European Sociological Review volume 27 | number 2 | 2011 180-195 DOI:10.1093/esr/jcp068, available online at www.esr.oxfordjournals.org Online publication 2 February 2010 180 Who Contacts Whom? Educational Homophily in Online Mate Selection Jan Skopek, Florian Schulz and Hans-Peter Blossfeld Data from an online dating platform are used to study the importance of education for initiating and replying to online contacts. We analyse how these patterns are influenced by educational homophily and opportunity structures. Social exchange theory and mate search theory are used to explain online mate selection behaviour. Our results show that educational homophily is the dominant mechanism in online mate choice. Similarity in education significantly increases the rate of both sending and replying to initial contacts. After controlling for the opportunity structure on the platform, the preference for similar educated others is the most important factor, particularly among women. Our results also support the exchange theoretical idea that homophily increases with educational level. If dissimilarity contacting patterns are found, women are highly reluctant to contact partners with lower educational qualifications. Men, in contrast, do not have any problems to contact lower-qualified women. Studies of educational homogamy generally show that couples where women have a higher level of education are rare. Our study demonstrates that this is mainly the result of women's reluctance to contact lower qualified men. Introduction The remarkable individual propensity to associate with a partner who has similar characteristics is a recurrent empirical finding in the study of mate selection (see, for a recent study, Blossfeld, 2009). From a social structural point of view, this homophily has far-reaching consequences for the reproduction of social inequalities in modern society. One aspect that is particularly important for the process of homophily is an individual's education. More than ever before, education has become the pivotal determinant of occupational success, and it also reflects the cultural resources influencing individuals' preferences for specific partners. Therefore, educational homophily suggests that the degree of social inequality engendered in individuals' life courses will be further enhanced through their marriage choices, because the advantageous (and disadvantageous) economic and sociocultural resources of two individuals are then pooled and cumulated (e.g. Mare, 1991; Blossfeld and Timm, 2003). Several recent empirical studies show that educational homogamy1 has even increased in many industrialized countries in recent decades (see, for an overview, Mare, 1991; Kalmijn, 1998; Blossfeld, 2009). Blossfeld and Timm (2003) argued that the formation of similarly educated couples is largely influenced by structural contact opportunities in the educational system. In addition, women's changing economic role in dual-earner societies has increased the importance of their education and labour force attachment (Blossfeld and Drobnic, 2001). As a consequence, men in more recent birth cohorts should increasingly prefer highly qualified women, and this should accentuate the level of homogamy even more as the traditional marriage pattern of the male breadwinner declines. © The Author 2010. Published by Oxford University Press. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com educational homophily in online mate selection | 181 Despite numerous studies reporting empirical evidence favouring either the structural or the intentional mechanisms of mate selection (see, for recent reviews, e.g. Kalmijn, 1998; Blossfeld, 2009), we still do not know how structural opportunities and individual strategies actually operate together: is educational homogamy primarily a consequence of the structure of the educational system, and thus, simply an institutionally preformed phenomenon? Or is educational homogamy more a result of homophily, implying that it is actors making intentional choices who systematically want to associate with similarly educated partners? The present study addresses this question by analysing the process of mate selection in online dating. We focus on education-specific patterns of assortative mating and examine whether and to what extent similarly educated participants in online dating platforms contact each other on a specific German Internet dating website and also whether there are gender-specific variations in their decisions. The Internet marriage market has grown rapidly in the last few years, and is now regularly used for mate search by about 5.5 million Germans (Schulz et al., 2008). Since becoming a mass phenomenon, online dating is often regarded in the literature as a driving force in the reduction of social inequalities, because such online platforms are less restricted and quite open (Illouz, 2006). The implication is that people meet and mate beyond traditional social barriers. If this was indeed the case, the recent macro-level findings indicating high levels of homogamy should primarily be an outcome of structural or institutional selection effects on local marriage markets rather than homo-phile individual preferences and strategies. Up to now, there have been only a few empirical studies of online dating (e.g. Fiore and Donath, 2005; Hitsch et al., 2009, in press; Lee, 2008), and there is need for more empirical evidence to assess the relevance of preferences in the process of mate selection. Our choice of the digital marriage market to analyse mate selection has at least one substantive and one methodological advantage. Substantively, access to online dating platforms is neither restricted, nor is the context institutionally prestructured or selective compared with many contexts in everyday life.2 Although there might still be a remarkable digital divide in modern societies (e.g. Dewan and Riggins 2005; Schulz et al, 2008), once this barrier has been passed, the digital marriage market is an open meeting space compared with the restricted contact opportunities in everyday life. If users continue to frame their decisions along educational divides, we can conclude that this is an expression of individual preference. Methodologically, online dating offers unique data for sociological analyses, because it enables us to explicitly study social interactions on the micro level. Particularly when