Social Learning, Influence

Social learning, influence, and ethnomedicine: individual, neighborhood and social network influences on attachment to an ethnomedical cultural model in rural Senegal

Abstract

The preference in many parts of the world for ethnomedical therapy over biomedical

alternatives has long confounded scholars of medicine and public health. In the

anthropological literature cultural and interactional contexts have been identified as

fundamental mechanisms shaping adherence to ethnomedical beliefs and health seeking

behaviors. In this paper, we examine the association between individual, neighborhood,

and social network characteristics and the likelihood of attachment to an ethnomedical

cultural model encompassing beliefs about etiology of disease, appropriate therapeutic

and preventative measures, and more general beliefs about metaphysics and the efficacy

of health systems in a rural population in Eastern Senegal. Using data from a unique

social network survey, and supplemented by extensive qualitative research, we model

attachment to the ethnomedical model at each of these levels as a function of

demographic, economic and ideational characteristics, as well as perceived effectiveness

of both biomedical and ethnomedical therapy. Individuals’ attachment to the

ethnomedical cultural model is found to be strongly associated with characteristics of

their neighborhoods, and network alters. Experiences with ethnomedical care among

neighbors, and both ethnomedical and biomedical care among network alters, are

independently associated with attachment to the ethnomedical model, suggesting an

important mechanism for cultural change. At the same time, we identify an independent

association between network alters’ cultural models and those of respondents, indicative

of a direct cultural learning or influence mechanism, modified by the degree of global

transitivity, or ‘connectedness’ of individuals’ networks. This evidence supports the long

held theoretical position that symbolic systems concerning illness and disease are shared,

reproduced, and changed through mechanisms associated with social interaction. This has

potentially important implications not only for public health programming, but for the

understanding of the reproduction and evolution of cultural systems more generally.

Key words: Social learning, Cultural Models, Social Networks, Ethnomedicine, Population

Health

1. Introduction

In large parts of world, established ethnomedicines compete in the health services

market with those provided under the biomedical model. How do we account for the

ethnomedical beliefs of members of these populations? It could be, as is often assumed, that

individual characteristics are determinative. Individuals with less formal education, or with limited exposure to Western ideas, for example, may be ignorant of the biomedical model.

Religious affiliation and experience with treatment by healers associated with either the biomedical or ethnomedical health system may also be influential. The most persuasive explanation however, a mainstay in medical anthropology since at least Evans-Pritchard’s seminal work with the Azande (Evans-Pritchard, 1937), is that in the cultural and interactional

knowledge context in which individuals find themselves, ethnomedicine should be treated as

rational, at least in the sense that “all beliefs are on a par with one another with respect to the

causes of their credibility” (Barnes & Bloor, 1982, p. 23). In places where biomedical health

systems are relatively recent and less extensive, biomedical models for common illnesses are

often not widely diffused, while ethnomedical models, including culturally specific natural and

supernatural explanations for disease and associated therapeutic regimes, are supported by a

broad and interconnected system of beliefs, norms and behaviors reinforced through social

learning and influence taking place through interaction (Fabrega, 1975; Kleinman, 1978b;

Ngokwey, 1988; Pachter, 1994; Yoder, 1997).

1.1 Cultural models of health, disease and illness

The use of culturally constructed cognitive schemas, or abstract, flexible representational and interpretive frameworks for understanding and acting in complex situations, as a key explanatory framework for health belief and behavior has found wide traction within studies of ethnomedicine (Angel & Thoits, 1987; McKee, 2003; Nichter, 1991; Vecchiato, 1997;

Yoder, 1997). The organizing framework we use to conceptualize and identify cultural elements related to health, the connectionist schema model, is well suited to identifying social learning and influence mechanisms operating through interaction. Schemas define for an individual what exists symbolically in a particular context (for example, agents of disease), and structure perceived possibilities for action (such as therapeutic options). Schemas, in what is known as the

‘connectionist’ framework, are learned as variably weighted aggregate representations of general contexts with associated possibilities and constraints on action drawn from of repeated imprints of different, particular experiences across heterogeneous yet analogous situations (Strauss 1992;

Smith & Queller 2004; Strauss & Quinn 1998).

Cultural models, as developed in cognitive anthropology, are hierarchically nested sets of such schemas that reciprocally shape one another (D’Andrade, 1992). Cultural models in this

tradition share much with Kleinman’s explanatory model (EM) framework (Kleinman, 1978a).

The term ‘cultural model’ is often used in applied research to reference aggregates of belief and behavior using that framework. Though other differences exist, cultural models in the connectionist schema framework are more general, and go further than the EM model in explicitly theorizing mechanisms by which schemas are associated and meaning is constructed at

the individual level through interaction.

Higher level schemas in the framework applied here include those concerning ontology

and metaphysics, associated ultimate causes of illness, and the relative efficacy of competing

systems to address it. Mid-level schemas concern the proximal, contextually specific causes of

particular illnesses, while lower level schemas concern therapeutic or preventative options.

Hierarchical association between these levels occurs where more abstract schemas call upon and

constrain the range of lower level schemas available. Lower level schemas and attendant action

associated with them in turn shape the context for reproduction and evolution of higher level

schemas (Strauss 1992; D’Andrade 1995; D’Andrade 1992). Perceptions of the efficacy of

treatment received under one model or the other, for example, may inform beliefs about the

proximal cause of illness (either to support or undermine them). These in turn may reinforce or

weaken attachment to higher level schemas which support them.

1.2 Social learning and influence

Social interaction is a key element of the repeated imprinting of particular experience

fundamental to the development and evolution of schemas under the connectionist model.

Individual experience (such as with medical care) plays a fundamental role, but is limited in its

frequency and variability. By far the most frequent stimuli we are exposed to come through

interaction with others, either interpersonal or mediated. Cultural models (including attendant

institutions and norms) can be conceptualized as arising from, and evolving, as a function of the heterogeneous distributions of stimuli individual members of a population are exposed to through interaction, and the situated actions they take (c.f. Blau, 1994; Mead, 1967; Tarde,

1895). Central tendencies of these distributions and the resources (or capital, material or human)

they reference signal shared experience, or intersectionality. Variability in these distributions as experienced and acted on by individuals creates variation in subjective perception and contingent

action, which, over time, allows cultural models to evolve.

In health and health behavior, as well as demography, insights concerning social learning

and evaluation mechanisms are beginning to expand our understanding of a wide variety of

critical issues (Behrman, Kohler, & Watkins, 2002; Berkman & Kawachi, 2014; Sandberg, 2006;

Smith & Christakis, 2008; Valente, 2010). Exploiting knowledge of social learning and behavior change processes, particularly through social networks, is currently thought to hold great potential for the development of more efficient public health interventions as well (Latkin &

Knowlton, 2015; Valente, 2012; Valente, Palinkas, Czaja, Chu, & Brown, 2015).

No research to date, however, has attempted to directly model the influence of social learning on belief systems and cultural models related to health and illness, the goal of this investigation. In this paper, using unique social network data from a small population in

Senegal, we test hypotheses concerning the influence of social learning through interaction on attachment to an ethnomedical cultural model through the structural characteristics, health care experiences, and cultural models of health and illness held by respondents’ social network members and neighborhood co-residents. Our aim here is to identify interactional mechanisms through which ethnomedical belief systems are supported and potentially change, but from a more general perspective, of course, how social learning contributes to our understanding of the

5 evolution of culture through interaction in this special case may yield insights into broader processes in other cultural domains.

2. Methods

2.1. Setting

The data used for the present analysis come from the Niakhar Social Networks and

Health Project (NSNHP). The NSNHP is a large scale social network project with multiple qualitative, survey, and methodological components conducted in collaboration with the Niakhar

Demographic and Health Surveillance System (NDHSS). 1 The NDHSS is a prospective longitudinal surveillance system maintained by the Institut de Recherche pour le Développement

(IRD) in the Fatick region of Senegal. The NDHSS study zone comprises 30 villages with a total population in 2014 of 43,664 whose members have been under surveillance for over 50 years.

Residence is organized in kin-based residential household groups, or compounds, known as concessions composed of one or more households, or hearths. Economic production and consumption is organized at the household level. The economy is largely agro-pastoral, with staple crops millet grown for consumption and peanuts grown for the cash market. Households also keep small livestock and raise cattle.

