121

2.3 Theoretical Framework...... 122

xviii 2.3.1 Preferences...... 123

2.3.2 Income Dynamics and Budget Constraints...... 124

2.3.3 Problem of Divorce before the Reform...... 124

2.3.4 Problem of Divorce after the Reform...... 125

2.4 Empirical Approach...... 126

2.4.1 Data...... 126

2.4.2 Empirical Approach...... 128

2.5 Results...... 130

2.5.1 Effect of the Reforms on Divorce...... 130

2.5.2 Main Results: Effect of the Reforms on Women’s Outcomes...... 131

2.5.3 Robustness Checks, Heterogeneity and Interpretations...... 134

2.6 Conclusions...... 140

References...... 141

Appendices 147

2.A Descriptive Statistics and Raw Difference-in-Difference...... 147

2.B Cross-Sectional Estimations (Correlations)...... 151

2.C Placebo and Speciﬁcation Checks...... 154

2.D Proofs of Propositions 1 and 2...... 160

3 Women’s Empowerment and Husband’s Migration: Evidence from Indonesia 165

3.1 Introduction...... 167

3.2 Background on Social Norms and Legal Reforms...... 170

3.2.1 National Access-to-Justice Strategy: a Natural Experiment...... 170

3.2.2 Traditional Residence Norms: the Indonesian Context...... 172

3.2.3 Migration in Indonesia...... 173

3.3 Data and Empirical Strategy...... 173

3.3.1 IFLS Data...... 173

3.3.2 Empirical Strategy...... 176

3.4 Results...... 177

3.4.1 Main Results on Split-Migration...... 177

3.4.2 Interpretation and Heterogeneity...... 180

xix 3.4.3 Mechanisms: Theoretical Background...... 184

3.5 Conclusions...... 187

References...... 188

Appendices 193

3.A Descriptive Statistics and Robustness Checks...... 193

General Conclusion 199

xx List of Tables

1.1 The Effect of Ancestral Matrilineality on Female HIV (OLS)...... 33

1.2 Robustness of OLS Estimates to Additional Controls...... 41

1.3 Geographic RD Estimates...... 49

1.4 IV Estimates of the Effect of Ancestral Matrilineality on Female HIV...... 51

1.5 Nearest Neighbor Matching (ATT)...... 52

1.6 Ancestral Matrilineality and Female Sexual Autonomy (OLS)...... 55

1.7 Ancestral Matrilineality and Sexual Behaviour (OLS)...... 58

1.8 Ancestral Matrilineality and Contraception (OLS)...... 61

1.9 Simulated HIV Rates...... 70

1.C.1 Summary Statistics (Final Regression Sample)...... 95

1.C.2 Ancestral Matrilineality, Genetic Diversity and Father Diversiﬁcation (OLS)..... 96

1.C.3 Selection And Falsiﬁcation Tests...... 97

1.C.4 Heterogeneous Effects by Subsamples: Common Law vs. Civil Law Countries /

Polygynous vs. Non Polygynous Individuals...... 98

1.C.5 Considering Ancestral Matrilocality...... 99

1.C.6 Robustness of OLS Estimates to Gender-Speciﬁc Alternative Channels...... 100

1.C.7 Assessing the importance of bias from unobservables by controlling for observable

characteristics...... 101

1.C.8 Ancestral Matrilineality and Couples HIV Discordance (OLS)...... 102

1.C.9 Ancestral Matrilineality and Fertility (OLS)...... 103

1.C.10Heterogeneityin Condom Use by Perception of Risk of HIV Transmission (OLS).. 104

1.C.11AncestralMatrilineality, Acknowledgment of HIV Risks and Access to Condom (OLS)105

1.C.12AncestralMatrilineality and Sexual Debuts (OLS)...... 106

1.C.13Summaryof Model’s Parameters...... 107

xxi 1.C.14SimulatedHIV Rates (Alternative Scenarios)...... 108

2.1 Worldwide Correlations between Ancestral Relative Matrilocality and Contempora-

neous Attitudes towards Women’s Role...... 118

2.1 Effect of Legal Reforms on Women’s Divorce Probability...... 132

2.2 Effect of Legal Reforms on Women’s Well-Being and Empowerment (Stable Couples,

2007-14)...... 133

2.3 Heterogenous Effect of the Reform according to the Distance to District Capital... 139

2.A.1Determination of Traditional Post-Marital Residence Norm by Ethnicity...... 147

2.A.2Traditional vs. Actual Matrilocality in Indonesia (2014)...... 148

2.A.3Descriptive Statistics of Control Variables (Stable Couples, 2007-14)...... 149

2.A.4Raw Difference-in-Differences of Outcome Variables (Stable Couples, 2007-14)... 150

2.B.1 Villages’ Post-Marriage Residence Norm and Divorce related Adat Traditional Norms 151

2.B.2 Correlations between Matrilocality and Women’s Well-Being and Empowerment (Sta-

ble Couples, 2007-14)...... 152

2.B.3 Effect of Legal Reforms on Women’s Well-Being and Empowerment (Pooled Cross-

Sections)...... 153

2.C.1 Women’s Divorce Probability: Placebo Estimations...... 154

2.C.2 Well-Being and Empowerment (Stable Couples): Placebo Estimations...... 155

2.C.3 Effect of Legal Reforms on Women’s Well-Being and Empowerment: Robustness

Checks...... 156

2.C.4 Alternative Deﬁnitions of Treatment: Matrilocal Intensity...... 157

2.C.5 Alternative Deﬁnitions of Treatment: Including Mixed Couples...... 158

2.C.6 Effect on Empowerment: Alternative Deﬁnitions based on Final Say Answers.... 159

3.1 Difference-in-difference Estimations on Husband’s Split Migration...... 178

3.2 Difference-in-difference Estimations on Joint Migration...... 179

3.3 DD Estimations on Husband’s Split Migration - Heterogeneity by Natural Disaster. 182

3.4 DD Estimations on Female and Child Outcomes - Heterogeneity by Migration Decision183

3.A.1Determination of Traditional Post-Marital Residence Norm by Ethnicity...... 193

3.A.2Descriptive Statistics of Main Variables...... 194

3.A.3Placebo DID Estimations on Husband’s Split Migration...... 195

xxii 3.A.4DID Estimations on Husband’s Split Migration (Panel with Fixed Effects)...... 196

3.A.5Placebo DID Estimations on Husband’s Split Migration (Panel with Fixed Effects). 197

xxiii xxiv List of Figures

1.1 Diagrams of kinship systems (source: Lowes, 2018a)...... 18

1.2 Ancestral Ethnic Group Boundaries and Matrilineal Belt...... 19

1.3 Ancestral Ethnic Group Boundaries and Contemporary Female HIV Rates (Final sam-

ple)...... 21

1.4 Simulation - Scenario 1 (휇푚 = 휏푚 = 4)...... 71 1.B.1 Ancestral Ethnic Group Boundaries and Contemporary Gender Differences in HIV

Rates...... 87

1.B.2 Ancestral Ethnic Group Boundaries and Contemporary Male HIV Rates...... 88

1.B.3 Location of DHS Villages...... 89

1.B.4 Individual’s Matrilineality and Distance to Nearest Ancestral Matrilineal Ethnic Bound-

ary (RDD)...... 90

1.B.5 Female HIV Rate and Distance to Nearest Ancestral Matrilineal Ethnic Boundary

(RDD)...... 91

1.B.6 Simulation - Scenario 2 (휇푚 = 4 and 휏푚 = 2)...... 92

1.B.7 Simulation - Scenario 3 (휇푚 = 휏푚 = 2)...... 92

2.1 Village-level Traditional Post-Marital Residence (IFLS data)...... 120

2.2 Divorce Trends around Reform Time...... 122

3.1 Divorce Trends around Reform Time...... 171

3.2 Village-level Traditional Post-Marital Residence (IFLS data)...... 172

xxv xxvi General Introduction

Understanding individual’s preferences and subsequent behaviour has always been a question of considerable importance in the economic discipline. Recently, accumulating evidences have highlighted that our preferences have been shaped in the long-run by cultural context (Falk, Becker, Dohmen,

Enke, Huffman and Sunde, 2018), and in particular by kinship organizations adopted by ancestral societies (Enke, 2019). Further, recent evidences show that these ancestral kinship organizations have had a long-term impact on a wide range of contemporary individual’s behaviours, from propensity to engage into conﬂict (Moscona, Nunn and Robinson, 2020) to spousal cooperation (Lowes, 2018) and women’s empowerment and well-being (Lowes and Nunn, 2017).

Along these lines, researchers are increasingly providing evidences that cultural context matters in determining the efﬁciency of development policies (Ashraf, Bau, Nunn and Voena, 2020, La Ferrara and Milazzo, 2017), all the more so as it comes to improving women’s economic rights and opportunities. Indeed, while it is well established that the legal and institutional framework of a country can greatly contribute to women’s autonomy (Duﬂo, 2012), policy-makers need to move beyond the

“one-size-ﬁts-all” strategy, especially in countries with wide cultural diversity.1

This thesis, which includes three chapters, lies at the intersection of these two streams of research, as it investigates how cultural norms and development policies interact in shaping women’s empowerment, with a speciﬁc focus on developing countries, characterized by both cultural diversity and room for improving the situation of women.

In chapter 1, entitled “Women’s Position in Ancestral Societies and Female HIV: The Long-

Term Effect of Matrilineality in Sub-Saharan Africa”, I explore the long-lasting effect of ancestral

(i.e. dating back to before colonization) ethnic groups’ matrilineal kinship structures on contemporary female HIV in Sub-Saharan Africa.2 Building on theories from evolutionary anthropology as well as

1Interested reader may have a look at the talk given by Nathan Nunn on this topic here: https://www. youtube.com/watch?v=stK1oHDYs-0 2In matrilineal kinship structures lineage and inheritance are traced through female members and children integrate the kin group of their mother rather than their father.

1 General Introduction evolutionary psychology, I expect females originating from ancestrally matrilineal ethnic groups to adopt more promiscuous sexual and contraceptive behaviours, and therefore to suffer from more HIV.

Using data from 32 Demographic Health Surveys (DHS) covering 18 Sub-Saharan African countries,

I ﬁnd empirical evidences that, indeed, matrilineal females suffer from HIV today. In line with my conceptual framework, I further ﬁnd that matrilineal women’s adoption of riskier sexual and contraceptive behaviours are the main mechanisms explaining this result.

In chapter 2, entitled “Traditional Norms, Access to Divorce and Women’s Empowerment:

Evidence from Indonesia” and which is a joint work with Olivier Bargain and Roberta Ziparo, we explore how nation-wide policies aimed at easing women’s access to justice courts in Indonesia had differentiated impact across traditional modes of post-marital cohabitation on women’s propensity to divorce and subsequent women’s empowerment and well-being. Using data from the Indonesia

Family Life Survey (IFLS), we ﬁnd that women originating from traditionally matrilocal ethnic groups were more responsive to the reforms than their patrilocal counterparts.3 More precisely, in line with our theoretical predictions, we ﬁnd that matrilocal women were more likely to divorce following the reforms, and that those remaining married experienced a renegotiation of intra-household bargaining power in their favor, as well as an increase in well-being (i.e. health, fertility, value of assets, own and children’s subjective well-being).

In chapter 3, entitled “Women’s Empowerment and Husband’s Migration: Evidence from

Indonesia” and which is also a joint work with Olivier Bargain and Roberta Ziparo, we explore how women’s empowerment may act as a driver of husband’s split migration (i.e. migration leaving his family behind), which is characterized by anticipated asymmetry of information and moral hazard.

To do so, we build on Chapter 2 were we uncover an exogenous heterogenous empowerment between matrilocal and patrilocal women following the reforms mentioned above. In line with our theoretical predictions, we empirically ﬁnd that matrilocal husbands were more likely to split migrate following the reforms, and therefore that women’s empowerment is a driver of husband’s migration.

This thesis contributes to several strands of literature: development economics, cultural economics, economics of gender, household economics, political economy, health economics and economics of migration. Further, its contributions extend beyond the economic discipline: this thesis adopts an interdisciplinary approach as it builds on theories from the evolutionary anthropology and

3According to the matrilocal post-marital residence norm, married couples are live with or close to the wife’s family; while cccording to the patrilocal post-marital residence norm, married couples live with or close to the husband’s family.

2 General Introduction the evolutionary psychology literatures. Furthermore, this thesis also links with the medical literature, and the epidemiological literature in particular.

This thesis further offers a wide coverage along several dimensions. First, it offers a wide geographic and contextual coverage: it covers both Indonesia (a country with wide geographic and cultural diversity) and 18 Sub-Saharan African countries. Second, it explores the effect of several types of traditional norms: inheritence norms (i.e. matrilineality vs. patrilineality) and post-marital residence norms (i.e. matrilocality vs. patrilocality). Third, it emphasizes a wide spectrum of outcomes, from HIV, sexual and contraceptive behaviours to women’s intra-household decision-making, health, fertility, assets, subjective well-being and husband’s migration. Finally, this thesis uses a wide range of estimation techniques to provide empirical evidences that are theoretically grounded: OLS with Fixed Effects, Geographic RDD, IV strategy, Nearest Neighbor Matching and Difference-in-

Differences.

In the end, this thesis provides two main conclusions: (1) cultural context, and ethnic kinship norms in particular, matters in shaping women’s empowerment and well-being; (2) development policies aimed at improving women’s situation should be more tailored to speciﬁc cultural contexts.

3 General Introduction

4 References

Ashraf, Nava, Natalie Bau, Nathan Nunn, and Alessandra Voena, “Bride Price and Female Education,” Journal of Political Economy, 2020.

Duflo, Esther, “Women Empowerment and Economic Development,” Journal of Economic Lit- erature, 2012, 50 (4), 1051–1079.

Enke, Benjamin, “Kinship, Cooperation, and The Evolution of Moral Systems,” The Quarterly Journal of Economics, 2019, 134 (2), 953–1019.

Falk, Armin, Anke Becker, Thomas Dohmen, Benjamin Enke, David Huffman, and Uwe Sunde, “Global Evidence on Economic Preferences,” The Quarterly Journal of Economics, 2018, 133 (4), 1645–1692.

Ferrara, Eliana La and Annamaria Milazzo, “Customary Norms, Inheritance, and Human Capital: Evidence from a Reform of The Matrilineal System in Ghana,” American Economic Journal: Applied Economics, 2017.

Lowes, Sara, “Kinship Systems, Gender Norms, and Household Bargaining: Evidence from the Matrilineal Belt,” Working Paper, 2018.

and Nathan Nunn, “Bride Price and the Wellbeing of Women,” WIDER Working Paper, 2017.

Moscona, Jacob, Nathan Nunn, and James A Robinson, “Segmentary Lineage Organi- zation and Conﬂict in Sub-Saharan Africa,” Econometrica, 2020.

5 6 Chapter 1

Women’s Position in Ancestral Societies and Female HIV: The Long-Term Effect of Matrilineality in Sub-Saharan Africa

Abstract: Can contemporary female HIV rates be traced back to women’s position in ancestral societies ? In matrilineal kinship organizations, lineage and inheritance are traced through female members and children integrate the kin group of their mother rather than their father. A prediction that emerges from both the evolutionary psychology and the evolutionary anthropology literature is that matrilineal females should adopt more promiscuous sexual behaviours than their patrilineal counterparts, leading to higher HIV rates. Using within Sub-Saharan African countries variation in ethnic groups’ ancestral kinship organizations, I find that females originating from ancestrally matrilineal ethnic groups are today more likely to be infected by HIV. This finding is robust to the inclusion of subnational fixed effects, as well as a large set of cultural, historical, geographical, and environmental factors. I find consistent results using a number of alternative estimation strategies, including a geographic regression discontinuity design at ethnic boundaries and an instrumental variable strategy. Matrilineal females’ riskier sexual and contraceptive behaviours constitute the main explanatory mechanisms. Building an epidemiological model, I simulate how these differences in sexual and contraceptive behaviours translate into different gender-specific HIV rates dynamics.

7 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

Keywords: Kinship systems, matrilineality, cultural persistence, HIV, sexual behaviour, gender

JEL Classiﬁcation: D13, D91, I12, I15, J12, N37, Z13

8 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

1.1 Introduction

Every four minutes three young women become infected with HIV in the world. In 2018, the number of women in the world living with HIV amounted to 18.8 millions (UNAIDS, 2018b). With

14.5 millions of women living with HIV1, representing about 80% of the worldwide female HIV positive population, Sub-Saharan Africa is the most affected region in the world.2 It is the only place in the world where more women than men live with HIV: adult female HIV positive population was about 1.5 larger than male HIV positive population3, and adult females represented about 57% of new adult infections in 2017.4 This phenomenon has been referred to “Feminization of HIV” in Sub-

Saharan Africa. Widely-acknowledged, this phenomenon has recently raised researchers’ interest in understanding its long-term origins and subsequent variation across the continent (Anderson, 2018;

Bertocchi and Dimico, 2019).

In parallel, researchers are increasingly coming to understand the role of transmitted cultural traits on contemporary individuals’ behaviour (Giuliano and Nunn, 2019). Ancestral kinship organizations, which are fundamental social institutions determining group membership and social obligations, have been recently emphasized as deep-rooted determinants of individuals’ behaviour (Moscona, Nunn and

Robinson, 2020). For example, building on the anthropological literature, Lowes(2018a) provides experimental evidence that matrilineality undermines spousal cooperation within household. As she reports, while in patrilineal kinship organizations children integrate their father’s kin group and inheritance can only be passed on to children of male group members, in matrilineal kinship organizations group membership and inheritance are traced through female members. In other words, matrilineal males do not transmit to their biological children but to their sister’s children. Further, in patrilineal societies a wife is effectively incorporated into the lineage of her husband because she is not relevant to her kin group for determination of lineage or inheritance, thereby reducing her ability to rely on her own kin group in the case of separation or conﬂict. In matrilineal societies, husbands and wives maintain strong allegiances with their own kin group, allowing matrilineal wifes to beneﬁt from an improved outside option relative to their patrilineal counterparts, thereby reducing matrilineal men’s authority over their spouses (Lowes, 2018a).

111.2 millions in “Eastern and Southern Africa” and 3.3 millions in “West and Central Africa”. 2These numbers are from http://aidsinfo.unaids.org/ 39.5 millions males lived with HIV in Sub-Saharan Africa. 4Among the 1.02 millions newly infected adult individuals in Sub-Saharan Africa in 2017, 580,000 were females.

9 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

In this paper, I examine ancestral matrilineality as a deep-rooted determinant of contemporary

HIV prevalence among female populations in Sub-Saharan Africa. The latest theories from the evolutionary psychology literature suggest that gender-speciﬁc sexual preferences have ben shaped in the long-run by our ancestors reproductive success (Buss, 2016; Diamond, 1998). As such, producing large amounts of sperm, males exhibit on average a stronger preference for sexual variety than their female counterparts, who, producing limited number of ovula over their lifetime and bearing the larger share of the initial parental investment (i.e. gestating, bearing, nursing, nurturing and protecting a child) are seeking for a sexual mate that will be willing to commit in the long-term. However,

I argue that, relative to their patrilineal counterparts, matrilineal males have a weaker propensity to commit in the long-term to ensure the survival of their offsprings, since these latter will not integrate their own kin group. Therefore, ancestral matrilineal societies may have constituted specific contexts in which substituting long-term committed sexual relationships with sexual variety may have been a better sexual strategy for females’ reproductive sucess, allowing them to have access to better genes for their offsprings (“Better Genes Theories”, discussed in subsection 1.2.3). In the other way around, a theory from evolutionary anthropologists posits that the adoption of matrilineal kinship structures may have been an adaptative response to environments with low paternal certainty, and associated females’ sexual promiscuity. One hypothesis is that matrilineality would have emerged in environments with low paternal certainty because, while it is difficult to confirm paternity, maternity is easily observable. Thus, an inheritance system in which property passes from the mother’s brother to her sons may be optimal since the brother knows he is related to his sister, but cannot verify that he is related to his children (Fortunato, 2012). In the end, benefiting from a better marriage outside option as well as being inherently valued more relative to their patrilineal counterparts (Lowes, 2018a), I expect matrilineal females to benefit from a relatively greater sexual autonomy and ability to implement their own sexual preferences. Consequently, I test the hypothesis that ancestral matrilineality has shaped more promiscuous contemporary females’ sexual and contraceptive behaviour, and therefore regretfully drived higher HIV prevalence among matrilineal female populations.

To do so, I use a representative random sample of individuals reporting their ethnicity and who were tested for HIV across 18 Sub-Saharan African countries (in fact, when possible, I use several surveys for some countries, amounting to 32 country-Demographic Health Surveys, DHS). Linking individuals to their ancestral ethnic group in the Ethnographic Atlas, a worldwide anthropological database containing ethnographic information on cultural aspects and ways of life of ethnic groups

10 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa prior to industrilization and colonial contact, I estimate the probability of testing positive for HIV as a function of whether an individual originates from an ancestrally matrilineal ethnic group. I find a higher probability of testing positive for HIV for females originating from ancestrally matrilineal ethnic groups. Interestingly, and in line with my conceptual framework, I find that this effect is specific to female populations.

Concern remains that the variation in ethnic group’s ancestral kinship organizations exploited here may be correlated with important unobservables which are driving my main result, and has nothing to do with the channel of causality argued for in this paper. To address this concern, in my key identication strategy I exploit the variation in ethnic groups’ ancestral kinship organizations within subnational regions in Sub-Saharan African countries. This enables me to include subnational region- survey fixed effects into my estimations, and therefore control for a large set of potential national and subnational legal, institutional and economic confounding factors. Furthermore, the inclusion of individual-level controls allows me to account for differential socio-economic composition, in particular in education, marital behaviour and religion, which are plausible drivers of sexual and contraceptive behaviours (Case and Paxson, 2013). In addition, I include a large set of ethnic-group level historical controls, computed from the Ethnographic Atlas, which are aimed at capturing ancestral matrilineality’s ethnical covariates that may be other long-term confounding factors affecting contemporary gender norms and subsequent female’s sexual and contraceptive behaviour. Furthermore, exploiting village-level geographic informations provided in DHS, I also include a host of fine-grained geographical controls meant at capturing alternative geographical channels which have been shown to shape contemporary variation in HIV prevalence in Sub-Saharan Africa. Finally, I show that my main results are not driven by differential selection into HIV testing, nor reflect differences in general health status, but are specific to sexually transmitted diseases.

Despite the care provided in controlling for a large range of observables, I go one step further by formally testing for omitted bias, computing Altonji, Elder and Taber(2005) ratios and estimating Os- ter(2017) bias-adjusted lower bound coefﬁcients. These exercises provide no evidence of an omitted bias in my OLS estimates.

Reverse causality may remain a concern if matrilineality was an adaptative response in environments associated with more promiscuous sexual behaviours. In order to provide further support to the causal interpretation of my main individual-level OLS ﬁndings, I exploit data on the GPS location of

DHS villages as well as the digitized Murdock’s map of ancestral ethnic groups in Africa (Figure 1.2),

11 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa and I implement three alternative identiﬁcation strategies. These latter allow me to estimate the effect of the variation in cultural trait on individuals living in similar environments.

First, I conduct a geographic regression discontinuity (RD) design strategy at ethnic boundaries, exploiting the measure of distance that I computed between DHS villages and ancestrally matrilineal geographic areas. Restricting attention to individuals living in DHS villages located close to an ancestral matrilineal ethnic boundary, I estimate the effect of living in a village located on the matrilineal side of the ancestral border, while controlling for geographic location/distance running variables. I

ﬁnd that RD estimates are qualitatively similar to OLS (though smaller in magnitudes, due to presumably blurry ancestral ethnic boundaries and potential spillovers at the border), and highlight again ancestral matrilineality as a driver of contemporary female HIV rate.

Second, I implement an instrumental variable (IV) strategy, instrumenting individual’s ethnic group’s ancestral matrilineality with the distance between the location of DHS villages and the nearest ancestral matrilineal ethnic boundary, while controlling for a large array of potential confounders. My

IV-2SLS estimate provides additional support to my main OLS ﬁnding.

Finally, I estimate the average treatment over treated effect of being located on an ancestrally matrilineal geographic area on DHS villages’ proportion of HIV positive females, matching villages with their nearest neighbor in an ancestrally non-matrilineal geographic area based on a large array of geographic observables. Finding that villages located in ancestrally matrilineal areas have signiﬁcantly higher females HIV rates than their nearest neighbor in non-matrilineal areas, I provide additional support to my main OLS ﬁnding.

To explain such long-term effect of ancestral matrilineality on contemporary female HIV in Sub-

Saharan Africa, I provide evidence that, beneﬁting from a higher social status, bargaining power and subsequent higher sexual autonomy, matrilineal females adopt riskier sexual behaviour which are more conducive to HIV. I also show that women originating from ancestrally matrilineal ethnic groups are more likely to be HIV positive while having a HIV negative husband. This result is consistent with my conceptual framework and suggests extramarital channels of transmission of the virus.

In addition, I underline that matrilineal females’ higher contraception-related decision-making power translates into them substituting condom with long-term contraceptive methods. This means that, incidentally, matrilineal females substitute more protective contraceptive methods with less protective ones. Nevertheless, I also ﬁnd evidence that when they have internalized the risk of transmission of the virus, matrilineal individuals are more likely to adopt condom as a contraceptive method. This

12 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa last result calls for policies aiming at raising awareness about the risk associated with promiscuous behaviours, to ﬁght against the spread of HIV for this population at high risk. Finally, I provide evidence that discard differences in access to condom and in sexual debuts as other explaining mechanisms. Indeed, I ﬁnd that matrilineal females have in fact an easier access to condom and begin their sexual life later.

Finally, I build on the epidemiological literature (Tassier, 2013; Worden and Porco, 2017) to perform a numerical simulation exercise, aiming at illustrating how the differences between matrilineal and patrilineal populations in gender-specific sexual and contraceptive behaviours translate into different gender-specific HIV rates dynamics. Under credible parameter values, these simulations show dynamics that are consistent with my conceptual framework and my empirical findings.

This paper contributes to several strands of literature. It contributes ﬁrst to a recent literature exploring the long-term determinants of female HIV. Cagé and Rueda(Forthcoming) show that contemporary geographical variation in HIV across Sub-Saharan Africa was inﬂuenced by Protestant and

Catholic missions and their health investments in the early 20th century. They ﬁnd that regions close to historical mission stations exhibit higher HIV prevalence. They argue that this is likely because of a lower knowledge about condom use due to the persistent effect of conversion. However, they also ﬁnd that proximity to a mission with a health investment is associated with lower HIV prevalence nowa- days. They propose safer sexual behaviours as well as the persistence of health infrastructures around these missions as possible explanatory channel. Anderson(2018) explores the legal origins of female

HIV in Sub-Saharah Africa and shows that current HIV rates are higher for women living in common law countries, as compared to women living in civil law countries. She argues that female’s property rights being weaker in common law countries, women suffer from lower intrahousehold bargaining power and related ability to impose safe sexual practices to their husbands. Ultimately, these women are mainly infected by their husbands. Bertocchi and Dimico(2019) highlight historical slave trade as an other long-term determinant of female HIV in Sub-Saharan Africa. They expose that historical slave trade has fostered current polygynous practices, which are associated with more female’s marital dissatisfaction. Thus, they ﬁnd that females in polygynous union are more likely to adopt riskier sexual behaviour. In turn, this increases their likelihood to contract and transmit the virus, through the husband, to their faithful co-wives.

I show that contemporary variation in female HIV rate in Sub-Saharan Africa can be traced back to ancestral (i.e. before colonization) kinship organizations. I therefore explore a more deep-rooted

13 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa factor than explored so far in the literature. Further, I provide evidence that the long-term effect of ancestral matrilineality holds true within common law countries, and for non-polygynous couples. In addition, I highlight different routes of infection to explain the higher prevalence of HIV for matrilineal females, namely extramarital routes. However, I also show that matrilineal individuals are more likely to use condom when they have properly internalized the risk of transmission of the virus by being aware of their seropositive status. Finally, I draw new conclusions and policy recommendations. While female empowerment has been advocated recently in the literature as the main policy approach (Anderson, 2018; Bertocchi and Dimico, 2019), the ﬁndings of this paper call for information campaigns targeting empowered women and raising awareness about the risk associated with promiscuous sexual behaviours.

Contemporary determinants of the spread of HIV in Africa have also been emphasized. Oster

(2005) argues that Africa very high HIV transmission rates are likely due to high rates of other untreated sexually transmitted infections, and that within continent differences in HIV transmission rates can be attributed to differences in sexual behaviour and epidemic timing. In this paper, I explore sexual and contraceptive behaviours as main mechanisms explaining the higher rates of HIV found for matrilineal female populations. Oster(2012b) provides evidence of a fairly consistent positive relationship between exports and new HIV infections, suggesting that increased exports increase the movement of people (trucking), which increases sexual contacts. Corno and De Walque(2012) underline mine workers’ migration as a driver of HIV infection of both mine workers and their wives, showing that both of them are less likely to adopt safe sexual behaviours. More precisely, they find that miners are less likely to abstain and to use condoms, in particular during occasional sexual intercourse; while women with a miner as a partner are less likely to abstain, to be faithful or to use condom with their miner husband. I show in the paper that the long-term effect of ancestral matrilineality on contemporary female HIV is robust to these alternative channels. Case and Paxson(2013) explore the adverse role of girl’s education on female HIV in Africa and find that, by delaying teen marriage, increase in girls’ schooling in some regions triggered risky adolescent sexual behaviour, more conducive to HIV. Consequently, they find that regions that had higher rates of female education in 1980’s have higher HIV rates today. I provide evidence in the paper that age at sexual debuts can be discarded as an explanatory mechanism of the effect of ancestral matrilineality.

This paper further contributes to the literature aimed at understanding the inﬂuence of HIV risk perception on sexual and contraceptive behaviour. Oster(2012a) investigates the lack of behavioural

14 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa change despite the high prevalence of HIV in Sub-Saharan Africa. She ﬁnds reduction in risky sexual behaviour only in areas with higher life expectancy. Consistent with optimizing behaviour, she explains that high rates of non-HIV mortality suppress behavioural response. However, she does not

ﬁnd evidence of greater behavioural change in areas with higher knowledge of the epidemic. Björk- man Nyqvist, Corno, De Walque and Svensson(2018) provide experimental evidence that behavioural response from lottery incentives on HIV prevention are higher for risk-lover individuals, who are also individuals with ex-ante riskier sexual behaviour. Thornton(2008) provides experimental evidence that learning HIV positive status induces an increase in condom use, while learning HIV negative status does not affect condom use behaviour. Paula, Shapira and Todd(2014) show that downward revisions in the belief assigned to being HIV positive increase risky behaviour, while upward revisions decrease it. Similarly, Delavande and Kohler(2015) show that risky sexual behaviour is inﬂuenced by individuals’ expectations about survival and future HIV status. These latter in turn depend on the perceived impact of HIV on survival, on expectations about own and partner’s current HIV status, and expectations about HIV transmission rates. The result I provide in the paper on the heterogeneity in condom use by perception of the risk of transmitting the virus is fully in line with the results of this literature.

My paper also relates to the literature emphasizing cultural factors shaping contraception use, crucial for limiting the spread of HIV. In Sub-Saharan African context, Cordero-Coma and Breen

(2012) emphasize ﬁdelity norm and reproduction norm as two of the fundamental elements that guide spouses’ condom use behaviour. Focusing on rural Malawi, Chimbiri(2007) documents that condom use is negligible inside marriage, and that initiating a discussion about condom use for preventing infection in marriage is like bringing an “intruder” into the domestic space. In the context of Bangladesh,

Islam, Islam and Banowary(2009) underline that matrilineal Garo women contraceptive behaviour differs from their patrilineal Bengali counterparts, in that their current use of contraceptives is higher than at national level, but their use of condom is lower than at national level. I ﬁnd a very similar result in the paper when I explore differences in contraceptive methods used by matrilineal and patrilineal individuals.

More broadly, this papers speaks to the literature exploring the long-term determinants of contemporary gender outcomes, which can be found summarized in Giuliano(2017). Alesina, Giuliano and Nunn(2013) ﬁnd that descendants of societies that traditionally practiced plough agriculture, characterized by gender division of labor, today have less equal gender norms. Alesina, Brioschi and

15 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

Ferrara(2016) provide evidence that pre-colonial customs about marriage patterns, living arrangements and the productive role of women are associated with contemporary violence against women.

Becker(2018) highlights that pre-industrial economic production relying on pastoralism favored the adoption of customs aimed at constraining female sexuality, such as female genital cutting and restric- tions on women’s mobility. Teso(2019) shows that women whose ancestors were more exposed to the transatlantic slave trade, which leaded to female-biased sex ratio during centuries in Sub-Saharan

Africa, are today more likely to be in the labor force, have lower levels of fertility, and are more likely to participate in household decisions. In the paper, I show that women originating from ancestrally matrilineal ethnic groups in Sub-Saharan Africa beneﬁt from higher marriage outside option, intrahousehold bargaining power and sexual autonomy.

Finally, this paper also links directly to the burgeoning literature assessing the inﬂuence of ancestral norms on women’s well-being. Lowes(2018a) provides evidence from Democratic Republic of

Congo that men and women from matrilineal ethnic groups cooperate less with their spouse in a lab experiment. She also ﬁnds that children of matrilineal women are healthier and better educated, and that matrilineal women experience less domestic violence. Ashraf, Bau, Nunn and Voena(2020b) highlight the positive role of bride price custom on female education in Indonesia and Zambia. In the same vein, Bargain, Loper and Ziparo(2020) show the complementarity between women’s access to judicial institutions and traditional matrilocality on women’s empowerment and well-being in Indone- sia. In this paper, I show that ancestral matrilineality has an ambiguous effect on women’s well-being.

Indeed, although I ﬁnd that matrilineal women have a higher sexual freedom, I also ﬁnd that they are more at risk of being HIV positive.

The remainder of the paper is organized as follows. In section 1.2, I provide an overview of ancestral matrilineality as well as contemporary female HIV in Sub-Saharan Africa, and I expose my conceptual framework. Then, I provide a description of the data and I discuss my empirical strategy in section 1.3. My main results, robustness checks and alternative identiﬁcation strategies are presented in section 1.4. I then explore the mechanisms in section 1.5, and I conduct an epidemiological simulation exercise in section 1.6. Finally, I provide concluding thoughts in section 1.7.

16 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

1.2 Context and Conceptual Framework

1.2.1 Ancestral Matrilineality in Sub-Saharan Africa

Kinship relations are important in the context of Sub-Saharan Africa, where kinship groups form a basic political unit in which members recognize each other as kin and often have certain obligations toward each other, such as land sharing, contribution to bride price payments for lineage members, provision of ﬁnancial support (school fees, burial payments, etc.) (Fox, 1934; Lowes, 2018a). In matrilineal kinship system, individuals trace lineage and descent through women. As such, while bio- logically related to family of both their mother’s side and their father’s side, individuals are considered kin only if they share a common female ancestor (Lowes, 2018a).

Figure 1.1 (a) is from Lowes(2018a) and illustrates the structure of matrilineal kinship systems.

As she explains: “In the diagram, men are represented by triangles and women are represented by circles. Membership in the same matrilineal group is denoted with red. Children are in the same matrilineal group as their mothers. Likewise, a mother is in the same matrilineal group as her male and female siblings. In many matrilineal societies, the mother’s brother has an important role relative to his sister’s children. His inheritance and lineage will be traced through his sister’s children, and he has obligation to ﬁnancially support her children. Importantly, husband and wife do not share the same lineage”. Even once married, the wife remains in her lineage of origin: for all married couples one spouse is red and the other spouse is blue. Consequently, as noted by anthropologists, “husbands are less able to mistreat their spouses in matrilineal systems since these latter have greater support from their kin groups.” (Lowes, 2018a).

Figure 1.1 (b) is also from Lowes(2018a) and presents the structure of patrilineal kinship. As she explains: “Now, children are in the same group as their father, as denoted in blue. In a patrilineal society, rather than maintaining strong ties with her own lineage, a woman is effectively incorporated into the lineage of her husband upon marriage. This is because once she is married, she is not relevant for determining descent and inheritance for her lineage. This is illustrated in the patrilineal kinship diagram by the married women denoted in grey, while the unmarried daughter shares the same color as her father.” (Lowes, 2018a).

Historically, matrilineal kinship systems are correlated with other cultural traits, which have been shown to have long-term impact on gender roles and economic development.5 For example, Lowes

5Interested reader may find an exhaustive overview of the origins of matrilineal kinship systems andrelated women’s empowerment in Lowes(2018a) appendix.

17 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

Figure 1.1 – Diagrams of kinship systems (source: Lowes, 2018a)

(2018a) shows that, in Africa6, matrilineality is negatelively correlated with the practice of bride price, the use of the plough as well as animal husbandry. As reported in Table 1.C.1 in appendix, I

ﬁnd similar evidence looking at my main sample’s descriptive statistics. Therefore, as described in subsection 1.3.3, it will be crucial to control for all these historical correlates.

15 percent of the 527 Sub-Saharan societies recorded in the Ethnographic Atlas are matrilineal; while 70 percent are patrilineal. Furthermore, as noted in Lowes(2018a), the vast majority of these matrilineal societies are distributed across the center of Africa in the so called “Matrilineal

Belt”; which intersects present day Angola, Republic of Congo, DRC, Gabon, Malawi, Mozambique,

Namibia, Tanzania and Zambia. Figure 1.2 illustrates the matrilineal belt across Africa, with matrilineal groups denoted in blue, patrilineal groups denoted in green, and bilateral and other groups in beige7. In my ﬁnal sample, about 67% of individuals originating from an ancestrally matrilineal ethnic groups are indeed living in one of the countries crossed by the matrilineal belt. Interestingly, the highest female HIV rates in Sub-Saharan Africa are also found in the countries crossed by the matrilineal belt. In fact, ones talk about an “AIDS Belt”. I describe it in more detail in the next subsection.

1.2.2 Female HIV in Sub-Saharan Africa

In 2018, 14.5 millions of women lived with HIV in Sub-Saharan Africa, representing about 80% of the worldwide female HIV positive population (UNAIDS, 2018b). Uniquely, it is the only place in

6Using data from the Ethnographic Atlas. 7This map is based on Murdock’s map of ethnic group’s ancestral boundaries. Note that Murdock’s classification of ethnic groups slightly differs from the Ethnographic Atlas (EA)’s classification. Ithereforeuse Michalopoulos, Putterman and Weil(2019) and Teso(2019) mappings of Murdock-EA ethnic groups.

18 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

Figure 1.2 – Ancestral Ethnic Group Boundaries and Matrilineal Belt

the world where more women than men live with HIV: adult female HIV positive population is about

1.5 larger than male HIV positive population8, and adult females represented about 57% of new adult infections in 2017.9 This phenomenon has been referred to “Feminization of HIV” in Sub-Saharan

Africa.10

Interestingly, these numbers hide substantial within Sub-Saharan Africa variation in female HIV rates. Indeed, among the 14.5 millions of women living with HIV in Sub-Saharan Africa in 2018, 11.2 millions lived in “Eastern and Southern Africa”, against 3.3 millions in “West and Central Africa”.

The geographical variation in female HIV rates in Sub-Saharan Africa, computed from observations in my ﬁnal sample using DHS data and linking individuals to their ancestral ethnicity, is represented

89.5 millions males live with HIV in Sub-Saharan Africa. 9Among the 1.02 millions newly infected adult individuals in Sub-Saharan Africa in 2017, 580,000 were females. 10These numbers are from http://aidsinfo.unaids.org/

19 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

in Figure 1.311 and shows high prevalence rates in Eastern and Southern Africa, the so-called “AIDS

Belt”.12 Zambia and Malawi, two of the countries with highest female HIV rates are also two of the

countries with the highest proportions of individuals originating from matrilineal ethnic groups.13

As detailed in section 1.1, several factors have been emphasized to explain variation in female

HIV rates within Sub-Saharan Africa. Institutions have highlighted gender inequality and disempow-

erment as key barriers to progress against the HIV epidemic (UNAIDS, 2018a), and urged for effort to

address these issues. A widespread conjecture is that strengthening women’s property and inheritance

rights will prevent the spread of HIV/AIDS by promoting women’s economic security and empower-

ment. Anderson(2018) provides the ﬁrst empirical evidence of a causal relationship between female

bargaining power and female HIV infection rates in Sub-Saharan Africa. Exploiting variation in legal

origins of Sub-Saharan countries, she ﬁnds that HIV prevalence is today higher for women living in

common law countries, where code of law is associated with weaker female property rights, as com-

pared to women living in civil law countries. She proposes women’s lower intrahousehold barganing

power and ability to impose safe sexual pratices to their husband as main explanatory mechanisms.

However, this mechanism does not explain the geographical correlation between the so-called

“Matrilineal Belt” and the so-called “AIDS Belt” in Africa: the highest rates of female HIV can be

found in ancestrally matrilineal geographical areas, correlated with higher women’s status and sexual

autonomy. I explore this puzzle in this paper. To do so, I begin with a description of my conceptual

framework in the next subsection.

1.2.3 Conceptual Framework

1.2.3.1 Gender Differences in Sexual Strategies

Humans, like other sexually reproducing species, do not choose mate randomly. According to the

latest theories from evolved psychologists, our mating is strategic, and the sexual strategies we de-

veloped were shaped in the (very) long run by natural selection through our ability to survive and

11This map is based on Murdock’s map of ethnic group’s ancestral boundaries. Note that Murdock’s classification of ethnic groups slightly differs from the Ethnographic Atlas (EA)’s classification. Ithereforeuse Michalopoulos et al.(2019) and Teso(2019) mappings of Murdock-EA ethnic groups. 12Interested reader might have a look at https://www.hsph.harvard. edu/news/magazine/spr08circumcisionmap/; and/or https://www.prb.org/ thestatusofthehivaidsepidemicinsubsaharanafrica/ 13In my final sample of matrilineal versus patrilineal females, proportions of matrilineal females areabout 87% in Malawi and about 60% in Zambia; female HIV rates are about 12% in Malawi and about 14% in Zambia. These numbers are computed using the following DHS surveys: Malawi (2004, 2010, 2014) and Zambia (2007, 2013). See subsection 1.3.1 for more details. This compares to the proportion of about 15% of matrilineal females, and the average female HIV rate of about 5% in my full final sample of matrilineal versus patrilineal females.

20 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

Figure 1.3 – Ancestral Ethnic Group Boundaries and Contemporary Female HIV Rates (Final sample)

reproduce successfully. As Buss(2016) states: “Those in our evolutionary past who failed to mate successfully failed to become our ancestors. All of us descend from a long and unbroken line of ancestors who competed successfully for desirable mates, attracted mates who were reproductively valuable, retained mates long enough to reproduce, fended off interested rivals, and solved the prob- lems that could have impeded reproductive success. We carry in us the sexual legacy of those success stories.” Further, underlying each sexual strategy14 are psychological adaptations (e.g. preferences for a mate, feelings of love, desire for sex, sexual jealousy, etc.), that are sensitive to the information or cues from the external world.

These theories provide an interesting framework to understand why males and females exhibit on average marked different sexual preferences. According to numerous psychological studies15, males

14“Sexual strategies do not require conscious planning or awareness. [...] Indeed, just as pianist’s sudden awarenes of her hands may impede performance, most human sexual strategies are most successfully carried out without the awareness of the actor.” (Buss, 2016) 15See Buss(2016) for an extensive review.

21 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa exhibit a stronger preference for casual relationships and sexual variety; while females tend to be more choosy in their mate as well as more looking for long-term committed relationships. Indeed, producing million of sperms, wich are replenished at a rate of roughly 12 millions per hour (Buss, 2016), males can expect at most to reproduce as many times as there are available fertile females willing to have sex. Additionally, males only bear minimal initial parental investment in case of fecondation.

Consequently, preference for sexual variety and casual sexual relationships is a psychological trait that has been relatively more developed in male evolved sexual psychology. On the contrary, producing only a ﬁxed limited amount of reproductive cells (i.e. only 400 ova mature to the point where they are capable of being fertilized), females face a limited number of reproductive opportunities. In addition, bearing the larger share of the initial parental investment (i.e. gestating, bearing, nursing, nurturing and protecting a child), females provide extraordinarily valuable, but limited, reproductive resources. Therefore, because females in our evolutionary past risked enormous investment from having sex, evolution favored females who were highly selective about their mates. In particular, evolution favored females capable of reproducing with long-term committed and reliable provider mates, ensuring the survival and reproduction success of their offsprings (Diamond, 1998).

Nevertheless, while the reproductive beneﬁts of casual relationships for males are large and direct, women may also reap beneﬁts from short-term mating, according to the latests evolutionary psychology theories. As a matter of fact, even though having access to more sperm would not increase a woman’s reproductive success, through casual sex it is possible for a woman to gain superior genes that are passed on to their children. This is the so-called “Better Genes Theory”16 (Buss, 2016; Greil- ing and Buss, 2000). As such, a women might try to secure the investment of a lower-ranking man by marrying him, for example, while simultaneously securing the genes of a higher-ranking man by mating with him casually (the maing marketplace rending far easier for a woman to get a men with better genes to have sex with her than to get him to marry her) (Buss, 2016). One version of this “Better

Genes Theory” has been called the “Sexy Son Hypothesis”, according to which women prefer to have casual sex with men who are attractive to other women because they will have sons who possess the same charming characteristics and therefore will enjoy greater mating success in the next generation.

However, adopting such extra-pair mating strategies may come at a cost. In particular, an unfaith- ful married woman risks the withdrawal of resources by her husband, reputational damages as well as getting pregnant and bearing a child without the beneﬁts of an investing partner. To sum up, short-

16I refer reader interested in alternative hypothesis on women’s extra-pair mating function to Buss(2016) and Greiling and Buss(2000), which provide extensive discussions.

22 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa term mating poses hazards, but has powerful beneﬁts as well, and women have evolved psychological mechanisms to select circumstances in which the cost of short-term mating are minimized and the beneﬁts increased.

Matrilineal kinship organizations, relative to patrilineal ones, constitute environments in which extra-pair mating sexual relationships should be beneficial for women. Indeed, as detailed in subsection 1.2.1, in matrilineal kinship organizations males’ biological children do not integrate their lineage but the lineage of their mother. In addition, matrilineal males do not transmit their wealth to their biological children but to their sisters’ children. In other words, females in matrilineal societies should have lower expectations regarding their mate’s propensity to engage in long-term committed relationships to maximize the reproductive success of their offsprings. Consequently, for their reproductive success they should expect relatively greater benefits from gaining access to better genes, through extra-pairing mating. Furthermore, contrary to patrilineal societies where females integrate their husband’s lineage once married, matrilineal females remain in their lineage of origin after marriage and keep strong bounds with their family. In other words, matrilineal females benefit from higher marriage outside option, relative to their patrilineal counterparts, but also from relatively lower opportunity cost associated with extra-pair mating sexual strategy. Finally, being inherently valued more through their position in their societies, matrilineal females should also benefit from a greater sexual autonomy than their patrilineal counterparts, and a greater ability to pursue their preferred sexual strategy.

If, according to my conceptual framework, ancestral matrilineality fostered females to seek for better genes through sexual variety, matrilineality should be correlated with greater genetic diversity as well as father diversification. This is exactly what I find in Table 1.C.2 in appendix. As reported in the two first columns, I find a positive country-level association between the estimated proportion of countries’ citizens with ancestors that had matrilineal inheritence rule (using data from Giuliano and Nunn, 2018) and countries’ indexes of genetic diversity (computed by Ashraf and Galor, 2013b), robust when including continent fixed effects as well as a large array of controls, including distance to Addis Ababa, the cradle of humankind. Further, as reported in the two last columns, running individual-level regressions following my main identification strategy on my final sample from DHS

(described in more detail in subsection 1.3.1), I ﬁnd that women originating from ancestrally matrilineal ethnic groups are signiﬁcantly more likely to diversify the fathers of their biological children.

Alternatively, the adoption of matrilineal kinship structures may have been an adaptative response to environments with low paternal certainty. Indeed, evolutionary anthropologists explain the exis-

23 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa tence of matrilineal societies as the result of an evolutionary process that created institutions suitable for the ecological and social environment (Lowes, 2018a). One hypothesis is that matrilineality would have emerged in environments with low paternal certainty because, while it is difﬁcult to conﬁrm paternity, maternity is easily observable. Thus, an inheritance system in which property passes from the mother’s brother to her sons may be optimal since the brother knows he is related to his sister, but cannot verify that he is related to his children (Fortunato, 2012). However, as noted by Lowes

(2018a), this model alone would require that paternity certainty be below .268, a value that is unrealis- tically low even for matrilineal societies. Still, developping a more complex model Holden, Sear and

Mace(2003) argue that daughter-biased investment (leading to matrilineality) may be adaptive when the marginal beneﬁt of investing in sons (relative to daughters) is not sufﬁcient to offset by the risk of non-paternity of the son’s children.17 In sum, ancestral matrilineality would be associated with ancestrally higher paternal uncertainty and more promiscuous sexual behaviour and, ultimately, with more promiscuous contemporaneous sexual behaviours.

All in all, my conceptual framework points toward a main prediction, that I empirically test in this paper: I expect to ﬁnd that, as of today, matrilineal females adopt sexual behaviours that are more promiscuous than their patrilineal counterparts, and therefore suffer from more HIV.

1.2.3.2 Women’s Empowerment and Contraceptive Use

As discussed above, I expect matrilineal women to be more empowered than their patrilineal counterparts, which may also affect their contraceptive behaviour. Indeed, when it comes to the choice of contraceptive use within a couple, intra-household bargaining matters. As such, and because men’s preferences may constrain household adoption (Ashraf, Field, Voena and Ziparo, 2020a), women’s empowerment acts a driver of contraceptive use (Anderson, 2018, Cassidy, Groot Bruin- derink, Janssens and Morsink, 2018, Kibira, Ndugga, Nansubuga, Sewannonda and Kwagala, 2014).

However, what mainly matters to restrain the spread of HIV is not the use of contraceptives per se,

17The evolutionary anthropology literature provides alternative hypothesis on the ecological origins of matrilineality. Aberle(1961) posits that matrilineality would be more beneficial with certain types of production such as hoe agriculture, while patrilineality would be more beneficial with hunting, which requires skill development and male cooperation. Holden and Mace(2003) argue that with the rise of moveable heritable wealth, such as cows, the marginal benefits of investing in sons increases, leading to the demise of matrilineal societies. Pastoralism, and the spread of cattle in particular, would have therefore led to the loss of matrilineality against the adoption of patrilineality and mixed descent rules. BenYishay, Grosjean and Vecci (2017) find that coral reef density predicts the prevalence of matrilineality in the Solomon Islands. Further, they find that reef density explains as much as 10% of the variation in inheritance rules across villagesinthe Solomon Islands, suggesting an adaptation of institutions to ecological conditions. Interested reader may find a more exhaustive overview of the literature on the origins of matrilineal kinship systems in Lowes(2018a) appendix.

24 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa but the type of contraceptive technology used. This latter is also affected by women’s empowerment.

Indeed, bearing the responsability of contraception and reproduction, more empowered women may be more likely to substitute short-term contraceptive methods such as male condom with long-term contraceptive methods (e.g. intra-uterine device (IUD), female sterilization, implants, pills or injec- tions) giving them long-term control over their fertility (Islam et al., 2009, Palamuleni and Adebowale,

2014, Samari, 2018).

As detailed in subsection 1.2.1 and reported in Lowes(2018a), due to their central position in their kinship structure and the support they receive from their relatives, matrilineal women have a higher social status than their patrilineal counterparts. Therefore, I expect them to to beneﬁt from a higher intra-household bargaining power, and to be consequently more likely to substitute male condom with long-term contraceptive methods. By doing so, they would substitute more protective contraceptive methods with less protective ones, which would contribute to explain why they suffer from more HIV.

1.3 Data and Empirical Strategy

To study the long-term impact of ancestral matrilineality on female HIV, I match contemporary individual-level data from the Demographic and Health Surveys (DHS) with historical ethnic group- level data from the Ethnographic Atlas (EA). This section describes the data and the empirical strategy.

1.3.1 Contemporary Data

Data on HIV infection come from the DHS, which have been conducted in sub-Saharan African countries since the 1990’s. The DHS household surveys typically interview a nationally representative sample of between 10,000 to 20,000 women (aged 15-49) and men (aged 15-59). By collecting blood for HIV testing from representative samples of the population, the DHS Program provides nationally representative estimates of HIV prevalence rates. The testing is simple: the interviewer collects dried blood spots (DBS) on filter paper from a finger prick and the filter paper is transported to a laboratory for testing. The testing is anonymous, voluntary, and non-informative to respondents. The average response rate is extremely high; 93 percent for women (slightly lower for men).

I restrict my sample to DHS surveys containing both HIV testing and individual ethnicity information, as well as GPS data. I further restrict my sample to individuals originating from ethnic groups which are either ancestrally matrilineal, or patrilineal. This leaves me with a sample of 159,560

25 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa women across 18 Sub-Saharan African countries (i.e. 32 country-surveys18). The proportion of men tested for HIV is lower and I have a sample of 120,580 men. As a main outcome, I construct an individual-level indicator variable, HIV, that takes value one if the respondent is seropositive according to the DHS HIV Test.

On average, 4.8 percent of women in my initial sample are HIV positive (this compares to 3.1 percent of men). The average HIV infection rate of women originating from ancestrally matrilineal ethnic groups is approximately 11.4 percent. It is close to one-fourth, at 3.7 percent, for women originating from ancestrally patrilineal ethnic groups. This compares to 7.8 percent for matrilineal males vs. 2.1 percent for patrilineal males. Though only correlational, these numbers provide a ﬁrst evidence of the higher prevalence of HIV among individuals originating from ancestrally matrilineal group, women in particular.

