1 Introduction

Assortative matching and persistent inequality: Evidence from the world’s most exclusive marriage market

Marc Goni˜ 1

This Version September 2015

Abstract: Why do people marry similar others? Using data on peerage marriages in nineteenth century Britain, I find that search costs and segregation in the marriage market can generate sorting without resort to preferences. Peers typically courted in the London Season, a centralized marriage market taking place each year. As Queen Victoria went into mourning for her husband, the Season was interrupted (1861–63), raising search costs and reducing market segregation. I find that the cohort of women affected by the interruption were 80% more likely to marry a commoner. Sorting along landed wealth decreased by 50% within the peerage. My second contribution is to show that marital sorting can contribute to wealth inequality in the long-run. In 1870, landownership was more concentrated in counties where intermarriage had been lower since c.1500. Finally, I discuss how inequality delayed the provision of public schooling in England and Wales. JEL Codes: J12, C78, D63, N33. Keywords: Marital assortative matching, search costs, marriage market segmentation, wealth inequality.

It is a truth universally acknowledged, that a single man in possession of a good fortune, must be in want of a wife. Jane Austen, Pride and Prejudice.

1 Introduction

One of the hallmarks of modern marriage in OECD countries is that people tend to marry other people who have similar education, income, or social status (Chen et al. 2013). Marital assortative matching has important implications for inequality and income redistribution (Fernandez and Rogerson 2001, Fernandez et al. 2005,

1I am grateful to Hans-Joachim Voth for his advice. I also thank Ran Abramitzky, Dan Bogart, Davide Cantoni, Greg Clark, Deborah Cohen, Mauricio Drelichman, Jan Eeckhout, Nancy Ellen- berger, Gabrielle Fack, Ray Fisman, Sebastian Galiani, Gino Gancia, Albrecht Glitz, Regina Grafe, Alfonso Herranz-Loncan, Boyan Jovanovic, Peter Koudijs, Lee Lockwood, Joel Mokyr, Thomas Piketty, Giacomo Ponzetto, Jaume Ventura, John Wallis, and seminar participants at CREI, UPF, Northwestern, Paris School of Economics, University of Vienna, Leicester, HSE, Bank of Serbia, the 2012 MOOD, the 2012 EEA, the 2013 EHS, the 2013 Cliometrics, the 2013 SAEe, the 2015 RCI, and the 2014 EHA for helpful comments and suggestions. The Cambridge Group for the History of Population and Social Structure kindly shared data.

1 Greenwood et al. 2014). To investigate these implications further, it is crucial to first understand why do people marry similar others. Homophily — a preference for others who are like ourselves — is only one reason for assortative matching. In addition, the people we meet also circumscribes the set of mates we can chose from. In other words, every relationship not only reflects who we chose but also depends on who we meet. A robust prediction of marriage models is that search frictions affect marriage outcomes (Burdett and Coles 1997, Eeckhout 1999, Shimer and Smith 2000, Bloch and Ryder 2000, Adachi 2003, Atakan 2006, Jacquet and Tan 2007). Therefore, any process that reduces search costs for partners or affects choice sets of potential mates (e.g., online dating) could well lead to greater sorting and hence greater inequality. Confirming this prediction with data is not straightforward. Recent empirical work has used speed dating (Fisman et al. 2008), marriage ads in newspapers (Banerjee et al. 2013), or dating websites (Hitsch et al. 2010). Results are at odds with the theory — preferences appear to be an important determinant of sorting, but the matching technology does not seem to clearly affect all outcomes even when it reduces search costs and affects choice sets of potential mates. Does this discrepancy reflect flaws in search theory or in modern-day data? Dating is very different from marriage. In most cases, it does not reflect the long-term partnership formation at the core of search and matching theory. Relating marriage outcomes to the matching technology is also complicated by the fact that the latter is hard to measure. In modern marriage markets, members of the opposite sex continuously interact in a multitude of settings. As a consequence, it is virtually impossible to isolate a particular matching technology from other courtship processes. In this paper, I use a unique historical setting to investigate the causes of sorting in marriages and its implications for the distribution of wealth. In the nineteenth century, from Easter to August each year, a string of social events was held in London to “aid the introduction and courtship of marriageable age children of the nobility and gentry”2 — the London Season. Courtship in noble circles was largely restricted to London; in most cases, the only place where a young aristocrat could speak with a girl was at a ball during the Season. Guests were carefully selected by social status, and the high cost involved in participating even excluded aristocrats if they were pressed for money.3 In economic terms, the Season effectively reduced search costs

2Motto of the London Season at londonseason.net 3The cost was driven by the need to host large parties in a stately London home; only those who

2 for partners and, by segmenting the market, restricted the choice set of potential mates. Crucially, the Season was interrupted by a major, unanticipated, exogenous shock: the death of Prince Albert. When Queen Victoria went into mourning, all royal dinners, balls, and luncheons were cancelled for three consecutive years (1861– 63). I use this large shock to identify the effects of the Season on marriage outcomes. If search costs and marriage market segmentation can generate assortative matching, the interruption of the Season should reduce marital sorting. If, instead, sorting is mainly driven by preferences, marriage outcomes should remain largely unaffected. I find a clear, strong link between search costs. marriage market segmentation, and marital sorting. Using a combination of hand-collected and published sources on peerage marriages, I find that the cohort of women affected by the interruption of the Season were 80 percent more likely to marry a commoner, sorting along landed wealth decreased by 50 percent, and spouses came from geographically adjacent places. In addition to this quasi-experimental approach, I estimate an instrumental variables model with the interruption of the Season and changes in the size of the marriageable cohort as sources of exogenous variation in attendance at royal parties in 1851–75. I find that a 5 percent increase in attendance to the Season decreases by 1 percentage point the probability of marrying a commoner, increases by 1.5–2 percentage points the probability of marrying a spouse from a family with similar landed wealth, and increases in 3.4 miles the distance between spouses’ family seats. My second main contribution is to examine whether these sorting patterns hamper social mobility and contribute to wealth inequality. A counterfactual analysis shows that if the Season had not existed, marriages between peers’ daughters and commoners’ sons would have been 30 percent higher between 1851–75. The institutional innovation of the Season, thus, helped the British political and economic elite erect an effective barrier that kept out newcomers. Because marriage is important for the intergenerational transmission of property, marital sorting significantly contributed to wealth inequality. I link information on 10,364 marriages spanning 11 generations between 1531 and 1880 to county-level measures of landownership concentration in the 1870s. I find that a 30 percent increase in intermarriage — the estimated effect of suppressing the Season — would be associated to a 3 points reduction in the Gini index for land inequality and to and to a 50 percent reduction in the share of land in the hands of peers. issued invitations to balls, dinners, and luncheons could expect to receive them.

3 The top of the wealth distribution in England has remained largely unchanged from 1800 (Clark and Cummins 2015). Compared to the elite of many other countries, the British aristocracy not only “held the lion’s share of land, wealth, and political power in the world’s greatest empire” (Cannadine 1990),4 its members towered over their continental cousins in terms of exclusivity, riches, and political influence. Did this persistent political and economic inequality — result of entry barriers like the Season — have any lasting consequences? I compare the outcomes of Forster’s Edu- cation Act (1870) — the first attempt to introduce state education in England and Wales — across counties. In line with the predictions of Sokoloff and Engerman(2000) and Galor et al.(2009), the evidence suggests that counties where landownership was more concentrated systematically under-invested in state education. The empirical setting I examine offers a number of advantages. First, it allows me to open the “black box” of the matching technology. Marriage markets today are typically informal. We can only guess who is on the market and who meets whom. In contrast, in the Season, young ladies were even presented to the Queen at court, announcing of who was on the marriage market each year. Second, Prince Albert’s death generates as good as random assignment into the Season. His demise and especially Queen Victoria’s long mourning were unexpected (Ellis 1977, Hobhouse 1983). Furthermore, cohorts that did and did not marry during the interruption of the Season are essentially identical in terms of observables. Selection based on unobservables is unlikely, as there was a high pressure to marry young in Victorian Britain: if a lady was not engaged two or three Seasons after “coming out” into society, she was written off as a failure (Davidoff 1973: 52). Finally, in contrast to studies using modern-day data, this historical case study allows me to examine whether marital sorting contributes to wealth inequality in the long-run. FigureI illustrates the approach of the paper. The chart shows the effect of the interruption of the Season (1861–63) on the rate of intermarriage between peers and commoners. Bars plot the number of people attending royal parties in the Seasons between 1859 and 1867. The solid line indicates the percentage of marriages outside the peerage. It is presented as a ratio of the rate for women older than 22 in 1861 relative to women below this cutoff age. I separate these two groups because one would not expect very young ladies to be severely affected by the interruption of the

4Around 1880, fewer than 5,000 landowners still owned more than 50 percent of all land (Can- nadine 1990).

4 Season; they could simply delay their choice of husband until everything went back to normal. However, women aged 22 and over in 1861 could not wait long if they wanted to avoid being written off as failures based on the social norms of the time.5 Before Albert’s death and after the Season resumed, women in both age groups were equally likely to marry a commoner. However, a great gap between the two opens after 1861. Those who had to marry when the Season was disrupted were 80 percent more likely to marry a commoner than younger ladies who could wait for the Season to resume. This suggests that the Season was highly effective as a matching technology — by announcing who was on the market, creating multiple settings for the opposite sexes to meet, and segregating the rich and powerful from the poor and insignificant, it crucially determined who married whom. 5000 1.8 4000 1.6 3000 22 over < 22) 1.4 ≥ attendees 2000 1.2 ratio (age 1000 1 0 1859 1860 1861 1862 1863 1864 1865 1866 1867 year

Royal parties % marrying outside peerage

Figure I: The interruption of the Season on the relative probability of marrying a commoner. Shaded bars show the number of attendees at royal parties in the Season (left axis). For the connected line, the sample is all 276 peers’ daughters ﬁrst marrying in 1859–67. Diamonds display the rate of intermarriage with commoners for women older than 22 on January 1, 1861, relative to women below this cutoﬀ age (right axis).

This paper can be seen as integrating two literatures. The ﬁrst is the rich literature on assortative matching and the importance of search costs. Starting with the seminal works of Gale and Shapley(1962) and Becker(1973), economists have devoted considerable attention to understanding the conditions for positive assorta-

5In the early 1860s most ladies married when between the ages of 22 and 25. Since the older cohort would be 25 or more when the Season resumed in 1864, waiting was not an option for them.

5 tive matching. A classic insight from the assignment literature is that a reduction in search costs increases sorting in the matching process (Collin and McNamara 1990, Burdett and Coles 1997, Eeckhout 1999, Bloch and Ryder 2000, Shimer and Smith 2000, Adachi 2003, Atakan 2006). In addition, Bloch and Ryder(2000) and Jacquet and Tan(2007) conclude that limiting people’s choice set to those who are most similar reduces the congestion externality — the time an agent spends meeting people with whom she will never match — which in turn increases sorting. Surprisingly, these well-accepted theoretical insights lack clear-cut empirical support. Hitsch et al.(2010) estimate mate preferences from a dating website and conclude that “sorting [in dates] arises in the absence of search frictions” (p.162). Fur- ther, online daters sort less along education and ethnicity than the general population searching offline. This suggests that lower search costs online and search filters, which may restrict choice sets, actually decrease sorting. Lee(forthcoming) obtains similar results in the context of a Korean online matchmaking agency. Using marital adver- tisements in Indian newspapers, Banerjee et al.(2013) conclude that search frictions play little role in explaining caste-endogamy on the arranged marriage market. Fis- man et al.(2008) finds, similarly, very few matches between individuals from different ethnic groups in speed dating, even if the events were designed such that people of different ethnic groups met at a high rate. These results, at odds with the predictions of search theory, may be explained by the fact that the preferences of online dating users or speed-dating participants differ from those of the general population. A few studies use observational data on marriages to show that, to a large extent, marital sorting is the result of marriage market segmentation and search costs. Using Danish administrative data, Nielsen and Svarer(2009) find that around half of educational sorting is explained by the tendency to marry someone who went to the same or a nearby school. Abramitzky et al.(2011) show that after World War I, French males married up6 to a greater extent in regions where more men had died in the trenches, i.e., where women encountered men at a lower speed because of biased sex ratio.7

6That is, they married spouses of higher socio-economic status. 7Another set of related papers test whether marriage markets display increasing returns to scale. Gautier et al.(2010) find that cities are a more attractive place to live for singles because they offer more potential partners. In contrast, Botticini and Siow(2011) find that sorting patterns do not differ between city and countryside in the United States, China, and early renaissance Tuscany. In contrast to cities, the Season not only pooled large numbers of eligible singles together, but it was also meant to facilitate courtship. In this respect, the London Season provides a better test bed for

6 The second literature that motivates this paper focuses on the relation between marital sorting and inequality. Kremer(1997) concludes that even large increases in marital and residential sorting would not affect the distribution of income. This result has been contested by Cancian and Reed(1998), Fernandez and Rogerson (2001), Schwartz(2010), and Fernandez et al.(2005), who establish a theoretical and empirical correlation between sorting in the marriage market and income inequality. In a recent paper, Greenwood et al.(2014) document an increase in marital sorting since the 1960s. According to the authors, this trend is behind the much talked-about increase in income inequality in the United States. Whether this will crystallize into wealth inequality remains an open question, as the limited time-span of these studies precludes them to predict what will happen in three or four generations. Because I consider a historical setting and my data spans 11 generations from 1531 and 1880, my paper is the first to evaluate whether sorting contributes to wealth inequality in the very long-run. My paper is not the first to analyze long-run inequality in Britain. Using historical tax records, Piketty and Saez(2006) find a trend of increasing inequality over the last 25 years. Miles(1993), Mitch(1993), and Long and Ferrie(2013) show low rates of occupational mobility in the nineteenth century. Consistently, Clark and Cummins (2015) analyze wealth at death for people with rare surnames to show that wealth is persistent across generations. My paper suggests that marital sorting is behind the persistence at the top of the wealth distribution of several families. In this paper, I also discuss whether wealth inequality is harmful for growth. In- equality may lead to excessive taxation (Persson and Tabellini 1994), affect individual investments in human capital (Galor and Zeira 1993, Fernandez and Rogerson 2001), or delay the provision of public education (Sokoloff and Engerman 2000, Galor et al. 2009). I shed light on this channel by discussing how landownership concentration affected the provision of public schooling in England and Wales after 1870s. Finally, this study is also important because it adds to our understanding of elites. When sufficiently political powerful, elites may support policies pernicious to economic growth (Acemoglu 2008). Thus, understanding the institutions that help elites persist is crucial. The Season often serves as an example of the taste for leisure matching technologies that are increasingly common today, such as online matching sites or speed dating events. The Online Appendix Section B.1 shows that the matching technology embedded in the Season displays increasing returns to scale.

7 cultivated by upper-class families (Doepke and Zilibotti 2008). Allen(2009) suggests that the pomp associated with the aristocratic lifestyle was in fact a sunk investment to ensure trustworthy service to the Crown. This allowed the peerage to effectively rule England from 1550 to 1880. In line with this hypothesis, I argue that the Season was a crucial institution allowing the elite to erect entry barriers against newcomers (Stone and Stone 1984, Spring and Spring 1985, Wasson 1998). Relative to the existing literature, I make the following contribution: first, this paper is one of the first to provide empirical evidence that search costs reduce sorting in marriages. Marriage market segmentation also plays an important role, which calls for further research on this element by the theoretical search literature. Second, my results suggest that marital sorting contributes not only to income inequality — a well-established result in the literature — but also to a more unequal distribution of wealth in the long-run. This may have lasting economic consequences by distorting the provision of public goods such as education. Third, this paper unveils one of the mechanisms that shaped the class structure in Victorian Britain, helping sustain the British nobility’s role as an unusually small, exclusive, and rich elite. The article proceeds as follows. Section2 portraits the Season and the historical background. Section3 describes the data. Section4 presents the empirical results using the interruption of the Season (1861–63) and changes in the size of the cohort as sources of identifying variation. In Section5 I investigate the implications of marital sorting for social mobility and wealth inequality in the long-run. I also discuss how inequality affected the provision of schooling. Finally, Section6 concludes.

2 Historical Background

In this section, I describe the institutional arrangements that, in combination, con- stituted the London Season. The London Season arose sometime in the seventeenth century (Davidoﬀ 1973). British peers typically resided in isolated manors on their countryside estates. From February to August, however, they moved to London with their families to attend Parliament. This period was known as the London Season. Although Paris and Vienna developed similar arrangements (e.g., the Rallye), they never eclipsed the London Season. Why did such a Season not emerge in continental Europe? Continental noblemen were not as rural as British peers. Also, most parliaments did not meet as regularly as in Britain, so continental aristocracy did

8 not annually migrate to the capital. In addition, primogeniture and entailment allowed the peerage and gentry to remain small enough that these meetings in London were possible. Around 1900, only 1 in 3,200 people in Britain was an aristocrat. In comparison, the proportion in continental Europe was 1 in 100 (Beckett 1986: 35-40). The Season peaked between the 1800s and the 1870s (Ellenberger 1990). Figure II plots more than 4,000 movements into and out of London by members of the “fashionable world,” as was reported in the Morning Post in 1841. At the beginning of the year, most people were out of town. The biggest influx came at the end of January when Parliament convened and anyone who was anyone in the elite moved from their country seats to London. On April 20, the Queen joined, and the first debutante was presented at court. This marked the commencement of the main Season; it was the most crowded time of year. Many social events designed to introduce bachelors to debutantes took place. For example, on May 15 — the day of the royal ball at Buckingham — more than 800 “fashionable” families were in London. The Season was officially over by August 12, when the shooting season started and most peers moved back to their country estates. This seasonal migration was repeated annually. What was the purpose of the Season? Although in 1841 seasonal migrations co- incided with the Parliamentary calendar, cumulative inflows peaked between Easter and August, when most of the social events crucial to the “matching process” took place. In addition, Sheppard(1977) notes that families that were not prominent in politics, such as the earls of Verulam and Wilton, also showed the same migration pattern, indicating that the Season provided opportunities other than political lob- bying.8 The unspoken purpose of these festivities was to bring together the right sort of people, thus “providing the setting for the largest marriage market in the world” (Aiello 2010). The Season became crucial in the nineteenth century, when

arranged marriages were no longer acceptable so that individual choice must be carefully regulated to ensure exclusion of undesirable partners Under such a system it was vital that only potentially suitable people should mix. To meet these ends, balls and dances became the particular place for a girl to be introduced into Society. (Davidoﬀ 1973: 49)9

8One can presume that the Parliamentary motive played a secondary role. Parliament sessions were adjourned when the Derby took place. As Harper’s Monthly Magazine stated once, “The Season depends on Parliament, and Parliament depends on sport” (May issue, 1886; quoted in Aiello 2010). 9Dowries also became old-fashioned in the nineteenth century. For example, they disappeared from Edward Cave’s Gentleman’s Magazine, which used to report them and any gossip that could capture the reader’s attention (Cannon 1984: 73).

9 Royal ball 800 600

Queen arrives to London End of the Season 400 Opening of Parliament 200 Number of families 0 −200 Jan Feb Mar Apr May Jun Jul Ago Sep Oct Nov Dec

Arrivals Departures Cummulative

Figure II: The Season in 1841. This ﬁgure plots over 4,000 movements into and out of London of families participating in the Season of 1841, as reported in The Morning Post. The solid line shows the total number of families in London in each week, i.e., arrivals after subtraction of departures. Since departures were underreported, I multiply them by a factor of 1.6. Source: Sheppard(1977).

To restrict the pool of singles, most of the social events took place in the homes of the elite, who carefully selected their guests based on status (Davidoff 1973). Public meeting places like Ranelagh or Hurlingham closed down, and the “fashionable world” put a stop to masked balls, easily gate-crashed by commoners (Ellenberger 1990: 636). The expenses required to participate in the Season also restricted access. Very few could afford renting a house in Grosvenor Square or organizing a ball for hundreds of guests. The Duke of Northumberland spent £20,000 in the Season of 1840 (Sheppard 1971), at a time when a bricklayer earned 6 shillings (3/10 of a pound) a day (Porter 1998: 176). The arrangement also excluded impoverished peers who, after generations of gambling or mismanagement, were hard-pressed for money. The race to find a proper husband started with presentation at court. Lucy, daughter of the 4th Baron Lyttelton, described it as “a very memorable day” and a “moment of great happiness.”10 It was, in practice, a public announcement of who was on the marriage market each year. To be eligible, a lady had to be sponsored by

10Diary of Lady Cavendish, June 11, 1859.

10 someone who had already been accepted in the royal circle, usually her mother. After being presented at Court and before returning to Hagley Hall, the family seat in Worcestershire, Lucy attended countless evening parties, concerts, and balls, where she danced with the most eligible bachelors. She even participated in a royal ball at Buckingham Palace, where she thought her heart “would crack with excitement!”11 This stressful routine was not uncommon. Lady Dorothy Nevill attended “50 balls, 60 parties, 30 dinners and 25 breakfasts” (Nevill 1920) in her first Season. It was usual to start the day with a ride across Hyde Park at 10 am and end up at 3 am at a ball (Malheiro 1999). This whirlwind of social events facilitated encounters between singles. Ascot races were “the Eden of debutantes and the milliners’ harvest”.12 Royal parties were the most exclusive events, giving “a stamp of authority to the whole fabric of Society” (Davidoff 1973: 25). Many ladies met their future husbands at these balls, which were described as “mating” rituals (Inwood 1998). The pressure to get married was enormous. Ladies had only two to three Seasons to get engaged to a suitable partner. If they failed, they were considered “on the shelf” (Davidoff 1973: 52).13 If they succeeded, they would be fully integrated in society, enjoy future Seasons with their husband, and eventually, take their daughters to their own court presentation. Marriages were not typically love matches but based on money or eligibility. This is reflected in decorum rules that prevented a girl from dancing more than three times with one particular partner or sitting out a dance with a bachelor (Davidoff 1973: 49). Adultery was commonplace. Oscar Wilde wrote that the Season “is too matrimonial. People are either hunting for husbands, or hiding from them.” (An Ideal Husband: first act). Marriage was considered

[n]ot so much an alliance between the sexes as an important social definition; serious for a man but imperative for a girl. It was part of her duty to enlarge her sphere of influence through marriage. (Davidoff 1973: 50) The demise of the Season is inextricably linked with the decline of the nobility. Their immense economic power rested on a simple foundation: landed wealth. Ac- cording to Cannadine(1990), protection from foreign competition and light taxes 11Diary, June 29, 1859. 12Harper’s Monthly Magazine, 1886; quoted in Aiello(2010) 13The fate of Georgiana Longestaffe, a lady in her late 20s in Trollope’s The Way We Live Now, illustrates how much marriage prospects deteriorated as years went by. Georgiana “had meant, when she first started on her career, to have a lord; but lords are scarce [...] She had long made up her mind that she could do without a lord, but that she must get a commoner of the proper sort [...] But now the men of the right sort never came near her” (Ch. 32).

11 made agriculture very proﬁtable until the 1870s. After that, estates that could once support their mortgages — and their proprietors’ opulent lifestyles — fell into ruin. It was the death of the Season. Events became public, commoners were presented at court (Ellenberger 1990), and ﬁnally, “love laughed at lineage” (Turner 1954: 184).

3 Data sources

I use four data sources, two of which are newly computerized, and one of which is based on hand-collected archival documents. In detail, I complement the Hollingsworth genealogical data on the British peerage (2001) with geo-referenced information on family seats, landholdings, and gross annual rents from land from two published sources: Burke’s Heraldic Dictionary (1826) and Bateman’s Great Landowners (1883). Finally, to measure when the Season worked smoothly and when it was disrupted, I construct a new series of attendance at royal parties from the British National Archives.