The population is 96.7% ethnically Sereer, and the dominant western religion is Islam, followed by Christianity. A significant syncretism exists, however with an indigenous monotheism practiced to a greater or lesser extent by most of the population. This religion centers around the intercessory power of a complex system of ancestral entities, known as

1 For a complete description of project components, visit the project website at www.nsnhp.org.

Pangools who act as a conduit for the power of the remote, omnipotent sky-god, Roog Sene

(Dupire, 1994; Gravrand, 1990).

2.1.1 Ethnomedical causes of illness

Qualitative work conducted as part of the NSNHP, in the form of 98 in-depth disease narratives from a random sample of the population, suggests that schemas concerning causation of illness among the population may be conventionally categorized, for the most part, as either naturalistic, having a natural, mechanistic explanation, or personalistic, caused by somebody, usually a supernatural being or human with supernatural powers (Snow 1974).

As with other populations in West Africa (Green, 1999; Nichter, 2008), in addition to naturalistic causes associated with the biomedical model, the Sereer have a variety of culturally- specific naturalistic theories of disease causation. These include diseases related to infectious agents not recognized in the biomedical model, diseases associated with heredity, with food handling and preparation, with agricultural labor and a variety of environmental exposures.

A large number of illnesses, however, both physical and psychological are believed to be caused by supernatural non-human entities, including animal spirits, Pangools (as punishment for insufficient supplication) and genies (djinn). Alongside the non-human supernatural agents recognized by the Sereer as capable of causing illness, there are several human (or recently human) ‘agents of evil’ or ‘persons of the night’ (Sr. o wiin yeng) who are also capable of inflicting illness, and often, death. The most common supernaturally caused illness, however, is a mystical ill-wind (Sr. nguegne), a personalistic environmental contagion causing febrile illness, exposure to which is facilitated by violation of specific behavioral norms (c.f. Green, 1999).

2.1.2 Diagnosis and therapy

Appropriate diagnosis and therapy for an illness depends on its suspected cause.

Members of the population distinguish diseases with ethnomedical personalistic and naturalistic causes that can only be treated by a traditional healer, those of biomedical origin that can be

treated only by Western medical means, and a small subset of ethnomedical naturalistic diseases

that can be treated with either traditional or biomedical therapy. In the first category are illnesses

caused by various ‘agents of evil’, those caused by neglect of Pangools, or violation of animal

spirits or totems. The second category includes the majority of infectious diseases not otherwise ethnomedically diagnosed, chronic diseases such as hypertension and diabetes, and injuries. The third category is relatively small, to our knowledge, with variation from individual to individual.

As discussed above, the hierarchical association of schemas at different levels is not deterministic, but heterogeneous across individuals. Exposed to a heterogeneity of stimuli through interaction, individuals may retain elements of both biomedical and ethnomedical

models simultaneously, in opposition, or in synthesis. Attachment to lower level schemas

associated with one model or another under particular circumstances depends on the cognitive availability of various options constrained by higher level schemas and perceived efficacy of therapy (both potential and experienced). For these reasons, diagnosis and attendant therapy of illness often shifts between ethnomedical and biomedical models serially. At times, biomedical and ethnomedical therapies are pursued simultaneously.

Most often in the case of less severe illness, biomedical (over-the-counter or preserved prescription medicine) or ethnomedical (ethnobotanical or supplication of Pangools) auto- medication is attempted. Depending on the perceived success of these efforts, or the severity of the illness, other forms of auto-medication may be attempted, or biomedical or ethnomedical

8 clinical help may be sought. Direct experience with the efficacy of these treatments plays a key role in supporting belief in the efficacy of the system implicated. So do the experiences of family, neighbors, and members of social networks, either directly through the health seeking decision making process, or through their input in shaping individuals cultural models.

2.2 Survey data

The NDHSS collects data, longitudinally, on a variety of demographic, social and health phenomena. In linking respondents and members of their social networks to this vast data, the

NSNHP has been able to collect information on larger and more comprehensive social networks

(across more types of ties, with unlimited, ‘free-choice’ citation of network alters) than has previously been possible. The NDHSS data also allow for the identification of kin and members of a variety of residential and geographic groups, from the household through the neighborhood and village.

The first panel of the main NSNHP survey, collected in 2014 and analyzed here, solicited information on the presence, characteristics and strength of network ties in 15 distinct types of interaction across four theoretically key domains (affective, exchange, temporal co-presence and role relational) of association (Sandberg, Rytina, Delaunay, & Marra, 2012). The main survey has two components. The first is a complete census of individuals aged 16 and above from one village, Yandé. The second is a random sample of individuals in the same age group from the rest of the population of the NDHSS zone.

In addition to information on respondents’ social networks, the survey also contains an extensive respondent questionnaire covering a number of substantive topics. The largest module, based on our prior qualitative research, concerns belief, ideation, and behavioral reports related

to high, mid- and lower-level schemas associated with etiology, diagnosis, and treatment of

disease as well as perceived quality of both biomedical and traditional clinical health care.

In the Yandé data, used in the present analysis, the response rate was 95.4%. In total,

1310 individuals were interviewed in 203 concessions across 10 neighborhoods. Nine respondents’ interviews were lost or invalidated due to errors in the CAPI administration. A further 14 respondents were dropped for reporting 5 or fewer network alters in total, deemed implausible and indicative on non-cooperation. Two respondents were missing information on the wealth of their residential compound and one on the health ideation of their network members. Removing these, we have a final analytic sample size of 1284 for the present analysis.

Respondents in Yandé named on average 40 network members (or ‘alters’). Of these, 24 alters were uniquely identified (removing multiplexity, or nomination in multiple name generators).

2.3 Measures and model modeling strategy

In the following analysis, we use a nested modelling strategy to estimate the association of attachment to an ethnomedical cultural model and characteristics of individuals, their neighborhoods, and their social networks. All models presented here are estimated in Stata 14 via logistic regression with a random effect at the household group (concession as defined above) level. This latter element is included to estimate within-concession variation in attachment to the ethnomedical cultural model, which, due to the presence of some relatively small concessions, were inestimable here as fixed effects.

2.3.1 Dependent Variable

A latent class analysis of sixteen items concerning health ideation and behavior from the main survey module of the respondent questionnaire was used to classify respondents according

to their relative attachment to biomedical and ethnomedical cultural models. Higher level

schemas concerning religious belief and cosmology are represented in these models by items

related to ultimate causes of illness, efficacy of treatment and the power of supernatural

intercession in human affairs. Mid-level schemas concerning proximal causes of disease are

represented by items concerning the hypothetical attribution of cause to either ethnomedical or

biomedical domains in specific scenarios. Lower level schemas concerning prevention and

therapy are represented by items concerning appropriate health behaviors in specific scenarios.

These are described in depth in the appendix provided as online supplemental material. Table A2

presents exact question wordings and response categories, recoding and distribution for the

present analysis, as well as descriptions of the rationale for the inclusion for each.

The six questions associated with higher level schemas concerned respondents’

evaluation of the most important cause of illness among the Sereer, belief in the efficacy of

ethnomedical and biomedical healers, and belief in traditional versus Western religious leaders.

Response choices for these variables were forced preference rank order. For the first two

questions, the first and second choices were used to categorize responses as belief in solely

ethnomedical causation/efficacy, ethnomedical causation/efficacy given priority, biomedical

causation/ efficacy given priority, and solely biomedical causation/efficacy. The third question

was categorized along an analogous spectrum from traditional to Western belief. The next three

questions were simple single response format. The first concerned belief in the efficacy of

supplication to Pangools, the second whether supernatural illnesses could be treated by a medical doctor. The final indicator is behavioral, an interviewer observation as to whether the respondent was wearing an amulet of mystical protection at the time of the interview.

For mid-level schemas, we employ measures of four illness scenarios represented by

symptoms and context prevalent in popular discourse as revealed from our qualitative work.

Respondents were asked to assess the likely cause of each. Response choices, again in forced-

preference rank order format, included both biomedical and ethnomedical options, with the first two choices used to categorize responses as belief in solely ethnomedical causation, ethnomedical causation given priority, biomedical causation given priority, and solely biomedical causation.