Further, DHS surveys present questions that are useful measures of female sexual autononomy, actual sexual behaviour, contraception use, acknowledgment about HIV risks and protective methods, and sexual debuts, which I investigate in section 1.6.

Finally, I also exploit information on geographic covariates, computed at the village level from numerous data sources, and provided by the DHS. Additionally, I exploit data on the location of large- scale active mines in Africa, provided by S&P Global Market Intelligence19 and used in Berman,

Couttenier and Girard(2019), and I compute measures of presence of active mines near DHS villages.

Village-level geographic controls are described in subsection 1.3.3.

1.3.2 Historical Data

Data on pre-colonial ethnic groups’ traits come from the Ethnographic Atlas (EA, 1967), a worldwide ethnicity-level database constructed by George Peter Murdock. This database covers 1,265 ethnic groups in the world20, and contains detailed ethnographic information on cultural aspects and ways of life of the portrayed ethnic groups, prior to industrialization and colonial contact21, such as

18Burkina-Faso (2003, 2010); Cameroon (2004, 2011); Chad (2014); Congo Democratic Republic (2007, 2013); Ethiopia (2005, 2011, 2016); Gabon (2012); Ghana (2003, 2014); Guinea (2005, 2012); Ivory Coast (2011); Kenya (2003, 2008); Liberia (2013); Malawi (2004, 2010, 2014); Mali (2006, 2012); Senegal (2005, 2010); Sierra Leone (2008, 2013); Togo (2013); Uganda (2011); Zambia (2007, 2013). 19Interested reader may find their website here: https://www.spglobal.com/marketintelligence/en/ campaigns/metals-mining 20The majority of the ethnicities sampled are in Africa. 21“For the parts of the world without a written history, the information is from the earliest observers of these cultures. For some cultures, the first recorded information is from the 20th century, but even for these cultures, the data capture as much as possible the characteristics of the ethnic group prior to European contact.” (Alesina et al., 2013)

26 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa kinship and family organization, settlement patterns, political organization, institutional complexity, historical mode of subsistence, etc.

To match individual-level contemporary data with historical data, I use information provided in

DHS on individual’s ethnicity. However, the classification of respondents’ ethnic groups in DHS does not always coincide with the Ethnographic Atlas’ classification, requiring a matching between the two datasets.22 Therefore, I first follow Michalopoulos et al.(2019) matching, which enables me to do the matching for most of the individuals in my sample. I then follow Teso(2019) matching 23 to match some remaining individuals. Finally, I build on online sources24 to match ethnic groups not matched in previous procedures. I end up with a sample of 349,895 individuals in DHS matched with an ancestral ethnic group in the Ethnographic Atlas (EA).

I then discard individuals originating from ancestral ethnic groups with inheritance rule and kinship organization which were neither matrilineal nor patrilineal (such as bilaterality of ambilinearity), or with missing information in the EA, representing 69,755 out of 349,895 individuals in my sample.

I therefore restrict my sample to individuals originating from an either ancestrally matrilineal or patrilineal ethnic group, and I end up with a sample of 280,140 individuals (159,560 women and 120,580 men).

Finally, I construct my main explanatory variable, Matrilineality, as an indicator of ethnic-group’s ancestral matrilineality, from EA information on ethnic group’s inheritance rule. In my analysis, I additionally use a wide array of historical controls varying at the ethnic group level and computed from the EA. I describe these controls in the next subsection.

1.3.3 OLS Empirical Strategy

I explore the gender-speciﬁc long term effect of ancestral matrilineality on HIV by estimating the following equation:

22For many of the groups, the matching is straightforward as the name used in the DHS is the same or very similar to the one used in the Ethnographic Atlas (EA). When the name of an ethnic group in DHS is not found in EA’s classification, this is typically because an alternative group’s name isused. 23Teso(2019) matching procedure largely builds on Michalopoulos et al.(2019) matching procedure. 24One of the most useful sources of information on alternative ethnic groups’ names is the Joshua Project website (http://www.joshuaproject.net/). For most of the unmatched ethnicities, the respondent lists her nationality as ethnicity.

27 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

푦푖푒푣푟푡 = 훼 + 훽1푀푎푡푟푖푙푖푛푒푎푙푖푡푦푒 + 훽2퐹 푒푚푎푙푒푖 + 훽3퐹 푒푚푎푙푒푖 × 푀푎푡푟푖푙푖푛푒푎푙푖푡푦푒+

′ ′ ′ XievrtΔ + XertΩ + XvrtΠ + 휆푟푡 + 휀푖푒푣푟푡 (1.1)

with 푦푖푒푣푟푡 denoting an individual-level outcome (for example an indicator for whether the individual is HIV positive) for individual 푖 from ethnic group 푒, living in village 푣 in within-country DHS region 푟, and surveyed at year 푡. 푀푎푡푟푖푙푖푛푒푎푙푖푡푦푒 is an indicator for whether an individual originates from an ancestrally matrilineal ethnic group (an ancestrally patrilineal ethnic group otherwise);

퐹 푒푚푎푙푒푖 is an indicator for whether an individual is a female. 훽1 is intended to capture a “Matrilineal

25 effect”; and 훽2 is intended to capture a “Gender effect” on HIV. 훽3 is my main coefﬁcient of interest and captures the effect of ancestral matrilineality on female HIV once the “Matrilineal effect” and the

“Gender effect” have been controlled for. The inclusion of this interaction term is motivated by my conceptual framework, according to which originating from an ethnic group with an ancestral matrilineal kinship organization should significantly influence contemporary sexual behaviour for female individuals specifically. This hypothesis is further motivated by the first descriptive statistics reported in subsection 1.3.1 on heterogeneity of contemporary HIV rates by gender and individual’s ethnic group’s ancestral kinship organization.

′ Xievrt represents a set of individual-level covariates which includes indicators of marital status; a dummy for whether an individual is in a polygynous union; number of children; age; age squared; a dummy for whether an individual lives in an urban location; education in number of years; a dummy for whether an individual is currently working; wealth index indicators and religion indicators.

′ Xert represents a set of ethnic group-level ancestral covariates which includes intensity of women’s historical participation in agriculture; ancestral polygyny; ancestral bride price; ancestral plough use; ancestral pastoralism; ancestral presence of clans; indicators of ancestral settlement patterns; indicators of ancestral juridictional hierarchies beyond local communities; ancestral reliance on hunting; ancestral reliance on ﬁshing; ancestral reliance on gathering; ancestral reliance on animal husbandry; ancestral reliance on agriculture; ancestral presence of large domesticated animals; indicators of intensity of ancestral agriculture; and year of observation of the ethnic group in the Ethnographic Atlas.

25According to the medical literature, females are more likely than males to be infected when exposed to HIV, due to physiological and anatomical reasons, such as larger surface area of mucosal HIV exposure (Yi, Shannon, Prodger, McKinnon and Kaul, 2013).

28 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

′ Xvrt represents a set of village-level geographic covariates which includes latitude; longitude; altitude; nightlight composite; population density (2010); distance to lake or coastline; distance to nearest international border; average time (minutes) required to reach a high-density urban center

(2015); malaria incidence (2010); vegetation index; indicators for the length of the growing season; distance to the nearest active mine; a dummy indicating the presence of an active mine within 1000 km; an index of ethnic fractionalization; and an index of ethnic polarization. Summary statistics of all my control variables are reported in Table 1.C.1 in appendix.

Crucially, I also include (within-country) DHS region-survey (year) ﬁxed effects, 휆푟푡, to take into account time trends, as well as unobserved country-level and within-country level institutional, economic and geographic factors that could potentially affect contemporary HIV prevalence and also be correlated with the geographical distribution of ancestral matrilineality. Doing so allows me to assess the inﬂuence of matrilineal versus patrilineal ancestral kinship organizations on individuals located in the same institutional, economic, and geographic environment. This is the epidemiological approach of the cultural economics literature (Fernández, 2011).

Finally, since variation in the main explanatory variable occurs at the ethnic group level, observations of outcomes of individuals of the same ethnic group may not be independent. Consequently, in order to account for potential within-group correlation of the residuals (휀푖푒푣푟푡), throughout the analysis standard errors are clustered at the ethnic group level.

A crucial concern for the causal interpretation of the OLS estimates is the possible presence of omitted variables that are correlated with both contemporary HIV prevalence and with ancestral matrilineality of ethnic groups. For instance, if ancestrally matrilineal ethnic groups were more likely to have social organizations and institutions26, as well as modes of production more conducive of

27 equal gender norm or of the spread of the virus, this would translate in an estimate of 훽3 that is biased upward. The ethnicity-level controls are meant to alleviate these concerns. Additionally, I include ethnic group’s year of observation in Ethnographic Atlas to alleviate the concern that some groups were portrayed later than others and might therefore have been more developed, and hence potentially more gender equal, for example.

Likewise, 훽3 might be biased upward if ancestral matrilineality was correlated with geographic factors that are conducive of HIV. For example, geographic characteristics such as the type of vege-

26In this vein, Ashraf et al.(2020b) document the positive role of ethnic groups’ practice of bride price on female education. 27In this vein, Alesina et al.(2013) document the negative role of ancestral plough use on contemporary female empowerment.

29 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa tation or the altitude may be correlated with the spread of the virus. I directly control for that, and

I also control for malaria incidence, a proxy measure of virus spread. Further, following Corno and

De Walque(2012), mine workers’ international migration is an other driver of the spread of HIV. I therefore control for the village’s distance to the nearest active mine, as well as for the presence of an active mine within 1000km. In the same vein, Oster(2012b) highlights exports and road networks, and subsequent increase in movements of people and sexual contacts, as an other factor of HIV infection. Controlling notably for distance to international borders as well as average time to reach a high-density urban center as a proxy allows me to alleviate these concerns. An other long-term determinant of female HIV in Sub-Saharan Africa put forward by Cagé and Rueda(Forthcoming) is the geography of Protestant and Catholic missions in the early 20th century, as well as their health investments. In the same vein, Teso(2019) shows the long term effect of the slave trade in Sub-Saharan

Africa on contemporary gender norms. Again, my OLS estimates would be biased if ancestral matrilineality was correlated with such factors. The inclusion of numerous geographic covariates, such as latitude, longitude as well as distance to lake or coastline and average time to reach a high-density urban center, is meant to alleviate these concerns.

Along the same lines, 훽3 might also be biased upward if ancestral matrilineality was correlated with ethnic fractionalization and/or ethnic polarization. Indeed, Tequame(2012) provides evidence that these latter are associated with higher information asymmetry and lower social sanctions within communities, which are more conducive to risky sexual behaviour and therefore HIV. Computing indexes of ethnic fractionalization and polarization at village level and including them in my regressions is aimed at controlling for such potential omitted bias.

Ethnic groups’ economic prosperity may be an other possible omitted variable, potentially correlated with both ancestral matrilineality and contemporary HIV prevalance. To account for this, I control for village’s population density (2010) and nightlight composite, two proxies of economic prosperity. The inclusion of historical ethnic groups’ modes of production and institutional controls are also meant to capture historical economic prosperity. Again, including ethnic group’s year of observation as a control allows me to alleviate the concern that some groups were portrayed later than others and might therefore have been more developed. Related to this concern, ethnic groups’ access to contemporary health infrastructures could be an other possible omitted variable: controlling notably for average time (minutes) required to reach a high-density urban center (2015) is meant to control for this.

30 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

Finally, institutional factors may also drive female HIV. Anderson(2018) underlines that legal origins of Sub-Saharan African countries signiﬁcantly determine current female HIV rates: she namely

finds that female HIV rates are significantly higher in common law countries, relative to civil law countries. The inclusion of (within-country) DHS region x survey (year) fixed effects is notably meant to capture such country (and within-country) institutional factors, in addition of time trends.

1.4 The Long-Term Effect of Ancestral Matrilineality on Female HIV

1.4.1 Main Results - OLS Estimates

Table 1.1 presents the OLS estimates of the effect of ancestral matrilineality on contemporary female HIV, once gender and ancestral matrilineality have been controlled for. While in column 1 I do not include any control, I include (within-country) region-survey (year) ﬁxed effects in column

2. The coefficient of the interacted variable 퐹 푒푚푎푙푒 × 푀푎푡푟푖푙푖푛푒푎푙푖푡푦 is positive and statistically significant (row 3), and is not affected by the inclusion of region-survey fixed effects. Further, following the inclusion, of individual controls to account for differences in socio-demographic composition of matrilineal and patrilineal individuals, the main estimate of interest remains unchanged in column

3. To alleviate omitted variable concerns detailed in subsection 1.3.3, I subsequently include ethnic group’s controls in my regressions (column 4), and village level geographic controls (column 5). The coefﬁcient of the interacted variable 퐹 푒푚푎푙푒 × 푀푎푡푟푖푙푖푛푒푎푙푖푡푦 remains very consistent and of large magnitude across the speciﬁcations. Women originating from an ancestrally matrilineal ethnic group are 1.6 to 2.2 percentage points more likely to be HIV positive than their patrilineal counterparts. This effect is large in magnitude, corresponding to 70% to 105% of the average HIV prevalence (2.1% to

2.3%) among patrilineal males (control group) in my full regression sample. This result supports the main prediction of my conceptual framework.

Interestingly, I also ﬁnd a consistent and statistically signiﬁcant positive estimate of being a female on the likelihood of being HIV positive (row 2). This is an additional evidence of the well-documented

“Feminization of HIV in Sub-Saharan Africa” discussed in section 1.1. As highlighted in Greenwood,

Kircher, Santos and Tertilt(2019) and in the medical literature, for physiological and anatomical reasons, women are at higher risk than men to contract the virus when exposed to it (Yi et al., 2013).

Importantly, the “Matrilineal effect” (row 1) becomes non-signiﬁcant and very close to zero, once region-survey ﬁxed effects are included. Ancestral matrilineality in Sub-Saharan Africa being essen-

31 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa tially geographically located in countries of the so-called “Matrilineal Belt”, this suggests that the unconditional positive signiﬁcant effect of 푀푎푡푟푖푙푖푛푒푎푙푖푡푦 found in column 1 may in fact capture the effect of other country and within-country level factors. Among them, legal systems and codes of law could be plausible candidates. Indeed, Anderson(2018) highlights that female HIV rates are signicantly higher in common law countries. Since most of countries of the “Matrilineal Belt” are common law countries, the “Matrilineal effect” found in column 1 might in fact be a “Common Law

Countries effect” (I discuss it in more details in the following subsection). However, this effect dis- appears once (within-country) region-survey fixed effects are included, while my main coefficient of interest remains consistent, positive and statistically significant at the 1% level.

All in all, the absence of significance of the “Matrilineal effect” as well as the positive significant effect of my main variable of interest (퐹 푒푚푎푙푒×푀푎푡푟푖푙푖푛푒푎푙푖푡푦) is confirmed by the analysis made by gender subsamples, and presented in columns 6 and 7 of Table 1.1: I find that women originating from an ancestral matrilineal ethnic group are 1.2 percentage points (column 6) significantly more likely to be HIV positive (representing 31% of the average HIV prevalence of patrilineal females in my regression sample); while I do not find any significant effect for their male counterparts (column

7).

This gender-speciﬁc effect can be rationalized by the interaction of two phenomena. First, according to my conceptual framework, ancestral matrilineality has a long-lasting effect speciﬁc on female’s contemporary sexual behaviour. Second, the medical literature highlights that, for physiological and anatomical reasons (e.g. larger surface area of mucosal HIV exposure; increased mucosal expression of the HIV co-receptor CCR5; and a greater probability of virus exposure on the rectal mucosa among other factors), women are at higher risk than men to contract the virus when exposed to it (Yi et al.,

2013).28 In the end, the fact that matrilineal women adopt riskier sexual behaviours than their patrilineal counterparts combined with the higher suceptibility of women to contract the virus when exposed to it lead to the gender-speciﬁc long-lasting effect of matrilineality on HIV.29

28Building on several medical studies, Greenwood et al.(2019) assume that women are 75% more likely to get infected than males for physiological and anatomical reasons. 29I illustrate this in section 1.6, performing a simulation based on an epidemiological model assuming behavioural differences between matrilineal and patrilineal women as well as gender differences inHIV susceptibility.

32 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

Table 1.1 – The Effect of Ancestral Matrilineality on Female HIV (OLS)

HIV (1) (2) (3) (4) (5) (6) (7)

Matrilineality 0.058*** 0.000 0.001 -0.000 -0.001 0.012* -0.004 (0.011) (0.004) (0.005) (0.006) (0.005) (0.007) (0.005) Female 0.016*** 0.012*** 0.009*** 0.009*** 0.008*** (0.002) (0.002) (0.002) (0.002) (0.002) Female × Matrilineality 0.020*** 0.022*** 0.019*** 0.016*** 0.017*** (0.006) (0.006) (0.005) (0.006) (0.006) Ind. Controls Yes Yes Yes Yes Yes Ethnic Group Controls Yes Yes Yes Yes Village-Geographic Controls Yes Yes Yes Region-survey FE Yes Yes Yes Yes Yes Yes

Gender Both Both Both Both Both Female Male Observations 280,140 280,140 273,417 193,991 182,312 105,964 76,348 Adj. R-squared 0.018 0.050 0.078 0.076 0.078 0.087 0.065 Clusters 172 172 172 100 100 100 82 Mean Dep. Var. (Patri. Males) 0.021 0.021 0.021 0.023 0.023 0.039 0.023 Notes: OLS estimates are reported with standard errors clustered at the ethnicity level in brackets. The unit of observation is an individual originating from either a traditionally matrilineal or a traditionally patrilineal ethnic group (it therefore excludes individuals originating from ethnic groups with alternate inheritance rules (ambilineality, bilinearity, duolinearity, etc.)). “Matrilineality” indicates (dummy) whether an individual belongs to a traditionally matrilineal ethnic group. “Female” indicates (dummy) whether an individual is a female. “Female × Matrilineality” indicates (dummy) whether an individual is a female belonging to a traditionally matrilineal ethnic group. “HIV” is a dummy indicating whether an individual is HIV positive (from DHS HIV Tests). The Individual Controls are computed from DHS and include: indicators of marital status; a dummy for polygynous union; number of children; age; age squared; a dummy for living in an urban location; education in number of years; a dummy for currently working; wealth index indicators and religion indicators. The (Ancestral) Ethnic Group Controls are computed from the Ethnographic Atlas (EA) and include: intensity of women’s historical participation in agriculture; ancestral polygyny; ancestral bride price; ancestral plough use; ancestral pastoralism; ancestral presence of clans; indicators of ancestral settlement patterns; indicators of ancestral juridictional hierarchies beyond local communities; ancestral reliance on hunting; ancestral reliance on fishing; ancestral reliance on gathering; ancestral reliance on animal husbandry; ancestral reliance on agriculture; ancestral presence of large domesticated animals; indicators of intensity of ancestral agriculture; and year of observation of the ethnic group. The Village-Geographic Controls are computed at the village level and include: latitude; longitude; altitude; nightlight composite; population density (2010); distance to lake or coastline; distance to nearest international border; average time (minutes) required to reach a high-density urban center (2015); distance (in km) to nearest active mine; a dummy indicating whether an active mine is located within a distance of 1000 km max. of the village; malaria incidence (2010); vegetation index; indicators for the length of the growing season; an index of ethnic fractionalization; and an index of ethnic polarization. Region-survey is a subnational region defined in DHS, interacted with its survey-year. Sample in column 6 consists of female individuals. Sample in column 7 consists of male individuals. * p<0.10,** p<0.05,*** p<0.01

1.4.2 Robustness Checks

1.4.2.1 Selection Analysis

The blood test being not compulsory, selection might arise in the sample. However, the DHS program reports that the average response rate, for those who are eligible for the test, is extremely high and that a comparison between the characteristics of those who agreed to be tested and those who

33 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa refused testing shows minimal bias.30 Moreover, it is reasonable to expect that any selection would cause a downward bias since infected individuals should be less keen to be tested.31 As a result, if female HIV infection is positively affected by ancestral matrilineality, more underreporting among females originating from ancestrally matrilineal ethnic groups should be expected. In order to test this hypothesis, I test the potential effect of ancestral matrilineality on the probability of underreporting

HIV infection, which in practice amounts to a refusal to consent to the blood test and non take-up of it.

Further, given the persistence of mistrust in medicine in regions where colonial medical campaigns were established (Lowes and Montero, 2018), lack of consent in regions close to missions could pose a potential threat to the estimation. The inclusion of (within-country)-survey ﬁxed effects as well as numerous village-level geographical controls allows me to rule out such potential threat.

Column 1 and column 2 of Table 1.C.3 report estimates on the probability that DHS respondant consent to take HIV test, and on the probability that she actually takes it. The estimate of 퐹 푒푚푎푙푒 ×

푀푎푡푟푖푙푖푛푒푎푙 is non signiﬁcant and very close to zero in both cases, suggesting that potential selection can be ruled out (in fact, descriptive statistics show that about 95% of patrilineal females against 93% of matrilineal females consent to HIV test in my ﬁnal sample, and these numbers are similar for actual take-up.)

While I do not find evidence of selection into DHS HIV testing between matrilineal and patrilineal women, selective attrition due to HIV mortality might affect my estimates. The direction of this bias is unclear. On one hand, if we assume that seropositive individuals have a lower life expectancy, the higher HIV prevalence among matrilineal women should bias downward my main results.32 On the other hand, using country-level data from UNAIDS, I find a positive association between matrilineality and antiretroviral treatment (ART) coverage (see my discussion on ART in subsection 1.5.4), and a negative association between matrilineality and HIV mortality.33 While time-varying country-level variations are explicitly controlled for in my regressions through the inclusion of (within-country) region-survey (year) fixed effects, these country-level correlations suggest that bias due to selective mortality could also go in the other way.

30See https://dhsprogram.com/topics/HIV-Corner/hiv-prev/index.cfm 31This is confirmed by Mishra, Vaessen, Boerma, Arnold, Way, Barrere, Cross, Hong and Sangha(2006), who find that the rate of HIV infection among individuals not tested for HIV is systematically largerthanthe rate among those not tested. 32The average time from infection to the outbreak of symptoms is equal to 10 years; and the average time from the outbreak of symptoms to death is 2 years (Greenwood et al., 2019). 33By lowering the viral load, ART makes the seropositive person taking the drugs feel healthier and live longer, notably.

34 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

1.4.2.2 Other Health Outcomes as a Falsiﬁcation Test

HIV is a highly infectious, largely sexually transmitted disease. Using DHS data on level of anemia in respondents, a not sexually transmitted disease34, as well as BMI and Rohrer index as other objective measures of health status35, I perform falsiﬁcation tests by estimating the long-term effect of ancestral matrilineality on such health outcomes.36 This allows me to rule out the possibility that differences in HIV rates found previously hide more general differences in economic development and/or health infrastructures, and therefore differences in overall health status. Results are reported in column 3 to 5 in Table 1.C.3. I ﬁnd that ancestral matrilineality is not associated with an increase in the prevalence of anemia, neither with Body Mass Index (BMI)37 or Rohrer index38 of female individuals, corroborating the hypothesis that sexual behaviour is the actual driver of my results on

HIV.

1.4.2.3 Robustness to Alternative Channels

The previous results demonstrate that ancestral matrineality is a long-term determinant of contemporary prevalence of HIV among female populations in Sub-Saharan Africa. However, other long- term determinants of female HIV in Sub-Saharan Africa have also been recently highlighted in the literature. I discuss here the robustness of my main ﬁndings to these competing channels.

Common Law vs. Civil Law Countries. Anderson(2018) has recently highlighted the legal origins of female HIV in Sub-Saharan Africa. In particular, exploiting variation in legal origins of formerly colonized countries and the fact that common law is associated with weaker female marital property laws as compared to civil law, she finds higher prevalence of female HIV in common law countries. Lower bargaining power of women in these countries and therefore lower ability to impose safe sexual pratices to their husbands constitute her main mechanism. Interestingly, in her identification strategy she exploits geographical variation in common law vs. civil law countries within ethnic group, by including ethnic group fixed effects, and therefore rule out any ethnical effect. In some sense, my identification strategy in symmetric since I exploit ethnic norm (and related kinship organi-

34Bertocchi and Dimico(2019) select anemia for their falsification test because of its relevance, given the association between anemia and malaria, another vast-scale health problem in Africa. 35This measures are only available for female respondants in DHS. 36I measure the severity of anemia with a dummy variable taking value one when an individual is diagnosed with either mild, moderate, or severe anemia, and zero otherwise. 37Body Mass Index (BMI) is computed as follows: 푚푎푠푠/ℎ푒푖푔ℎ푡2 38Rohrer index is computed as follows: 푚푎푠푠/ℎ푒푖푔ℎ푡3

35 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa zation) variation within (within-country) region - survey (year), therefore allowing me to rule out any legal/institutional effect.

As a further robustness check, I also perform an heterogeneous analysis, estimating Equation 1.1 on the subsample of individuals residing in common law countries versus individuals residing in civil law countries.39 The two ﬁrst columns of Table 1.C.4 reports the result of these estimations and provide evidence of an heterogeneity: the effect of ancestral matrilineality on female HIV holds true within common law countries only. In fact, looking at the magnitude, it seems that the average effect found in subsection 1.4.1 is mostly driven by individuals residing in common law countries. However, this is not a surprising result since most of the variation in ancestral inheritance norm is found within common law countries.40

Polygyny. Bertocchi and Dimico(2019) have recently highlighted contemporary polygyny as an other driver of female HIV in Sub-Saharan Africa. Females’ riskier sexual behaviour triggered by the absence of the husband, and subsequent multiplicative virus transmission by the husband to his other wives constitute their main mechanism. The inclusion of a dummy for actual polygynous union in my individual-level controls, as well as a dummy for ancestral polygyny norm in my ethnic-group controls are meant to capture this alternative transmission channel.

Again, I also peform an heterogeneous analysis as an additional robustness check, by estimating

Equation 1.1 on the subsamples of individuals who are not currently in a polygynous union versus individuals who are currently in a polygynous union. The results are reported in column 3 and 4 of

Table 1.C.4 and indicate an heterogeneity in the long-term effect of ancestral matrilineality on female

HIV in Sub-Saharan Africa: these effect holds true for non-poygynous individuals only. Similarly to the previous heterogenous analysis, the magnitude of the effect found for non-polygynous individuals suggest that these latter are in fact mainly driving the average effect found in subsection 1.4.1. How- ever this is not a surprising result since most of the variation in ancestral inheritance norm is found among non-polygynyous females.41

39Following Anderson(2018), I use the dataset from La Porta, Lopez-de Silanes and Shleifer(2008) to identifiy common law and civil law countries. 40In my final sample, in common law countries about 29% of females originate from an ancestrally matrilineal ethnic group vs. about 5% in civil law countries. 41In my final sample, about 17% of non-polygynous females originate from an ancestrally matrilineal ethnic group vs. about 7% of polygynous females.

36 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

Ancestral Matrilocality. Matrilocality, an ethnic-group traditional norm according to which a married couple is supposed to live with or close to the wife’s family, is highly correlated with matrilineality in Sub-Saharan Africa. In my main regression sample, about 92% of matrilocal individuals are matrilineal and about 94% of non-matrilocal individuals are non-matrilineal. Hence, ancestral matrilocality is hardly separable from ancestral matrilineality in my conceptual framework, and can be seen as an additional dimension of women’s position in ancestral matrilineal societies. For these reasons, I decide to not control for it in my preferred speciﬁcations. Nevertheless, as reported in

Table 1.C.5 in appendix, I perform a robustness check exercise to check that ancestral matrilocality is not driving my main results.

Several lessons can be drawn from this exercise. First, looking at column 1, my main findings are unchanged when I additionally control for ancestral matrilocality. Unsurprisingly, the adjusted r-squared remains also unchanged. Second, looking at column 2, my main findings remain valid for non-matrilocal individuals, who constitute the majority of my regression sample. All in all, these results suggest that it is very unlikely that my main findings are driven by differences between ancestrally matrilineal and patrilineal ethnic groups’ adoption of a matrilocal traditional norm.

Geographic Channels. Several geographic factors of the spread of HIV have been highlighted in the literature. Corno and De Walque(2012) show that mine workers’ international migration is a driver of the spread of HIV. Along the same line, Oster(2012b) highlights exports and road networks, and subsequent increase in movements of people and sexual contacts, as an other factor of HIV infection. An other long-term determinant of female HIV in Sub-Saharan Africa put forward by Cagé and Rueda(Forthcoming) is is the geography of Protestant and Catholic missions in the early 20th century, as well as their health investments. In the same vein, Teso(2019) shows that the slave trade in

Sub-Saharan Africa had long-term effects on contemporary gender norms, and Bertocchi and Dimico

(2019) underline that the slave trade was a driver of actual polygyny in Sub-Saharan Africa.

In my main speciﬁcation, and in addition of (within-country) DHS survey region ﬁxed effects which should capture most of these geographical variations, I also include a host of village-level geographical controls, including latitude, longitude, altitude, nightlight composite, population density

(2010), distance to lake or coastline, distance to nearest international border, average time (minutes) required to reach a high-density urban center (2015), distance to nearest active mine, an indicator for the presence of an active mine within 1000km of the village, malaria incidence (2010), vegetation

37 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa index, and indicators for the length of the growing season (see subsection 1.3.3). My main results being robust and almost unchanged following the inclusion of these controls (see Table 1.1, column

5), I can reasonably rule out the possibility that such competing geographic channels are driving my main results.

Ethnic Fractionalization and Polarization. Ethno-linguistic diversity has been emphasized by

Tequame(2012) as a driver of risky sexual behaviour and subsequent HIV. As a matter of fact, as she argues, one mechanism might be that social sanction due to risky behaviour is less costly in heterogeneous societies rather than homogeneous ones. A second mechanism might be that information spreads more easily in homogeneous than in heterogeneous societies, because the former are more likely to have the same language, culture and networks. Since to be subject to social sanction individuals should be detected as having risky sexual behaviuor, individuals who want to keep risky sexual behaviour secret might ﬁnd heterogeneous societies more favorable. In addition, social interactions, which might differ by the degree of ethnical homogeneity, provide information about the level of

HIV/AIDS at community level, including information on infectious status and risky behaviour of partners. Therefore, to account for the possibility that ancestral matrilineality might be associated with different degrees of ethnical heterogeneity, I control for an index of ethnic fractionalization and an index of ethnic polarization, which I compute at the village level using information on individual’s ethnicity, following Montalvo and Reynal-Querol(2005) formulas. 42

Economic Development. Differences in economic development may trigger differences in HIV rates. For example, legal systems and institutions are well-know driver of economic development (see

La Porta et al., 2008 for a review). Such differences are controlled for both at the country and within- country level with the inclusion of (within-country) region fixed effects in my regressions. I also control for potential differences at the village-level by including, as proxies of village-level economic development, village’s nighlight composite as well as village’s population density. I finally control for such diferences at the ethnic group level by controlling for ancestral settlement patterns; ancestral juridictional hierarchies beyond local communities; ancestral reliance on hunting; ancestral reliance on fishing; ancestral reliance on gathering; ancestral reliance on animal husbandry; ancestral reliance on agriculture; ancestral presence of large domesticated animals; intensity of ancestral agriculture;

42These formulas and their interpretations are reported in subsubsection 1.A.2.3 in appendix.

38 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa and year of observation of the ethnic group (to alleviate the concern that some groups were portrayed later than others and might therefore have been more developed).

I ﬁnd robust and almost unchanged estimates of the long-term effect of ancestral matrilineality on female HIV once these controls are included (see column 5 in Table 1.1), thus alleviating the concern that differences in economic development may drive my results.

Controlling for gender-specific alternative channels. So far, my OLS estimates are robust to a wide range of alternative channels that may spuriously drive my results on HIV. As a further robustness check, I check that my results are not driven by gender-specific alternative channels. In other words, I further check that alternative channels discussed so far are not affecting female populations differently. To do so, I additionally control for my set of controls and fixed effects interacted with a

“Female” dummy. Results from this robustness exercise are reported in Table 1.C.6. While I ﬁnd that the effect of being a female is affected by the inclusion of this new host of gender-speciﬁc controls, my main result on the long-lasting effect of matrilineality on female populations remains unaffected.

Controlling for additional observables. Despite the emphasis put so far in controlling for numerous alternative channels, based on observables computed at either the individual, Ethnographic

Atlas (EA) ethnic groups or DHS village level, I intend here to control more directly for alternative channels discussed above, adding covariates computed at the Murdock’s ethnic group level, and based on Nunn(2010), Nunn and Wantchekon(2011) and Teso(2019) datasets. The limitation of this exercise is that I cannot match DHS ethnic groups with Murdock’s ethnic groups as extensively as I did when I matched DHS ethnic groups with EA ethnic groups. Therefore, adding these new covariates will restrict my sample size. However, the value of this exercise is to assess the robustness and the stability of my estimates when explicitly controlling for additional alternative channels. In this way, I more explicitly control for the slave trade alternative channel by controling for the logarithm of 1 plus the number of slaves taken in the transatlantic and in the indian slave trade from the Murdock ancestral ethnic group, divided by the area of the land historically inhabited by the ethnic group. Then, I aim to control for ethnic group’s contact with colonizers during colonization by computing a dummy for whether a European explorer’s route traveler crossed the historical land of the ethnic group, and a dummy for whether part of the railway network built by the Europeans was on the land of the ethnic group. Further, I also control for the differential effects of the different types of religious missions

39 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

(Cagé and Rueda, Forthcoming) by including in my regression the number of catholic, protestant and

British and Foreig Bible Society (BFBS) missions per square kilometer of an ethnic group’s land during the colonial period. Additionally, I directly control for the minimum distance of the centroid of the land historically inhabited by the ethnic group from the routes of the Saharan trade, and from the closest city of the Saharan trade. Finally, I also estimate my regressions by controlling for the number of cities with more than 20,000 inhabitants that were present in 1400 on the land inhabited by the ethnic group, and for the number of conﬂicts between 1400 and 1700 in the area inhabited by the ethnic group, based on Besley and Reynal-Querol(2014) original dataset.

In the same spirit, I also control for additional covariates, computed at the DHS village level, which likewise reduce my sample size but allow me to assess the robustness of my estimates. More speciﬁcally, using answers from male respondants in DHS43, I compute village’s proportion of male circumcised as well as village’s proportion of males who report having paid for sex in the last 12 months. These two covariates are intended to control for circumcision and prostitution44, which have both been emphasized in the literature as an impediment and a driver of HIV, respectively.

Table 1.2 reports the OLS estimates of this additional robustness exercise and underlines that my

OLS estimates are remarkably consistent when controlling for all of these covariates both separately and simultaneously. Thus, according to my fully-controlled OLS regression reported in column 9, I still find that females are more likely to suffer from HIV than their male counterpart. Further, I also still find that female originating from an ancestrally matrilineal ethnic group are 1.7 percentage point more likely to suffer from HIV than their patrilineal counterparts, an effect of very similar magnitude than the effect I found in my main specification, in column 5 of Table 1.1.

43The reduction in sample size when adding these controls is notably explained by the fact that only females are interviewed in some DHS villages. 44Unfortunately for this study, female respondants have not been interrogated in DHS about prostitution.

40 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa Yes Yes Yes Yes Yes Yes Yes Yes Yes , “Ind. Controls”, “Ethnic Group Controls”, “Village- Yes Yes HIV Matrilineality” × “Female * p<0.10,** p<0.05,*** p<0.01 Robustness of OLS Estimates to Additional Controls (1) (2) (3) (4) (5) (6) (7) (8) (9) (0.006) (0.005)(0.002) (0.006) (0.002)(0.006) (0.006) (0.002) (0.006) (0.006) (0.002) (0.006) (0.006) (0.002) (0.006) (0.007) (0.002) (0.006) (0.006) (0.002) (0.006) (0.010) (0.002) (0.006) (0.002) (0.005) (0.006) 0.017*** 0.017*** 0.017*** 0.016*** 0.016*** 0.016*** 0.017*** 0.017*** 0.017*** , “Matrilineality”, “Female”, “HIV” Table 1.2 – Matrilineality × OLS estimates are reported with standard errors clustered at the ethnicity level in brackets. The unit of observation is an individual originating from either ObservationsAdj. R-squaredClustersMean Dep. Var. (Patri. Males) 0.024 167,485 0.024 0.080 167,485 0.024 167,485 0.080 169,150 83 0.024 0.080 169,150 0.024 169,150 0.080 83 115,620 0.024 0.080 136,586 83 0.026 106,363 0.080 0.023 0.090 84 0.026 0.086 84 0.092 84 73 74 58 Precolonial conflicts (1400-1700) DHS Villages Contemporary Controls: Village’s prop. of malesVillage’s circumcised prop. of males who paid for sex MatrilinealityContact explorer routeColonial railwayCatholic missions/areaProtestant 0.001 missions/areaBFBS missions/areaDistance -0.002 Saharan routeDistance Saharan nodeCities in 0.002 1400 0.001 0.001 0.002 -0.007 Yes -0.007 Yes -0.013 Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes FemaleFemale Ind. ControlsEthnic Group ControlsVillage-Geographic ControlsRegion-survey FEMurdock Ethnic Groups Historicalln(1 Controls: + Total slave exports/area) Yes 0.008*** Yes 0.008*** Yes 0.008*** Yes Yes 0.008*** Yes Yes 0.008*** Yes 0.008*** Yes 0.009*** Yes Yes 0.008*** Yes 0.009*** Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Notes: Geographic Controls”, and “Region-surveyrespondent’s FE” ethnic are group defined indummy in taking theTable value transatlantic 1.1 . one and/or ifbuilt “ln(1 the an by + the indian European Total Europeans slave explorer slaveand was trade route exports/area)” Foreign on crossed divided is Bible the the by the Society land landSaharan the logarithm (BFBS) of of area node” of the missions the ethnic per 1 of are ethnic group. plus square group. land thecity kilometer the “Catholic historically of minimum “Colonial of missions”, number inhabited railway” distance the “Protestant an of is by of Saharanethnic missions” ethnic a slaves the the trade, and group’s group. dummy taken group. “BFBS centroid respectively. taking land from missions” value of “Precolonial during the “Contact “Citiescircumcised” are one the the conflicts explorer in is the if colonial land route” number (1400-1700)” 1400” the a period, of historically is part within ismale catholic, respectively. of inhabited a DHS the is respondants protestant, “Distance the by village’s number number British whothe railway Saharan the proportion of report network route” ethnic of cities having and conflicts of group male with paid “Distance between from respondants for more circumcised. the sex than routes in 20,000 “Village’s1400 the of prop. inhabitants last and the that of 12 Saharan 1700 were males months. trade present who in andin the area paid 1400 from for on the sex” inhabited the closest is land the by the inhabited within by DHS ethnic group. the village’s “Village’s proportion of prop. of males a traditionally matrilineal or a traditionally patrilineal ethnic group (it therefore excludes individuals originating from ethnic groups with alternate inheritance rules (ambilineality, bilinearity, duolinearity, etc.)).

41 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

1.4.2.4 Assessing Selection on Unobservables

Despite my attempts to control for a large set of observable factors, at both the individual, ethnic group and village level, the estimates reported in Table 1.1 may still be biased by unobservable factors correlated with both ancestral matrilineality and females’ contemporary sexual and contraceptive behaviours, and subsequently HIV. A priori, the direction of the potential omitted variable bias is not clear. Consider for instance the possibility that ethnic groups where women historically beneﬁted from more sexual freedom adopted matrilinelity as a resulting adaptative kinship organization. As such, some evolutionary anthropologists explain the existence of matrilineal societies as the result of an evolutionary process that created institutions suitable for the ecological and social environment.45

In particular, matrilineality may be advantageous in environments with low paternal certainty since while it is difﬁcult to conﬁrm paternity, maternity is easily observable. Thus, an inheritance system in which property passes from the mother’s brother to her sons may be optimal since the brother knows he is related to his sister, but cannot verify that he is related to his children (Fortunato, 2012). This could drive the OLS estimates away from zero. On the other side, mating process might be inherently more assortative in matrilineal societies, relative to patrilineal societies, yielding matrilineal females suffering from relatively lower marital dissatisfaction. If that was the case, this might lead to matrilineal females being relatively less likely to adopt risky sexual behaviours46, and therefore drive the

OLS estimates towards zero. In this subsection I consequently assess the likelihood that the OLS estimates might be biased by unobservables.

Coefﬁcients Ratio Tests (Altonji et al., 2005). I start by assessing the sensitivity of the OLS estimates to controlling for observable characteristics. To do so, I ﬁrst employ the strategy adapted by

Nunn and Wantchekon(2011) from Altonji et al.(2005) that allows to determine how much stronger selection on unobservables would have to be compared to selection on observables to fully explain ^ ^ ^ ^ away my results. To perform this test, I calculate the ratio 훽퐹 /(훽푅 − 훽퐹 ), where 훽퐹 is my coefﬁcient ^ of interest from a regression that includes my full set of controls, while 훽푅 is my coefﬁcient of interest from a regression that includes a restricted set of controls. The intuition behind the formula ^ ^ is straightforward. First, consider why the ratio is decreasing in (훽푅 − 훽퐹 ). The smaller is the ^ ^ difference between 훽푅 and 훽퐹 , the less the estimate is affected by selection on observables, and the

45Lowes(2018a) appendix provides an extensive overview of the hypothesized origins of matrilineal kinship systems. 46In fact Bertocchi and Dimico(2019) show that, due to marital dissatisfaction, women in polygynous union in Sub-Saharan Africa adopt riskier sexual behaviours which are more conducive to HIV.

42 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa stronger selection on unobservables needs to be (relative to observables) to explain away the entire ^ ^ effect. Next, consider the intuition behind 훽퐹 in the numerator. The larger 훽퐹 , the greater is the effect that needs to be explained away by selection on unobservables, and therefore the higher is the ratio

(Nunn and Wantchekon, 2011).

The results are reported in columns 1 of Table 1.C.7, where each row reports result for different set of restricted covariates. This yields four ratios that range from -509.27 to 167.42. In some cases, the coefficient in the fully-controlled model is larger than that of the uncontrolled model, giving a negative ratio. In general, these ratios are far greater than 1 in absolute value, and therefore suggest that the influence of unobservable characteristics would have to be far greater than the influence of observable characteristics to fully account for my OLS findings.

Minimum Coefﬁcient Lower Bound (Oster, 2017). Further, I also use the method from Oster

(2017) to calculate a bias-adjusted lower bound of my coefﬁcient of interest. Oster shows that if one assumes that observables and unobservables have the same explanatory power in the outcome vari- * ^ ^ ^ 2 ^2 ^2 ^2 able, then the following is a consistent estimator: 훽 = 훽퐹 −(훽푅 −훽퐹 )×((푅푀푎푥 −푅퐹 )/(푅퐹 −푅푅)), ^ ^ ^2 ^2 where 훽푅 and 훽퐹 are deﬁned as above; 푅퐹 is the R-squared from the fully-controlled regression; 푅푅 2 is the R-squared from the restricted regression; and 푅푀푎푥 is the R-squared from a regression that includes all observable and unobservable controls. Although in theory the maximum possible value

2 of 푅푀푎푥 is one, as underlined by González and Miguel(2015), in the real world, where there is sig- 2 nificant measurement error, the value of 푅푀푎푥 should be much lower than one. In fact, by definition, 2 ^2 2 푅푀푎푥 ∈ [푅퐹 ; 1]. Oster(2017) provides some insights on how 푅푀푎푥 should be chosen, showing 2 that 푅푀푎푥 = 1 may lead to over-adjustment in many cases. I follow her procedure and present bias- 2 ^2 47 2 ^2 adjusted lower bound coefficients for 푅푀푎푥 = 푚푖푛(1.3푅퐹 ; 1) , 푅푀푎푥 = 푚푖푛(1.5푅퐹 ; 1), and 2 ^2 푅푀푎푥 = 푚푖푛(2푅퐹 ; 1) in column 2, 3 and 4 respectively (I also report bias-adjusted lower bound 2 coefficients for 푅푀푎푥 = 1 in column 5 for informational purpose). All bias-adjusted lower bound estimates from this exercise are reported in Table 1.C.7 and remain positive and, taken at face value, still imply a sizeable estimated effect of ancestral matrilineality on female HIV, of same order of magnitude than previously found in my OLS regressions. Further, it is worth noting that the full set of these biased-adjusted lower bound estimates (column 2 to 5) falls within the 99.5% confidence interval of my fully-controlled OLS estimate (column 7), which suggests

47 2 ^2 Oster(2017) suggests using 푅푀푎푥 = 1.3푅퐹 as a cutoff to test for the stabillity of a treatment effect consistent with randomized treatment.

43 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa that the size of the estimate from the OLS regression with full controls is similar to the bias-adjusted estimates. All in all, these tests suggest that my fully-controlled OLS estimates are very unlikely to be affected by omitted variable bias, and therefore support a causal interpretation of my OLS ﬁndings.

1.4.3 Alternative Identiﬁcation Strategies

Though my OLS estimates are robust to a large set of observables that could be potential confounders, and are unlikely to be affected by omitted variable bias, reverse causality may remain a concern. For example, matrilineality might have been an adaptative response in environments with low paternal certainty (see discussion in subsubsection 1.4.2.4). If environments with low paternal certainty were associated with more promiscuous sexual behaviours, OLS estimates would be biased away from zero.

In order to provide further support to the causal interpretation of my main individual-level OLS

ﬁndings, I implement three alternative identiﬁcation strategies. These strategies have the common feature to allow me to estimate the effect of the variation in cultural trait within similar environments.

To do so, I exploit data on the GPS location of DHS villages and the digitized Murdock’s map of ancestral ethnic groups in Africa.

1.4.3.1 Accounting for Unobservables: Geographic RD Estimates

As a first alternative identification strategy, I undertake a geographic regression discontinuity analysis. More precisely, I examine and compare individuals living in villages geographically close, but where some villages are located within the ancestral boundaries of an ancestrally matrilineal ethnic group and the other within the ancestral boundaries of an ancestrally non-matrilineal (i.e. patrilineal or other) ethnic group. In this framework, the ancestral matrilineal ethnic boundary is the delineation created by the ancestral borders of ethnic groups that practiced matrilineal descent alongside groups that practiced patrilineal or alternate descent (based on digitized Murdock’s map of ancestral ethnic groups in Africa, see Figure 1.2). The intuition behind this specification is that the ancestral matrilineal ethnic boundary is determined by the ancestral borders of multiple matrilineal and non-matrilineal ethnic groups. The boundaries between these multiple ethnic groups are arbitrary, and along the border geographic areas are quite similar.48

48Note that the ancestral matrilineal ethnic boundary does not coincide with any actual border.

44 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

Therefore, my strategy is to use a regression discontinuity (RD) estimation method that restricts the sample to individuals living in villages that are sufﬁciently close to the ancestral matrilineal ethnic boundary and estimate the causal effect of living in a village located on the matrilineal side on female

HIV. Importantly, I do not estimate the effect of originating from an ancestrally matrilineal ethnic group anymore, but the effect of living in a village located in an ancestrally matrilineal geographic area. The benefit of this strategy is that it accounts for unobservable factors that vary smoothly across space. Therefore, as long as the determinants of unobservable traits (e.g. geography, history, idiosyncratic shocks, state presence etc.) vary smoothly, the unobservable traits will be accounted for by the RD strategy. Further, in order to get the more conservative estimates, I directly control for the large array of individual, ethnic group’s historical and village-level geographic control, as well as (within-country) region × survey (year) fixed effects described in the previous sections. More specifically, my RD specification takes the following form:

푦푖푒푣푟푡 = 훼 + 훽1푀푎푡푟푖푙푖푛푒푎푙푖푡푦푣 + 훽2퐹 푒푚푎푙푒푖 + 훽3퐹 푒푚푎푙푒푖 × 푀푎푡푟푖푙푖푛푒푎푙푖푡푦푣+

′ ′ ′ XievrtΔ + XertΩ + XvrtΠ + 푓(푙표푐푎푡푖표푛푣) + 휆푟푡 + 휀푖푒푣푟푡 (1.2)

with 푦푖푒푣푟푡 denoting an individual-level outcome (indicator for whether the individual is HIV positive) for individual 푖 from ethnic group 푒, living in village 푣 in within-country DHS region 푟, and surveyed at year 푡. 푀푎푡푟푖푙푖푛푒푎푙푖푡푦푣 is an indicator for whether an individual lives in a village located on an ancestrally matrilineal geographic area; 퐹 푒푚푎푙푒푖 is an indicator for whether an individual is a female. 훽1 is intended to capture a “Matrilineal effect”; and 훽2 is intended to capture a “Gender effect” on HIV. The coefﬁcient of interest 훽3 captures the effect of living in a village located in an ancestrally matrilineal geographic area on female HIV once the “Matrilineal effect” and the “Gender effect” have been controlled for. 푓(푙표푐푎푡푖표푛푣) denotes a a RD polynomial that controls for a smooth function of the geographic location of DHS villages. In my speciﬁcations I alternatively use the minimum distance to the nearest ancestral matrilineal ethnic boundary (in km) and the gps coordinates (latitude and longitude) of the village as running variables. Further, I use several functional forms of the polynomial, using polynomials of different orders, and alternatively estimating them separately on each side

′ ′ ′ of the boundary (“flexible polynomials”). Xievrt; Xert, and Xvrt represent a set of individual-level, ethnic group-level ancestral, and village-level geographic covariates respectively, which are defined in subsection 1.3.3. 휆푟푡 denotes (within-country) DHS region-survey (year) fixed effects. Standard

45 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

errors 휀푖푒푣푟푡 are clustered at the ethnic group level, and the sample is restricted to individuals living in villages that are within a certain distance to the ancestral matrilineal ethnic boundary, either 100,

150, or 200 kilometers.

Validating the Assignment of Matrilineal Individuals. The boundaries used for my RD estimates are from Murdock (1959, see Figure 1.2), a source that has been used previously in a number of studies that use a similar RD approach (see Moscona et al.(2020) for a recent and related example).

However, an important assumption when using the ethnic boundaries is that they accurately reﬂect true discontinuities of ethnic afﬁliation today. This is particularly important since, in reality, one may not observe clear borders between ethnic groups, and instead only a gradual change of the mix of ethnicities over space. Further, an additional assumption that I have made so far is that the matching between individual’s self-reported ethnicity in DHS and ancestral ethnic groups in Ethnographic At- las is accurate.49 Therefore, I now check the validity of my use of Murdock’s ethnic boundaries by examining how individual’s ancestral matrilineality varies at ancestral matrilineal ethnic boundaries.

This is shown in Figure 1.B.4 in appendix, which reports the bivariate relationship between distance from the ancestral matrilineal ethnic border and individual’s ethnic group’s ancestral matrilineality.

The 푦-axis displays the fraction of the population in a 5 km bin that reports that they are a member of an ancestrally matrilineal ethnic group, and the 푥-axis is distance in kilometers from the border. A positive distance indicates a location within the ancestral territory of an ancestrally matrilineal ethnic group and a negative distance indicates a location within the ancestral territory of an ancestrally non- matrilineal (i.e. patrilineal or other) ethnic group. Reassuringly, I ﬁnd that there is a discontinuous change at the border in the fraction of the population that report being member of an ancestrally matrilineal ethnic group.50

Geographic RD Estimates. Before turning to my estimates I ﬁrst examine the raw data of the

RD sample. Figure 1.B.5 in appendix shows a bin scatterplots of the predicted HIV rate for females living in villages located within 150 km of the ancestral matrilineal ethnic border, using a ﬂexible

49This is important since individual’s ethnic group’s ancestral matrilineality variable is based on kinship organization of ancestral ethnic groups reported in the Ethnographic Atlas. 50It is however important to note that, while information on individual’s ethnic group’s ancestral matrilineality is based on the matching between individual’s self-reported ethnicity in DHS and ancestral ethnic groups in Ethnographic Atlas, these latter slightly differ from ethnic groups classification in Murdock’s Map of ancestral ethnic groups (1959). This might partially explain why the discontinuous increase very close to the boundary is of limited size.

46 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa third-order RD polynomial conditioned on region (within-country) × survey (time) FE and estimated separately on each side of the border. Positive values, on the horizontal axis, reﬂect 5 km bins in ancestrally matrilineal geographic areas and negative values reﬂect 5 km bins in ancestrally non-matrilineal geographic areas. Even in the raw data, a discontinuity at the border is apparent: a discontinuous increase in female HIV rate on the matrilineal side of the border can be observed. I next turn to my more formal RD estimates.

Table 1.3 reports the geographic RD estimates for different bandwidths: 100 km (columns 1 and

2), 150 km (columns 3 and 4), and 200 km (columns 5 and 6); different running variables: minimum distance in km between DHS village and nearest ancestral matrilineal ethnic boundary (columns 1,

3 and 5), as well as village’s latitude and longitude (columns 2, 4 and 6); and different polynomial specifications: linear polynomial, flexible linear polynomial, quadratic polynomial, flexible quadratic polynomial, cubic polynomial and flexible cubic polynomial (with “flexible” standing for polynomials estimated separately at each side of the boundary). Several lessons can be drawn from these estimates.

First, I ﬁnd a remarkably consistent and signiﬁcantly positive “Gender effect”, consistent with the

“feminization of HIV” in Sub-Saharan Africa, already extensively discussed in the literature and in section 1.4. Second, and more importantly, I find a positive estimate of 퐹 푒푚푎푙푒 × 푀푎푡푟푖푙푖푛푒푎푙푖푡푦, significant when analysing the samples of individuals living in villages located within 150 or 200 km of the nearest ancestral matrilineal ethnic border. This result corroborates my OLS finding that females originating from ancestrally matrilineal ethnic groups suffer from significantly more HIV today.