3.1 Peerage records

Peerage records have chronicled the family histories of the British aristocracy since Arthur Collins published his Peerage in 1756. Hollingsworth(1964) collected all this genealogical material. He tracked all peers who died between 1603 and 1938 (primary universe) and their offspring (secondary universe).14 In 2001, the Cambridge Group for the History of Population and Social Structure re-digitised the 30,000 handwritten original index sheets. In its current form, the data comprises approximately 26,000 individuals. Each entry provides information about spouses’ vital events (date of birth, marriage, and death), social status, whether the husband was heir-apparent at age 15, and the status of the highest ranked parent. Status is presented in five categories: (1) duke, earl, or marquis, (2) baron or viscount, (3) baronet, (4) knight, and (5) commoner. The entries state whether a title is an English, Scottish, or Irish peerage. There is also information on the number of births and birth order. In this paper, I consider 273 peers’ daughters marrying in 1859–67 and 1,762 peers and peers’ offspring marrying in 1851–75.15 For my long-run analysis, I consider

14Online Appendix A details the sources used by Hollingsworth(1964) and shows the entry for Charles George Lyttelton from Cokayne’s Complete Peerage for the sake of illustration (Figure A.I). 15Note that the Hollingsworth(2001) dataset excludes the landed gentry, who also participated in the Season. The gentry were considered commoners and did not always attend the same parties; the

12 10,364 marriages between 1531 and 1880. Tables A.I and A.III in the Online Appendix provide summary statistics for these samples.

3.2 Family Seats

Unfortunately, the Hollingsworth(2001) dataset does not provide the birthplace or residence of individuals. To resolve this, I exploit the fact that each titled family was required to build a seat in their estate and to live there for most of the year.16 Family seats are recorded in heraldic dictionaries. The most relevant source for my purposes is Burke’s Heraldic Dictionary (1826). Most of the young aristocrats who married between 1851 and 1875 were recorded as presumptive heirs in this source. Therefore, the family seats in Burke(1826) correspond in general to the seats where the individuals under analysis grew up and lived most of the year. I gathered information on 694 country seats for 498 families linked to the peerage. Then, I geo-referenced these seats. Figure A.IV in the Online Appendix shows that the nobility was well dispersed all over the British Isles and that seats were quite isolated from each other. Merging this information with the Hollingsworth(2001) dataset gives me 351 couples that married in 1851–1875 for whom both seats are recorded.17 Note that, by construction, this sample only includes individuals marrying in the peerage. I construct two measures of geographic endogamy: ﬁrst, the probability of marrying in the same region. Regions are ﬁrst-level NUTS for England and Wales, electoral regions for Scotland, and provinces for Ireland. Second, the distance between spouses’ seats. When one or both spouses have more than one seat I take the minimum distance. Table A.I, Panel B in the Online Appendix provides summary statistics.

3.3 The Great Landowners

In Jane Austen’s Pride and Prejudice, Mr. Darcy is described as a wealthy gentleman with an income exceeding £10,000 a year and proprietor of Pemberley, a large estate

Season was not a uniform event but consisted of many “layers” (Wilkins 2010: 30). In this paper, thus, I focus on the layer for which marriage has the highest stakes — the peerage. 16On the importance of seats for the British aristocracy, see Stone and Stone(1984). 17Speciﬁcally, I match individuals using own title for men and parental title for women. When parental (own) title of a woman (man) is not available, I try to match it using own (parental) title. Moreover, some entries in the Hollingsworth dataset are labeled with two titles (e.g., James Richard Stanhope, 7th Earl of Stanhope and 13th Earl of Chesterﬁeld). In these cases, I record the country seats of both titles. With this methodology, all but four titles from Burke(1826) are matched.

13 in Derbyshire. The land rents and wealth of nonﬁctional aristocrats were also public knowledge thanks to Bateman(1883)’s The Great Landowners of Great Britain and Ireland. The book consists of a list of all owners of 3,000 acres and upwards by 1876, worth £3,000 a year. Also, 1,300 owners of 2,000 acres and upwards are included. Each entry states acreage and gross annual rents.18 The book also reports the alma mater of the landowner, the clubs to which he belonged, if he was in possession of woods, whether his family held land since Henry VIII, whether he took his seat in Parliament, and other services he provided to the Queen.19 I created a computerized database of all relevant information for the 558 men who appeared both in Bateman(1883) and in the Hollingsworth dataset and for the families of 130 women marrying in 1859–67. Table A.I, Panel C in the Online Appendix provides detailed summary statistics. For their spouses, I computerized the landholdings of any close relative listed in Bateman(1883). Using this procedure, I matched 130 husbands and 353 wives.20 The book also provides a table for each county showing the number of landowners and their cumulative acreage, all divided into eight classes: peers or peers’ sons (3,000 acres and upwards), commoner great landowners (3,000 acres and upwards), squires (1,000 to 3,000 acres), greater yeomen (300 to 1,000 acres), lesser yeomen (100 to 300 acres), small proprietors (one to 100 acres), cottagers (less than one acre), and public bodies. I use this to construct two county-level measures of wealth inequality: the Gini index for the distribution of land and the share of land owned by peers. Summary statistics are in Table A.III, Panel B in the Online Appendix.

3.4 Lord Chamberlain’s invitations

The British National Archives keep the original invitations issued to parties at Buck- ingham or St. James’s Palace during the London Season.21 I created a computerized database with the total number of invitations issued in 1851–75, the number of invi- tees attending and excused to each party, the type of event, and its date.

18Lord Overstone was outraged by Bateman’s underassessment of his landholdings (Bateman 1883, 348). In the 1870s, with a rising public clamor about what was called the “monopoly of land,” these complaints cannot but subscribe the view of landholdings as a signal of social position. 19Appendix A shows the entry for Charles George Lyttelton (Figure A.III). 20Seventy-two percent of the matched wives are daughters and sisters of great landowners. Acreage is similar across family relations. See Table A.II in the Online Appendix for details. 21British National Archives, LC 6/31-55 (1839–76), LC 6/127-156 (1877–1902), and LC 6/157-164.

14 On average, approximately 5,000 invitations were issued each year. FigureIII plots the numbers attending over time by type of event. The initial year, 1851, displays unusually high attendance rates, explained by the Crystal Palace Exhibition. After that, there seems to be an increasing trend. The largest party was on June 24, 1874, when almost 2,000 people attended a ball at Buckingham. Crucially, this evidence reveals a huge disruption to the Season between 1861 and 1863. This was the result of Queen Victoria’s mourning for the death of her husband, Prince Albert. 8,000 6,000 4,000 attendees 2,000 0 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875

Ball Breakfast and Afternoon party Children’s Ball Concert and Evening party Court

Figure III: Attendees at royal parties, by type of event. The sample comprises all 126 parties held at Buckingham, St. James’ Palace, and Windsor in the Seasons between 1851 and 1875. Balls include state and costume balls. Court refers to the Queen’s diplomatic and oﬃcial court.

4 Empirical analysis

This section presents the empirical results. First I describe marriage outcomes in 1851–75. Next, I establish a causal link between search costs, market segmentation, and sorting by exploiting the interruption of the Season after the death of Prince Albert. Finally, I estimate the magnitude of the eﬀects with an instrumental variables approach. I use two sources of exogenous variation in attendance to the Season: its interruption in 1861–63 and changes in the size of the marriageable cohort. To guide the empirical analysis, the Online Appendix D develops a basic conceptual two-sided search framework for understanding the the eﬀects of search costs and

15 market segmentation on sorting. The model builds on Burdett and Coles(1997) and incorporates endogenous market segmentation.22

4.1 Data descriptives

From about 1850 to the 1870s the Season was at its peak, balls were crowded, and presentation at court was considered the most important day in a girl’s life (Davidoff 1973, Ellenberger 1990). What did marriage outcomes look like? TableI describes sorting by social class for all peers and peers’ offspring marrying in 1851–75. The table cross-tabulates husbands’ and wifes’ social position. Each cell compares the observed percentage of marriages in each category with the expected percentage if matching was random. Most people married into their own class: 97 percent of dukes’ heirs married the daughter of a duke (Panel A). If matching was random only 41 percent of them would have married such an eligible bride. The same is true on the lower end: commoners at age 15 — but who ended their lives as peers — were 69 percentage points more likely to marry the daughter of a commoner than under random matching. A similar pattern emerges for peers’ daughters (Panel B). Overall, chi-squared tests confirm a positive significant relation between husband and wife’s rank. In other words, there was positive assortative matching in social status. Great landowners also tended to marry the daughters of great landowners. Figure IV shows the results of a kernel-weighted local polynomial regression of wifes’ family landholdings on husbands’ landholdings. Results suggest that both in terms of acreage (Panel A) and in terms of land rents (Panel B), richer individuals were more likely to marry spouses from well-accomplished families.23 In the Online Appendix A, I present additional descriptives. Table A.I, Panel B shows that in 1851–75 there was little geographic endogamy among peers. Only 21 percent of spouses came from the same region and, on average, their seats were 150 miles apart.24 Figure A.V shows that ladies were pressured to marry quickly as suggested by anecdotal evidence. A woman’s “market value” — measured as the rate

22In the Online Appendix, I also test for increasing returns to scale, evaluate the eﬀects of the Season on match quality, and examine the role of preferences (Section B). Further, I examine the robustness of the results (Section C). 23Furthermore, members of ancient families (or a junior branch of it) that held land in England since the time of Henry VIII were 15 percent less likely to marry a commoner than the average peer. 24Lucy Lyttelton described the journey from Hagley Hall to London (105 miles) as “most smutty,” facing “wind, rain, and dirt on the box [of the open britschka].” Diary, May 18, 1859.

16 of marriage with peers and dukes’ heirs and by husbands’ landholdings — decreased with age. Moreover, the devaluation for dukes’ daughters was steeper, suggesting that independently of social position, delaying marriage was not a good option.

Table I

Husband and wife’s social status (1851–1875)

Panel A. Marriage outcomes for peers and peers’ sons

Wife’s parental rank Com. Knight Baronet Baron Duke N Common 15 obs 84.6 5.1 10.3 0.0 0.0 175 exp 15.5 0.9 2.9 39.5 41.0 diff 69.0*** 4.2*** 7.4*** -39.5*** -41.0*** Baronet 15 obs 8.3 0.0 91.7 0.0 0.0 12 exp 15.8 0.8 3.3 39.2 40.8 diff -7.5 -0.8 88.3*** -39.2** -40.8** Peer son obs 0.0 0.0 0.0 50.2 49.8 574 exp 15.6 0.9 2.9 39.5 41.0 diff -15.6*** -0.9*** -2.9*** 10.6*** 8.8*** Baron heir obs 2.7 0.0 0.0 91.1 6.3 112 exp 15.5 0.9 2.9 39.6 41.1 diff -12.9*** -0.9 -2.9* 51.5*** -34.8*** Duke heir obs 1.7 0.0 0.0 0.9 97.4 116 exp 15.6 0.9 2.9 39.6 41.0 diff -13.9*** -0.9 -2.9* -38.7*** 56.4*** N 154 9 29 391 406 989 Pearson chi-squared (16): 1,500*** (0.00) Gamma test: 0.72** (0.03)

Panel B. Marriage outcomes for peers’ daughters

Husband’s rank at age 15 Com. Gentry Peer son Baron heir Duke heir N Baron dau. obs 61.1 14.2 6.3 10.8 7.6 380 exp 50.3 13.9 9.0 13.3 13.4 diff 10.7*** 0.3 -2.7* -2.5*** -5.8*** Duke dau. obs 39.8 13.7 11.6 15.8 19.1 387 exp 50.3 14.0 9.0 13.3 13.4 diff -10.5*** -0.3 2.6* 2.5*** 5.7*** N 386 107 69 102 103 767 Pearson chi-squared (4): 45.7*** (0.00) Gamma test: 0.34** (0.05) Note: This table cross-tabulates wives’ and husbands’ social position for all 989 peers and peers’ sons first marrying in 1851–75 (Panel A) and all 767 peers’ daughters marrying in 1851– 75 (Panel B). Wife’s social position is her father’s rank, husband’s is his rank at age 15. Common 15 and Baronet 15 were commoners at age 15 but ended their lives as peers. The Com. column category also includes individuals who were always commoners. Peer son excludes heirs. Baron includes baronies and viscountcies, and Duke includes dukedoms, earldoms, and marquisates. Each cell contains observed percentages, expected percentages if matching was random, and the difference below. P-values are in parentheses; *** p<0.01, ** p<0.05, * p<0.1

17 .This 1983 ). (Hobhouse duties her of and most grief-stricken, over was took Victoria Albert, died. Prince mother husband, Victoria’s her Queen 1861, 16, March (1859–67) On results Quasi-experimental Albert. Prince 4.2 of death the interrupted shock: was exogenous I Season and so, the unanticipated, do To when major, years costs. a three search by the reduced during and outcomes segregation market marriage of examine was result He the His extent, Ireland. large Britain. and a in Yorkshire, landowners exceed Derbyshire, to greatest Middlesex, said in was the income acres of 198,572 the one of was Devonshire, possession He of in 22. Duke was the She of Cavendish. Frederick son general Lord these married mirrors Lucy marriage and Lyteltton’s patterns. Ireland Lucy had marriage and young. families Britain marry over spouses’ to all and pressed from were class came women social spouses own addition, In their in landholdings. similar married they others: similar with intervals. conﬁdence 95% are lines Dashed (1883). Bateman in on listed also landholdings is worth wives’ upwards, and smoothing: acres polynomial (1851–75). landholdings local husbands’ Kernel-weighted IV: Figure nsm uigtegle g fteSao h hlrno h oiiybonded nobility the of children the Season the of age golden the during sum, In

wife’s family 30 40 50 60 70 0 A. Acreage(percentiles) 20 £ ,0 erb 86 rtmryn n15–5t iewoefamily whose wife a to 1851–75 in marrying ﬁrst 1876, by year a 2,000 40 husband £ 8,0 er et hwta uhmrigswr,to were, marriages such that show I Next, year. a 180,000 60 80 h apecmrssal15pesi ossino 2,000 of possession in peers 175 all comprises sample The 100 18

wife’s family 30 40 50 60 70 0 B. Landrents(percentiles) 20 40 husband 60 80 100 was the start to a disastrous year that would end with Albert’s death on December 14. Victoria plunged into deep grief. She wore black for the rest of her life and avoided public appearances as much as she could. From 1861 to 1863, most royal parties in the Season were cancelled. In 1862, the Queen suspended all court presentations (Ellenberger 1990). In sum, young ladies were not announced at court; poor and insigniﬁcant suitors were not fully screened out; and, because royal parties were cancelled, encounters became more costly.25 My identiﬁcation strategy is a quasi-experimental design that is based on compar- ing female cohorts marrying in the interruption of the Season (1861–63) to cohorts marrying in the years before and after (1859–60 and 1864–67). Formally, I estimate:

(4.1) χ(yi,T =1, yj,T =1|T = 1) − χ(yi,T =0, yj,T =0|T = 1) ,

where χ(yi, yj) is a measure of association between wife’s (i) and husband’s (j) characteristic y. For category-based characteristics (e.g, social class), χ is a chi-squared test of association.26 For continuos variables (e.g., landed wealth), χ is a moment in the distribution of |yi − yj|. To compare diﬀerent moments, I use Kolmogorov- Smirnov tests of the equality of distributions across the treatment. The treatment, T , is equal to 1 if an individual married in the London Season, and 0 if she did not; The exogenous interruption of the Season after the death of Prince Albert gives me the appropriate counterfactual for χ(yi,T =0, yj,T =0|T = 1). As with every quasi-experimental approach, the identifying assumption is that subjects in the treatment and control groups are essentially identical. The principal threat to identiﬁcation, thus, is that least attractive women might select themselves into marriage during the interruption of the Season while attractive women anticipate their marriage plans or wait for the Season to resume. I can assess the relevance of selection based on observables. TableII shows that women that did and did not marry during the interruption of the Season are essentially identical in terms of observables. The proportion of dukes’ daughters marrying during the interruption of the Season was the same as in normal years. On average, ladies came from a family on the same percentile of the distribution of acreage. The proportion of daughters of English peers,

25Of course, the London Season was not restricted to royal parties and court presentations. How- ever, these events were central to the well-functioning of the marriage market. 26I use Pearson’s chi-square, likelihood ratio chi-square, Cramer’s V, G-test, and Kendall’s tau-c.

19 age at marriage, life expectancy, or birth order also do not vary across cohorts. In addition, the table suggests that Queen Victoria’s mourning was the only disruption to the marriage market in 1861–1863. Neither the size of the marriageable cohort nor sex ratios were distorted during this period.

Table II

Balanced cohorts: interruption of the Season versus normal years

No Season Normal years Diﬀerence 1861–63 1859–67† A. Characteristics at marriage (women) Age at ﬁrst marriage 24.7 24.4 0.4 (0.6 ) (0.4) (0.7) Life expectancy 69.0 67.5 1.5 (1.9) (1.4) (2.4) Birth order (excluding heir) 3.7 3.9 0.2 (0.3) (0.2) (0.4) Duke daughters 0.51 0.52 -0.01 (0.05) (0.04) (0.06) Baron daughters ref. ref.

Peerage of England 0.65 0.59 0.06 (0.05) (0.04) (0.06 ) Peerage of Scotland and Ireland ref. ref.

Family acreage (percentile) 49.4 50.6 -1.2 (3.1) (2.4) (4.0) B. Cohort characteristics Female cohort size (18–24) 264 261 3 (1.93) (3.06) (3.46) Sex ratio (men/women) 1.111 1.107 0.005 (0.010) (0.024) (0.021) Note: Panel A is for all 276 peers’ daughters ﬁrst marrying in 1859–67. Duke daughters and Peerage of England are proportions. Panel B are yearly averages. Sex ratio is the number of peers and peers’ sons aged 19–25 to peers’ daughters aged 18–24. When women are underreported, I use the sex ratio at birth (i.e., wom = 0.95 × men). Standard errors are in parentheses; † excludes the years of the interruption; *** p<0.01, ** p<0.05, * p<0.1

It might be that selection is based on unobservables. This scenario is unlikely for two reasons. First, in order to anticipate marriage plans or wait for the Season to resume, one should be able to predict the timing and the duration of the interruption. Prince Albert’s death was unexpected. He was only 42 when he died. Albert’s doctors diagnosed typhoid fever as the cause of his death. Only recently it has been discovered

20 that he had been ill for two years (Hobhouse 1983: 150–151). In addition, Albert took on important government duties until one month before his death.27 On the other hand, the peerage was also surprised by the length of the mourning. They complained that “after the lamented death of Prince Albert, the Queen came less and less to London, and the palace was more and more deserted” (Ellis 1977: 361). The second reason why selection is unlikely is the high pressure to marry young. If a lady was not engaged two or three Seasons after “coming out”, she was written off as a failure independently of her social status or her beauty (Davidoff 1973: 52). Figure A.V in the Online Appendix shows that marriage prospects after 22 deteriorate for all ladies. This fear was so deeply internalized that, as Queen Victoria’s mourning extended over time, ladies panicked and got married even if the Season was cancelled. FigureV confirms this. It plots hazard rates — the percentage of single women who get married at each age — for the marriageable cohort in 1861, separating younger (aged 18 to 22) from older ladies (aged 23 to 27). I also include the hazard rates of the marriageable cohort in 1851, when the Season worked smoothly. The figure suggests that only those who were very young in 1861 deferred marriage. The cohort aged 22 or more in 1861, instead, behave like the benchmark cohort of 1851. 15 10 hazard rate (%) 5

18−19 20−21 22−23 24−25 26−27 28−29 30−31 age

Marriageable cohort in 1861, young ladies Marriageable cohort in 1861, old ladies Marriageable cohort in 1851

Figure V: Deferred marriage decisions. This ﬁgure plots the marriage hazard rate at each age group for three cohorts of peers’ daughters: blue shows ladies aged 18 to 22 in 1861 (N=254), red shows ladies aged 23 to 27 in 1861 (N=262), and grey shows ladies aged 18 to 27 in 1851 (N=462).

27For example, on November 8, 1861, Union forces intercepted the British RMS Trent and removed two Confederate envoys. Albert intervened to soften the diplomatic response, lowering the threat of war (Hobhouse 1983: 154–155).

21 In sum, given that cohorts that did and did not marry during the interruption of the Season are identical in terms of observables, that the timing of the interruption was unanticipated, and the high pressure to marry young, I conclude that Queen Victoria’s mourning provides as good as random assignment into the Season. TableIII examines the effect of the interruption of the Season on sorting by social class. The table cross-tabulates husbands’ and wives’ social position for all peers’ daughters marrying in 1859–67. Panel A is for ladies marrying during the interruption of the Season, Panel B for those marrying in the years before and after. When the Season was working smoothly, peers’ daughters sorted in the marriage market: 21 percent of dukes’ daughters married a duke heir. If matching was random, we would only expect 13 percent of these marriages to take place. In contrast, barons’ daughters were 10 percentage points more likely to marry a commoner than under random matching. The patterns are clearly different when the Season was interrupted in 1861–63. Now, only 16 percent of dukes’ daughters marry dukes’ heirs, a very similar percentage to that obtained under random matching. For barons’ daughters, the pattern is similar. Panel C presents non-parametric tests of association between husbands’ and wives’ social position in the interruption of the Season and in normal years. The Person’s chi-squared statistic tests the null hypothesis that husbands’ and wives’ social position are independent. The null is rejected only when the Season was at place. In other words, without the Season marriage was random with respect to social position. The likelihood ratio test, which is more accurate for small samples, confirms this result.28 Cramer’s V measures the strength of this association on a 0 to 1 scale.29 In normal years, the relation between husbands’ and wives’ social position was large. The gamma-test and Kendall’s tau-c asses the direction of this relationship. In normal years, the relation was positive: higher ranked wives married better ranked spouses, and vice versa. That is, there was positive assortative matching. These results vanish when there is no Season: Cramer’s V indicates a weak relationship between spouses’ ranks, while the g-test and Kendall’s tau-c are not significantly different from zero. In sum, positive assortative matching disappeared when the Season was interrupted.

28In n by m contingency tables, the Person’s chi-squared test requires an expected cell count of 5 or more in 80 percent of cells, and no cells with zero expected count. This condition is satisﬁed. 29The eﬀect sizes for 4 degrees of freedom are: 0–0.15 “no relation”; 0.15–0.25 “weak”; 0.25–0.4 “strong”; > 0.4 “worrisomely strong” (the two variables are measuring the same).