Lower level schemas are represented by two questions concerning treatment of common symptoms associated with both ethnomedical and biomedical causation, one question concerning an ideal-typical attack from a supernatural agent, one concerning the most appropriate venue for childbirth (a period of increased risk of supernatural attack), and two questions concerning

preventative measures to be taken to avoid illness. As with most of the other indicators discussed here, these questions were also forced preference rank order. They were recoded along the solely ethnomedical to solely biomedical continuum described above.

Latent class estimation resulted in a preferred a three-class solution (appendix table A1).

We have labeled these classes as ‘ethnomedical’ (30% of the sample), ‘biomedical’ (32%), and

‘liminal’, or holding an intermediate position between the ethnomedical and biomedical

classifications (38%). The estimated likelihood of class membership for each response category of the recoded indicators for the preferred model is presented graphically in the left panel of appendix table A2.

On the higher level, those classified as ethnomedical hold the most traditional religious

and spiritual beliefs and have the greatest belief in the efficacy of traditional and religious

healers. Respondents classified as biomedical are the most adherent to Western monotheism, the

most likely to believe in the efficacy of medical doctors and least likely to have been wearing

mystical protection during the interview, though the liminal class is almost identical in these

respects.

The ethnomedical class is the one most likely to cite only, or to prioritize ethnomedical

causes of disease, prevention and treatment, and the least likely to cite biomedical options. The

biomedical class is more likely to cite only or prioritize biomedical options for these indicators

of mid-and lower-level schemas. The liminal class, as suggested by the name, falls between these

two. They are less likely to give either purely ethnomedical or biomedical responses and more

likely to consider causes from the alternative model (biomedical if their first choice was

ethnomedical, ethnomedical if their first choice was biomedical) as secondary than the other

classes. They are also more likely to consider alternative treatments than either of the other two

classes, in some cases prioritizing ethnomedical, others, biomedical options. In terms of

prevention, however, they are nearly identical to the biomedical class. For the purposes of this

analysis, we dichotomize this classification, indicating ethnomedical classification, the reference

being biomedical or liminal classification for use as the dependent variable. This was done to

model clearly the contrast between those holding the most purely ethnomedical schemas, and the

reference classes, which share relatively higher levels of biomedically oriented schemas.

2.3.2 Individual level independent variables:

To estimate unbiased social learning or influence effects, it is necessary to control for

individual characteristics that may also influence the outcome, and which may at the same time

be associated with information available from network alters (Palloni, 2001). In large part these

biases stem from homophily – the strong structural and psychological tendency toward

association among those who are more alike than not. When these are uncontrolled, endogeneity

results from the association between respondents’, neighbors’ and network alters’ characteristics

(Shalizi & Thomas, 2011). Who individuals interact with and learn from is constrained by these

structural elements and shaped by associational choice within them. The structural aspects,

including demographic and spatial constraints on who individuals have the possibility of

interacting with have been labeled as producing ‘baseline’ homophily; associational choice

within these constraints ‘inbreeding’ homophily (McPherson, Smith-Lovin, & Cook, 2001).

Social networks have been shown to be strongly homophilous with regard to, age,

religion, education, occupation, and gender, either due to baseline homophily, inbreeding

homophily or some combination of the two (McPherson et al., 2001). If factors influencing the

structure of social association through homophily or otherwise are uncontrolled in a model of

social learning or influence, we risk over-estimating these effects. Simultaneous identification of

learning (or contagion) and homophily effects is notoriously problematic. Solutions that have

been offered include conditioning and propensity score matching on observable individual and

network characteristics as well as fixed and random intercept models (Aral, Muchnik, &

Sundararajan, 2009; Goldsmith-Pinkham & Imbens, 2013; Lin, 2010; Shalizi & Thomas, 2011).

In our baseline multivariate model, we take the first approach, conditioning on individual

characteristics likely to both have direct effects on the likelihood of attachment to an

ethnomedical model and to structure social association through homophily. These include the

sex, age, educational attainment, and religious affiliation of the respondent, and levels of

agricultural investment and material wealth of their households. To the extent that these may

cause endogeneity in the identified network effect, or are correlated with other factors that may

do so, that potential bias is controlled, but potential bias due to latent homophily from other,

unobserved sources remains (Shalizi & Thomas, 2011).

Women in this population are generally responsible for organizing the health treatment of

members of their households, leading to wider experience with the biomedical system (Franckel

& Lalou, 2009). We expect them to be less attached to the ethnomedical model than men. In our

qualitative work, older people were generally more attached to traditional belief systems,

including ethnomedical models. Age is specified in the all models as a continuous variable.

Education is an important stratifying dimension in the population under study; less than half of

the population has any formal education. We expect that greater exposure to formal education

will be associated with a lower likelihood of ethnomedical classification. We code individual

level educational attainment with four categories, at most primary education, secondary

education and higher than secondary education; no formal education is the reference category.

Religious identification is obviously central to higher order schemas. Those identifying with the

traditional Sereer religion should be more likely to be classified as having an ethnomedical

orientation. Aside from that, we have no firm expectation about the association between religious

affiliation and ethnomedical belief classification, but it is possible that Christians will be less

likely to hold ethnomedical schemas due to exposure to a private Catholic hospital in the

surveillance zone, Muslims more likely to due to the prominent practice of Islamic spiritual

medicine (Syed, 2003). Religion is specified as a categorical variable, with indicators for

Catholic/Christian and traditional religious belief. Islam is the reference category. Members of

families whose economic activities are mostly agricultural were seen in our qualitative work to

hold more traditional values concerning health and illness. We expect agricultural wealth to be

positively associated with ethnomedical classification. Material wealth in the form of improved

housing and household amenities, in contrast, may proxy a tighter integration into the cash

economy, and exposure or aspiration to Western ideals of material well-being. We expect

material wealth to be negatively associated with ethnomedical classification. An exploratory

factor analysis was performed using twelve variables measured at the household level from a

census of household wealth conducted simultaneously with the first panel of the NSNHP. Two

negatively correlated factors with eigenvalues greater than one were extracted, corresponding to

agricultural investment and material wealth (analysis not shown, available on request). Factor

scores were standardized relative to all households in Yandé, with a mean of zero and standard

deviation of 1.

To model effects associated with individual level feedback from lower level schemas,

potentially critical in the evolution of mid-level schemas regarding causation of disease, we include four measures of respondents’ experiences with both the biomedical and ethnomedical

health systems. Respondents were asked whether the last time they or a family member had

consulted with either a medical doctor or a traditional healer they were healed, and to assess the

quality of care at health posts and from traditional healers, respectively. These questions were recoded to indicate positive responses, other responses as the reference category. We expect positive evaluations of ethnomedical (biomedical) treatment and care quality will be positively

(negatively) associated with ethnomedical classification. There is a likelihood, however, that

responses to concerning care quality will be to some degree endogenous, in that attachment to a

particular cultural model likely shapes perceptions of the quality of care received. Though when

specified with other individual level controls this bias may be reduced, inferences about their

associations with ethnomedical classification should be made cautiously.

2.3.3 Neighborhood level independent variables

There are ten neighborhoods in Yandé, with a median 106 respondents. On the

neighborhood level, we measure many of the same characteristics as at the individual level,

aggregated as the arithmetic mean across all members of the population above the age of 16

years old. For educational attainment, we include two continuous variables for the proportion of

residents in each neighborhood with at most primary education and, because higher education is

relatively rare, the proportion with secondary and higher education. Agricultural investment and

material wealth are aggregated to and standardized at the neighborhood level. Neighborhood

religious context is measured as the proportion of residents who are Catholic or another Christian denomination. Experience with clinical biomedical and ethnomedical treatment is measured as the aggregate proportion of neighborhood residents who indicated the last time they or a family member had consulted with either a practitioner they were healed. Measures of general health

care quality were omitted at the neighborhood level because they were highly correlated with the

measures of healing efficacy, raising issues of collinearity. For each of these measures, we

expect the direction of the association with ethnomedical classification to be the same as for their

individual level analogues. The mechanisms by which these associations are expected to be

produced, however, are indirect, operating through learning and influence from the aggregate

ideational and experiential context in which respondents live.

2.3.4 Social network level independent variables

Network measures were aggregated across all unique alters cited above the age of 16

cited in the name generators into a ‘synthetic’ network, in which duplicates were removed, an

approach that has been seen as acceptable in such situations (Knoke and Burt 1983). All models

with network specifications include a control for the number of alters cited, as the absolute size

of personal network may proxy a number of unobserved mechanisms related to diffusion and

social learning (Valente, 2010).