Noteworthy, the magnitudes of my geographic RD estimates of 퐹 푒푚푎푙푒 × 푀푎푡푟푖푙푖푛푒푎푙푖푡푦 are lower than the magnitudes of my OLS estimates. A ﬁrst explanation stems from the fact that in this analysis 푀푎푡푟푖푙푖푛푒푎푙푖푡푦 indicates whether an individual lives in a village located within the ancestral boundaries of an ancestrally matrilineal ethnic group, instead of an individual’s ethnic group’s ancestral matrilineality. In addition, although one other explanation for this could be a potential bias from unobservables present in my OLS estimates (which is however unlikely according to the tests I performed in subsubsection 1.4.2.3), the difference in magnitudes might also be explained by the fact that within an ancestral matrilineal ethnic group’s territory, and close to the border, only a fraction of the population is today likely to belong to an ancestrally matrilinal ethnic group. As shown in

Figure 1.B.4, close to the border on the matrilineal side approximately 60% to 40% of the population does not belong to an ancestrally matrilineal ethnic group. This suggests that the magnitude of the

47 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

RD estimates could be biased downwards by this amount as well. In addition, ancestral ethnic groups boundaries are susceptibly blurry, and spillovers from matrilineal group to patrilineal group may arise close to the boundary, especially when it comes to sexual relationships and sexually transmitted diseases. Higher ethnic fractionalization and/or polarization at the boundary may help explain such spillovers since, as underlined by Tequame(2012), riskier sexual behaviours are more easy to conceal in fragmented societies. Alternatively, migration at the boundary could also increase sexual mixing between individuals originating from ancestrally matrilineal ethnic groups and those originating from ancestrally patrilineal ethnic groups. All in all, such spillovers would cause estimated effects at the border to be muted, and partly explain why the estimated effects fall above conventional signiﬁcance levels when restricting the sample to individuals living very close to the border (i.e. within 100km).51

1.4.3.2 Instrumenting for Ancestral Matrilineality

As a second alternative identiﬁcation strategy, I perform an instrumental variable (IV) strategy, instrumenting individual’s ethnic group’s ancestral matrilineality (푀푎푡푟푖푙푖푛푒푎푙푖푡푦 regressor) with a measure of the minimum distance (in km) between the individual’s DHS village to the nearest ancestral matrilineal ethnic boundary (based on digitized Murdock’s map of ancestral ethnic groups in Africa, see Figure 1.2). Also, I instrument my main regressor of interest 퐹 푒푚푎푙푒×푀푎푡푟푖푙푖푛푒푎푙푖푡푦 with a variable interacting 퐹 푒푚푎푙푒 dummy with this measure of distance.52

The relevance of this instrument is a priori straightforward and validated by the bivariate relationship between distance from ancestral matrilineal ethnic border and individual’s ethnic group’s ancestral matrilineality (see subsubsection 1.4.3.1 and Figure 1.B.4). The critical issue is whether the distance between an individual’s location and ancestral matrilineal ethnic border is uncorrelated with factors, other than individual’s ethnic group’s ancestral matrilineality, that may have affected individual’s sexual and contraceptive behaviour and therefore HIV susceptibility. Therefore, there remain a number of other reasons why the exclusion restriction may not be satisﬁed. First, distance between contemporary individual’s location and ancestral matrilineal ethnic boundary may be correlated with geographic characteristics (e.g. vegetation, altitude, remoteness, etc.) which might affect the spread

51The reduction in sample size is also likely at play in explaining why the effect is less precisely estimated. 52I assign positive values to this measure of distance for DHS villages located within the boundaries of a matrilineal ancestral ethnic groups, while I assign negative values for villages located within the boundaries of a non-matrilineal (i.e. patrilineal or other) ancestral ethnic group. Therefore, a distance of +100 means that a village is located within an ancestrally matrilineal area, 100 km away from the nearest matrilineal/non- matrilineal ancestral ethnic boundary; while a distance of -100 means that a village is located within an ancestrally non-matrilineal area (i.e. patrilineal or other), 100 km away from the nearest matrilineal/non- matrilineal ancestral ethnic boundary.

48 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

Table 1.3 – Geographic RD Estimates

(1) (2) (3) (4) (5) (6) Dep. Var.: HIV Distance from matrilineal boundary: 100km 150km 200km Running variable: Distance Lat./Long. Distance Lat./Long. Distance Lat./Long. Panel A: Linear Polynomial Matrilineality -0.000 0.000 -0.001 -0.001 -0.002 -0.001 (0.004) (0.004) (0.004) (0.004) (0.004) (0.004) Female 0.011*** 0.011*** 0.010*** 0.010*** 0.011*** 0.011*** (0.003) (0.003) (0.002) (0.002) (0.002) (0.002) Female × Matrilineality 0.008 0.008 0.009** 0.009** 0.009** 0.009** (0.005) (0.005) (0.004) (0.004) (0.004) (0.004) Panel B: Flexible Linear Polynomial Matrilineality 0.005 0.002 0.000 0.001 -0.001 0.000 (0.006) (0.006) (0.005) (0.005) (0.005) (0.006) Female 0.011*** 0.011*** 0.010*** 0.010*** 0.011*** 0.011*** (0.003) (0.003) (0.002) (0.002) (0.002) (0.002) Female × Matrilineality 0.008 0.008 0.009** 0.009** 0.009** 0.009** (0.005) (0.005) (0.004) (0.004) (0.004) (0.004) Panel C: Quadratic Polynomial Matrilineality 0.000 0.001 -0.001 0.000 -0.001 -0.000 (0.004) (0.004) (0.004) (0.004) (0.004) (0.004) Female 0.011*** 0.011*** 0.010*** 0.010*** 0.011*** 0.011*** (0.003) (0.003) (0.002) (0.002) (0.002) (0.002) Female × Matrilineality 0.008 0.008 0.009** 0.009** 0.009** 0.009** (0.005) (0.005) (0.004) (0.004) (0.004) (0.004) Panel D: Flexible Quadratic Polynomial Matrilineality 0.005 0.001 0.005 0.001 0.001 0.001 (0.007) (0.009) (0.006) (0.008) (0.005) (0.008) Female 0.011*** 0.011*** 0.010*** 0.010*** 0.011*** 0.011*** (0.003) (0.003) (0.002) (0.002) (0.002) (0.002) Female × Matrilineality 0.008 0.008 0.009** 0.009** 0.009** 0.009** (0.005) (0.005) (0.004) (0.004) (0.004) (0.004) Panel E: Cubic Polynomial Matrilineality -0.000 0.001 -0.001 0.000 -0.001 -0.000 (0.004) (0.004) (0.004) (0.004) (0.004) (0.004) Female 0.011*** 0.011*** 0.010*** 0.010*** 0.011*** 0.011*** (0.003) (0.003) (0.002) (0.002) (0.002) (0.002) Female × Matrilineality 0.008 0.008 0.009** 0.009** 0.009** 0.009** (0.005) (0.005) (0.004) (0.004) (0.004) (0.004) Panel F: Flexible Cubic Polynomial Matrilineality 0.007 -0.005 0.000 -0.007 0.005 -0.008 (0.010) (0.011) (0.007) (0.010) (0.006) (0.010) Female 0.011*** 0.011*** 0.010*** 0.010*** 0.011*** 0.011*** (0.003) (0.003) (0.002) (0.002) (0.002) (0.002) Female × Matrilineality 0.008 0.008 0.009** 0.009** 0.009** 0.009** (0.005) (0.005) (0.004) (0.004) (0.004) (0.004)

Baseline Controls + FE Yes Yes Yes Yes Yes Yes Observations 48,947 48,947 71,333 71,333 89,982 89,982 Clusters 68 68 73 73 76 76 Mean Dep. Var. 0.034 0.034 0.033 0.033 0.032 0.032

Notes: Geographic RD estimates based on Equation 1.2 are reported with standard errors clustered at the ethnicity level in brackets, for different bandwidths, running variables and polynomial specifications. The unit of observation is an individual originating from either a traditionally matrilineal or a traditionally patrilineal ethnic group (it therefore excludes individuals originating from ethnic groups with alternate inheritance rules (ambilineality, bilinearity, duolinearity, etc.)). “Matrilineality” indicates (dummy) whether an individual lives in a village located within the ancestral boundaries of an ancestrally matrilineal ethnic group (based on Murdock’s map of ancestral ethnic groups, see Figure 1.2). “HIV” and “Female” are defined as in Table 1.1. “Baseline Controls + FE” are the “Ind. Controls”, “Ethnic Group Controls”, “Village-Geographic Controls”, and “Region-survey FE” defined in Table 1.1. * p<0.10,** p<0.05,*** p<0.01 49 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa of HIV. Second, this contemporary measure of distance may also be correlated with the geographical distribution of economic activities, infrastructures, transportation networks, etc. and therefore reﬂects migration patterns that appeared in the meantime (i.e. since pre-colonial period) and which might have affected contemporary HIV rates. To alleviate such concerns I include in my IV regressions region (within-country) × survey (time) ﬁxed effects, and I directly control for such alternative channels with a large array of ethnic group’s historical controls and village-level geographic controls, which are already detailed in subsection 1.3.3 and subsubsection 1.4.2.3.

Table 1.4 presents estimates of my instrumental variable strategy. The OLS estimate of regressing my measure of distance on individual’s ethnic group’s ancestral matrilineality, as well as F-Stat of the test that the coefficient of the instrument is null are reported in column 2. They suggest that my instrument is a strong predictor of individual’s ethnic group’s ancestral matrilineality. Similar conclusions can be drawn from the OLS estimate and F-Stat of regressing 퐹 푒푚푎푙푒×푀푎푡푟푖푙푖푛푒푎푙푖푡푦 on the variable interacting the 퐹 푒푚푎푙푒 dummy with my instrument, reported in column 3. More importantly, the IV-2SLS estimate reported in column 1 confirms my OLS result and indicate that women originating from ancestrally matrilineal ethnic groups significantly suffer from more HIV today.

Noteworthy, instrumenting two potentially endogenous regressors with two instruments and assuming clustered standard errors, I follow Andrews(2018) and compute weak instrument-robust 95% conﬁdence interval of my IV estimate.53 This latter is reported in square brackets in Table 1.4 and suggests a non-null positive IV estimate of 퐹 푒푚푎푙푒 × 푀푎푡푟푖푙푖푛푒푎푙푖푡푦.54

1.4.3.3 Nearest Neighbor Matching

As a ﬁnal alternative identiﬁcation strategy to individual-level OLS regressions, I use nearest neighbor matching to compare each DHS village located in an ancestrally matrilineal area to the DHS village located in an ancestrally non-matrilineal (i.e. patrilineal or other) area55 that is the most similar in terms of geographic characteristics. The matching average treatment over treated effects are reported in Table 1.5, with village pairs being matched56 on the full range of village geographic con-

53I use “twostepweakiv”, a Stata package developed by and presented in Sun(2018). I compute two-step weak instrument-robust 95% confidence interval, projected on 퐹 푒푚푎푙푒 × 푀푎푡푟푖푙푖푛푒푎푙푖푡푦 regressor. 54Calculating 95% two-step weak instrument-robust CI based on Andrews(2018) and Sun(2018), I find a 12% distortion cutoff, which suggests that my instruments are strong instruments, in the sense thatsize distortions are below 12% for conventional 95% confidence intervals. 55Based on Murdock’s map of ancestral ethnic groups (see Figure 1.2). 56I use nearest neighbor matching based on Mahalanobis distance. Estimates are corrected for bias due to matching on multiple continuous variables, based on Abadie and Imbens(2006) and Abadie and Imbens(2011).

50 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

Table 1.4 – IV Estimates of the Effect of Ancestral Matrilineality on Female HIV

(1) (2) (3) IV-2SLS OLS (First Stages) HIV Matrilineality Female × Matri.

Matrilineality -0.010 (0.036) Female 0.006*** Yes Yes (0.002) Female × Matrilineality 0.029*** (0.010) [0.013;0.069] Distance to matri. boundary 0.0003*** (0.0001) Female × Distance to matri. boundary 0.0008*** (0.0002) Ind. Controls Yes Yes Yes Ethnic Group Controls Yes Yes Yes Village-Geographic Controls Yes Yes Yes Region-survey FE Yes Yes Yes

Observations 179,312 179,312 179,312 Adj. R-squared 0.032 0.835 0.617 Clusters 100 100 100 Mean Dep. Var. (Patri. Males) 0.023 0.158 0.085 F-Stat (test coeff. = 0) 10.60 10.04 Notes: IV-2SLS estimates are reported in column 1; first stage OLS estimates are reported in column 2 and 3; with standard errors clustered at the ethnicity level in brackets. Weak instrument-robust 95% confidence interval (Andrews, 2018, Sun, 2018) is reported in square brackets (calculating 95% two-step weak instrument-robust CI, the distortion cutoff is 12%, indicating strong instruments). The unit of observation is an individual originating from either a traditionally matrilineal or a traditionally patrilineal ethnic group (it therefore excludes individuals originating from ethnic groups with alternate inheritance rules (ambilineality, bilinearity, duolinearity, etc.)). “Distance to matri. boundary” is computed on QGIS as the minimum distance between DHS village and the nearest ancestral matrilineal ethnic boundary (based on Murdock’s map of ancestral ethnic group boundaries, see Figure 1.2). This takes negative values for DHS villages located within boundaries of non-matrilineal ancestral ethnic groups. “HIV”, “Matrilineality”, “Female”, “Female × Matrilineality”, “Ind. Controls”, “Ethnic Group Controls”, “Village-Geographic Controls”, and “Region-survey FE” are defined in Table 1.1. * p<0.10,** p<0.05,*** p<0.01

trols of Equation 1.1. As reported in column 1, villages located in ancestrally matrilineal areas are characterized by average female HIV rates that are 2.9 percentage points higher than their nearest neighbor village located in an ancestrally non-matrilineal area. This effect is signiﬁcant at 1% and of large magnitude, representing about 78% of the average village’s female HIV rate of villages located in an ancestrally non-matrilineal area. Interestingly, no such effect is found on village’s proportion of HIV positives males57, consistent with my hypothesis that the effect of ancestral matrilineality on contemporary HIV rates is speciﬁc to female populations.

57The smaller sample size is explained by the fact that in some DHS villages only females were interviewed.

51 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

Table 1.5 – Nearest Neighbor Matching (ATT)

(1) (2) Village’s proportion of Village’s proportion of HIV positive females HIV positive males

Ancestrally matrilineal area 0.029*** 0.007 (0.011) (0.011)

Observations 13,176 12,710 Mean Dep. Var. (Villages in non-matri. area) 0.037 0.023 Notes: This table reports the average treatment effect on the treated (ATT) on the proportion of village’s HIV positive females in column 1, and males in column 2, between DHS villages located on an ancestrally matrilineal area and DHS villages located on an ancestrally non-matrilineal area (patrilineal or other), using nearest neighbor matching. Villages are matched using the Mahalanobis distance function based on all Village-Geographic Controls detailed in Table 1.1. Estimates are corrected for bias due to matching on multiple continuous variables (Abadie and Imbens(2006); Abadie and Imbens(2011)). Abadie and Imbens robust standard errors are reported in brackets. * p<0.10,** p<0.05,*** p<0.01

1.5 Mechanisms

The previous section has uncovered a robust relationship between ancestral matrilineality and the contemporary spread of the HIV epidemic among female individuals in Sub-Saharan Africa. Accord- ing to my conceptual framework, the legacy of ancestral matrilineality on women’s sexual preferences and ability to implement them has led matrilineal women to adopt sexual and contraceptive behaviours that favor a higher rate of transmission of HIV today. In this section, I turn to a more direct investigation of these channels, by exploring the empirical relationship between ancestral matrilineality and female’s sexual autonomy, sexual and contraceptive behaviour

1.5.1 Female Sexual Autonomy

First, I explore whether women originating from ancestrally matrilineal ethnic groups beneﬁt from a higher social status, and consequent higher bargaining power, which would allow them to beneﬁt from a higher sexual autonomy. First, building on the household economics literature highlighting the role of marriage outside option on intrahousehold bargaining power (Baland and Ziparo, 2018;

Bargain et al., 2020), I create a dummy equals to one if a female is currently divorced. Then, following

Anderson(2018), I use information provided in DHS on land and house ownership, as a measure of female property rights in case of divorce (restricting my sample to divorced women), and I create a dummy which is equal to one if a divorced female reports owning a house and/or a land. Finally, I

52 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa construct a dummy equals to one if there is at least one wife’s parent in the household.58 Working with these outcomes allow me to check whether women originating from ancestrally matrilineal ethnic groups beneﬁts from better marriage outside options and, consequently, higher bargaining power within marriage.

To further explore how this translates into female sexual autonomy, I emphasize a second set of outcomes which pertain to women’s decision-making regarding contraception. In particular, I use answer to the question asked to women in union currently using contraception about who decides on the use of contraception, and I create a dummy equals to one if a woman reports being decision- maker. Other questions in DHS directly cover sex negotiations and autonomy and asks whether a woman could get a male condom herself, whether a wife is justified to ask husband to use condom if he has a STI (Sexually Trasmitted Infection); and whether it is justified for a women to refuse sex with her husband if he has another women. For each of these questions, I create a dummy equals to 1 if a woman answers affirmatively.

For the analysis in this subsection, I focus on females and estimate the following equation:

′ ′ ′ 푦푖푒푣푟푡 = 훼 + 훽1푀푎푡푟푖푙푖푛푒푎푙푖푡푦푒 + XievrtΔ + XertΩ + XvrtΠ + 휆푟푡 + 휀푖푒푣푟푡 (1.3)

′ ′ ′ where Xievrt, Xert, Xvrt, 휆푟푡 and 휀푖푒푣푟푡 are individual-level, ethnic group-level and village-level controls and (within-country) DHS region-survey (year) ﬁxed effects, deﬁned as in subsection 1.3.3.

Standard errors are clustered at the ethnic group level. My coefﬁcient of interest is now 훽1 and captures the long-term effect of ancestral matrilineality on female populations.59

In Table 1.6, I report results from estimating Equation 1.3 on these outcomes. I find that ancestral matrilineality is indeed significantly positively associated with all these dimensions of female’s social status, bargaining power and sexual autonomy. More specifically, according to the estimates reported in column 1, matrilineal women are 1.5 percentage points significantly more likely to be divorced, meaning twice more likely to be divorced than patrilineal females.60 Also, according to the estimate reported in column 2, I find that divorced women originating from ancestrally matrilineal ethnic

58Bargain et al.(2020) show the influence of the presence of wife’s family in the household onwife’s intrahousehold decision-making, in the Indonesian context. 59Remember that I restrict my sample to individuals originating from an either ancestrally matrilineal or an ancestrally patrilineal ethnic group. Therefore, 푀푎푡푟푖푙푖푛푒푎푙푖푡푦 = 0 means that the individual originates from an ancestrally patrilineal ethnic group (comparison group). 60The mean proportion of divorced patrilineal females in my regression sample is 1.5%.

53 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa group are 23.5 percentage points (representing about 65% of the mean probability of the patrilineal females in my sample) more likely to own a house and/or a land, as compared to their divorced patrilineal counterparts. I additionally ﬁnd in column 3 that matrilineal females are 0.08% more likely to have at least one parent present in their household, an effect of large magnitude given that only 2% of patrilineal females beneﬁt from the presence of at least one of their parents in their household.

This result is consistent with the correlation underlined in subsubsection 1.4.2.3 between ancestral matrilineality and ancestral matrilocality. Further, estimates reported in column 3 to 6 indicate, in the same order, that matrilineal women are 4.3 percentage points more likely to be decision-maker regarding the use of contraception; 2.6 percentage points more likely to answer that they could get

(male) condom themselves; 3 percentage points more likely to find justified for a woman to ask her husband to use a condom if he has a STI; and 2.8 percentage points more likely to find justified for a woman to refuse sex with her husband if he has another women.

All in all, in line with my conceptual framework, I can rule out Anderson(2018) mechanism, according to which HIV prevalence should be higher for less empowered females since they should be less able to impose safe sexual practices to their husbands. I propose two main other mechanisms in the two next subsections. First, I explore whether the higher sexual autonomy of women originating from ancestrally matrilineal ethnic groups translates into them implementing their preferred sexual strategy, and thereby translates into them adopting riskier sexual behaviours. Further, I explore how matrilineal females’ higher decision-making power regarding contraception translates into contraceptive behaviours that are incidentally less protective.

54 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa refuse sex if husband has other women Wife justified has STI if husband is a subnational region defined ask condom Wife justified condom Can get Region-survey Decide Contraception in HH Wife’s Parents (divorced) Own house and/or land * p<0.10,** p<0.05,*** p<0.01 Ancestral Matrilineality and Female Sexual Autonomy (OLS) (1) (2) (3) (4) (5) (6) (7) (0.004) (0.108) (0.003) (0.010) (0.013) (0.011) (0.013) Divorced Table 1.6 – OLS estimates are reported with standard errors clustered at the ethnicity level in brackets. The unit of observation is MatrilinealityInd. ControlsEthnic Group ControlsVillage-Geographic ControlsRegion-survey FE Yes 0.015***Observations YesAdj. R-squared 0.235**Clusters Yes YesMean Dep. 0.008*** Var. (Patri.) Yes Yes 0.043*** Yes Yes 105,964 0.015 0.026** Yes Yes 0.041 1,191 0.030*** Yes 0.359 Yes 0.316 Yes 100 Yes 66,902 0.028** 0.020 0.026 Yes Yes 13,709 Yes 54 Yes 0.871 0.054 Yes Yes 56,270 Yes Yes 100 0.632 0.123 91,420 Yes Yes 0.772 Yes 79 0.157 Yes 95,604 Yes 0.622 Yes 0.139 99 100 99 Notes: a woman originating from either a traditionally matrilinealan or individual a is traditionally currentlyhousehold. patrilineal divorced. ethnic “Own “Wife’s group. house Parents “Matrilineality” and/or infemales indicates land HH” sample). (divorced)” is “Decide is a contraception” a dummy is indicating dummy a indicating the dummy whether presence indicatinghusband an of has whether at individual STI” an least owns is individual a oneSTI. a is house wife’s “Wife dummy decision-maker and/or parents justified indicating regarding in a refuse whether contraception. refuse land an sex (divorced individualsex with find if husband her justifiedin husband has DHS, if for other interacted he awomen” with has wife its other is survey-year. women.to a dummy herask Controlsindicating husband are defined whether in a respondant to useTable 1.1 . find a condom if he has justified for a wife to (dummy) whether an individual belongs to a traditionally matrilineal ethnic group. “Divorced” is a dummy indicating whether “Can get condom” is a dummy indicating whether an individual can get herself a male condom. “Wife justified ask condom is

55 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

1.5.2 Sexual Behaviour

Reproductive Evolved Psychology and Sexual Behaviour. In order to capture sexual behaviours that are prominent risk factors of HIV contagion, I follow the well-established medical and economic literature (e.g. Bertocchi and Dimico, 2019), and I create a dummy equals to one if an individual had any sexual activity in the last 4 weeks, a dummy equals to one if an individual reports having any extramarital affair in the last 12 months, and the number of extramarital partners in the last 12 months that she reports (I focus on formally married individuals for these two latter outcomes).

Several lessons can be drawn from Table 1.7, which presents results from estimating Equation 1.1 on these sexual behaviour outcomes. To begin with, according to the estimates reported in the second row, sub-saharan african women are signiﬁcantly and sizably less likely to report adopting risky sexual behaviours than their male counterparts. This result is consistent with the evolutionary psychology theory I build on in my conceptual framework, as well as with numerous psychological studies

(Buss, 2016) according to which, relative to females, males exhibit a stronger preference for casual relationships and sexual variety. However I also ﬁnd that, among females, women originating from ancestrally matrilineal ethnic groups have signiﬁcantly riskier sexual behaviours, that are more conducive to HIV. More precisely, estimates reported in the third row indicate that matrilineal married women are 4.1 percentage points more likely to have had a sexual activity in the last 4 weeks (column

2). Interestingly, I ﬁnd qualitatively similar results when working on individuals being in any marital status, while controlling for this latter (column 1). Further, I ﬁnd that matrilineal married females are 2.6 percentage points (representing about 20% of the patrilineal married males’ mean probability) more likely to report any extramarital affair in the last 12 months. In addition, estimate from column

4 indicates that married women originating from an ancestrally matrilineal ethnic group had 0.068 more extramarital partners in the last 12 months, an effect of large magnitude representing about 33% of the mean number of extramarital partners of patrilineal married males. Finally, it is worth noting that according to the estimates reported in the ﬁrst row, and contrary to their female counterparts, matrilineal males do not have an overall signﬁcantly different propensity to adopt risky sexual behaviour than their patrilineal counterparts. All these results are fully consistent with my conceptual framework and highlight mechanisms at play in explaining the long-term effect on ancestral matrilineality of contemporary female HIV in Sub-Saharan Africa.

56 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

Routes of Infection. Building on my conceptual framework and the results on sexual behaviours highlighted above, I expect matrilineal females to be mainly infected through extramarital routes of infection. This is exactly what I ﬁnd in the analysis on couple’s serodiscordancy status, performed in

Table 1.C.8.

Restricting my attention to non-polygynous formally married couples with both wife and husbands tested for HIV in DHS and both of them originating from either a traditionally matrilineal or a traditionally patrilineal ethnic group, I ﬁnd that when the wife originates from an ancestrally matrilineal ethnic group, she is about twice more likely (i.e. 1.3 percentage points) to be HIV positive while having a HIV negative husband, relative to wives originating from ancestrally patrilineal ethnic groups. Importantly, this result remains marginally signiﬁcant when I also include individual’s, (ancestral) ethnic group’s and village-geographic controls computed for the husband, in addition of controlling for whether the wife and the husband originate from different ancestral ethnic groups (“Mixed ethnicity”), as well as controlling for whether the wife and the husband originate from ethnic groups with different ancestral kinship organizations (matrilineal vs. patrilineal) (“Mixed matrilineality”).

All in all, these results indicate that, in line with their riskier sexual behaviour outside the domestic sphere, matrilineal women are signiﬁcantly more likely to be infected by HIV through extramarital channels.

Moreover, results in the two last columns indicate that matrilineal couples are relatively less likely to be HIV+ seroconcordant. This ﬁnding may be interpreted as an evidence that the mechanisms highlighted by Anderson(2018) is also at play, namely that patrilineal females are more likely to be infected by their husband due to their lower sexual autonomy. Alternatively, this ﬁnding may also be explained by differential couple formation/dissolution dynamics between matrilineal and patrilineal groups (e.g. a matrilineal female may be relatively more likely to divorce61 if her husband is HIV positive. In this case, the couple is dropped from my sample of couples, leading to non-random attrition).

While DHS data do not allow me to directly empirically test for the effect of matrilineality on gender-related social behaviour, it is worth noting that a recent literature has highlighted the inﬂuence of matrilineality on such gender related behaviour. In particular, Lowes(2018a) provides experimental evidence from Democratic Republic of Congo (DRC) that spouses from matrilineal ethnic groups cooperate less with each other than their patrilineal counterparts. Further, Lowes(2018b) provides, in

61See my result on divorce in Table 1.6.

57 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa a similar context, experimental evidence that matrilineality closes the gender gap in risk-preference, with matrilineal women having a higher preference for risk than their patrilineal counterparts. These differences in contemporary behaviours could be additional factors shaping riskier sexual behaviour found here, and within marriage contraceptive behaviour explored in the next subsection.

Table 1.7 – Ancestral Matrilineality and Sexual Behaviour (OLS)

All Sample Formally Married Individuals

Nb. of Sexual Sexual Infidelity extramarital activity activity partners (1) (2) (3) (4)

Matrilineality -0.005 -0.017 -0.006 -0.038 (0.011) (0.016) (0.010) (0.024) Female -0.079*** -0.094*** -0.127*** -0.196*** (0.006) (0.007) (0.014) (0.027) Female × Matrilineality 0.032* 0.041** 0.026* 0.068** (0.017) (0.020) (0.015) (0.027) Ind. Controls Yes Yes Yes Yes Ethnic Group Controls Yes Yes Yes Yes Village-Geographic Controls Yes Yes Yes Yes Region-survey FE Yes Yes Yes Yes

Observations 181,108 104,483 104,489 104,489 Adj. R-squared 0.310 0.087 0.109 0.022 Clusters 100 99 99 99 Mean Dep. Var. (Patri. Males) 0.537 0.768 0.136 0.204 Notes: OLS estimates are reported with standard errors clustered at the ethnicity level in brackets. The unit of observation is an individual originating from either a traditionally matrilineal or a traditionally patrilineal ethnic group. “Matrilineality” indicates (dummy) whether an individual belongs to a traditionally matrilineal ethnic group. “Female” indicates (dummy) whether an individual is a female. “Female × Matrilineality” indicates (dummy) whether an individual is a female belonging to a traditionally matrilineal ethnic group. “Sexual activity” is a dummy indicating whether an individual reports having had any sexual activity in the last 4 weeks. “Infidelity” is a dummy indicating whether an individual reports having had any extramarital partner in the last 12 months. “Nb. of extramarital partners” is the number of extramarital partners in the last 12 months reported by an individual. Controls are defined in Table 1.1. Region-survey is a subnational region defined in DHS, interacted with its survey-year. Samples in columns2 to 4 consist of formally married individuals. * p<0.10,** p<0.05,*** p<0.01

1.5.3 Contraceptive Behaviour

Condom versus Long-Term Contraceptive Methods. The results highlighted in subsection 1.5.1 indicated that matrilineal females were more likely to be decision-maker regarding contraception, suggesting that matrilineal females may also differ from their patrilineal counterparts in their contraceptive behaviour. This is what I investigate in this subsection.

58 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

To do so, I exploit information in DHS on current contraception method used by respondant to create (1) a dummy for whether the respondant reports using condom as her current contraceptive method; (2) a dummy for whether the respondant reports using a long-term contraceptive method

(i.e. “pill”, “IUD”, “injection”, “female sterilization”, “implants/norplant”, or “lactational amenorrhea (LAM)”); (3) and a dummy for whether the respondant reports using any contraceptive method.

Incidentally, long-term contraceptive methods are less protective methods against HIV, relative to condom (Anderson, 2018).

Table 1.8 presents results from estimating Equation 1.1 on these contraceptive use outcomes.

These results reveal that, while not having a significantly different contraceptive behaviour at the extensive margin (i.e. the estimate of 퐹 푒푚푎푙푒 × 푀푎푡푟푖푙푖푛푒푎푙푖푡푦 on the probability of using any contraception method is not significant (column 3)), matrilineal women significantly differ from their patrilineal counterparts in the contraceptive methods they use. More specifically, matrilineal women are 3.1 percentage points less likely to report using condom as a current contraceptive method, an effect representing about 22% of the mean probability that patrilineal males report using condom as a contraceptive method (column 1).62 This very low level of condom use within marriage in

Sub-Saharan Africa is documented in the literature.63 On the contrary, I also ﬁnd in column 2 that matrilineal women are 4.7 percentage points more likely than their patrilineal counterparts to report using a long-term contraceptive method, representing about 60% of the mean of patrilineal males.

In sum, it seems that, having a higher status and being more likely to be decision-maker regarding contraception (see subsection 1.5.1), matrilineal women are more likely to bear the responsability of contraception, and substitute condom with long-term contraceptive methods. Doing so, matrilineal females incidentally substitute protective contraceptive methods with less protective ones. This is in line with Islam et al.(2009), who find similar patterns in the context of Bangladesh between matrilineal Garo and patrilineal Bengali women. Indeed, they find that while matrilineal Garo women’s use of contraceptive is higher than the national level, condom use among Garo women is lower than at national level. On the contrary, Garo women’s contraceptive methods are highly skewed towards female long-term methods, in comparison to national levels. In line with my findings, they argue that this may be due to the matrilineal nature of the Garo society where wives take most of the family planning responsibilities and rely less on their husbands.

62Note that this mean probability is even lower for females, with only 3.1% (not reported) of females in my final sample reporting condom as a current contraceptive method. 63Chimbiri(2007) calls condom an “intruder in marriage”.

59 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

An other driver of such contraceptive behaviour may be the desired fertility of matrilineal females.64 Indeed, as reported in Table 1.C.9 in appendix, I find a higher gender discrepancy in desired fertility for the matrilineal group. In other words, I find that matrilineal females are relatively more likely to desire more children than their patrilineal counterparts, while the opposite is true for male individuals. Along this line, I find a similar result regarding actual fertility. These results may be fully rationnalized in my conceptual framework: matrilineal male’s biological children will never integrate his lineage in addition of not inheriting from him. On the contrary, as highlighted in Lowes(2018a), by integrating her lineage and allowing her to benefit from support from her brothers, children constitute assets for a matrilineal woman, which may drive her fertility preferences.

All in all, differences in contraceptive behaviours highlighted in this subsection, and in particular matrilineal females’ lower propensity to use condom, constitute an additional mechanism that help explaining the highest contemporary rates of HIV found within matrilineal female populations.

64Note that I do control for the number of children in my regressions.

60 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

Table 1.8 – Ancestral Matrilineality and Contraception (OLS)

Long-term Any Condom contraceptive contraceptive method method (1) (2) (3)

Matrilineality 0.018* -0.017 0.005 (0.010) (0.011) (0.014) Female -0.106*** 0.046*** -0.058*** (0.017) (0.008) (0.019) Female × Matrilineality -0.031* 0.047*** 0.017 (0.018) (0.014) (0.020) Ind. Controls Yes Yes Yes Ethnic Group Controls Yes Yes Yes Village-Geographic Controls Yes Yes Yes Region-survey FE Yes Yes Yes

Observations 147,175 146,185 147,175 Adj. R-squared 0.155 0.155 0.156 Clusters 84 84 84 Mean Dep. Var. (Patri. Males) 0.138 0.078 0.245 Notes: OLS estimates are reported with standard errors clustered at the ethnicity level in brackets. The unit of observation is an individual originating from either a traditionally matrilineal or a traditionally patrilineal ethnic group. “Matri- lineality” indicates (dummy) whether an individual belongs to a traditionally matrilineal ethnic group. “Female” indicates (dummy) whether an individual is a female. “Female × Matrilineality” indicates (dummy) whether an individual is a female belonging to a traditionally matrilineal ethnic group. “Condom” is a dummy indicating whether an individual reports “condom (male)” as her current contraceptive method. “Long-term contraceptive method” is a dummy indicating whether an individual reports a long-term contraceptive method (i.e. “pill”, “IUD”, “injection”, “female sterilization”, “implants/norplant”, or “lactational amenorrhea (LAM)”)). “Any contraceptive method” is a dummy indicating whether an individual reports currently using any contraception method. Con- trols are defined in Table 1.1. Region-survey is a subnational region defined in DHS, interacted with its survey-year. * p<0.10,** p<0.05,*** p<0.01

Condom Use and Internalized Risk. My previous finding that matrilineal females have a lower use of condom may hide substantial heterogeneity. I investigate in Table 1.C.10 in appendix whether matrilineal individuals adopt different condom use behaviour when they have properly internalized the risk of HIV transmission. More precisely, I first investigate heterogeneity by perception of the condom as a mean of reducing HIV transmission. I find that the negative effect of matrilineality on condom use that I previously found is null for individuals who have a correct belief about the role of condom. Moreover, I find that when they are seropositive, matrilineal individuals are in fact more likely to report using condom as a contraceptive method. Even more strikingly, I find that this heterogeneity is mainly driven by individuals who have ever been tested for HIV before DHS, and who are therefore presumably aware of their serostatus.

61 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

All in all, it seems that, when they have correctly internalized the risk of transmitting the virus, matrilineal females take advantage of their higher ability to impose safe sexual practices. However, this result has to be interpreted cautiously as some biases may arise here. For example, individuals who are HIV positive in my sample may systematically differ from HIV negative individuals in unobservables that are not fully accounted for in my regressions (think about unobserved sexual promiscuity, selection into HIV testing or selective mortality). The same limitations arise when I compare individuals who have been tested in the past with those who have never been tested.

Still, the heterogeneities I uncover are in line with Thornton(2008) who provides experimental evidences from Malawi that HIV-positive individuals learning their status are more likely to purchase condoms, while learning HIV-negative status has no effect. Further, these ﬁndings are encouraging as they suggest that policies targeting empowered women should raise their awareness about the actual riskiness of promiscuous sexual behaviour in order to induce behavioural changes restraining the spread of HIV.

1.5.4 Discarded Mechanisms

Acknowledgment of HIV Risks and Access to Condom. To explore whether matrilineal females’ adoption of riskier sexual and contraceptive behaviours may result from a lower acknowledgment of

HIV risks and access to condom, I create several outcomes from information in DHS on acknowledgment of HIV risks and role of condom in reducing these risks. I create a dummy indicating whether an individual has ever heard about AIDS (“Heard of AIDS”); a dummy indicating whether an individual has ever heard about any STI (“Heard of STI”); a dummy indicating whether an individual thinks that always using condom during sex reduces chance of getting HIV (“Think condom reduces HIV”); and a dummy indicating whether an individual thinks that having only one sexual partner reduces chance of getting HIV (“Think having one partner reduces HIV”). Further, I also create outcomes related to access to condom, namely a dummy indicating whether an individual knows a source to get male condoms (“Know a source to get condom”); as well as a dummy indicating whether an individual can get herself a male condom (“Can get condom”).

Results from regressing Equation 1.1 on these outcomes are reported in Table 1.C.11. Two main conclusions can be drawn from this analysis: ﬁrst, I do not ﬁnd any evidence that matrilineal females suffer from a lower acknowledgment of risks than their patrilineal counterparts (in fact, the vast

62 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa majority of both matrilineal and patrilineal females acknowledge such risks65). Consequently, I can rule out differential acknowledgment of HIV risks as a factor explaining matrilineal females’ adoption of more promiscuous sexual behaviour. Second, matrilineal females’ lower use of condom cannot be explained by a lack of access to condom. If any, I conversely ﬁnd that, in line with their higher sexual autonomy, matrilineal women have an overall signiﬁcantly easier access to condom.

These results, as well as those found in the previous subsection on condom use behaviour suggest that this is not HIV awareness per se that induces behavioural changes, but rather the perception of the risk of being HIV positive, and transmitting the virus afterwards. This is lin ine with the ﬁndings of the literature exploring the inﬂuence of HIV risk perception on sexual and contraceptive behaviours

(Oster, 2012a, Paula et al., 2014, Delavande and Kohler, 2015), which is detailed in the introduction.

Sexual Debuts. For biological reasons, young women constitute a population at high risk of contracting HIV when exposed to it (Yi et al., 2013). As such, increase in early and/or premarital sexual activity is a driver of the spread of the virus (Oster, 2005, Case and Paxson, 2013). To test whether this could be a driver of matrilineal females’ highest rate of HIV, I exploit information in DHS on reported age at first sexual intercourse, as well as age at first marriage. Interestingly and as reported in Table 1.C.12 in appendix, I find that matrilineal females start their sexual life later than their patrilineal counterparts, suggesting that my main result is in fact mitigated by this mechanism. Further, my results on age at first marriage also suggest a relatively lower age gap between matrilineal spouses, which additionally mitigates my main findings on female sexual behaviour and HIV. As a matter of fact, a higher age gap with her husband may constitute a source of dissatisfaction for a married wife, potentially encouraging her to engage into extramarital relationships (Bertocchi and Dimico, 2019).

Furthermore, if one assumes that an older husband has a higher likelihood of being HIV positive

(because of a longer sexual life) and is consequently more likely to infect his wife, this result may partially explain why, contrary to Anderson(2018), I mainly capture a female’s extramarital route of infection.

Antiretroviral Therapy (ART). The introduction of the Antiretroviral Therapy (ART) has been a major policy in Africa to ﬁght against the spread of HIV. However, evidences of the efﬁciency of

ART are mixed, since it has two opposite effects on the prevalence of the virus. On one hand, ART is

65In my sample, about 98% of matrilineal females and about 94% of patrilineal females have ever heard of HIV.

63 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

expected to lower the prevalence of HIV by lowering the viral load of a serepositive person and makes

her feel healthier, live longer, and be less likely to pass the virus on. Greenwood et al., 2019 call it

an “Equilibrium Effect”. On the other hand, ART may lead to what Greenwood et al., 2019 call a

“Behavioural Effect”: since the existence of ART makes life with HIV more tolerable, this may lead

healthy-feeling infected people to engage in riskier behaviour (see De Walque, Kazianga and Over,

2012 for evidences from Mozambique; see Cohen, Montandon, Carrico, Shiboski, Bostrom, Obure,

Kwena, Bailey, Nguti and Bukusi, 2009 and Friedman, 2018 for evidences from Kenya). Moreover,

beneﬁting from a longer life-expectancy, HIV-infected people have more time to pass the virus on.

All in all, according to Greenwood et al., 2019, which effect dominates depends on the fraction of

people treated.66

Unfortunately DHS does not provide data on ART. I am therefore not able to more formally

explore ART as a mechanism of my main result on HIV. Using country-level data from UNAIDS

on ART coverage in the 18 countries covered in this paper, I ﬁnd a signiﬁcant positive association

between the proportion of countries’ citizens having a matrilineal ancestor and ART coverage (not

reported).67 While illustrative, this ﬁnding does not affect my results as time-varying country-level

ART coverage is explicitly controlled for in my regressions through the inclusion of (within-country)

region-survey (year) ﬁxed effects.

1.6 An Epidemiological Approach

In this section, I adopt an epidemiological approach to conduct a simulation exercise. The aim

of this simulation exercise is not to embrace the full complexity of the world, but rather to provide

some insights on how the differences in sexual behaviours found in the previous section translate

into different gender-speciﬁc HIV rates dynamics, between ancestrally matrilineal and ancestrally

patrilineal societies.

1.6.1 The Model

To illustrate the differential effects of being matrilineal or patrilineal on female and male HIV prop-

agation rates, through the two main routes of infection highlighted in the empirical results, namely

66According to their model, the “Equilibrium Effect” dominates when large proportions of the population are treated, while the “Behavioural Effect” dominates when only small proportions of the population are treated. 67UNAIDS data portal can be found using the following link: https://aidsinfo.unaids.org/. Data on matrilineality are from Giuliano and Nunn, 2018.

64 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa extra-couple casual sexual relationships vs. intra-couple long-term sexual relationships, I adopt a compartmental SI (“Susceptible to Infected”) epidemic model, that I adapt from Worden and Porco

(2017) and where I add heterogeneity by gender.

I consider a closed population composed of female and male individuals (denoted with subscript f and m respectively). Further, each of these gendered populations are composed of a fraction of

“susceptible” (i.e. not infected) individuals, and a fraction of “infected” individuals, denoted 푠푓;푚 and 푖푓;푚 respectively.

Sexual behaviours are represented by the following parameters: 1 is an indicator of a long-term committed relationship; 훼푓;푚 is the number of extramarital partners; 휇푓;푚 is the number of sexual intercourses with the long-term committed partner; 휏푓;푚 is the number of sexual intercourses with each extramarital partner. Further, the probability that a male is infected per each sexual intercourse with a HIV positive female is 훿푚 = 훾푚 × 휌푚, namely the product of the probability that he does not use condom during sexual intercourse with the transmission risk for one-time male unprotected sex

(assuming for simplicity that condom is totally efﬁcient against HIV transmission). The probability that a female is infected per each sexual intercourse with a HIV positive male is 훿푓 = 훾푓 × (휌푚 × 휔푓 ), with 휔푓 capturing gender difference in biological susceptibility of contracting the virus when exposed to it.68 Finally, I assume random mixing, namely that the probability that an individual faces a HIV positive sexual encounter (should it be her long-term committed partner or an extramarital casual partner) is equal to the proportion of infected individuals in the population of the other sex (i.e. 푖푚 for females’ probability and 푖푓 for males’ probability).

Female propagation rates (휆푓 ) can thereby be modeled as follows:

휆푓 = 1휇푓 훿푓 푖푚 + 훼푓 휏푓 훿푓 푖푚 (1.4) ⏟ ⏞ ⏟ ⏞ Intra-couple Extra-couple infection infection

68The medical literature highlights that females have a relatively higher susceptibility of contracting HIV when exposed to it, for biological reasons (i.e. larger surface area of mucosal HIV exposure; increased mucosal expression of the HIV co-receptor CCR5; and a greater probability of virus exposure on the rectal mucosa among other factors (Yi et al., 2013)).

65 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

⇔

휆푓 = ( 1휇푓 + 훼푓 휏푓 )훿푓 푖푚 (1.5) ⏟ ⏞ ⏟ ⏞ Intra-couple Extra-couple intercourses intercourses

Symmetrically, male’s propagation rate (휆푚) can be modeled as follows:

휆푚 = (1휇푚 + 훼푚휏푚)훿푚푖푓 (1.6)

From Equation 1.5 and Equation 1.6, gender difference (휆푓 − 휆푓 ) in propagation rates can be modeled as follows:

휆푓 − 휆푚 = (1휇푓 + 훼푓 휏푓 )훿푓 푖푚 − (1휇푚 + 훼푚휏푚)훿푚푖푓 (1.7)

⇔

휆푓 − 휆푚 = 1(휇푓 훿푓 푖푚 − 휇푚훿푚푖푓 ) + 훼푓 휏푓 훿푓 푖푚 − 훼푚휏푚훿푀 푖푓 (1.8) ⏟ ⏞ ⏟ ⏞ ⏟ ⏞ Gender difference Female Male in intra-couple extra-couple extra-couple infection infection infection

Interestingly, if one assumes that 휇푓 = 휇푚, gender difference in intra-couple infection reduces simply to the gender difference in biological susceptibility of contracting the virus when exposed to it, and the difference in the probability that the husband versus the wife is infected.

It is worth noting that, for simplicity, I assume exogenous and time invariant sexual behavioural parameters (1; 훼푓;푚; 휇푓;푚; 휏푓;푚 and 휆푓;푚). Further, I assume random mixing. In other words, I assume that the probability that and individual face a HIV positive partner is equal to the proportion of HIV positive individuals in the pool of the opposite sex (푖푓;푚). Additionally, I also assume that this probability is the same for a long-term committed partner or an extramarital partner. Moreover,

I assume time-invariant contagiousness. Those simplifying assumptions might be relaxed if one is interested in embracing with greater care the full complexity of the world, which is beyond the scope of this simulation exercise.

66 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

1.6.2 The Effect of Matrilineality: Comparative Statics

Females propagation rates (휆푓 ). From Equation 1.5, it is straightforward to see that 휆푓 is increasing in all females’ sexual behavioural parameters. Further, according to my conceptual framework, as well as my empirical results on sexual behaviour and contraceptive use, the value of these parameters should be higher for matrilineal females, relative to their patrilineal counterparts, leading to a higher HIV propagation rate for matrilineal female populations. Further, 휆푓 is also increasing in 푖푚.

Therefore, assuming that 푖푓 and 푖푚 coevolve in the same direction, I expect a self-reinforcing effect for matrilineal females: since matrilineal females are characterized by higher infection rates, more matrilineal males should be contaminated to, leading in turn to even more contaminated matrilineal females. All in all, this leads to the following prediction:

푀 푃 69 Prediction 1: 휆푓 > 휆푓 , the HIV propagation rate is higher for matrilineal females, relative to their patrilineal counterparts.

Males propagation rates (휆푚). From Equation 1.6, it is straightforward to see that 휆푚 is increasing in all males’ sexual behavioural parameters. However, since ancestral matrilineality does not yield signiﬁcant sexual behavioural differences among males populations according to my conceptual framework, as well as my empirical results, the value of these parameters should be the same for both matrilineal and patrilineal males. Nevertheless, 휆푚 is also increasing in 푖푓 , which should be higher for matrilineal populations (i.e. infection rate is higher for matrilineal females than their patrilineal counterparts). Consequently, relative to patrilineal male populations, I expect the HIV propagation rate of matrilineal male populations to be only weakly higher. This leads to the following prediction:

푀 푃 Prediction 2: 휆푚 ≥ 휆푚, the HIV propagation rate is weakly higher for matrilineal males, relative to their patrilineal counterparts.

Gender differences in propagation rates (휆푓 − 휆푚). From Equation 1.8, it is straightforward to see that 휆푓 − 휆푚 is increasing in females extramarital promiscuous sexual behaviour, and decreasing in males extramarital promiscuous sexual behaviour. As discussed earlier, and in line with my conceptual framework as well as my empirical ﬁndings, matrilineal females adopt more promiscuous extramarital sexual behaviours, while matrilineal males adopt as promiscuous extramarital sexual behaviours as their patrilineal counterparts. This leads to the following prediction:

69Superscripts 푀 and 푃 denote Matrilineality and Patrilineality, respectively.

67 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

푀 푀 푃 푃 Prediction 3: 휆푓 − 휆푚 > 휆푓 − 휆푚, the gender difference in HIV propagation rates is higher for matrilineal populations.

1.6.3 Simulation

Building on my SI compartmental model of infection discussed earlier, I perform a numerical simulation of the gender-speciﬁc HIV rates dynamics in matrilineal versus patrilineal contexts. These

HIV rates dynamics can be represented by the following system of equations:

⎧ ⎪ 푠 = 푠 − (1휇 + 훼 휏 )훿 푖 푠 ⎪ 푓;푡+1 푓;푡 푓 푓 푓 푓 푚;푡 푓;푡 ⎪ ⎪ ⎨⎪ 푖푓;푡+1 = 푖푓;푡 + (1휇푓 + 훼푓 휏푓 )훿푓 푖푚;푡푠푓;푡

⎪ 푠 = 푠 − (1휇 + 훼 휏 )훿 푖 푠 ⎪ 푚;푡+1 푚;푡 푚 푚 푚 푚 푓;푡 푚;푡 ⎪ ⎪ ⎩ 푖푚;푡+1 = 푖푚;푡 + (1휇푚 + 훼푚휏푚)훿푚푖푓;푡푠푚;푡

with 푠푓;푡 and 푠푚;푡 denoting the proportions of “susceptible” (i.e. not infected) individuals among female and male populations respectively; and 푖푓;푡 and 푖푚;푡 denoting the proportions of “infected” individuals among female and male populations respectively.

According to my conceptual framework as well as my empirical results, the highest rates of HIV among matrilineal females are mainly driven by matrilineal females adopting more promiscuous extramarital sexual behaviours. Further, I found that matrilineal females are mainly contaminated through extramarital routes of infection. To illustrate how this leads to higher gender differential in HIV rates among matrilineal populations, I assume a hypothetical closed population (i.e. no births, no deaths, no migration) of 100 females and 100 males who are in a long-term committed sexual relationship

(e.g. formal marriage), with 1 infected female (1%) and 1 infected male (1%) at the initial period, and

I simulate HIV rates dynamics on a monthly basis, over a 10 years period.

Parameters values. 1 is an indicator of a long-term committed sexual relationship (e.g. formal marriage). I assign to the parameters 훼푓 and 훼푚 the values I have estimated in column 4 of Table 1.7.

푀 More precisely, I assume no difference between matrilineal and patrilineal males and assign 훼푚 = 푃 훼푚 = 훼푚 = 0.204 for both matrilineal and patrilineal males, based on the mean of patrilineal males in 푀 푃 my regression sample; and I assign 훼푓 = 0.076 for matrilineal females; and 훼푓 = 0.008 for patrilineal females.70

70As for the other females parameter values based on my estimations, I assign values as follows: mean of patrilineal males + “Female” estimate for patrilineal females; mean of patrilineal males + “Female” estimate +

68 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

The probability that a male is infected per each sexual intercourse with a HIV positive female is 훿푚 = 훾푚 × 휌푚, namely the product of the probability that he does not use condom during sexual intercourse with the transmission risk for one-time male unprotected sex (assuming for simplicity that condom is totally efﬁcient against HIV transmission). Therefore, the value of 훾푚 is simply 1 minus the probability that a male uses condom. Consequently, based on the mean probability that a patrilineal male reports using condom as a contraceptive method in my regression sample of column

1 of Table 1.8, 훾푚 = 1 − 0.138 = 0.862. Again, I assign the same value for both matrilineal and

71 patrilineal males. 휌푚 is taken from Greenwood et al.(2019) and is set equal to 0.0045. As a result,

훿푚 = 0.862 × 0.0045 = 0.003879.

The probability that a female is infected per each sexual intercourse with a HIV positive male is

훿푓 = 훾푓 × (휌푚 × 휔푓 ), with 휔푓 capturing gender difference in biological susceptibility of contracting

푀 the virus when exposed to it. Following my estimates in column 1 of Table 1.8, 훾푓 = 1 − (0.138 − 푃 0.106 − 0.031) = 0.999, and 훾푓 = 1 − (0.138 − 0.106) = 0.968, for matrilineal and patrilineal females respectively. Further, building on several medical studies, Greenwood et al.(2019) assume that women are 75% more likely to get infected than males for physiological and anatomical reasons.

푀 I follow them and assign 휔푓 = 1.75. Therefore, 훿푓 = 0.999 × 0.0045 × 1.75 ≈ 0.007867 for 푃 matrilineal females; and 훿푓 = 0.968 × 0.0045 × 1.75 = 0.007623 for patrilineal females. Unfortunately, the DHS does not directly provide data on the frequency of sexual intercourses within couples. Therefore, I perform the simulation exercise for different values of 휇푚, 휏푚, 휇푓 , and

휏푓 . For consistency purpose, I assume that 휇푚 = 휇푓 in both matrilineal and patrilineal context. In other words, wife and husband have the same number of intra-couple sexual intercourses per time period. Further, for simplicity I assume that in matrilineal context 휏푚 = 휏푓 (i.e. males and females have the same number of sexual intercourses per time period with each of their extramarital partner72). Fi- nally, given that according to my empirical ﬁndings patrilineal females report a lower sexual activity

푃 푀 than their matrilineal counterparts, to determine 휏푓 I lower the value of 휏푓 in the same proportion than patrilineal females reported sexual activity. According to my estimate in column 2 of Table 1.7, patrilineal females are 4.1 percentage points less likely to report any sexual activity in the last month

푃 푀 than matrilineal females. Accordingly, I set 휏푓 = 0.959 × 휏푓 . As detailed below, I perform the

“Female x Matrilineal” estimate for matrilineal females. Here, it gives 0.204-0.196=0.008 for patrilineal females; and 0.204-0.196+0.068=0.076 for matrilineal females, based on my estimates in column 4 in Table 1.7. 71This number falls in the range of estimates reported by a variety of studies, according to Greenwood et al. (2019). 72This does not mean that females and males have the same number of extramarital partners. These latter are captured by 훼푓 and 훼푚 respectively.