22 Table III

The Season and sorting by social position (1859–67)

Panel A: Married in the Season, 1859–60 and 1864–67

Husband’s rank Wife: Com. Gentry Peer son Baron heir Duke heir N Baron’s dau. observ. 63.7 15.4 7.7 7.7 5.5 91 expect. 53.4 13.1 10.2 10.2 13.1 diﬀ. 10.3 2.3 -2.5 -2.5 -7.6** Duke’s dau. observ. 42.4 10.6 12.9 12.9 21.2 85 expect. 53.4 13.1 10.2 10.2 13.1 diﬀ. -11.1 -2.5 2.7 2.7 8.1** N 94 23 18 18 23 176

Panel B: Married during the interruption of the Season, 1861–63

Husband’s rank Wife: Com. Gentry Peer son Baron heir Duke heir N Baron’s dau. observ. 53.5 14.0 7.0 16.3 9.3 43 expect. 46.3 18.4 9.8 12.8 12.8 diﬀ. 7.2 -4.4 -2.8 3.5 -3.5 Duke’s dau. observ. 40.0 22.0 12.0 10.0 16.0 50 expect. 46.2 18.2 9.6 13.0 13.0 diﬀ. -6.2 3.8 2.4 -3.0 3.0 N 43 17 9 12 12 93

Panel C: Aggregate chi-square statistics

Season, 1859–67 No Season, 1861–63 Pearson chi2 15.18*** 3.84 Pr=0.004 Pr=0.43 df=4 df=4 Likelihood ratio chi2 15.68*** 3.89 Pr=0.003 Pr=0.42 df=4 df=4 Cramer’s V 0.29 0.20 Gamma test 0.41 0.17 ASE=0.104 ASE=0.157 Kendall’s tau-b 0.25* 0.101 ASE=0.066 ASE=0.094 Note: Panel A is a contingency table of wife’s parental rank and husband’s rank at age 15 for 176 peers’ daughters marrying when the Season was working smoothly in 1859–60 and 1864–67. Panel B is the corresponding contingency table for 93 peers’ daughters marrying during the interruption of the Season in 1861–63. Categories are described in TableI. Each cell contains observed percentages, expected percentages if matching was random, and the diﬀerence below. Panel C presents the aggregate chi-squared statistics; *** p<0.01, ** p<0.05, * p<0.1

Section C.1 in the Online Appendix tests the robustness of this result. I conduct a placebo test using cohorts marrying 10, 11, 12, ... 50 years before the death of Prince Albert. In contrast to the results in FigureI and TableIII, I ﬁnd that these cohorts

23 do not change their marriage behavior in the years corresponding to the placebo interruption of the Season (see Figure C.I). During the interruption of the Season, women also married markedly poorer spouses. FigureVI considers all peers’ daughters marrying in the peerage or the gentry in 1859–1867.30 Panel A plots the distribution of acreage of their husbands’ families. Women marrying when the Season was disrupted wed a husband in the 30th percentile of the landed wealth distribution. The years before and after the interruption, instead, the mode is in the 80th percentile. Sorting by landed wealth was also distorted. Panel B plots the distribution for the difference between husband and wife’s acreage, in absolute value. Consider the median couple when the Season worked smoothly: one spouse held 18,000 more acres than the other. Without the Season, the difference is 27,000 acres. That is, sorting along landed wealth decreased by 50 percent. The Kolmogorov-Smirnov test suggest that, overall, the distributions are significantly different. This evidence powerfully suggests that the Season — by reducing search costs and segregating commoners, induced the children of the nobility to sort more in the marriage market.

Panel A. Husband’s family acreage Panel B. Mismatch in acres .015

Season .01 Density .005

No Season 0 0 10 20 30 40 50 60 70 80 90 100 husband’s acreage (pecentiles) 0 20,000 40,000 60,000 80,000 Season (1859−67) No Season (1861−63) Absolute value of the difference in spouses’ family acreage Kolmogorov−Smirnov statistic = 0.2454* (p−value = 0.067) Kolmogorov−Smirnov statistic = 0.3458* (p−value = 0.014)

Figure VI: The effects of the interruption of the Season on sorting by landholdings. The sample includes 105 women: all peers’ daughters ﬁrst marrying in 1859–67 except women whose husband’s family is not listed in Bateman(1883) and women marrying commoners, because land does not proxy husbands’ wealth. Panel A plots the kernel density of husband’s family acreage in percentiles. In Panel B, boxes display the distribution of the diﬀerence between husband and wife acreage in absolute value. The adjacent lines are the lower and upper adjacent value of the distribution. The bottom and top of the box are the 25th and 75th percentile. The central line is the median. Outside values are excluded; *** p<0.01, ** p<0.05, * p<0.1

30Commoner husbands are excluded because land does not accurately proxy their wealth.

24 The interruption of the Season had large eﬀects on sorting by social position and landholdings without resort to preferences. Does this mean that preferences played a secondary role in pairing? Section B.3 in the Online Appendix explores this. Table B.II indicates that husbands from a liberal (tory) club married wives whose relatives were also members of a liberal (tory) club.31 Sorting in this dimension was not as strong as sorting by socio-economic status. However, it was not aﬀected by the interruption of the Season (Figure B.II), suggesting that it was driven by homophilic preferences, independent of search costs and marriage market segmentation. By centralizing the marriage decisions in London, the Season allowed singles from all over the country to court. Do peers turn back to the area around their country seats to search for a spouse during the mourning period? FigureVII suggests the answer is yes. The chart plots the number of attendees at royal parties along with the average distance between spouses’ family seats. In 1862 and 1863, spouses came from much closer places than in years when the Season worked smoothly. 32 200 5000 180 4000 160 3000 miles attendees 140 2000 120 1000 0 100 1859 1860 1861 1862 1863 1864 1865 1866 year

Royal parties Distance between spouses’ seats

Figure VII: The effects of the interruption of the Season on the distance between spouses’ family seats. Shaded bars show the number of attendees at royal parties in the Season (left axis). For the connected line, the sample is all 57 marriages in 1859–66 between peers’ sons and peers’ daughters whose family seats were listed in Burke(1826). The year 1867 is excluded because I could only calculate distance for 4 marriages.

31Brook’s, Reform, and Devonshire were liberal clubs; Carlton, Junior Carlton, Conservative, and St. Stephens were tories (Bateman 1883: 497) 32When a couple did not produce a son estates were inherited by a distant relative. Therefore, choosing a partner from the vicinity was not necessarily advantageous for estate consolidation. It only reﬂected whether peers searched for a partner in the local or in a centralized market.

25 These results suggest that search costs and market segmentation can generate sorting by socio-economic status without resort to preferences. When the Season announced who was on the market, created multiple settings for the opposite sexes to meet, and segregated the rich and powerful from the poor and insigniﬁcant, it strengthened the degree of sorting by socio-economic status. In contrast, when the Season was interrupted after Prince Albert’s death, local marriage markets became a more important marriage medium. These markets were more shallow, reducing the degree to which the children of the nobility could sort in the marriage market.

4.3 Instrumental variables results (1851–75)

In this subsection, I estimate the eﬀects of the Season using an instrumental variables approach. Yearly variation in the number of attendees at royal parties in the Season can be explained by many factors, some of which might be endogenous to marriage outcomes.33 Here I exploit two sources of exogenous variation in attendance: the interruption of the Season after Prince Albert’s death and changes in the size of the marriageable cohort. Formally, the number of attendees at royal parties in a given year, At, is treated as an endogenous variable and models as

0 0 (4.2) At = Ztρ + Vtη + d1851t + νt , where Zt is the vector of instruments. It includes an indicator for the interruption of the Season (1861–1863) and the size of the marriageable cohort at t. To measure it, I compute the number of peers’ daughters between ages 18 and 24 each year 34 from the Hollingsworth dataset. The vector Vt includes alternative predictors for attendance such as the sex ratio or the length of the railway network, which proxies for commuting costs. It also includes a time trend and decade fixed effects to account for the time effects described in FigureIII. Finally, I include an indicator for 1851, when the Crystal Palace Exhibition brought in lots of people to London. The magnitude of the effect of the Season on sorting along discrete characteristics

33It could be argued, for example, that whenever marriage outcomes got worse from the perspective of the nobility, more parties were organized in order to bring back social sorting. In addition the relation between the Season and intermarriage could be driven by underlying economic factors such as land prices. If economic conditions undermined the prosperity of the nobility and the royalty, they might have needed to marry wealthy commoners to alleviate debts. 34Eighteen was the earliest age at which a girl was presented at court. After 24, the hazard rate for women decreased sharply. See Online Appendix A.2 for details.

26 (e.g., social class) is captured by coeﬃcient β in

ˆ ˆ 0 0 (4.3) P r yi,t = yj,t|At, Vt, Xi,t = Φ β At + Vtλ + Xi,tδ ,

where y is a discrete characteristic of wife i and husband j and φ is the cumulative distribution function of the standard normal distribution. For continuous characteristics (e.g., wealth), the eﬀect is captured by coeﬃcient β in

ˆ 0 0 (4.4) |Yi,t − Yj,t| = β At + Vtλ + Xi,tδ + i,t ,

where Y is a continuos characteristic of wife i and husband j. In both models, Vi,t is the aforementioned vector of time varying controls. Xi,t is a vector of individual controls, including diﬀerent measures of socio-economic status, age at marriage, birth order (for women only), and the number of similar potential partners in the market.

Note that I am using a triangular IV model in which both the treatment (At) and the instruments (Zt) only vary at the year level whereas marriage outcomes (yi,t and

Yi,t) are measured at the individual level. To fit this model, I estimate the recursive equation systems (4.2) and (4.3) and (4.2) and (4.4) by maximum likelihood.35 As with every IV approach, the identifying assumption is that the instruments are relevant and that the exclusion restriction is satisfied. In the previous section I argued the the interruption of the Season after Prince Albert’s death satisfies these assumptions. The size of the cohort is also a relevant instrument: whenever a boom cohort entered the marriage market, more girls were presented at court, more events were organized, and the number of attendees to royal parties increased. The exclusion restriction would be violated if the size of the cohort was endogenous to marriage outcomes or if it affected marriage outcomes outside the London Season. This scenario is unlikely for four reasons: first, no one plans how many children to have based on projections of marriage market conditions 20 years in the future. Second, the size of the cohort does not vary much locally. Only when aggregated nationwide is the variation in cohort size meaningful.36 Therefore, marriage behavior would not be affected unless the British marriage market was centralized. Third, Gautier et al. (2010) and Botticini and Siow(2011) show that local, decentralized marriage markets such as the ones set up in the countryside were not subject to increasing returns to

35I use the STATA user-written command cmp and cluster errors by year (Roodman 2015). 36In 1851–75, the standard deviation of the size of the cohort is 14.8 nationwide and 0.1 per county.

27 scale. In other words, in these alternative markets, changes in the size of the cohort should not affect marriage behavior.37 Finally, as I have two sources of exogenous variation in attendance to the Season, I can use the Sargan test. Across specifications, I cannot reject that the cohort size instrument is exogenous. Panel B of TableIV presents the first-stage results. In column [2] I only include one instrument: the interruption of the Season. This shock accounts for much of the variation in attendance at royal parties between 1851 and 1875. The F-test is large enough to eliminate concerns about weak instruments. In column [3], I also include variation in the size of the cohort, which strengthens the first-stage. I find a positive, significant relation between the size of the cohort and attendance at royal parties. A single additional woman of marriageable age attracts 67 people to royal parties. Panel A presents the probit and IV-probit estimates for the effect of the Season on the rate of intermarriage with commoners. I find that the Season was a key determinant of sorting in this dimension. In particular, using exogenous variation from the interruption of the Season (column [2]), I find that increasing the number of attendees by 4 percent — 200 additional attendees —38 would decrease the probability of the average peer daughter marrying a commoner by one percentage point. Results are very similar when I include exogenous variation in attendance from changes in the size of the cohort (column [3]). For peers’ sons, instead, the effect is slightly lower and not significant in the IV specification. Although in Panel A I do not report the coefficients from control variables, they all have expected the signs. Consistent with the evidence from TableI, higher ranked individuals were less likely to marry commoners. The relative size of the class did not play any role, indicating that marriages were not randomly set. Older girls were less selective: for the average peer daughter, growing a year older increased the chances of marrying a commoner by 2 percent, reflecting the social pressure to get engaged shortly after coming out (Davidoff 1973: 52). Birth order does not have a significant effect. The children of families in the Scottish or Irish peerage were more likely to marry commoners. Finally, imbalances in the sex ratio and the length of the railway network do not seem to play a relevant role in this context.

37Cohort size variation may aﬀect the sex ratio. Given that men tend to marry younger spouses, if the population is growing, the relative number of men in the marriage market decreases, producing a marriage squeeze (Bhrolchain 2001). To account for that, I include sex ratios as a control. 38Given that the average number of attendees to royal parties was 4,641.2, 200 guests are 4 percent.

28 Table IV

The Season and sorting by social position (1851–75)

[1] [2] [3] [4] [5] [6]

Panel A Dep. Variable: Spouse is a commoner Women Men probit IVpobit IVpobit probit IVpobit IVpobit

Attendees Season -0.003*** -0.005*** -0.004*** -0.002* -0.002 -0.002 (0.001) (0.001) (0.001) (0.001) (0.002) (0.002) Observations 775 781 781 979 981 981 % correctly pred. 69.2 69.3 69.4 75.4 75.3 75.4 Sargan test 1.47 1.40 (p=0.5) (p=0.5) Individual controls YES YES YES YES YES YES Marriage mrkt co. YES YES YES YES YES YES Decade FE YES YES YES YES YES YES Trend YES YES YES YES YES YES

Panel B: First-stage Dep. Variable: Attendees at the Season (100’s)

Interruption (1861–63) – -27.5** -31.2*** – -27.5** -31.2*** – (10.1) (9.1) – (10.1) (9.1) Marriage cohort size – 0.7** – 0.7** – (0.3) – (0.3) Observations – 25 25 – 25 25 F-test – 18.0 20.9 – 18.0 20.9 Controls – YES YES – YES YES Decade FE – YES YES – YES YES Trend – YES YES – YES YES

Note: The sample for Panel A is all peers and peers’ offspring first marrying in 1851–75. The variable capturing the effect of the Season is the number of attendees at royal parties (in 100s). Individual controls include indicators of social position at age 15 (i.e., Commoner, Barons’ offspring, Dukes’ offspring, Barons’ heir and Dukes’ heir), age at marriage, and an indicator for English titles. For women I also include birth order excluding heirs. Marriage market controls include sex ratio, the relative size of class, and railway length. See Online Appendix A.2 for detailed descriptions. Cells report marginal effects at the mean. Standard errors are clustered by year. For Panel B, the sample is the years 1851–75 and regressions include an indicator for the 1851 Crystal Palace fair, sex ratio, and railway length as controls. Constants are not reported. Standard errors are in parentheses; *** p<0.01, ** p<0.05, * p<0.1.

Overall, the model correctly predicts the probability of marrying a commoner around 70 percent of the cases.39 The IV and probit marginal eﬀects are very similar, indicating that the endogeneity bias might be small. Finally, the Sargan test for overi- dentifying restrictions implies that I cannot reject the exogeneity of the instruments. The Season not only aﬀected the rate of intermarriage with commoners; it also helped to strengthen sorting by landed wealth. TableV considers peers and peers’ 39The remaining 30 percent might be explained by less observable factors, such as beauty or love.

29 sons who were listed as great landowners in Bateman 1883. Panel A, columns [1] to [3] look at the probability of marrying a spouse from the same wealth class. Classes are defined according to Bateman’s categorization: the first class are owners of 100,000 acres and upwards; then owners of 50,000 to 100,000 acres; 20,000 to 50,000 acres; 10,000 to 20,000 acres; 6,000 to 10,000 acres; and 2,000 to 6,000 acres (Bateman 1883: p. 495). Using variation from the interruption of the Season (column [2]) and from the interruption and changes in the size of the cohort (column [3]), I find strong significant causal effects: every 125 additional attendees at royal parties would increase by 1 percentage point the chances of a great lord marrying within the same wealth class. These effects are not the result of an arbitrary definition of classes. In columns [4] and [5], I consider a continuous measure of sorting: the absolute value of the difference between spouses’ acreage in percentiles. Every 100 additional attendees at royal parties would decrease by 0.2 this measure of mismatch in landed wealth. Panel B examines the effect of the Season on sorting by rents from land. Columns [1] to [3] report the coefficients from a regression of equation (4.3)’s form, where the dependent variable indicates whether spouses’ are in the same decile of the distribution of land rents, one decile above, or one decile below. Columns [4] and [5] consider the absolute value of the difference between spouses’ land rents in percentiles. The effects are similar in magnitude to the ones in Panel A: a 100 additional attendees increase by almost 1 percentage point the probability that both spouses’ earn similar rents from their family estates and decreases mismatch by 0.2 points.40 Results are robust to controlling for acreage and land rents, age at marriage, an indicator for English titles, railway length, relative size of land class, and the sex ratio. Interestingly, I find that the length of the railway is negatively associated with sorting by landholdings. Railways reflected the riches of industrialists. As the railway network expanded, their daughters became more attractive in the marriage market despite their lack of landholdings.

40Why is the effect of the Season on sorting by landholdings larger than the effect on the probability of marrying a commoner? One possible explanation are differences in market depth: while there were very few great landowners in the marriage market, the number of peers’ offspring, albeit small, was larger. When the Season did not facilitate encounters between singles, waiting for a great lord was unsuited. In contrast, even when the Season was disrupted, there was still hope to meet a peers’ offspring. This finding is in line with Banerjee et al.(2013). They find that, although individuals seem to be willing to disregard beauty and education to marry within their caste, they do not have to do so in equilibrium because the market is sufficiently deep, meaning there is a high probability of eventually encountering someone within your caste who is highly educated and/or handsome.

30 Table V

The Season and other sorting patterns (1851–75)

[1] [2] [3] [4] [5] [6]

Panel A. Sorting by acres Wife in same land class Mismatch probit IVprobit IVprobit OLS IV IV

Attendees Season 0.005*** 0.007*** 0.008*** -0.205 -0.266** -0.225** (0.002) (0.002) (0.002) (0.122) (0.116) (0.112) Observations 257 257 257 171 171 171 % correctly pred. 84.4 84.4 84.4 Sargan test 1.3 0.05 (p=0.3) (p=0.8)

Panel B. Sorting by land rents Wife in same land class Mismatch probit IVprobit IVprobit OLS IV IV

Attendees Season 0.008*** 0.008*** 0.009*** -0.209** -0.196** -0.167* (0.001) (0.002) (0.002) (0.101) (0.088) (0.086) Observations 257 257 257 171 171 171 % correctly pred. 77.0 77.4 77.4 Sargan test 1.4 0.4 (p=0.2) (p=0.5)

Panel C. Sorting by socio-economic status, constructed using PCA Mismatch OLS IV IV

Attendees Season - - - -0.621* -0.891* -0.712** - - - (0.329) (0.511) (0.351) Observations - - - 170 170 170 Sargan test - - 0.1 (p=0.7)

Individual controls YES YES YES YES YES YES Marriage market co. YES YES YES YES YES YES Decade FE and trend YES YES YES YES YES YES Interruption IV - YES YES - YES YES Cohort size IV - NO YES - NO YES F-stat ﬁrst-stage - 18.0 20.9 - 18.0 20.9

Note: The sample is all peers and peers’ sons first marrying in 1851–75 in possession of more than 2,000 acres, worth £2,000 a year by 1876. The variable capturing the effect of the Season is the number of attendees at royal parties (in 100s). Columns [1] to [3] report marginal effects at the mean for the probability that the wife’s family is in the same land class. In Panel A, classes are: >100,000 acres; 50,000 to 100,000; 20,000 to 50,000; 10,000 to 20,000; 6,000 to 10,000; and 2,000 to 6,000 (Bateman 1883: 495). In Panel B, classes are defined in annual land rents: in the same decile, one decile above, or one decile below. Columns [4] to [6] restrict the sample to those marrying into families listed in Bateman(1883). Panel A reports marginal effects for the absolute value of the difference in spouses’ acreage percentile, Panel B for the absolute value of the difference in spouses’ land rents percentile, and Panel C for the absolute value of the difference in a socio-economic ranking based on the principal component analysis. See the text for details. Individual controls include age at marriage and an indicator for English titles. Panels A and B also include acreage and land rent’s percentile. Panel C includes all the remainder variables used in the PCA. Marriage market controls include railway length at marriage year, the percentage of women aged 18–24 belonging to his land class (in terms of acres in Panel A, in terms of rents in Panels B and C), and sex ratio. See Online Appendix A.2 for detailed descriptions. IV and IV probit use first-stages reported in TableIV, Panel B. Standard errors clustered by year are in parentheses; *** p <0.01, ** p<0.05, * p<0.1

31 Both models assess sorting by landholdings accurately. Between 75 and 85 percent of the observations are correctly predicted. Again, probit and IV models produce similar results and the Sargan tests cannot reject the exogeneity of the instruments. To more precisely estimate the effects of the Season on sorting I define a pecking order, i.e, I order men and women according to an index of socio-economic status. The index is the first principal component from six variables that reflect socio-economic status: social position at age 15, acreage, land rents, ownership of woods, whether family held land since the time of Henry VIII, and the peerage of the title (England, Ireland, or Scotland). For women, I also considers her status as heir or coheir. The first principal component explains 26 percent of the sample variance for both men and women. My measure of sorting is the absolute value of the difference between husband’s and wife’s ranking, i.e., when the measure is 0, it indicates that the ith man with more socio-economic status is married to the ith woman with more socio- economic status. When the measure takes larger values, it indicates less sorting. TableV, Panel C examines the effect of the Season on this measure of sorting by socio-economic status. Whether I use exogenous variation in attendance from the interruption of the Season (column [5]), or the full set of instruments including changes in the size of the cohort, I find that attendance to the Season increases sorting. Increasing attendance in 800–1200 people would bring spouses 8 positions closer in the pecking order, which would imply a 10 percent increase in sorting.41 These results suggest that the Season, by reducing search costs and segmenting the marriage market, generated marital sorting along status and wealth. Did this come at the cost of unhappy marriages? Section B.2 in the Online Appendix examines the effect of the Season on the quality of the match, proxied by the degree of consanguinity between spouses and by their fertility. The evidence powerfully suggest that when the Season worked smoothly, it increased match quality along these dimensions. The pattern of geographic endogamy is also consistent with the centralization of the marriage market in London. In TableVI I show that a well-attended Season allowed aristocratic singles from further away to meet, to court and eventually to marry. For every 100 additional attendees, the probability of marrying a spouse in the same region decreases by almost 1 percentage point (column [3]) and the distance between spouses’ seats increases by 1.25 miles (column [6]).42

41The mean diﬀerence between spouses’ socio-economic rank is 81.3 positions. 42Results are not strong because the sample is smaller. Note also that once attendance is in-

32 Table VI

The Season and geographic endogamy (1851–75)

[1] [2] [3] [4] [5] [6]

Dep. Variable: Wife in same region Distance in miles probit IVprobit IVprobit OLS IV IV

Attendees Season -0.018 -0.004 -0.008* 1.1 1.2 1.7** (0.012) (0.005) (0.004) (0.7) (0.8) (0.8) Observations 169 169 169 169 169 169 % correctly pred. 78.7 78.1 78.1 . . Sargan test 1.11 . 2.28 (p=0.3) . (p=0.14) Individual controls YES YES YES YES YES YES Marriage mrkt co. YES YES YES YES YES YES Decade FE YES YES YES YES YES YES Year trend YES YES YES YES YES YES Interruption IV - YES YES - YES YES Cohort size IV - NO YES - NO YES F-stat ﬁrst-stage - 18.0 20.9 - 18.0 20.9

Note: The sample is all peers and peers’ sons first marrying in 1851–75. I exclude marriages if a spouse’s family seat is not listed in Burke(1826). The variable capturing the effect of the Season is the number of attendees at royal parties (in 100s). Columns [1] and [2] report marginal effects at the mean for the probability that the wife is from the same region. Regions are first- level NUTS for England and Wales, electoral regions for Scotland, and provinces for Ireland. In Columns [3] and [4], the dependent variable is distance between spouses’ seats. When spouses have more than one seat, I take the minimum distance. Individual controls include indicators of social position at age 15, age at marriage, and an indicator for English titles. Marriage market controls include sex ratio, railway length at marriage year, and the percentage of women aged ± 2 years whose family seat is in the same region. See Online Appendix A.2 for detailed descriptions. IV and IV probit use first-stages reported in TableIV, Panel B. Standard errors clustered by year are in parentheses; *** p<0.01, ** p<0.05, * p<0.1

One important question is whether the matching technology displays increasing returns to scale (see Botticini and Siow 2011 Gautier et al. 2010 and reference therein). Does the number of users of a matching technology increase the magnitude of its eﬀects on sorting? Section B.1 in the Online Appendix examines this issue using the case study of the London Season. Probit regressions of equation (4.3)’s form allow strumented, the magnitude of the coeﬃcient increases. The endogeneity bias may be important for geographic sorting.