Network educational attainment is operationalized with two continuous measures, the

proportion of alters with at most primary education and the proportion with secondary education

or higher. Average network agricultural investment and material wealth scores are aggregated

and standardized at the individual level. The proportions of network alters who were Christian,

and who responded they were healed by a traditional healer or medical doctor are all continuous

measures. The expected direction of the association of each of these variables with ethnomedical

classification is identical to that for their individual and neighborhood level analogues. When

specified jointly with the neighborhood level variables, coefficients associated with these may,

with caution, be interpreted as operating through this specific set of interpersonal channels net of

broader contextual learning effects.

The final network variable we include here is simply the average probability, derived

from the latent class analysis, that respondents’ network alters will be classified as adhering to

the ethnomedical model. Labeled here ‘classifcation endogeneity’, when specified with controls

for sources of endogeneity arising from homophily at the individual and neighborhood levels,

estimates associated with this variable can be interpreted as the (relatively) unbiased direct

association between alters’ cultural models and respondents’ own.

2.4 Results

Descriptive statistics for all variables used in the following analysis are presented in table 1.

[Table 1 about here]

Table 2 presents the results of the logistic regression models, with coefficients transformed to

marginal/discrete changes in probability of ethnomedical classification. It should be noted that

since the data we are using come from a census of Yandé, the marginal and discrete differences

in probability presented here are population parameters, not estimates.

[Table 2 about here]

Model zero presents estimates associated with separate, zero-order bivariate regressions

of ethnomedical classification on each of the independent variables. Among the individual level

covariates, we see strong associations in the predicted directions for age, educational attainment,

agricultural investment and material wealth. Christians are estimated to be less, those citing

traditional religious affiliation more likely, to be classified as ethnomedical than are Muslims.

The coefficients for individual experience with ethnomedical and biomedical health systems are

also in the predicted directions.

The marginal/discrete change in probability associated with each of the variables at the

neighborhood and network levels are also in the predicted directions. Higher neighborhood and

network educational attainment, material wealth, proportion of residents/alters who were

Christian and proportion of residents/alters saying they were healed by a doctor are negatively

associated with respondents’ ethnomedical classification. Higher average neighborhood and

network agricultural investment and proportions of residents/alters who say they were healed by

a traditional healer are positively associated with ethnomedical classification. Finally, the

average probability of being classified as supportive of the ethnomedical model among network

alters (the classification endogeneity measure) is strongly associated with respondents’ own

classification as such, the marginal change in probability approaching unity.

Model one presents the first multivariate specification, with just the individual level

covariates. Here we see, as would be expected, a reduction in the magnitude of most estimates

relative to the zero-order models. The associations between the categories of educational

attainment and ethnomedical classification, though still important, are reduced by half or more

relative to their zero-order analogues. The positive association of agricultural investment is

dramatically reduced (as are the associations with the religious indicators), but that of material

wealth remains. The health care experience variables are all slightly reduced in magnitude, with

the exception of health post quality.

Model two introduces the neighborhood level variables to the specification from model

one. In the presence of the individual level covariates, only the proportions of neighborhood

residents with primary education, who identify as Christian, and who state that they’ve been

healed by a traditional healer remain strongly associated with respondents’ ethnomedical

classification. None of the individual level marginal effects change in substantively important

ways. Taken together, these results suggest that while baseline homophily associated with

geographic proximity explains a substantial amount of the zero-order neighborhood associations,

independent, if muted, associations between ethnomedical belief and the educational and

religious composition of neighborhoods, as well as perceived efficacy of traditional healthcare

among respondents’ co-residents, remain. Also of note is a reduction in household group level

variance captured by the random effect at the household group level, suggesting that a substantial

proportion of this variance is captured by structural characteristics of neighborhoods.

Model three releases the constraints on the social network variables from the

specification of model 1, with the exception of classification endogeneity measure. As in model

two, the associations between the network covariates and ethnomedical classification are muted

relative to the zero-order results, with the exception of proportion of alters with primary

education, the proportion of alters who are Christian, and the two health care experience

variables. Again, none of the individual level marginal effects are substantively changed. This

suggests that homophily explains the majority of social network level associations, though some

independent effects, particularly with regard to education and health care experience among

network alters remain. Relative to model one, the network covariates also explain a substantial

proportion of the variance of the random effect, capturing effects associated with within household-group interaction.

Model four takes the specification from model three, adding only the classification endogeneity measure. This association, as described above, was by far the strongest in the zero- order models. This association is dramatically reduced in model four but is still substantively important. Comparing model four and model three, a substantial proportion of the other network characteristics’ associations with ethnomedical classification are explained. This includes the proportion of alters with primary education, the proportion of alters who are Christian, and approximately ¼ of the association between the network health care experience variables and ethnomedical classification. Perhaps as importantly, in this model the variance of the household- group random effect is further reduced relative to model three, to the point where including it in the specification would no longer be a significant improvement over a simple logit model at the .01 alpha level.

Of course, as described above, neighborhood associations with ethnomedical classification presented thus far could be due to interaction with network alters from within the neighborhood, and network level associations could be due to baseline homophily, inbreeding homophily, or both. Model five presents the first combined specification of covariates at all three levels, excluding the classification endogeneity measure. While the network level effects remain relatively stable, if slightly diminished, in the presence of neighborhood characteristics, the

marginal effects associated with important neighborhood level associations increase in

magnitude relative to model two. This suggests that a small part of the network level effects are

driven by baseline homophily of association at the neighborhood level.

Model six adds the network classification endogeneity measure to the specification from model five. The marginal change in probability of ethnomedical classification associated with average probability of network members being themselves classified as such remains strong.

Again, as in model four, inclusion of this measure explains a substantial degree of the other network level associations, including education, religious affiliation and those of the health care experience measures. It also explains a substantial proportion of the neighborhood level associations seen in model five, including the proportion of neighborhood residents saying they were healed by a traditional healer. This suggests that some of these neighborhood contextual effects are in fact due to direct network interaction between respondents and residents of their neighborhoods. As importantly, in this model, the magnitude of the marginal probabilities for the proportions of network alters having been healed by a traditional healer and a medical doctor are slightly reduced, but not eliminated, suggesting an independent effect net of alters’ own cultural models.

Having established evidence for social learning effects associated with this cultural model among network members, a final analytic possibility presents itself that holds the potential to further illuminate the mechanisms by which it may operate. A fundamental hypothesis of social network models of learning and influence is that such effects may be dependent on the structure of network ties in interaction with the content of information exchanged within them

(Merton & Kitt, 1950). Network theorists, for example, suggest that networks that are more cohesive, either through density (the ratio of ties between two dyads to the total number of dyads

possible) or transitivity (the degree to which relationships between one individual and two others

imply a tie between the latter two, or how ‘cliquish’ the network is) tend to be more

homogeneous in terms of their beliefs and behaviors and engender a strong normative context.

To test whether such interactions exist between network structure and the classification

endogeneity measure, we estimated two final models, one interacting this measure with ego

network density, the other with global ego network transitivity (not shown). Though there was no

substantively important interaction of classification endogeneity with network density, that with

transitivity significantly improved model fit over model six. Figure 1. Presents the predicted probability of respondent ethnomedical classification as a function of the average network probability of classification at the 5th (.24), 50th (.45) and 95th (.68) percentiles of the distribution

of global ego network transitivity in our analytic sample. As can be seen there, more highly

transitive, or ‘cliquish’ networks are more responsive to the signal concerning the ethnomedical

cultural model embedded in them than are less transitive networks.

2.5 Discussion and Conclusions

In this paper we have presented evidence to support the long held theoretical position that

symbolic systems, or cultural models, concerning illness and disease instantiated through

ideation and associated behavior are shared and reproduced differentially between individuals

through social interaction. We are not unaware that attempting to statistically identify casual

factors associated with attachment to an ethnomedical cultural system may be perceived as

epistemologically anathema to some in the field of medical anthropology. While acknowledging

that such an analysis makes at times gross, simplifying assumptions about the complex symbolic

relationships entailed, we believe it adds to the existing body of ethnographic work by

approaching the topic from a fundamentally different, but not completely discordant perspective.

Such a perspective can, it is hoped, contribute to understanding of the social mechanisms

involved in their reproduction and evolution.