69 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

exercise for different values of 휇푚 and 휏푚. The values and deﬁnitions of parameters are summarized in Table 1.C.13 in appendix.

Simulation results. The main results of this simulation exercise73, computed under a scenario where 휇푚 = 휏푚 = 4, are presented numerically in Table 1.9, and graphically in Figure 1.4. Several lessons can be drawn. First, at any time period matrilineal female HIV rate is higher than patrilineal female HIV rate, and this gap is further increasing through time (comparing column 1 and 4). This is consistent with the ﬁrst prediction of my model. Second, the gap in HIV rates between matrilineal and patrilineal males is of weaker magnitude. This is consistent with the second prediction of my model. Finally, in line with these two previous results and as indicated in the last column, at any time period the gender gap in HIV rates is relatively higher in the matrilineal context, and is further increasing through time. This result is consistent with the third prediction of my model.

Table 1.9 – Simulated HIV Rates

Scenario 1 (휇푚 = 휏푚 = 4) Matrilineality Patrilineality Raw Months Females Males Difference Females Males Difference Diff-in-Diff

1 0.010 0.010 0.000 0.010 0.010 0.000 0.000 25 0.021 0.017 0.004 0.020 0.017 0.003 0.001 50 0.039 0.030 0.009 0.036 0.029 0.007 0.002 75 0.071 0.054 0.017 0.064 0.050 0.014 0.003 100 0.126 0.095 0.031 0.111 0.087 0.027 0.004 120 0.193 0.146 0.047 0.167 0.132 0.035 0.012 Notes: This table reports the simulated HIV rates, based on the compartmental SI epidemic model detailed in section 1.6, and assuming that 휇푚 = 휏푚 = 4.

As a robustness check, in Table 1.C.14 as well as Figure 1.B.6 and Figure 1.B.7 in appendix,

I perform the same simulation exercise under alternative scenarios, namely 휇푚 = 4 and 휏푚 = 2 to allow for the possibility that the frequency of sexual intercourse per extramarital partner is lower than the frequency with long-term committed partner; and 휇푚 = 2 and 휏푚 = 2 to emphasize an overall lower frequency of sexual intercourses within both long-term committed and extramarital couples. The conclusions drawn from these alternative exercises are qualitatively similar to my main simulation exercise.

73I only report proportions of infected individuals. According to my SI model, the proportion of susceptible individuals is simply the proportion of non-infected individuals (= 1 - proportion of infected individuals).

70 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

Figure 1.4 – Simulation - Scenario 1 (휇푚 = 휏푚 = 4)

Matrilineality Patrilineality .2 .2 .15 .15 .1 .1 % % of the Population % of the Population .05 .05 0 0 0 50 100 150 0 50 100 150 Month Month

% Infected Females % Infected Males % Infected Females % Infected Males

To conclude, this simulation exercise illustrates the different HIV propagation dynamics at play and conﬁrms the main ﬁnding of my paper: ancestral matrilineality is, through its impact on females’ sexual and contraceptive behaviour, a driver of contemporary female HIV in Sub-Saharan Africa.

1.7 Conclusions

In this paper I build on the latest evolutionary psychological theories as well as on the anthropological literature to test the hypothesis that females originating from ancestrally matrilineal ethnic groups are more likely to be infected by HIV today than their patrilineal counterparts. Indeed, ancestral matrilineal kinship organizations constituted environments in which substituting long-term committed sexual relationships with sexual variety may have been more beneficial for females’ reproductive success. This strategy would have allowed them to substitute matrilineal males’ lower propensity to commit in the long-term with an increased access to better genes for their offsprings. Furthermore, benefiting from better marriage outside option as well as being inherently valued more, matrilineal females are expected to benefit from a greater sexual autonomy and ability to implement their sexual preferences. Consequently, I test the hypothesis that ancestral matrilineality has shaped more promiscuous contemporary females’ sexual and contraceptive behaviour, and therefore regretfully drived higher HIV prevalence among these populations.

Using data from 18 countries and exploiting within-country variation in ethnic groups’ ancestral kinship organizations, I find that women originating from ancestrally matrilineal ethnic groups are significantly more likely to be HIV positive today. This result is robust to a large set of cultural, historical, geographical and environmental factors, as well as the inclusion of (within-country) region- survey (year) fixed effects. In addition, I show that this long-lasting “matrilineal effect” on female

71 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

HIV is not driven by any differential selection into DHS HIV testing, nor by differences in overall women’s health status, but is speciﬁc to sexually transmitted diseases.

I go one step further by formally testing for omitted bias, computing Altonji et al.(2005) ratios and estimating Oster(2017) bias-adjusted lower bound coefﬁcients. These latter provide very little support for the presence of an omitted bias in my OLS estimates.

Going beyond in identifying causal relationships and correcting for potential reverse causality, I exploit GPS location of DHS villages as well as digitized Murdock’s map of ancestral ethnic groups in

Africa to compute the distance between DHS villages and nearest ancestral matrilineal ethnic boundary. I use this measure to conduct a geographic regression discontinuity design estimation strategy at ethnic boundaries, as well as an instrumental variable strategy. I show that my main estimates are qualitatively robust across all these speciﬁcations. Finally, I estimate the average treatment over treated effect of being located on an ancestrally matrilineal geographic area on DHS villages’ proportion of

HIV positive females, matching villages with their nearest neighbor in an ancestrally non-matrilineal geographic area based on a large array of geographic observables. I ﬁnd that villages located in ancestrally matrilineal areas have signiﬁcantly higher females HIV rates, supporting my main OLS

ﬁnding.

Consistent with my conceptual framework, I find that the higher HIV rates found for matrilineal females can be explained by these latter benefiting from a higher sexual autonomy and ability to implement their sexual preferences, and therefore adopting more promiscuous sexual behaviours. Along these lines, I find that matrilineal females are mainly infected through extramarital routes of infection.

Further, I highlight differences in contraceptive behaviour as a second main mechanism. More precisely, I ﬁnd evidence that matrilineal females higher contraception-related decision-making power translates into them substituting condom with long-term contraceptive methods. By doing so, matrilineal women incidentally substitute protective contraceptive methods with less protective ones.

Nevertheless, I ﬁnd evidence that when they have internalized the risk of HIV transmission, matrilineal individuals are then more likely to adopt condom as a contraceptive method. Finally, I discard differences in access to condom as well as in sexual debuts as additional mechanisms.

In the end, I build on the epidemiological literature to perform a numerical simulation exercise, aiming at illustrating how the differences between matrilineal and patrilineal populations in gender- speciﬁc sexual and contraceptive behaviours translate into different gender-speciﬁc HIV rates dynam-

72 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa ics. Under credible parameter values, these simulations show dynamics that are consistent with my conceptual framework and my empirical ﬁndings.

While recent research literature and policy recommendations have put the emphasis on the need to empower women to reduce the spread of HIV among female populations in Sub-Saharan Africa, this paper highlights mechanisms which call for complementary policies. As a matter of fact, my results have shown that matrilineal women, despite their higher sexual autonomy and ability to impose safe sexual practices to their husbands, are more likely to adopt risky sexual and contraceptive behaviours.

Nevertheless, I have also highlighted that, once beliefs about actual risks of HIV transmission are correctly internalized, matrilineal women take advantage of their higher decision-making power to increase their condom use. This promising result calls for complementary policies targeting empowered women, matrilineal women speciﬁcally, and raising their awareness about the actual riskiness of adopting promiscuous sexual behaviours. It is hoped that these policies will induce behavioural changes that will restrain the spread of HIV in Sub-Saharan Africa.

73 Chapter 1. Ancestral Matrilineality and Female HIV in Sub-Saharan Africa

74 References

Abadie, Alberto and Guido W Imbens, “Large Sample Properties of Matching Estimators for Average Treatment Effects,” Econometrica, 2006, 74 (1), 235–267.

and , “Bias-Corrected Matching Estimators for Average Treatment Effects,” Journal of Busi- ness & Economic Statistics, 2011, 29 (1), 1–11.

Aberle, David F, “Matrilineal Descent in Cross-Cultural Perspective,” Matrilineal Kinship, 1961, pp. 655–727.

Alesina, Alberto, Benedetta Brioschi, and Eliana La Ferrara, “Violence Against Women: a Cross-Cultural Analysis for Africa,” NBER Working Paper, 2016.

, Paula Giuliano, and Nathan Nunn, “On the Origins of Gender Roles: Women and the Plough,” The Quarterly Journal of Economics, 2013, 128 (2), 469–530.

Altonji, Joseph G, Todd E Elder, and Christopher R Taber, “Selection on Observed and Unobserved Variables: Assessing the Effectiveness of Catholic Schools,” Journal of Political Economy, 2005, 113 (1), 151–184.

Anderson, Siwan, “Legal Origins and Female HIV,” The American Economic Review, 2018, 108 (6), 1407–39.

Andrews, Isaiah, “Valid Two-Step Identiﬁcation-Robust Conﬁdence Sets for GMM,” Review of Economics and Statistics, 2018, 100 (2), 337–348.

Ashraf, Nava, Erica Field, Alessandra Voena, and Roberta Ziparo, “Maternal Mor- tality Risk and Spousal Differences in Desired Fertilitye,” Working Paper, 2020.

, Natalie Bau, Nathan Nunn, and Alessandra Voena, “Bride Price and Female Educa- tion,” Journal of Political Economy, 2020.

Ashraf, Quamrul and Oded Galor, “Genetic Diversity and the Origins of Cultural Fragmen- tation,” The American Economic Review: Papers and Proceedings, 2013, 103 (3), 528–33.

and , “The "Out of Africa" Hypothesis, Human Genetic Diversity, and Comparative Economic development,” The American Economic Review, 2013, 103 (1), 1–46.

Baland, Jean-Marie and Roberta Ziparo, “Intra-Household Bargaining in Poor Countries,” Towards Gender Equity in Development, 2018.

Bargain, Olivier, Jordan Loper, and Roberta Ziparo, “Traditional Norm, Access to Divorce and Women’s Empowerment: Evidence from Indonesia,” Working Paper, 2020.

Becker, Anke, “On the Economic Origins of Constraints on Women’s Sexuality,” Working Paper, 2018.

75 References

BenYishay, Ariel, Pauline Grosjean, and Joe Vecci, “The Fish is the Friend of Matriliny: Reef Density and Matrilineal Inheritance,” Journal of Development Economics, 2017, 127, 234– 249.

Berman, Nicolas, Mathieu Couttenier, and Victoire Girard, “Natural Resources and Ethnic Identity,” Working Paper, 2019.

Bertocchi, Graziella and Arcangelo Dimico, “The Long-Term Determinants of Female HIV Infection in Africa: The Slave Trade, Polygyny, and Sexual Behavior,” Journal of Development Economics, 2019, 140, 90–105.

Besley, Timothy and Marta Reynal-Querol, “The Legacy of Historical Conﬂict: Evidence from Africa,” American Political Science Review, 2014, 108 (2), 319–336.

Buss, David M, The Evolution of Desire - Strategies of human mating, Basic Books, 2016.

Cagé, Julia and Valeria Rueda, “Sex and the Mission: The Conﬂicting Effects of Early Chris- tian Investments on the HIV Epidemic in Sub-Saharan Africa,” Journal of Demographic Eco- nomics, Forthcoming.

Case, Anne and Christina Paxson, “HIV Risk and Adolescent Behaviors in Africa,” Ameri- can Economic Review: Papers & Proceedings, 2013, 103 (3), 433–38.

Cassidy, Rachel, M Groot Bruinderink, Wendy Janssens, and Karlijn Morsink, “The Power to Protect: Household Bargaining and Female Condom Use,” Working Paper, 2018.

Chimbiri, Agnes M, “The Condom is an "Intruder" in Marriage: Evidence from Rural Malawi,” Social Science & Medicine, 2007, 64 (5), 1102–1115.

Cohen, Craig R, Michele Montandon, Adam W Carrico, Stephen Shiboski, Alan Bostrom, Alfredo Obure, Zachary Kwena, Robert C Bailey, Rosemary Nguti, and Elizabeth A Bukusi, “Association of Attitudes and Beliefs Towards Antiretroviral Ther- apy with HIV-Seroprevalence in the General Population of Kisumu, Kenya,” PloS one, 2009, 4 (3), e4573.

Cordero-Coma, Julia and Richard Breen, “HIV Prevention and Social Desirability: Husband-wife Discrepancies in Reports of Condom Use,” Journal of Marriage and Family, 2012, 74 (3), 601–613.

Corno, Lucia and Damien De Walque, “Mines, Migration and HIV/AIDS in Southern Africa,” Journal of African Economies, 2012, 21 (3), 465–498.

Delavande, Adeline and Hans-Peter Kohler, “HIV/AIDS-Related Expectations and Risky Sexual Behaviour in Malawi,” The Review of Economic Studies, 2015, 83 (1), 118–164.

Diamond, Jared M, Why is sex fun?: the evolution of human sexuality, Basic Books, 1998.

Fernández, Raquel, “Does Culture Matter?,” in “Handbook of social economics,” Vol. 1, Elsevier, 2011, pp. 481–510.

Fortunato, Laura, “The Evolution of Matrilineal Kinship Organization,” Proceedings of the Royal Society London, 2012.

Fox, Robin, Kinship and Marriage: An Anthropological Perspective, Cambridge: Cambridge University Press, 1934.

76 References

Friedman, Willa Helterline, “Antiretroviral Drug Access and Behavior Change,” Journal of Development Economics, 2018, 135, 392–411.

Giuliano, Paola, “Gender: An Historical Perspective,” The Oxford Handbook of Women and the Economy, 2017.

and Nathan Nunn, “Ancestral Characteristics of Modern Populations,” Economic History of Developing Regions, 2018, 33 (1), 1–17.

and , “Understanding Cultural Persistence and Change,” Working Paper, 2019.

González, Felipe and Edward Miguel, “War and Local Collective Action in Sierra Leone: A Comment on the Use of Coefﬁcient Stability Approaches,” Journal of Public Economics, 2015, 128, 30–33.

Greenwood, Jeremy, Philipp Kircher, Cezar Santos, and Michèle Tertilt, “An Equi- librium Model of the African HIV/AIDS Epidemic,” Econometrica, 2019, 87 (4), 1081–1113.

Greiling, Heidi and David M Buss, “Women’s Sexual Strategies: The Hidden Dimension of Extra-Pair Mating,” Personality and Individual Differences, 2000, 28 (5), 929–963.

Holden, Clare Janaki and Ruth Mace, “Spread of cattle led to the loss of matrilineal descent in Africa: a coevolutionary analysis,” Proceedings of the Royal Society of London B: Biological Sciences, 2003, 270 (1532), 2425–2433.

, Rebecca Sear, and Ruth Mace, “Matriliny as Daughter-Biased Investment,” Evolution and Human Behavior, 2003, 24 (2), 99–112.

Islam, Mohammad Amirul, M Rakibul Islam, and Banya Banowary, “Sex Preference as a Determinant of Contraceptive Use in Matrilineal Societies: a Study on the Garo of Bangladesh,” The European Journal of Contraception & Reproductive Health Care, 2009, 14 (4), 301–306.

Kibira, Simon PS, Patricia Ndugga, Elizabeth Nansubuga, Andrew Sewannonda, and Betty Kwagala, “Contraceptive Uptake Among Married Women in Uganda: Does Em- powerment Matter?,” African Population Studies, 2014, 28, 968–975.

Lowes, Sara, “Kinship Systems, Gender Norms, and Household Bargaining: Evidence from the Matrilineal Belt,” Working Paper, 2018.

, “Matrilineal Kinship and Gender Differences in Competition,” Working Paper, 2018.

and Eduardo Montero, “The Legacy of Colonial Medicine in Central Africa,” Working Paper, 2018.

Michalopoulos, Stelios, Louis Putterman, and David N Weil, “The Inﬂuence of Ances- tral Lifeways on Individual Economic Outcomes in Sub-Saharan Africa,” Journal of the European Economic Association, 2019.

Mishra, Vinod, Martin Vaessen, J Boerma, Fred Arnold, Ann Way, Bernard Barrere, Anne Cross, Rathavuth Hong, and Jasbir Sangha, “HIV Testing in National Population-Based Surveys: Experience from the Demographic and Health Surveys,” Bulletin of the World Health Organization, 2006, 84, 537–545.

Montalvo, José G and Marta Reynal-Querol, “Ethnic Polarization, Potential Conﬂict, and Civil Wars,” The American Economic Review, 2005, 95 (3), 796–816.

77 References

Moscona, Jacob, Nathan Nunn, and James A Robinson, “Segmentary Lineage Organi- zation and Conﬂict in Sub-Saharan Africa,” Econometrica, 2020.

Nunn, Nathan, “Religious Conversion in Colonial Africa,” The American Economic Review: Papers & Proceedings, 2010, 100 (2), 147–52.

and Leonard Wantchekon, “The Slave Trade and the Origins of Mistrust in Africa,” The American Economic Review, 2011, 101 (7), 3221–52.

Nyqvist, Martina Björkman, Lucia Corno, Damien De Walque, and Jakob Svens- son, “Incentivizing Safer Sexual Behavior: Evidence from a Lottery Experiment on HIV Preven- tion,” American Economic Journal: Applied Economics, 2018, 10 (3), 287–314.

Oster, Emily, “Sexually Transmitted Infections, Sexual Behavior, and the HIV/AIDS Epidemic,” The Quarterly Journal of Economics, 2005, 120 (2), 467–515.

, “HIV and Sexual Behavior Change: Why not Africa?,” Journal of Health Economics, 2012, 31 (1), 35–49.

, “Routes of Infection: Exports and HIV Incidence in Sub-Saharan Africa,” Journal of the Euro- pean Economic Association, 2012, 10 (5), 1025–1058.

, “Unobservable Selection and Coefﬁcient Stability: Theory and Evidence,” Journal of Business & Economic Statistics, 2017, pp. 1–18.

Palamuleni, Martin Enock and Ayo Stephen Adebowale, “Women Empowerment and the Current Use of Long Acting and Permanent Contraceptive: Evidence from 2010 Malawi De- mographic and Health Survey,” Malawi Medical Journal, 2014, 26 (3), 63–70.

Paula, Áureo De, Gil Shapira, and Petra E Todd, “How Beliefs About HIV Status Affect Risky Behaviors: Evidence from Malawi,” Journal of Applied Econometrics, 2014, 29 (6), 944–964.

Porta, Rafael La, Florencio Lopez de Silanes, and Andrei Shleifer, “The Economic Consequences of Legal Origins,” Journal of Economic Literature, 2008, 46 (2).

Samari, Goleen, “Women’s Empowerment and Short-and Long-Acting Contraceptive Method Use in Egypt,” Culture, health & sexuality, 2018, 20 (4), 458–473.

Sun, Liyang, “Implementing Valid Two-Step Identiﬁcation-Robust Conﬁdence Sets for Linear Instrumental-Variables Models,” The Stata Journal, 2018, 18 (4), 803–825.

Tassier, Troy, The Economics of Epidemiology, Springer, 2013.

Tequame, Miron, “HIV, Risky Behavior and Ethno-Linguistic Heterogeneity,” Journal of Eco- nomics and Statistics, 2012, 232 (6), 606–632.

Teso, Edoardo, “The Long-Term Effect of Demographic Shocks on the Evolution of Gender Roles: Evidence from the Transatlantic Slave Trade,” Journal of the European Economic Association, 2019.

Thornton, Rebecca L, “The Demand for, and Impact of, Learning HIV Status,” The American Economic Review, 2008, 98 (5), 1829–63.

UNAIDS, “Miles to Go - Closing Gaps. Breaking Barriers. Righting Injustices.,” Joint United Nations Programme on HIV/AIDS, 2018.

78 References

, “Women and Girls and HIV,” Joint United Nations Programme on HIV/AIDS, 2018.

Walque, Damien De, Harounan Kazianga, and Mead Over, “Antiretroviral Therapy Perceived Efﬁcacy and Risky Sexual Behaviors: Evidence from Mozambique,” Economic Devel- opment and Cultural Change, 2012, 61 (1), 97–126.

Worden, Lee and Travis C Porco, “Products of Compartmental Models in Epidemiology,” Computational and Mathematical Methods in Medicine, 2017, 2017.

Yi, Tae Joon, Brett Shannon, Jessica Prodger, Lyle McKinnon, and Rupert Kaul, “Genital Immunology and HIV Susceptibility in Young Women,” American Journal of Reproductive Immunology, 2013, 69, 74–79.

79 References

80 Appendix

1.A Data Description

1.A.1 List of the countries included in the analysis

All Demographic Health Surveys (DHS) from Sub-Saharan African countries whith data on both HIV, individual’s ethnicity as well as village’s GPS information are included in the analysis. Based on this selection criterion, the country surveys used include 32 surveys from 18 countries as follow:

• Burkina-Faso (2003, 2010)

• Cameroon (2004, 2011)

• Chad (2014)

• Congo Democratic Republic (2007, 2013)

• Ethiopia (2005, 2011, 2016)

• Gabon (2012)

• Ghana (2003, 2014)

• Guinea (2005, 2012)

• Ivory Coast (2011)

• Kenya (2003, 2008)

• Liberia (2013)

• Malawi (2004, 2010, 2014)

• Mali (2006, 2012)

• Senegal (2005, 2010)

• Sierra Leone (2008, 2013)

• Togo (2013)

• Uganda (2011)

• Zambia (2007, 2013)

81 References

1.A.2 Description of the main controls 1.A.2.1 Individual-level Data

• Marital status: Categorical variable of the current marital status of the respondent (0 “Never married”; 1 “Married”; 2 “Living together”; 3 “Widowed”; 4 “Divorced”; and 5 “Not living together”). Source: DHS (variable v501)

• Actual polygyny: Dummy variable indicating whether an individual is in a polygynous union. Source: DHS (computed from variable v505)

• Age: Individual’s Age in completed years. Source: DHS (variable v012)

• Age squared: Square of the individual’s age in completed years. Source: DHS (computed from variable v012)

• Number of children: Individual’s total number of children ever born. Source: DHS (variable v201)

• Urban: Dummy variable indicating whether an individual resides in an urban location. Source: DHS (variable v025)

• Education: Individual’s education in single years. Source: DHS (variable v133)

• Working: Dummy variable indicating whether an individual is currently working. Source: DHS (variable v714)

• Wealth: Categorical variable of wealth index (a composite measure of a household’s cumu- lative living standard) (1 “Poorest”; 2 “Poorer”; 3 “Middle”; 4 “Richer”; and 5 “Richest”). Source: DHS (variable v190)

• Religion: Categorical variable of religion (1 “Christian”; 2 “Muslim”; 3 “Other religion”; and 4 “No religion”). Source: DHS (computed and harmonized between countries by the author, from variable v130)

1.A.2.2 Ethnicity-level Data

These variables are computed from the Ethnographic Atlas (E.A.), a worldwide anthropological database containing ethnographic information on cultural aspects and ways of life of ethnic groups prior to industrilization and colonial contact.

• Women’s historical participation in agriculture: Variable increasing in women’s historical participation in agriculture, and ranging from 1 “Males only or almost alone” to 6 “Females only or almost alone”. Source: Ethnographic Atlas (variable v54)

• Polygyny: Dummy variable indicating whether an ethnic group practiced polygyny (v8 = 2; 4 or 5 in the E.A.). Source: Ethnographic Atlas (variable v8)

• Bride Price: Dummy variable indicating whether an ethnic group practiced bride price (v6 not equal to 6 or 7 in the E.A.). Source: Ethnographic Atlas (variable v6)

• Plough: Categorical variable of plough use (1 “No plough”; 2 “Not aboriginal but well established plough use at period of observation”; and 3 “Aboriginal plough use prior to contact”). Source: Ethnographic Atlas (variable v39)

82 References

• Pastoralism: Index of ethnic group’s dependence on pastoralism, computed following Becker (2018) by interacting the index of ethnic group dependence on animal husbandry (v4 in the E.A.) with a dummy indicating whether herd animals were the predominant ethnic group’s type of animal husbandry (i.e v40 = 3, 4, 5, 6 or 7 in the E.A.). Source: Ethnographic Atlas (computed from variables v4 and v40)

• Clans: Dummy variable indicating whether an ethnic group was organized in clan communities (v15 = 6 in the E.A.). Source: Ethnographic Atlas (computed from variable v15)

• Settlement patterns: Variable increasing in the complexity of ethnic group’s settlement patterns, ranging from 1 “Nomadic or fully migratory” to 8 “Complex settlements”. Source: Ethnographic Atlas (variable v30)

• Jurisdictional Hierarchies: Categorical variable of the number of levels in the ethnic group’s jurisdictional hierarchy beyond local community (= v33 - 1 in the E.A.). Source: Ethnographic Atlas (variable v33)

• Reliance on hunting: Categorical variable of ethnic group’s reliance on hunting, ranging from 0 “0 - 5% dependence” to 9 “86 - 100% dependence”. Source: Ethnographic Atlas (variable v2)

• Reliance on ﬁshing: Categorical variable of ethnic group’s reliance on ﬁshing, ranging from 0 “0 - 5% dependence” to 9 “86 - 100% dependence”. Source: Ethnographic Atlas (variable v3)

• Reliance on gathering: Categorical variable of ethnic group’s reliance on gathering, ranging from 0 “0 - 5% dependence” to 9 “86 - 100% dependence”. Source: Ethnographic Atlas (variable v1)

• Reliance on animal husbandry: Categorical variable of ethnic group’s reliance on animal husbandry, ranging from 0 “0 - 5% dependence” to 9 “86 - 100% dependence”. Source: Ethnographic Atlas (variable v4)

• Reliance on agriculture: Categorical variable of ethnic group’s reliance on agriculture, ranging from 0 “0 - 5% dependence” to 9 “86 - 100% dependence”. Source: Ethnographic Atlas (variable v5)

• Large domesticated animals: Dummy variable indicating whether an ethnic group practiced animal husbandry (v40 > 1 in the E.A.). Source: Ethnographic Atlas (computed from variable v40)

• Intensity of agriculture: Categorical variable increasing in ethnic group’s intensity of agriculture, and ranging from 1 “No agriculture” to 6 “Intensive irrigated agriculture”. Source: Ethnographic Atlas (variable v28)

• Year of observation in the E.A.: Year of observation of the ethnic group in the Ethnographic Atlas. Source: Ethnographic Atlas (variavle v102)

1.A.2.3 Village-level Data Except when otherwise indicated, most of village-level data are computed from DHS geospatial covariates. Documentation on DHS geospatial covariates can be found here: https://spatialdata. dhsprogram.com/covariates/.

• Latitude: Latitude of DHS village (in decimal degrees). Source: DHS Geospatial Covariates

83 References

• Longitude: Longitude of DHS village (in decimal degrees). Source: DHS Geospatial Covari- ates • Altitude: Altitude of DHS village (in meters). Source: DHS Geospatial Covariates • Nightlight composite: The average nighttime luminosity, of the area within the 2 km (urban) or 10 km (rural) buffer surrounding the DHS village location (2015). Source: DHS Geospatial Covariates • Population density (2010): The average UN-adjusted population density of the area within the 2 km (urban) or 10 km (rural) buffer surrounding the DHS village location (2010). Source: DHS Geospatial Covariates • Distance to lake or coastline: The geodesic distance to either a lake or the coastline. Source: DHS Geospatial Covariates • Distance to international border: The geodesic distance to the nearest international borders. Source: DHS Geospatial Covariates • Average time urban center: (2015): The average time (minutes) required to reach a high- density urban center, from the area within the 2 km (urban) or 10 km (rural) buffer surrounding the DHS village location, based on year 2015 infrastructure data. Source: DHS Geospatial Covariates • Malaria incidence (2010): The average number of people per year who show clinical symptoms of plasmodium falciparum malaria within the 2 km (urban) or 10 km (rural) buffer surrounding the DHS village location (2010). Source: DHS Geospatial Covariates • Vegetation index (2010): The average vegetation index value within the 2 km (urban) or 10 km (rural) buffer surrounding the DHS village (2010). Source: DHS Geospatial Covariates • Length of the growing season: Categorical variable of the length of the growing season in days for the area within the 2 km (urban) or 10 km (rural) buffer surrounding the DHS village location; ranging from 1 “0 days” to 16 “ > 365 days”. Source: DHS Geospatial Covariates • Distance to nearest active mine: Distance (in km) to the nearest active mine (active the same year than the village is surveyed in DHS). Source: Calculation from the author, exploiting data on the location of large-scale active mines in Africa, provided by S&P Global Market Intelli- gence (https://www.spglobal.com/marketintelligence/en/campaigns/metals-mining). • Active mine within 1000 km: Dummy indicating whether the distance to the nearest active mine is lower or equal to 1000 km. Source: Calcultation from the author, exploiting data on the location of large-scale active mines in Africa, provided by S&P Global Market Intelligence (https://www.spglobal.com/marketintelligence/en/campaigns/metals-mining). • Ethnic fractionalization: Index of ethnic fractionalization computed at the DHS village level using information on individual’s (who reside in the village) ethnicity, following the formula of Montalvo and Reynal-Querol(2005) as follows:

푁 푁 ∑︁ 2 ∑︁ 퐸푡ℎ푛푖푐_퐹 푟푎푐푡푖표푛푎푙푖푧푎푡푖표푛 = 1 − 휋푖 = 휋푖(1 − 휋푖) (1.9) 푖=1 푖=1

where 휋푖 is the proportion of people who belong to the ethnic group 푖, and 푁 is the number of thnic groups. The index of ethnic fractionalization, which is a Herﬁndhahl index, has a simple interpretation as the probability that two randomly selected individuals from a given geographic area do not belong to the same ethnic group. Source: Calculation from the author

84 References

• Ethnic polarization: Index of ethnic polarization computed at the DHS village level using information on individual’s (who reside in the village) ethnicity, following the formula of Mon- talvo and Reynal-Querol(2005) as follows:

푁 (︂ )︂2 푁 ∑︁ 1/2 − 휋푖 ∑︁ 퐸푡ℎ푛푖푐_푃 표푙푎푟푖푧푎푡푖표푛 = 1 − 휋2 = 4 휋2(1 − 휋 ) (1.10) 1/2 푖 푖 푖=1 푖=1 The original purpose of this Ethnic Polarization (EP) index is to capture how far the distribution of the ethnic groups is from the (1/2, 0, 0, ... 0, 1/2) distribution (bipolar), which represents the highest level of polarization. Ranging between 0 and 1, a higher value of the EP index indicates a higher ethnical polarization, with EP equal to 0 indicating an ethnical homogeneity, and an EP equal to 1 for two ethnic groups of the same size. Source: Calculation from the author

85 References

1.B Additional Figures

86 References

Figure 1.B.1 – Ancestral Ethnic Group Boundaries and Contemporary Gender Differences in HIV Rates

87 References

Figure 1.B.2 – Ancestral Ethnic Group Boundaries and Contemporary Male HIV Rates

88 References

Figure 1.B.3 – Location of DHS Villages

89 References

Figure 1.B.4 – Individual’s Matrilineality and Distance to Nearest Ancestral Matrilineal Ethnic Boundary (RDD)

This graph presents the unconditional relationship between individual’s ethnic group’s ancestral matrilineality and individual’s DHS village geographic location, for which a linear polynomial is estimated separately at each side of the boundary. The sample is limited to individuals living in villages located within 150 km of an ancestral matrilineal ethnic boundary. The x-axis reports geographic distance. Positive values are kilometers into the territory of an ancestrally matrilineal ethnic group and negative values are kilometers into the territory of an ancestrally non-matrilineal (i.e. patrilineal or other) ethnic group. The y-axis measures the fraction of the population at each distance that originates from an ancestrally matrilineal ethnic group.

90 References

Figure 1.B.5 – Female HIV Rate and Distance to Nearest Ancestral Matrilineal Ethnic Boundary (RDD)

This graph presents the relationship between female HIV rate and individual’s DHS village geographic location, for a specification that conditions on region (within-country) × survey (time) FE, and for which a cubic polynomial is estimated separately at each side of the boundary. The sample is limited to females living in villages located within 150 km of an ancestral matrilineal ethnic boundary. The x-axis reports geographic distance. Positive values are kilometers into the territory of an ancestrally matrilineal ethnic group and negative values are kilometers into the territory of an ancestrally non- matrilineal (i.e. patrilineal or other) ethnic group. The y-axis measures the fraction of the female population at each distance that is HIV positive.

91 References

Figure 1.B.6 – Simulation - Scenario 2 (휇푚 = 4 and 휏푚 = 2)

Matrilineality Patrilineality .2 .2 .15 .15 .1 .1 % % of the Population % of the Population .05 .05 0 0 0 50 100 150 0 50 100 150 Month Month

% Infected Females % Infected Males % Infected Females % Infected Males

Figure 1.B.7 – Simulation - Scenario 3 (휇푚 = 휏푚 = 2)

Matrilineality Patrilineality .05 .05 .04 .04 .03 .03 % % of the Population % of the Population .02 .02 .01 .01 0 50 100 150 0 50 100 150 Month Month

% Infected Females % Infected Males % Infected Females % Infected Males

92 References

1.C Additional Tables

93 References

94 References

Table 1.C.1 – Summary Statistics (Final Regression Sample)

Females Males All Sample

Matrilineal Patrilineal Difference Matrilineal Patrilineal Difference Matrilineal Patrilineal Difference

HIV 0.098 0.039 0.059*** 0.066 0.023 0.043*** 0.083 0.033 0.051*** (0.194) (0.298) (0.002) (0.150) (0.249) (0.002) (0.277) (0.177) (0.001)

Individual-level variables Marital status 1.204 1.082 0.122*** 0.810 0.786 0.025** 1.021 0.960 0.061*** (1.228) (1.037) (0.009) (0.934) (0.888) (0.008) (1.119) (0.989) (0.006) Actual polygyny 0.106 0.272 -0.166*** 0.055 0.121 -0.066*** 0.082 0.210 -0.128*** (0.308) (0.445) (0.004) (0.227) (0.326) (0.003) (0.275) (0.407) (0.002) Age 28.236 28.937 -0.701*** 30.205 31.277 -1.073*** 29.151 29.896 -0.745*** (9.410) (9.970) (0.085) (11.498) (12.096) (0.113) (10.478) (10.952) (0.069) Age squared 885.820 936.766 -50.946*** 1044.514 1124.572 -80.058*** 959.575 10103.752 -54.177*** (578.551) (635.385) (5.416) (776.859) (835.740) (7.790) (682.563) (730.113) (4.599) Nb. children 3.062 3.283 -0.221*** 2.960 3.155 -0.195*** 3.014 3.230 -0.216*** (2.790) (3.126) (0.027) (3.560) (4.114) (0.038) (3.172) (3.565) (0.022) Urban 0.274 0.310 -0.035*** 0.262 0.341 -0.078 0.269 0.322 -0.054*** (0.446) (0.462) (0.004) (0.474) (0.440) (0.004) (0.003) (0.001) (0.003) Education 5.338 4.185 1.153*** 6.419 5.516 0.903*** 5.841 4.731 1.110*** (3.956) (4.349) (0.037) (3.907) (4.757) (0.044) (3.970) (4.567) (0.029) Working 0.575 0.636 -0.061*** 0.782 0.788 -0.006 0.671 0.698 -0.027*** (0.494) (0.481) (0.004) (0.412) (0.409) (0.004) (0.470) (0.459) (0.003) Wealth 3.013 2.997 0.016 3.038 2.993 0.045*** 3.025 2.995 0.029** (1.416) (1.430) (0.012) (1.385) (1.424) (0.013) (1.402) (1.428) (0.009) Religion 1.237 1.520 -0.283*** 1.314 1.615 -0.300*** 1.273 1.559 -0.286*** (0.565) (0.651) (0.006) (0.715) (0.689) (0.007) (0.640) (0.668) (0.004)

Ethnicity-level variables Women’s participation in agri. 4.624 3.758 0.866*** 4.647 3.600 1.047*** 4.635 3.693 0.942*** (1.023) (1.236) (0.010) (1.000) (1.224) (0.011) (1.012) (1.233) (0.008) Polygyny 0.770 0.524 0.247*** 0.788 0.484 0.304*** 0.779 0.508 0.271*** (0.421) (0.499) (0.004) (0.409) (0.500) (0.005) (0.415) (0.500) (0.003) Bride Price 0.867 1.000 -0.132*** 0.880 1.000 -0.119*** 0.873 1.000 -0.126*** (0.339) (0.022) (0.001) (0.325) (0.022) (0.001) (0.332) (0.022) (0.001) Plough 1.000 1.019 -0.019*** 1.000 1.000 0.000 1.000 1.011 -0.011*** (0.000) (0.137) (0.001) (0.000) (0.022) (0.000) (0.000) (0.106) (0.001) Pastoralism 0.087 0.208 -0.121*** 0.089 0.195 -0.107*** 0.088 0.203 -0.115*** (0.045) (0.131) (0.122) (0.043) (0.131) (0.001) (0.044) (0.131) (0.001) Clans 0.234 0.262 -0.028*** 0.223 0.259 -0.035*** 0.229 0.261 -0.032*** (0.423) (0.440) (0.004) (0.417) (0.438) (0.004) (0.002) (0.001) (0.003) Settlement patterns 6.679 6.628 0.051*** 6.699 6.828 -0.129*** 6.688 6.710 -0.022*** (0.610) (1.129) (0.009) (0.599) (1.044) (0.009) (0.605) (1.100) (0.007) Jurisdictional hierarchies 1.470 1.574 -0.103*** 1.509 1.525 -0.016* 1.488 1.554 -0.066*** (0.637) (0.973) (0.008) (0.631) (0.935) (0.008) (0.635) (0.958) (0.006) Reliance on hunting 15.451 8.723 6.729*** 15.809 8.877 6.932*** 15.618 8.786 6.831*** (6.505) (5.432) (0.048) (6.585) (5.604) (0.055) (6.545) (5.504) (0.036) Reliance on fishing 14.762 9.229 5.533*** 14.629 9.312 5.317*** 14.700 9.263 5.437*** (7.758) (7.429) (0.065) (7.732) (7.517) (0.071) (7.746) (7.466) (0.048) Reliance on gathering 7.198 5.863 1.336*** 7.237 6.416 0.821*** 7.216 6.089 1.127*** (4.414) (4.779) (0.041) (4.357) (4.979) (0.046) (4.388) (4.870) (0.031) Reliance on animal husbandry 9.666 20.651 -10.985*** 9.720 19.472 -9.752*** 9.691 20.168 -10.477*** (2.655) (12.684) (0.101) (2.524) (12.571) (0.108) (2.595) (12.651) (0.074) Reliance on agriculture 54.339 58.996 -4.657*** 53.969 59.235 -5.266*** 54.167 59.094 -4.927*** (5.659) (9.489) (0.078) (5.552) (10.006) (0.088) (5.613) (9.705) (0.058) Large domesticated animals 0.839 0.960 -0.122*** 0.856 0.946 -0.090*** 0.847 0.954 -0.108*** (0.368) (0.195) (0.002) (0.351) (0.227) (0.002) (0.360) (0.209) (0.002) Intensity of agriculture 3.016 3.449 -0.432*** 3.015 3.348 -0.332*** 3.016 3.407 -0.391*** (0.180) (0.889) (0.007) (0.175) (0.830) (0.007) (0.178) (0.867) (0.005) Year of observation in E.A. 1918.87 1916.973 1.896*** 1919.177 1917.302 1.875*** 1919.013 1917.108 1.905*** (11.451) (18.575) (0.153) (11.297) (15.702) (0.142) (11.380) (17.455) (0.106)

Village-level variables Latitude -10.973 5.025 -15.998*** -11.046 5.667 -16.714*** -11.007 5.288 -16.295*** (6.955) (7.643) (0.065) (6.805) (8.235) (0.075) (6.885) (7.898) (0.049) Longitude 27.569 10.705 16.864*** 27.709 5.996 21.713*** 27.634 8.775 18.860*** (11.371) (19.016) (0.156) (11.196) (17.501) (0.067) (11.290) (18.555) (0.112) Altitude 856.539 656.037 200.502*** 879.954 534.783 345.171*** 867.421 606.332 261.089*** (389.289) (535.854) (4.461) (386.619) (516.611) (4.679) (388.220) (531.406) (3.252) Nightlight composite 2.563 1.601 0.962*** 2.448 1.720 0.728*** 2.510 1.650 0.860*** (6.897) (4.658) (0.044) (6.808) (4.857) (0.050) (6.856) (4.741) (0.033) Pop. density (2010) 1297.833 1056.708 241.125*** 1050.502 958.689 91.813*** 1182.883 1016.528 166.355*** (5170.735) (3149.892) (30.431) (4312.437) (3103.504) (31.634) (4792.504) (3131.321) (21.984) Distance (geodesic) lake/coastline 103,597 117,080 -13,483*** 105,490 131,855 -26,365*** 104,477 123,137 -18,660*** (113,291) (111,648) (966) (109,640) (112,878) (1,060) (111,611) (112,388) (715) Distance (geodesic) international border 46,026 81,903 -35,877*** 46,250 81,936 -35,687*** 46,130 81,817 -35,787*** (41,586) (74,721) (611) (40,845) (79,587) (700) (41,243) (76,752) (460) Average time urban center (2015) 77.479 78.707 -1.228*** 78.590 86.703 -8.113*** 77.995 81.985 -3.989*** (76.205) (96.880) (0.812) (76.041) (108.687) (0.978) (76.130) (101.961) (0.625) Malaria incidence (2010) 0.379 0.368 0.011*** 0.380 0.349 0.031*** 0.379 0.360 0.019*** (0.171) (0.213) (0.002) (0.174) (0.215) (0.002) (0.172) (0.214) (0.001) Vegetation index (2010) 3066.037 3233.712 -167.676*** 3077.646 3143.876 -66.230*** 3071.432 3196.886 -125.454*** (582.642) (975.880) (8.012) (590.357) (987.358) (8.764) (586.259) (981.593) (5.914) Length growing season 8.008 9.24895 -1.240*** 7.967 8.769 -0.802*** 7.989 9.052 -1.062*** (1.511) (2.733) (0.022) (1.481) (2.601) (0.023) (1.497) (2.690) (0.016) Distance mine 3480.909 2399.391 1081.518*** 3494.092 2107.588 1386.504*** 3487.036 2279.775 1207.261 (1018.664) (1094.061) (9.352) (993.290) (1027.745) (9.641) (5.869) (2.755) (6.785) Mine within 1000 km 0.034 0.041 -0.007*** 0.035 0.054 -0.019*** 0.035 0.047 -0.012*** (0.182) (0.199) (0.002) (0.185) (0.227) (0.002) (0.183) (0.211) (0.001) Ethnic fractionalization 0.370 0.270 -0.100*** 0.363 0.270 0.093*** 0.367 0.270 0.097*** (0.280) (0.265) (0.002) (0.281) (0.265) (0.003) (0.280) (0.265) (0.002) Ethnic polarization 0.622 0.486 0.136*** 0.611 0.487 0.125*** 0.617 0.486 0.131*** (0.370) (0.393) (0.003) (0.372) (0.396) (0.004) (0.371) (0.394) (0.002)

Observations 15,758 90,206 13,683 62,665 29,441 152,871 Notes: This table reports summary statistics (means; and standard deviations in parenthesis) computed on my main regression sample (column 5 in Table 1.1). Standard errors of t-test of equality of means are reported in parenthesis in columns reporting mean differences. A description of variables is provided in subsection 1.A.2 of this appendix. References

Table 1.C.2 – Ancestral Matrilineality, Genetic Diversity and Father Diversification (OLS)

Country-Level Individual-Level

Mobility index- Mobility index- Women’s children’s predicted genetic Women’s children’s predicted genetic number of diversity (ancestry father diversification diversity different fathers andjusted) (1) (2) (3) (4)

Matrilineality 0.005** 0.007* 0.007** 0.014** (0.002) (0.004) (0.003) (0.007) Controls Yes Yes Yes Yes Continent FE Yes Yes No No Region-survey FE No No Yes Yes

Observations 124 124 66,895 66,895 R-squared 0.994 0.897 0.023 0.333 Clusters N.A. N.A. 100 100 Mean Dep. Var. 0.711 0.722 1.019 0.537 Notes: In columns 1 and 2, OLS estimates are reported with robust standard errors in brackets. The unit of observation is a country. “Matrilineality” is the estimated proportion of countries’ citizens with ancestors that had matrilineal inheritence rule (source: Giuliano and Nunn, 2018). This variable ranges from 0 to 1. Outcomes are from Ashraf and Galor(2013b) and are computed at the country level. The “Mobility index-predicted genetic diversity” is the expected heterozygosity (genetic diversity) of a given country as predicted by migratory distance (human mobility index), calculated for the journey from Addis Ababa (Ethiopia) to the country’s modern capital city. “Mobility index-predicted genetic diversity (ancestry adjusted)” is the expected heterozygosity (genetic diversity) of a given country as predicted by migratory distance (human mobility index), calculated for the journey from Addis Ababa (Ethiopia) to each of the year 1500 CE locations of the ancestral populations of the country’s component ethnic groups in 2000 CE, as well as for the journey between each pair of these ancestral populations. The Controls are from Ashraf and Galor(2013b) and Ashraf and Galor(2013a), computed at the country level and include: the “aerial” great circle distance “as the crow flies” from Addis Ababa (Ethiopia) to the country’s modern capital city; the square of this previous measure of distance; the “migratory” great circle distance from Addis Ababa (Ethiopia) to the country’s modern capital city along a land-restricted path forced through one or more of five intercontinental waypoints; the square of this previous measure of distance; the absolute value of the latitude of a country’s geodesic centroid; a geospatial index of the suitability of land for agriculture (based on ecological indicators of climate suitability for cultivation); the percentage of a country’s arable land; the country’s mean distance to nearest waterway; the country’s total land area; the average monthly temperature of a country; the average monthly precipitation of a country; the fraction of a country’s land area that is located in tropical and subtropical climate zones; the total number of different types of infectious diseases in a country; a dummy indicating whether a nation is an island; and a dummy indicating whether a nation is landlocked. In column 3 and 4, OLS estimates are reported with standard errors clustered at the ethnicity level in brackets. The unit of observation is a woman surveyed in DHS and being part of my final sample, originating from either a traditionally matrilineal or a traditionally patrilineal ethnic group. “Matrilineality” indicates (dummy) whether a woman belongs to a traditionally matrilineal ethnic group. “Women’s children’s number of different fathers” is the number of different males with whom a female had her biological children (aged less than 17 and member of the household at the time of the survey). “Women’s children’s father diversification” is the the number of different males with whom a female had her biological children divided by the number of a female’s biological children (aged less than 17 and member of the household at the time of the survey). The controls consist of the Individual controls, the (Ancestral) Ethnic Group Controls as well as the Village-Geographic Controls defined in subsection 1.3.3 and Table 1.1. The number of female’s biological children is not controlled for anymore in column 4. Region-survey is a subnational region defined in DHS, interacted with its survey-year. In columns 3 and 4, “R-squared” reported is adjusted; and “Mean Dep. Var.” reported is the mean of patrilineal women in the regression samples. * p<0.10,** p<0.05,*** p<0.01

96 References

Table 1.C.3 – Selection And Falsification Tests

Selection Falsification Tests

Consent HIV Take HIV Rohrer Anemia BMI Test Test Index (1) (2) (3) (4) (5)

Matrilineality -0.028*** -0.030*** -0.008 -0.105 -0.042 (0.006) (0.007) (0.018) (0.113) (0.075) Female 0.018*** 0.020*** (0.003) (0.003) Female × Matrilineality -0.000 0.001 (0.005) (0.006) Ind. Controls Yes Yes Yes Yes Yes Ethnic Group Controls Yes Yes Yes Yes Yes Village-Geographic Controls Yes Yes Yes Yes Yes Region-survey FE Yes Yes Yes Yes Yes

Gender Both Both Female Female Female Observations 145,619 145,877 69,999 83,908 83,908 Adj. R-squared 0.053 0.053 0.062 0.197 0.167 Clusters 74 74 63 81 81 Mean Dep. Var. (Patri. Males) 0.941 0.934 0.473 22.195 13.980 Notes: OLS estimates are reported with standard errors clustered at the ethnicity level in brackets. The unit of observation is an individual originating from either a traditionally matrilineal or a traditionally patrilineal ethnic group. “Matrilineality” indicates (dummy) whether an individual belongs to a traditionally matrilineal ethnic group. “Female” indicates (dummy) whether an individual is a female. “Female × Matrilineality” indicates (dummy) whether an individual is a female belonging to a traditionally matrilineal ethnic group. “Consent HIV Test” is a dummy indicating whether an individual consented to take the DHS HIV test. “Take HIV Test” is a dummy indicating whether an individual actually took the DHS HIV test. “Anemia” is a dummy indicating whether an individual has any level of anemia (available only for women). “BMI” is the Body Mass Index (available only for women). “Rohrer Index” is available only for women. Controls are defined in Table 1.1. Region-survey is a subnational region defined in DHS, interacted with its survey-year. * p<0.10,** p<0.05,*** p<0.01

97 References

Table 1.C.4 – Heterogeneous Effects by Subsamples: Common Law vs. Civil Law Countries / Polygynous vs. Non Polygynous Individuals

HIV

Ind. not in Ind. in Common Law Civil Law Sample Polygynous Polygynous Countries Countries Union Union (1) (2) (3) (4)

Matrilineality -0.001 -0.002 -0.000 0.002 (0.009) (0.006) (0.005) (0.019) Female 0.009*** 0.008*** 0.011*** -0.001 (0.003) (0.002) (0.002) (0.003) Female × Matrilineality 0.015** 0.002 0.014** 0.007 (0.008) (0.007) (0.006) (0.015) Ind. Controls Yes Yes Yes Yes Ethnic Group Controls Yes Yes Yes Yes Village-Geographic Controls Yes Yes Yes Yes Region-survey FE Yes Yes Yes Yes

Observations 81,422 100,890 147,799 34,513 Adj. R-squared 0.089 0.027 0.083 0.060 Clusters 57 57 100 99 Mean Dep. Var. (Patri. Males) 0.043 0.014 0.023 0.024 Mean Female x Matri. 0.158 0.029 0.095 0.049 Std. Dev. Female x Matri. 0.365 0.167 0.294 0.215 Notes: OLS estimates are reported with standard errors clustered at the ethnicity level in brackets. The unit of observation is an individual originating from either a traditionally matrilineal or a traditionally patrilineal ethnic group. Common Law Countries and Civil Law Countrie classification is from La Porta et al.(2008) dataset. “Matrilineality” indicates (dummy) whether an individual belongs to a traditionally matrilineal ethnic group. “Female” indicates (dummy) whether an individual is a female. “Female × Matrilineality” indicates (dummy) whether an individual is a female belonging to a traditionally matrilineal ethnic group. “HIV” is a dummy indicating whether an individual is HIV positive (from DHS HIV Tests). Controls are defined in Table 1.1. Actual polygyny is not included in the controls in columns 3 and 4. Region-survey is a subnational region defined in DHS, interacted with its survey-year. * p<0.10,** p<0.05,*** p<0.01

98 References

Table 1.C.5 – Considering Ancestral Matrilocality

HIV

Full Not Sample (Controlling for Matrilocal Matrilocal Matrilocality) (1) (2) (3)

Matrilineality -0.003 0.001 (0.007) (0.007) Female 0.008*** 0.009*** -0.027* (0.002) (0.002) (0.013) Female × Matrilineality 0.017*** 0.014* 0.048*** (0.006) (0.007) (0.013) Ind. Controls Yes Yes Yes Ethnic Group Controls Yes Yes Yes Village-Geographic Controls Yes Yes Yes Region-survey FE Yes Yes Yes Ancestral Matrilocality Yes

Observations 182,312 160,079 22,233 Adj. R-squared 0.078 0.067 0.107 Clusters 100 90 10 Mean Dep. Var. (Patrilineal Males) 0.023 0.023 0.021 Mean Female x Matrilineality 0.086 0.031 0.488 Std. Dev. Female x Matrilineality 0.281 0.172 0.500 Notes: OLS estimates are reported with standard errors clustered at the ethnicity level in brackets. The unit of observation is an individual originating from either a traditionally matrilineal or a traditionally patrilineal ethnic group. The sample in column 2 is composed of individuals originating from an ethnic group which was not ancestrally matrilocal; while the sample in column 3 is composed of individuals originating from an ethnic group which was ancestrally matrilocal. “Matrilineality” indicates (dummy) whether an individual belongs to a traditionally matrilineal ethnic group (this is dropped in column 3 for collinearity reason). “Female” indicates (dummy) whether an individual is a female. “Female × Matrilineality” indicates (dummy) whether an individual is a female belonging to a traditionally matrilineal ethnic group. “HIV” is a dummy indicating whether an individual is HIV positive (from DHS HIV Tests). Controls are defined in Table 1.1. Ancestral matrilocality is additionally controlled for in column 1. Region-survey is a subnational region defined in DHS, interacted with its survey-year. * p<0.10,** p<0.05,*** p<0.01

99 References

Table 1.C.6 – Robustness of OLS Estimates to Gender-Specific Alternative Channels

HIV (1) (2) (3) (4)