33 me to calculate nonlinear marginal effects. Figure B.I suggests that the matching technology embedded in the Season displays increasing returns to scale: the larger the royal parties were, the greater the effect of bringing in additional guests on sorting.43 In the Online Appendix, I also test the robustness of these results. In Section C.2, I stratify my dataset by observables. For sorting by social status and landholdings, I find stronger and more tightly identified effects for individuals of higher socio- economic position. Section C.3 shows that results are robust to using alternative measures of the London Season such as the number of invitations issued instead of the number of attendees. Section C.4 uses an insight from Altonji et al.(2005) to gauge the potential effect of unobserved variables. Results suggest that, on average, the effect of unobservables would have to be approximately 5 times larger than the effect of the observed covariates to explain away the estimated effects. Section C.5 shows that the bias of the estimates would be small even if the cohort size instrument has some correlation with unobservables that are influencing marriage outcomes. Altogether, the evidence indicates that the Season, by reducing search costs and segmenting the marriage market, played a crucial role in determining who married whom. In normal years, following a “boom” cohort the Season was well-attended and the children of the nobility married others with similar social position, landholdings, and socio-economic status. Also, these spouses came from more distant places. In- stead, when attendance to the Season was disrupted after the death of Prince Albert, these sorting patterns were reversed. Next, I shed light on the implications that these sorting patterns had for social mobility and wealth inequality in the long-run.

5 Long-run economic implications

This section examines the implications of marital assortative matching for social mobility and wealth inequality. I also discuss the pernicious eﬀect of inequality on the provision of public schooling in late-Victorian England.

43For example, when royal parties were attended by less than 2,000 guests, the probability of marrying a spouse with similar landholdings increased by 0.25 percent for every additional 100 attendees. When the Season gathered more than 4,000 people, the same marginal eﬀect jumps to 0.5 percentage points, and when royal parties reach 7,000 attendees, it increases to 1 percent.

34 5.1 Social mobility

How would social mobility look like in the absence of the Season? To answer this question, I use the estimated coeﬃcients in TableIV, Panel A (column [3]) to predict the probability of marrying a commoner setting the number of attendees to zero for all years. Multiplying this predicted probability by the number of marriages, I get a counterfactual for the total number of commoners marrying peers’ daughters if the Season had not taken place between 1851 and 1875.44 FigureVIII compares observed and counterfactual marriage outcomes. Between 1851 and 1875, the rate of intermarriage between peers’ daughters and commoners would have been 30 percent higher without this institution. Given that the observed rate of intermarriage was already around 60 percent, it could be said that most of the marital segregation can be explained by the London Season. In other words, without the Season many newcomers would have married into the nobility; England would have looked much more like continental countries with large and not very rich aristocracies. 800 600 400 200 marriages outside the peerage (cumulative) 0 1850 1855 1860 1865 1870 1875 year

Season No Season (counterfactual)

Figure VIII: Peer-commoner intermarriage without the Season. I predict the counterfactual probability of marrying outside the peerage using the estimated coefficients from TableIV, Panel A (column [3]). I set the number of attendees to zero and the values of the remaining variables at their yearly means. This probability is then multiplied by the number of marriages per year, which I assume to be fixed. Dashed lines are 95% confidence intervals. The hollow diamond line displays the actual number of peers’ daughters marrying commoners.

44I assume the total number of marriages would not change in the absence of the Season. Given that the proportion of peers’ daughters never marrying was small, this is a reasonable assumption.

35 5.2 Wealth inequality

Does positive assortative matching in the marriage market lead to an unequal distribution of wealth? Cancian and Reed(1998), Schwartz(2010), Fernandez et al. (2005) and Greenwood et al.(2014) show both theoretically and empirically that sorting and income inequality reinforce one another. However, whether this crystal- lizes into wealth inequality remains an open question, as modern-day data can hardly speak to the long-run consequences of marital sorting. In this subsection, I shed light on this issue in the context of my historical setting. In a cross-section of English and Welsh counties, I document a strong and significant correlation between sorting and wealth inequality over the very long-run. I focus my attention on landed wealth. Specifically, I assign each noble family to the county where their principal estate was located. Then I compute the dynastic intermarriage rate: the percentage of members of a dynasty that first married a commoner, from the first dynasty member in the Hollingsworth dataset to the member listed in Bate- man(1883). Heirs are excluded from this calculation to avoid the endogeneity that may arise if they married strategically to consolidate their estates. This practice was common in the late seventeenth century (Mingay 1963). In total, I consider 10,364 individuals from 553 noble dynasties marrying in 1531–1880. FigureIX plots county- level averages of dynastic intermarriage against the Gini index for the distribution of land (Panel A) and the share of land owned by peers (Panel B). I find that in counties where noble dynasties intermarried less with commoners over time, land was more unequally distributed. Consider, for example, the county of Wiltshire. There were 15 noble families. On average, only 40 percent of the members of these families married a commoner since the sixteenth century. In the 1870s, the Gini index for the distribution of land was 0.93 and almost 30 percent of the county was in the possession of these 15 families. Overall, the correlation coefficient is -0.45 between dynastic intermarriage and the Gini index, and -0.33 between dynastic intermarriage and the share of land owned by peers. Given these numbers, a 30 percent increase in intermarriage — the estimated effect of suppressing the Season — would have reduced the Gini index for land inequality by 3 points and the share of land in the hands of peers by 50 percent. This evidence does not allow for causal inference. However, given the importance of marriage for the intergenerational transmission of wealth, the mechanism behind

36 this correlation seems obvious. By segregating the rich and powerful from the poor and insignificant, the Season prevented wealth from trickling down. To illustrate this point, consider a society divided into two groups. In the initial period, members of the first group possess all the wealth in the economy. Wealth is fixed and bequeathed from generation to generation. In this simple case, the only way in which society will become more equitable is if at some point individuals from the two classes intermarry. Any institution that prevents this from happening will perpetuate inequality.45

Panel A Panel B 1 .4

NBL CAEWAR DUR CAE GLA WIL RUT DOR NTT ERY HRTHAM LAN AGY WIL DBY CHS STS SSX .3 .95 GLSWRY YKSNTT KEN NTHSRY OXFHEFSAL NBL DUR BKMBDF CHS DEN WOR STSNTH DBY NFK FLN ESS BRKSFK SAL LEIDEVNRY MER CON SSX MON HRT MON DOR CMN WAR SOM ERY BKM OXFNRYGLA .9 .2 BDFLEI AGY gini HUN BRE HUN LIN YKSFLN WOR LIN WRYNFK MGYWES GLS SFK CAM LAN HAM DEV MGY KEN SRY SOM CAM BRK

MDX share of land owned by peers CON MDX CGN .1

.85 WES CMN CGN PEM ESS RAD BRE PEM DEN HEF MER

RAD 0 .8 30 40 50 60 70 80 90 100 30 40 50 60 70 80 90 100 % dynastic intermarriage (1531−1880) % dynastic intermarriage (1531−1880) Correlation = −0.45*** Correlation = −0.33**

Figure IX: Sorting over generations and inequality in landownership (1531–1880). This ﬁgure plots measures of landownership concentration against the rate of intermarriage of noble dynasties for all historic counties in England and Wales. See the text for descriptions. Solid lines show the ﬁtted values; *** p<0.01, ** p<0.05, * p<0.1

The aforementioned mechanism had important consequences for the distribution of land, especially in the eighteenth century. At that time, credit constraints on smaller landowners generated “a drift in property ... in favor of the large estate and the great lord” (Habakkuk 1940: 2, 4). These constraints might have been relaxed

45In this simple example, I assume that wealth can be accumulated but that there is no technology generating new wealth. If wealth can be generated, society may become more equitable (even under perfect segregation) if poorer individuals generate wealth at a higher rate. In Britain, the Industrial Revolution might have played this role.

37 if wealth had trickled down from the children of the nobility to commoners through marriage. Crucially, wealth could also trickle down through non-heirs. Although the costume was primogeniture, second sons and daughters were not completely excluded from inheritances. On the day of their marriage, heirs typically signed a marriage set- tlement, agreeing to provide an annual income from the family estate to their younger brothers and sisters (Habakkuk 1940). By marrying these non-heirs, commoners could access to the wealth of the peerage. One of the limitations of this exercise is that it only considers landed wealth. How important were other sources of wealth, especially after the Industrial Revolution? Online Appendix A.3 examines this in detail. Figure A.VI shows that, although the share of millionaires who earned their fortunes in manufacture, food, drink, or tobacco industries grows over time, it is not until the twentieth century that this source of wealth became more important than landownership. From 1800 to the 1870s, 80–95 percent of millionaires were still landowners (Rubinstein 1977: 102). This evidence is consistent with Clark et al.(2014). Tracking rare surnames over generations, they conclude that the Industrial Revolution did not accelerate social mobility. These results suggest that marital sorting can contribute to wealth inequality in the long-run. For England, the pattern was striking. Despite being the cradle of the Industrial Revolution, the top of the wealth distribution remained largely unchanged. The Season, and its implied sorting patterns, partly accounts for this immobility, helping to sustain the peerage’s role as an small, exclusive, and persistent elite.

5.3 Provision of public schooling

Was this huge wealth inequality harmful? Here I discuss the relation between land inequality and public schooling in England and Wales. Sokoloﬀ and Engerman(2000) and Galor et al.(2009) suggested that while emerging capitalists might be willing to support and subsidize education, entrenched landowners opposed educational reforms due to the lack of complementarity between human capital and agrarian work, and to reduce the mobility of the rural labor force. This is particularly suited to explaining why the provision of education in England lagged 50 years behind Prussia and the United States (Stephens 1998). Go˜ni(2015) examines these issues using evidence from School Boards. School Boards were introduced in 1870 following Forster’s Education Act. Each Board had plenty of powers to raise funds from rates (i.e., local taxes on

38 property) for education. In theory, thus, where landowners held the lion’s share of land, they had to eﬀectively pay for most of the education provision. TableVII suggests that, in practice, where these elites were particularly powerful they managed to take over School Boards. The table shows the results of regressions of investments in education in 1871–95 on landownership concentration. The correlation is negative: for the average county, a one percentage point increase in the share of land owned by peers is associated to 23.7 fewer funds raised from property taxes.

Table VII

Investments in education (1871–95)

[1] [2] [3] [4]

Dep. Variable: Funds raised from rates (pence p.c.) South and All Midlands North Wales

Share of land peer great landowners -23.68*** -24.55*** -190.29* -28.00*** (7.93) (8.41) (94.56) (8.28) “ “ comm. “ “ -21.97 -52.65** 157.31 25.21* (15.97) (22.51) (96.43) (13.92) log income (pence) -2.46 -3.08 -21.38 -12.87** (2.82) (2.44) (35.40) (5.33) % working on manufacturing 0.11* 0.22*** 0.66 0.03 (0.06) (0.06) (0.61) (0.08) city size 0.07*** 0.04*** -0.25 0.09 (0.02) (0.01) (0.26) (0.08) % voting conservative 0.07 0.10* -0.36 -0.19 (0.04) (0.05) (0.94) (0.22) % non-conformist 0.23*** 0.05 -0.02 0.05 (0.07) (0.16) (0.81) (0.17) % religious -0.11 -0.17 -0.19 0.30 (0.08) (0.11) (0.75) (0.20) Observations 104 66 14 24 Adjusted-R2 0.400 0.530 0.454 0.472 Note: The sample is all counties in England and Wales in 1871–95. All variables are decade averages. In columns [2] to [4], I split the sample by geographic area. South and Midlands are English counties south of Cheshire, and Yorkshire (West Riding and East Riding). North is Cheshire, Cumberland, Durham, Lancashire, Northumberland, Westmoreland, and Yorkshire. Funds raised from rates are from Go˜ni(2015). County controls are from Hechter(1976). Constants not reported. Standard errors clustered by county in parentheses; *** p<0.01, ** p<0.05, * p<0.1

39 Of course, differences in education provision could be explained by differences in income, culture, religiosity, or geographic conditions. However, this result is robust to the inclusion of various county-level controls: the share of land owned by commoner great landowners, income, the percentage of people working on manufacturing, city size, the percentage voting conservative, religiosity, or the percentage of non-conformists. In columns [1] to [4], I split the sample into South and Midlands, North, and Wales to account for geographical differences across counties. The negative correlation between land inequality and education investments is present in all areas, although it is particularly strong in the pastoral north, where family farms dominated (Clark and Gray 2014). Altogether, this suggests that elites in England were still sufficiently powerful in the late nineteenth century to oppose policies favorable to sustained economic growth, like public education provision (Acemoglu 2008). The Season, and it implied sorting patterns, were crucial in sustaining the English nobility’s role as an exclusive and persistent political elite.

6 Conclusion

In nineteenth century Britain, from Easter to August each year, the children of the nobility engaged in a whirl of social events. From presentations at court to royal parties, the objective was to pull together the right sort of suitors and to aid its introduction and courtship. In this article I show that by reducing search costs and segregating poor and “undesirable” suitors, this institution crucially contributed to marital sorting. To establish causality, I focus on three years when the Season was interrupted by the death of Prince Albert in 1861. Marriage behavior changed dramatically. The generation of ladies affected by the disruption were 80 percent more likely to marry a commoner and married spouses with smaller landholdings, reducing sorting by landed wealth by 50 percent. Moreover, geographical distance between spouses’ family seats shrunk, indicating that the partner selection problem shifted to the local marriage markets temporarily. To quantify the magnitude of these effects, I use an IV model with the interruption of the Season and changes in the size of the marriageable cohort as sources of identifying variation. I find that when the Season was well-attended, marital sorting strengthened. A 5 percent increase in attendance to the Season reduced by 1 percent the probability of marrying a commoner for the

40 average peer daughter. For great lords, it increased by 1.5–2 percent the probability of marrying endogamously in terms of landed wealth and annual rents from land. A classic insight from the assignment literature is that search frictions in the matching process and market segregation affect the degree of sorting (Burdett and Coles 1997, Eeckhout 1999, Bloch and Ryder 2000, Shimer and Smith 2000, Adachi 2003, Atakan 2006). This well-founded theoretical result, however, lacks strong empirical support (Fisman et al. 2008, Banerjee et al. 2013, Hitsch et al. 2010). My findings suggest that search costs and marriage market segmentation can generate sorting without resort to preferences, reconciling the empirical evidence with the theoretical search literature. Marital sorting also contributed significantly to the extreme concentration of wealth that characterized Victorian Britain. Using a counterfactual analysis, I find that between 1851 and 1875 the rate of intermarriage between peers’ daughters and commoners would have been 30 percent higher in the absence of the Season. Linking information on 10,364 marriages in 1531–1880 to cross-county estimates of the land distribution, I show that where noble dynasties intermarried less with commoners over the centuries, landed wealth became more unequally distributed. In line with my findings, Fernandez and Rogerson 2001 and Fernandez et al. 2005 establish a theoretical and empirical correlation between sorting and inequality. Greenwood et al. 2014 suggests that marital sorting has contributed significantly to the raising income inequality in the United States. The advantage of my historical case study is that it allows me to address whether this will crystalize into higher wealth inequality and to examine the lasting economic consequences of these ten- dencies. In detail, in this paper I show evidence suggestive that the Season, and its implied sorting patterns, contributed to wealth inequality, granting the English landed elite sufficient political and economic power to oppose policies that are favorable to sustained economic growth (Acemoglu 2008), such as public schooling provision. How much of the trends in marriage markets and inequality in Victorian Britain has any relevance for the world as it is now? The peerage does not look so different from individuals at the top 1 percent of the income distribution in the United States. They both marry at high rates, sort in the marriage market (Bakija et al. 2012), and earn immense annual incomes (Wolff 2012). Even their support for liberal policies is similar (see Table A.IV in the Online Appendix for details). Furthermore, Piketty and Saez(2006) show that over the last 50 years, income inequality has risen and that

41 this trend is mainly driven by only the top 1 percent of the population. In nineteenth century England, despite the advent of the Industrial Revolution, the lion’s share of land and wealth was also in the hands of the peerage (Cannadine 1990). In addition, today’s marriage market is not free of segregation. Laumann et al. (1994), Table 6.1, document that 60 percent of all married couples in the United States met in school, at work, at a private party, in church, or at a gym/ social club. All of these are, to some extent, segregative marriage markets where entry is restricted to similar others. Furthermore, the increasingly popular dating apps and websites have reduced search costs and have restricted choice sets for partners through search filters or customized recommendations by the platform. Many of these matchmaking services not only guarantee you will find love, but also that you will not waste time meeting people with whom you would never match.46 My findings suggest that if the very rich, this top 1 percent are somehow involved in someof these new forms of segregative matchmaking, the effects on the degree of marital sorting will be dramatic. Over the long-run, this may crystalize the rising income inequality into wealth inequality, with important implications for broader political and economic outcomes, including taxation and the provision of public goods. Piketty(2014) claims that we are living in a second Gilded age and that, unless some measures are taken, we will see a future dominated by a class of rentiers like the ones in the novels of Jane Austen. By analyzing the marriage decisions of these rentiers, this paper suggests that search costs and segregation in the marriage market are central to this debate.

References

Abramitzky, R., A. Delavande, and L. Vasconcelos (2011). “Marrying Up: The Role of Sex Ratio in Assortative Matching”. American Economic Journal: Applied Economics 3(3): 124–57. Acemoglu, D. (2008). “Oligarchic Versus Democratic Societies”. Journal of the European Economic Association 6(1): 1–44. Adachi, H. (2003). “A Search Model of Two-Sided Matching under Nontransferable Utility”. Journal of Economic Theory 113(2): 182–198. Aiello, D. (2010). “The London Season”. Last modiﬁed November 22, 2010. Accessed

46Gray and Farrar, for example, an exclusivist matchmaker operating in London for the last 23 years, only accepts “the most eligible singles.” The cheapest fee is of 15,000 pounds. As their motto says, “this service is not right for everybody” (grayandfarrar.com). Across the pond, Kelle- her International, a long-running, high-end matchmaking service, is targeting Silicon Valley with particular vigor (see D. Crane’s New York Times article on October 11, 2013).

42 May 9, 2011. http://www.thehistorybox.com/ny_city/society/printerfriendly/nycity_ society_london_season_article0008.htm. accessed May 9, 2011. Allen, D. W. (2009). “A Theory of the Pre-Modern British Aristocracy”. Explorations in Economic History 46(3): 299–313. Altonji, J. G., T. E. Elder, and C. R. Taber (2005). “Selection on Observed and Unobserved Variables: Assessing the Effectiveness of Catholic Schools”. Journal of Political Economy 113: pp. 151–184. Atakan, A. E. (2006). “Assortative Matching with Explicit Search Costs”. Econometrica 74(3): 667–680. Bakija, J., A. Cole, and B. T. Heim (2012). “Jobs and Income Growth of Top Earners and the Causes of Changing Income Inequality: Evidence from U.S. Tax Return Data’. Techni- cal report, Unpublished manuscript. Available at https://web.williams.edu/Economics/wp/ BakijaColeHeimJobsIncomeGrowthTopEarners.pdf. Banerjee, A., E. Duflo, M. Ghatak, and J. Lafortune (2013). “Marry for What? Caste and Mate Selection in Modern India”. American Economic Journal: Microeconomics 5(2): 33–72. Bateman, J. (1883). The Great Landowners of Great Britain and Ireland. London: Harrison and Sons. Becker, G. S. (1973). “A Theory of Marriage: Part I”. Journal of Political Economy 81(4): 813–46. Beckett, J. V. (1986). The Aristocracy in England 1660-1914. Oxford: Blackwell. Bhrolchain, M. N. (2001). “Flexibility in the Marriage Market”. Population: An English Selec- tion 13(2): 9–47. Bloch, F. and H. Ryder (2000). “Two-Sided Search, Marriages, and Matchmakers”. International Economic Review 41(1): 93–115. Botticini, M. and A. Siow (2011). “Are There Increasing Returns to Scale in Marriage Markets?”. Working Papers 395, IGIER (Innocenzo Gasparini Institute for Economic Research), Bocconi University. Burdett, K. and M. G. Coles (1997). “Marriage and Class”. Quarterly Journal of Economics 112(1): 141–168. Burke, J. (1826). A General and Heraldic Dictionary of the Peerage and Baronetage of the United Kingdom. London: Henry Colburn. Cancian, M. and D. Reed (1998). “Assessing the Effects of Wives’ Earnings on Family Income Inequality”. Review of Economics and Statistics 80(1): 73–79. Cannadine, D. (1990). The Decline and Fall of the British Aristocracy. New Haven: Yale University Press. Cannon, J. (1984). Aristocratic Century: The Peerage of Eighteenth-Century England. Cambridge: Cambridge University Press. Chen, W.-H., M. Forster, and A. Llena-Nozal (2013). “Demographic or Labour Market Trends”. OECD Journal: Economic Studies 1: 179–207. Clark, G. and N. Cummins (2015). “Intergenerational Wealth Mobility in England, 1858–2012: Surnames and Social Mobility”. Economic Journal 125(582): 61–85. Clark, G., N. Cummins, Y. Hao, D. D. Vidal, T. Ishii, Z. Landes, D. Marcin, F. Abu-Sneneh, W. Chow, K. M. Jung, A. M. Marek, and K. M. Williams (2014). The Son also Rises: Surnames and the History of Social Mobility. Princeton: Princeton University Press. Clark, G. and R. Gray (2014). “Geography is not destiny: geography, institutions and literacy in England, 1837–63”. Oxford Economic Papers. Collin, E. J. and J. M. McNamara (1990). “The Job Search Problem as an Employer-Candidate Game”. Journal of Applied Probability 28: 815–827. Davidoff, L. (1973). The Best Circles: Society, Etiquette, and the Season. London: Cromm Hel. Doepke, M. and F. Zilibotti (2008). “Occupational Choice and the Spirit of Capitalism”. Quarterly Journal of Economics 123(2): 747–793.