If we accept the legitimacy of this approach, there remain obvious limitations to the

analysis presented here. The measures used to derive the dependent variable are neither

theoretically exhaustive, representative of other populations (even within Senegal), nor necessarily comprehensive. They are however, fine-grained indicators based on in-depth

qualitative research of popular beliefs and experiences among members of this population, a

necessity for representing cultural models. In this analysis, we have dichotomized the

classification results to isolate those with more purely ethnomedical cultural models in order to

draw a contrast with those classified as adhering to biomedical, or partially biomedical (liminal)

models. While we believe for the purposes of this analysis this is justified, future analyses could

profitably explore the associations between network processes and these two other classifications

independently. A closer understanding of processes associated with liminal classification,

potentially (though not certainly) transitional between ethnomedical and biomedical models

could be of particular interest in understanding ideational and behavioral change in such a

context. Perhaps the most obvious critique is simply that there remains uncontrolled endogeneity

through baseline or in-breeding homophily, where individuals are constrained in their interaction

with, or choose neighborhoods or network members based on their shared experience or beliefs.

We have explicitly designed the conditioning strategy used to eliminate, or at least minimize, this

source of bias. For estimates of neighborhood or network level characteristics to be biased in this

way, one would have to believe that individuals select their neighborhoods or network alters

based on their ideational characteristics independently of the association of this ideation and all

the individual level covariates (and covariates at the complementary level) in the models. We

believe, given the traditional social organization of the present context, this is highly unlikely.

Though much of the association between attachment to the ethnomedical cultural model and neighborhood and social network characteristics seen in the simple bivariate associations is explained by individual level characteristics, likely as the result of baseline or in-breeding homophily, and in the case of neighborhoods, direct interaction with network alters within them, some important associations remain. Net of their individual characteristics, respondents living in neighborhoods or having social networks with higher percentages of Christians and those with at

least some formal education are less likely to be classified as holding the ethnomedical cultural

model.

More importantly, however, are experiences with clinical care. These may have a direct

influence on clinical decisions, as well as an indirect one, through their influence on individuals’

identification with and application of particular cultural models. As discussed above, such effects

may be critical in the evolution and relative distribution of schemas. That neighbors’ experiences

with traditional healers, and the experiences of social network alters with both traditional and

biomedical practitioners maintain an association with respondents’ cultural models when controlling for network members’ own classification provides support for the hypothesized mechanism linking cultural models to observed or learned experience net of ideational context, potentially an important driver of cultural evolution.

We have also provided evidence in support of a substantial independent influence of cultural models held by network alters. To the extent that we have controlled for factors associated with alters’ own ideation – their educational attainment, their religion and experience with clinical care – and respondent characteristics, these effects may be, with caution, interpreted

as evidence of direct learning and influence of cultural models through interaction. That the

cultural models held by network alters further explain substantial proportions of other identified

network associations (including those related to alters’ clinical experiences) instead of being explained by them, suggests they exert an independent effect, operating through social interaction, a suggestion supported by the finding that this association works strongly in

interaction with the connectedness (as measured through transitivity) of individuals’ social

networks.

Table 1. Means, proportion, and standard deviations of variables in multivariate analysis of ethnomedical classification in Yandé City: (n=1,284)

Mean SD Individual Characteristics Sex (reference: male) Female 0.533 0.499 Age 35.901 16.294 Education (reference: no education) Primary education 0.184 0.387 Secondary and higher 0.164 0.371 Higher than secondary education 0.111 0.314 Agricultural Investment 0 1 Material Wealth 0 1 Religion (reference: Muslim) Catholic/other Christian 0.107 0.309 Traditional 0.014 0.118 Missing/other 0.014 0.118 Experience Healed by traditional healer (reference: all others) Yes 0.503 0.500 Healed by doctor (reference: all others*) Yes 0.921 0.269 Health post care quality (reference: all others**) Good/very good 0.952 0.214 Traditional health care quality (reference: all others**) Good/very good 0.514 0.500 Neighborhood Characteristics Proportion of residents with primary education 0.223 0.053 Proportion of residents with secondary and higher education 0.263 0.078 Residents' agricultural investment (factor analysis) 0 1 Residents' material wealth (factor analysis) 0 1 Proportion of residents are Christian 0.104 0.140 Experience Healed by traditional healer (reference: all others*) Yes 0.402 0.070 Healed by doctor (reference: all others*) Yes 0.734 0.056 Social Network Characteristics Number of Network Alters 23.050 7.885 Proportion of alters with primary education 0.248 0.134 Proportion of alters with secondary and higher education 0.228 0.170 27

Mean SD Alters' agricultural investment (factor analysis) 0 1 Alters' material wealth (factor analysis) 0 1 Proportion of alters are Christian 0.119 0.167 Experience Healed by traditional healer (reference: all others*) Yes 0.364 0.153 Healed by doctor (reference: all others*) Yes 0.652 0.152 Average probability of alters supporting ethnomedicine 0.297 0.168 Source: compiled by the author. Notes: final analytic size: 1,284, restricted the prediction to the estimation subsample * Reference: all others (no, not completely, don’t know, nobody in the household has ever consulted a traditional healer, no response, missing.) ** Reference: all others (very bad, bad, acceptable/mediocre, don’t know, no response, missing.)

Table 2. Likelihood of ethnomedical classification on individual, neighborhood and social network characteristics: marginal and discrete change in probability, Yandé City: (n=1,284) Model 0 Model 1 Model 2 Model 3 Model 4 Model 5 Model 6

Individual Characteristics Woman (reference = male) -0.026 -0.055* -0.056* -0.065** -0.057* -0.067** -0.062* Age 0.005*** 0.003*** 0.003*** 0.003*** 0.003*** 0.003*** 0.003*** Education (reference = no education) Primary education -0.100** -0.044† -0.043 -0.038 -0.039 -0.041 -0.041 Secondary education -0.195*** -0.083* -0.086* -0.087* -0.095* -0.093* -0.098* Higher than secondary education -0.295*** -0.170*** -0.171*** -0.166*** -0.173*** -0.171*** -0.173*** Agricultural investment 0.048*** 0.000 -0.018 -0.025† -0.024† -0.025† -0.024† Material wealth -0.105*** -0.064*** -0.057*** -0.058** -0.058** -0.057** -0.057** Religion (reference = Muslim) Catholic/other Christian -0.134*** -0.053 -0.023 -0.004 -0.003 0.01 0.008 Traditional 0.246* 0.066 0.045 0.066 0.07 0.048 0.054 Missing/other -0.143 -0.072 -0.069 -0.054 -0.044 -0.052 -0.044 Experience Healed by traditional healer (yes)* 0.212*** 0.123*** 0.122*** 0.119*** 0.114*** 0.116*** 0.114*** Healed by doctor (yes)* -0.164*** -0.094* -0.091* -0.101* -0.100* -0.096* -0.096* Health post care quality (good/ very good)** -0.118* -0.144* -0.140* -0.130* -0.131* -0.127* -0.129* Traditional health care quality (good/ very good)** 0.215*** 0.103*** 0.092*** 0.090*** 0.088*** 0.084** 0.084** Neighborhood Characteristics Proportion of residents with primary education -1.150*** -1.299* -1.366* -1.180* Proportion of residents with secondary and higher education -0.755*** -0.525 -1.065† -0.930† Residents' agricultural investment (factor analysis) 0.099*** -0.148 -0.254* -0.218 Residents' material wealth (factor analysis) -0.074*** -0.022 -0.038 -0.033 Proportion of residents who are Christian -0.368*** -0.582† -0.739* -0.643* Experience Healed by traditional healer (yes)* 1.385*** 1.523* 1.675** 1.356* Healed by doctor (yes)* -0.039 -0.475 -0.35 -0.153 Social Network Characteristics 29

Model 0 Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 Total number of network alters 0.002 0.001 0.001 0.001 0.001 Proportion of alters with primary education -0.486*** -0.274** -0.199* -0.267* -0.225* Proportion of alters with secondary and higher education -0.461*** 0.055 0.127 0.062 0.107 Alters' agricultural investment (factor analysis) 0.089*** 0.027 0.024 0.015 0.015 Alters' material wealth (factor analysis) -0.111*** 0.002 0.008 0.005 0.009 Proportion of alters are Christian -0.395*** -0.229* -0.192† -0.207† -0.182† Experience Healed by traditional healer (yes)* 0.482*** 0.391*** 0.289** 0.341*** 0.283** Healed by doctor (yes)* -0.249** -0.403*** -0.332*** -0.393*** -0.355*** Average probability of alters supporting ethnomedicine 0.845*** 0.291** 0.197*