Matrilineality -0.000 -0.003 0.001 -0.004 (0.006) (0.005) (0.005) (0.005) Female -0.066*** -0.187 0.248 -0.276 (0.015) (0.299) (0.276) (0.319) Female × Matrilineality 0.015** 0.016*** 0.010* 0.016** (0.007) (0.005) (0.006) (0.007) Ind. Controls Yes Yes Yes Yes Ethnic Group Controls Yes Yes Yes Yes Village-Geographic Controls Yes Yes Yes Yes Region-survey FE Yes Yes Yes Yes Ind. Controls × Female Yes Yes Yes Yes Ethnic Group Controls × Female Yes Yes Yes Village-Geographic Controls × Female Yes Yes Region-survey × Female FE Yes

Observations 182,312 182,312 182,312 182,312 Adj. R-squared 0.081 0.081 0.081 0.082 Clusters 100 100 100 100 Mean Dep. Var. (Patri. Males) 0.023 0.023 0.023 0.023 Notes: OLS estimates are reported with standard errors clustered at the ethnicity level in brackets. The unit of observation is an individual originating from either a traditionally matrilineal or a traditionally patrilineal ethnic group (it therefore excludes individuals originating from ethnic groups with alternate inheritance rules (ambilineality, bilinearity, duolinearity, etc.)). “HIV”, “Matrilineality”, “Female”, “Female × Matrilineality”, “Ind. Controls”, “Ethnic Group Controls”, “Village-Geographic Controls”, and “Region- survey FE” are defined in Table 1.1. * p<0.10,** p<0.05,*** p<0.01

100 References is 퐹 ^ 훽 2.8 S.E. + / − 퐹 ^ 훽 of the regression with 2 1 푅 = 2 푀푎푥 is the 푅 2 퐹 ^ 푅 ; 1) = 2 퐹 ^ 푅 2 . Column 6 reports the bounds of the 99.5% ( 2 푀푎푥 )) 푅 is the coefficient in the restricted set and 2 푅 ^ 푚푖푛 푅 푅 ^ 훽 − 2 퐹 ^ ; 1) 푅 ( 2 퐹 = ^ / ) 푅 5 2 퐹 . ^ 푅 1 2 푀푎푥 ( 푅 − in two regressions; in one regression a “restricted” set of controls 푚푖푛 2 푀푎푥 푅 (( 퐹 푒푚푎푙푒 ; 1) × 2 퐹 = × ^ ) 푅 퐹 3 ^ . 훽 1 2 푀푎푥 Minimum Coeff.Lower Bound ( Oster , 2017 ) ( − . 푅 푅 ^ 훽 푚푖푛 ( − 퐹 푒푚푎푙푒 퐹 ^ 훽 푀푎푡푟푖푙푖푛푒푎푙푖푡푦 × of the regression with the restricted set of controls, and 2 푅 (see Altonji et al. , ).2005 Each cell in columns 2-5 report bias-adjusted coefficient lower bounds of ) (1) (2) (3) (4) (5) (6) -4.89 0.014 0.012 0.008 -0.088 [0.0004 ; 0.0328] Test 15.79 0.017 0.017 0.018 0.031 [0.0004 ; 0.0328] 퐹 ^ 167.42 0.019 0.020 0.024 0.099 [0.0004 ; 0.0328] -509.27 0.016 0.016 0.015 -0.001 [0.0004 ; 0.0328] is the 훽 2 푅 − ^ Coeff. Ratio 푀푎푡푟푖푙푖푛푒푎푙푖푡푦 푅 푅 ^ 훽 ( / ( Altonji et al. , 2005 ) 퐹 ^ 훽 Assessing the importance of bias from unobservables by controlling for observable characteristics Controls in Full (F) set from Equation 1.1 from Equation 1.1 from Equation 1.1 from Equation 1.1 Full set of controls Full set of controls Full set of controls Full set of controls based on Oster ( 2017 ). If Table 1.C.7 – 퐹 푒푚푎푙푒 × Robustness Tests: Each cell in column 1 report ratios based on the coefficient of Controls in Ind. Controls Ind. Controls; Region-survey FE Restricted (R) set Region-survey FE; Region-survey FE; Matrilineality; Female Matrilineality; Female; Matrilineality; Female; Matrilineality; Female; Ethnic Group Controls Notes: is included and inthe the left other, side a of “full”the the set coefficient table of (see in controlsTable isthe 푀푎푡푟푖푙푖푛푒푎푙푖푡푦 1.1 included.fullfor set, a In then full both the descriptionthe regressions, ratio is of full the the set sample full of sizesconfidence set controls, areinterval of then the controls the same. of from minimumthe fully TheEquation coefficient controlled controls). 1.1 lower estimate included bound If of in is: each set are listed on

101 References consist of Region-survey ). 0.01 < (Ancestral) Ethnic Group , Husband’s Controls 0.05,*** p < defined in Table 1.1 and computed “Mixed matrilineality” 0.10,** p < Individual controls Note: * p Village-Geographic Controls Dep. Var.: Couple’s Serostatus and ), as well as a dummy indicating whether the wife and the husband originate from are computed for the wife and are defined in Table 1.1 . Ancestral Matrilineality and Couples HIV Discordance (OLS) (1) (2) (3) (4) (5) (6) (7) (8) Wife - / Husband - Wife + / Husband - Wife -(0.0086) / Husband + (0.0096) (0.0070) Wife + / Husband (0.0076) + (0.0066) (0.0068) (0.0062) (0.0058) “Mixed ethnicity” (Ancestral) Ethnic Group Controls , Table 1.C.8 – Village-Geographic Controls and OLS estimates are reported with standard errors clustered at the ethnicity level in brackets. The unit of observation Individual controls MatrilinealityInd. ControlsEthnic Group ControlsVillage-Geographic ControlsRegion-survey FEHusband’s Yes Controls 0.0103 YesObservations 0.0107 YesAdj. R-squared YesClusters 0.0133* YesMean Yes Dep. 0.0131* Var. Yes (Patri.)Prop. Mixed Yes Ethnicity YesProp. -0.0018 Mixed Matrilineality Yes Yes 0.955 0.0025 25,272 Yes Yes 0.048 0.100 Yes 0.193 Yes -0.0218*** 23,917 0.957 Yes -0.0263*** 0.044 0.106 Yes 0.156 25,272 Yes 0.014 Yes 82 0.048 0.021 23,917 Yes 0.193 Yes 0.013 Yes 0.044 25,272 Yes 0.024 Yes 82 0.156 Yes 0.016 Yes 23,917 0.048 0.027 Yes 0.193 Yes 82 0.015 25,272 0.044 0.030 Yes Yes 0.156 Yes Yes 0.015 23,917 82 Yes 0.048 0.056 0.193 Yes 0.015 Yes 82 0.044 0.061 0.156 82 Yes 82 82 Notes: is a non-polygynous formallyor married a couple traditionally with patrilinealmatrilineal ethnic both ethnic group. wife group. and “Matrilineality” “Wife is husband - a originating / dummy from Husband indicating eitherHusband -” whether a +” is the traditionally is wife a matrilineal a belongs dummyis dummy indicating to a indicating whether a dummy whether both traditionally indicating the the whether wife wife both is and the HIV the wifethe negative husband and while are the the HIV husband husbandfor negative. are is the HIV HIV husband, positive. positive. in addition “Wife of + a / dummyethnic Husband indicating groups +” whether with the differentis wife ancestral a and kinship subnational the organizations region husband (matrilineal originate defined from vs. patrilineal) ain DHS, ( interacted different Ethnographic with survey-year. its Atlas ancestral ethnic group ( Controls “Wife + / Husband -” is a dummy indicating whether the wife is HIV positive while the husband is HIV negative. “Wife - /

102 References

Table 1.C.9 – Ancestral Matrilineality and Fertility (OLS)

Want another child Nb. children (1) (2)

Matrilineality -0.070*** -0.129*** (0.020) (0.057) Female -0.247*** 0.188*** (0.012) (0.041) Female × Matrilineality 0.127*** 0.153*** (0.030) (0.051) Ind. Controls Yes Yes Ethnic Group Controls Yes Yes Village-Geographic Controls Yes Yes Region-survey FE Yes Yes

Observations 120,501 182,312 Adj. R-squared 0.310 0.643 Clusters 84 100 Mean Dep. Var. (Patri. Males) 0.787 3.155 Notes: OLS estimates are reported with standard errors clustered at the ethnicity level in brackets. The unit of observation is an individual originating from either a traditionally matrilineal or a traditionally patrilineal ethnic group. “Matrilineality” indicates (dummy) whether an individual belongs to a traditionally matrilineal ethnic group. “Female” indicates (dummy) whether an individual is a female. “Female × Matrilineal- ity” indicates (dummy) whether an individual is a female belonging to a traditionally matrilineal ethnic group. “Want another child” is a dummy indicating whether the individual wants another child. “Nb. children” is the number of biological children ever born. Controls are defined in Table 1.1 (However, number of children is not controlled for anymore). Region-survey is a subnational region defined in DHS, interacted with its survey-year. * p<0.10,** p<0.05,*** p<0.01

103 References is a Never tested Ever tested Region-survey Full sample Condom Never tested Ever tested Females Males (0.006) (0.009)(0.004) (0.007) (0.007) (0.006) (0.016) (0.023) (0.008) (0.021) (0.016) (0.011) Full sample (1) (2) (3) (4) (5) (6) (7) (8) (0.005)(0.005) (0.009) (0.009) * p<0.10,** p<0.05,*** p<0.01 Heterogeneity in Condom Use by Perception of Risk of HIV Transmission (OLS) No HIV -0.007* -0.003 -0.005 0.003 -0.008 0.006 Condom reduces HIVCondom not reduces HIV -0.005 -0.011**HIV 0.021*** -0.006 0.009 0.034*** -0.003 0.048*** 0.084*** -0.008 × × × × Table 1.C.10 – OLS estimates are reported with standard errors clustered at the ethnicity level in brackets. The unit of observation is an Matrilineality Ind. ControlsEthnic Group ControlsVillage-Geographic ControlsRegion-survey FEObservationsAdj. R-squaredClustersMean Dep. Var. (Patri.)F-Test Equality of Yes coeff. (p-value) Yes Yes Yes Yes Yes 0.310 Yes Yes Yes 0.000 Yes 0.038 72,855 Yes 0.073 Yes Yes 89,081 0.031 Yes Yes 0.073 30,795 Yes Yes0.000 0.055 84 0.079 54,814 Yes Yes 0.241 Yes 0.021 52,796 Yes 0.067 Yes 84 0.000 0.151 58,094 Yes Yes 0.237 Yes 17,683 0.001 0.138 Yes Yes 0.239 82 39,954 Yes Yes 0.000 0.414 0.236 Yes 0.240 84 Yes 0.106 0.231 Yes 74 74 74 73 Matrilineality Matrilineality Matrilineality Notes: individual originating from either a traditionally matrilineal or a traditionallyindicating patrilineal whether ethnic an group. individualindividual “Matrilineality” is reports indicates HIV “condom (negative)subnational (male)” positive region according as defined to her DHShave currentin DHS, ever test. interacted contraceptive been method. “Condom” tested with isbeen for survey-year. its Controls a tested AIDS are dummy for beforeSamples indicating AIDS defined DHS “Ever whether before survey. in an DHS SamplesTabletested” survey. “Never 1.1 . tested” in incolumn column 3 4 and and 7 consist 8 individuals who of consist of individuals who have never whether an individual thinks that always using condoms during sex (does not) reduces chance of getting HIV. “(No) HIV” is a dummy (dummy) whether an individual belongs to a traditionally matrilineal ethnic group. “Condom (not) reduces HIV” is a dummy indicating

104 References

Table 1.C.11 – Ancestral Matrilineality, Acknowledgment of HIV Risks and Access to Condom (OLS)

Acknowledgment of Risks Access to condom

Think having Know a Heard of Heard of Think condom Can get one partner source to AIDS STI reduces HIV condom reduces HIV get condom (1) (2) (3) (4) (5) (6)

Matrilineality -0.007 -0.006 0.005 -0.002 -0.026 0.007 (0.005) (0.004) (0.009) (0.007) (0.017) (0.017) Female -0.027*** -0.025*** -0.048*** -0.031*** -0.191*** -0.286*** (0.005) (0.004) (0.007) (0.006) (0.014) (0.025) Female × Matrilineality 0.022*** 0.021*** 0.007 -0.007 0.092*** 0.037 (0.006) (0.005) (0.013) (0.009) (0.023) (0.030) Ind. Controls Yes Yes Yes Yes Yes Yes Ethnic Group Controls Yes Yes Yes Yes Yes Yes Village-Geographic Controls Yes Yes Yes Yes Yes Yes Region-survey FE Yes Yes Yes Yes Yes Yes

Observations 182,284 182,218 156,133 164,673 164,368 103,316 Adj. R-squared 0.121 0.104 0.048 0.051 0.337 0.161 Clusters 100 100 100 100 100 99 Mean Dep. Var. (Patri. males) 0.966 0.974 0.852 0.913 0.761 0.878 Notes: OLS estimates are reported with standard errors clustered at the ethnicity level in brackets. The unit of observation is a woman originating from either a traditionally matrilineal or a traditionally patrilineal ethnic group. “Matrilineality” indicates (dummy) whether an individual belongs to a traditionally matrilineal ethnic group. “Female” indicates (dummy) whether an individual is a female. “Female × Matrilineality” indicates (dummy) whether an individual is a female belonging to a traditionally matrilineal ethnic group. “Heard of AIDS” is a dummy indicating whether an individual has ever heard of AIDS. “Heard of STI” is a dummy indicating whether an individual has ever heard of any STI. “Think condom reduces HIV” is a dummy indicating whether an individual thinks that always using condoms during sex reduces chance of getting HIV. “Think having one partner reduces HIV” is a dummy indicating whether an individual thinks that having only one sexual partner reduces chance of getting HIV. “Know wource to get condom” is a dummy indicating whether an individual knows a source to get male condoms. “Can get condom” is a dummy indicating whether an individual can get herself a male condom. Controls are defined in Table 1.1. Region-survey is a subnational region defined in DHS, interacted with its survey-year. * p<0.10,** p<0.05,*** p<0.01

105 References

Table 1.C.12 – Ancestral Matrilineality and Sexual Debuts (OLS)

Age first sex Age first marriage (1) (2)

Matrilineality -0.783*** -0.739*** (0.197) (0.216) Female -1.807*** -4.947*** (0.257) (0.251) Female × Matrilineality 1.257*** 1.405*** (0.298) (0.336) Ind. Controls Yes Yes Ethnic Group Controls Yes Yes Village-Geographic Controls Yes Yes Region-survey FE Yes Yes

Observations 150,197 128,860 Adj. R-squared 0.238 0.444 Clusters 100 100 Mean Dep. Var. (Patri. Males) 18.602 23.996 Notes: OLS estimates are reported with standard errors clustered at the ethnicity level in brackets. The unit of observation is an individual originating from either a traditionally matrilineal or a traditionally patrilineal ethnic group. “Matrilineality” indicates (dummy) whether an individual belongs to a traditionally matrilineal ethnic group. “Female” indicates (dummy) whether an individual is a female. “Female × Matrilineality” indicates (dummy) whether an individual is a female belonging to a traditionally matrilineal ethnic group. “Age first sex” is the age at first sexual intercourse. “Age first marriage” is theage at start of first marriage or union. Controls are defined in Table 1.1. Region-survey is a subnational region defined in DHS, interacted with its survey-year. * p<0.10,** p<0.05,*** p<0.01

106 References

Table 1.C.13 – Summary of Model’s Parameters

Parameter Value Interpretation Source

푀 훼푓 0.076 Matrilineal female’s number of extramarital partners The author (Table 1.7) 푃 훼푓 0.008 Patrilineal female’s number of extramarital partners The author (Table 1.7) 훼푚 0.204 Male’s number of extramarital partners The author (Table 1.7) 푀 훾푓 0.999 Matrilineal female’s probability of not using condom The author (Table 1.8) 푃 훾푓 0.968 Patrilineal female’s probability of not using condom The author (Table 1.8) 훾푚 0.862 Male’s probability of not using condom The author (Table 1.8) 휌푚 0.0045 Male’s contamination risk for one-time unprotected sex Greenwood et al.(2019) 휔푓 1.75 Gender difference in biological susceptibility to contract the virus Greenwood et al.(2019)

푀 Matri. female’s proba. of infection per each contact with a HIV+ 훿푓 0.007867 푀 The author (Calculation) male (= 훾푓 × 휌푚 × 휔푓 )

푃 Patri. female’s proba. of infection per each contact with a HIV+ 훿푓 0.007623 푃 The author (Calculation) male (= 훾푓 × 휌푚 × 휔푓 ) Male’s probability of infection per each contact with a HIV+ 훿푚 0.003879 The author (Calculation) female (= 훾푚 × 휌푚)

휇푚 4 Male’s nb. of sexual intercourses with LT committed partner The author (Assumption) 휇푓 4 Female’s nb. of sexual intercourses with LT committed partner The author (Assumption) 휏푚 4 Male’s nb. of sexual intercourses with each extramarital partner The author (Assumption) Matrilineal female’s nb. of sexual intercourses with each 휏 푀 4 The author (Assumption) 푓 extramarital partner

푃 Patrilineal female’s nb. of sexual intercourses with each 휏푓 3.836 푀 The author (Assumption) extramarital partner (= 0.959 × 휏푓 )

푖푓,0 0.01 Initial female’s HIV rate The author (Assumption) 푖푚,0 0.01 Initial male’s HIV rate The author (Assumption)

107 References

Table 1.C.14 – Simulated HIV Rates (Alternative Scenarios)

Scenario 2 (휇푚 = 4 and 휏푚 = 2) Matrilineality Patrilineality Raw Months Females Males Difference Females Males Difference Diff-in-Diff

1 0.010 0.010 0.000 0.010 0.010 0.000 0.000 25 0.020 0.016 0.004 0.019 0.016 0.003 0.001 50 0.037 0.028 0.009 0.035 0.027 0.008 0.001 75 0.065 0.048 0.017 0.061 0.046 0.015 0.002 100 0.112 0.082 0.030 0.103 0.078 0.025 0.005 120 0.168 0.124 0.044 0.153 0.116 0.037 0.007

Scenario 3 (휇푚 = 휏푚 = 2) Matrilineality Patrilineality Raw Months Females Males Difference Females Males Difference Diff-in-Diff

1 0.010 0.010 0.000 0.010 0.010 0.000 0.000 25 0.015 0.013 0.002 0.014 0.013 0.001 0.001 50 0.021 0.017 0.004 0.020 0.017 0.003 0.001 75 0.029 0.022 0.007 0.027 0.022 0.005 0.002 100 0.039 0.030 0.009 0.036 0.029 0.007 0.002 120 0.050 0.038 0.012 0.046 0.036 0.010 0.002 Notes: This table reports the simulated HIV rates, based on the compartmental SI epidemic model detailed in section 1.6, assuming 휇푚 = 4 and 휏푚 = 2 in the upper part of the table; and 휇푚 = 휏푚 = 2 in the lower part of the table.

108 109 References

110 Chapter 2

Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia

This chapter is based on a joint research with Olivier Bargain (Bordeaux University) and Roberta

Ziparo (Aix-Marseille University).

Abstract: Social norms can interact with formal institutions in shaping women’s autonomy. We examine this question in the context of legal reforms and their differentiated impact on divorce and empowerment across traditional modes of post-marital cohabitation. Global evidence first shows that the degree of ancestral matrilocality (i.e. the practice of living with the bride’s relatives after marriage) correlates with contemporaneous opinions about gender role. This is especially the case in countries with low divorce rates such as Indonesia. We then exploit a policy experiment for this country, which exogenously fosters women’s access to justice and ability to divorce. We theoretically establish how women originating from matrilocal ethnic groups should respond to the reform compared to those from patrilocal ethnicities. We confirm the model predictions using a panel difference-in-difference approach: the former divorce more and, when in stable marriages, experience a significant improvement in well-being and empowerment. This result is consistently obtained for a broad range of outcomes including women’s health, fertility control, asset value, women’s and children’s well-being as

111 Chapter 2. Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia well as women’s ﬁnal say over key decisions. Modern legal reforms compound with ancestral norms and exacerbate potential inequalities between women of different ethnic origins. This conclusion calls for better tailored policies that can transcend cultural contexts and overcome the adherence to informal laws.

Keywords: Legal Reforms, Divorce, Ethnic Norms, Intra-Household Decision-Making

JEL Classiﬁcation: D13, I15, I38, J16, K36, Z13

112 Chapter 2. Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia

2.1 Introduction

It is well established that the legal and institutional framework of a country can greatly contribute to women’s autonomy (Duﬂo, 2012). In particular, changes in divorce laws and legal rights can improve women’s empowerment within marriage (Voena, 2015). However, cultural norms can mitigate policy effects, especially in the context of developing countries characterized by great ethnic diversity. Ad- mittedly, traditional norms per se have garnered considerable attention in both anthropological and economic studies, notably their association with gender-related development outcomes.1 Yet, less is known about the way ancestral social norms interact with formal institutions.

With this question in mind, some authors have examined the impact of large-scale development policies depending on different cultural contexts stemming from ethnical diversity (Ashraf, Bau,

Nunn, and Voena, 2020, La Ferrara and Milazzo, 2017).2 The present paper contributes to this nascent literature by offering evidence on the role of matrilocality – the practice of living with or near the bride’s parents after marriage – regarding the effectiveness of pro-women laws. We start with a simple motivation: ancestral norms of post-marriage residence are still relevant today. Using global variation, we ﬁnd a strong association between the degree of ancestral matrilocality and contemporaneous opinions about women’s autonomy. This is especially the case for countries with a low access to divorce such as Indonesia. We then move to the context of this country.3 Comprising more than 300 different ethnic groups, it exhibits a wide heterogeneity in terms of kinship norms, with large pockets of matrilocal communities (Rammohan and Robertson, 2012). Ethnic-based customs, the so- called Adat system, still guide family life and often prevail over religious and legal laws (Buttenheim and Nobles, 2009). In particular, contemporaneous attitudes towards divorce and gender rights are strongly associated with traditional post-marriage arrangements as prescribed by the local Adat.

1Jayachandran(2015) describes how various cultural practices – such as patrilocality, patrilinearity or the payment of bride price or dowry – may affect women’s outcomes. Several studies examine in particular the effect of bride price on early marriages (Corno, Voena, et al., 2016) and women’s well-being (Lowes, Nunn, Robinson, and Weigel(2017) or the effect of matrilinearity on household cooperation and children’s outcomes (La Ferrara, 2007, Lowes, 2018). Recent studies explore the origins of cultural norms using ethnic diversity (Alesina and Ferrara, 2005) – in particular how socioeconomic conditions and agricultural practices may have shaped gender roles (Alesina, Giuliano, and Nunn, 2013, Doepke and Tertilt, 2009, Alesina, Brioschi, and Ferrara, 2016) and norms such as matrilinearity (BenYishay, Grosjean, and Vecci, 2017). 2Extending Duflo(2001), Ashraf et al.(2020) find that Indonesian ethnic groups that traditionally engage in bride price payments at marriage increase female enrollment in response to a large school construction program. La Ferrara and Milazzo(2017) study the introduction of quotas for the land that parents should devolve on their children in Ghana and find a negative impact of the reform on educational outcome ofboys originating from matrilineal ethnic group. Hence, the effectiveness of development policy may strongly vary with traditional rules or even crowd the practices of the norm in some cases (Bau, 2016). 3Despite recent improvements in women’s legal rights, Indonesia was ranked 110th by the United Nations (Gender Inequality Index of 0.494) and 92nd by the World Economic Forum (Global Gender Gap Index of 0.681) in 2015.

113 Chapter 2. Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia

To research what cultural norms imply for policy effectiveness in a context of pro-women legislation, we exploit a series of country-wide reforms that have fostered women’s access to justice and their ability to divorce in 2008-2010. A simple household model with limited commitment and changes in outside options predicts that women of matrilocal tradition are more likely to divorce and, when the marriage lasts, beneﬁt from renegotiation. The natural experiment allows us to test these predictions using difference-in-difference estimations with ﬁxed effects on the Indonesia Family Life

Survey (IFLS). We confirm that Indonesian women originating from matrilocal ethnic groups are more responsive to the reform than those from patrilocal ancestry. They tend to divorce more and, if remaining with their partner, experience a significant improvement in their bargaining position and welfare, as shown through a broad array of outcomes including health status, control over their own fertility, asset accumulation, own and child well-being as well as final say over key life decisions.

This paper contributes to several strands of the literature. First, we bring new evidence on the role of access-to-justice legal reforms, and divorce-related laws, on marital breakdown, households’ intertemporal behavior and women’s empowerment. Most of this literature examines how unilateral divorce, in Western countries, makes the threat of divorce credible and hence possibly affects intra- household allocation during marriage.4 This type of mechanism is rarely studied in the context of poorer countries, especially in muslim settings where divorce is not common.5 We show that in the

Indonesian case, a change in legal rights in favor of women can be perceived as sufﬁciently effective

– at least among ethnicities of matrilocal tradition – to affect the probability of divorce and the degree of women’s empowerment. Second, the reform examined in this paper is different from, and complementary to, the policies studied in related contributions including educational programs in Ashraf et al.(2020) and wealth transmission control in La Ferrara and Milazzo(2017). Legal laws are also important for they can be very effective in changing people’s living arrangements. We show that they can pervade the private sphere and inﬂuence intra-household mechanisms in some segments of the population. Third, this paper is one of the few that test how the efﬁcacy of pro-women policies can vary with customs. We use explicit measures of traditional norms about post-marriage residence to

4Stevenson and Wolfers(2006) evaluate how the adoption of laws allowing unilateral divorce across the United States changed family violence and whether the option provided by unilateral divorce reduced suicide and spousal homicide. As in Stevenson and Wolfers(2006), we expect changes in women’s position in marriage to be driven both by selection into marriage and change in the pareto-weight of women that stay married. From a theoretical point of view, our analysis complement their analysis as we study the interaction of formal and informal laws in shaping the evolution of women’s outside option. From an empirical point of view, the structure of our database allows us to dig more into the proposed mechanisms: the use of household fixed effects allows us to measure the effect of the reforms on stable couples, as wediscussin section 2.4. 5Some evidence exists for middle income countries. Sun and Zhao(2016) document how China’s pro-women divorce reform has empowered women within marriage and reduced health-damaging sex-selective abortion.

114 Chapter 2. Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia assess the contribution of culture to the effectiveness of legal reforms. It turns out that legal reforms cannot target the most disadvantaged women if policies are not better tailored to transcend cultural contexts and overcome the adherence to informal laws.

2.2 Background on Social Norms and Legal Reforms

We provide some background on traditional residence norms and their correlation with women’s empowerment using global evidence and a focus on Indonesia. The section ends with a description of the reforms under study.

2.2.1 Traditional Norms and Female Empowerment: Global Evidence

Post-marital residence norms have long been emphasized by anthropologists and sociologists as a social structure shaping household organisation. As such, different ethnic groups engage in different practices, which have been categorized as follows: matrilocality (married couples live with or near the bride’s family), patrilocality (they live with or near the groom’s family), ambilocality (they can live with or near either spouse’s parents) and neolocality (they can set their own household, i.e. the basis of most developed nations).

Traditional versus Actual Residence Norms and Women’s Outcomes. Many anthropological and economic studies attribute lower education, a lower marriage age and low levels of autonomy to women in groups adopting patrilocality.6 The basic explanation recalled by Sundaram and Vanneman,

2008 pertains to the fact that parents are dissuaded from investing in their daughters’ education if they leave the home after marriage. Selective abortion in East/South Asia and South Caucuses is also linked to the fact that daugthers in patrilocal families cannot provide care once the parents are old

(Ebenstein, 2014). Patrilocal contexts may also increase husbands’ outside options due to the pressure exerted by the presence of their own relatives on the wife. Yet, living with his or her relatives is not the only aspect affecting spouses’ relative empowerment. More generally, matrilocality or patrilocality are salient features of a broader ethnic diversity in family customs and gender roles. For this reason, we will focus on the heterogeneity in terms of traditional rather than actual post-marriage residence in our empirical approach. Another reason to choose traditional norms is that actual arrangements may

6This is the case for instance in Dyson and Moore(1983) for northern India, Garg and Morduch(1998) for Ghana and, in the Indonesian context, Buttenheim and Nobles(2009), Rammohan and Johar(2009) and Rammohan and Robertson(2012).

115 Chapter 2. Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia be highly correlated with a particular couple’s unobserved heterogeneity and, hence, reﬂect more than what the norm entails.

Origins of Residence Norms. Several explanations have been given by anthropolgists for the emergence of residence norms.7 Patrilocality might have originated from a greater productive role attributed to sons, from a larger bargaining power given to sons, or from the need to locate multiple women within a husband’s household in a polygynous setup (Edlund, 2001). Botticini and Siow,

2003 explain how patrilocal societies transfer wealth via dowries for their daughters and via bequests for their sons in order to maintain the incentives of the latter to exert effort on the family farm. Sons’ incentives may also depend on paternity uncertainty, which may be reduced if their parents monitor the sexual behavior of the son’s wife, hence explaining how patrilineal inheritance and patrilocal exogamy reinforce each other (Guha, 2010).

Evolutionary anthropologists further highlight the coevolution of inheritence and post-marital residence practices. This can be rationalized by the fact that matrilocality would allow children to grow up with their mother’s family, which is their lineage group under matrilineality. While the so-called

“Main Sequence Theory” posits that changes in residence rules precede change in other social structures, Opie, Shultz, Atkinson, Currie, and Mace(2014) challenge this theory. Using linguistic data from 542 Bantu languages to construct a Bantu phylogenetic tree and using Bayesian models, they uncover the history of Bantu kinship patterns and trace the interplay between descent and residence systems. Their results suggest that when groups adopt matrilineal systems, they then subsequently adopt matrilocal practices (rather than ﬁrst becoming matrilocal and then adopting matrilineal inheritance).8 The origins of matrilocality may therefore also be linked to the origins of matrilineality.9

One hypothesis is that matrilineality would have emerged in environments with high sexual promiscuity and low paternal certainty because, while it is difﬁcult conﬁrm paternity, maternity is easily observable. Thus, an inheritance system in which property passes from the mother’s brother to her sons may be optimal since the brother knows he is related to his sister, but cannot verify that he is related to his children (Fortunato, 2012). Alternative theories highlight that types of production are factors contributing to the emergence of descent rules. Jones(2011) posits that matrilineality tends to occur in horticultural societies where women often have a more dominant role in agriculture, while

7Interested reader may find an overview in Bau(2016). 8In line with these two competing theories, Bau(2016) finds a worldwide positive correlation between matriliocality and matrilineality. 9Interested reader may find overviews in Lowes(2018) and Loper(2020).

116 Chapter 2. Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia patrilineality would have been favored when main types of production required skills and male coop- erations, such as pastoralism (the cattle in particular) where males have a larger role in agricultural production (Holden and Mace, 2003).

Overall, ancestral residence norms carry a lot of information about gender rights and roles within an ethnic group, as the emergence of norms has been closely associated with the values driving inheritance, marriage age, the practice of dowry/bride price or the timing of fertility.10

10Bau(2016) finds that, worldwide, matrilocality is correlated with the practice of dowry, while patrilocality is correlated with the practice of bride price.

117 Chapter 2. Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia Share of firms with female ownership Survival Health and Economic Participation Divorce justifiable men rights as the same Women have better business executives Men make (1) (2) (3) (4) (5) (6) (7) (0.109)(0.117) (0.113)(0.126) (0.111) (0.214) (0.148) (0.192) (0.255) (0.423) (0.279) (0.043) (0.531) (0.035) (0.004) (0.035) (0.003) (0.069) (0.003) (0.063) (0.070) (0.089) (0.085) (0.160) (0.244) (0.033) (0.002) (0.060) Men have a job when more rights to jobs are scarce 2nd Tertile3rd Tertile -0.054 -0.025 -0.071 -0.133 0.200 0.254 0.423 0.211 0.082** 0.084** 0.004 -0.001 0.126* 0.048 1st Tertile -0.359*** -0.390*** 0.495** 0.604** 0.163*** 0.009** 0.179** × × × Worldwide Correlations between Ancestral Relative Matrilocality and Contemporaneous Attitudes towards Women’s Role (Relative Matri.) (Relative Matri.) R-squaredTertile 1 = Tertile 2Tertile 2 (p-value) = Tertile 3 (p-value) 0.023 0.045 0.024 0.181 0.673 0.215 0.647 0.685 0.497 0.450 0.592 0.030 0.044 0.768 0.230 0.053 0.544 0.376 0.259 0.060 0.294 Average Effect Relative MatrilocalityHeterogenous Effect by Countries’(Relative Intensity Matri.) Divorce of -0.167* -0.205** 0.322** 0.457*ObservationsMean Dep. Var. 0.116*** 0.005** 0.140** 1.891 73 2.370 72 8.093 71 4.762 73 0.643 0.972 70 0.347 70 51 R-squared 0.618 0.636 0.580 0.764 0.496 0.204 0.252 Country-level linear estimates of contemporary women’s outcomes on the relative degree of matrilocality, calculated as the proportion of citizens from ancestral matrilocality minusfrom the the proportion World from Value ancestral Surveys (WVS) patrilocality modules (source: onAlesinaessential self-report et characteristics attitudes al. , of towards2013 ). democracy genderindices (approval Outcomes roles, ranging on in asking from a columns whether: 0 1-10 1-4include: to scale), (1) are 1, (5) (4) men drawn drawn an divorce have from index justifiable morein the (approval based rights sex 2016 to on ratio World on a the at Economic a job participation birth Forum1-10 scale). gap, (columns and the Outcomes 5-6) in remuneration andsuitability healthy gap in the for life and World column agriculture, expectancy, Bank the 5-7 are fraction (7) Enterprisebased advancement of the Surveys gap, on ancestral (column percentage (6) the land 7). of an WVS thatthe firms index They also was same with measuring include tropical estimates some gender average or conditional respondants’ female difference Indonesia subtropical, on characteristics ownership ancestral is the (gender,(country domestication ranked age, intensity surveys 8th of of age conducted from large divorce squared over the animals, in and and bottom the education). estimations and country In (tertiles hence the of below lower the to panel, country the we proportion first provide of tertile. divorced-separated Robust respondants). standard errors are reported in brackets. Significance levels: 2003-2010). All the regressions include country-level controls: log GDP/capita and its square, proportion of pre-industrial plough use, ancestral when jobs are scarce (from 1-disagree to 3-agree), (2) men make better business executives (from 1-disagree to 4-agree), (3) gender equality as Table 2.1 – * p<0.10, ** p<0.05, *** p<0.01. Equality tests show that in 5/7 cases, the first tertile is significantly different from the rest of the distribution.

118 Chapter 2. Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia

Global Evidence. We also provide further motivation for studying the importance of ancestral residence norms. In the upper panel of Table 2.1, we use country variation of indices drawn from the

World Value Surveys, the World Economic Forum and the World Bank Enterprise Surveys on attitudes towards gender rights and women’s outcomes (economic activity, health, female ﬁrm ownership). We regress these indices on the country-level relative prevalence of traditional matrilocality and a broad set of controls. We ﬁnd a systematic relation between pro-women outcomes and the ancestral degree of matrilocality. In the lower panel, we show that these results are mainly driven by countries in the lowest tertile of the divorce rate distribution. In words, ethnic cultural diversity transpires in terms of gender attitudes in more traditional countries such as Indonesia (this country is ranked 8th from the bottom of the divorce intensity distribution).

2.2.2 Traditional Residence Norms: the Indonesian Context

Against this background, Indonesia is a particularly relevant ﬁeld of investigation given its extraor- dinary ethnic diversity and the noticeable variety of social norms that derive from it. Three points should be made.

First, residence norms are not geographically polarized. This can be seen on the map of Figure 2.1, where we show the location of villages interviewed in the IFLS and the prevailing residence norm of each village. In most of the regions, villages of both patrilocal and matrilocal ancestries are surveyed.

This is good news because otherwise, a different exposure to the reform under study could simply be related to spatial heterogeneity (or other aspects related to living in speciﬁc areas). We will actually explore this type of confounders more deeply later by additionally controlling in our estimations for the distance to cities where the new legal rights can be exerted.

Second, we have discussed the relevance of traditional norms in general and stress here the role of informal laws (Adat) in the Indonesian context. They shape many aspects of family life and are historically associated with ethnic differences in family-related behavior including marriage, inheritance, land-holding and dispute resolution. In particular, divorce and polygyny are relevant cases of tensions between Adat customs, religious laws and state laws that have emerged during the colonial period and still persist (Buttenheim and Nobles, 2009). Traditional residence norms, as the salient part of Adat rules, will be exploited in our empirical analysis as they potentially ﬁlter the effect of legal reforms

119 Chapter 2. Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia

Figure 2.1 – Village-level Traditional Post-Marital Residence (IFLS data)

while reﬂecting couples’ attitude towards divorce. It is possible that women of matrilocal tradition beneﬁt more from the new legal support and the increased possibility of divorce.11

Finally, we document a similar type of association between ancestral matrilocality and attitudes towards gender roles as we have done using cross-country variation. We rely on the description of traditional norms by village-level Adat experts as provided in the 1997 IFLS. Simple estimations reported in Appendix Table 2.B.1 show that in case of divorce, villages of matrilocal traditions tend to favor women: the divorce ruling more often takes place in a religious/civil court, the husband has less often the right to claim pre-marriage assets or assets acquired since marriage, young children are less likely to follow the husband or his relatives.12 Note however that divorce is relatively infre- quent so that there is no clear pattern of matrilocal-advantage in marriage during the pre-reform years

(2000 and 2007 IFLS), at least concerning the empowerment outcomes used in our double difference approach, as will be seen thereafter (and also found in Levine and Kevane 2003). In fact, a clear advantage for women of matrilocal tradition will appear after the reforms.

11In a study of Adat marriage patterns among the Sasak, Grace(2004) contends that interpretations of both Islamic law and state law related to marriage are often shaped by local Adat. Consistently, an access-to-justice reform may not be appreciated equally by women of different ethnic traditions. 12We do not find any village-level correlation between traditional matrilocality and traditional practice of bride price (not reported). This is in line with Bau(2016) who does not find any significant correlation between matrilocality and the practie of bride price in Indonesia using alternative datasets. Unfortunately, Adat questionnaire in IFLS does not provide information on descent rules. Note however that Bau(2016) finds a strong correlation between matrilocality and matrilineality in Indonesia using alternative datasets

120 Chapter 2. Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia

2.2.3 National Access-to-Justice Strategy: a Natural Experiment

Indonesia is characterized by an Islamic justice system in which religious courts have exclusive jurisdiction over cases where the parties are Muslim and which involve marriage-related cases (1974

Marriage Law). These cases mostly concern divorce and related matters including property division, child custody or spousal maintenance. Given the prevalence of Islam, around 98% of divorces are pronounced by religious courts (the remaining cases are heard by general courts). A critical aspect is the ability to exercise one’s rights. In particular, women’s access to justice is positively related to gender equity developments (Alﬁtri, 2011), especially by making them more assertive about their right to divorce. However, these rights may be constrained by a lack of information, by the cost of court cases (half of Indonesia citizens live below $2 a day) or by the social consequences for women who institute divorce proceedings through the formal legal system. These conclusions have been reached by the ‘Access and Equity’ report by the Family Court of Australia, AusAID and other stakeholders.

In this context, and following this report, the Indonesian government has launched the National ‘Ac- cess to Justice’ Strategy during 2008-2010. This program aimed at increasing access to the courts for women and disadvantaged groups (Sumner, Zurstrassen, and Lister, 2011). It comprises three pillars.

First, in 2008 and 2009, the Justice for the Poor scheme has been put in place, ﬁnancially supported by AusAID, the Family Court of Australia and the World Bank. To alleviate the ﬁnancial constraint of poor households, it has substantially increased religious courts’ budgets in order to waive legal fees. It has also increased the capacities of circuit courts, namely courts travelling to subdistricts in order to hold hearings for family law cases in rural and remote areas. Second, the laws 48, 49 and

50 on Judicial Authority and General/Religious Courts were passed in 2009, requiring both types of courts to provide an extensive range of services that could improve women’s access to courts (court fee waivers, legal aid services, legal assistance to clients who cannot afford lawyers). Finally, the

Presidential regulation n.5 passed in 2010 has provided additional budget for fee-waiver schemes, circuit courts and legal aid services.

The effect of this series of reforms has been documented in several reports, showing a signiﬁcant increase in the ability of women, the poor or those living in remote areas to access courts and exercise their rights. Sumner et al.(2011) explain that the awareness and public conﬁdence in the legal system has also increased dramatically. The authors show that the access to courts has substantially increased, i.e. the number of people accessing religious courts through fee waiver (resp. circuit courts) has been multiplied by 20 (resp. 6) in 2011 compared to 2007. As illustrated in Figure 2.2, an increase in

121 Chapter 2. Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia divorce is observed around the time of the reform. Maybe more impressive is the rise in divorce cases initiated by the wives, which exceed 70% of all divorces in the recent years. Finally, courting reforms have also helped women in case of domestic violence and to break cycles of illegal marriage, divorce or births. That is, they can legally register births, marriages and divorces, which are important steps to establish their legal identity and to enforce their rights, for instance their entitlement to poverty alleviation or healthcare programs for themselves or their children (Sumner and Lindsey, 2011). Since these evolutions improve their outside options in case of divorce, the reforms may also improve women’s and children’s situation within marriage, as we shall see.13

Figure 2.2 – Divorce Trends around Reform Time

2.3 Theoretical Framework

We suggest a conceptual framework to pinpoint the main mechanisms at stake. It is based on a simple dynamic model with limited commitment (Mazzocco, 2007, Chiappori and Mazzocco, 2019), more speciﬁcally derived from Voena(2015) and aimed to elucidate the channels through which traditional norms and divorce laws may affect the probability of divorce and the position of women in marriage.

We consider a household made of two individuals 푗 = 퐻, 푊 , husband 퐻 and wife 푊 , who are

13Note that these evolutions might also affect women’s and child’s well-being through additional channels, such as increased access to healthcare programs. Still, the predicted effect of the reforms through these other channels should go in the same direction than the main predicted effect.

122 Chapter 2. Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia married at time 1 and live T periods. At each period 푡, they must decide whether to stay married, and how to allocate resources between the private consumption of each spouse and savings. The household, denoted ℎ = 푀, 푃 , can either belong to a matrilocal 푀 or a patrilocal 푃 ethnic group.

2.3.1 Preferences

Both spouses derive utility from their own private consumption 푐푗ℎ and joint consumption of a public good 푄ℎ. For the latter, the production function is written:

[︁ 1 1 ]︁휌 푄ℎ = (푥퐻ℎ) 휌 + (푥푊 ℎ) 휌 퐸(푘) with 휌 > 1 and 퐸(푘) ≥ 1, implying household returns to scale, and 푥푗ℎ the contribution of 푗 to the public good. Adding a time subscript, we write spouse 푗’s utility functions under marriage and divorce respectively as:

푗ℎ 푗ℎ ℎ 푗ℎ 푗ℎ ℎ 푗ℎ 푗ℎ 푗ℎ 푗ℎ 푗ℎ 푗ℎ 푗ℎ 푈푚푎푟푟푖푒푑(푐푡 , 푄푡 ) = 푢 (푐푡 , 푄푡 ) + 휒 and 푈푑푖푣표푟푐푒푑(푐푡 , 푥푡 ) = 푢 (푐푡 , 푥푡 ) with 휒푗ℎ denoting the subjective taste for marriage of each spouse, which may evolve over time.

Assume that taste shocks follow a random walk stochastic process, which captures the persistence in taste for the current marriage:

푗ℎ 푗ℎ 푗ℎ 푗ℎ 푗ℎ 휒푡 = 휒푡−1 + 휖푡 and 휒1 = 휖1

푗ℎ 2 with 휖푡 i.i.d. as 푁(0, 휎 ). We now introduce here the main dimension of heterogeneity between matrilocal and patrilocal ethnic groups. We assume patrilocal women to suffer from a stronger social stigma from divorce in the household they live in. Formally:

푊 푀 푊 푃 Assumption I: 휒1 < 휒1 .

We also assume that men suffer from a lower social stigma than women, irrespective of the ethnic group they belong to:

퐻푀 퐻푃 푊 푀 Assumption II: 휒1 = 휒1 < 휒1 .

123 Chapter 2. Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia

2.3.2 Income Dynamics and Budget Constraints

Assume spouses have a permanent income 푦푗ℎ that may be modiﬁed in each period 푡 by a random shock so that the income dynamic of each spouse follow a random walk process (Voena, 2015):

푗ℎ 푗ℎ 푗ℎ 푗ℎ 푗ℎ 푦푡 = 푦푡−1 + 휁푡 and 푦1 = 휁1

푗ℎ 2 with 휁푡 i.i.d. as 푁(0, 휎 ). The two groups have the same expected total permanent household income.

For both ethnic groups the household budget constraint in each period is as follows:

ℎ ℎ 퐻ℎ 퐻ℎ 푊 ℎ 푊 ℎ 퐻ℎ 푊 ℎ 퐴푡+1 − (1 + 푟푡)퐴푡 + 푐푡 + 푥푡 + 푐푡 + 푥푡 ≤ 푦푡 + 푦푡 .

We discuss the problem of divorce for a community property regime where assets are shared equally between spouses, but we can extend result for context in which title-based and equitable sharing are applied. In a community property regime, assets accumulation is deﬁned jointly between spouses.

Under divorce, the budget constraint is modiﬁed as follows:

푗ℎ 푗ℎ 푗ℎ 푗ℎ 푗ℎ 퐴푡+1 − (1 + 푟푡)퐴푡 + 푐푡 + 푥푡 ≤ 푦푡 .

2.3.3 Problem of Divorce before the Reform

To understand how the reform has modiﬁed access to divorce, we make the following assumption:

Assumption III: Before the reform, the household probability to access divorce, denoted 푑, is equal to 0.

A zero 푑 means that the spouses have no possibility to exit marriage. This approximation ﬁts well the reality of many couples in Indonesia and the very low divorce rate recalled earlier.14 At each period

ℎ 퐻ℎ 퐻ℎ 퐻ℎ 푊 ℎ 푡, the couple’s problem is determined by the state variables 휔푡 = {퐴푡 , 푦푡 , 푦푡 , 휒푡 , 휒푡 } that maximize the following value function:

퐻ℎ 퐻ℎ 퐻ℎ ℎ 퐻ℎ 푊 ℎ 푊 ℎ 푊 ℎ ℎ 푊 ℎ 푉푡(휔푡) = 훾 [푢 (푐푡 , 푄푡 ) + 휒푡 ] + 훾 [푢 (푐푡 , 푄푡 ) + 휒푡 ] + 훽퐸[푉푡+1(휔푡+1)] 14This assumption also implies that women’s position in marriage before the reform does not differ between couples of matrilocal versus patrilocal tradition. Again, this fact is verified empirically (there is little correlation between women’s outcomes and matrilocal ancestry in 2007, i.e. before the reform, as detailed in Table 2.B.2).

124 Chapter 2. Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia where 훾퐻ℎ and 훾푊 ℎ are the Pareto weights of husband and wife respectively, which are determined by social norms and do not change with outside options.

2.3.4 Problem of Divorce after the Reform

We assume that the reform increases the accessibility of divorce equally for matrilocal and patrilocal women. Formally:

Assumption IV: After reform, the household probability to access divorce 푑 is randomly drawn from the following probability distribution:

⎧ ⎪ ⎨⎪1 with prob. 푝 푑ℎ = ⎪ ⎩⎪0 with prob. 1 − 푝.

When 푑 = 1, spouses have the possibility to exit marriage. Since we are in a limited commitment set-up, we assume that divorce can be initiated unilaterally by each spouse. Thus, in each period, the household will stay in the marital union if there exists an allocation that makes both spouses better- off than the divorce allocation. In this case, the within-period Pareto weight (bargaining power) also

ℎ 퐻ℎ 푊 ℎ 퐻ℎ 푊 ℎ 퐻ℎ 푊 ℎ enters the vector of state variables that now becomes 휔푡 = {퐴푡 , 푦푡 , 푦푡 , 휒푡 , 휒푡 , 훾˜푡 , 훾˜푡 } 푗ℎ 푗ℎ 푗ℎ 푗ℎ where for each spouse 훾˜푡+1 = 훾˜푡 + 휇푡 . The parameter 휇푡 is the Lagrange multiplier associated with the following participation constraint:

푗ℎ 푗ℎ ℎ 푗ℎ 푗ℎ 푈푚푎푟푟푖푒푑(푐푡 , 푄푡 ) + 훽퐸[푉푡+1(휔푡+1)|(휔푡)] ≥ 푉퐷푖푣표푟푐푒푑(휔푡)

As shown by Voena(2015), this limited commitment set-up implies that:

푊 ℎ 푊 ℎ ℎ 휕푈푚푎푟푟푖푒푑(푐푡 ,푄푡 ) 푊 ℎ 퐻ℎ 퐻ℎ 휕푐푡 훾˜푡 + 휇푡 퐻ℎ 퐻ℎ ℎ = 푊 ℎ 푊 ℎ 휕푈푚푎푟푟푖푒푑(푐푡 ,푄푡 ) 훾˜ + 휇 퐻ℎ 푡 푡 휕푐푡 Combining assumptions I-IV, we can formulate the following predictions:

Prediction 1. For households that have access to divorce after the reform, matrilocal women have in expectation a higher probability of divorcing than patrilocal women.

푊 푀 Prediction 2. For households that have access to divorce after the reform, 퐸(훾˜푡 ) > 푊 푃 퐸(훾˜푡 ), i.e. women from matrilocal origins have in expectation a higher bargaining power in marriage than patrilocal women.

125 Chapter 2. Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia

Proofs are provided in Appendix D. These two predictions are precisely what we aim to test with the difference-in-difference analysis that follows.

2.4 Empirical Approach

2.4.1 Data

IFLS Data. The empirical analysis draws on data from the Indonesia Familiy Life Survey (IFLS).

It is particularly well-suited for our study, as it contains extensive socioeconomic data at the individual level (including information on individuals’ ethnicity, marital history, health status and subjective well-being) and at household level (including decision-making questions, household composition and economy). IFLS samples also contain village-level information, notably the ethnic composition and prevalent kinship norms (including inheritance and post-marital residence norms), as provided by

Adat experts or community leaders. Even though repeated cross-sections would be sufﬁcient for our double difference analysis, the panel dimension is appreciable as it allows us to control for household

ﬁxed effects. Moreover, the IFLS beneﬁts from an exceptionally low attrition rate so the sample remains representative at every wave.15

Selection. We select non-polygamous households and exclude ‘mixed’ couples, i.e. when spouses originate from two different ethnic groups with different post-marital residence norms. We focus on the 2007 and 2014 waves, two years surrounding the Access-to-Justice reforms (2008-10). Our main analysis considers women’s well-being and empowerment outcomes within stable marriages (i.e. not remarried) over 2007-2014.16 To check the parallel trend assumption over the pre-reform period

(2000-2007), we will use couples married from 2000 to 2014 with the same partner (or, alternatively, couples married in both 2000 and 2007 with the same partner). We will also use a double difference approach to characterize the effect of the reform on women’s divorce rates. In that case, we will use different alternative samples as described hereafter.

15The IFLS is based on an initial sample representing about 83% of the Indonesian population living in 13 of the 27 Indonesian provinces in 1993. Extensive efforts were provided to track respondants when collecting data in each of the five waves (1993, 1997, 2000, 2007 and 2014) to reach a recontact rate of 92% in the lastwave (Strauss, Witoelar, and Sikoki, 2016). 16We decide to focus on stable couples over 2007-2014 rather than stable couples over 2000-2014 since working with this latter subsample implies working with a sample that is about 44% smaller and raises stronger sample selection issues (e.g. aging sample, longer-lasting marriages, etc.). Note however that we get similar conclusions on the differential effects of the reforms when we work with this subsample (not reported).

126 Chapter 2. Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia

Outcome Variables. We consider a series of outcomes pertaining to a women’s empowerment and to her or a children’s well-being, as reported in the IFLS. Well-being variables include a dummy indicating whether a woman experienced at least one morbidity symptom in the last 4 weeks preceding the survey, the woman’s number of living births, the adequacy of her own and her children’s standard of living and food consumption on 1-3 scales, and the value of assets owned by the wife (in thousands of rupiah).17 Empowerment variables are dummies indicating whether the wife and/or her potential relatives have the ﬁnal say regarding key dimensions of household choices including contraception and large household expenditures.18

Treatment Intensity: Ethnicities’ Traditional Post-marital Residence. As previously discussed, we focus on ethnic heterogeneity in terms of traditional residence norms. The traditional norm of each ethnicity is not explicitly reported but can be proxied following a simple methodology, as already suggested by Buttenheim and Nobles(2009). The IFLS contains information on individual ethnicity and we can categorize villages according to their main ethnicity. For each group of villages of the same prevailing ethnicity, we observe the distribution of Adat experts’ answer to the question about “where the newly married couple lives after wedding according to the traditional law”, as reported in Table 2.A.1. To attribute to each woman the traditional residence norm of her ethnicity, we retain the modal answer. Note that it is systematically matrilocality or patrilocality (rather than neolocality or ambilocality), so that an ethnic group’s traditional norm will be a binary information as indicated in the last column. We obtain a proportion of about 83% (17%) individuals with a matrilocal

(patrilocal) ethnic heritage. As reported in Table 2.A.2, we ﬁnd that the traditional residence norm is still a very signiﬁcant predictor of a couple’s actual household composition in 2014.19

Descriptive Statistics and Raw Difference-in-Differences. Descriptive statistics for couples from matrilocal or patrilocal ethnic traditions are reported in Table 2.A.3 for both 2007 and

17Other child outcomes could be considered, for instance the gender of the last child as a proxy for son preferences (and the stopping rule when couple obtain the desired number of boys, see Jayachandran and Kuziemko 2011). 18These dimensions seem relevant and have been used in previous studies on muslim countries – see for instance Sadania(2016) or Lépine and Strobl(2013) for Egypt and Senegal respectively. Other aspects are deemed less relevant, such as decisions upon daily purchase and cooking, since they may reflect delegation of responsibility rather than women’s genuine autonomy (Baland, Boltz, Catherine, Seleck, and Ziparo, 2020). 19A possible alternative strategy would consist in instrumenting the actual residence choice by the traditional norm. Table 2.A.2 shows that a first-stage estimation of traditional matrilocality on presence of wife’s relatives in the household (plus standard controls) yields a F-statistic well above 10. We do not follow this path because, as explained in sections 2.1 and 2.2, we want to use the traditional norm directly, for it carries more information than actual residence choices on gender rights and roles within an ethnic group.