43 Eeckhout, J. (1999). “Bilateral Search and Vertical Heterogeneity”. International Economic Re- view 40(4): 869–887. Ellenberger, N. W. (1990). “The Transformation of London ‘Society’ at the End of Victoria’s Reign: Evidence from the Court Presentation Records”. Albion: A Quarterly Journal Concerned with British Studies 22(4): 633–653. Ellis, G. (1977). “Buckingham Palace”. In G. Ellis (Ed.), Royal Places of England. London: Hollings. Fernandez, R., N. Guner, and J. Knowles (2005). “Love and Money: A Theoretical and Empirical Analysis of Household Sorting and Inequality”. Quarterly Journal of Economics 120(1): 273–344. Fernandez, R. and R. Rogerson (2001). “Sorting and Long-Run Inequality”. Quarterly Journal of Economics 116(4): 1305–1341. Fisman, R., S. S. Iyengar, E. Kamenica, and I. Simonson (2008). “Racial Preferences in Dating”. Review of Economic Studies 75(1): 117–132. Gale, D. and L. S. Shapley (1962). “College Admissions and the Stability of Marriage”. American Mathematical Monthly 69(1): 9–15. Galor, O., O. Moav, and D. Vollrath (2009). “Inequality in Landownership, the Emergence of Human-Capital Promoting Institutions, and the Great Divergence”. Review of Economic Stud- ies 76(1): 143–179. Galor, O. and J. Zeira (1993). “Income Distribution and Macroeconomics”. Review of Economic Studies 60(1): 35–52. Gautier, P. A., M. Svarer, and C. N. Teulings (2010). “Marriage and the City: Search Frictions and Sorting of Singles”. Journal of Urban Economics 67(2): 206 – 218. Goñi,M. (2015). “Landed Elites and Education Provision in England and Wales. Evidence from School Boards, 1870–99”. Manuscript, University of Vienna. Greenwood, J., N. Guner, G. Kocharkov, and C. Santos (2014). “Marry Your Like: Assortative Mating and Income Inequality”. American Economic Review 104(5): 348–53. Habakkuk, H. J. (1940). “English Landownership, 1680–1740”. Economic History Review 10(1): 2–17. Hechter, M. (1976). “U.K. County Data, 1851–1966”. Data set SN: 430, Colchester, Essex: UK Data Archive. Hitsch, G. J., A. Hortacsu, and D. Ariely (2010). “Matching and Sorting in Online Dating”. Amer- ican Economic Review 100(1): 130–63. Hobhouse, H. (1983). Prince Albert: His Life and Work. London: Hamish Hamilton. Hollingsworth, T. H. (1964). “The Demography of the British Peerage”. Supplement to Population Studies. Hollingsworth, T. H. (2001). “Hollingsworth Genealogical Data on the British Peerage”. Database. Re-digitised from handwritten original index sheets and archived by the Cambridge Group for the History of Population and Social Structure. Inwood, S. (1998). A History of London. London: Papermac. Jacquet, N. L. and S. Tan (2007). “On the Segmentation of Markets”. Journal of Political Econ- omy 115(4): 639–664. Kremer, M. (1997). “How Much Does Sorting Increase Inequality?”. Quarterly Journal of Eco- nomics 112(1): 115–39. Laumann, E. O., J. H. Gagnon, R. T. Michael, and S. Michaels (1994). The Social Organization of Sexuality: Sexual Practices in the United States. Chicago: University of Chicago Press. Lee, S. (forthcoming). “Effect of Online Dating on Assortative Matching: Evidence from South Korea”. Journal of Applied Econometrics. Long, J. and J. Ferrie (2013). “Intergenerational Occupational Mobility in Great Britain and the United States since 1850”. American Economic Review 103(4): 1109–37. Malheiro, B. (1999). “1876 Victorian England Revisited”. Last modified 1999. Accessed May 15, 2011. http://logicmgmt.com/1876/splash.htm.

44 Miles, A. (1993). “How Open Was Nineteenth-Century British Society? Social Mobility and Equality of Opportunity, 1839–1914”. In A. Miles and D. Vincent (Eds.), Building European Society: Oc- cupational Change and Social Mobility in Europe, 1840–1940. Manchester: Manchester University Press. Mingay, G. E. (1963). English Landed Society in the Eighteenth Century. London: Routledge & Kegan Paul. Mitch, D. F. (1993). “Inequalities Which Every One May Remove: Occupational Recruitment, En- dogamy, and the Homogeneity of Social Origins in Victorian England”. In A. Miles and D. Vincent (Eds.), Building European Society: Occupational Change and Social Mobility in Europe, 1840– 1940. Manchester: Manchester University Press. Nevill, D. (1920). The Reminiscences of Lady Dorothy Nevill. London: Thomas Nelson and Son. Nielsen, H. S. and M. Svarer (2009). “Educational Homogamy: How Much Is Opportunities?”. Journal of Human Resources 44(4): 1066–1086. Persson, T. and G. Tabellini (1994). “Is Inequality Harmful for Growth?”. American Economic Review 84(3): 600–621. Piketty, T. (2014). Capital in the Twenty-First Century. Harvard: Harvard University Press. Piketty, T. and E. Saez (2006). “The Evolution of Top Incomes: A Historical and International Perspective”. American Economic Review 96(2): 200–205. Porter, D. (1998). The Thames Embankment: Environment, Technology, and Society in Victorian London. Akron: University of Akron Press. Roodman, D. (2015). Cmp: Stata module to implement conditional (recursive) mixed process estimator. http://EconPapers.repec.org/RePEc:boc:bocode:s456882. Rubinstein, W. D. (1977). “Wealth, Elites and the Class Structure of Modern Britain”. Past and Present (76): 99–126. Schwartz, C. R. (2010). “Earnings Inequality and the Changing Association between Spouses Earn- ings”. American Journal of Sociology 115(5): 1524–1557. Sheppard, F. H. W. (1971). London, 1808–1870: The Infernal Wen. London: Secker and Warburg. Sheppard, F. H. W. (1977). “The Social Character of the Estate: The London Season in 1841”. Survey of London: volume 39: The Grosvenor Estate in Mayfair, Part 1 (General History): 89–93. http://www.british--history.ac.uk/report.aspx?compid=41842. Shimer, R. and L. Smith (2000). “Assortative Matching and Search”. Econometrica 68(2): 343–369. Sokoloﬀ, K. L. and S. L. Engerman (2000). “Institutions, Factor Endowments, and Paths of Devel- opment in the New World”. Journal of Economic Perspectives 14(3): 217–232. Spring, E. and D. Spring (1985). “The English Landed Elite, 1540–1879: A Review”. Albion: A Quarterly Journal Concerned with British Studies 17(2): 149–166. Stephens, W. (1998). Education in Britain, 1750–1914 (Social History in Perspective). New York: St. Martin’s Press. Stone, L. and J. F. Stone (1984). An Open Elite? England 1540–1880. Oxford: Clarendon Press. Turner, E. (1954). A History of Courting. London: Ian Henry. Wasson, E. A. (1998). “The Penetration of New Wealth into the English Governing Class from the Middle Ages to the First World War”. Economic History Review 51(1): 25–48. Wilkins, K. (2010). “Elite Networks of the London Season: Perspectives from the New Mobilities Literature”. Technical report, Durham theses, Durham University. Available at Durham http: //etheses.dur.ac.uk/689/. Wolﬀ, E. N. (2012). “The Asset Price Meltdown and the Wealth of the Middle Class”. Working Paper 18559, National Bureau of Economic Research.

45 Online Appendix

A Data description

This appendix describes in detail the original data sources used in this paper. Then it presents descriptive statistics for the main variables used in the empirical analysis and it details how I constructed the marriage market controls. Finally, I present additional descriptive statistics showing the high pressure to marry young in Victorian Britain, the importance of land as a source of wealth after the Industrial Revolution, and a comparison between the peerage and the top one percent in the United States today.

A1 Original sources for the Peerage

Peerage records have chronicled the family histories of the peerage of Britain and Ireland since Arthur Collins published his Peerage in 1710. Many follow-up genealogical studies have updated his work. Among them, three peerage records stand out: Burke’s Peerage and Baronetage, Debrett’s The Peerage of the United Kingdom and Ireland, and Cokayne’s Complete Peerage. The genealogist John Burke wrote Landed Gentry, a similar record for knights and baronets. This last piece tends to be quite mythological, the result of centuries of word-of-mouth information. Oscar Wilde once said, “It is the best thing the English have done in fiction” (Burke et al. 2005). For the sake of illustration, Figure A.I shows the entry for Charles George Lyt- telton from Cokayne’s Complete Peerage. He was born in 1842 and that he married when he was almost 36. He held the titles of Baron of Frankeley and Baron of West- coote of Ballymore. He was also Viscount Cobham and Baron Cobham, but only received this honor at age 46 on the death of a distant cousin. Importantly, the entry provides similar information for his wife, Mary Susan Caroline Cavendish. She was 11 years younger and she was the daughter of the 2nd Baron Chesham. Hollingsworth(1964) collected this genealogical material for his study of the British peerage. He tracked all peers who died between 1603 and 1938 (primary universe) and their offspring (secondary universe). The primary universe was defined from Cokayne’s Complete Peerage. The universe of children was found from a va- riety of sources: Collins’ Peerage of England, Lodge’s Peerage of Ireland, Douglas’ Scots Peerage, Burke’s Extinct Peerage and modern peerage editions by Burke and

46 Figure A.i: Charles Lyttelton, Cockayne’s Complete Peerage

Debrett. The remaining gaps were filled from a large list of sources, among which Burke’s Landed Gentry stands out. See Hollingsworth(1964) for details. The dataset is now owned by the Cambridge Group for the History of Population and Social Structure. They re-digitised the 30,000 handwritten original index sheets in 2001. In its current form, the data comprises approximately 26,000 individuals. I complement the Hollingsworth(2001) dataset with information on family seats. Family seats are recorded in heraldic dictionaries. These dictionaries are summarized peerage records that contain additional information at a family level: religious affilia- tion, motto, coat of arms, and family seats. The most relevant source for my purposes is Burke’s Heraldic Dictionary (1826). Most of the young aristocrats who married between 1851 and 1875 were recorded as presumptive heirs in this source. Therefore, the family seats in Burke(1826) correspond in general to the seats where the individuals under analysis grew up and lived most of the year. Moreover, country seats were expensive to build and representative of long lasting lineages. They generally remained in the hands of the same family generation-after-generation until the 1870s, when the aristocracy started its decline. Therefore, the seats in Burke(1826) can also be used for individuals born before or after this source was published. Figure A.II shows the entry in Burke(1826) for the Baron Lyttelton. The Lyttelton’s family seat was Hagley Park, an eighteenth century house in Hagley, Worcestershire I also compliment the Hollingsworth(2001) dataset with information on landholdings from Bateman(1883). The book consists of a list of all owners of 3,000 acres and upwards by 1876, worth £3,000 a year. Also, 1,300 owners of 2,000 acres and upwards are included. It is based in the Return of Owners of Land, 1873; the first complete picture of the distribution of landed property in the British Isles since William the Conqueror commissioned the Domesday Book in 1086. The objective of the Return was to counter the rising public clamor about the over-concentration of land ownership within the elite. Although the House of the Lords stressed the importance of

47 Figure A.ii: Lord Lyttelton, Burke’s Heraldic Dictionary (1826) using reliable and independent data to refute the attacks, the Return was not without inaccuracies (Bateman 1883: preface). Bateman revised and corrected these errors. Several great landowners, however, felt disparaged. They wrote letters with outrage and demands for the immediate correction of the acres and rents assessed to them. Lord Overstone, for example, complained that “this list is so fearfully incorrect that it is impossible to correct it” (Bateman 1883: 348). Bateman’s estimates, therefore, should be taken as a lower bound for the wealth of great landowners in late nineteenth century Britain and Ireland. For the sake of illustration, Figure A.III shows the entry for Charles George Lyttel- ton from Bateman(1883). He was in possession of 6,939 acres scattered throughout his estates in Worcestershire and Herefordshire. His gross annual rents from land exceeded £10,200 a year. Charles George attended Eton and Trinity College, Cam- bridge. He was member of Brook’s, a gentleman’s liberal club. He sat in Parliament for Worcestershire. The letter S denotes that he is the head of, or head of a junior branch of, a family who held land in England since the time of Henry VIII. He

48 succeeded to the Lyttelton barony in 1876.

Figure A.iii: Charles Lyttelton, Bateman’s Great Landowners (1883)

A2 Data overview

Table A.I shows the summary statistics for the three samples I use in the empirical analysis: peers’ daughters marrying in 1859–67, peers’ daughters marrying in 1851– 75, and peers and peers’ sons marrying in 1851–75. PanelA describes the variables in the Hollingsworth(2001) dataset. Across samples, the proportion of dukes’ and barons’ oﬀspring is similar: half of the women marrying in 1859–67 and 1851–75 are dukes’ daughters, the other half are barons’ daughters; 11 percent of men marrying in 1851–75 are dukes’ heirs, 12 percent are barons’ heirs; 29 percent are dukes’ younger sons, 29 percent are barons’ younger sons. Men who were commoners at age 15 but ended their lives holding a peerage — either by creation or by inheriting the title from a distant relative — comprise 18 percent of the sample. The proportion of individuals from families in the English peerage is also stable across samples. To measure sorting by social position, I use information on the social status of spouses. The proportion of individuals marrying a commoner, that is, marrying outside the peerage, was 66 percent for women and 75 percent for men. This is not a large number given that around 1900, only 1 in 3,200 people in Britain was an aristocrat (Beckett 1986: 35-40).

49 Table A.i Summary statistics

marrying in 1851–75 marrying in 1859–67 women men women mean se mean se mean se A. Hollingsworth demographic data Commoner at 15∗ - - 0.18 0.01 - - Baron oﬀspring (not heir) 0.5 0.02 0.29 0.01 0.49 0.03 Duke oﬀspring (not heir) 0.5 0.02 0.29 0.01 0.51 0.03 Baronet - - 0.01 0 - - Baron heir - - 0.11 0.01 - - Duke heir - - 0.12 0.01 - - Heir - - 0.23 0.01 - - Peerage of England 0.6 0.02 0.63 0.02 0.61 0.03 Peerage of Scotland 0.12 0.01 0.12 0.01 0.12 0.02 Peerage of Ireland 0.28 0.02 0.25 0.01 0.27 0.03 Age at marriage 24.99 0.23 29.74 0.24 24.47 0.35 Life expectancy 69.01 0.67 67.06 0.49 67.99 1.16 Birth order (excluding heirs) 3.81 0.10 - - 3.85 0.17 Births 3.74 0.12 - - 3.95 0.19 Births per fertile marriage year† 0.28 0.01 - - 0.29 0.01 Births relative to Hutterites† 0.58 0.02 - - 0.6 0.03 Prop. marrying a commoner 0.66 0.02 0.75 0.01 0.66 0.03 Observations 781 981 273 B. Family seats from Burke(1826) Prop. marrying in same region‡ 0.21 0.03 idem idem 0.21 0.05 Distance btw spouses’ seats‡ 152.2 9.2 idem idem 152.3 16.8 Observations 170 idem idem 61 C. Bateman’s Great Landowners (1883) Acres > 100,000 (class 1) 0.05 0.01 0.06 0.02 Acres in 50,000–100,000 (class 2) 0.07 0.02 0.09 0.02 Acres in 20,000–50,000 (class 3) 0.27 0.03 0.25 0.03 Acres in 10,000–20,000 (class 4) 0.27 0.03 0.23 0.03 Acres in 6,000–10,000 (class 5) 0.16 0.02 0.13 0.02 Acres in 2,000–6,000 (class 6) 0.19 0.02 0.23 0.03 Acreage 26,495 2,500 28,447 2,497.1 Land rents 23,517 1,465 24,864 1,721 Ownership of woods 0.14 0.02 0.09 0.02 Family held land when Henry VII 0.24 0.03 0.26 0.03 Member of a liberal cluba 0.21 0.03 - - Member of a tory cluba 0.56 0.03 - - Prop. marrying in class (acres) 0.17 0.02 0.09 0.02 Prop. marrying in class (rents)b 0.23 0.03 0.06 0.02 Observations 257 236 Mismatch in acresc 27.64 1.72 32.67 1.92 Mismatch in land rentsc 27.18 1.48 32.5 1.97 Observations 171 130

∗ Received peerage later, either by creation or by inheriting a distant relative’s title. † See for detailed descriptions. ‡ Restricted to individuals whose spouses’ family is listed in Burke(1826). a Brook’s, Reform, and Devonshire are liberal Clubs; Carlton, Junior Carlton, Conservative, or St. Stephen’s are tories (Bateman 1883: 497). b Proportion marrying in the same decile, one decile above, or one decile below of the land rents distribution. c Mismatch is |p(yi)−p(yj )|, where y are the acres or land rents of the peer or peers’ oﬀspring i and the spouse j; and p stands for percentile. Only considers individuals whose spouses’ family is listed in Bateman(1883).

50 Men marry significantly later than women. The average age at marriage for peers’ daughters marrying in 1859–67 and 1851–67 is around 25 years old. Men typically marry around 30. Independently of the sample and the gender, individuals lived between 67 and 69 years and had around 4 children. The number of births per fertile marriage year is obtained by dividing the number of children by the total number of years a women was married between ages 15 and 40. On average, women produced 0.3 children per fertile marriage year. Finally, to take into account the biological maximum fertility of women, I also calculate the number of births relative to what a Hutterite women would have produced in the same marriage fertile lifetime. Hutterites are a traditionalist Christian church fellowship that reject birth control, and thus can be taken as a proxy for unconstraint fertility (Clark 2007 and Voigtlander and Voth 2013).47 According to this measure, the peerage reached 60 percent of their biological maximum fertility. As described in Section 3.2 in the paper, I complemented the Hollingsworth(2001) dataset with information on 694 country seats for 498 families linked to the peerage from Burke(1826). I then geo-coded each of these seats using GeoHack. When a set was not found, I used the coordinates from the nearest town, which sometimes was listed in Burke(1826). Figure A.IV shows the geographical distribution of family seats. The nobility was well dispersed all over the British Isles and seats were quite isolated from each other. The majority of seats are located in England, especially in the South East region. Scotland is less populated, although the density rises around the frontier with England. Finally, in Ireland most of the seats are concentrated in the eastern shore. Panel B in Table A.I summarizes the two variables that I constructed to measure geographic endogamy. The first is the probability of marrying in the same region. Regions are first-level NUTS for England and Wales, electoral regions for Scotland, and provinces for Ireland. Only 21 percent of the individuals married within their region. The second measure is distance between spouses’ seats. Again, the estimates are remarkably similar across samples. Women marrying in 1859–67 and women marrying in 1851–75 were both, on average, raised in a country house 152 miles away from their husband’s seat. Note that, by construction, only matrimonies formed by

47Speciﬁcally, Hutterites marital fertility rates are 0.55 for ages 20–24, 0.502 for 25–29, 0.447 for 30–34, 0.406 for 35–39, and 0.222 for 40–44 (Clark 2007). Using these numbers I calculate how many children a Hutterite couple would have produced in the same marital fertile lifetime as each peer daughter. Then I divide the actual number of children that each peer daughter had by this number.

51 My Map 18/10/12 18:34

My Map

Figure A.iv: Country seats. This figure shows the locationCopyright: of 694©2012 country Esri, DeLorme, seats fromNAVTEQ 498 families holding a peerage listed in Burke’s Heraldic Dictionary. peers’ offspring are included. Therefore, in this case, the variables take the same values for men and women. Note also that these estimates are lower bounds, as when one or both spouses have more than one seat I take the minimum distance. Finally, Panel C in Table A.I summarizes all the relevant information that I computerized from Bateman(1883). Note that, in this case, I only considered peers and peers’ sons marrying in 1851–75 who were listed as great landowners; and peers’ daughters marrying in 1850–67 with a close relative listed as a great landowner. Again, variables are stable across samples: only 5–6 percent of individuals are in the first land class (i.e., in possession of more than 100,000 acres), 19–23 percent belong to the lesser land class (i.e., in possession of 2,000–6,000 acres). The average land- holding is around 25,000 acres and yields £23,000–24,000 a year. The proportion of owners of woods and members of families in possession of land at the time of Henry VII is also similar. In contrast, 56 percent of noblemen were members of a tory club, while only 21 percent of them belonged to a liberal club.48

48Brook’s, Reform, and Devonshire were liberal clubs, and Carlton, Jr. Carlton, Conservative,

http://gpefm.maps.arcgis.com/home/webmap/print.html Page 1 of 1 To measure sorting by landholdings I computerized all the relevant information for the spouses of these individuals. In detail, I included the landholdings of any close relative listed in Bateman(1883). Using this procedure, I matched 353 wives for the sample marrying in 1851–75 (42.8 percent) and 130 husbands for the sample marrying in 1859–67 (64.4 percent). Table A.II, Panel A reports that most of the matched wives are sisters (43.6 percent) and daughters (28.3 percent) of great landowners. Acreage and land rents are similar across family relations. The exception are wives who aunts of great landowners, as their acreage seems to be significantly smaller but yields larger land rents. Panel B reports similar statistics for the matched husbands of peers’ daughters marrying in 1859–67. Most of them were landowners themselves (60.8 percent), followed by fathers (14.6 percent) and brothers (13.1 percent) of landowners listed in Bateman(1883). Differences across family relations are more pronounced than in the case of matched wives: the acreage and land rents attached to landowners is lower, while fathers of great landowners on average came from richer families. Using the information on the landholdings of peers and peers’ offspring (Table A.I) and their spouses (Table A.II), I compute two different measures of sorting: the first is the probability of marrying in the same land class. For acres, I use the land classes described in Table A.I. For land rents, I use deciles (i.e., the probability of marrying in the same or a contiguous decile of the distribution of land rents). Both for acres and land rents, if an individual marries a spouse whose family was not listed in Bateman(1883), this marriage is considered exogamous. In 1851–75, 17 percent of peers’ sons listed as great landowners married endogamously with respect to acreage, and 23 percent with respect to land rents. This is a considerable number, considering that land classes are on average very small. The second measure of sorting by landholdings is continuous: the absolute value of the difference between husband and wife’s landholdings in percentiles. I label this measure as mismatch. The advantage of this measure is that it does not hinge on an arbitrary definition of land classes. The downside is that it only considers individuals whose spouse family is also listed in Bateman(1883). 49 On average, mismatch was around 27 percentile points for men marrying in 1851–75 and around 32 percentile points for women marrying in 1859–67. and St. Stephen’s were tory clubs (Bateman 1883: 497). 49One can assign landless spouses to a specific category but cannot assume their landholdings to be 0, as Bateman only listed those in possession of 2,000 acres and upwards.

53 Table A.ii

Relation to landowner of matched wives

Panel A: Matched wives (1851–75) Acreage (1,000’s) Land rents (1,000’s) N % mean s.e. mean s.e. Sister 154 43.6 62 17.3 28 2.4 Daughter 100 28.3 42 14.2 25.3 2.9 Aunt 35 9.9 28.8 3.4 40.2 5.9 Second cousin 22 6.2 22.4 8.3 21.5 4.4 Cousin 17 4.8 25.7 4.3 26.4 4.1 Niece 13 3.7 24.9 4.7 15.9 2.4 Granddaughter 7 2 30.7 10.9 23.4 6.9 Other 5 1.4 65 23 61.6 27.2 Total 353 100 47 8.6 27.9 1.6 Percentage matched : 42.8 %

Panel B: Matched husbands (1859–67) Acreage (1,000’s) Land rents (1,000’s) N % mean s.e. mean s.e. Himself 79 60.8 17.5 2.4 18.4 2.4 Father 19 14.6 61.4 24.7 35.6 24.7 Brother 17 13.1 26.4 6.1 35.1 6.1 Son 8 6.2 13 4.7 13.2 4.7 Nephew 5 3.9 36 14.7 36.4 14.7 Other 2 1.5 38 15.1 33 15.1 Total 130 100 25.9 4.2 23.7 4.2 Percentage matched : 64.4 % Note: The sample for Panel A is the matched wives of peers and peers’ sons marrying in 1851–75 who are listed as great landowners in Bateman(1883). The sample for Panel B is the matched husbands of peers’ daughters marrying in 1859–67 whose family head was listed as great landowner in Bateman(1883). Standard deviations are in parentheses; *** p<0.01, ** p<0.05, * p<0.1

Table A.III shows the summary statistics for the variables used to evaluate whether sorting contributes to wealth inequality in Section5 in the paper. PanelA presents the demographic characteristics of the sample used to characterize sorting in the long-run. It comprises 10,364 individuals spanning about 11 generations. The ﬁrst individual in the sample is Thomas Butler, 10th Earl of Ormond, born in 1531; the last is Charles Rawdon-Hastings, 11th Earl of Loudoun, married in 1880. These individuals belong to 553 diﬀerent dynasties. On average, these dynasties span 259 years and include 28 individuals; 23 after excluding heirs. In my analysis I exclude heirs to avoid the

54 endogeneity that may arise if they married strategically to consolidate their estates. This practice was common in the late seventeenth century (Mingay 1963). My main measure of sorting in the long-run is the proportion of non-heirs in this sample who married outside the peerage, i.e., who married a commoner husband or a commoner’s daughter. On average, 60 percent married out, a relatively small percentage taking into account the small relative size of the British nobility.