Constant 2.152*** 1.784** 1.722** 1.283† 1.609** 1.345*

LL -649.48 -637.15 -627.68 -624.08 -621.13 -619.66 Wald 138.85*** 155.5*** 166.59*** 186.45*** 174.12*** 186.35*** AIC 1331 1320 1303 1298 1304 1303 BIC 1413 1439 1427 1427 1464 1468 Variance household group (std. error) .767 .579 .544 .249 .476 .296 (.219) (.188) (.188) (.162) (.168) (.171) LR v. logit (p) 37.65 (.000) 25.43 20.43 3.86 17.05 5.18 (.000) (.000) (.025) (.000) (.011) Source: The Niakhar Social Network and Health Project (2014) Notes: † p<0.1, * p<0.05, ** p<0.01, *** p<0.001, final analytic size: 1,284, restricted the prediction to the estimation subsample * Reference: all others (no, not completely, don’t know, nobody in the household has ever consulted a traditional healer, no response, missing.) ** Reference: all others (very bad, bad, acceptable/mediocre, don’t know, no response, missing.

Figure 1. Predicted probability of ethnomedical classification by average network likelihood of classification and global ego network clustering, interaction model 0.6 global 0.5 transitivity 0.4 0.3 0.24

classification 0.2 0.45 0.1 0.68 0

Predicted probabiltyPredicted of etnomedical 0 0.2 0.4 0.6 0.8 Average network probability of ethnomedical classification

References

Angel, R., & Thoits, P. (1987). The impact of culture on the cognitive structure of illness. Culture, Medicine and Psychiatry, 11(4), 465–494. Aral, S., Muchnik, L., & Sundararajan, A. (2009). Distinguishing influence-based contagion from homophily-driven diffusion in dynamic networks. Proceedings of the National Academy of Sciences of the United States of America, 106(51), 21544–21549. Barnes, B., & Bloor, D. (1982). Relativism, Rationalism and the Sociology of Knowledge. In M. Hollis & S. Lukes (Eds.), Rationality and Relativism. Blackwell. Behrman, J. R., Kohler, H.-P., & Watkins, S. C. (2002). Social networks and changes in contraceptive use over time: evidence from a longitudinal study in rural Kenya. Demography, 39(4), 713–738. Berkman, L. F., & Kawachi, I. (2014). Social Network Epidemiology. Social Epidemiology (2nd ed., pp. 234–289). New York: Oxford University Press. Blau, P. M. (1994). Structural Contexts of Opportunities (1st ed.). University of Chicago Press. D’Andrade, R. G. (1992). Schemas and Motivation. In R. G. D’Andrade & C. Strauss (Eds.), Human Motives and Cultural Models (pp. 23–44). Cambridge : Cambridge University Press. Dupire, M. (1994). Sagesse sereer: essais sur la pensée sereer ndut. KARTHALA Editions. Evans-Pritchard, E. E. (1937). Witchcraft, oracles and magic among the Azande. The Clarendon Press. Fabrega, H. (1975). The need for an ethnomedical science. Science, 189(4207), 969–975. Franckel, A., & Lalou, R. (2009). Health-seeking behaviour for childhood malaria: household dynamics in rural senegal. Journal of Biosocial Science, 41(01), 1–19. Goldsmith-Pinkham, P., & Imbens, G. W. (2013). Social Networks and the Identification of Peer Effects. Journal of Business & Economic Statistics, 31(3), 253–264. Gravrand, H. (1990). La civilisation Sereer - Pangool (Vols. 1-2, Vol. 2). Dakar: Nouvelles Editions africaines du Senegal. Green, E. C. (1999). Indigenous Theories of Contagious Disease. Rowman Altamira. Kleinman, A. (1978a). Concepts and a model for the comparison of medical systems as cultural systems. Social Science & Medicine. Part B: Medical Anthropology, 12, 85–93. Kleinman, A. (1978b). Culture, Illness, and Care: Clinical Lessons from Anthropologic and Cross-Cultural Research. Annals of Internal Medicine, 88(2), 251. Latkin, C. A., & Knowlton, A. R. (2015). Social Network Assessments and Interventions for Health Behavior Change: A Critical Review. Behavioral medicine (Washington, D.C.), 41(3), 90–97. Lin, X. (2010). Identifying Peer Effects in Student Academic Achievement by Spatial Autoregressive Models with Group Unobservables. Journal of Labor Economics, 28(4), 825–860. McKee, L. A. (2003). Ethnomedicine and enculturation in the Andes of Ecuador. In J. Koss- Chioino, T. L. Leatherman, & C. Greenway (Eds.), Medical Pluralism in the Andes. Psychology Press. McPherson, M., Smith-Lovin, L., & Cook, J. M. (2001). Birds of a Feather: Homophily in Social Networks. Annual Review of Sociology, 27(1), 415–444. Mead, G. H. (1967). Mind, Self, and Society: From the Standpoint of a Social Behaviorist. University Of Chicago Press.

Merton, R. K., & Kitt, A. S. (1950). Contributions to the theory of reference group behavior. In R. K. Merton & P. Lazarsfeld (Eds.), Continuities in Social Research: Studies in the Scope and Method of the American Soldier (pp. 40–105). New York: Free Press. Ngokwey, N. (1988). Pluralistic etiological systems in their social context: A Brazilian case study. Social Science & Medicine, 26(8), 793–802. Nichter, M. (1991). Ethnomedicine: Diverse trends, common linkages. Commentary. Medical Anthropology, 13(1–2), 137–171. Pachter, L. M. (1994). Culture and Clinical Care: Folk Illness Beliefs and Behaviors and Their Implications for Health Care Delivery. JAMA, 271(9), 690–694. Palloni, A. (2001). Diffusion in sociological analysis. In J. B. Casterline (Ed.), Diffusion Processes and Fertiltiy Transition. Washington, D.C.: National Academy Press. Sandberg, J. (2006). Infant Mortality, Social Networks, and Subsequent Fertility. American Sociological Review, 71(2), 288–309. Sandberg, J., Rytina, S., Delaunay, V., & Marra, A. S. (2012). Social learning about levels of perinatal and infant mortality in Niakhar, Senegal. Social Networks, 34(2), 264–274. Shalizi, C. R., & Thomas, A. C. (2011). Homophily and Contagion Are Generically Confounded in Observational Social Network Studies. Sociological methods & research, 40(2), 211– 239. Smith, K. P., & Christakis, N. A. (2008). Social Networks and Health. Annual Review of Sociology, 34(1), 405–429. Syed, I. B. (2003). Spiritual medicine in the history of Islamic medicine. J Int Soc History Islamic Med, 2(1), 45–9. Tarde, G. de. (1895). Les lois de l’imitation étude sociologique. Paris: F. Alcan. Valente, T. W. (2010). Social Networks and Health: Models, Methods, and Applications (1 edition.). Oxford ; New York: Oxford University Press. Valente, T. W. (2012). Network Interventions. Science, 337(6090), 49–53. Valente, T. W., Palinkas, L. A., Czaja, S., Chu, K.-H., & Brown, C. H. (2015). Social Network Analysis for Program Implementation. PLOS ONE, 10(6), e0131712. Vecchiato, N. L. (1997). Sociocultural Aspects of Tuberculosis Control in Ethiopia. Medical Anthropology Quarterly, 11(2), 183–201. Yoder, P. S. (1997). Negotiating relevance: belief, knowledge, and practice in international health projects. Medical anthropology quarterly, 11(2), 131–146.

Online Supplemental Material

Appendix: Latent classification of cultural models

As discussed in the main body of the paper, measures from the NSNHP survey module

concerning health ideation and behavior are used in this analysis to measure higher level

schemas – concerning religious belief and cosmology, mid-level schemas concerning cause of

disease, and lower level schemas concerning prevention and therapy. This appendix describes

these measures, their coding, and estimation of the latent class model from which the dependent

variable, ethnomedical classification, and the likelihood of ethnomedical classification among

social network alters are derived.