127 Chapter 2. Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia

2014. The upper panel describes standard socio-demographic characteristics (the rest of the table focuses on additional controls used in the robustness checks analysis and commented later). This set of characteristics, denoted 푋푖푡 hereafter, is important as it will be used as controls in the estimations. It includes dummies for being a university graduate, currently working, living in rural areas, being muslim as well as for age groups (using 5-year steps) to capture life cycle effects. We see that women of ethnic matrilocal customs tend to work less and be more often urban and muslim than their patrilocal counterparts. These differences seem to be constant over time. The last column reports the time difference of the matrilocal-patrilocal difference, i.e. a raw difference-in-difference (DD) calculations on these characteristics. It conﬁrms that there is no signiﬁcant change over time, which is reassuring on the absence of confounding factors. For instance, the reform could affect muslims more than non- muslims, which would be what we capture by confronting matrilocal and patrilocal groups. While it does not seem to be the case, our estimations will control for the whole set of characteristics and their differentiated effect over time.

Regarding outcomes, statistics are reported in Table 2.A.4. We distinguish well-being variables from

final say measures on key decisions. There is little difference between matrilocal and patrilocal groups in the pre-reform period (this is the case for 8 out of 9 variables and a F-test cannot reject the equality of both sets of mean values). Interestingly, strong differences emerge in the end period, both in terms of well-being and empowerment. The raw DD actually indicates that relative to the pre-reform period, the situation of women of matrilocal tradition significantly improves: less morbidity symptoms, more control over fertility (both number of birth and final say on contraception), better living conditions and nutrition for them and their children, more asset accumulation, more control over large household expenditure. DD estimations will refine these raw calculations but both sets of results are consistent

(we will verify that raw DD and actual DD estimates are in line both in terms of signiﬁcance and orders of magnitude). We also carry out simple estimations of these outcomes on a dummy for belonging to an ethnicity of matrilocal tradition (and additional controls as described in the double difference approach), for each year separately. Results in Table 2.B.2 show that there is not much difference between ethnic groups in 2007 but very strong differences in 2014, after the reforms.

2.4.2 Empirical Approach

Difference-in-Difference Estimations. We denote 푦푖 the outcome (female well-being / empowerment measures), for a woman in household 푖 observed at time 푡. The treatment variable,

128 Chapter 2. Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia

푀푎푡푟푖푙표푐푎푙푖, is equal to 1 (0) if the woman’s ethnic group is traditionally matrilocal (patrilocal). Importantly, nothing precludes women of patrilocal tradition to be affected by the reform. Hence, what our difference-in-difference approach captures is a potential difference in the intensity of treatment between the two groups. Nonetheless, we will comment on ﬁrst-difference effects of the reform regarding women of patrilocal tradition, which give very suggestive indication of the potential direction of the reform overall. Given that we pool only two years, time effects are simply denoted by

푃 표푠푡푡, which is equal to 1 for the period following the Access-to-Justice reform (year 2014) and 0 for the base period (year 2007). The estimation conducted on pooled years is described by the following equation:

푦푖푡 = 훼 + 훽푃 표푠푡푡 × 푀푎푡푟푖푙표푐푎푙푖 + 훾푃 표푠푡푡 + 훿푋푖푡 + 휂푃 표푠푡푡 × 푋푖푡 + 휑푖 + 휀푖푡. (2.1)

The coefﬁcient 훾 on 푃 표푠푡푡 captures the time trend in the outcome, which includes the effects of the reforms common to all the Indonesian ethnicities, as identiﬁed on the households of patrilocal tradition

(the ‘untreated’). The coefficient 훽 on the interaction term is the difference-in-difference estimator, representing the extra effect of belonging to an ethnicity of matrilocal tradition (the ‘treated’) once the reforms are in place. Household fixed effects 휑푖 implicitly pick the average time-invariant difference between ethnic groups (and notably the difference between matrilocal and patrilocal traditions, i.e. 푀푎푡푟푖푙표푐푎푙푖 is absorbed by the fixed effects). Covariates 푋푖푡 may improve the precision of the model but also control for the difference in time-varying observables between matrilocal and patrilocal ethnic groups, as previously discussed. Standard errors are clustered at the level of the women’s village of origin to correct for potential geographical correlation in error terms.20

Identification Issues. First, in the context of a DD analysis, treated and control groups are not randomly chosen and may be very different. We have seen that women of matrilocal ethnic groups work signiﬁcantly less, are more often muslim and are more urban. Hence, the measured effect might be due to different time trends in these variables – for instance differentiated responses to the reform between muslim and non-muslim – rather than between traditional residence norms.

We have partly ruled out this possibility when presenting the descriptive statistics for 푋푖푡. However,

20Villages of origin are where the woman’s household lived in the first wave of the IFLS (1993). It seems reasonable to use such a variable for it is time-invariant in our estimations and presumably closer to the time of marriage.

129 Chapter 2. Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia

we also include interactions 푃 표푠푡푡 × 푋푖푡 to completely account for (possibly confounding) time trends pertaining to speciﬁc religion or geographical groups in the DD estimation. Second, another usual concern is the potential presence of unobservables, which would affect the outcome trends of matrilocal and patrilocal ethnic groups differently. A minimum requirement in this respect will consist in checking whether the outcomes of the two groups show parallel trend prior to the reforms under study, namely between 2000 and 2007. Third, we have extensively checked that throughout the

2007-2014 period, there was no major (potentially confounding) policy or social change that could have affected women’s empowerment in a differentiated way between couples of patrilocal versus matrilocal origins. Fourth, we consider the possibility of endogeneous policy changes (Bertrand,

Duflo, and Mullainathan, 2004, Besley and Case, 2000). This would be the case if the series of reforms was triggered by a pre-existing rise in power by women who could best benefit from policy changes, namely women of matrilocal ethnicities. Several points mitigate this concen: (i) the Access- to-Justice strategy was implemented at a national level, not targeting any particular social or ethnic groups; (ii) it was prompted by international influence, notably that of the Family Court of Australia,

AusAID and other stakeholds, as previously described; (iii) common trends in the pre-reform period will guarantee that there is a priori no differentiated trends in empowerment between ethnic groups that would eventually lead to the reforms.

2.5 Results

2.5.1 Effect of the Reforms on Divorce

A preliminary set of results pertains to the effect of the reform on divorce. We have conjectured that an enhanced access to justice may increase the mere possibility of divorce and that it would occur relatively more among ethnicities of matrilocal tradition. We check this prediction with estimations reported in Table 2.1. We ﬁrst apply a difference-in-difference approach to women’s marital status for the pooled years 2007 and 2014, using the same speciﬁcation as outlined in the previous section.

To assess the mere transitions from marriage to divorce in both years, we first consider a sample of women previously married (model 1). Alternatively, we keep the whole sample of women observed in both 2007 and 2014 but selecting only those married or divorced, i.e. excluding the singles and widowed (model 2). We also combine both selection criteria (model 3) or keep the first one while considering a status as divorced or separated (model 4). All the models leads to a significant effect

130 Chapter 2. Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia of the reform on the probability of divorce for women of matrilocal ethnicities compared to patrilocal counterparts. Relative to the pre-reform divorce rate in the control group, the effect represents an increase of 40%-66% in divorce rates across the different models. The coefﬁcient on Post shows no

ﬁrst-difference effect of the reform for the control group, suggesting that in the absence of counteract- ing time changes, the burden of divorce must have remained heavy for women of patrilocal heritage.

We also estimate a simple difference model for the year 2014, using the subsample of women married in 2007. We consider both outcomes: divorce (model 5) or divorce/separation (model 6). Being from a matrilocal tradition leads again to a signiﬁcant relative impact of the reforms and of a very similar magnitude compared to DD estimations. In Appendix Table 2.C.1, we test the parallel trend assumption for the DD or a direct change in marital breakdown for the simple difference approach.

We use a placebo sample pooling years 2000-2007 and the very same types of speciﬁcations. In all of them, placebo estimates are not statistically different from zero, i.e. there is no sign of speciﬁc trends in divorce among ethnic groups of matrilocal tradition before the reforms.

2.5.2 Main Results: Effect of the Reforms on Women’s Outcomes

Previous results establish that by fostering access to justice, the reform has signiﬁcantly increased divorce rates among women in matrilocal ethnicities, verifying the ﬁrst theoretical prediction. The second prediction entails a relative gain in bargaining power among women of matrilocal groups

(section 3). This result stems both from a more frequent renegotiation among matrilocal ethnic groups and because of a selection effect (matrilocals divorce more so that those who stay in marriage have obtained favorable renegotiations, cf. appendix D).

We check this prediction formally with the DD approach laid out in equation (1). Our baseline estimations are conducted on stable marriages (spouses observed married in both 2007 and 2014) but we also provide results for a broader selection hereafter. Baseline results are presented in Table 2.2. For most of the outcomes, the coefficient on Post is not statistically different from zero. It suggests that, if no other forces affect women’s position within the private sphere over time, women of patrilocal ethnicities do not benefit much from the reform. In contrast, the DD estimates confirm that matrilocal women experience a relative improvement in their living conditions. As seen in descriptive statistics, their morbidity symptoms decrease, the number of births is reduced, women’s and children’s standard of living and food consumption improve, the wives’ asset accumulation increases, and the final say over contraception and large expenditure is in progress. The magnitude of the effect, relative to the

131 Chapter 2. Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia pre-reform control group outcome, is substantial. In absolute values, it ranges from an increase of

7-10% in mothers and children’s standard of living and nutrition up to a near doubling of the ﬁnal say variables (+85% and 104% on decisions upon contraception and large household expenditures respectively).21

Table 2.1 – Effect of Legal Reforms on Women’s Divorce Probability

Divorced Divorced Dep. Var. Divorced Divorced Divorced or Divorced or Separated Separated Estimator Diff-in-Diff Simple Diff.

Excluding Excluding singles & Married singles Married Married Married Samples widowed, before 2007 and before 2007 in 2007 in 2007 married widowed before 2007 (1) (2) (3) (4) (5) (6) Post 0.0225 0.00575 0.0164 0.000335 (0.0231) (0.00907) (0.0160) (0.0297) Post × Matrilocal 0.0103** 0.0125*** 0.0107** 0.0122** (0.00434) (0.00472) (0.00486) (0.00512) Matrilocal 0.00879*** 0.00928*** (0.00287) (0.00336) Rel. effect 66.0% 40.2% 63.3% 53.0% 56.3 % 40.3%

Observations 12,752 17,390 10,892 12,752 8,801 8,801 R-squared 0.007 0.008 0.008 0.006 0.005 0.005 Clusters 318 319 318 318 319 319 Individual FE Yes Yes Yes Yes No No Controls Yes Yes Yes Yes Yes Yes Post × Controls Yes Yes Yes Yes No No Linear estimations of women’s divorce status (dummy for divorced, or divorced/separated). We apply the difference-in-difference approach to a selection of women observed in both 2007 and 2014, who weremarried in 1997 and/or 2000 (columns 1, 3 and 4); and a selection of women being married, divorced or separated in 2007 and 2014 (columns 2 and 3). For them, Post is equal to 1 for observations in 2014 (post-reform) and 0 in 2007 (pre-reform). We also estimate the potential increase in divorce using women observed in 2014 who were married in 2007 (columns 5 and 6). Matrilocal is a dummy indicating whether an individual belongs to a traditionally matrilocal ethnic group. Estimations include individual FE (absorbing Matrilocal and muslim - column 1-4), time-varying controls (women’s characteristics: university graduate, currently working, living in rural areas and age group dummies using 5-year steps), + a muslim dummy in columns 5 and 6, and interactions between Post and controls (including Post interacted with a muslim dummy) in columns 1-4 as indicated. Standard errors are reported in brackets and clustered at village of origin level. Significance levels: * p<0.10, ** p<0.05, *** p<0.01.

21Relative effects tend to be large given that we start from very low rates of female say before the reform. For instance, decisions upon large expenditure are broadly in the hands of men. As indicated in Table 2.A.4, women of matrilocal (patrilocal) ethnicities have the final say in this domain in only 6.4% (5.3%) of thecases in 2007. DD results indicate a relative increase of 5.5 percentage points for the matrilocals.

132 Chapter 2. Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia Large expenditures Contraception Wife’s assets value Children’s food conso. Children’s std. of living Food consumption Women’s and Child’s Well-Being Women’s Empowerment of living Standard Number of births Effect of Legal Reforms on Women’s Well-Being Empowerment and (Stable Couples, 2007-14) (1) (2) (3) (4) (5) (6) (7) (8) (9) -12.6% -7.6% 9.7% 9.3% 7.6% 6.9% 47.1% 85.3% 103.6% (0.529) (3.411) (0.505) (0.401) (0.431) (0.565) (33,483) (0.239) (0.140) (0.0319) (0.0915) (0.0536) (0.0507) (0.0541) (0.0500) (5,286) (0.0333) (0.0159) Morbidity symptoms Table 2.2 – Matrilocal -0.0887*** -0.233** 0.191*** 0.188*** 0.152***Controls 0.140*** Yes 10,689** 0.163*** Yes 0.0549*** Yes Yes Yes Yes Yes Yes Yes × × Post Relative effect ObservationsR-squaredClusters 11,840 11,490 0.047 11,546 318 0.348 11,548 0.043 318 0.059 6,730 318 6,728 0.039 318 11,870 0.073 316 10,884 0.083 316 10,884 0.065 318 0.088 317 317 Post 1.327** 2.577Household FEControlsPost 0.467 Yes -0.563 Yes Yes -0.0617 Yes Yes 0.267 Yes Yes 10,203 Yes Yes 0.321 Yes Yes 0.0863 Yes Yes Yes Yes Yes Yes Yes Difference-in-difference estimations of well-being empowerment and indicators on a sample of stable couples surveyed in both and 2014. Post 2007 is equal to 1matrilocal for ethnic 2014 group. (post-reform)characteristics: and All university 0 estimations graduate, for include currently 2007controls working, household (including (pre-reform). living FE Post in Matrilocal interacted (absorbing ruralsymptom with is Matrilocal areas in a a and and the muslim dummy age dummy). muslim), last indicating groupof 4 time-varying Well-being whether dummies her weeks outcomes: controls an using standard preceding (women’s dummy individual 5-year of theof and indicating steps) belongs living survey whether living and husband’s and to (‘Morbidity a interactions and food a Symptoms’), between woman food the traditionally consumptionthousands Post experienced consumption woman’s on of and at number on a rupiah). least of 1-3 1-3 one Empowermentdoes living scale scales outcomes: morbidity births not (‘Standard (‘Children’s dummies (‘Number have of std. any of indicating living’calculated say) Births’), of whether and in regarding the the living’ ‘Food % key adequacy consumption’), wife and of dimensions thelevel. and/or ‘Children’s mean of adequacy her food Significance outcome household of potential levels: for conso.’), choices her and relatives patrilocal including children’s have the contraception group * p<0.10, standard the value and in final of large 2007 ** p<0.05,say assets household (pre-reform). (while owned expenditures *** p<0.01. Standard by . errors husband the the are The wife reported relative (in effect in is brackets and clustered at village of origin

133 Chapter 2. Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia

These magnitudes are very much in line with the raw DD calculations presented above. For instance, the raw DD for the women’s living standard of (.150) is of a similar order as the DD estimate (.190).

It is also similar to the difference in the matrilocal effect in 2014 compared to that effect in 2007 in the cross-sectional estimations of Table 2.B.2 (-.084-.107=.191). The latter comparison boils down to a DD on stable couples 2007-2014 but ignoring the panel dimension and household ﬁxed effects.

For a robustness check and to verify that our results are not associated with the speciﬁc focus on stable couples, we also run DD estimation on repeated cross-sections for 2007 and 2014 whereby the only selection criteria is to focus on married women. As can be seen in Table 2.B.3, results are again similar to the ﬁxed-effect DD estimates (for instance, .203 in our example on women’s living standards).22

2.5.3 Robustness Checks, Heterogeneity and Interpretations

Common Trends. For both well-being and empowerment measures, common trends are verified in placebo estimations reported in Table 2.C.2. With one exception, placebo estimates obtained over the pre-reform period 2000-2007 are insignificant. We also conduct an extensive series of sensitivity checks. We focus hereafter on two robustness analyses on the sensitivity to model specification and to the definitions of empowerment outcomes.

Sensitivity Analysis: Alternative Pathways. We check if our results are sensitivite to the set of controls included in the model. This is important because some variables, such as being muslim or rural, are highly correlated with traditional matrilocality and may be seen as the pathway through which the reform impacts upon women’s conditions. In our descriptive statistic analysis, we have partly addressed this concern, showing that the main characteristics have not changed differently between matrilocal and patrilocal ethnic groups over time. Nonetheless, we check whether our DD estimates are sensitive to the inclusion of 푃 표푠푡푡 × 푋푖푡 terms. Table 2.C.3 reports the relative effects from alternative DD estimations. The ﬁrst row reproduces our benchmark estimates. The second row

22Note that in additional, unreported estimations, we also test the effect of the reform on female labor market participation and find no differential effect among women of matrilocal ancestry. Admittedly, it is unclearwhich effect to expect. An increased risk of divorce may benefit to women in stable marriage and increase theirleisure time – this is the finding of Voena(2015) for women in US states that shift to unilateral divorce, when property is divided equally. Alternatively, the risk of divorce may also push them to take up a job as a security in case of divorce, which is particularly true when female participation is low (see Bargain, González, Keane, and Özcan 2012 in the context of divorce legalization). In the case of Indonesia, the pre-reform participation rate of about 50% was not particularly low compared to other muslim countries. Moreover, work is not necessarily seen as an insurance against poverty in case of divorce, since women’s activities are typically informal and bring only complementary income (Verick, 2018).

134 Chapter 2. Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia

shows the series of estimates from a model where 푃 표푠푡푡 × 푋푖푡 are fully withdrawn. Relative effects change quantitatively for some of the outcomes but the order of magnitude is broadly preserved and our conclusions are unchanged. Women of matrilocal ethnic groups are more often muslim and urban than their patrilocal counterparts. Removing only 푃 표푠푡푡 × 푅푢푟푎푙푖 hardly affects baseline estimates

(3rd row). Excluding 푃 표푠푡푡 × 푀푢푠푙푖푚푖 in the 4th row leads to some variation (as in the second row) but does not fundamentally change our results – the baseline effects of the reform are not driven by different time trends between muslims and non-muslims.

In addition, we also include new controls to the model. Javanese represent the main ethnic group and account for 56.5% of matrilocal individuals in our sample. To check whether specific time trends in this group may drive our results, we include 푃 표푠푡푡 × 퐽푎푣푎푛푒푠푒푖 in the model: results are barely changed (5th row). Alternatively, we run our baseline specification on a sample excluding the javanese (6th row): despite a sharp reduction in sample size, we find very similar results compared to the baseline. Couples of mixed ethnicity being more often represented among matrilocal couples, we explicitly control for 푃 표푠푡푡 × 푀푖푥푒푑 퐸푡ℎ푛푖푐푖푡푦푖 (7th row) and find similar results. Another threat to our identification would be a differential exposure to unobserved shocks between matrilocal and patrilocal group. While we are not aware of any shock that would have differently affected matrilocal or patrilocal individuals, we additionally control for 푁푎푡푢푟푎푙 퐷푖푠푎푠푡푒푟푖푡, a dummy indicating whether an individual lives in a village having experienced a natural disaster in the 5 years preceding the sur-

23 vey, and 푃 표푠푡푡 × 푁푎푡푢푟푎푙 퐷푖푠푎푠푡푒푟푖푡. Our results remain similar (8th row). Finally, geographical polarization of certain ethnic groups could also be envisaged as a potential alternative explanation to our results. We have previously checked that ethnicities of matrilocal tradition do not necessarily live closer to the capital of the district, where religious and civil courts are located. Moreover, we have also documented the fact that part of the reform consisted in increasing the frequency of ‘circuit courts’ visiting people in remote areas. Nonetheless, we also provide additional estimations where we now control for 푃 표푠푡푡 × 퐶푙표푠푒푖 with the 퐶푙표푠푒푖 dummy indicating whether the household lives in the two ﬁrst tertiles of the distance to the district capital. Results are reported in the upper panel of Table 2.3, which is dedicated to a complete analysis on the role of distance. This check shows no major difference with the baseline estimates.

23We take advantage of information recorded in IFLS on whether villages experienced a natural disaster in the 5 years preceding the survey. Unfortunately, this information is only provided for villages of origin (those surveyed since IFLS 1), explaining a decrease in our sample size when adding these controls. It is worth noting that looking at descriptive statistics, we do not find significant differences between matrilocal and patrilocal, neither differential time trends between these two groups (see Table 2.A.3).

135 Chapter 2. Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia

Furthermore, as exposed in subsection 2.2.1, matrilocality may be correlated with other cultural traits that may affect gender roles. While time-invariant unobservables such as ethnic groups’ traditional norms are explicitly controlled for through the inclusion of fixed effects in our regressions, such cultural traits may have time specific effects that need to be further controlled for. Matrilineality or the practice of bride price may notably constitute plausible alternative pathways of the effect of the reforms. On one hand, matrilineality may strengthen responsiveness to the reforms, as women in matrilineal structures benefit from higher marriage outside option through notably a higher support from their relatives and a more central social position in the kinship structures (Lowes, 2018, Loper, 2020).

Unfortunately, the Adat questionnaire in IFLS does not provide information on descent rules. We are therefore not able to formally disentangle the effect of matrilocality from the effect of matrilineality.24

On the other hand, the practice of bride price may mean that the wife’s family has to pay the bride price back following a divorce. This would weaken responsiveness to the reforms. Using information from the Adat questionnaire in IFLS, we do not ﬁnd any signiﬁcant willage-level correlation between traditional matrilocality and traditional practice of bride price (not reported).25 Further, our results remain unchanged when we additionally control for 푃 표푠푡푡 × 퐵푟푖푑푒 푃 푟푖푐푒푖 (9th row).

Sensitivity Analysis on Empowerment Measures. Two outcomes correspond to final say measures pertaining to key decisions that may not be associated to delegation and rather interpreted as genuine autonomy for women. We provide some sensitivity checks regarding their definition and measure. For both types of decisions, Table 2.C.6 first reproduces the baseline estimates (columns 1 and 5). They are obtained for final say outcomes defined as dummies equal to 1 if the wife and/or her relatives decide (while the husband does not have any say), 0 otherwise. We then show results corresponding to a more restrictive definition whereby empowerment means that the wife decides alone

(columns 2 and 6). This is an interesting check given the fact that our main source of heterogeneity pertains to traditions regarding coresidence with spouses’ relatives. We ﬁnd hardly any change compared to the initial estimates. Next, we address the fact that ﬁnal say outcomes are recorded by the husband. It may vary compared to the wife’s answer about who decides on each item in the household.

The heterogeneity can go both ways. Our baseline may understate women’s decision power if men do not recognize when their wife has the say. Inversely, empowerment may be more ﬁrmly established when the husband admits that she is the decision maker in a given domain. We simply reestimate the

24Note however that matrilocality and matrilineality are highly correlated, especially in the context of Indonesia, where Bau(2016) finds a significant correlation of .996 using alternative datasets. 25In fact, we find that about 86% of both matrilocal and patrilocal villages traditionally practice bride price.

136 Chapter 2. Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia model for outcomes based on the wife’s answers (columns 3 and 7). Estimates are reduced by around a third but remain very signiﬁcant. The last set of results combine both variants (columns 4 and 8) and leads to the same conclusions.

Sensitivity Analysis on Treatment Measures. So far, following Buttenheim and Nobles

(2009) we have used the modal answer of villages of same main ethnicity to uncover ethnic groups’ post-marriage residence norm to define our treatment group. Doing so, we have ended up with a binary variation between traditional matrilocality and patrilocality. While convenient for the interpretation of our results, such measure may hide some heterogeneity. Indeed, looking at Table 2.A.1, traditional practice of matrilocality seems a bit more nuanced for some ethnic groups. Therefore, as a further sensitivity exercise, we use the proportion of villages of same main ethnic group practicing matrilocality as a measure of treatment, that we call “Matrilocal Intensity”.26 As reported in Table 2.C.4, using this alternative measure of treatment we draw similar conclusions regarding the differential effects of the reforms. Another sensitivity exercise pertains to the issue of mixed couple. So far we have excluded couples with spouses originating from ethnicities with different post-marital residence norms.27 The main motive for doing so was to facilitate the interpretation of our results. Nevertheless, we check that our results are robust to the inclusion of mixed couples in our sample, while controlling for their time-specific effect.28 Reassuringly, Table 2.C.5 suggests that our results are not confounded by this sample selection.29 Further, results are qualitatively similar when we alternatively assign the husband’s ethnic group’s post-marital residence norm to define our treatment.30

Heterogeneity We ﬁnally derive heterogeneous effects by interacting the treatment variable with key characteristics (such as religion). Unreported estimations show no particular pattern in general but one dimension comes out quite interestingly: the distance to the capital of the local district. Re- sults are reported in Table 2.3. We have already commented on the upper part where we check that our estimates are not affected by potentially different time trends for those living with an easier access to religious/civil courts, which would confound our interpretation if women of matrilocal traditions

26As an example, we find in Table 2.A.1 that about 64% of Javanese villages report traditional matrilocality. Therefore, Matrilocal Intensity = 0.64 for Javanese individuals. 27Note that we show in a robustness check exercise in Table 2.C.3 that our results are unaffected when controlling for couples with spouses originating from different ethnicities of same post-marital residence norms. 28The time-invariant effect of being in a mixed couple is already captured by fixed effects in our regressions. 29Sample size is increased by about 4% when including mixed couples. 30In our sample, about 98% of matrilocal wives have a matrilocal husband and about 89% of patrilocal wives have a patrilocal husband.

137 Chapter 2. Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia systematically lived nearer to the district capital.31 The lower panel of this table suggests heterogeneous effects of the reform. We interact 푃 표푠푡 × 푀푎푡푟푖푙표푐푎푙 with ‘close’ and ‘far’ dummies defined as belonging to a village located in the two first and in the third tertiles of distance to the district capital respectively. For almost all outcomes (except wife’s assets value and final say on contraception), the effect of the reform is larger for those living further away. Reported t-tests do not reject equality between distance groups in general (it does for children’s food consumption) but this pattern is intriguing. In fact, if we take into account structural differences between groups – for instance the fact that women’s decision power on contraception is smaller in far away villages – then we yield contrasted relative effects that are systematically larger (except for wife’s assets value) for women living far from administrative centers.32 These results are purely suggestive but may indicate that the set of reforms tends to benefit poor villager women more, possibly as an effect of circuit courts, which especially help those in remote regions, or the reduced cost of legal procedures through the court fee-waivers.

31Note that the sample is reduced by around 20% due to the fact that distance information is available only for those who lives in the village of origin. Results in the upper panel of Table 2.3 show the robustness of our results not only to the control for 푃 표푠푡 × 퐶푙표푠푒 but also to the use of this smaller sample. 32Relative effects are calculated as the DD estimate relative to mean pre-reform control group outcomes while distinguishing according to distance to district capital.

138 Chapter 2. Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia Large expenditures Contraception Wife’s assets value Children’s food conso. Children’s std. of living Food consumption Women’s and Child’s Well-Being Women’s Empowerment of living Standard Close) × Heterogenous Effect of the Reform according to the Distance to District Capital Number of births (1) (2) (3) (4) (5) (6) (7) (8) (9) -11.8% -7.4% 10.0% 10.0% 9.0% 6.9% 54.7% 93.9% 139.1% (0.0400) (0.130)(0.0482) (0.0662) (0.175) (0.0617) (0.102) (0.0862) (0.0762) (0.0807) (0.0704) (0.0862) (7,840) (5,375) (0.0394) (0.0593) (0.0218) (0.0234) (0.0335) (0.106) (0.0631) (0.0570) (0.0646) (0.0646) (5,738) (0.0365) (0.0169) Morbidity symptoms Table 2.3 – Close -0.0752*Far -0.226* -0.0924* 0.170** -0.279 0.195*** 0.236** 0.144* 0.210** 0.237*** 0.0734 0.239*** 16,119** 4,093 0.203*** 0.0507** 0.161*** 0.0763*** × × Matrilocal -0.0821** -0.248** 0.197*** 0.201***Matri. Matri. 0.181*** 0.140** 11,265* 0.186*** 0.0612*** ControlsClose Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes × × × × × Post Relative effect ObservationsR-squaredClusters 8,564 0.045 8,228 314 0.304 8,310 0.046 314 8,310 0.054 314 4,444 0.054 314 4,444 0.069 302 8,566 0.086 7,910 302 0.073 7,910 314 0.096 314 314 Average Effect (Controlling for Post Heterogenous Effect Post Post Relative effect: CloseFarObservationsR-squaredClustersT-Test Equal. (p-val.)Household FEControls -10.1% 8,564 0.765Post -14.5% -6.8% 0.045 8,228 0.806 -8.2% 314 8.6% Yes 0.304 8,310 0.549 12% Yes 0.046 314 Yes 9.5% 8,310 0.872 10.5% Yes 0.054 Yes 314 7.1% 4,444 0.310 11.8% Yes 0.054 Yes 314 3.6% 4,444 0.086 11.8% Yes 0.071 Yes 61.1% 302 8,566 0.139 30% 0.086 Yes 87.9% Yes 7,910 302 0.513 101.3% 0.074 Yes 126.2% Yes 159.7% 7,910 314 0.399 0.097 Yes Yes 314 Yes Yes 314 Yes Post DID estimations on a sample of stable couples surveyed in both 2007 and 2014. Post is equal to 1 for 2014 (post-reform) and 0 for 2007 (pre-reform). Outcomes and ‘Matrilocal’ are defined in Tabledummy 2.2 . indicating Estimations individuals include livingfor household far patrilocal FE to group (absorbing the in matrilocal districtorigin 2007 and capital level. (pre-reform) (i.e. muslim Significance in dummies), above levels: eachcapital controls the group (p-value). defined 2nd of * p<0.10, in tertile distance of to ** p<0.05, distance). district capital. The *** relative Standard p<0.01. effect errors T-test areis calculated reported treatment of equal in bracketseffects and in between clustered % of individuals at mean outcome village of that are close and far from district Table 2.2 and Post interacted with ‘Close’, a dummy indicating individuals living close to the district capital (i.e. below the 2nd tertile of distance). ‘Far’ is a

139 Chapter 2. Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia

2.6 Conclusions

Social scientists increasingly recognize the role played by traditional norms in shaping individual behavior and affecting economic development (Bank, 2015). Very recent studies actually show that customs and social norms can strongly mitigate the impact of development programs aimed at supporting disadvantaged groups such as women. This paper contributes to this nascent literature by focusing on legal reforms, which may also promote access to justice and embolden women in the exercise of their rights. We examine such a reform in the context of Indonesia, where ethnic heterogeneity in terms of gender roles transpires through the type of traditional residence norms after marriage, namely matrilocality or patrilocality. We document a strong global correlation between ancestral matrilocality and contemporaneous perceptions about gender role, especially in low-divorce regions such as Indonesia. For this country, we test whether deeply rooted norms also affect the exposure to the access-to-justice reform. We ﬁnd that women originating from customary matrilocal ethnic groups tend to divorce more after the reform, relative to those from patrilocal tradition. It also appears that a subsequent renegotiation takes place in stable marriages so that women from matrilocal groups experience a signiﬁcant increase in well-being and empowerment.33

Our study sheds some light on how cultural norms may interact with development policies and legal changes. We suggest that the progressive legal reforms compound with social norm in a way that make them effective only for some segments of the population, exacerbating the inequality of treatment between women. This statement must slightly be nuanced, because women of matrilocal tradition were not homogeneously better off than those from patrilocal ethnicities before reform. Further analyses suggest that within the matrilocal groups, the most disadvantaged women – living in remote areas and possibly in more conservative milieu – have caught up a little in terms of empowerment, possibly thanks to specific features of the reform such as circuit courts. Anyhow, the main implication of these results is that legal reforms can exclude subpopulations who, because of social norms, cannot takeup legal services as much as others. Policies, even when designed nationally, should be tailored to specific cultural contexts. More specifically, our results contribute to the analysis of patrilocality per se,

33A potential limit of our empirical approach might be that the main measures under study took place between 2008 and 2010 while our post-reform period of observation is 2014. On the contrary, what we capture in our estimates is likely to be a middle-term consequences of the reforms, which is arguably more interesting than the situation in 2010. Indeed, it allows a possible period of adaptation for the new measures to become fully operational and women to adapt to new legal opportunities.

140 Chapter 2. Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia a norm that not only reduces parents’ incentive to invest in daughters’ human capital, but also tends to have persistent consequences by limiting women’s legal opportunities in the longer run.

141 Chapter 2. Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia

142 References

Alesina, A., B. Brioschi, and E. L. Ferrara (2016). Violence against women: a cross-cultural analysis for africa. NBER Working Paper.

Alesina, A. and E. L. Ferrara (2005). Ethnic diversity and economic performance. Journal of Eco- nomic Literature 43(3), 762–800.

Alesina, A., P. Giuliano, and N. Nunn (2013). On the origins of gender roles: Women and the plough. The Quarterly Journal of Economics 128(2), 469–530.

Alﬁtri (2011). Legal reform project, access to justice and gender equity in indonesia. Indonesian Journal of International Law 9, 292.

Ashraf, N., N. Bau, N. Nunn, and A. Voena (2020). Bride price and female education. Journal of Political Economy.

Baland, J.-M., M. Boltz, G. Catherine, M. Seleck, and R. Ziparo (2020). Optimal intra-household decision structure. Working Paper.

Bank, W. (2015). World development report 2015: Mind, society, and behavior. Washington, DC: World Bank Publications.

Bargain, O., L. González, C. Keane, and B. Özcan (2012). Female labor supply and divorce: New evidence from ireland. European Economic Review 56(8), 1675–1691.

Bau, N. (2016). Can policy crowd out culture? Working Paper.

BenYishay, A., P. Grosjean, and J. Vecci (2017). The ﬁsh is the friend of matriliny: Reef density and matrilineal inheritance. Journal of Development Economics 127, 234–249.

Bertrand, M., E. Duﬂo, and S. Mullainathan (2004). How much should we trust differences-in- differences estimates? The Quarterly Journal of Eeconomics 119(1), 249–275.

Besley, T. and A. Case (2000). Unnatural experiments? estimating the incidence of endogenous policies. The Economic Journal 110(467), 672–694.

Botticini, M. and A. Siow (2003). Why dowries? The American Economic Review 93(4), 1385–1398.

Buttenheim, A. M. and J. Nobles (2009). Ethnic-based nuptial regimes and marriage behaviour. Population studies 63(3), 277.

Chiappori, P.-A. and M. Mazzocco (2019). Static and intertemporal household decisions. Journal of Economic Literature.

Corno, L., A. Voena, et al. (2016). Selling daughters: age of marriage, income shocks and bride price tradition. Working Paper.

143 References

Doepke, M. and M. Tertilt (2009). Women’s liberation: What’s in it for men? The Quarterly Journal of Economics 124(4), 1541–1591.

Duﬂo, E. (2001). Schooling and labor market consequences of school construction in indonesia: Evidence from an unusual policy experiment. The American Economic Review 91(4), 795–813.

Duﬂo, E. (2012). Women empowerment and economic development. Journal of Economic Litera- ture 50(4), 1051–1079.

Dyson, T. and M. Moore (1983). On kinship structure, female autonomy, and demographic behavior in india. Population and Development Review, 35–60.

Ebenstein, A. (2014). Patrilocality and missing women. Working Paper.

Edlund, L. (2001). Dear son-expensive daughter: Do scarce women pay to marry? Manuscript, Columbia Univ.

Fortunato, L. (2012). The evolution of matrilineal kinship organization. Proceedings of the Royal Society London.

Garg, A. and J. Morduch (1998). Sibling rivalry and the gender gap: Evidence from child health outcomes in ghana. Journal of Population Economics 11(4), 471–493.

Grace, J. (2004). Sasak women negotiating marriage, polygyny and divorce in rural east lombok. Intersections: Gender, History and Culture in the Asian Context 10.

Guha, B. (2010). Patrilocal exogamy as a monitoring mechanism: How inheritance and residence patterns co-evolve.

Holden, C. J. and R. Mace (2003). Spread of cattle led to the loss of matrilineal descent in africa: a coevolutionary analysis. Proceedings of the Royal Society of London B: Biological Sci- ences 270(1532), 2425–2433.

Jayachandran, S. (2015). The Roots of Gender Inequality in Developing Countries. Annual Review of Economics 7(1), 63–88.

Jayachandran, S. and I. Kuziemko (2011, 08). Why Do Mothers Breastfeed Girls Less than Boys? Evidence and Implications for Child Health in India. The Quarterly Journal of Economics 126(3), 1485–1538.

Jones, D. (2011). The matrilocal tribe: An organization of demic expansion. Human Nature 22(1-2), 177–200.

La Ferrara, E. (2007, 02). Descent rules and strategic transfers. evidence from matrilineal groups in ghana. Journal of Development Economics 83, 280–301.

La Ferrara, E. and A. Milazzo (2017). Customary norms, inheritance, and human capital: evidence from a reform of the matrilineal system in ghana. American Economic Journal: Applied Economics, Forthcoming.

Lépine, A. and E. Strobl (2013). The effect of women’s bargaining power on child nutrition in rural Senegal. World Development 45, 17–30.

Levine, D. and M. Kevane (2003). Are investments in daughters lower when daughters move away? evidence from indonesia. World Development 31(6), 1065–1084.

144 References

Loper, J. (2020). Women’s position in ancestral societies and female hiv: The long-term impact of matrilineality in sub-saharan africa. Working Paper.

Lowes, S. (2018). Kinship systems, gender norms, and household bargaining: Evidence from the matrilineal belt. Working Paper.

Lowes, S., N. Nunn, J. A. Robinson, and J. L. Weigel (2017). The evolution of culture and institutions: Evidence from the kuba kingdom. Econometrica 85(4), 1065–1091.

Mazzocco, M. (2007). Household intertemporal behaviour: A collective characterization and a test of commitment. The Review of Economic Studies 74(3), 857–895.

Opie, C., S. Shultz, Q. D. Atkinson, T. Currie, and R. Mace (2014). Phylogenetic reconstruction of bantu kinship challenges main sequence theory of human social evolution. Proceedings of the National Academy of Sciences 111(49), 17414–17419.

Rammohan, A. and M. Johar (2009). The determinants of married women’s autonomy in indonesia. Feminist Economics 15(4), 31–55.

Rammohan, A. and P. Robertson (2012). Do kinship norms inﬂuence female education? evidence from indonesia. Oxford Development Studies 40(3), 283–304.

Sadania, C. (2016). Working and women’s empowerment in the egyptian household: The type of work and location matter. AMSE Working Papers 1722.

Stevenson, B. and J. Wolfers (2006). Bargaining in the shadow of the law: Divorce laws and family distress. The Quarterly Journal of Economics 121(1), 267–288.

Strauss, J., F. Witoelar, and B. Sikoki (2016). The ﬁfth wave of the indonesia family life survey: Overview and ﬁeld report.

Sumner, C. and T. Lindsey (2011). Courting reform: Indonesia’s islamic courts and justice for the poor. International Journal for Court Administration 4(1), 3–16.

Sumner, C., M. Zurstrassen, and L. Lister (2011). Increasing access to justice for women, the poor, and those living in remote areas: An indonesian case study. World Bank Justice for the Poor Brieﬁng Note 6(2).

Sun, A. and Y. Zhao (2016). Divorce, abortion, and the child sex ratio: The impact of divorce reform in china. Journal of Development Economics 120, 53–69.

Sundaram, A. and R. Vanneman (2008). Gender differentials in literacy in india: The intriguing relationship with women’s labor force participation. World Development 36(1), 128–143.

Verick, S. (2018). Female labor force participation and development. IZA World of Labor, 87v2.

Voena, A. (2015). Yours, mine, and ours: Do divorce laws affect the intertemporal behavior of married couples? The American Economic Review 105(8), 2295–2332.

145 References

146 Appendix

2.A Descriptive Statistics and Raw Difference-in-Difference

Table 2.A.1 – Determination of Traditional Post-Marital Residence Norm by Ethnicity

Ethnicity # Villages Matrilocal Patrilocal Ambi/Neolocal Norm (%) (%) (%) Jawa 109 64.22 17.43 18.35 Matrilocality Sunda 40 67.50 7.50 25.00 Matrilocality Bali 15 0.00 86.67 13.33 Patrilocality Minang 12 100.00 0.00 0.00 Matrilocality Banjar 10 100.00 0.00 0.00 Matrilocality Betawi 10 70.00 20.00 10.00 Matrilocality Bugis 9 77.78 11.11 11.11 Matrilocality Sasak 9 0.00 100.00 0.00 Patrilocality Madura 6 83.33 16.67 0.00 Matrilocality Melayu 6 50.00 16.67 33.33 Matrilocality Batak 4 25.00 75.00 0.00 Patrilocality Bima 4 50.00 25.00 25.00 Matrilocality Cirebon 2 100.00 0.00 0.00 Matrilocality Makassar 2 100.00 0.00 0.00 Matrilocality Nias 2 0.00 100.00 0.00 Patrilocality Palembag 2 100.00 0.00 0.00 Matrilocality South Sumatra 2 0.00 100.00 0.00 Patrilocality Toraja 2 100.00 0.00 0.00 Matrilocality Dayak 1 100.00 0.00 0.00 Matrilocality Sumbawa 1 0.00 100.00 0.00 Patrilocality Tionghoa 1 0.00 100.00 0.00 Patrilocality Villages are grouped according to their dominant ethnic group. The table reports, for each ethnic group, the distribution of villages’ traditional norms of post-marriage residence (matrilocal, patrilocal or ambilocal/neolocal). Traditional norms are drawn from the declaration of local Adat experts in the 1997 IFLS. We attribute a residence norm to each ethnic group, defined as the modal answer from this distribution.

147 References

Table 2.A.2 – Traditional vs. Actual Matrilocality in Indonesia (2014)

Presence of Spouse’s Relatives in Household

Gap between Share of Share of Wife’s Husband’s Wife’s and Wife’s Husband’s Relatives Relatives Husband’s Relatives Relatives Relatives (1) (2) (3) (4) (5)

Matrilocality 0.0868*** -0.0439*** 0.0284*** -0.0129** 0.0412*** (0.0167) (0.0160) (0.00439) (0.00530) (0.00747) Relative effect 84.3% -22.2% 136.8% -24.4% -128.7% Ind. Controls Yes Yes Yes Yes Yes

Observations 5,880 5,880 5,880 5,880 5,880 R-squared 0.039 0.047 0.057 0.062 0.029 F-stat 26.98 7.55 41.81 5.90 30.43 Clusters 318 318 318 318 318 Linear estimations of contemporaneous co-residence practices on “matrilocality”, i.e. a dummy indicating that a woman belongs to a traditionally matrilocal ethnic group. Dependant variables include dummies for the presence of at least one wife’s relative in the household, for the presence of at least one husband’s relative in the household, share of wife’s relatives in the household, share of husband’s relatives in the household. ‘Gap between Wife’s and Husband’s Relatives’ is the difference between the number of wife’s relatives and the number of husband’s relatives in the household, divided by the size of the household. All estimations control for women’s characteristics: university gratuate, currently working, lives in a rural area, muslim, dummies for age category (by 5 years). Standard errors clustered at the village of origin level in brackets. Significance levels: * p<0.10, ** p<0.05, *** p<0.01.

148 References

Table 2.A.3 – Descriptive Statistics of Control Variables (Stable Couples, 2007-14)

2007 2014 Raw Matri. Patri. Diff. Matri. Patri. Diff. DD

Wife’s Control Variables Age Category 4.856 4.891 -0.035 6.174 6.169 0.005 0.040 (2.309) (2.354) (0.081) (2.307) (2.355) (0.081) (0.114) University 0.066 0.065 0.001 0.078 0.079 -0.001 -0.002 (0.248) (0.246) (0.009) (0.268) (0.270) (0.009) (0.013) Work 0.594 0.686 -0.091*** 0.667 0.741 -0.074*** 0.017 (0.491) (0.465) (0.017) (0.471) (0.438) (0.017) (0.023) Rural 0.506 0.575 -0.069*** 0.417 0.518 -0.101*** -0.031 (0.500) (0.495) (0.017) (0.493) (0.500) (0.017) (0.024) Muslim 0.969 0.540 0.429*** 0.969 0.540 0.429*** 0.000 (0.172) (0.499) (0.009) (0.172) (0.499) (0.009) (0.013)

Husband’s Control Variables Age Category 5.797 5.679 0.117 7.111 6.965 0.146 0.029 (2.469) (2.481) (0.086) (2.486) (2.490) (0.086) (0.122) University 0.073 0.097 -0.024** 0.082 0.109 -0.027*** -0.004 (0.261) (0.296) (0.010) (0.274) (0.312) (0.010) (0.013) Work 0.956 0.958 -0.001 0.921 0.930 -0.009 -0.008 (0.204) (0.202) (0.008) (0.270) (0.255) (0.008) (0.012) Rural 0.503 0.576 -0.073*** 0.414 0.519 -0.104*** -0.031 (0.500) (0.494) (0.017) (0.493) (0.500) (0.017) (0.024) Muslim 0.968 0.537 0.431*** 0.968 0.537 0.431*** 0.000 (0.177) (0.499) (0.009) (0.177) (0.499) (0.009) (0.013) Number of observations 4954 989 5943 4954 989 5943 11886 Prop. Matri/Patri 83.4% 16.6% 83.4% 16.6%

Additional Controls (Robustness Checks) Close District Capital 0.575 0.545 0.030 0.575 0.545 0.030 0.000 (0.494) (0.498) (0.020) (0.494) (0.498) (0.020) (0.028) Number of observations 3568 727 4295 3568 727 4295 8590 Javanese 0.565 0.000 0.565*** 0.565 0.000 0.565*** 0.000 (0.496) (0.000) (0.016) (0.496) (0.000) (0.016) (0.022) Number of observations 4954 989 5943 4954 989 5943 11886 Mixed Ethnicity 0.132 0.015 0.117*** 0.132 0.015 0.117*** 0.000 (0.339) (0.122) (0.011) (0.340) (0.122) (0.011) (0.015) Number of observations 4954 989 5943 4954 989 5943 11886 Natural Disaster 0.557 0.580 -0.023 0.557 0.565 -0.008 0.015 (0.497) (0.494) (0.021) (0.497) (0.496) (0.021) (0.030) Number of observations 3326 676 4002 3326 676 4002 8004 Statistics based on main DID sample in Table 2.2 (stable couples 2007-2014). Matri/patri: individual of ethnicities from matrilocal/patrilocal tradition. Diff.: time difference, Raw DD: absolute difference- in-difference (with ***, **, * indicating significance at 1%, 5%, 10% levels). Standard deviations are reported in brackets in columns 1, 2, 4 and 5. Standard errors are reported in brackets in columns 3, 6 and 7.

149 References

Table 2.A.4 – Raw Difference-in-Differences of Outcome Variables (Stable Couples, 2007-14)

2007 2014 Raw Matri. Patri. Diff. Matri. Patri. Diff. DD

Dep. Variables: Well-Being Measures Morbidity Symptoms 0.761 0.705 0.055*** 0.843 0.839 0.004 -0.051*** (0.427) (0.456) (0.014) (0.363) (0.368) (0.014) (0.020) Number of Births 2.628 3.085 -0.457*** 3.287 4.018 -0.731*** -0.274** (2.300) (2.662) (0.084) (2.287) (2.838) (0.084) (0.119) Standard of Living 1.948 1.976 -0.028 2.052 1.929 0.123*** 0.150*** (0.531) (0.592) (0.021) (0.656) (0.656) (0.021) (0.030) Food Consumption 2.010 2.026 -0.016 2.148 2.018 0.130*** 0.146*** (0.511) (0.566) (0.020) (0.606) (0.636) (0.020) (0.028) Ch. Std. of Living 2.027 2.010 0.016 2.134 2.009 0.126*** 0.109*** (0.530) (0.556) (0.027) (0.656) (0.656) (0.027) (0.039) Ch. Food Conso. 2.066 2.033 0.033 2.226 2.091 0.135*** 0.102*** (0.516) (0.554) (0.026) (0.620) (0.645) (0.026) (0.037) Wife Assets 25,218 22,697 2,521 54,433 40,960 13,473*** 10,952** (58,793) (72,253) (3,540) (132,600) (116,218) (3,540) (5,006)

Dep. Variables: Empowerment (Final Say) Contraception 0.191 0.191 0.000 0.342 0.220 0.122*** 0.122*** (0.393) (0.393) (0.016) (0.474) (0.415) (0.016) (0.022) Large Expenditures 0.064 0.053 0.011 0.191 0.115 0.076*** 0.064*** (0.245) (0.224) (0.012) (0.393) (0.320) (0.012) (0.016) Statistics based on main DID sample in Table 2.2 (stable couples 2007-2014). Matri/patri: individuals of ethnicities from matrilocal/patrilocal tradition. Diff.: time difference, Raw DD: absolute difference-in-difference (with ***, **, * indicating significance at 1%, 5%, 10% levels). Standard deviations are reported in brackets in columns 1, 2, 4 and 5. Standard errors are reported in brackets in columns 3, 6 and 7.

150 References

2.B Cross-Sectional Estimations (Correlations)

Table 2.B.1 – Villages’ Post-Marriage Residence Norm and Divorce related Adat Traditional Norms

Divorce settled Husband takes Husband takes Young children in religious all assets from all assets acquired live with the man or civil courts before marriage during marriage or his parents (1) (2) (3) (4) (5) (6) (7) (8) Matrilocal Village 0.102 -0.043 -0.056** -0.173*** (0.065) (0.027) (0.026) (0.046) Patrilocal Village -0.247*** 0.064* 0.074** 0.252*** (0.071) (0.038) (0.037) (0.062)

Observations 247 247 249 249 249 249 249 249 R-squared 0.010 0.044 0.013 0.021 0.027 0.037 0.071 0.116 Village-level linear estimations of situations in case of divorce on either matrilocal or patrilocal traditional norm of post-marriage residence. If matrilocal=1 (patrilocal=1), 0 corresponds to patrilocal (matrilocal), neolocal and ambilocal. The norm is obtained from the Adat questionnaire (answers by Adat experts in each village) in 1997 IFLS data. Columns 1 & 2: if a divorce happens, the decision-making process used in the divorce is 1: religious/civil courts or 0: family discussion. Columns 3 & 4: if a divorce occurs, 1: the husband has the right to claim those assets that existed before marriage, 0 otherwise. Columns 5 & 6: if a divorce occurs, 1: the husband has the right to claim those assets obtained since the couple was married, 0 otherwise. Columns 7 & 8: after a divorce, young children go 1: with the husband or husband’s parents, 0 otherwise. Robust standard errors in brackets. Significance levels: * p<0.10, ** p<0.05, *** p<0.01.

151 References Large expenditures Contraception Wife’s assets value Children’s food conso. Children’s Panel B: 2007 Panel A: 2014 std. of living Food consumption Women’s and Child’s Well-Being Women’s Empowerment of living Standard Number of births Correlations between Matrilocality and Women’s Well-Being and Empowerment (Stable Couples, 2007-14) (1) (2) (3) (4) (5) (6) (7) (8) (9) (0.0164) (0.154) (0.0390) (0.0353) (0.0443)(0.0348) (0.0422) (0.114) (5,033) (0.0607) (0.0538) (0.0275) (0.0631) (0.0185) (0.0580) (2,156) (0.0231) (0.0125)) Morbidity symptoms Table 2.B.2 – Matrilocal -0.00598Observations -0.900***R-squared 0.108*** 5,920Clusters 0.119*** 0.018 5,745 0.116***Matrilocal 318 0.201 5,773 0.118*** 0.0880**Observations 14,495*** 0.075 -0.640*** 318R-squared 5,774 -0.0847 5,920Clusters 0.129***Controls 0.078 318 -0.0691 0.018 5,745 0.0663*** 3,365 318 0.409 -0.0375 0.072 5,773 Yes 318 3,364 0.048 -0.0239 318 5,774 0.093 Yes 5,935 316 3,725* 0.045 318 0.080 Yes 3,365 5,442 -0.0301 0.049 316 318 3,364 0.027 Yes 0.0128 5,442 0.047 318 5,935 316 0.022 Yes 0.111 5,442 317 316 Yes 0.018 5,442 317 318 Yes 0.007 317 Yes 317 Yes Cross-sectional linear estimations of women’s well-being and empowerment outcomes (defined in the footnote of Table 2.2 ) on a matrilocal dummy (indicating whether ancharacteristics individual (university belongs graduate, to currently aerrors working, traditionally are living matrilocal reported in ethnic rural in group) areas, brackets as and muslim well clustered and as at age women’s village group and of dummies women’s origin using spouse’s level. 5-year steps). Significance levels: Standard * p<0.10, ** p<0.05, *** p<0.01.