Table A.iii

Summary statistics for long-run analysis

mean sd min max

A. Dynastic intermarriage in 1531–1880 (Hollingsworth 1964) Birth year of ﬁrst dynasty member 1641 0.59 1531 1860 Marriage year of last dynasty member 1857 0.7 1814 1880 Duration of the dynasty∗ 259 0.6 45 388 Number of dynasty members 28 0.11 1 62 Proportion marrying commoners 0.6 0 0 1 Num. of dynasty members (excl. heirs) 23.6 0.09 0 55 Prop. marrying commoners (excl. heirs) 0.6 0 0 1 Observations (individuals) 10,364 Observations (dynasties) 553

B. Land distribution c.1870 (Bateman 1883) Acreage Peer great landowners 114,553 12,776 13,789 571,456 Commoner great landowners 172,067 20,209 9,640 965,985 Squires 85,492 8,299 7,471 382,500 Greater yeomen 94,053 9,085 4,577 389,500 Lesser yeomen 82,222 8,850 5,440 377,910 Small proprietors 79,646 9,410 6,782 421,966 Cottagers 3,125 517 90 20,737 Public bodies 28,837 3,557 2,030 140,452 Waste 34,127 8,381 254 351,369 Total 694,122 73,485 93,489 3,621,875 Gini†‡ 0.926 0.005 0.812 0.979 Share of land by great landowners∗ 0.414 0.014 0.161 0.709 Observations (counties) 55 ∗ Time span between the birth of the ﬁrst recorded member of the dynasty and the death of the member listed in Bateman(1883). † Excludes waste and land owned by public bodies. ‡ This index only captures inequality across groups. Within-group data is only available for great landowners.

55 PanelB presents summary statistics for the county-level measures of landownership concentration computerized from Bateman(1883)’s appendix. Land was ex- tremely concentrated in late nineteenth century Britain: the two smallest groups of landowners, peers and commoner great landowners, owned on average c.300,000 acres, which amounts to 40 percent of the arable land. To calculate the Gini index for the land distribution for each county, I assume no inequality within the seven categories of landowners considered by Bateman(1883) (i.e., peers, commoner great landowners, squires, greater yeomen, lesser yeomen, small proprietors, and cottagers). I exclude waste and land owned by public bodies from the calculations. The average county in the sample presents a Gini index of 0.926. In England, Westmoreland and Middlesex present the lowest Gini indices while Northumberland was the county with higher landownership concentration.

Finally, the regression analysis in Section 4.3 in the paper includes marriage market controls: the relative size of the social class, the relative size of the land class, the relative size of the marriageable cohort near one’s country seat, sex ratios, and the length of the railway network. Next, I describe how I calculated these covariates.

Relative size of class: The relative size of the social class for individual i is the percentage of people of the opposite sex aged ± 2 years her own age who belong to the same social class. Social classes are (i) duke, earl, or marquis, (ii) baron or viscount, (iii) baronet, (iv) knight, and (v) commoner. For a landowner i, the relative size of class is the percentage of women aged 18–24 with a relative listed in Bate- man(1883) whose acreage or land rents are in the same class as i. Acreage classes are (i) more than 100,000 acres; (ii) 50,000 to 100,000 acres; (iii) 20,000 to 50,000 acres; (iv) 10,000 to 20,000 acres; (v) 6,000 to 10,000 acres; and (vi) 2,000 to 6,000 acres (Bateman 1883: 495). Land rents’ classes are deﬁned in terms of deciles, i.e., two individuals are in the same class if land rents are in the same decile, one decile above, or one decile below. Finally, to compute the relative size of the marriageable cohort around i’s country seat I use the percentage of women aged ± 2 years i’s age and whose family seat is in the same region as i’s. Regions are ﬁrst-level NUTS for England and Wales, electoral regions for Scotland, and provinces for Ireland.

Sex ratio: I compute the ratio of men to women in the marriage market as the

56 number of peers and peers’ sons aged 19–25 to peers’ daughters aged 18–24. Eighteen was the earliest age at which a girl was presented at court. After 24, the hazard rate for women sharply decreases (see Figure A.V). The year lag for men accounts for the fact that they married later. When the number of women is underreported, I use the sex ratio at birth, i.e., I assume that the number of women in 0.95 times the number of men born in the same year.

Railway length: This measure is the number of miles of railway in Britain and Ireland each year. Data is from Mitchell (1988), Ch.10, Table 5. Railway length proxies for commuting costs the year that a marriage took place, but also for the power and riches of industrialists.

A3 Additional descriptive statistics

A3.1 Pressure to marry young The principal threat to identification to my quasi-experimental approach is that least attractive women may select themselves into marriage during the interruption of the Season while attractive women may wait for it to resume. In Section 4.2 in the paper I argue that one of the reasons why this scenario is unlikely is the high pressure to marry young in Victorian Britain. Independently of their status, wealth, or attractiveness, ladies had only two to three Seasons to get engaged after being presented at court. After that, they were written off as failures (Davidoff 1973: 52). This pressure to marry young was so deeply internalized that, as Queen Victoria’s mourning extended over time — far more than expected — ladies panicked and got married even if the Season was cancelled. Were ladies pressured to marry quickly as suggested by the anecdotal evidence? Figure A.V shows that women’s implied market value, measured as the rate of intermarriage with peers’ (Panel A) and dukes’ heirs (Panel B) indeed decreased with age. The same holds in terms of husbands’ landholdings (Panel C). Specifically, the decline starts at age 22. At this age, the probability of marrying a peers’ son drops from 45 to 30 percent, the probability of marrying a duke heir from 20 to less than 10 percent, and on average, husbands are poorer by 20,000 acres. Moreover, it seems that as a woman approached the “Rubicon” of 30 years, her attractiveness fell dramatically in the eyes of her suitors in the Season. Furthermore, Panel D suggests that, independently of one’s social position, delay-

57 ing marriage was not a good option. The panel plots the probability of marrying in the peerage for dukes’ and barons’ daughters separately. Clearly, the depreciation of a woman’s attractiveness crucially depended on her implicit “quality:” The devaluation for dukes’ daughters was much steeper than that of barons’ daughters.

Panel A: Marrying in the peerage Panel B: Marrying a duke heir 30 50 40 20 30 % % 20 10 10 0 0 18−20 21−23 24−26 27−29 30−32 33−35 36−38 18−20 21−23 24−26 27−29 30−32 33−35 36−38 Age at marriage Age at marriage

Panel C: Husband’s landholdings Panel D: Marrying in the peerage 50 70 60 40 50 40 30 % 30 20 20 thousands of acres 10 10 0 18−20 21−23 24−26 27−29 30−32 33−35 36−38

0 Age at marriage 18−21 22−25 26−29 Age at marriage Duke daughters Baron daughters

Figure A.v: Implied market value of women, by age group (1851–75). The sample is all 796 peers’ daughters ﬁrst marrying in 1851–75. In Panel C, the sample is restricted to all 113 peers’ daughters marrying a peer in possession of 2,000 acres and upwards, worth £2,000 a year by 1876. In Panels A and D diamonds indicate the percentage marrying a peer or a peer son.

A3.2 Other sources of wealth Section 5.2 in the paper shows that marital sorting is important for wealth inequality. One of the limitations of the analysis is that it only considers landed wealth. How important were other sources of wealth, especially after the Industrial Revolution? Figure A.VI plots the percentage of people leaving £1,000,000 or more, by source of wealth. The data is from Rubinstein(1977). In the ﬁrst half of the nineteenth century, 95 percent of all millionaires were landowners. The percentage remains

58 around 80 percent until the 1870s. By the turn of the century, land was still the principal source of wealth. Although the share of millionaires who earned their fortunes in manufacture, food, drink, or tobacco industries grows over time, it is not until the twentieth century that this source of wealth became more important than landownership. The percentage of people earning their fortunes in commerce follows a similar pattern. Finally, professionals and public administrators always represented the smallest share of millionaires. 100 80 60 40 % of Millionaires 20 0 1809−58 1858−79 1880−99 1900−14

Manufacture, food, drink, tobacco Commercial Professional, public administration, defence Land

Figure A.vi: Top wealth-holders by source of wealth. The sample is all persons leaving £1,000,000 or more and landowners leaving estates worth over £33,000 per year (equivalent to £1,000,000 in 30 years), based upon Bateman(1883). Individuals are assigned to the occupation in which they made their fortune based on biographical evidence. Source: Rubinstein(1977)

There are several anecdotes that illustrate this pattern. In the 1890s, the London estates of the duke of Westminster were worth £14 million; at a time when the typical manufacturer left around £100,000 and the richest businessman left no more than £6 million. The catch-up process took a long time: businessmen began to leave estates exceeding that of Westminster only in the 1920s (Rubinstein 1977: 104). This evidence is consistent with previous work analyzing the evolution of inequality in Britain in the long-run. Miles(1993), Mitch(1993), and Long and Ferrie(2013) show low rates of occupational mobility in the nineteenth century. Clark et al.(2014) follow a diﬀerent approach. Tracking wealth at death of people with rare surnames, they conclude that the top of the wealth distribution in England has remained largely

59 unchanged from 1800. According to this view, the Industrial Revolution did not seem to accelerate social mobility.

A3.3 External validity: the peerage vs. the top 1 percent Admittedly, this paper looks at a very specific population living in a very specific historical setting. Can the study of the British peerage in the nineteenth century tell us something about modern-day elites? Table A.IV compares the nineteenth century British peerage to the people at the top 1 percent of the income distribution in the United States today. Both elites are immensely rich. The average annual income of an individual in the top 1 percent is $1,300 thousand (Wolff 2012). The average British peer listed in Bateman(1883) as a great landowner earned £23,400 a year. To calculate the equivalent in current US dollars, I first multiply this number by the percentage increase in the retail price index from 1870 to 2008. Then I use a 1.5 pound to dollar exchange rate. This gives me an annual income of about $2,500 thousand. Peers’ incomes almost doubled the ones of the top 1 percent. In terms of marriage outcomes, both elites are remarkably comparable. According to Bakija et al.(2012), Table 4, 12.5 percent of members of the top 1 percent never marry. The corresponding percentage for peers and peers’ sons is similar: 15.2. In- terestingly, the degree of marital sorting is also similar across groups. Among the top 1 percent, 24 percent marry spouses in a top-notch job: executives, managers, super- visors, financial professions, lawyers, business operations, professors and scientists, entrepreneurs, and skilled sales (excluding real estate). This percentage is almost the same as the proportion of peers and peers’ sons marrying a peer daughter between 1851 and 1875, 25.8 percent. Finally, the top 1 percent and the British peerage show similar support for liberal policies (by the standards of their times). Combined Gallup polls between 2009 and 2011 indicate that, among the very wealthy, 20 percent say their political views are liberal (Saad 2011). Similarly, 20.7 percent of peers listed in Bateman(1883) as great landowners were members of liberal clubs like Brook’s, Reform, and Devonshire (Bateman 1883: 497). The support for conservative policies, however, was clearer among peers; 55.2 percent of them belonged to tory clubs (Carlton, Junior Carlton, Conservative, or St. Stephen’s), while only 39 percent of the top 1 percent call themselves conservatives.

60 Table A.iv

The British peerage and the top 1 percent in the United States today

US top 1 percent British peerage

Not married (%) 12.5 15.2 Spouse in top-notch job / peer daughter (%) 24 25.8 Liberal / member of a liberal club (%) 20 20.7 Conservative / member of a tory club (%) 39 55.2 Average annual income / land rents $1,300k 23.4k (in 1870 £) 2,500k (in 2008 $)

Notes for the top 1 percent: The sample for rows 1 and 2 is all households with incomes above $295,000 (excluding capital gains), as reported by the Statistics of Income division of the Internal Revenue Center in 2005 (Bakija et al. 2012, Table 4). Row 2 restricts the sample to married individuals. Rows 3 and 4 considers households surveyed in the Gallup polls in 2009–11 with incomes of at least $516,633, which corresponds to the top 1 percent in 2010 based on data from the Tax Policy Center. Row 4 uses households in the top 1 percent from the 2010 Survey of Consumer Finances (Wolff 2012) Notes for the peerage: The sample for row 1 is peers and peers’ sons born in 1820–45 living more than 35 years (i.e., marriageable in 1851–75). Row 2 considers peers and peers’ sons first marrying in 1851–75. Rows 3 to 5 consider peers and peers’ sons first marrying in 1851–75 and listed in Bateman(1883) as great landowners. See the text for details.

B Extensions

In this section, I evaluate whether the marriage market displays increasing returns to scale. Then, I look at fertility rates of matched couples and their degree of consanguinity to examine whether the Season had any eﬀects on match quality. Finally, I examine the role of homophilic preferences as an important determinant of marital sorting.

B1 Increasing returns to scale

Does the marriage market display increasing returns to scale? Gautier et al.(2010) and Botticini and Siow(2011) have used implicit diﬀerences in marriage market depth between the city and the countryside to answer this question. While these papers analyze whether an agglomeration makes matching more eﬃcient, I consider a matching technology that not only pooled large numbers of eligible singles together, but it was also meant to facilitate their introduction and courtship. In this respect, the Lon- don Season provides a better test bed for matching technologies that are increasingly

61 common today, such as online matching sites or speed dating events. To test for increasing returns to scale, I estimate the effect of the Season on sorting at different levels of attendance. Probit regressions of equation (4.3)’s form allow me to calculate nonlinear marginal effects. Figure B.I plots the results for the probability of marrying a commoner, the probability of marrying in the same class in terms of acres and land rents, and the probability of marrying a spouse from the same region. Results suggest that the larger the royal parties were, the greater the effect of bringing in additional guests on sorting. This effect is significant for sorting by acres and land rents. Thus, the Season was subject to increasing returns: as more people attended, singles met at a higher speed, pickiness increased, and sorting strengthened.

Dep. Variable: Husband is a commoner Dep. Variable: Wife in same acres class .001 .015 0 −.001 .01 −.002 −.003 .005 −.004 −.005 mg. effect of 100 attendees mg. effect of 100 attendees 0 −.006 0 1,000 2,000 3,000 4,000 5,000 6,000 0 1,000 2,000 3,000 4,000 5,000 6,000 Attendees at the Season Attendees at the Season

Dep. Variable: Wife in same rents class Dep. Variable: Wife in same region 0 .012 .01 −.005 .008 −.01 .006 .004 mg. of 100 attendees −.015 mg. of 100 attendees .002 −.02 0 1,000 2,000 3,000 4,000 5,000 6,000 0 1,000 2,000 3,000 4,000 5,000 6,000 Attendees at the Season Attendees at the Season

Figure B.i: Increasing returns to scale. This figure plots the marginal effect of 100 additional attendees to royal parties on various measures of sorting. I use the probit IV model in TablesIV to VI. Marginal effects are evaluated at different values of attendance (x-axis) and at the means of all other variables.

62 B2 Match quality

Does the matching technology affect the quality of the match? In this subsection, I evaluate the effects of the London Season on two measures that proxy for match quality: the degree of consanguinity between spouses and fertility. Consanguinity is a good proxy for match quality. Although the genetic risks associated to consanguinity are debated (Bittles 2012), there is an old stigma on cousin marriage. For example, Charles Darwin, who was married to his first cousin Emma Wedgwood, was deeply concerned with the health of his 10 children. In Goni (2015b), I show that the London Season prevented consanguinity. When the Season was interrupted after Prince Albert’s death, cousin marriage blossomed. In detail, as marriage decisions shifted to the local marriage markets, the children of the nobility faced a reduced pool of proper singles. In these circumstances, many considered marrying within their extended family to secure a noble match. The percentage of people marrying their second cousin increased by a factor of 5, from a 0.5 percent the years before and after the interruption (i.e., 1859–60 and 1864–67), to almost 3 percent in 1861–63. These matches were of poorer quality: children born to consanguineous unions were more likely to die before reaching marriage age, had fewer children, and were 50 percent more likely to be childless. Alternatively, one could evaluate the quality of the match by looking at how many children couples produced. Among noblemen, producing enough children to ensure succession is paramount. Furthermore, fertility can also proxy for marital happiness. Take, for example, the marriage of the Duke of Northumberland to Edith Campbell: they had 13 children, 10 of them after a heir was born. Other couples may not have been so eager to visit the spouses’ chambers.50 Table B.I reports the coefficients from a regression of fertility on attendance to the Season. In columns [1] and [2], the dependent variable is the number of children born to each couple. Results are modest and not significant when I instrument attendance to the Season with its interruption in 1861–63 and with the size of the cohort. In columns [3] and [4] I consider the number of births per year of fertile marriage lifetime (i.e., between ages 15 and 40). Results are small but significant: a 100 ad-

50Marital happiness is diﬃcult to measure, all the more as divorce was not common at the time. Although the Divorce and Matrimonial Causes Act of 1857 allowed divorce on the grounds of adultery, my data indicates that between 1851–75 only 31 peers and peers’ oﬀspring actually separated. The number of children born to a couple, instead, can serve as a proxy for marital happiness.

63 ditional guests in the Season increase in 0.002 the number of births per year, which is equivalent to 0.1 standard deviations. In columns 5 and 6, I look at a measure that takes into account the biological maximum fertility of women: births relative to what a Hutterite women would have produced. Hutterites are a traditionalist Chris- tian church fellowship that reject birth control, and thus can be taken as a proxy for unconstraint fertility (Clark 2007 and Voigtlander and Voth 2013). OLS and IV results suggest that when the Season worked smoothly, matched couples were closer to the Hutterites, that is, to the biological maximum fertility. In particular, 250 additional guests to royal parties increase by one percentage point the number of children relative to Hutterites.

Table B.i

The Season and the quality of the match (1851–75)

[1] [2] [3] [4] [5] [6]

Dep. Variable: Number of births total per year relative to Hutterites OLS IV OLS IV OLS IV

Attendees Season 0.025** 0.020 0.002*** 0.002*** 0.004** 0.004** (0.011) (0.014) (0.001) (0.001) (0.002) (0.002) Observations 466 479 466 479 462 474 R-squared 0.098 0.087 0.068 Sargan test . 1.30 . 1.10 . 1.18 . p=0.5 . p=0.6 . p=0.9 Individual controls YES YES YES YES YES YES Marriage market co. YES YES YES YES YES YES Decade FE YES YES YES YES YES YES Year trend YES YES YES YES YES YES Interruption IV - YES - YES - YES Cohort size IV - YES - YES - YES F-stat from ﬁrst-stage - 20.9 - 20.9 - 20.9

Note: The sample is all peers and peers’ daughters first marrying in 1851–75. I exclude women marrying over 30 and women having one or no children. The former might be hard-pressed to have children before 35, when fertility sharply declines, biasing the estimates upwards. The later may be infertile or have difficulties to produce a heir, biasing the estimates downwards. Columns [1] and [2] report marginal effects of 100 additional attendees at the Season on the total number of births. In Columns [3] and [4], the dependent variable is the number of births per year, considering the number of years a woman was effectively married between ages 15–40. Columns [5] and [6] consider the number of births relative to the births that an average Hutterite woman would have had in her place. Hutterites marital fertility rates are 0.55 for ages 20–24, 0.502 for 25–29, 0.447 for 30–34, 0.406 for 35–39, and 0.222 for 40–44 (Clark 2007). I exclude women for which the Hutterite counterfactual does not produce more than one child. Individual controls include indicators of social position at age 15, age at marriage, birth order (excluding heirs), and an indicator for English titles. Marriage market controls include sex ratio and railway length. See Online Appendix A.2 for detailed descriptions. Columns [2], [4], and [6] use the first-stage reported in TableIV, Panel B. Standard errors clustered by year are in parentheses; *** p<0.01, ** p<0.05, * p<0.1

64 All together, this suggests that the matching technology embedded in the Season had positive effects on match quality: it reduced the degree of consanguinity between spouses and matched couples present slightly better fertility records. The estimated effects, however, are much smaller than the effects on sorting by social position or riches. Marriages in the Season were, after all, not so much a matter of love as an important social and economic arrangement.

B3 Homophilic preferences

Homophily — a preference for others who are like ourselves — may lead to assortative matching in the marriage market. In this subsection I show that, in contrast to sorting by socio-economic status, sorting with respect to political ideology is explained by homophilic preferences and is not affected by search costs or marriage market segmentation. Most peers belonged to political clubs: Brook’s, Reform, and Devonshire were liberal clubs, and Carlton, Jr. Carlton, Conservative, and St. Stephen’s were tory clubs (Bateman 1883: 497). Table B.II cross-tabulates the political ideology of spouses proxied by the ideological allegiance of the club they, or their families, belonged to. In most cases, husband and wife shared the same political views. For example, 39 percent of liberal husbands married wives from liberal families, while under a random assignment only 29.5 percent of them would marry a wife with the same ideology. Aggregate statistics confirm these patterns, although they also suggest that sorting in this dimension was not as strong as sorting by social position. Figure B.II plots the probability of marrying a like-minded partner over time. Po- litical endogamy seems to be independent of the number of attendees to royal parties, and it was not affected by the interruption of the Season during Queen Victoria’s mourning. Therefore, I conclude that in contrast to sorting by socio-economic status, sorting by political ideology was mainly driven by homophilic preferences, independent of search costs and marriage market segmentation.

65 Table B.ii

Marriage and political preferences (1817–1875)

Wife’s father or brother

Husband: Liberal club Tory club N

Liberal club observed 39.5 60.5 43 expected 29.3 70.7 diﬀerence 10.2* -10.2* Tory club observed 24.7 75.3 99 expected 29.3 70.7 diﬀerence -4.6* 4.6* N 42 100 142

Pearson chi-squared (1): 3.14* (0.08) Gamma test: 0.33 (0.17) Note: The sample comprises all 142 peers and peers’ sons who (1) ﬁrst married in 1817–75, (2) are listed in Bateman(1883), (3) belonged to a political club, and (4) married a wife who had a relative in a political club. I include individuals marrying from 1817 onwards to increase the sample size. Brook’s, Reform, and Devonshire are liberal Clubs; Carlton, Junior Carlton, Conservative, or St. Stephen’s are tories (Bateman 1883: 497); *** p<0.01, ** p<0.05, * p<0.1 100 120000 80 80000 % 60 attendees 40000 40 0 20 1851−55 1856−60 1861−65 1866−70 1871−75

Royal parties Political endogamy c.i. (95%)

Figure B.ii: Political endogamy over time. Shaded bars show the number of attendees at royal parties over 5-year intervals (left axis). For the connected line, the sample is all 92 peers who (1) ﬁrst married in 1851–75, (2) were listed by Bateman(1883), (3) belonged to a political club, and (4) married a wife with a relative in a political club. Diamonds display the percentage marrying a wife with a relative in a club with the same ideology (right axis). Brook’s, Reform, and Devonshire are liberal; Carlton, Junior Carlton, Conservative, or St. Stephen’s are tories Bateman(1883): 497. 66 C Robustness

In this section, I conduct a placebo test using cohorts not affected by the interruption of the Season in 1861–63. I then stratify my dataset by observables in order to identify the segments of the peerage for which the effects of the Season are more pronounced. I also examine the robustness of my results to using alternative measures of the London Season. Finally, I explore the validity of the cohort size instrument. First, I gauge the potential effect of unobserved variables using an insight from Altonji et al. (2005). Second, I assess the bias of the estimates in case the cohort size instrument is “plausibly” exogenous, i.e., it has some correlation with unobservables that are influencing marriage outcomes.