Higher-level schemas

The first question representing higher level schemas simply asked respondents what the most

important cause of illness among the Sereer people was, with ethnomedical and biomedical

choices. The second gauged general belief in the relative efficacy of ethnomedical and

biomedical healers, the third belief in traditional versus Western (Islamic and Christian) religious

leaders. Response choices for the first two questions included both biomedical and ethnomedical

causes, and respondents were asked to list the first, second and third most likely causes in a

forced preference rank order format. For the purposes of the latent class analysis, the first two

responses were used (the distribution of the third response being sparse for most variables) and

classified as purely ethnomedical when two ethnomedical options were chosen,

ethnomedical/biomedical when respondents selected an ethnomedical cause first, a biomedical cause second, ethnomedical/biomedical when an ethnomedical cause was chosen first, a

biomedical one second, and purely biomedical.

The third question, also forced preference rank order, was coded along an analogous spectrum from traditional to Western belief representing traditional belief only, traditional belief priority over Western religion, Western religious belief prioritized over traditional, and Western belief only. Among the remaining three single response format questions, the first asked respondents whether they believed in the efficacy of supplication to the Pangools for a good harvest, indicative of their general power outside the specific health context. The second asked whether supernatural (Satanic) illnesses could be treated by a medical doctor. As noted above, in our qualitative work, it was suggested by a number of respondents that this was possible, and the question aims to assess the degree to which this type of potential syncretism is associated with broader belief systems. The final indicator is an interviewer observation as to whether the respondent was wearing a mystical protection (in the form of a bracelet or necklace) at the time of the interview. Such amulets are believed in traditional religion to protect wearers from attacks from supernatural entities.

Mid-level schemas

The first two indicators of mid-level schemas are context/symptom scenarios concerning dizziness and headaches for a non-pregnant woman, and fatigue and muscle aches. The third asks respondents to assign responsibility for the death of a child, which, in our qualitative interviews was often seen as fundamentally unnatural and necessarily the responsibility of another, either through malevolence or neglect. These measures were chosen due to substantial heterogeneity in attribution to biomedical and ethnomedical causes as seen in our qualitative work. A fourth

question concerned an ideal-typical scenario of nguegne, the mystical ill-wind. All of these

questions, as the first three higher-level schema indicators, were force response rank ordered, and

coded so as to fall along the continuum from solely ethnomedical to solely biomedical.

Lower level schemas

Lower level schemas are represented by two questions concerning treatment of common symptoms, dry skin accompanied with a rash, and intestinal distress. The third question concerns the proper treatment the ideal-typical attack from a djinn or encounter with an embodied

Pangool, which our qualitative work suggests is accompanied by a cold sensation and chills. The

question concerning the most appropriate venue for childbirth was included because childbirth is

seen as a high-risk period of supernatural attack against both the mother and newborn, the best

prevention against with is seclusion of the mother during birth. The two questions concerning

preventative measures to be taken for good health include one which is general and includes options representing hygiene on the biomedical side, mystical protection on the ethnomedical side. The final question concerns preventative measures to be taken for pregnant women, which, for reasons detailed above, may be either biomedical or ethnomedical.

Table A1. Latent Class Model Fit, Health Ideation Schemas, Adults Age 16+, Yandé City (n=1310)

Number of VLMR-LR Latent AIC BIC l (df) Classes 764.91 (44) 2 41159 41620 -20490 (p=0.000) 358.38 (45) 3 40889 41583 -20311 (p=0.000) 199.27 (45) 4 40779 41706 -20211 (p=0.664) Source: NSNHP Panel 1, 2014: compiled by author

Table A2. Latent class probabilities, question wordings and response frequencies, rationale and biomedical/ethnomedical classification for variables in preferred 3-class solution. Mid-level schemas: diagnosis of disease as ethnomedical or biomedical Response categories/ Scale frequency Estimated probability of latent classification Question/Description (%) If a woman suffers from 1. Satanic diseases (eth.) Diagnosis: headaches, dizziness headaches, dizziness and she 2. Diseases associated with is not pregnant. In general, Pangools (eth.) what is the most probably 3. Natural infectious diseases (bio.) Ethnomedical 11% 20% 37% 32% cause among the list I am 4. Anemia or tension (bio.) about to read out to you? (and 5. Other diseases that are Sereer after that, and after that?) and non-Satanic (eth.) Liminal 6% 15% 34% 46% 6. Other diseases that can be healed Headaches and dizziness are by medical staff (bio.) alternatively perceived as Biomedical 3% 9% 22% 66% symptoms of attacks by 1. ethnomedical 5% Pangools, malevolent others, 2. ethnomedical/biomedical 14% 0% 20% 40% 60% 80% 100% the result of infectious disease, 3. biomedical/ethnomedical 29% anemia or hypertension or Ethnomedical Ethno/bio Bio/ethno Biomedical 4. biomedical 51% general fatigue If a child gets very ill and dies, 1. Nobody, it was a natural illness Diagnosis: child death according to you who is most (bio.) likely to blame for misconduct 2. It was his mother's fault (eth.) in the matter? (and after that, 3. It is the doing of bad (malevolent) Ethnomedical 39% 5% 27% 30% and after that?) people in the neighborhood (eth.) 4. It is the doing of other bad A pronounced cultural (malevolent) people (eth.) Liminal 25% 17% 25% 33% perception is that child death is unnatural, and may be caused 1. Others 29% by pollution of the mother 2. Others/no one 11% Biomedical 26% 9% 13% 51% through violation of totemic 3. No one/others 22% rituals, the intervention of 4. No one 37% 0% 20% 40% 60% 80% 100% malevolent others such as

agents of envious neighbors or Others Others/no one No one/others No one baby snatchers (koumoulass).

Table 1b. continued Response categories/ Scale frequency Estimated probability of latent classification Question/Description (%) A person complains about 1. A biff (nguegne) (eth.) Diagnosis: fatigue, muscle aches general fatigue and muscle 2. Sibirou (bio.) aches. According to you, what 3. Other non-Satanic diseases of the disease is she most likely Sereres (eth.) Ethnomedical 3% 23% 29% 45% suffering from? (and after that, 4. Other biomedical diseases (bio.) and after that?) 1. ethnomedical 0% Liminal 4% 11% 28% 57% Common symptoms of febrile 2. ethnomedical/biomedical 13% illness, they are often 3. biomedical/ethnomedical 14% associated with both 4. biomedical 72% Biomedical 4% 16% 15% 65% ethnomedical personalistic disease, ethnomedical 0% 20% 40% 60% 80% 100% naturalistic disease (Sibirou) as well as malaria and Ethnomedical Ethno/bio Bio/ethno Biomedical biomedical infectious disease. During rainy season, someone 1. Nguenge (eth.) Diagnosis: Nguegne complains about headaches 2. Sibirou (bio.) after having spent time by 3. Other non-Satanic diseases of the pools of rainwater while the Sereer (eth.) Ethnomedical 9% 29% 24% 38% night fell. What is the most 4. Other biomedical diseases (bio) probable cause of this person's disease? (and after that, and Liminal 4% 22% 29% 45% after that?) 1. ethnomedical 3% This is an ideal-typical 2. ethnomedical/biomedical 26% Biomedical 4% 24% 17% 54% description of symptoms 3. biomedical/ethnomedical 20% associated with nguegne, or 4. biomedical 50% 0% 20% 40% 60% 80% 100% ‘bad wind’, a supernatural

disease mixing personalistic Ethnomedical Ethno/bio Bio/ethno Biomedical and pollution aspects