152 References Large expenditures Contraception Wife’s assets value Children’s food conso. Children’s std. of living Food consumption Women’s and Child’s Well-Being Women’s Empowerment of living Standard Number of births Effect of Legal Reforms on Women’s Well-Being Empowerment and (Pooled Cross-Sections) (1) (2) (3) (4) (5) (6) (7) (8) (9) (0.585) (0.617) (0.496) (0.516) (0.480) (0.854) (35,631) (0.344) (0.101) -11,.5% 1.8% 10.4% 10% 9.4% 9.5% 28.5% 76.8% 79.5% (0.0332) (0.0973) (0.0471) (0.0474) (0.0512) (0.0462) (3,571) (0.0299) (0.0125) Morbidity symptoms Table 2.B.3 – Matrilocal -0.0809** 0.0605 0.203*** 0.201*** 0.189***Controls 0.194*** Yes 6,514* Yes 0.136*** 0.0437*** Yes Yes Yes Yes Yes Yes Yes × × Post PostRelative effect Observations 0.212R-squaredClusters 3.255***Controls 18,243Post 0.624 17,929 0.029 1.590*** 17,930 319 0.377 Yes -0.706 17,932 0.073 319 Yes -0.121 12,253 0.087 319 107,716*** Yes 12,249 0.473 0.087 319 Yes 18,271 -0.0569 0.103 17,189 319 Yes 0.098 17,189 319 Yes 0.045 319 Yes 0.043 319 Yes 319 Yes Difference-in-difference estimations of well-being empowerment and indicators on a sample of couples surveyed in 2007 couples and surveyed in individual belongs to a traditionally matrilocal ethnic group.as Post Outcomes interacted are with defined all in theseTable. 2.2 controls. Allp<0.01. We estimations report include relative controls effects defined that in are calculated in % of mean outcome for patrilocal group in 2007 2014 (pooled cross-sections). Post is equal to 1 for 2014 (post-reform) and 0 for 2007 (pre-reform). Matrilocal is a dummy indicating whether an Table 2.2 + a dummy indicating matrilocality, a dummy indicating muslim religion, a dummy indicating a spouse of muslim religion, as well (pre-reform). Standard errors are reported in brackets and clustered at village of origin level. Significance levels: * p<0.10, ** p<0.05, ***

153 References

2.C Placebo and Speciﬁcation Checks

Table 2.C.1 – Women’s Divorce Probability: Placebo Estimations

Divorced Divorced Dep. Var. Divorced Divorced Divorced or Divorced or Separated Separated Estimator Diff-in-Diff Simple Diff.

Excluding Excluding Married singles & Married singles Married Married Samples before widowed, before and in 2000 in 2000 2000 married 2000 widowed before 2000 (1) (2) (3) (4) (5) (6) Post 0.0117 0.0253 0.00916 0.0271 (0.0101) (0.0299) (0.0180) (0.0274) Post × Matrilocal -0.00167 0.00589 -0.00180 -0.00763 (0.00483) (0.00615) (0.00546) (0.00603) Matrilocal 0.00122 -0.00240 (0.00454) (0.00599)

Observations 10,772 13,790 9,524 10,772 7,147 7,147 R-squared 0.019 0.011 0.021 0.022 0.006 0.016 Clusters 320 320 320 320 320 320 Individual FE Yes Yes Yes Yes No No Controls Yes Yes Yes Yes Yes Yes Post × Controls Yes Yes Yes Yes No No Linear estimations of women’s divorce status (dummy for divorced, or divorced/separated). We apply the difference-in-difference approach to a selection of women observed in both 2000 and2007, who were married in 1997 (columns 1, 3 and 4); and a selection of women being married, divorced or separated in 2000 and 2007 (columns 2 and 3). For them, Post is equal to 1 for observations in 2007 and 0 in year 2000. We also estimate the potential increase in divorce using women observed in 2007 who were married in 2000 (columns 5 and 6). Matrilocal is a dummy indicating whether an individual belongs to a traditionally matrilocal ethnic group. Estimations include individual FE (absorbing Matrilocal and muslim - column 1-4), time-varying controls (women’s characteristics: university graduate, currently working, living in rural areas and age group dummies using 5-year steps), + a muslim dummy in columns 5 and 6, and interactions between Post and controls (including Post interacted with a muslim dummy) in columns 1-4 as indicated. Standard errors are reported in brackets and clustered at village of origin level. Significance levels: * p<0.10, ** p<0.05, *** p<0.01.

154 References Large expenditures Contraception Wife’s assets value Children’s food conso. Children’s std. of living Food consumption Women’s and Child’s Well-Being Women’s Empowerment of living Standard Well-Being and Empowerment (Stable Couples): Placebo Estimations Number of births Table 2.C.2 – (1) (2) (3) (4) (5) (6) (7) (8) (9) (0.293) (3.301) (0.833) (0.514) (0.592) (0.631) (42,050) (0.381) (0.333) (0.0391) (0.0807) (0.0762) (0.0717) (0.0716) (0.0708) (3,027) (0.0485) (0.0174) Morbidity symptoms Matrilocal 0.0396 -0.125 0.0704 0.0987Controls -0.00807 Yes 0.0601 Yes 2,353 Yes 0.00179 Yes -0.0299* Yes Yes Yes Yes Yes × × Post PostObservationsR-squared -0.215ClustersHousehold FE 6,640 5.367ControlsPost 2.023** 0.038 6,358 Yes 1.018** 317 0.375 6,438 Yes Yes -0.986* 0.063 317 6,436 Yes Yes -0.385 0.077 316 2,586 Yes 60,677 Yes 0.075 316 2,584 Yes 0.483 0.061 Yes 6,642 293 0.556* 0.178 Yes Yes 6,084 293 0.047 Yes Yes 6,084 317 0.025 Yes Yes 316 Yes Yes 316 Yes Placebo difference-in-difference estimations well-being of empowerment and indicators on a sample of stable couples (2000-2014) surveyed in both 2000 and 2007. Post is equalreported to in 1 brackets for and 2007 clustered and at 0 village for of 2000. origin Other level. variables Significance are levels: described in * p<0.10,Table 2.2 . ** p<0.05, Estimations include *** p<0.01. household FE (absorbing Muslim), basic controls, and control interactions with Post (including Post interacted with a musim dummy). Standard errors are

155 References we Large Controlling expenditures Sample excluding we do not control for Natural Disaster is our main specification × Rural × Baseline Contraception No Post Wife’s we additionally control for Post interacted with assets value we additionally control for Post interacted with ‘bride price’, Children’s food conso. Mixed Ethnicity × Controlling for Natural Disaster and Post Bride Price × Children’s std. of living we do not control for Post interacted with ‘muslim’ dummy anymore. In Controlling for Post Food Muslim × consumption Controlling for Post Natural Disaster × Women’s and Child’s Well-Being Women’s Empowerment No Post of living Standard we do not control for Post interacted with time varying controls anymore. In Effect of Legal Reforms on Women’s Well-Being Empowerment: and Robustness Checks Number of births Javanese Mixed Ethnicity Bride Price Controls × × × × (1) (2) (3) (4) (5) (6) (7) (8) (9) we additionally control for Post interacted with ‘javanese’, a dummy indicating an individual of javanese ethnicity. In Morbidity symptoms Table 2.C.3 – No Post Muslim Controls Rural × × × Javanese we exclude javanese individuals from our main sample. In × Rel. effect (%)ObservationsSample -13.4*** excluding Javanese Rel. effect 11,840 (%) -6.5**ObservationsControlling -16.5*** for 11,490 Post Rel. 9.7*** effect 6,252 (%) -6.8**Observations 11,546Controlling -12.4*** for Natural 10.7*** DisasterRel. 6,062 and effect 11,840 (%) Post -7.7**Observations 11,548Controlling -14.6*** 10.2*** for 11,490 6,088 Post Rel. 9.5*** effect9.3*** 7,978 (%) -6.8**Observations 11,546 6,730 -13.3*** 6,094 8.8*** 7,656 11,840 -7.9*** 7.4** 11,548 10.1*** 11,490 6,728 7,734 7.7**9.1*** 3,778 6.7** 11,546 6,730 11,870 7,734 7.7***9.7*** 7.0*** 73.4*** 3,776 11,5489.5*** 6,72810.2*** 10,884 69.0** 4,084 6.3** 72.8*** 120.0*** 6,276 6,730 7.5*** 11,870 7.4***83.2*** 122.8*** 10,884 4,084 44.1* 7.0*** 5,766 6,728 10,884 63.5** 7,990 83.2*** 99.6*** 11,870 49.8** 98.0*** 136.1*** 5,766 10,884 7,41485.3*** 92.5*** 10,884 7,414 10,884 Baseline Rel. effect (%)ObservationsNo -12.6*** Post Rel. effect 11,840 (%) -7.6**ObservationsNo Post 11,490 -7.0*Rel. 9.7*** effect 11,840 (%)Observations 11,546No -8.2*** -12.7*** Post 11,490 Rel. effect 11,840 (%) 7.9*** -7.6**Observations 11,548 11,546Controlling for 11,490 Post -6.6 9.8***9.3*** 11,840 11,548 11,546 6,730 -8.6*** 11,490 7.6***7.3*** 7.6*** 11,548 11,546 6,730 6,7289.5*** 6.9*** 11,548 5.7*** 6,730 6,728 11,870 7.7*** 47.1** 5.2**6.9*** 6,730 6,728 11,870 10,884 7.2***85.3*** 103.6*** 47.2** 5.5** 6,728 11,870 10,884 49.1** 10,884 65.4*** 118.7*** 4.3** 11,870 10,88486.4*** 109.4*** 10,884 48.8** 10,884 10,884 66.0*** 111.9*** 10,884 Difference-in-difference estimations of well-being empowerment and indicators on a sample of stable couples surveyed in both 2007 and 2014. Outcomes are additionally control forthe ‘natural survey, disaster’, and a Posta dummy interacted dummy indicating with indicating an ‘natural anoutcome individual disaster’. individual for living belonging In patrilocal to inp<0.01. group an a in ethnic village 2007 group having (pre-reform). traditionally experienced practicing Standard a bride errors natural price. are disaster We reported in report in the relative effects brackets 5 that and years clustered preceding are calculated at village of originin % of mean level. * p<0.10, ** p<0.05, *** defined in Table 2.2 . All estimations include household FE (absorbing matrilocal and muslim) and controls defined in Table 2.2 . in Table 2.2 .Post In interacted withfor ‘rural’ Post dummy anymore. In Javanese ‘mixed ethnicity’, a dummy indicating a couple with spouses of different ethnicities. In

156 References Large expenditures Contraception Wife’s assets value Children’s food conso. Children’s std. of living questionnaire in IFLS. Standard errors are reported in brackets and clustered at Food Adat consumption Women’s and Child’s Well-Being Women’s Empowerment Alternative Definitions Treatment:of Matrilocal Intensity of living Standard Number of births Table 2.C.4 – (1) (2) (3) (4) (5) (6) (7) (8) (9) (0.528) (3.532)(0.037) (0.512) (0.106) (0.066) (0.402) (0.062) (0.431) (0.068) (0.563) (0.064) (74,582) (7,564) (0.240) (0.043) (0.141) (0.022) Morbidity symptoms Matrilocal Int. -0.153*** -0.0956 0.283*** 0.242*** 0.188***Controls 0.159** 23,585*** Yes 0.184*** Yes 0.0564** Yes Yes Yes Yes Yes Yes Yes × × PostPost ObservationsR-squaredClusters 1.316**Household FEControls -3.419 11,840Post 0.481 11,490 0.050 Yes 11,546 -0.536 318 0.346 11,548 Yes Yes 0.045 318 -0.012 Yes Yes 6,730 0.059 318 0.320 Yes 6,728 Yes 180,859** 0.039 318 11,870 Yes 0.338 0.073 Yes 316 10,884 0.094 0.085 Yes Yes 316 10,884 0.064 Yes Yes 318 0.088 Yes Yes 317 Yes Yes 317 Yes Difference-in-difference estimations of well-being empowerment and indicators on a sample of stable couples surveyed in both 2007 and 2014. Outcomes are defined in Table 2.2 . Allthe estimations main include ethnic household group) FE reporting (absorbing traditional matrilocal matrilocality intensity in and muslim) and controls defined in village of origin level. Significance levels: * p<0.10, ** p<0.05, *** p<0.01. Table 2.2 . ‘Matrilocal Int.’ indicates individual’s ethnic group’s matrilocal intensity, defined as the proportion of villages (where this ethnic group is

157 References Large expenditures Contraception Wife’s assets value Children’s food conso. Children’s std. of living Food consumption Women’s and Child’s Well-Being Women’s Empowerment Alternative Definitions Treatment:of Including Mixed Couples of living Standard Panel A: Matrilocality defined based on Wife’s Ethnic Group’s Norm Panel B: Matrilocality defined based on Husband’s Ethnic Group’s Norm Number of births Table 2.C.5 – (1) (2) (3) (4) (5) (6) (7) (8) (9) -11.4% -6.1% 9.0% 8.9%-11.4% 8% -7.4% 9.0% 7.2% 7.9% 40.8% 5.6% 76.0% 5.4% 75.3% 45.0% 70.8% 107.7% (0.030) (0.083) (0.049) (0.046) (0.050)(0.029) (0.082) (0.047) (0.050) (5,219) (0.048) (0.030) (0.052) (0.016) (0.046) (4,930) (0.033) (0.015) Morbidity symptoms Mixed Couple Yes Yes Yes Yes Yes Yes Yes Yes Yes Matrilocal -0.081*** -0.184** 0.177*** 0.180*** 0.163***Matrilocal 0.148*** -0.081*** -0.223*** 9,187* 0.178*** 0.159*** 0.149***Controls 0.113** 0.043*** 0.111** Yes 10,317** Yes 0.138*** Yes 0.057*** Yes Yes Yes Yes Yes Yes × × × × Post Relative effect ObservationsR-squaredClusters 12,346 0.048 11,988Relative effect 318 12,042 0.354ObservationsR-squared 12,044Clusters 0.042 319 12,328 0.058 7,062 319 0.048 11,970 7,062 0.042 318 319 12,024 0.354 12,376 12,026 0.043 319 0.073 316 11,356 0.056 7,050 319 0.086 316 11,356 7,050 0.039 319 0.064 319 12,358 0.071 0.086 316 11,334 0.087 318 316 11,334 0.063 318 319 0.087 318 318 Post Post Household FEControlsPost Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Difference-in-difference estimations of well-being empowerment and indicators on a sample of stable couples surveyed in both 2007 and 2014, including couples with spouses originatingestimations from include ethnic household groups FE withdummy (absorbing different indicating matrilocal, post-marital whether muslim residence an and norm individual mixed (‘Mixed originates couple Couples’). from dummies) an Outcomes and ethnic controls group are are with defined in the defined aTable husband in . 2.2 different originatesTable post-marital All 2.2 . from residence a norm ‘Mixed traditionally Couple’ matrilocalthat her/his is ethnic spouse. a group. The Post relativeis equal to effect is calculated in % of outcome mean for patrilocal group in 2007 wife originates from a traditionally matrilocal ethnic group. In Panel B, Matrilocal is a dummy indicating whether an individual belongs to a couple where 1 for 2014 (post-reform) and 0 for 2007 (pre-reform). In Panel A, Matrilocal is a dummy indicating whether an individual belongs to a couple where the (pre-reform). Standard errors are reported in brackets and clustered at village of origin level. Significance levels: * p<0.10, ** p<0.05, *** p<0.01.

158 References Wife alone & decides respondant Wife respondant Wife alone decides Husband (baseline) respondant Wife alone & decides respondant Wife respondant Wife alone decides Effect on Empowerment: Alternative Definitions based Final on Say Answers (1) (2) (3) (4) (5) (6) (7) (8) 85.3% 85.3% 64.2% 64.4% 103.6% 105.3% 69.1% 74.3% (0.0333) (0.0333) (0.0263) (0.0263) (0.0159) (0.0159) (0.0183) (0.0187) Husband (baseline) respondant Table 2.C.6 – Matrilocal 0.163*** 0.163*** 0.0956*** 0.0953*** 0.0549*** 0.0558***Controls 0.0373** Yes 0.0401** Yes Yes Yes Yes Yes Yes Yes × × Relative effect ObservationsR-squaredClusters 10,884 0.065 10,884 317 11,384 0.065 11,384 317 0.091 10,884 0.091 318 10,884 0.088 318 11,384 0.090 11,384 317 0.108 317 0.110 318 318 Dep. Var.Specification Post ContraceptionHousehold FEControlsPost Yes Yes Yes Large expenditures Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Difference-in-difference estimations of empowerment on a sample of stable couples surveyed in both 2007 and 2014. Post is equal to 1 for 2014FE (post-reform) (absorbing and matrilocal 0 and fordummy). muslim), 2007 The basic (pre-reform). baseline controls, Other definition allher variables controls relatives are interactedof empowerment present defined with outcomes in in Post reliesanswer the (includingTable or household 2.2 . post defined have the husband’s interacted Estimations empowerment on theof with include say origin a answer as household (while level. muslim her making the and Significance husband gives does levels: the decision not have * p<0.10, Standard alone. a value any say). ** p<0.05, errors the wife and/or 1 if Alternative are definitions *** p<0.01. usereported in the wife’s andbrackets clustered at village

159 References

2.D Proofs of Propositions 1 and 2

We study the expected bargaining power and the probability of divorce of matrilocal and patrilocal women. To solve the model, we follow the algorithm in Voena(2015). At each period 푡, the spouses maximize the following value function:34

퐻ℎ 퐻ℎ 퐻ℎ ℎ 퐻ℎ 푊 ℎ 푊 ℎ 푊 ℎ ℎ 푊 ℎ 푉푡(휔푡) = 훾푡 [푢 (푐푡 , 푄푡 ) + 휒푡 ] + 훾푡 [푢 (푐푡 , 푄푡 ) + 휒푡 ] + 훽퐸[푉푡+1(휔푡+1)] s.t.

ℎ ℎ 퐻ℎ 퐻ℎ 푊 ℎ 푊 ℎ 퐻ℎ 푊 ℎ 퐴푡+1 − (1 + 푟푡)퐴푡 + 푐푡 + 푥푡 + 푐푡 + 푥푡 ≤ 푦푡 + 푦푡

푗ℎ *푗 푗ℎ Once the within-marriage allocation is established, we can deﬁne 푉푚푎푟푟푖푒푑,푡(휔푡) = 푢(푐푡 ; 휒푡 ). Then, the following algorithm applies:

푗ℎ 푗ℎ 푗ℎ 푗ℎ 1. if 푉푚푎푟푟푖푒푑,푡(휔푡) ≥ 푉푑푖푣표푟푐푒푑,푡(휔푡) for both 푗 = 퐻, 푊 , then 푉푡+1(휔푡) = 푉푚푎푟푟푖푒푑,푡(휔푡) and the couple remains married.

푗ℎ 푗ℎ 푗ℎ 푗ℎ 2. if 푉푚푎푟푟푖푒푑,푡(휔푡) < 푉푑푖푣표푟푐푒푑,푡(휔푡) for both 푗 = 퐻, 푊 , then 푉푡+1(휔푡) = 푉푑푖푣표푟푐푒푑,푡(휔푡) and the couple divorces.

푗ℎ 푗ℎ 푖ℎ 푖ℎ 3. 푉푚푎푟푟푖푒푑,푡(휔푡) < 푉푑푖푣표푟푐푒푑,푡(휔푡) and 푉푚푎푟푟푖푒푑,푡(휔푡) ≥ 푉푑푖푣표푟푐푒푑,푡(휔푡) for 푗 = 퐻, 푊 , 푖 = 퐻, 푊 and 푖! = 푗.

In case 3, the allocation shifts. A new marriage allocation is determined through the maximisation of 푗ℎ 푗ℎ the value function as deﬁned above but with renegotiated bargaining weights 훾푡 + 휇푡 . The 휇s are such that 푗 stays in marriage, i.e. they are deﬁned as Lagrangian multipliers of the binding version 푗ℎ 푗ℎ of the participation constraint 푉푚푎푟푟푖푒푑,푡(휔푡) ≥ 푉푑푖푣표푟푐푒푑,푡(휔푡), and that the value of marriage for 푖 is high enough after renegotiation. If no such changes in bargaining weights can be found, the couple divorces.

Divorce occurs for two main reasons. First, one of the two spouses has a negative preference shock for marriage, so that no internal allocation allows her/him to stay without making the other spouse too worse off. Second, one of the spouses get an income shock so bad that the other spouse, given his/her tastes for marriage, does not ﬁnd optimal to stay in marriage anymore. Renegotiation of the Pareto weights occur for very similar reasons, with the exception that they could be driven also by spouses’

34 The model is solved backward. In the final period 푇 , the asset constraint is the following: 퐴푇 +1 ≥ 0

160 References positive income shocks. Renegotiations occur in case 3 above when there exist feasible allocations within marriage that allow the spouse who would like to divorce to gain power without making the other spouse unwilling to stay.

Divorce. We study the wife’s incentives to divorce in terms of her and her husband’s marriage

푊 ℎ 푊 ℎ preferences and their income at time 푡. For each combination of {푦푡 , 휒푡 }, we can deﬁne a 퐻ℎ husband’s resource level 푦푡 that makes the wife indifferent between staying and leaving. This level is independent of the state realisation of the bargaining power, as redistribution may occur as soon as divorce is preferable for at least one spouse.

퐻ℎ Also, this 푦푡 corresponds to a minimum level of wife’s resources share, associated with the Pareto 푊 ℎ 푊 ℎ 푊 ℎ 푊 ℎ weight level 휏푡 = 훾푡 + 휇푡 , that she needs to reach not to divorce. If 훾푡 + 휇푡 ≥ 휏푡, the wife accepts to stay married. Whether this bargaining weight is feasible depends on the above mentioned

퐻ℎ parameters plus the husband preference parameters 휒푡 . So for each combination of state variables 푊 ℎ 퐻ℎ 퐻ℎ 푊 ℎ {푦푡 , 푦푡 , 휒푡 }, we can deﬁne a minimum level of taste for marriage of the wife 휒푑푖푣표푟푐푒푑,푡 below which divorce occurs.

We want to compare the probability of divorce and the expected bargaining power of matrilocal and patrilocal women. Since the utility functions of the two types of women are assumed to be the same,

푊 ℎ both women have the same taste for marriage threshold 휒푑푖푣표푟푐푒푑,푡, below which they divorce. We 푊 푀 푊 푀 푊 푃 푊 푃 then have to compare 푃 (휒푡 < 휒푑푖푣표푟푐푒푑,푡|Ω푡) to 푃 (휒푡 < 휒푑푖푣표푟푐푒푑,푡|Ω푡). All the stochastic variables follow a random walk process. The income variables have the same expected value and variance in both groups. The preference parameter of the husband has the same distribution in both groups. What differs between group is the distribution of the wife marriage parameter. In particular,

푊 푀 푊 푀 2 푊 푃 푊 푃 2 at time 푡 we have that 휒푡 ∼ 푁(휒0 , 푡휎 ) and 휒푡 ∼ 푁(휒0 , 푡휎 ). To know who has the 푊 푀 푊 푃 higher probability of divorce we compare the CDF of the two distribution. Since 휒0 < 휒0 , it 푊 푀 푊 푃 follows that Φ푡 > Φ푡 and matrilocal women are expected to divorce more frequently in every period 푡. This leads to prediction 1.

Bargaining power. Similarly, we can study the evolution of bargaining power when the couple stay married. Given a starting bargaining power 훾0, renegotiations will occur each time the existing bargaining power does not provide enough resources in marriage as compared to divorce.

푊 ℎ From the wife’s point of view, for every 훾푡 we can deﬁne a minimum level of taste for marriage,

161 References

푊 ℎ 푊 ℎ 퐻ℎ 푊 ℎ 휒푟푒푛푒푔표푡푖푎푡푖표푛,푡, given {푦푡 , 푦푡 }, below which renegotiation occurs. Given the distribution of 휒푡 for matrilocal and patrilocal women ℎ = 푀, 푃 , renegotiation takes place more often for matrilocal couples. Starting from 훾0, the bargaining power of matrilocal women will thus increase more often than for patrilocal ones.

However, this will in turn make the husbands of matrilocal women also willing to renegotiate more, whenever their divorce constraint binds. Then, in the two groups, we can compute at every 푡 the

퐻ℎ probability that a woman accepts the husband’s renegotiation. For every realisation of 휒푡 < 퐻ℎ 푊 ℎ 퐻ℎ 휒푟푒푛푒푔표푡푖푎푡푖표푛,푡, i.e. the husband threshold for renegotiation given {푦푡 , 푦푡 }, the couple will stay 푊 ℎ 푊 ℎ married if 휒푡 > 휒푑푖푣표푟푐푒,푡. This will happen more often for patrilocal women than matrilocal women, implying that patrilocal women will accept a deterioration of they bargaining power, while matrilocal women will divorce. So, in expectation, the matrilocal women selected into marriage are those with a higher bargainig power. This yields prediction 2.

162 163 References

164 Chapter 3

Women’s Empowerment and Husband’s Migration: Evidence from Indonesia

This chapter is based on a joint research with Olivier Bargain (Bordeaux University) and Roberta

Ziparo (Aix-Marseille University).

Abstract: Migration is an important risk-coping mechanism for poor households in developing countries. However, migration decisions may be sub-optimal in the presence of limited commitment between spouses. In this paper, we examine the link between the distribution of power in marriage and the decision to split-migrate (one spouse migrates alone) in Indonesia. We exploit a national policy experiment that has exogenously increased women’s bargaining power among ethnic groups of matrilocal tradition (the couple lives with the bride’s relatives) relative to patrilocal groups. The propensity of matrilocal husbands to split-migrate, relative to patrilocal husbands, increases by 2-3.4 percentage points, i.e. a rise of 41-76%, following the reform. We suggest that empowered women may have ex-ante gained control over outcomes that are costlier to monitor for husbands when they migrate. Hence, empowerment restores some efﬁciency in migration decisions by reducing the anticipated information asymmetry and the moral hazard associated with migration. Consistently, we show that households with empowered women are more able to cushion shocks due to natural disasters and, among all households experiencing split-migration, matrilocal women are better off than their patrilocal counterparts. We provide a theoretical framework that rationalizes the intra-household mechanisms behind these intuitions.

165 Chapter 3. Women’s Empowerment and Husband’s Migration: Evidence from Indonesia

Keywords: Migration, Female Empowerment, Intra-Household Decision-Making, Ethnic Norms,

Gender, Natural Experiment

JEL Classiﬁcation: D13, J1, K38, K4, O15, R23, Z1

166 Chapter 3. Women’s Empowerment and Husband’s Migration: Evidence from Indonesia

3.1 Introduction

Migration ﬂows have sharply increased in the last decades.1 In absence of formal insurance and credit markets, migration has become a frequent strategy for low-income households to cope for several risks (e.g. natural disasters, crop failure, price ﬂuctuations, etc.). It provides an ex post solution to cushion negative income shocks as well as an ex-ante strategy to diversify future income streams and to deal with uncertainty (Amuedo-Dorantes and Pozo, 2011). Migration is a costly investment and frequently entails the departure of only some of the family members, often those with the highest expected returns. As a result, migration is characterized by a rising proportion of transnational or so- called split households. For instance in Indonesia, one of the main migrant-sending countries in Asia, the proportion of split-migrating men was 49.2% over the period 2000-07 and 57.6% over 2007-14.2

For couples, whether migration is joint or split is an important decision as it does not only in-

fluence their living conditions and the total resources available to the household: it also affects the distribution of resources within the household. Efficient decisions require implicit contracts to be sustained over time.3 Yet, when families are separated, as in the case of split migration, the existence of information asymmetries and incomplete contracts often lead to inefficient decisions (Ashraf, Ayci- nena, Martínez A, and Yang, 2015). Beyond this result, little is known about the process leading to migration choices and in particular about the selection of specific individuals as potential migrants within the family. A nascent literature investigates these questions but rarely when taking into account the fact that spouses may have different views on the value of migration and different levels of bargaining power regarding these decisions (see the discussion in Chort and Senne 2018).4

The present paper sheds light on the link between women’s empowerment and husbands’ propensity to split migrate. We ﬁrst build on the strategy developed by Bargain, Loper, and Ziparo(2020), which exploit a natural experiment in Indonesia, namely a series of legal reforms that have fostered women’s access to justice in 2008-10. We take advantage of ethnical heterogeneity in post-marriage residence norms, i.e. matrilocality (a couple lives with the bride’s relatives) and patrilocality (with

1244 million individuals were characterized as international migrants in 2015, i.e. a 41% increase compared to 2000 and a faster growth than the world population growth, while 763 million individuals were internal migrants within their country (UN, 2015). 2These figures are drawn from IFLS data used in the present study. The total rate of families declaring the migration of at least one member between 2000 and 2007 was 13% (among which 6.4 percentage points corresponding to split migration) and 12% between 2007 and 2014 (7.2 ppt for split migration). 3For instance, if the husband migrates alone, he must be insured by his family in the short-run while he can repay later by insuring the family against the risks of undertaking new investments in physical or human capital (Stark and Taylor, 1991). 4The migration literature has acknowledged the collective dimension of the migration decision, but mainly through the lens of remittances motives (see Rapoport and Docquier 2006 for a review).

167 Chapter 3. Women’s Empowerment and Husband’s Migration: Evidence from Indonesia the groom’s family). Our companion paper shows that the reform has exogenously increased empowerment among women of matrilocal ethnic groups relative to their patrilocal counterparts. With the

Indonesia Family Life Survey (IFLS), we use this result to test the role of women’s empowerment on split migration decisions. Our difference-in-difference (DD) approach conveys that husbands of matrilocal ethnicities are more likely to split migrate after the reform than those from patrilocal groups.

Their relative propensity to split migrate increases by 2−3.4 percentage points, i.e. a rise of 41−76%.

Consistently with our interpretation, no such change is found for the probability of migrating jointly.

The intuition of this result goes as follows. More empowered women may have ex-ante gained control over outcomes that are costlier to monitor for husbands in case they migrate. If increased power guarantees that women already control the type of resources that would be tapped in case the husband is gone, then anticipated asymmetry of information and the risk of moral hazard in migration are reduced. As a result, husbands’ migration opportunities are improved and split migration is more likely to occur. There are two collaries that we can test. First, households with empowered women should be more able to cushion a shock such as natural disasters. Following the reform, we find that the relative increase in split-migration rates is indeed driven by matrilocal households living in villages recently exposed to natural shocks. Second, the well-being of matrilocal women should not vary much whether the husband migrates or not, but should be higher than among patrilocal families after reform. Indeed, we confirm that among couples experiencing split migration, matrilocal women observed after reform have a higher autonomy (a relatively higher propensity to work, spend money on food and healthcare and send children to school), suggesting shifts in the control of resources when the husband migrates. Finally, we provide a theoretical framework that rationalizes the central intuition and our empirical findings. We discuss migration decisions in the context of limited commitment issues for migrating households and show that women’s empowerment facilitates migration as husbands need to commit less to accept to leave. It follows that households in which women are more empowered are also more able to cope with risks thanks to higher migration opportunities for husbands.

Our paper contributes to the literature on household decision-making and migration. First, we shed some light on living arrangements in the context of low-income countries, often characterized by the frequent inefﬁcient nature of household decisions (see the critical survey by Baland and Ziparo

2017). While many studies focus on the effect of migration and remittances on the left behind (e.g.

Amuedo-Dorantes and Pozo 2010, Antman 2011a, Acharya and Leon-Gonzalez 2014, Bargain and

168 Chapter 3. Women’s Empowerment and Husband’s Migration: Evidence from Indonesia

Boutin 2015), the decision to remit, and how much to remit, is often suboptimal. Indeed, potential migrants expect a misuse of their transfers by their left-behind relatives (and may even reconsider the mere decision to migrate by anticipation). Many studies estimate the extent of non-cooperative behavior between migrants and the left-behind due to information asymmetries (Chen, 2013, Azam and Gubert, 2005, Antman, 2011b, 2015). Migrants often attempt to monitor their relatives (Chen,

2006, De Laat, 2014) or, when offered a greater degree of control over the money use, opt more for investment and savings rather than for consumption expenditure on their family (Ashraf et al., 2015).

In this context, our results suggest an interesting selection mechanism that impacts upon migration decisions and potentially affects efﬁciency.5

Our ﬁndings also contribute to the large literature exploring the determinants and nature of migration. The bulk of the literature focuses on the role of earnings (e.g. Borjas 1987, Bertoli, Moraga, and

Ortega 2013, Foged 2016) and networks (e.g. Azam and Gubert 2006, McKenzie and Rapoport 2010,

Chort, Gubert, and Senne 2012). Our setting is closely related to the recent studies on migrant selection within the household. In particular, Chort and Senne(2015, 2018) show how households opt for family members who maximize potential remittances. Bratti, Fiore, and Mendola(2016) explore the role of household structure and ﬁnd a limited role of sibship size but a higher propensity to migrate among older and ﬁrst-born siblings. To our knowledge, our paper is one of the very few to study the determinants of husbands’ split-migration, in relationship with women’s empowerment.6

The paper is structured as follows. Section 2 provides insights of the Indonesian cultural and institutional contexts. In Section 3, we describe the data and our empirical strategy. Section 4 presents

5More generally, the asymmetry of information between spouses is studied in family economics as a general source of inefficiency, especially for long-term or irreversible decisions (Lundberg and Pollak, 2003, 2008). It is possibly more acute in the context of migration, all the more so as the difficulty may pertain more to the ability to share information about a person’s preferences (the migrant) than to respect these preferences (Ambler, 2015). Conversely, there are examples where household members are willing to hide information about their own income to their relatives to escape from (sometimes abusive) redistribution (Baland, Guirkinger, and Mali, 2011). Along these lines, our results also contribute to the hold-up problem literature (Grossman and Hart, 1986; Hart and Moore, 1990) as we show that women’s empowerment helps spouses to agree ex-ante on a commitment that lowers anticipated moral hazard and asymmetry of information during husband’s migration, therefore restoring some efficiency in the household’s migration decision. 6Nobles and McKelvey(2015) analyze this question in the context of Mexican migrants in the US, using PROGRESA transfers as a shift in women’s bargaing power. The difficulty is that recipient households do not only experience women empowerment but also a large income effet that may changed the incentive to use migration as an income-generating or diversification strategy. This is not the case in our context sinceboth matrilocal (empowered) and patrilocal women are eligible to the income effect of the reforms under study (fee waiver for legal costs) or subject to a negative income shock (natural disaster) in our heterogeneity analysis. We also suggest a complete analysis of the intra-household mechanisms at stake, which is not present in the aforementioned study. Gemici(2011) is another rare example of study linking household migration decisions and intra-household bargaining. While she models household migration decisions in a dynamic framework with intra-household bargaining, the scope of her study differs from ours as she assesses the implications of joint location constraints on the migration patterns, labor market outcomes, and marital stability of men and women in the United States.

169 Chapter 3. Women’s Empowerment and Husband’s Migration: Evidence from Indonesia the main results and the mechanism behind them. In Section 5, the theoretical framework is used to rationalize our results. Section 6 concludes.

3.2 Background on Social Norms and Legal Reforms

We describe the reforms under study and we provide some background on traditional residence norms in Indonesia. The section ends with a summary of the differential effects of the reforms on women’s empowerment found in our companion paper (Bargain et al., 2020).

3.2.1 National Access-to-Justice Strategy: a Natural Experiment

Indonesia is characterized by an Islamic justice system in which 98% of divorces are pronounced by religious courts (the remaining cases are heard by general courts). They have exclusive jurisdiction over marriage-related cases in general, including divorce, property division, child custody or spousal maintenance. Women’s access to justice is positively related to gender equity developments (Alﬁtri,

2011), especially by making them more assertive about their right to divorce. However, these rights may be constrained by a lack of information, by the cost of court cases or by the social consequences for women who institute divorce proceedings through the formal legal system. These conclusions have been reached by the ‘Access and Equity’ study supported by the Family Court of Australia,

AusAID and other stakeholders. In this context, the Indonesian government has launched the National

‘Access to Justice’ Strategy during 2008-2010, aimed at increasing access to the courts for women and disadvantaged groups (Sumner, Zurstrassen, and Lister, 2011). Supported by AusAID, the Family

Court of Australia and the World Bank, it comprises budgets to waive legal fees, an increased capacity for circuit courts (i.e. courts travelling to subdistricts in order to hold hearings for family law cases in rural and remote areas) and increased legal aid services and assistance to women who cannot afford lawyers.

The effect of this series of reforms has been documented in several reports, showing a signiﬁcant increase in the ability of women, especially in remote areas, to access courts and exercise their rights

(Sumner et al., 2011).7 As illustrated in Figure 3.1, an increase in divorce is observed around the time of the reform. Maybe more impressive is the rise in divorce cases initiated by the wives, which exceed 70% of all divorces in the recent years. In other words, as we document in Bargain et al.(2020)

7The number of people accessing religious courts through fee waiver (resp. circuit courts) has been multiplied by 20 (resp. 6) in 2011 compared to 2007.

170 Chapter 3. Women’s Empowerment and Husband’s Migration: Evidence from Indonesia these reforms have increased women’s marriage outside options by lowering the cost associated with divorce, and subsequently triggered renegotiation of bargaining powers in favor of women within couples remaining married.

By increasing women’s access to the courts, these courting reforms not only increased women’s access to divorce, but they more generally helped women in case of domestic violence and allowed them to break cycles of illegal marriages, divorce and births which prevented them to enforce their rights, for instance their entitlement to healthcare programs for themselves or their children (Sumner and Lindsey, 2011). Further, women’s increased access to courts also allowed them to be legally documented as head of their household (i.e. in case of divorce for example), which can play a major role in Indonesia in them being able to access poverty alleviation programs (subsidised rice and cash transfers), free healthcare access, and access to free education programs for themselves, their children and their extended family (Sumner and Lindsey, 2011). In sum, these reforms might have also triggered husband’s migration through additional channels that would go in the same direction than the main channel we explore (i.e. increase in women’s intra-household bargaining power).

Figure 3.1 – Divorce Trends around Reform Time

171 Chapter 3. Women’s Empowerment and Husband’s Migration: Evidence from Indonesia

3.2.2 Traditional Residence Norms: the Indonesian Context

Post-marital residence norms have been categorized as follows: matrilocality (married couples live with or near the bride’s family), patrilocality (they live with or near the groom’s family), ambilocality

(they can live with or near either spouse’s parents) and neolocality (they can set their own household, i.e. the basis of most developed nations). The rule adopted by households traditionally varies with the ethnic group the household belongs to. We exploit the ethnic diversity, as well as the geographical dispersion of ethnic groups in Indonesia (Figure 3.2) to study the link between the residential norm and household behaviour. In the Indonesian context, informal laws (Adat) shape many aspects of family life and are historically associated with ethnic differences in family-related behaviors including marriage, inheritance, land-holding and dispute resolution. Traditional residence norms are the salient part of Adat rules: women in groups of matrilocal tradition tend to have higher levels of autonomy

(Rammohan and Robertson, 2012). In particular, we show in our companion paper (Bargain et al.,

2020) that women originating from matrilocal ethnic groups were more responsive to the reforms than their patrilocal counterparts. More precisely, they were relatively more likely to divorce following the reforms and, in stable marriages, to renegotiate and gain empowerment. The companion paper documents a substantial increase in ﬁnal say regarding key life decisions (contraception and large household expenditure) as well as in a broad range of well-being measures (health condition, control over fertility, living standards and food expenditure of wives and their children, the value of assets controlled by women). Hereafter, we build on these results to conduct our empirical strategy regarding split-migration decisions.

Figure 3.2 – Village-level Traditional Post-Marital Residence (IFLS data)

172 Chapter 3. Women’s Empowerment and Husband’s Migration: Evidence from Indonesia

3.2.3 Migration in Indonesia

Indonesia is one of the biggest migrant-sending countries in Asia (Hugo, 2009), with 3.4 million international migrants in 2010.8 Most of this international migration takes place within Asia, with

Saudi Arabia and Malaysia being the main destinations of Indonesian migrants. (IOM, 2020). Internal migration outweighs international migration in Indonesia, with an estimated number of 9.8 million temporary internal migrants in 2010 (Sukamdi and Mujahid, 2015). Java is the most attractive island with more than 550,000 in-migrants in 2010, but has also the highest number of out-migrants (nearly a million). Further, migration in Indonesia tends to be mostly rural-urban, with internal migrants increasingly moving further away from their home. These trends are in part explained by transport and communications improvements.9 Importantly, Southern Asian populations, and Indonesian in particular, are particularly vulnerable to natural disasters and climate change. Therefore, internal displacements are very often triggered by such events, and especially by coastal ﬂooding and cyclonic activity Édes, Gemenne, Hill, and Reckien(2012). In 2018, the estimated number of new internal displacements due to natural disasters amounted to 853,000 in Indonesia (IOM, 2020).

3.3 Data and Empirical Strategy

3.3.1 IFLS Data

Data. The empirical analysis draws on data from the Indonesia Familiy Life Survey (IFLS). It contains extensive socioeconomic data at the individual level (including information on individuals’ ethnicity, and marital history) and at household level (including household composition) as well as village-level information. Notably the ethnic composition and prevalent kinship norms (including inheritance and post-marital residence norms) at the village level are provided by Adat experts or community leaders. The IFLS is based on an initial sample representing about 83% of the Indonesian population living in 13 of the 27 Indonesian provinces in 1993. Extensive efforts were provided to track respondants when collecting the data in each of the ﬁve waves (1993, 1997, 2000, 2007 and

2014): 92% of the households are still in the database in the last wave (Strauss, Witoelar, and Sikoki,

2016).

8Data from the Migration Data Portal of the International Organization for Migration (IOM): https: //migrationdataportal.org/?i=stock_abs_&t=2019 9The proportion of urban population has increased from 42.4% in 2000 to 48.8% in 2010, which is partly the outcome of increased migration and, to some extent, to other factors such as the re-classication of rural areas as urban due to development (Sukamdi and Mujahid, 2015).

173 Chapter 3. Women’s Empowerment and Husband’s Migration: Evidence from Indonesia

Selection. Our analysis focus on 2007 and 2014 waves, the two years surrounding the Access-to-

Justice reforms (2008-10). We exclude polygamous households. Migration decisions are recorded as family members migrating between two waves of the IFLS. In our main analysis, we consider husband’s split migration corresponding to migration between 2000-2007 (recorded in 2007) and migration between 2007-2014 (recorded in 2014). Hence, it is important to focus on couples that have been stable over the potential migration period, i.e. between 2000-2007 and between 2007-2014.

To check the parallel trend assumption, we also need stable couples for the periods 1993-2000 (with potential migration recorded in 2000) and 2000-2007. As a further robustness check, we also exploit the panel structure of the IFLS to control for household ﬁxed effects. For this, we select couples that are stable over the whole period 2000 to 2014, using recorded information between 2007-2014 for our main analysis (for placebo checks, we use stable couples over 1993-2007 using data for 2000 and 2007). Note that our favorite approach is the one based on pooled cross-sections since using the panel implies working on an older sample and longer-lasting couples, which necessarily limits the interpretation of the results. Further, using the panel also raises sample selection issues: ﬁrst, as with this sample we are working on stable couples over 2000-2014 for which we have information on husband’s migration over this entire period, we get a sample that is about 35% smaller than the pooled cross-section sample. Second, we might get differential attrition between matrilocal and patrilocal group as some couples with husband’s migration might not be surveyed in the subsequent wave (e.g. if the wife joined him meanwhile and was not traced back). In this case such couples would be dropped from our panel sample.10

Outcome and treatment. Our main outcome variable is husband split migration dummy. As mentioned above, this variable recorded in 2014 indicates whether the husband has migrated without his wife since the preceding wave of the survey, i.e. between 2007 and 2014. The same reasoning applies for the other years. As for treatment, we aim to capture the exogenous change in bargaining power induced by the reform. Hence, we directly use as treatment variable the ethnic-group traditional norm of post-marriage residence, i.e. whether the couple is from patrilocal or matrilocal heritage.11 The traditional norm of each ethnicity is not known and we proxy it follow the methodology used in Bar-

10In our panel sample, matrilocal husband’s propensity to split migrate over 2007-2014 is about 17% smaller than in the pooled cross-sections sample, while it is about 26% for patrilocal, suggesting potential differential attrition. 11As explained in detail in Bargain et al.(2020), we refrain from using households’ actual post-marriage residence for two reasons. First, actual arrangements may be highly correlated with a particular couple’s unobserved heterogeneity and, hence, reflect more than what the norm entails. Second, empowerment isnot only affected by the fact that one lives with her relatives or her partner’s relatives. Ancestral residence norms are a salient feature of ethnic heterogeneity in many dimensions of gender rights and roles.

174 Chapter 3. Women’s Empowerment and Husband’s Migration: Evidence from Indonesia gain et al.(2020) and inspired by Buttenheim and Nobles(2009). 12 We obtain a proportion of about

83% (17%) individuals with a matrilocal (patrilocal) ethnic heritage.

Additional information. For control variables and heterogeneity analyses, we mobilize a range of household characteristics including spouses’ age, education, labor market status, religion, as well as geographical information (notably whether they are Javanese and whether they live close to a district capital, i.e. where courts are located). We also create a dummy for ‘natural disaster’. It takes a value of 1 if the household lives in a village that has experienced a severe shock in the 5 years preceding the survey. We will also check the results of estimations with female and child outcomes as dependent variables, including work (female labor participation dummy), women’s living standards and food consumption, child outcomes (living standards, food consumption, healthcare, education, health) and child investiment (school attendance, school hours, child labor).

Descriptive Statistics. Statistics for couples from matrilocal and patrilocal ethnic traditions are reported in Table 3.A.2 for both 2007 and 2014. The upper panel describes standard socio-demographic characteristics, denoted 푋푖푡 hereafter and used as main controls in our analysis. The rest of the table focuses on additional controls used in the robustness checks. In both years, women of ethnic matrilocal customs tend to be more often urban and muslim and are less likely to work than their patrilocal counterparts. These differences seem to be broadly constant over time, which is reassuring regarding the absence of confounding factors. For instance, the reform could affect muslims more than non- muslims, which would be what we capture by confronting matrilocal and patrilocal groups. It does not seem to be the case (except maybe for rural), as conveyed by the raw DD calculations applied to control variables (last column). Nonetheless, our estimations will control for both the whole set of characteristics 푋푖푡 and their differentiated effect over time. A foretaste of the main results is given in the ﬁrst row: the raw DD indicates that after the reform, the propension to split migrate increases for matrilocal husbands relative to their patrilocal counterparts. DD estimations shall reﬁne these basic calculations.

12To attribute to each woman in our sample the traditional post-marital residence norm of her ethnicity, we first rely on IFLS information about individual ethnicity and, then, identify traditional norms as follows.We categorize villages according to their main ethnicity. For each group of villages of the same prevailing ethnicity, we observe the distribution of Adat experts’ answer to the question about “where the newly married couple lives after wedding according to the traditional law”, as reported in Table 3.A.1. We retain the modal answer: it is systematically matrilocality or patrilocality (rather than neolocality or ambilocality), so that an ethnic group’s traditional norm will be either one or the other option as indicated in the last column. In Bargain et al.(2020) we show that the traditional residence norm is still a very significant predictor of actual household composition in Indonesia in 2014.

175 Chapter 3. Women’s Empowerment and Husband’s Migration: Evidence from Indonesia

3.3.2 Empirical Strategy

Main Estimation. We want to study the effect of an exogenous increase in women’s power within marriage on split migration. To answer this question, we estimate the following equation:

푚푖푡 = 훼 + 훽푃 표푠푡푡 × 푀푎푡푟푖푙표푐푎푙푖 + 푍푖푡 + 휖푖푡 (3.1)

with 푍푖푡 = 휌푃 표푠푡푡 + 휆푀푎푡푟푖푙표푐푎푙푖 + 휑푋푖푡 + 휓푃 표푠푡푡 × 푋푖푡

In this model, 푚푖푡 is a dummy variable taking value 1 if household 푖 reports that the husband has migrated alone during the 7-year period preceding wave 푡. The variable 푃 표푠푡푡 is equal to 1 for the period following the Access-to-Justice reform (year 2014) and 0 for the base period (year 2007).

The variable 푋푖푡 is a vector of controls that include religion (muslim dummy), education (university degree dummy), work, rural area and age groups dummies (using 5-years steps). We do not directly measure the bargaining power of women and refrain from aggregating a set of ﬁnal say answers for that purpose.13 Instead, we draw from the conclusions of our previous study that shows theoretically, and veriﬁes empirically, an increase in the risk of divorce and, for stable marriages, an empowerment among matrilocal women relative to patrilocal women following the reform. Thus, we use the variable

푀푎푡푟푖푙표푐푎푙푖, equal to 1 (0) if the woman’s ethnic group is traditionally matrilocal (patrilocal), as a reduced-form measure of a post-reform change in matrilocal women’s empowerment and autonomy.

Mechanisms. We are interested in the risk-coping function of migration in relation with the bal- ance of power in Indonesian couples. We want to test whether households with more empowered women are also more able to react to shocks via increased male migration. To answer this question, we estimate the following variant of Equation 3.1 that introduces heterogenous effects:

퐷 퐷 푚푖푡 = 훼 + 훽1 푃 표푠푡푡 × 푀푎푡푟푖푙표푐푎푙푖 × 퐷푖푠푎푠푡푒푟푖푡 + 훽0 푃 표푠푡푡 × 푀푎푡푟푖푙표푐푎푙푖 × (1 − 퐷푖푠푎푠푡푒푟푖푡)

+훿푃 표푠푡푡 × 푃 푎푡푟푖푙표푐푎푙푖 × 퐷푖푠푎푠푡푒푟푖푡 + 휂퐷푖푠푎푠푡푒푟푖푡 + 푍푖푡 + 휖푖푡 (3.2)

In this speciﬁcation, 퐷푖푠푎푠푡푒푟푖푡 is the dummy variable indicating if a household has experienced

퐷 퐷 a natural disaster in the 5 years preceding the survey. Comparing 훽1 and 훽0 allows us to test the 13In Bargain et al.(2020), we only use final say variables regarding contraception and large household expenditures, as key dimensions of women’s autonomy. Other aspects are deemed less relevant, such as decisions upon daily purchase and cooking, since they may reflect delegation of responsibility rather than women’s genuine autonomy (Baland, Boltz, Catherine, Seleck, and Ziparo, 2020).

176 Chapter 3. Women’s Empowerment and Husband’s Migration: Evidence from Indonesia migration response among empowered women (i.e. matrilocal women after the reform) in case of a

퐷 climate shock. Comparing 훽1 and 훿 allows checking whether households with empowered women are more able to cushion this shock through split migration compared to patrilocal households. Note that the model controls for 푍푖푡 as usual.

Finally, we investigate the dynamics of women’s outcomes 푦푖푡 in relation with the reform and migration choices. We estimate the following equation:

푚 푚 푦푖푡 = 훼 + 훽1 푃 표푠푡푡 × 푀푎푡푟푖푙표푐푎푙푖 × 푚푖푡+ + 훽0 푃 표푠푡푡 × 푀푎푡푟푖푙표푐푎푙푖 × (1 − 푚푖푡) (3.3) +훿푃 표푠푡푡 × 푃 푎푡푟푖푙표푐푎푙푖 × 푚푖푡 + 휂푚푖푡 + 푍푖푡 + 휖푖푡

with 푍푖푡 as above. The outcome variable 푦푖푡 represents several indicator of women’s and children’s well-being in household 푖 at time 푡. Among empowered women (i.e. matrilocal groups after the reform), there should not be much difference between those who split migrate (푚푖푡=1) and those who do not (푚푖푡=0) if empowerment reduces the extent of information asymmetries leading to suboptimal transfers. We can also check that among households experiencing split migration, empowered

푚 women are better off than patrilocal women after the reform (i.e. 훽1 larger than 훿). Even though we are aware that selection into migration is endogenous (as it varies with matrilocal status and the matrilocals’ response to the reform), this model aims to provide suggestive evidence on the potential mechanisms linking migration and intra-household decision making.

3.4 Results

3.4.1 Main Results on Split-Migration

Baseline results. Table 3.1 reports the effect of belonging to a matrilocal ethnic group after the reform, i.e. the estimates of 훽 in equation (1). The first column reports the raw DD estimate (no controls) that was previously discussed when looking at descriptive statistics. The second model simply controls for 푋푖푡, the third model for both 푋푖푡 and 푃 표푠푡푡 ×푋푖푡, and the fourth one for Javanese and time interaction (Javanese are the largest groups and are matrilocal). All four models convey a significant DD estimate of similar magnitude across specifications, i.e. the split migration rate increases by 2−2.3 percentage points among couples of matrilocal tradition. Relative effects are also reported, namely the estimates over the control group mean outcome prior to the reform. Relative to

177 Chapter 3. Women’s Empowerment and Husband’s Migration: Evidence from Indonesia that backdrop, split migration increases by 42 − 48% (across models) among families of matrilocal ethnic groups after the reform. In the next models, we control for geographical indications, namely whether the household is close to the district capital or whether it has experienced a natural disaster, or for all the controls simultaneously. In these three cases, the estimate increases by around 50%, i.e. split migration relatively increases by 3.1 − 3.4 points.