C1 Placebo test

Imagine Prince Albert died 10 years before he actually did. The Season would have been interrupted between 1851 and 1853. Ladies marrying during this “placebo” interruption, however, should not marry worse than those marrying in the years before and after, as it actually happened when the Season was truly suspended in 1861–63. I perform this placebo test in Figure C.I. Panel A shows the true effects of the interruption of the Season in 1861–63. Panel B looks at the average effects of a placebo interruption of the Season 10, 11, 12, ... and 50 years before it actually happened. Clearly, the synthetic cohorts do not change their marriage behavior in the years of the placebo interruption of the Season. Women above and below 22 years old are similarly likely to marry a commoner at all times. Only in 1861–63, when the Season was truly interrupted, a great gap between the two groups opens. Those ladies aged 22 or more who had to marry when the Season was disrupted were more likely to marry a commoner. This provides support for the identification assumptions of the quasi-experimental approach in Section 4.2 in the paper. In detail, the results of the placebo test suggest that the differential effects of the interruption of the Season on younger and older ladies is not the result of unobservable covariates correlated with age.

67 aso ihrscoeooi oiin hnteSao a eoeosy well- (exogenously) individ- was for Season effects the identified When tightly more position. and socio-economic stronger higher find of I uals (4.4). and (4.2) the and below versus median socio- the a above with PCA) individuals using and median. (calculated median, rents earning the index lords below status posses- great versus economic median, in median the the landowners below above non-heirs, versus land versus the median from the heirs of above into segments acreage different sample of across the sion Season subdivide the I of effects peerage. the compares subsection This stratification Sample 1861–63). before the C2 years in to 50 50 plotted corresponding ..., and are years 12, ..., ratios the 11, average indicates 12, 10, The area (i.e., 11, relative shaded at Season marrying averages. 1851, The the look take individuals 1, of and axis). then for interruption January (right A, I exercise placebo on line Panel this connected 1849–57. in 22 individuals repeat diamond in analyzed than take I the cohort is older I the that age. women before A, cutoff Season. for this years Panel this the reproduces commoners in below of B with women analyzed Panel interruption intermarriage to cohort the of axis). the marrying by (right rate 22 before first affected age than the years daughters cutoff not older 10 peers’ this women cohort marrying 276 below for synthetic first all women commoners a is to with for relative intermarriage sample exercise of 1861, the rate 1, line, the January connected display on the Diamonds For 1859–67. axis). in (left year per C. Figure n (4.3) and (4.2) systems equation the estimating of results the presents C.I Table i lcb test Placebo : attendees 0 1000 2000 3000 4000 5000 1859 1860 Panel A:Affectedcohort 1861 % marryingoutsidethepeerage Attendees atroyalparties 1862 1863 nPnlA hddbr hwtenme fatnesa oa parties royal at attendees of number the show bars shaded A, Panel In . 1864 1865 1866 1867

1 1.2 1.4 1.6 1.8 2 ratio (age ? 22 over < 22) 68 −2 −1 Panel B:Syntheticcohorts % marryingoutsidethepeerage(av.) Placebo interruptionyears 0 0 0 +1 +2 +3 +4

1.25 1.5 1.75 2 ratio (age ? 22 over < 22) attended, sorting by landholdings (rows [1] and [3]), land rents (rows [2] and [4]), and socio-economic ranking (row [5]) in general increased more for peers’ heirs and for landowners in possession of larger estates.51 In contrast, the eﬀect of the Season on geographic endogamy, that is, on the distance between spouses’ family seats, seems to come from individuals of lower status. Non-heirs, lesser landowners, and individuals with lower socio-economic index marry spouses from further away when the Season works smoothly. This suggests that although the London Season allowed heirs from highly accomplished families to marry better, their younger brothers were not reduced to staying at their country seats. They also participated in the Season, and consequently, they courted and married ladies from all over Britain and Ireland.

C3 Alternative measures of the Season

Table C.II examines the robustness of my results to using alternative measures of attendance to the London Season. Column [1] reports the baseline estimates using attendees at all royal parties as independent variable. Alternatively, column [2] uses the number of attendees at balls and concerts, the quintessence of the Season. The reported marginal effects are slightly stronger, except for those capturing the effect on geographic sorting (i.e., the probability that the wife is in the same region and distance between spouses’ family seats). A potential weakness of my analysis is that noblemen who were hard-pressed to marry into well-positioned families could have also been more eager to attend the Season. If this happened more when the size of the marriageable cohort was larger, my baseline estimates from the equation systems (4.2) and (4.3) and (4.2) and (4.4) would be biased. To account for this possibility, columns [3] and [4] use invitations issued to royal parties instead of the actual number of attendees. Marginal effects and standard errors are robust to this alternative measure of the Season.

51There are only two exceptions. The eﬀect of sorting by socio-economic index is stronger for non-heirs, and the eﬀect on the probability of marrying in your same land rents class is stronger for those with a smaller index of socio-economic status.

69 Table C.i Sample stratification

[1] [2] [3] [4] [5] [6] [7] [8] [9] Acres Land rents SES, using PCA

Above Below Above Below Above Below Baseline Heir No-heir median median median median median median Dep. Variable:

Wife in acres class 0.008∗∗∗ 0.018∗∗∗ 0.003 0.010∗∗∗ 0.005 0.011∗∗∗ 0.002 0.012∗∗∗ 0.008∗∗∗ (0.002) (0.003) (0.003) (0.004) (0.003) (0.004) (0.004) (0.003) (0.003) 257 126 131 130 127 128 129 129 128 Wife in rents class 0.009∗∗∗ 0.009∗∗ 0.006∗∗ 0.010∗ 0.008∗∗ 0.012∗∗ 0.006∗ 0.008 0.009∗∗∗ (0.002) (0.004) (0.002) (0.005) (0.003) (0.005) (0.004) (0.005) (0.003) 70 257 126 131 130 127 128 129 129 128

Missmatch in acres -0.23∗∗ -0.44 -0.14 -0.51∗∗ 0.24 -0.72∗∗∗ 0.54** -0.34 -0.33 (0.11) (0.33) (0.23) (0.20) (0.24) (0.22) (0.25) (0.34) (0.26) 171 96 75 94 77 100 71 87 84 Missmatch in rents -0.17∗ -0.19 0.035 -0.209 -0.109 -0.268 0.054 -0.479∗∗∗ 0.039 (0.09) (0.16) (0.227) (0.281) (0.207) (0.196) (0.166) (0.164) (0.164) 171 96 75 94 77 100 71 87 84 Missmatch in SES -0.68∗∗ -0.14 -0.90∗∗ -0.76 -0.69 -1.12∗∗∗ -0.54 -0.60 -0.51 (0.34) (0.68) (0.46) (0.59) (0.47) (0.38) (0.59) (1.17) (0.40) 170 95 75 94 76 100 70 87 83 Distance btw seats 1.7∗∗ -0.6 3.5∗∗∗ 1.9∗∗ 8.2∗∗∗ 1.1 7.8∗∗∗ 1.2 14.1∗∗∗ (0.8) (1.1) (1.3) (0.9) (2.1) (0.9) (1.6) (0.7) (5.2) 169 80 89 146 23 147 22 143 26 Note: This table reports IV marginal eﬀects of the number of attendees at royal parties (in 100s of attendees) on the corresponding marriage outcome in the rows. Each column divides the sample by heir status, acreage, land rents, and the SES index constructed using PCA. Regression samples and covariates are described in TablesIV toVI. First-stage results are reported in TableIV, Panel B. Standard errors clustered by year are in parentheses; *** p<0.01, ** p<0.05, * p<0.1 Table C.ii

Alternative measures of attendance to the Season

[1] [2] [3] [4] Independent variable capturing the eﬀect of the Season: Attendees at Invited at Attendees at Invited at royal parties royal parties balls and balls and (baseline) concerts concerts Dependent variable: Husband is a commoner -0.004*** -0.003*** -0.005*** -0.005*** (0.001) (0.001) (0.001) (0.001) Wife in same land class (acres) 0.008*** 0.007*** 0.008*** 0.007*** (0.002) (0.002) (0.002) (0.002) Wife in same land class (rents) 0.009*** 0.008*** 0.009*** 0.008***

71 (0.002) (0.002) (0.002) (0.002) Mismatch in acres -0.225** -0.21** -0.266** -0.246** (0.112) (0.107) (0.117) (0.11) Mismatch in land rents -0.167* -0.158* -0.204** -0.191** (0.086) (0.082) (0.083) (0.077) Mismatch in SES ranking, using PCA -0.713** -0.66** -0.855* -0.788* (0.351) (0.334) (0.482) (0.454) Wife in same region -0.008* -0.007* -0.006 -0.005 (0.004) (0.004) (0.006) (0.006) Distance between spouses’ seats 1.679** 1.565** 1.252 1.116 (0.806) (0.779) (0.896) (0.834) Note: This table reports IV marginal effects of the number of guests at the Season (in 100s of guests) on the corresponding marriage outcome in the rows. Each column defines guest in a different manner. Regression samples and covariates are described in TablesIV toVI. First-stage results are reported in TableIV, Panel B. Standard errors clustered by year are in parentheses; *** p<0.01, ** p<0.05, * p<0.1 C4 Assessing selection on unobservables

This subsection evaluates the size of the endogeneity bias in the absence of instru- mentation. Note that the IV and raw marginal effects reported in TablesIV andV are quite similar, suggesting that the bias is in fact small. Only when it comes to geographical endogamy does the need for an instrument stand out. The endogeneity bias will be large if there is substantial unobserved heterogeneity. To assess the potential effect of unobserved variables, I use the insight from Altonji et al.(2005) that selection on observables can be used to gauge the potential bias from unobservables. The strategy involves examining how much the coefficient of interest changes as control variables are added and then inferring how strong the effect of unobservables has to be to explain away the estimated effect. Formally, consider two 0 0 individual regressions of the form yi,t = β At + Vtλ + Xi,tδ + i,t. In one regression,

Vt and Xi,t only include a subset of control variables. Call the coefficient of interest in this “restricted” regression βR. In the other regression, covariates include the “full set” of controls. The corresponding coefficient is βF . The ratio βF /(βR − βF ) reflects how large the selection on unobservables needs to be (relative to observables) for results to become insignificant. Table C.III presents the estimated ratios. Of the 24 ratios reported,52 none is less than one. The ratios range from 1.2 to 68.5, with a mean ratio of 5.8. For example, when the restricted set of controls only includes time effects (i.e., a linear time trend and decade fixed effects), the effect of unobservables would have to be 2 times larger than the effect of the covariates to explain away the impact of the Season on the probability of peers’ daughters marrying commoners. In sum, results suggest that the endogeneity bias arising from unobserved heterogeneity is small in this context.

52Ratios for the distance between spouses’ seats are not reported because TableVI already makes clear that the endogeneity bias is strong in this dimension.

72 Table C.iii

Using selection from observables to assess the selection on unobservables

[1] [2] [3] [4] [5] [6] Husband is Wife in same class Mismatch SES ranking a commoner acres rents acres rents using PCA Controls in restricted set Controls in full set None All 1.2 1.8 1.2 1.5 1.2 2.5 Time effects All 2 2.5 2 3.9 1.5 4.1 Time effects + cohort controls All 9.5 5.8 6.1 68.5 3.4 3.9 Time effects + class controls All 2.2 2.5 1.9 5.3 1.5 3.8 0 0 Note: Each cell reports ratios based on the coefficients for Yi,t = β At + Xi,tλ + Vtδ + i,t from two individual-level linear regressions. Yi,t is the marriage outcome in the columns. At is the number of attendees at royal parties (in 100s of attendees). In one regression, the covariates Xi,t and Vi,t include the “restricted set” of control variables, described in each row. Call the coefficient of interest in this “restricted” regression βR. In the other regression, covariates include the full set of controls. Call the coefficient of interest in this full F F R F 73 regressions β . The reported ratio is the absolute value of β /(β − β ) Altonji et al.(2005). Time effects are a linear trend and decade fixed effects. Cohort controls are sex ratio and the relative size of class. Class controls include the relative size of class, indicators of social position at age 15 in col. [1], acreage and land rents percentile in cols. [2] to [5], and all the socio-economic variables used in the PCA in col. [6]. Regression samples and the full set of controls are described in TablesIV andV. C5 Plausibly exogenous instrument

A key assumption of my identification strategy is that my instruments affect marriage outcomes only through the Season. In section 4.2 in the paper, I argue that the interruption of the Season in 1861–63 satisfies the exclusion restriction. Here, I evaluate whether this is also the case for variation in the size of the cohort. On the one hand, note that the Sargan tests reported in TablesIV toVI cannot reject exogeneity of the instruments. The test is based on the assumption that at least one instrument is valid. Since the interruption of the Season after Prince Albert’s death is arguably an exogenous and excludable shock to the Season, the Sargan test is very informative. On the other hand, my results are not too sensitive to a hypothetical correlation between the size of the cohort and unobservables affecting marriage outcomes. To see this, I rewrite equations (4.2) to (4.4) to estimate the system in a two-stage least-squares framework:

0 0 At = ρ Cohort sizet + P (1861 − −63)t + Vtη + Xi,tδ + νt

ˆ 0 0 yi,t = β At + Vtλ + Xi,tδ + γ Cohort sizet + i,t, where γ is the direct effect of the size of the cohort on marriage outcomes; i.e., the effect that does not go through attendance at royal parties, captured by coefficient γ γ ρ. In this simple case, β(γ) = β(γ = 0) + , where is the bias from violating the ρ ρ exclusion restriction. Table C.IV reports the effects of the Season on marriage outcomes for different values of γ. Across specifications, the estimated coefficients and the standard errors do not vary much when γ < 0.1 · β and γ < 0.5 · β. The bias is meaningful only under a large violation of the exclusion restriction, i.e., when the direct effect of the cohort is almost the same as the effect of the Season. Although these results do not allow me to make inference about my estimates, they suggest that for plausible small violations of the exclusion restriction, the cohort size instrument would still be valid.

74 Table C.iv IV estimates for plausibly exogenous cohort size instrument

[1] [2] [3] [4] [5] Baseline γ = .1 · β γ = .25 · β γ = .5 · β γ = .75 · β

βˆ(γ) for Husband is comm. -0.011*** -0.013*** -0.015** -0.019** -0.023** (0.004) (0.005) (0.006) (0.008) (0.01)

βˆ(γ) for Wife in acres class 0.034*** 0.038*** 0.046*** 0.057*** 0.069*** (0.01) (0.011) (0.014) (0.018) (0.023)

βˆ(γ) for Wife in rents class 0.031*** 0.036*** 0.043*** 0.054*** 0.065*** (0.006) (0.007) (0.009) (0.013) (0.017)

βˆ(γ) for Mismatch in acres -0.225** -0.253** -0.293* -0.362* -0.43* (0.112) (0.127) (0.151) (0.191) (0.232)

βˆ(γ) for Mismatch in rents -0.167* -0.187* -0.217* -0.268* -0.318* (0.086) (0.098) (0.115) (0.145) (0.175)

βˆ(γ) for Mismatch in SES -0.715** -0.802** -0.933** -1.15* -1.367* (0.351) (0.399) (0.474) (0.6) (0.729)

βˆ(γ) for Wife in same region -0.028* -0.032* -0.038* -0.048* -0.057* (0.016) (0.018) (0.021) (0.026) (0.031)

βˆ(γ) for Distance btw seats 1.679** 1.91* 2.255 2.831 3.407 (0.806) (0.924) (1.107) (1.422) (1.744) Note: This table reports IV point estimates βˆ(γ) for the effects of the number of the number of attendees at royal parties (in 100s of attendees) on the corresponding marriage outcome in the rows. Each column assumes different values for γ, the direct effect of the cohort size instrument on marriage outcomes, i.e., Yi,t = βAˆt + 0 0 Xi,tλ + Vtδ + γ Cohortt + i,t in the second-stage described in Section 4.3 in the paper. Regression samples and covariates are described in TablesIV toVI. First-stage results are reported in TableIV, Panel B. Robust standard errors are in parentheses; *** p<0.01, ** p<0.05, * p<0.1

D Conceptual framework

This section develops a basic conceptual framework for understanding the determi- nants of marital sorting. I construct a two-sided search model that builds on Burdett and Coles(1997) and incorporate endogenous market segmentation. Using this model, I derive testable implications for the eﬀect of a reduction in search frictions and an increase in market segmentation on the strength of marital assortative matching.

D1 Baseline two-sided search model

Consider a market populated with a continuum of ex-ante heterogeneous men and women who wish to form long-term partnerships. Agents are characterized by their socio-economic status: x for men and y for women. Let x and y be distributed according to F (x) and G(y) over [0, 1]. The corresponding density functions are f(x) and g(y). All agents agree on how to rank one another. When a type x man matches

75 with a type y woman, the former receives utility y and the latter receives utility x,

ux(y) = y ∀x ∈ [0, 1]

uy(x) = x ∀y ∈ [0, 1]. Therefore, I follow Collin and McNamara(1990), Smith(1995), Bloch and Ryder (2000), Burdett and Coles(1997), and Eeckhout(1999) and assume utility to be nontransferable. Time is discrete and patience is determined by a discount factor β > 0. All men and women start their lives as singles, a state that yields no payoff. Because of search frictions, it takes time for agents to meet. The rate at which contacts are made is determined by a matching function. Given the measures of men (λm) and women (λw), the number of encounters is given by αM(λm, λw), where α is the efficiency of the matching function and M is increasing in both its arguments. I define αM(λm, λw) µ (λm, λw, α) = as the encounter rate for single women (analogous for w λw single men). When two singles meet, they decide whether to propose or not. A match is formed when both propose to each other. These agents then leave the pool of singles but are automatically replaced by two clones. This guarantees that the distributions G and F are time invariant.53 I now define optimal behavior for any woman y. Although being single is undesirable, it does not necessarily mean that a woman will match with the first person she meets. It might be wise to wait until a proper proposal comes. Thus, V (y), the value of being unmatched for women y, depends on the probability of eventually encountering “acceptable” agents. Formally,

Z 1 m w x (D.1) (1 − β)V (y) = β µw(λ , λ , α) Ω(y) max − V (y), 0 dF (x|y) , 0 1 − β

where Ω stands for the proportion of males who propose to her; F (x|y) is the distri- x bution of their socio-economic status; and is the value function for a woman 1 − β of type y married to a man of type x. x A woman y will agree to marry a man x if and only if ≥ V (y); that is, 1 − β she uses a reservation strategy. Note that reservation strategies will be nondecreasing

in type. Consider two women with types yh and yl, where yh > yl. Men accepting

53In the context of the Season, this assumption is justiﬁed by the fact that when the daughter of a noblemen gets married, her younger sister replaces her by coming out in the Season.

76 woman yl will also be willing to marry woman yh. Therefore, she cannot fare worse in the marriage market than any lower ranked women; i.e., V (yh) ≥ V (yl). Burdett and Coles(1997) used this argument to ﬁnd a unique equilibrium, which they called the Class Partition Equilibrium:

Proposition 1: (Class Partition Equilibrium.) The marriage equilibrium con- n N w sists of a sequence of strictly decreasing reservation strategies {a }n=0 for women and n N m 0 0 n n {b }n=0 for men, where a = b = 1 and for all n ≥ 1, a and b satisfy

Z an−1 n β m w n (D.2) a = µw(λ , λ , α) [x − a ]f(x)dx 1 − β an

and

Z bn−1 n β m w n (D.3) b = µm(λ , λ , α) [x − b ]g(x)dx . 1 − β bn

Proof: This proof follows Burdett and Coles(1997) and goes by induction. For the basis step, consider the problem faced by the most desirable woman (y = 1). All men will propose to her, so Ω(1) = 1 and F (x|1) = F (x) ∀x. Hence, her optimal β m m R 1 reservation match is r(1) = 1−β µw(λ , λ , α) r(1)[x − r(1)]f(x)dx. The reservation strategy for the most attractive man, ρ(1), is derived analogously. Note that r(1) < 1, ρ(1) < 1, r(1) = a1, and ρ(1) = b1 as deﬁned in Proposition1. Note also that if the most desirable woman (y = 1) or man (x = 1) is willing to accept an individual, then that individual shares the same reservation strategy as the most desirable of her sex. Consider a man of type x ∈ [r(1), 1]. Since the most desirable woman is willing to marry him, all women will be willing to marry him, and hence Ω(1) = 1 and G(y|x) = G(y) ∀y. This implies that ρ(x) = ρ(1). The same is true for women of type y ∈ [ρ(1), 1]. Redeﬁne a1 ≡ r(1) and b1 ≡ ρ(1). It follows clearly that men with x ∈ [a1, 1] and women with y ∈ [b1, 1] form an endogamic marriage class (class 1), in that agents in this class only marry members of this same class and reject all others. Now, assume that for n − 1, men with x ∈ [an−1, an−2] and women with y ∈ [bn−1, bn−2] form an endogamic marriage class (class n − 1), in that agents in this class only marry members of this same class, reject individuals of lower type, and are rejected by those in class n − 2.

77 For the inductive step, consider the most desirable women not in class n − 1, y0 + = bn−1 for an arbitrarily small > 0. By the inductive assumption, she is rejected by class n − 1 men. However, for all the men with x < an−1, she is the best available suitor. Thus, they all will propose to her. That is, Ω(y0) = F (an−1). The f(x) n−1 density function of these men under class n − 1 is given by F (an−1) for x ≤ a . Substituting this into equation (D.1) yields:

m w Z an−1 0 β αM(λ , λ ) 0 r(y ) = w (x − r(y ))f(x)dx . 1 − β λ r(y0)

0 n−1 0 n−1 g(y) n−1 Similarly, for men x + = a , Ω(x ) = G(b ) and G(bn−1) for y ≤ b . Thus,

m w Z bn−1 0 β αM(λ , λ ) 0 ρ(x ) = m (y − ρ(x ))g(y)dy . 1 − β λ ρ(x0)

Again, redeﬁne r(y0) ≡ an (ρ(x0) ≡ bn), which denotes the lowest type man (woman) acceptable to the most desired women (man) not in class n − 1. Since r(·)(ρ(·)) is nondecreasing, all women (men) not in class n − 1 will propose to a man (woman) with x ≥ an (y ≥ bn). Men satisfying x ∈ [an, an−1] and women with y ∈ [bn, bn−1] form marriage class n: they only accept each other, reject those of lower type, and are rejected by those in class n − 1. Q.E.D.