Table 1b. continued Lower-level schemas: treatment and prevention Response categories/ Scale Estimated probability of latent classification Question/Description frequency (%) You had a cold sensation, 1. You'll take some pills at home Treatment: cold, shivers accompanied by shivers after (bio.) meeting someone, what will 2. You'll take some herbs or a herbal you do first? (second, third?) drink at home (eth.) Ethnomedical 59% 30% 9% 2% 3. You'll consult a clairvoyant or a This is an ideal-typical traditional healer (eth.) description of illness following 4. You'll consult a nurse or doctor Liminal 38% 34% 21% 6% an encounter with a (bio.) malevolent supernatural entity, most often a Djinn 1. ethnomedical 42% Biomedical 35% 17% 11% 37% 2. ethnomedical/biomedical 27% 3. biomedical/ethnomedical 14% 0% 20% 40% 60% 80% 100% 4. biomedical 14%

traditional trad/bio bio/trad bio

You feel itchy, have a rash and 1. You'll take some pills at home Treatment: Itchy rash, dry skin dry skin, what will you do first (bio.) to cure this? (and second? and 2. You'll take some herbs or a herbal third?) drink at home (eth.) Ethnomedical 6% 20% 36% 38% 3. You'll consult a clairvoyant or a These symptoms are perceived traditional healer (eth.) to have numerous causes, most 4. You'll consult a nurse or doctor Liminal 1% 8% 73% 18% ethnomedical/naturalistic. The (bio.) most prominent is a congenital disease known as ‘southiet’, 1. ethnomedical 2% Biomedical 1%4% 7% 88% the symptoms of which, 2. ethnomedical/biomedical 11% though treatable through 3. biomedical/ethnomedical 41% 0% 20% 40% 60% 80% 100% biomedical therapy is only 4. biomedical 46% curable through ethnomedical Ethnomedical Ethno/bio Bio/ethno Biomedical therapy. Table 1b. continued

Response categories/ Scale Estimated probability of latent classification Question/Description frequency (%) What do you do first when you 5. You'll take some pills at home Treatment: intestinal distress have a sore stomach (burns, (bio.) aches, diarrhea)? (and second, 6. You'll take some herbs or an and third?) herbal drink at home (eth.) Ethnomedical 6% 46% 27% 22% 7. You'll consult a clairvoyant or a Common symptoms ascribed traditional healer (eth.) to both ethnomedical and 8. You'll consult a nurse or doctor Liminal 0% 23% 62% 14% biomedical etiology. (bio.)

1. ethnomedical 2% Biomedical 1% 11% 7% 80% 2. ethnomedical/biomedical 26% 3. biomedical/ethnomedical 34% 0% 20% 40% 60% 80% 100% 4. biomedical 37%

Ethnomedical Ethno/bio Bio/ethno Biomedical

Where would you prefer that 1. Alone at home (eth.) Treatment: childbirth location you (or your spouse) should 2. With help from other women at give birth? home (eth.) 3. With help from traditional Ethnomedical 9% 24% 5% 63% Cultural preference is midwife at home (eth.) traditionally for a woman to 4. At the clinic (bio.) give birth at home. Those Liminal 6% 8% 2% 84% more concerned with protection of infant from 1. Alone at home 7% personalistic attack, or without 2. With other women 13% Biomedical 5% 9% 0% 86% family are more likely to give 3. Midwife 2% birth alone. Other women in 4. Health clinic 77% 0% 20% 40% 60% 80% 100% the household and traditional midwives are intermediary Alone in home With household women with midwife medical center

Table 1b. continued Response categories/ Scale Estimated probability of latent classification Question/Description frequency (%) What are the best preventive 1. Get a talisman (gris-gris) to Prevention: pregnant woman practices a woman can adopt protect herself (and the infant) when she is pregnant? (eth.) 2. Limit her mobility (eth.) Ethnomedical 13% 30% 41% 17% Pregnant women and their 3. Conduct prenatal care visits at the unborn children are perceived clinic (bio.) to be particular targets of 4. Observe proper nutritional habits Liminal 1% 14% 33% 52% personalistic attack. Cultural (bio.) norms prescribe isolation and mystical protection against 1. ethnomedical 5% Biomedical 1% 17% 34% 48% these. The biomedical system 2. ethnomedical/biomedical 19% advocates for pre-natal testing 3. biomedical/ethnomedical 36% 0% 20% 40% 60% 80% 100% and nutritional 4. biomedical 40% supplementation. Ethnomedical Ethno/bio Bio/ethno Biomedical

What's the most effective way 1. A protective talisman (gris-gris) Prevention: illness to protect oneself against (eth.) illnesses? (and the next most 2. Washing your hands regularly effective after that?, and after (bio.) Ethnomedical 7% 20% 32% 41% that?) 3. Avoiding going out at wrong hours (eth.) Contrast between protection 4. Prepare meals with care (bio.) Liminal 1% 9% 20% 70% against personalistic attacks and nguegne and hygiene 1. ethnomedical 3% practices. 2. ethnomedical/biomedical 12% Biomedical 1% 10% 20% 70% 3. biomedical/ethnomedical 24% 4. biomedical 61%

0% 20% 40% 60% 80% 100%

Ethnomedical Ethno/bio Bio/ethno Biomedical

Table 1b. continued Higher-level schemas: treatment and prevention Response categories (coding)/ Scale Estimated probability of latent classification Question/Description frequency (%) In your opinion, what is the 1. Satanic diseases (eth.) Most important cause of illness most important disease among 2. Diseases associated with the Sereer in the list that I will Pangools (eth.) 3. Natural diseases that can be read to you now? (And what is Ethnomedical 65% 17% 13% 6% healed by traditional healers the next important one after (eth.) that?) 4. Diseases that can be healed by medical staff (bio.) Liminal 66% 19% 12% 3% Division of cause of illness between ethnomedical personalistic 1. ethnomedical 41% Biomedical 56% 20% 12% 11% (Satanic/Pangool), 2. ethnomedical/biomedical 32% ethnomedical naturalistic and 3. biomedical/ethnomedical 17% 0% 20% 40% 60% 80% 100% biomedical 4. biomedical 8%

Ethnomedical Ethno/bio Bio/ethno Biomedical

If you had to classify by order 1. Ancestors Religious affiliation of importance that which you 2. Islamic religious leader believe in most, what would it 3. The priest or pastor be? (and the next important, 4. The traditional healer Ethnomedical 7% 17% 50% 26% and the next important?) 5. Others

1. traditional 3% Liminal 1%2% 27% 69% Aimed at assessing 1. traditional/Western 7% cosmological beliefs, 2. Western/traditional 33% contrasting ancestors and 4. Western 56% Biomedical 1%3% 25% 71% healers (associated with traditional cosmology) with 0% 20% 40% 60% 80% 100% contemporary Abrahamic monotheisms. traditional trad/mono mono/trad montheist Table 1b. continued

Response categories (coding)/ Scale frequency (%) Response categories (coding)/ Response categories (coding)/ Scale frequency (%) Scale frequency (%) In your experience, who is the 1. Traditional healers (eth.) Therapeutic efficacy most effective healer? (and the 2. Religious healers (eth.) next? and after that?) 3. Doctors (bio.)

Ethnomedical 15% 18% 65% 2% General question concerning the efficacy of biomedical and ethnomedical clinical therapy 1. ethnomedical 6% Liminal 2%7% 82% 9% 2. ethnomedical/biomedical 11% 3. biomedical/ethnomedical 72% 4. biomedical 9%

Biomedical 3% 8% 75% 14%

0% 20% 40% 60% 80% 100%

Ethnomedical Ethno/bio Bio/ethno Biomedical

Can libation offer to calm 0. No Belief in Pangools ancestors bring about good 1. Yes rains and a good harvest? 3. Don't know

Ethnomedical 47% 45% 8% Belief in the efficacy of 0. No 63% Pangools to intervene in 1. Yes 24% human affairs. 3. Don’t know 12% Liminal 64% 21% 15%

Biomedical 78% 9% 13%

0% 20% 40% 60% 80% 100%

no yes dk

Table 1b. continued

Response categories (coding)/ Response categories (coding)/ Scale Response categories (coding)/ Scale frequency (%) Scale frequency (%) frequency (%) Can satanic diseases be healed 0. No Satanic illness healed by biomedicine by a medical doctor? 1. Yes 3. Don’t know In the qualitative data Ethnomedical 69% 25% 6% collected, a number of 0. No 75% respondents indicated that this 1. Yes 13% was possible, a position in 8. Don’t know 11% Liminal 82% 8% 10% which some illnesses with ethnomedical causes may be Biomedical 73% 10% 17% treated with biomedical therapy, suggesting possible 0% 20% 40% 60% 80% 100% syncretism

no yes dk

(Interviewer observation); 0. No Was the respondent wearing 1. Yes Wearing mystical protection magical talisman? 0. No 83% Mystical protection in the form 1. Yes 17% Ethnomedical 68% 32% of amulets or bracelets are common in the zone, and must Liminal 89% 11% be obtained from an ethnomedical practitioner, suggesting behavioral Biomedical 92% 8% adherence to an ethnomedical model. 0% 20% 40% 60% 80% 100%

no yes