Table 3.1 – Difference-in-difference Estimations on Husband’s Split Migration

Husband’s Split Migration (1) (2) (3) (4) (5) (6) (7) Post -0.006 -0.008 -0.019** -0.021** -0.002 -0.022 -0.011 (0.007) (0.008) (0.009) (0.009) (0.014) (0.015) (0.017) Post × Matrilocal 0.020** 0.021** 0.023** 0.023** 0.031*** 0.034*** 0.031** (0.008) (0.008) (0.010) (0.011) (0.011) (0.011) (0.012) Matrilocal -0.010 -0.019** -0.020** -0.027*** -0.020* -0.021* -0.029** (0.008) (0.009) (0.009) (0.010) (0.012) (0.012) (0.013) Relative effect 41.9% 44.4% 48.1% 47.9% 71.4% 76.1% 70%

Observations 13,939 13,682 13,682 13,682 10,068 9,881 9,665 R-squared 0.001 0.011 0.012 0.013 0.017 0.016 0.018 Clusters 320 320 320 320 317 318 317 T-Test Equal. (p-val.) 0.088 0.066 0.001 0.000 0.050 0.001 0.025 Controls Yes Yes Yes Yes Yes Yes Post × Controls Yes Yes Yes Yes Yes Javanese Yes Yes Post × Javanese Yes Yes Close Yes Yes Post × Close Yes Yes Natural Disaster Yes Yes Post × Nat. Disaster Yes Yes Difference-in-difference estimations of husband’s split migration on a sample of women in stable couples in 2007-2014 and surveyed in 2014, pooled with a sample of women in stable couples in 2000-2007 and surveyed in 2007. The outcome is a dummy variable indicating women’s husband’s split migration since the preceding wave of the survey (i.e. since 2007 when surveyed in 2014 and since 2000 when surveyed in 2007). Post is equal to 1 for 2014 (post-reform) and 0 for 2007 (pre-reform). ‘Controls’ include a dummy indicating muslim individuals; a dummy indicating individuals holding a university degree; a dummy indicating individuals being currently working; a dummy indicating individuals currently living in a rural area; and age groups dummies (using 5-years steps). ‘Javanese’ is a dummy indicating individuals of Javanese ethnicity. ‘Close’ is a dummy indicating individuals living in 2007 in a village located close to the district capital (i.e. below the 2nd tertile of distance). ‘Natural Disaser’ is a dummy indicating individuals living in a village having experienced a natural disaster in the 5 years preceding the survey. The relative effect is calculated in % of mean outcome for patrilocal group in 2007 (pre-reform). We report thep-values of T-Test of ‘Post’ = ‘Post × Matrilocal.’ Standard errors are reported in brackets and clustered at village of origin level. Significance levels: * p<0.10, ** p<0.05, *** p<0.01.

Robustness checks. We suggest several important checks. First, Table 3.A.3 shows the results of the placebo estimations using 2000-2007. The DD estimate is not significant over this period, confirming that the parallel trend assumption is verified. Second, results above are confirmed using the panel structure of the database. Table 3.A.4 presents estimates in which household fixed-effects are included. As noted, this sample is stable over time and, hence, is 7 years older in 2014 compared

178 Chapter 3. Women’s Empowerment and Husband’s Migration: Evidence from Indonesia to the pre-reform observations. In that sense, we can expect that migration behavior change in a way that relates to household ageing. Nonetheless, the treatment effect remains significant across the various specification and the order of magnitude is also similar (note that the common trend is also verified for this sensitivity analysis, as shown in Table 3.A.5). This suggests that our main results are not confounded by time-invariant unobservables.

Table 3.2 – Difference-in-difference Estimations on Joint Migration

Joint Migration (1) (2) (3) (4) (5) (6) (7) Post -0.000 -0.008 0.033 0.038* 0.024 0.014 0.020 (0.010) (0.010) (0.023) (0.020) (0.022) (0.019) (0.014) Post × Matrilocal 0.000 0.002 0.011 0.003 0.006 0.015 0.004 (0.011) (0.011) (0.014) (0.015) (0.013) (0.012) (0.012) Matrilocal -0.005 -0.021 -0.026 -0.005 -0.008 -0.008 0.002 (0.013) (0.016) (0.016) (0.017) (0.011) (0.011) (0.012) Rel. effect 0.1% 1.8% 9.5% 2.1% 12.9% 33.2% 9.2%

Observations 14,876 14,390 14,390 14,390 10,577 10,393 10,155 R-squared 0.000 0.058 0.059 0.061 0.054 0.046 0.047 Clusters 321 320 320 320 317 318 317 T-Test Equal. (p-val.) 0.996 0.616 0.397 0.140 0.423 0.970 0.353 Controls Yes Yes Yes Yes Yes Yes Post × Controls Yes Yes Yes Yes Yes Javanese Yes Yes Post × Javanese Yes Yes Close Yes Yes Post × Close Yes Yes Natural Disaster Yes Yes Post × Nat. Disaster Yes Yes Difference-in-difference estimations of joint migration on a sample of women in stable couplesin 2007-2014 and surveyed in 2014, pooled with a sample of women in stable couples in 2000-2007 and surveyed in 2007. The outcome is a dummy variable indicating joint migration (i.e. spouses migrate together) since the preceding wave of the survey (i.e. since 2007 when surveyed in 2014 and since 2000 when surveyed in 2007). Post is equal to 1 for 2014 (post-reform) and 0 for 2007 (pre-reform). ‘Controls’ are described in Table 3.1. The relative effect is calculated in % of mean outcome for patrilocal group in 2007 (pre-reform). We report the p-values of T-Test of ‘Post’ = ‘Post × Matrilocal.’ Standard errors are reported in brackets and clustered at village of origin level. Significance levels: * p<0.10, ** p<0.05, *** p<0.01.

Finally, to understand whether changes in migration patterns are speciﬁc to the split migration phenomenon or can be generalized to joint migration decisions, we estimate equation (1) using a dummy for joint migration as dependent variable (equal to 1 if the household reports a joint migration of the wife and husband since the previous IFLS wave). Results are presented in Table 3.2. There is no sign of a systematic and signiﬁcant change in migration patterns among matrilocal households following the reform (or among patrilocals). This result conveys that the empowerment of matrilocal women due to the reform for matrilocal women is not associated with an overall migration effect but only with an increase in husbands’ split migration.

179 Chapter 3. Women’s Empowerment and Husband’s Migration: Evidence from Indonesia

3.4.2 Interpretation and Heterogeneity

Interpretation. We provide the intuitions behind our main result, which will be further developed in the theoretical framework of the next section. Previous baseline results and sensitivity analyses convey that a relative increase in husbands’ split migration takes place in households where women are empowered thanks to the reform, namely families of matrilocal tradition. The role of empowerment can be explained by the fact that more empowered women may have ex-ante gained control over outcomes that are costlier to monitor for husbands in case they migrate. As recalled in the introduction, the recent literature on migration point to non-cooperative behavior due to information asymmetry and limited commitment in split-migration families. Left-behind relatives (often the wife) tend to over-invest in goods that are favorable to them but not observable by the migrant (often the husband), as shown for instance in Chen(2013). If increased power guarantees that women already control the type of resources that would be tapped in case the husband is gone, then anticipated asymmetry of information and the risk of moral hazard in migration are reduced. As a result, husbands’ migration opportunities are improved and split migration is more likely to occur. There are two nec- essary conditions to this interpretation that we are able to test. First, households with empowered women should be more able to cushion a shock such as natural disasters. Second, their position in marriage should not vary much whether the husband migrates or not, but should be higher than among patrilocal families after reform.

Exposure to natural shocks and migration. As recalled in section 2.3, Indonesia is particular vulnerable to natural disasters and climate events are a major cause of internal and external migration.

We conﬁrm that these are frequent events using the IFLS data: among matrilocal women observed after the reform, around half of them live in a village that has experienced a natural shock in the past

5 years. We then test our ﬁrst condition, corresponding to equation (2). Estimates are presented in

Table 3.3. We see that the increased likelihood of husbands’ split migration in couples with empowered women (post-reform matrilocals) is essentially driven by those experiencing a natural disaster.

This is not so surprising: as justified, climate shocks are a major source of migration in Indonesia and, according to our conjecture, empowered women are most likely to be able to trigger split migration to cushion this type of shock. For them, the rate of split migration increases by around 2.1 − 3 percentage points relative to patrilocal households and is significant. In matrilocal households not recently exposed to natural disasters, the effect is smaller (i.e. around 1.1 − 1.9), not significant and possibly

180 Chapter 3. Women’s Empowerment and Husband’s Migration: Evidence from Indonesia reflecting other (and less important) shocks.14 Then, comparing the first and third rows conveys that among households experiencing a natural disaster, those with empowered women are significantly more able to cushion the shock via split-migration (t-tests are reported and p-values are systematically close to zero). These results are suggestive of the ex-post risk-coping nature of migration and the fact that couples with more empowered women are also more able to activate this strategy.

Empowerment, migration and women’s and child’s outcomes. The other consequence of our interpretation of intra-household mechanisms pertains to living conditions of women after the reform. Namely, empowered women should be better off than patrilocal women when husbands migrate because women have already gained access to the resources that would otherwise be source of inefﬁciency for the migration decision. In Table 3.4, we provide estimates using a variety of outcomes including women’s autonomy (captured by a labor market status), her well-being in marriage

(standard of living, food consumption, healthcare and perception of family life) and child outcomes

(standard of living, food consumption, healthcare, health, education) or investment in children (school attendance, child labor). We provide three types of heterogeneous effects, interpreted in reference to pre-reform patrilocal households, as laid out in equation (3). Among matrilocal women observed post-reform, we observe very little differences between those with a migrating husbands and those without.15 Then, if we compare the first and third rows, i.e. matrilocal and patrilocal households with migrating husbands, the former are significantly better off (t-test are provided and show significant differences for most of the 14 outcomes). Patrilocal households with a migration experience post- reform are usually not different from the control group (patrilocals before reform). These results are purely indicative given the endogeneity of migration but confort our interpretation that empowered women have already gained control over the resources that typically lead to suboptimal migration behavior.

14We check that the absence of a significant increase in split-migration for the latter is not due to a differential impact of the reform across matrilocal women depending on the regions where they live in terms of climat shocks. To check this, we replicate the approach of Bargain et al.(2020) while introducing this heterogeneity and find no difference between exposed and non-exposed matrilocal women in that respect. 15Women’s and children’s outcomes of the former are slightly worse, given that they may experience an adverse shock (leading to the migration of husbands), while child labor shows a reverse trend (as an alternative source of income, it is logically less used when migration is a prefered strategy). Yet coefficients are rarely significantly different.

181 Chapter 3. Women’s Empowerment and Husband’s Migration: Evidence from Indonesia

Table 3.3 – DD Estimations on Husband’s Split Migration - Heterogeneity by Natural Disaster

Husband’s Split Migration (1) (2) (3) (4) (5) (6) Post × Matri. × Disaster 0.021 0.025* 0.030* 0.026* 0.027* 0.023 (0.015) (0.014) (0.016) (0.016) (0.015) (0.016) Post × Matri. × No Disaster 0.011 0.014 0.019 0.016 0.016 0.013 (0.014) (0.013) (0.015) (0.015) (0.015) (0.015) Post × Patri. × Disaster -0.021 -0.016 -0.018 -0.021 -0.021 -0.024 (0.017) (0.015) (0.015) (0.015) (0.016) (0.016)

Observations 9,996 9,881 9,881 9,881 9,665 9,665 R-squared 0.002 0.015 0.016 0.018 0.018 0.019 Clusters 318 318 318 318 317 317 T-Test P×M×Disaster = P×P×Disaster 0.000 0.000 0.000 0.001 0.000 0.001 Controls Yes Yes Yes Yes Yes Post × Controls Yes Yes Yes Yes Javanese Yes Yes Post × Javanese Yes Yes Close Yes Yes Post × Close Yes Yes Difference-in-difference estimations of husband’s split migration on a sample of women in stable couples in 2007-2014 and surveyed in 2014, pooled with a sample of women in stable couples in 2000-2007 and surveyed in 2007. The outcome is a dummy variable indicating women’s husband’s split migration since the preceding wave of the survey (i.e. since 2007 when surveyed in 2014 and since 2000 when surveyed in 2007). Post is equal to 1 for 2014 (post-reform) and 0 for 2007 (pre-reform). ‘Nat. Dis.’ is a dummy indicating individuals living in a village having experienced a natural disaster in the 5 years preceding the survey, while ‘No Nat. Dis.’ is a dummy indicating the opposite. ‘Matri’ is a dummy indicating individuals of matrilocal ethnicity; while ‘Patri’ is a dummy indicating individuals of patrilocal ethnicity. Other variables are described in Table 3.1. We control for dummies ‘Matrilocal’ and ‘Nat. Dis.’ in all our regressions in this table. We report the p-values of T-Test ‘Post × Matrilocal × Nat. Disaster’ = ‘Post × Patrilocal × Nat. Disaster’. Standard errors are reported in brackets and clustered at village of origin level. Significance levels: * p<0.10, ** p<0.05, *** p<0.01.

182 Chapter 3. Women’s Empowerment and Husband’s Migration: Evidence from Indonesia

Table 3.4 – DD Estimations on Female and Child Outcomes - Heterogeneity by Migration Decision

Std of Food Family Ch. Std Ch. Food Work Healthcare Living Conso Life Living Conso (1) (2) (3) (4) (5) (6) (7) Post × Matri. × Mig. -0.066 0.098 0.102 0.151** 0.106* 0.089 0.133* (0.046) (0.069) (0.065) (0.071) (0.064) (0.079) (0.072) Post × Matri. × No Mig. -0.004 0.171*** 0.171*** 0.197*** 0.170*** 0.184*** 0.173*** (0.022) (0.045) (0.046) (0.048) (0.043) (0.050) (0.049) Post × Patri. × Mig. -0.191*** -0.149 -0.270*** -0.234** -0.134* -0.128 -0.119 (0.068) (0.096) (0.096) (0.093) (0.077) (0.139) (0.117)

Observations 13,682 13,453 13,453 13,446 13,454 8,774 8,775 R-squared 0.058 0.060 0.069 0.064 0.057 0.068 0.087 Clusters 320 320 320 320 320 319 319 T-Test P×M×Mig = P×P×No Mig 0.070 0.020 0.000 0.000 0.010 0.155 0.070 Ch. Ch. Ch. Ch. Atten- Ch. Hours Ch. Hours Ch. Work Healthcare Education Health ding School at school at Work (8) (9) (10) (11) (12) (13) (14) Post × Matri. × Mig. 0.158* 0.109 0.141* 0.000 2.964** -0.153** -0.950 (0.084) (0.080) (0.077) (0.038) (1.315) (0.064) (0.792) Post × Matri. × No Mig. 0.196*** 0.136*** 0.0584 0.0252 3.571*** -0.0712* -0.417 (0.057) (0.050) (0.041) (0.025) (0.830) (0.039) (0.512) Post × Patri. × Mig. -0.106 -0.037 -0.067 -0.028 -1.196 -0.002 -0.986 (0.101) (0.114) (0.123) (0.066) (1.838) (0.063) (0.773)

Observations 8,774 8,752 9,962 9,928 9,928 9,167 9,167 R-squared 0.076 0.067 0.103 0.069 0.088 0.050 0.040 Clusters 319 319 320 320 320 320 320 T-Test P×M×Mig = P×P×No Mig 0.025 0.216 0.072 0.673 0.019 0.027 0.962 Controls Yes Yes Yes Yes Yes Yes Yes Post × Controls Yes Yes Yes Yes Yes Yes Yes Linear estimations on a sample of women in stable couples in 2007-2014 and surveyed in 2014, pooled with a sample of women in stable couples in 2000-2007 and surveyed in 2007. ‘Work’ is a dummy indicating a woman currently working; ‘Std of Living’, ‘Food Conso’, ‘Healthcare’, ‘Family Life’, ‘Ch. Std of Living’, ‘Ch. Food Conso’, ‘Ch. Healthcare’ and ‘Ch. Education’ are categorical variables built from answers (1: It is less than adequate for my needs, 2: It is just adequate for my needs or 3: It is more than adequate for my needs) to questions on subjective well-being regarding standard of living, food consumption, healthcare, family life, children standard of living, children food consumption, children healthcare and children education respectively. Outcomes in columns 10-14 are variables averaged among children in a household. ‘Ch. Health’ is a categorical variable relating to the evaluation of children’s health compared to 12 months ago, ranging from 1: Much Worse to 5: Much Better; ‘Ch. Attending School’ is a dummy indicating whether a child is currently attending school; ‘Ch. Hours at school’ is the number of effective hours a child attended school in the last week; ‘Ch. Work’ is a dummy indicating whether a child ever worked; and ‘Ch. Hours at Work’ is the number of hours a child spent working in the last week. Post is equal to 1 for 2014 (post-reform) and 0 for 2007 (pre-reform). ‘Mig’ is a dummy indicating women with husband having split migrated since last survey (i.e. in last 7 years), while ‘No Mig.’ is a dummy indicating the opposite. We control for dummies ‘Matrilocal’ and ‘Mig’ in all our regressions in this table. Other variables are described in Table 3.1. We report the p-values of T-Test ‘Post × Matrilocal × Mig’ = ‘Post × Patrilocal × Mig’. Standard errors are reported in brackets and clustered at village of origin level. Significance levels: * p<0.10, ** p<0.05, *** p<0.01.

183 Chapter 3. Women’s Empowerment and Husband’s Migration: Evidence from Indonesia

3.4.3 Mechanisms: Theoretical Background

We ﬁnally present a simple theoretical framework to elucidate why an increase of women bargaining power may lead to an increase of spouses’ split migration. We take a representative household composed by a husband and a wife (푗 = 퐻, 푊 respectively), possibly from matrilocal or patrilocal origin (ℎ = 푀, 푃 respectively). They are endowed with income 푦푗ℎ and live two periods: they decide about their location in the ﬁrst period and consume in the second period.16

Preferences and Budget Constraints. Both spouses derive utility from their own private consumption 푐푗ℎ and the joint consumption of a public good 푄ℎ. We write spouse 푗’s utility function under joint residence or split migration respectively as:

푗ℎ 푗ℎ ℎ 푗ℎ 푗ℎ ℎ 푗ℎ 푗ℎ 푗ℎ 푗ℎ 푗ℎ 푗ℎ 푈푗표푖푛푡(푐 , 푄 ) = 푢 (푐 , 푄 ) and 푈푠푝푙푖푡(푐 , 푥 ) = 푢 (푐 , 푥 ) with 푥푗ℎ denoting each spouse’s separate contribution to the public good. When spouses live together,

ℎ 푊 ℎ 퐻ℎ 푥푊 ℎ+푥퐻ℎ 푄 = 푓 (푥 , 푥 ) > 2 meaning that there are economies of scales of being together. The household budget constraint in each period is as follows:

퐻ℎ 퐻ℎ 푊 ℎ 푊 ℎ 퐻ℎ 푊 ℎ 푐푡 + 푥푡 + 푐푡 + 푥푡 ≤ 푦푡 + 푦푡

Consumption decisions. In the model, inefficiency comes from the migration decision, as defined later. Within each state of the second stage, i.e. split-migration or leaving together, household’s consumption decisions can be modeled as an efficient process (Chiappori, 1988, Chiappori et al.,

1992). Thus, consumption allocations when spouses live together are represented as following the maximization of a household welfare function:

ℎ 퐻ℎ 푊 ℎ 퐻ℎ 퐻ℎ ℎ 퐻ℎ 푊 ℎ 푊 ℎ 푊 ℎ ℎ 푉푗표푖푛푡 = 휇(푝, 푦푗표푖푛푡+푦푗표푖푛푡, z푗표푖푛푡)[푢 (푐 , 푄 )]+(1−휇(푝, 푦푗표푖푛푡, 푦푗표푖푛푡, z푗표푖푛푡))[푢 (푐 , 푄 )]

16Note that, while we focus on consumption to model household behaviour, women’s empowerment should correlate with other dimensions such as empowered women adopting more promiscuous sexual behaviours, therefore lowering anticipated moral hazard regarding wife’s faithfulness during husband’s split migration. In the end, this would trigger husband’s split migration. This mechanism would point in the same direction that our main mechanism.

184 Chapter 3. Women’s Empowerment and Husband’s Migration: Evidence from Indonesia

퐻ℎ 푊 ℎ where 휇(푝, 푦푗표푖푛푡, 푦푗표푖푛푡, z푗표푖푛푡) is the Pareto weight of the husband in the couple, determined by prices, incomes and a vector of distribution factors z푗표푖푛푡. When spouses opt for split migration, consumption allocations stem from the maximization of the following function:

ℎ 퐻ℎ 푊 ℎ 퐻ℎ 퐻ℎ 퐻ℎ 퐻ℎ 푊 ℎ 푊 ℎ 푊 ℎ 푊 ℎ 푉푠푝푙푖푡 = 휇(푝, 푦푠푝푙푖푡 +푦푠푝푙푖푡, z푠푝푙푖푡)[푢 (푐 , 푥 )]+(1−휇(푝, 푦푠푝푙푖푡, 푦푠푝푙푖푡, z푠푝푙푖푡))[푢 (푐 , 푥 )]

Whether the husband migrates or not, the consumption allocations in equilibrium are determined by the following optimality condition:

퐻ℎ 휕푢푡 퐻ℎ 휕푐푡 1 − 휇푡 푊 ℎ = 휕푢푡 휇푡 푊 ℎ 휕푐푡

implying that the higher the pareto weight of one spouse, the higher his/her private consumption, given total household income. To know whether each spouse prefer split migration or to stay together, we can compare the evolution of the indirect utility of each spouse in both situations.

Migration Decision. We model location decision following Lundberg and Pollak(2003). Spouses must agree on whether the household opt for split-migration or not. Both spouses must prefer migration for the husband to leave. Each spouse’s preferences regarding migration depends on the second stage consumption allocations, so split migration is possible if there exists a feasible resource sharing outcome that makes both spouses better off in migration rather than in status quo. When the husband migrates, total household income increases (and possibly compensate past income shocks due to natural disasters). Whether a Pareto improving allocation exists depends on the public good production or, said differently, on the economies of scale of the two spouses living together.17 We focus on cases in which such an allocation exists, so it is in the interest of the household to migrate.

푗ℎ ℎ ℎ 퐻ℎ 푊 ℎ We deﬁne 푣푡 (휇푡, 푌푡 ) the indirect utility of spouse 푗 in location 푡, where 푌푡 = 푦푡 + 푦푡 . 푗ℎ 푗ℎ ℎ 푗ℎ ℎ A spouse prefers migration if Δ푣푠푝푙푖푡−푗표푖푛푡 = 푣 (휇푠푝푙푖푡, 푌푠푝푙푖푡) − 푣 (휇푗표푖푛푡, 푌푗표푖푛푡) > 0. The difference between the two indirect utilities, for spouse 푗, can be written as:

휕푣푗ℎ(휇, 푌 ℎ) 휕푣푗ℎ(휇, 푌 ℎ) Δ푣푗ℎ = Δ휇 + Δ푌 ℎ . 푠푝푙푖푡−푗표푖푛푡 휕휇 푠푝푙푖푡−푗표푖푛푡 휕푌 ℎ 푠푝푙푖푡−푗표푖푛푡

17If the economies of scale are too important, a Pareto improving allocation does not exist.

185 Chapter 3. Women’s Empowerment and Husband’s Migration: Evidence from Indonesia

Given our assumptions, to know whether a spouse prefers migration or not, we must compare her

Pareto weight in the two situations. Since the marginal utility with respect to the weight has opposite sign for husband and wife, it is necessarily the case that one spouse loses some power after migration; whether the loosing spouse is the husband or the wife depends on the sign of Δ휇푠푝푙푖푡−푗표푖푛푡.

The Pareto weight is determined by the outside options of both husband and wife: as discussed in Browning, Chiappori, and Weiss(2014), the outside option is either divorce or a non-cooperative equilibrium in marriage. Migration is assumed to have a positive effect on the husband’s outside option by increasing his earning opportunities. At the same time, as shown by Chen(2013) and De Laat

(2014), it also increases the control of the wife on the allocation of resources for the public goods of the staying household and, thus, her outside option in the non-cooperative equilibrium. Therefore, we are agnostic about the sign of change in Pareto weights.

However, if commitment in the household was always possible, the winning spouse (the one for which the Pareto weight increases) could always assure to the other spouse to have at least as much utility in migration as in the joint location. Commitment seems feasible when the winning spouse is the husband, as he can commit to transfer resources to the wife on a regular basis. Instead, commitment seems more difﬁcult on the wife’s side, as the allocation of resources in the household is difﬁcult to observe by the husband (Chen, 2013, De Laat, 2014).

More formally, following Chen(2013) we can redeﬁne the Pareto weight in a household of type

ℎ as follows: 휇ℎ = 휇푐,ℎ + 휇˜ℎ where 휇푐 is the contractible part of the Pareto weight and 휇˜ is the less-contractible one. We can now introduce our ﬁrst and second assumptions.

Assumption I: When divorce is possible, 휇푃 > 휇푀 , with 휇˜푃 > 휇˜푀 .

This means that, after the reform that increases access to divorce, the Pareto weight of patrilocal husbands is higher than that of matrilocal ones (Bargain et al., 2020), and that the change happens in dimensions that are costlier to monitor, even when the husband is present.

ℎ ℎ Assumption II: Δ휇푠푝푙푖푡−푗표푖푛푡 = −휇˜ . The second assumption means that when the husband migrates, he loses all the non-contractible part of the Pareto weight.

Matrilocal men start from a worse off situation than patrilocal men, implying that the share of the

Pareto weight that need to be contractible for them to be willing to migrate is smaller. Furthermore, due to Assumption II, the decline in the Pareto weight following migration is smaller for matrilocal

186 Chapter 3. Women’s Empowerment and Husband’s Migration: Evidence from Indonesia men than for patrilocal ones. These two aspects taken together imply that migration need to assure a smaller increase in total household income for matrilocal men to be willing to migrate. This proves the following result:

Prediction 3. Matrilocal men accept to migrate more often than patrilocal men.

3.5 Conclusions

In this paper, we have explored how women’s empowerment may foster the migration of husbands alone (split migration). This question is rarely explored due to the potential presence of unobserved variables that may affect both husband’s attitudes toward women and their propensity to split migrate.

To deal with this endogeneity issue, we exploit an exogenous improvement in women’s position in marriage driven by a serie of national level policies. These policies, which ease women’s access to justice and facilitate divorce, beneﬁt especially to women from matrilocal ethnic groups who tend to experience a gain in empowerment and well-being. The present paper shows that relative to patrilocal families, they are also concerned by a relative increase in husbands’ split migration. Difference-in- differences estimation point to a sizeable effect, i.e. a relative increase in split migration rates by

2-3.4 percentage points (i.e. an increase of 41-76%).

We suggest a heterogeneous analysis and a theoretical framework that attempt to shed some light on the main mechanisms explaining this result. The role of empowerment can be explained by the fact that more empowered women may have ex-ante gained control over outcomes that are costlier to monitor for husbands (in case they migrate), so the anticipated asymmetry of information and risks of moral hazard in migration are reduced. As a result, husbands’ opportunity costs of migrating are lower and split migration is more likely to occur among these couples. This interpretation is corroborated by the correlation between different measures of women autonomy, well-being and investment in children and split-migration among matrilocal and patrilocal women: matrilocal women in migrating households show indeed a better position in marriage. Finally, women’s empowerment allows restoring some efﬁciency. Indeed, heterogeneous analyses show that the relative increase in husband’s split migration among post-reform matrilocal households is concentrated among those who have experienced a recent natural disaster. That is, women’s empowerment increases the possibility to cushion shocks through migration strategies.

187 Chapter 3. Women’s Empowerment and Husband’s Migration: Evidence from Indonesia

Further work should expand these results and test the role of empowerment in other settings. As motivated in the introduction, migration is historically rooted and widespread as a relevant strategy to cope with adverse shocks in poor households around the world. It has potential numerous beneﬁts for populations left behind and in contexts where ﬁnancial institutions and credit markets are lacking.

However, the circumstances in which migration can operate in an efﬁcient way and maximize the well- being of both migrants and their families are still broadly unknown. Inefﬁciency due to information asymmetry is well documented but the intra-household mechanisms driving migration choices – the mere migration decision and who to send abroad – are complex. We hope that the present paper contributes to a literature that is still in its infancy. Moreover, this study connects to another nascent literature, which examines how cultural norms may interact with development and gender policies.

Recent papers indeed show how educational programs (Ashraf, Bau, Nunn, and Voena, 2020), wealth transmission policies (La Ferrara and Milazzo, 2017) or legal reforms (Bargain et al., 2020) compound with traditional norms in a way that make them effective only for some segments of the population.

The present paper additionally shows that the ability to use migration as an income-generating and diversiﬁcation device also depends on the cultural, often ethnic context. Precisely, understanding how different family structures – with different traditional norms and different balances of power – shape migration implicit contracts within households, and subsequently drive self-selection patterns into migration, is a promising avenue of research.

188 References

Acharya, C. P. and R. Leon-Gonzalez (2014). How do migration and remittances affect human capital investment? the effects of relaxing information and liquidity constraints. Journal of Development Studies 50(3), 444–460.

Alﬁtri (2011). Legal reform project, access to justice and gender equity in indonesia. Indonesian Journal of International Law 9, 292.

Ambler, K. (2015). Don’t tell on me: Experimental evidence of asymmetric information in transnational households. Journal of Development Economics 113, 52–69.

Amuedo-Dorantes, C. and S. Pozo (2010). Accounting for remittance and migration effects on children’s schooling. World Development 38(12), 1747–1759.

Amuedo-Dorantes, C. and S. Pozo (2011, May). Remittances and income smoothing. American Economic Review 101(3), 582–87.

Antman, F. M. (2011a). The intergenerational effects of paternal migration on schooling and work: What can we learn from children’s time allocations? Journal of Development Economics 96(2), 200–208.

Antman, F. M. (2011b). International migration and gender discrimination among children left behind. The American Economic Review 101(3), 645–649.

Antman, F. M. (2015). Gender discrimination in the allocation of migrant household resources. Jour- nal of population economics 28(3), 565–592.

Ashraf, N., D. Aycinena, C. Martínez A, and D. Yang (2015). Savings in transnational households: a ﬁeld experiment among migrants from el salvador. Review of Economics and Statistics 97(2), 332–351.

Ashraf, N., N. Bau, N. Nunn, and A. Voena (2020). Bride price and female education. Journal of Political Economy.

Azam, J.-P. and F. Gubert (2005). Those in kayes. Revue économique 56(6), 1331–1358.

Azam, J.-P. and F. Gubert (2006). Migrants’ remittances and the household in africa: a review of evidence. Journal of African Economies 15(suppl 2), 426–462.

Baland, J.-M., M. Boltz, G. Catherine, M. Seleck, and R. Ziparo (2020). Optimal intra-household decision structure. Working Paper.

Baland, J.-M., C. Guirkinger, and C. Mali (2011). Pretending to be poor: Borrowing to escape forced solidarity in cameroon. Economic Development and Cultural Change 60(1), 1–16.

Baland, J.-M. and R. Ziparo (2017). Intra-household bargaining in poor countries. WIDER Working Paper.

189 References

Bargain, O. and D. Boutin (2015). Remittance effects on child labour: Evidence from burkina faso. The Journal of Development Studies 51(7), 922–938.

Bargain, O., J. Loper, and R. Ziparo (2020). Traditional norm, access to divorce and women’s empowerment: Evidence from indonesia. Working Paper.

Bertoli, S., J. F.-H. Moraga, and F. Ortega (2013). Crossing the border: Self-selection, earnings and individual migration decisions. Journal of Development Economics 101, 75–91.

Borjas, G. J. (1987). Self-selection and the earnings of immigrants. The American Economic Re- view 77(4), 531–553.

Bratti, M., S. Fiore, and M. Mendola (2016). Family size, sibling rivalry and migration: Evidence from mexico. Working Paper.

Browning, M., P.-A. Chiappori, and Y. Weiss (2014). Economics of the Family. Cambridge University Press.

Buttenheim, A. M. and J. Nobles (2009). Ethnic-based nuptial regimes and marriage behaviour. Population studies 63(3), 277.

Chen, J. J. (2006). Migration and imperfect monitoring: implications for intra-household allocation. The American Economic Review 96(2), 227–231.

Chen, J. J. (2013). Identifying non-cooperative behavior among spouses: child outcomes in migrant- sending households. Journal of Development Economics 100(1), 1–18.

Chiappori, P.-A. (1988). Rational household labor supply. Econometrica, 63–90.

Chiappori, P.-A. et al. (1992). " Collective" Models of Household Behavior: The Sharing Rule Ap- proach. DELTA.

Chort, I., F. Gubert, and J.-N. Senne (2012). Migrant networks as a basis for social control: Re- mittance incentives among senegalese in france and italy. Regional Science and Urban Eco- nomics 42(5), 858–874.

Chort, I. and J.-N. Senne (2015). Selection into migration within a household model: Evidence from senegal. World Bank Economic Review 29(suppl_1), S247–S256.

Chort, I. and J.-N. Senne (2018). You’ll be a migrant my son: Accounting for migrant selection within the household. Economic Development and Cultural Change, Forthcoming 66(2), 217–263.

De Laat, J. (2014). Household allocations and endogenous information: The case of split migrants in kenya. Journal of Development Economics 106, 108–117.

Édes, B. W., F. Gemenne, J. Hill, and D. Reckien (2012). Addressing climate change and migration in asia and the paciﬁc. Asian Development Bank.

Foged, M. (2016). Family migration and relative earnings potentials. IZA Discussion Paper.

Gemici, A. (2011). Family migration and labor market outcomes. Working Paper.

Grossman, S. J. and O. D. Hart (1986). The costs and beneﬁts of ownership: A theory of vertical and lateral integration. Journal of Political Economy 94(4), 691–719.

Hart, O. and J. Moore (1990). Property rights and the nature of the ﬁrm. Journal of Political Econ- omy 98(6), 1119–1158.

190 References

Hugo, G. (2009). Best practice in temporary labour migration for development: A perspective from asia and the paciﬁc. International migration 47(5), 23–74.

IOM (2020). World migration report 2020. Policy Report.

Lundberg, S. and R. Pollak (2003, September). Efﬁciency in Marriage. Review of Economics of the Household 1(3), 153–167.

Lundberg, S. and R. Pollak (2008). Family Decision Making, pp. 1–8.

McKenzie, D. and H. Rapoport (2010). Self-selection patterns in mexico-us migration: the role of migration networks. The Review of Economics and Statistics 92(4), 811–821.

Nobles, J. and C. McKelvey (2015). Gender, power, and emigration from mexico. Demography 52(5), 1573–1600.

Rammohan, A. and P. Robertson (2012). Do kinship norms inﬂuence female education? evidence from indonesia. Oxford Development Studies 40(3), 283–304.

Rapoport, H. and F. Docquier (2006). The economics of migrants’ remittances, Volume 1. Chap 17, Handbook on the Economics of Giving, Reciprocity and Altruism (Kolm, S. and Ythier, J. M., ed., Elsevier.

Stark, O. and J. E. Taylor (1991). Migration incentives, migration types: The role of relative depriva- tion. The economic journal 101(408), 1163–1178.

Strauss, J., F. Witoelar, and B. Sikoki (2016). The ﬁfth wave of the indonesia family life survey: Overview and ﬁeld report.

Sukamdi and G. Mujahid (2015). Internal migration in indonesia. UNFPA Indonesia Monograph Series.

Sumner, C. and T. Lindsey (2011). Courting reform: Indonesia’s islamic courts and justice for the poor. International Journal for Court Administration 4(1), 3–16.

UN (2015). Trends in international migration, 2015. United Nationa, Department of Economic and Social Affairs, Population Division, Technical Paper No. 2015/4.

191 References

192 Appendix

3.A Descriptive Statistics and Robustness Checks

Table 3.A.1 – Determination of Traditional Post-Marital Residence Norm by Ethnicity

193 References

Table 3.A.2 – Descriptive Statistics of Main Variables

2007 2014 Raw Matri. Patri. Diff. Matri. Patri. Diff. DD

Main Outcome Husband’s Split Migration 0.038 0.048 -0.010 0.053 0.043 0.010 0.020** (0.192) (0.215) (0.007) (0.223) (0.202) (0.006) (0.010) Number of observations 4964 1012 5976 6343 1363 7706 13682 Prop. Matri/Patri 83.1% 16.9% 82.3% 17.7%

Main Control Variables Age Category 6.143 6.286 -0.142* 6.080 6.073 0.007 0.149 (2.331) (2.425) (0.080) (2.294) (2.367) (0.069) (0.106) University 0.051 0.043 0.007 0.082 0.081 0.001 -0.006 (0.220) (0.204) (0.009) (0.275) (0.272) (0.007) (0.011) Work 0.624 0.700 -0.075*** 0.657 0.721 -0.064*** 0.011 (0.484) (0.459) (0.016) (0.475) (0.449) (0.014) (0.022) Rural 0.497 0.533 -0.035** 0.411 0.509 -0.099*** -0.063*** (0.500) (0.499) (0.017) (0.492) (0.500) (0.015) (0.023) Muslim 0.964 0.559 0.404*** 0.968 0.578 0.390*** -0.014 (0.187) (0.497) (0.009) (0.175) (0.494) (0.008) (0.012) Number of observations 4964 1012 5976 6343 1363 7706 13682 Prop. Matri/Patri 83.1% 16.9% 82.3% 17.7%

Additional Controls Javanese 0.549 0.000 0.549*** 0.546 0.000 0.546*** -0.003 (0.498) (0.000) (0.016) (0.498) (0.000) (0.014) (0.021) Number of observations 4964 1012 5976 6343 1363 7706 13682 Prop. Matri/Patri 83.1% 16.9% 82.3% 17.7%

Close District Capital 0.582 0.588 -0.007 0.578 0.561 0.017 0.024 (0.493) (0.492) (0.019) (0.494) (0.496) (0.018) (0.026) Number of observations 3843 797 4640 4473 955 5428 10068 Prop. Matri/Patri 82.8% 17.2% 82.4% 17.6%

Natural Disaster 0.563 0.576 -0.013 0.559 0.549 0.010 0.022 (0.496) (0.495) (0.019) (0.497) (0.498) (0.018) (0.026) Number of observations 3843 797 4640 4316 925 5241 9881 Prop. Matri/Patri 82.8% 17.2% 82.4% 17.6% Statistics based on main DID sample in Table 3.1. Matri/Patri: individual of ethnicities from matrilocal/patrilocal tradition. Diff.: difference between Matri and Patri, Raw DD: absolute difference-in-difference (with ***, **, * indicating significance at 1%, 5%, 10% levels). Standard deviations are reported in brackets in columns 1, 2, 4 and 5. Standard errors are reported in brackets in columns 3, 6 and 7.

194 References

Table 3.A.3 – Placebo DID Estimations on Husband’s Split Migration

Husband’s Split Migration (1) (2) (3) (4) (5) (6) (7) Post 0.000 -0.001 0.0179 0.003 0.010 0.024 -0.003 (0.007) (0.008) (0.014) (0.014) (0.017) (0.016) (0.019) Post × Matrilocal -0.012 -0.011 -0.018 -0.012 -0.015 -0.016 -0.010 (0.009) (0.009) (0.011) (0.012) (0.014) (0.014) (0.015) Matrilocal 0.002 -0.006 -0.003 -0.015 -0.005 -0.005 -0.020 (0.011) (0.013) (0.015) (0.015) (0.016) (0.016) (0.017)

Observations 10,465 10,366 10,366 10,366 8,066 8,066 8,066 R-squared 0.001 0.009 0.011 0.012 0.014 0.014 0.016 Clusters 321 320 320 320 317 317 317 Controls Yes Yes Yes Yes Yes Yes Post × Controls Yes Yes Yes Yes Yes Javanese Yes Yes Post × Javanese Yes Yes Close Yes Yes Post × Close Yes Yes Natural Disaster Yes Yes Post × Nat. Disaster Yes Yes Placebo difference-in-difference estimations of husband’s split migration on a sample ofwomen in stable couples in 2000-2007 and surveyed in 2007, pooled with a sample of women in stable couples in 1993-2000 and surveyed in 2000. The outcome is a dummy variable indicating women’s husband’s split migration since the preceding wave of the survey (i.e. since 2000 when surveyed in 2007 and since 1993 when surveyed in 2000). Post is equal to 1 for 2007 (placebo post-reform) and 0 for 2000 (placebo pre-reform). ‘Natural Disaster’ is a dummy indicating individuals living in a village having experienced a natural disaster in the 5 years preceding IFLS 4 (2007). Other variables are described in Table 3.1. Standard errors are reported in brackets and clustered at village of origin level. Significance levels: * p<0.10, ** p<0.05, *** p<0.01.

195 References

Table 3.A.4 – DID Estimations on Husband’s Split Migration (Panel with Fixed Effects)

Husband’s Split Migration (1) (2) (3) (4) (5) (6) (7) Post -0.013 -0.021 -0.287*** -0.286*** -0.280*** -0.281*** -0.271*** (0.009) (0.014) (0.028) (0.028) (0.030) (0.031) (0.032) Post × Matrilocal 0.018* 0.019* 0.028** 0.035** 0.040** 0.041** 0.046*** (0.010) (0.010) (0.013) (0.014) (0.016) (0.016) (0.017) Rel. effect 41.6% 42.5% 63.0% 79.1% 90.7% 92.0% 102.0%

Observations 9,072 9,072 9,072 9,072 7,068 6,710 6,710 R-squared 0.001 0.006 0.011 0.012 0.013 0.016 0.016 Clusters 318 318 318 318 314 313 313 T-Test Equal. (p-val.) 0.099 0.066 0.000 0.000 0.000 0.000 0.000 Household FE Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Post × Controls Yes Yes Yes Yes Yes Post × Javanese Yes Yes Post × Close Yes Yes Natural Disaster Yes Yes Post × Nat. Disaster Yes Yes Difference-in-difference estimations of husband’s split migration on a sample of women in stable couples in 2000-2014 and surveyed in both 2007 and 2014. The outcome is a dummy variable indicating women’s husband’s split migration since the preceding wave of the survey (i.e. since 2007 when surveyed in 2014 and since 2000 when surveyed in 2007). Post is equal to 1 for 2014 (post-reform) and 0 for 2007 (pre-reform). Other variables are described in Table 3.1. ‘Matrilocal’, ‘Muslim’, ‘Javanese’, and ‘Close’ dummy variables are now absorbed in the household FE. The relative effect is calculated in % of mean outcome for patrilocal group in 2007 (pre-reform). We report the p-values of T-Test of ‘Post’ = ‘Post × Matrilocal.’ Standard errors are reported in brackets and clustered at village of origin level. Significance levels: * p<0.10, ** p<0.05, *** p<0.01.

196 References

Table 3.A.5 – Placebo DID Estimations on Husband’s Split Migration (Panel with Fixed Effects)

Husband’s Split Migration (1) (2) (3) (4) (5) (6) (7) Post -0.005 -0.008 0.056 0.051 0.036 0.055 0.036 (0.011) (0.015) (0.037) (0.037) (0.045) (0.039) (0.046) Post × Matrilocal -0.016 -0.016 -0.025 -0.017 -0.016 -0.017 -0.009 (0.012) (0.012) (0.016) (0.017) (0.015) (0.015) (0.016)

Observations 6,536 6,536 6,536 6,536 5,718 5,718 5,718 R-squared 0.006 0.010 0.014 0.014 0.017 0.016 0.017 Clusters 318 318 318 318 316 316 316 Household FE Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Post × Controls Yes Yes Yes Yes Yes Post × Javanese Yes Yes Post × Close Yes Yes Post × Nat. Disaster Yes Yes Placebo difference-in-difference estimations of husband’s split migration on a sample ofwomen in stable couples in 1993-2007 and surveyed in both 2000 and 2007. The outcome is a dummy variable indicating women’s husband’s split migration since the preceding wave of the survey (i.e. since 2000 when surveyed in 2007 and since 1993 when surveyed in 2000). Post is equal to 1 for 2007 (placebo post-reform) and 0 for 2000 (placebo pre-reform). ‘Natural Disaster’ is a dummy indicating individuals living in a village having experienced a natural disaster in the 5 years preceding IFLS 4 (2007). Other variables are described in Table 3.1. ‘Matrilocal’, ‘Muslim’, ‘Javanese’, ‘Close’ and ‘Natural Disaster’ dummy variables are now absorbed in the household FE. Standard errors are reported in brackets and clustered at village of origin level. Significance levels: * p<0.10, ** p<0.05, *** p<0.01.

197 References

198 General Conclusion

This thesis presents three essays that explore how cultural norms and development policies interact in shaping women’s empowerment and well-being, with a speciﬁc focus on developing countries.

To investigate this research question, this thesis offers a wide coverage along several dimensions.

First, this thesis offers a wide geographic and contextual coverage: it covers both Indonesia (a country with wige geographic and cultural diversity) and 18 Sub-Saharan African countries. Second, it explores the effect of several types of traditional norms: inheritence norms (i.e. matrilineality vs. patrilineality) and post-marital residence norms (i.e. matrilocality vs. patrilocality). Third, it emphasizes a wide spectrum of outcomes, from HIV, sexual and contraceptive behaviours to women’s intra-household decision-making, health, fertility, assets, subjective well-being and husband’s migration. Finally, this thesis uses a wide range of estimation techniques to provide empirical evidences that are theoretically grounded: OLS with Fixed Effects, Geographic RDD, IV strategy, Nearest Neighbor

Matching and Difference-in-Differences. In the end, this thesis provides two main conclusions:

(1) Cultural context matters in shaping women’s empowerment and well-being. Building on the evolutionary anthropology and the evolutionary psychology literatures, and using data from

32 Demographic Health Surveys (DHS) covering 18 Sub-Saharan African countries, chapter 1 shows that ethnic group’s ancestral inheritence norms (i.e. matrilineality vs. patrilineality) have long-lasting effects on contemporary women’s sexual and contraceptive behaviours, leading to higher prevalence of HIV among matrilineal women. In the context of Indonesia, another country with wide cultural heterogeneity, chapter 2 exploits data from the Indonesia Family Life Survey and shows that traditional post-marital residence norms (matrilocality vs. patrilocality) are still signiﬁcant predictors of actual household composition in Indonesia and affect several dimensions of women’s well-being and empowerment.

(2) Development policies aimed at improving women’s situation should be more tailored to speciﬁc cultural contexts. chapter 1 uncovers heterogenous sexual and contraceptive behaviours be-

199 General Conclusion tween matrilineal and patrilineal populations, and therefore calls for policies that are more tailored to cultural contexts to induce behavioural changes restraining the spread of HIV. chapter 2 highlights that nation-wide policies in Indonesia aimed at easing women’s access to justice courts had differential effects across traditional modes of post-marital residence norm on women’s empowerment and well-being (health, fertility, assets, women’s and children’s subjective well-being, women’s intra- household decision-making). chapter 3 extends this result and further shows that these Indonesian reforms also affected another key household decision: husband’s migration. All in all, this thesis raises the need to move beyond the “one-size-ﬁts-all” strategy and calls for development policies that are more tailored to cultural contexts.

The originality of this thesis lies in its interdisciplinary approach: it uses a wide range of estimation techniques from the economic discipline to empirically assess hypothesis from evolutionary anthropology and evolutionary psychology literatures. Researchers are increasingly recognizing the key insights that can be brought by other disciplines (e.g. history, anthropology, psychology, etc.) when trying to understand the dynamic processess leading to economic development and to the current state of the world. This thesis contributes to this new stream, which is an exciting and promising future avenue of research.

200 Abstract

This thesis explores the long term effects of ancestral norms on contemporaneous outcomes, in the context of developing countries, with a specific focus on gender related outcomes. The 1st chapter is entitled “Women’s Position in Ancestral Societies and Female HIV: The Long-Term Effect of Matrilineality in Sub-Saharan Africa” and explores the long-term effect of matrilineality on contemporaneous female HIV in Sub-Saharan Africa. Using within Sub-Saharan African countries variation in ethnic groups’ ancestral kinship organizations, I find that females originating from ancestrally matrilineal ethnic groups are today more likely to be infected by HIV. This finding is robust to the inclusion of subnational fixed effects, as well as a large set of cultural, historical, geographical, and environmental factors. I find consistent results using a number of alternative estimation strategies, including a geographic regression discontinuity design at ethnic boundaries and an instrumental variable strategy. Matrilineal females’ riskier sexual and contraceptive behaviours constitute the main explanatory mechanisms. Building an epidemiological model, I simulate how these differences in sexual and contraceptive behaviours translate into different gender-specific HIV rates dynamics. The 2nd chapter is entitled “Traditional Norms, Access to Divorce and Women’s Empowerment: Evidence from Indonesia” and explores how cultural norms and development policies interact in shaping women’s empowerment and well-being in developing countries. My co-authors and I examine this question in the context of legal reforms and their differentiated impact on divorce and empowerment across traditional modes of post-marital cohabitation. We theoretically establish how women originating from matrilocal ethnic groups should respond to the reform compared to those from patrilocal ethnicities. We confirm the model predictions using a panel difference-in-difference approach: the former divorce more and, when in stable marriages, experience a significant improvement in well-being and empowerment. This conclusion calls for better tailored policies that can transcend cultural contexts and overcome the adherence to informal laws. The 3rd paper is entitled “Women’s Em- powerment and Husband’s Migration: Evidence from Indonesia” and examines the link between the distribution of power in marriage and the decision to split-migrate (one spouse migrates alone) in Indonesia. My co-authors and I exploit a national policy experiment that has exogenously increased women’s bargaining power among ethnic groups of matrilocal tradition relative to patrilocal groups. We find that the propensity of matrilocal husbands to split-migrate, relative to patrilocal husbands, increases by 2-3.4 percentage points, i.e. a rise of 41-76%, following the reform. We suggest that empowered women may have ex-ante gained control over outcomes that are costlier to monitor for husbands when they migrate. Hence, empowerment restores some efficiency in migration decisions by reducing the anticipated information asymmetry and the moral hazard associated with migration. Consistently, we show that households with empowered women are more able to cushion shocks due to natural disasters and, among all households experiencing split-migration, matrilocal women are better off than their patrilocal counterparts. We provide a theoretical framework that rationalizes the intra-household mechanisms lying behind these results. Keywords: Gender, cultural persistence, kinship systems, HIV, sexual behaviour, female empowerment, migration JEL classification: D13, D91, I15, J1, K38, N37, O15, R23, Z13

Résumé

Cette thèse explore les effets de long-terme des normes ancestrales sur l’émancipation et le bien-être des femmes dans les pays en développement. Le 1er chapitre est intitulé “Position des Femmes dans les Sociétés Ancestrales et VIH chez les Femmes : l’Effet de Long-Terme de la Matrilinéalité en Afrique Sub-Saharienne”. En utilisant la variation intra-pays dans les types de structures familiales des groupes ethniques, je trouve que les femmes originaires de groupes ethniques ancestralement matrilinéaires sont aujourd’hui plus susceptibles d’être séropositives. Ce résultat est robuste à l’inclusion d’effets fixes infranationaux et d’un large ensemble de contrôles culturels, historiques, géographiques et environnementaux. Je trouve un résultat similaire en utilisant des stratégies d’estimation dif- férentes, telle qu’une approche de discontinuité géographique de la régression, ou une stratégie de variable instrumentale. Le principal mécanisme explicatif est le suivant : les femmes matrilinéaires adoptent des comportements sexuels et de contraception plus à risque. Finalement, je construis un modèle épidémiologique et conduit une simulation permettant de comprendre comment les différences de comportements sexuels et de contraception trouvées entre les femmes matrilinéaires et patrilinéaires se traduisent en différentes dynamiques de transmission du VIH. Le 2e chapitre est intitulé “Normes Traditionnelles, Accès au Divorce et Emancipation des Femmes : Etude de Cas de l’Indonésie” et explore la manière dont les normes traditionnelles interagissent avec les politiques de développement visant à soutenir l’émancipation et le bien-être des femmes dans les pays en développement. Mes co-auteurs et moi-même examinons l’impact différencié de réformes légales sur le divorce et l’émancipation des femmes, en fonction de la norme traditionnelle de cohabitation post-mariage (matrilocalité et patrilocalité). Nous montrons théoriquement comment les femmes originaires de groupes ethniques matrilocaux répondent aux réformes, relativement aux femmes originaires de groupes ethniques patrilocaux. Nous confirmons ensuite empiriquement les prédictions du modèle avec une approche en double différences : nous trouvons que suite aux réformes les femmes matrilocales divorcent plus et, lorsqu’elles restent mariées, bénéficient d’une amélioration de leur bien-être et de leur pouvoir de déci- sion au sein du ménage. Cette conclusion souligne le besoin de penser des politiques qui soient plus adaptées aux différents contextes culturels. Le 3e chapitre s’intitule “Emancipation des Femmes et Migration du Mari : Etude de Cas de l’Indonésie” et examine le lien entre la distribution du pouvoir de décision entre les époux et la décision de migrer seul du mari. Mes co-auteurs et moi-même utilisons une série de réformes qui ont augmenté de façon exogène le pouvoir de décision des femmes matrilocales, relativement aux femmes patrilocales. Nous trouvons que suite à ces réformes la propension à migrer seul des époux matrilocaux a augmenté de 2-3,4 points de pourcentage relativement aux époux patrilocaux, une augmentation de 41-76%. Ce résultat suggère que les femmes émancipées détien- nent ex-ante le pouvoir de décision sur des aspects de la vie de famille qui sont difficiles à surveiller par le mari lorsqu’il migre seul. Ainsi, l’émancipation des femmes restaure de l’efficacité dans la décision de migration du mari, en réduisant l’asymétrie d’information anticipée et l’aléa moral associés à la migration. Nous montrons aussi que les ménages où les femmes sont plus émancipées sont davantage capables d’absorber les chocs liés aux catastrophes naturelles en envoyant un mari en migration. Nous trouvons également que parmi les ménages ayant un mari migrant, les ménages matrilocaux se retrouvent en meilleure posture que les ménages patrilocaux. Pour conclure, nous élaborons un cadre théorique afin de rationaliser les mécanismes intra-ménages se trouvant derrière ces résultats. Mots-Clés: Genre, persistence culturelle, systèmes de parenté, VIH, comportements sexuels, émancipation des femmes, migration Classification JEL: D13, D91, I15, J1, K38, N37, O15, R23, Z13