Under this simple preference specification in which one’s type affects her payoff only through whom she can match with, positive assortative matching arises natu- rally.54 The highest ranked men and women form endogamic marriage classes, while individuals in the lower tail of the socio-economic distribution, although preferring to marry top partners, are “forced” together. Sorting will therefore be stronger in equilibria with a larger number of smaller classes.55 I use this intuition to define the degree of sorting in a marriage equilibrium.

n N w n N m Definition 1: (Sorting) A marriage equilibrium {a }n=0, {b }n=0 displays a 54More generally, a class partition equilibrium arises when utility is non-transferable and preferences are multiplicatively separable, i.e., ui(x, y) = f1(x) · f2(y) for i = x, y (Eeckhout(1999)). 55To illustrate this, consider two extreme cases. If there is only one marriage class, all agents marry the ﬁrst person they meet; the characteristics of your spouse are completely independent of your own. That is, there is no sorting at all. Instead, consider an equilibrium in which people only marry those who look exactly like themselves. In this case, there are an inﬁnite number of “singleton” marriage classes, leading to perfect positive assortative matching.

78 n Nˆ w ˆn Nˆ m n n n ˆn larger degree of sorting than an equilibrium {aˆ }n=0, {b }n=0 if a ≥ aˆ and b ≥ b for all n, holding with inequality for some n, and N i ≥ Nˆ i for i = m, w.

D2 Implications of a reduction in search costs

Because of search frictions, it takes time for agents to meet. Here I consider two ways in which the matching technology can increase the rate at which contacts are made, and thus reduce search frictions: a more efficient matching technology and an m w increase in market depth. Formally, I assume that the encounter rate µi(λ , λ , α) = αM(λm, λw) for i = {w, m} is increasing in α and in λi. Thus, I depart from the λw standard assumption in the literature of constant returns to scale.56 57 How is marital sorting affected by such reductions in search frictions? The main trade-off that agents face in this model is between marrying sooner to enjoy marriage flow utility and waiting to get a proper match. As search frictions are reduced, the value of waiting increases; singles reject more offers and end up forming a larger number of smaller classes in equilibrium. In other words, sorting increases. Figure D.I gives an example of how the class equilibrium changes as the matching technology becomes more efficient and as participation rates increase. In this example, populations are symmetric, λm = λw = λ, and types are uniformly distributed on [0, 1], F (x) = G(x) = x. The matching function is M(λ) = λ2, so the encounter probability is subject to increasing returns to scale. Equilibrium classes are, thus, a0 = 1 , √ √ n n−1 1−β p n−1 a = a − βαλ 1 − β + 2βαλa − 1 − β . Assuming that λ = 1 and the discount factor β is 0.8, an increase in the efficiency of the matching technology from α = 0.5 to α = 1 increases the number and decreases the size of the equilibrium marriage classes (PanelA). Similarly, an increase in the mass of participants from

56Other models departing from the constant returns to scale assumption include Mortensen(1988), Chiappori and Weiss(2006), Anderberg(2002), and Gautier et al.(2010). In my context, this choice is justiﬁed by the fact that noble families from all over the country moved to London to get their oﬀspring married, which hints at the existence of some sort of increasing returns to scale in the matching technology embedded in the Season. 57The clone replacement assumption (i.e., the fact that matched agents are automatically replaced by two clones in the pool of singles) is crucial in order to avoid multiple equilibria once I introduce increasing returns to scale.

79 λ = 1 to λ = 1.5 also leads to more sorting in the marriage market (PanelB).

Panel A: Encounter speed (α) Panel B: Num. of attendees (λ) y y

x x β = 0.8; λ = 1; and α = 0.5 β = 0.8; λ = 1.0 ; and α = 0.5 β = 0.8; λ = 1; and α = 1.0 β = 0.8; λ = 1.5 ; and α = 0.5

Figure D.i: Comparative statics on search costs This ﬁgure displays the equilibrium deﬁned in Proposition1 with symmetric populations, λm = λw = λ, and types uniformly distributed on [0, 1], F (x) = G(x) = x. The matching function is M(λ) = λ2, so the encounter probability is subject to increasing returns to scale.

Proposition2 generalizes this result:

Proposition 2: As the matching technology becomes more eﬃcient (larger α) and as the measure of men and women increases (larger λm, λw), the degree of sorting in equilibrium increases.

Proof: This proof follows Bloch and Ryder(2000). According to Proposition1, class bounds are such that

m w Z an−1 n β αM(λ , λ ) n a − w (x − a )f(x)dx = 0 . 1 − β λ an

Using the implicit function theorem, the Leibniz integral rule, and some rearrange-

80 ment, I ﬁnd that

m w β M(λ , λ ) n−1 R a (x − an)f(x)dx ∂an 1 − β λw an = ≥ 0 . ∂α β αM(λm, λw) 1 + [F (an−1) − F (an)] 1 − β λw

Similarly, if the matching technology is subject to increasing returns to scale, i.e., ∂M(λm,λw)/λw ∂λw > 0 then

m w w β ∂M(λ , λ )/λ n−1 α R a (x − an)f(x)dx ∂an 1 − β ∂λw an = ≥ 0 . ∂λw β αM(λm, λw) 1 + [F (an−1) − F (an)] 1 − β λw

∂a1 The proof now goes by induction. For the basis step (n = 1), note that > 0 and ∂α ∂a1 ∂an−1 ∂an−1 > 0. Assume that for n − 1, ≥ 0 and ≥ 0. For the inductive step ∂λw ∂α ∂λw note that dan ∂an ∂an ∂an−1 = + dα ∂α ∂an−1 ∂α and dan ∂an ∂an ∂an−1 = + . dλw ∂λw ∂an−1 ∂λw ∂an−1 ∂an−1 By the inductive hypothesis, ≥ 0 and ≥ 0. Also, using the implicit ∂α ∂λw function theorem, Leibniz integral rule, and some rearrangement, it can be shown that β αM(λm, λw) (an−1 − an)f(an−1) ∂an 1 − β λw = ≥ 0 . ∂an−1 β αM(λm, λw) 1 + [F (an−1) − F (an)] 1 − β λw dan dan dbn Therefore, ≥ 0 and ≥ 0. A similar argument shows that ≥ 0 and dα dλw dα dbn ≥ 0 for all n = 1, ..., N m. Q.E.D. dλm

Furthermore, if the increase in the encounter rate is large enough, the equilibrium might reach perfect assortative matching, i.e., the nth ranked woman marries the nth ranked man.

81 Proposition 3: (Adachi 2003) As search costs become negligible, the set of equilibria converges to the set of stable matches derived under the deferred acceptance algorithm (Gale and Shapley 1962), with perfect assortative matching.

Proof: This proof follows Bloch and Ryder(2000). For ease of exposition, assume men and women are symmetric, i.e., λ ≡ λm = λw, and F (x) = G(y) ∀x = y ∈ [0, 1]. I start by deﬁning the set of stable matches under the deferred acceptance algorithm (Gale and Shapley 1962).

Definition 2: A matching is a one-to-one measure-preserving mapping from the set of men to the set of women. A matching is optimal if it maximizes total utility. A matching σ is unstable if there exists a blocking couple (x,y) in which both x and y are individually better oﬀ together than with the agent to which they are matched under σ, i.e., y > σ(x) and x > σ−1(y). The Gale-Shapley deferred acceptance algorithm yields a stable and optimal matching ν.

Lemma 1: Under the assumption than men and women are symmetric, the Gale- Shapley deferred acceptance algorithm yields a unique stable and optimal matching ν such that ν(x) = x.

To proof lemma 1, note that under symmetric populations and since one’s type does not aﬀects her payoﬀ, any measure-preserving mapping is optimal. Formally, R 1 R 1 Uν = 0 xf(x)dx = Uσ = 0 σ(x)f(x)dx for any measure-preserving matching σ, where U is the total utility. Consider any measure-preserving matching σ : [0, 1] → [0, 1] such that σ(x) 6= ν(x). To show that such mapping σ is not stable, I partition the set of men into three disjoint sets: those who are better or under σ, those who are assigned to the same women under σ and ν, and those that prefer their ν assignment.

X = {x ∈ [0, 1] : σ(x) > ν(x)}

Y = {x ∈ [0, 1] : σ(x) = ν(x)}

Z = {x ∈ [0, 1] : σ(x) < ν(x)}

Since σ and ν are measure preserving and σ(x) 6= ν(x), X and Z have a positive −1 −1 measure. Now note that σ (x0) = σ (ν(x0)) = x1 can be interpreted as a mapping

82 assigning to any man x0 the man x1 whom, under σ, is matched to x0’s partner under ν. Clearly, σ−1(Y ) = Y , since these are the men whose assigned women do no change under σ and ν. Hence, σ−1(X ∪ Z) = X ∪ Z. I now show that σ−1(X) 6= X. Suppose −1 x1 = σ (x0) ∈ X ∀x0 ∈ X. Then σ(x1) = ν(x0) > ν(x1). Since ν(x) = x ∀x, −1 −1 xo > x1. Hence, σ would map X into a proper subset of X. Therefore, for σ to be measure preserving, there must be a full measure x ∈ Z : σ−1(x) ∈ X. But if σ−1(x) ∈ X, then x > σ−1(x) so that woman ν(x) = x prefers x to her match according to σ. Further, since x ∈ Z, σ(x) < ν(x) so man x prefers woman ν(x) = x to his current match σ(x). This couple (x, x) is indeed a blocking couple, implying that σ 6= ν is unstable. Finally, to show that ν(x) = x is stable, consider any blocking couple (x, y): y 6= x. If y > x, then the women prefers ν−1(y) = y to x. If y < x, it is the man who prefers ν(x) = x to y. This implies that the set of blocking couples for ν(x) = x is empty. This concludes the proof of lemma 1. Once equipped with Lemma 1, it is straightforward to show that as search frictions disappear, the marriage equilibrium converges to ν(x) = x. According to Proposition 2, as α increases, marriage classes in equilibrium become smaller. Formally,

n−1 β αM(λ) Z a an = (x − an)f(x)dx 1 − β λ an

∂an is such that ≥ 0. Similarly, using the implicit function theorem, the Leibniz ∂α integral rule, and some rearrangement,

β αM(λ) n−1 R a (x − an)f(x)dx ∂an (1 − β)2 λ an = ≥ 0 . ∂β β αM(λ) 1 + [F (an−1 − F (an)] 1 − β λ

dan ∂a1 Now I show that ≥ 0 by induction. Clearly, for a1, > 0. For any n > 2, dβ ∂β dan ∂an ∂an ∂an−1 ∂an ∂an = + ≥ 0 since ≥ 0, ≥ 0 as shown in the proof of dβ ∂β ∂an−1 ∂β ∂β ∂an−1 ∂an−1 Proposition2, and ≥ 0 by the inductive hypothesis. ∂β As search frictions disappear, that is, as the matching eﬃciency α and the dis-

83 n count factor β increase, the class bounds a collapse to two sequences {x}x∈[0,1]. The highest-type men and women x = 1 consequently adopt a threshold strategy such that they only match with agents of type x = 1. The highest ranked men and women not in class 1 again adopt a threshold strategy such that they only match with the highest ranked agents not in class 1. Iteration of this argument gives rise to ν(x) = x, the unique stable and optimal matching derived by the Gale-Shapley deferred acceptance algorithm (Lemma A1). Q.E.D.

Two testable implications follow from propositions2 and3. When the London Season was well-attended, the children of the nobility should marry others who are closer in the “pecking order”. Instead, when the Season was disrupted after Prince Albert’s death, I expect to observe much less sorting in terms of social status and landed wealth.

D3 Implications of market segmentation

In this section, I introduce endogenous segregation in the model and evaluate its eﬀects on marital sorting. Although segregation in the marriage market is common, this feature is not usually incorporated into two-sided models of marriage search. Bloch and Ryder(2000) and Jacquet and Tan(2007) stand as notable exceptions. Henceforth, for ease of exposition, I assume that the male and female populations are symmetric, i.e., that λm = λw = 1 and F (x) = G(x) ∀x ∈ [0, 1]. I introduce a market maker to the economy who proposes excluding the least desirable suitors from the marriage market by charging a participation fee p. Each agent can then decide whether to go to the exclusive marketplace and avoid meeting these suitors at a cost p or to remain in the unrestricted marriage market. I call an equilibrium in which the least desirable suitors are excluded a segregation equilibrium.

Definition 3: A segregation equilibrium is a measurable subset (z, 1] such that for all x ∈ (z, 1], V˜ (x) − p ≥ V (x), where V˜ and V are the corresponding values of searching in the exclusive and the unrestricted marriage markets, respectively.

Since the matching technology has increasing returns to scale, this model is subject to multiple equilibria. I show that a segregation equilibrium exists by constructing one. I ﬁrst deﬁne the marriage equilibria in the unrestricted and exclusive markets under segregation. After that, I calculate the equilibrium fee p∗. Finally, I show that

84 under segregation no agent has an individual incentive to switch from the exclusive to the unrestricted market, or vice versa. Provided that the segregation equilibrium exists, the unrestricted marriage market f(x) is characterized by a mass F (z) of individuals distributed according to . The F (z) n N n equilibrium takes the form of a class partition {a }n=0 in which the cluster’s bounds a are deﬁned according to Proposition1. Similarly, the exclusive marriage market would f(x) be populated with 1 − F (z) individuals distributed over . The equilibrium 1 − F (z) N˜ will also take the form of a class partition {a˜}n=0. The participation fee p has to be such that agents of type z do not want to switch to the exclusive marriage market. Note that a type z agent would be the most desirable individual in the unrestricted market. Thus, her value of search there would correspond to the value of search in the top class [a1, z]

β αM(F (z)) Z z f(x) V (z) = (x − V (z)) dx . 1 − β F (z) V (z) F (z) In contrast, in the exclusive marriage market, z would be on the lowest class [z, a˜N˜ ], with a value of search of

N˜ β αM(1 − F (z)) Z a˜ f(x) V˜ (z) = (x − z) dx . 1 − β 1 − F (z) z 1 − F (z)

Therefore, for the segregation equilibrium to exist, the participation fee has to be such that p∗ = V ˜(z) − V (z). Now I show that with this p∗ and under the belief that types above z participate in the exclusive marriage market, all agents of type x < z have an individual incentive to remain in the unrestricted market. First, consider all agents in [a1, z). Following the intuition in Proposition1, they will behave in the same way as z in the unrestricted marriage market, since there they are desired by the highest-type of the opposite sex. So, the value of searching for a mate in the unrestricted market is such that V (x) = V (z) = a1 for all x ∈ [a1, z). Alternatively, if agents in [a1, z) switched to the exclusive marriage market, they would at most be included in the last marriage class, as agent z. It could even be the case thata ˜N˜ > x for some x ∈ [a1, z), which means that nobody in the exclusive marriage market would marry them. In such a case, she would only marry agents of type x < z who also had switched markets and therefore have a value of search V˜ (x) ≤ V˜ (z). Altogether, this implies that for all x ∈ [a1, z),

85 V (x) ≥ V˜ (x) − p∗, and thus they prefer the unrestricted market. This result is not so clear for men and women in the second class of the unrestricted market, i.e., x ∈ [a2, a1). If, for example, the exclusive marriage market is such that aÑ˜ < z, it might be that some of these individuals of type x ∈ [a2, a1) are x > aÑ˜ . In that case, they would be accepted by the lowest class within the exclusive marriage market, implying V (x) < V (z) = V˜ (z) − p∗ = V˜ (x) − p∗. Therefore, in order to have a segregation equilibrium, it must be that z =a Ñ˜ . If this assumption holds, then V˜ (x) < V˜ (z), implying that V (x) > V˜ (x) − p∗ for all x < a1. In other words, individuals of type x < a1 also prefer to remain in the unrestricted market. Finally, I show that no type with x > z has an incentive to switch markets. Con- sider first the individuals of type x ∈ [z, aÑ˜−1), that is, in the lowest marriage class of the exclusive market. For them, V˜ (x) =a Ñ˜−1 = V˜ (z). If they instead switch to the unrestricted marriage market, they will be the most attractive types there, in the top class. Thus, V (x) = V (z). It then follows that V˜ (x) − p = V (x). Since the equilibrium cluster’s boundsa ñ are nondecreasing in x, for all x > aÑ˜−1, the value of searching in the exclusive market is such that V˜ (x) > aÑ˜−1 = V˜ (z). Then, V˜ (x) − p > V˜ (z) − p = V (z) = V (x); that is, all types with x > z prefer to pay the fee p∗ and attend the exclusive market. This concludes the construction of the segregation equilibrium. Q.E.D

To produce clear-cut comparative statics, I need to impose more structure on the matching technology. To see why, consider a technology where the fraction of the population that is matched increases too fast with respect to the measure of agents. In such a case, segregation will have two effects: first, it will reduce the number of participants and consequently the speed of encounters between remaining singles. Second, segregation will restrict the choice set and soften the congestion externality imposed by agents who meet but will never match. Since I am interested in understanding the second effect, I impose a limit on the degree of increasing returns to scale:

M(λ) M(λ) (D.4) 2α ≥ αM (λ) > α . λ λ λ

Therefore, I assume that the matching technology is less than quadratic: the number of matches increases by a factor less than 4 when the number of participants in the

86 market doubles (Jacquet and Tan 2007). How would the marriage equilibrium in the exclusive marriage market be affected by an increase in segregation? Under equation D.4, segregation affects sorting through restricting the choice set to more similar individuals, but also by reducing the congestion externality. As agents do no longer meet others with whom they would never match, the rate at which singles meet proper types increases. This will increase the value of waiting, will lead to more rejections in the marriage market, and finally to an increase in sorting. Figure D.II gives an example of how the class equilibrium changes as lower-ranked individuals are excluded from the market. The model is calibrated for the case of symmetric populations (λm = λw) and uniform distributions on [0, 1], F (x) = G(x) = x. The matching function is M(λ) = λ1.1, so the encounter probability is subject to increasing returns to scale but matched agents do not increase M(λ) 0 M(λ) too fast with respect to the measure of agents, i.e., 2α λ ≥ αM (λ) > α λ . Equilibrium classes in the exclusive marriage market are, thus,a ˜0 = 1 ,a ñ = √ 2 q √ n−1 (1−z) 1−β βαM(1−z) n−1 a˜ − βαM(1−z) 1 − β + 2 (1−z)2 a˜ − 1 − β . Assuming that λ = 1 and the discount rate β is 0.8, an increase in the segmentation from α = 0 to α = 0.24 leads to more sorting: on the one hand, the choice set is restricted to more similar individuals. On the other hand, the congestion externality, the time agents spent meeting types x < z with whom they will never marry, is reduced. This leads to an increase in class bounds, and therefore to an increase in sorting in the exclusive marriage market. Proposition4 generalizes this result:

Proposition 4: As segregation increases (larger z), the degree of sorting in equilibrium increases.

Proof: From Proposition1, it is clear that marriage classes in the exclusive market are deﬁned such that:

Z a˜n−1 n β M(1 − F (z)) n a˜ − α 2 (x − a˜ )f(x)dx = 0 . 1 − β [1 − F (z)] a˜n

Using the implicit function theorem, Leibniz integral rule, and some rearrangement,

87 I ﬁnd that

β 1 2αM(1 − F (z)) n−1 f(z) − αM (1 − F (z)) R a˜ (x − añ)f(x)dx ∂añ 1 − β [1 − F (z)]2 1 − F (z) λ añ = . ∂z β M(1 − F (z)) n−1 1 + α R a˜ (x − añ)f(x)dx 1 − β [1 − F (z)]2 añ

2αM(1 − F (z)) ∂a˜n Since, by assumption ≥ αM (1 − F (z)), it follows that ≥ 0. 1 − F (z) λ ∂z ∂a˜1 The proof now goes by induction. For the basis step (n = 1), note that ≥ 0. ∂z ∂a˜n−1 Assume that for n − 1, ≥ 0. For the inductive step note that ∂z

dañ ∂añ ∂añ ∂añ−1 = + . dz ∂z ∂añ−1 ∂z

∂a˜n−1 By the inductive hypothesis, . Also, as shown in the proof of Proposition2, ∂z

β αM(1 − F (z)) (ãn−1 − añ)f(ãn−1) ∂añ 1 − β 1 − F (z) = ≥ 0 . ∂añ−1 β αM(1 − F (z)) 1 + [F (ãn−1) − F (ãn)] 1 − β 1 − F (z)

da˜n Therefore, ≥ 0. Q.E.D. dz

In the empirical analysis, I test Proposition4 by looking at how marriage behavior responded to the interruption of the Season after Prince Albert’s death. Because royal parties were cancelled, poor and insigniﬁcant suitors were not fully screened out, that is, marriage market segmentation was reduced.

88 y

β = 0.8; λ = 1; α = 1; and z = 0.00 β = 0.8; λ = 1; α = 1; and z = 0.24

Figure D.ii: Comparative statics on market segmentation This ﬁgure displays the equilibrium in the exclusive marriage market when a matchmaker can induce segregation by imposing a participation fee p. In this example, populations are symmetric, λm = λw = λ, and types are uniformly distributed on [0, 1], F (x) = G(x) = x. The matching function is M(λ) = λ1.1, so the encounter probability is subject to increasing returns to scale but matched agents M(λ) 0 M(λ) do not increase too fast with respect to the measure of agents, i.e., 2α λ ≥ αM (λ) > α λ .

ADDITIONAL REFERENCES

Anderberg, D. (2002). “An Equilibrium Analysis of Marriage, Divorce and Risk-Sharing”, Royal Economic Society Annual Conference 2002 6, Royal Economic Society. Bakija, J., A. Cole, and B. T. Heim (2012). “Jobs and Income Growth of Top Earners and the Causes of Changing Income Inequality: Evidence from U.S. Tax Return Data”, Unpublished manuscript. web.williams.edu/Economics/wp/BakijaColeHeimJobsIncomeGrowthTopEarners.pdf Bittles, A. H. (2012). Consanguinity in Context Cambridge: Cambridge University Press. Burke, J. (1898). Burke’s Landed Gentry London: Burke’s family et al. Burke, J. (2003). Burke’s Peerage, Baronetage and Knightage 107th edition. Edited by Charles Mosley. Wilmington: Burke’s Peerage (Genealogical Books). Burke’s Family (2005). The History of Burke’s Landed Gentry Scotland: Burke’s Peerage and Gentry. Chiappori, P. A. and Y. Weiss (2006). “Divorce, Remarriage, and Welfare: A General Equilibrium Approach”, Journal of the European Economic Association, 4(2-3), 415-426. Clark, G. (2007). A Farewell to Alms: A Brief Economic History of the World Princeton: Princeton University Press. Cokayne, G. E. (1913). The Complete Peerage of England, Scotland, Ireland, Great Britain and the United Kingdom London: The St. Catherine Press, Ltd. Debrett, J. (1814). The Peerage of the United Kingdom of Great Britain and Ireland, vol. 1. London:

89 G. Woodfall. Goñi,M. (2015). “Kissing cousins. Estimating the causal effect of consanguinity on fertility using evidence from the London Season”, Manuscript, University of Vienna. Mitchell, B. R. (1988). British Historical Statistics, Cambridge: Cambridge University Press. Mortensen, D. T. (1988). “Matching: Finding a Partner for Life or Otherwise”, American Journal of Sociology, 94. Saad, L. (2011). “U.S. ‘1%’ Is More Republican, but Not More Conservative”, Accessed September 7, 2015. Gallup.com. Smith, L. (1995). “Cross-Sectional Dynamics in a Two-Sided Matching Model”, Working papers 95-14, Massachusetts Institute of Technology (MIT), Department of Economics. Voigtlander, N. and H.-J. Voth (2013). “How the West ‘Invented’ Fertility Restriction”, American Economic Review, 103(6), 2227-64. Wolff, E. N. (2012). “The Asset Price Meltdown and the Wealth of the Middle Class”, Working Paper 18559, National Bureau of Economic Research.