investigating the very early phase of the mate selection process, this is an excellent source of information. It enables us to work with non-reactive observational data on mutual contact processes recorded without contacting the platform users. Every action and decision on this online dating platform, from a single mouse click to long e-mail messages, is logged in a database, allowing an accurate time-dependent reconstruction of the single phases of the mate selection process. Thus, we are able to analyse the very early decisions in the partner selection process in great detail by looking at who contacts whom first by e-mail (initial contact) and who replies to whose contact offers. Indeed, it is during this phase of the mating process that decisions are made about whether a resource relation between two users might be rewarding, and these early decisions condition every later phase. According to the hypothesis on the path dependence of social phenomena, these early decisions in the process of partner choice set the course for the further development of relationships. In the following, we will develop the theoretical framework for our analysis. This framework is based mainly on exchange theory, and allows us to derive hypotheses on education-specific contact behaviour for men and women within the context of online mate search. We will then present our empirical results, before finally drawing several conclusions on what can be learned from our study. Online Mate Selection: Theories and Hypotheses The choice of a (marriage) partner is based on a longer process involving many successive decisions. Conceptually, actors pass through many consecutive phases of filtering the field of eligible partners until eventually only one actual marriage partner remains (e.g. Kerckhoff and Davis, 1962; Murstein, 1970). From the perspective of a dynamic analysis of social structure, the very first steps of mating, and especially the initial contact between two potential mates, are of decisive importance for the emergence of collective patterns of mate selection and thus for the production and reproduction of social inequality in society (e.g. Mare, 1991; Blossfeld, 2009). Hence, in our first step, we concentrated on the actor's propensity to initially contact other users in the online dating environment. 182 I škopek, schulz and blossfeld Nonetheless, it is clear that studies of initial contacting behaviour in online dating such as Lee (2008) for Korea, Hitsch et al. (2009; in press) and Fiore and Donath (2005) for the United States, or Skopek et al. (2009) for Germany address just one side of the coin.3 Partner choices are consensual choices. Whereas the person who is initially contacting the other takes the necessary first step to start a potential relationship, the future of this process also depends on the contacted person's decision whether or not to reply. The necessary and sufficient premises for forming a stable relationship require both actors to approve the partnership and understand themselves as a couple. Hence, if an initial contact on the dating site remains unanswered, the process of partnership formation will simply cease, meaning that the individual, one-sided choice of one actor has failed. Therefore, the reply to an initial contact offer is the very first consensual decision setting the course for the further interaction process. Based on these considerations, our second step was to analyse the reply behaviour to initial contacts in online dating. This enables us to empirically assess how gender- and education-specific patterns of mate selection emerging from the first contacts are mutually reinforced or may change. Based on Korean data, Lee (2008) argued that the selection at the very beginning of the mating process, namely, the initial contact and the possible reply that follows, are quite good indicators for the actual preferences and aspiration levels of the actors that eventually account for their marriage decisions. Whereas the study of initial contact behaviour permits fairly good assessments of individual dispositions and preferences (see Skopek et al., 2009), an analysis of the reply patterns casts more light on online mating patterns by explicitly addressing the mutual consen-suality of the mating decision. The Structural Logic of the Marriage Market When looking for a partner, individuals are constrained by the structural conditions of their marriage markets (Becker, 1974; Blau, 1994). Individual freedom of choice, and thus the chance of realizing individual intentions, is limited by institutional and social structures, and mate selection is particularly constrained by the available opportunities to meet and interact with potential partners in everyday life (Verbrugge, 1977). Social, economic and cultural contexts like the educational or the employment system, neighbourhoods or the circle of friends, structure actors' social networks and have to be considered as selective, numerically limited and institutionally organized submarkets (Kalmijn and Flap, 2001). Hence, the chances of meeting potential mates in everyday life vary in different phases of the life course and their corresponding (social) spaces and activities (see Feld, 1981; Kalmijn and Flap, 2001). From a structural point of view, educational homo-gamy in couples can then be explained simply by the probabilistic logic of opportunities determined by population structures on the one hand, or by focused activities in educationally segregated marriage sub-markets on the other hand. However, the process of mate selection cannot be explained by social structures alone, because these do not operate directly on social behaviour. After all, no couples will be formed if the individuals in question do not want to take advantage of the possibilities in their marriage markets (e.g. Homans, 1985). This is especially relevant for the analysis of the digital marriage market, because online dating systems are primarily designed for mating purposes, providing special environments to facilitate the process of forming intimate relationships. Alongside other seminal features such as anonymity, the irrelevance of space and time for meeting online or the new possibilities of self-presentation, the presumably most important characteristic of the Internet compared to the organized settings of everyday life is the broad absence of institutional selection processes (e.g. Ben-Ze'ev, 2004; Geser, 2007; Skopek et al., 2009). Thus, online dating systems offer a great potential for sociologists to analyse individual partner choices in a rather 'open' marriage market, because users need to consciously select their 'favourites' from an observable, rather heterogeneous field of eligibles in terms of basic socioeconomic attributes like education.4 Mate Selection as Social Exchange Drawing on social exchange theory, we assume that actors looking for partners on the marriage market are trying to increase their expected subjective utility compared to living alone by (socially) exchanging resources with other actors. Accordingly, the process of establishing and maintaining an intimate relationship is characterized by a long-term, mutual arrangement of 'giving and taking' resources between the actors involved (see Blau, 1964; Edwards, 1969; Blossfeld and Timm, 2003; Skopek et al., 2009). Following Becker (1974), women and men will only form couples if they expect profitable gains from the corresponding exchange relation. To qualify as potential exchange partners on the marriage market, actors must signal educational homophily in online mate selection | 183 their specific resources (e.g. education) to members of the opposite gender (see, e.g. Edwards, 1969; Anderson and Hamori, 2000). Rationally acting individuals on this market will therefore strive only for relations with an optimal cost-benefit balance: the higher the value of one's own resources, the higher the value of resources one can reasonably demand from possible partners. Nonetheless, partner search is a rather difficult type of decision-making process under uncertainty because of, for instance, incomplete or asymmetrical information about the 'offerings' or the utility of further search activity. Thus, an optimal solution of the decision problem is hard to achieve in everyday life (e.g. Oppenheimer, 1988; Todd and Miller, 1999; Blossfeld and Timm, 2003). This suggests that rational actors with limited time and knowledge will normally base their mating decisions on fast and frugal heuristics (Todd and Miller, 1999), particularly by applying a kind of satisficing heuristic with a minimal aspiration level (see Simon, 1956). Such satisficing takes a shortcut by setting an adjustable aspiration level and ending the search as soon as one alternative is encountered that exceeds this standard. Hence, rational actors make feasible decisions on the marriage market by trying to ensure that they do not sell themselves at less than fair value, thereby systematically rejecting potential mates with lower resources. Following Todd and Miller (1999), we assume that the adjustable aspiration level is based on an individual's own mate value, which, in turn, is based on past life course experiences and the mate values of those who do or do not show interest. Thus, the more 'attractive' and 'desirable' an actor is on a marriage market, the higher her or his own aspiration can be, and the more restrictive will be the specific set of acceptable partners. Education-specific Mechanisms of Mate Selection Applying this model to our question of mate selection on the Internet, we assume that users of online dating systems only contact people or reply to contact offers by others if they subjectively anticipate reciprocity. Following our exchange theory model, this is most likely when both partners have similar resources. No rational actor will accept a negative cost-benefit ratio, and thus both will refrain from contacting people with lower resources and reject contact offers from other users with lower resources. Due to the competition on the marriage market, the dominant outcome pattern in mate selection should prove to be couples with similar resources (see Edwards, 1969; Becker, 1974; and for empirical findings, e.g., Kalmijn, 1998; Blossfeld and Timm, 2003). This mechanism should operate for different attributes (e.g. Kalmijn, 1994; Kalmijn and Flap, 2001), and is expected to be stronger when actors have more valuable resources. Furthermore, the mechanism is basically symmetrical and thus rather gender neutral (e.g. Skopek et al., 2009). The idea that 'like is attracted to like' is also the basis for Becker's (1974) hypothesis of positive sorting by non-market traits on the marriage market. Becker predicts that the gains from marriage, compared to living alone, are highest when men and women resemble each other as much as possible in all personal attributes. The history of mate selection research reveals an extensive discussion of education as one of these attributes (Blossfeld, 2009). Education is regarded as a rather enduring attribute that is also highly recognizable and intersubjectively comparable. Moreover, it is a very rich indicator for numerous dimensions of everyday life that are commonly accepted as being associated positively with stable and satisfying partnerships or marriages. Educational similarity makes it easier to establish a joint lifestyle (Kalmijn and Bernasco, 2001), it is normally accompanied by similar (cultural) interests (e.g., DiMaggio and Mohr, 1985), it increases the chance of conflict-free communication, and thus it is more likely to create positive emotions and social affirmation within intimate relationships (Kalmijn, 1994). This homophily hypothesis is the central theoretical concept that this study tested with empirical data from online dating. We focused on educational homophily, because earlier research has shown repeatedly that similarity in educational background positively facilitates meeting, mating and marrying (e.g. Kalmijn, 1998; Blossfeld, 2009). Applied to the contact and reply behaviour in online dating, we hypothesized that users would be (hypothesis a) more likely to send an initial contact offer to an equally educated user, and (hypothesis b) more likely to reply to a contact offer from an equally educated user. If this proved to be true, then the process of contacting and interacting in online dating would systematically enforce the selection of educationally similar couples, and reduce the percentage of educationally dissimilar couples over time. Indeed, first analyses of contact behaviour in online dating in the United States and Korea have suggested that mate selection on the Internet is indeed based on homophily, indicating strong binding effects of similar education (Fiore and Donath, 2005; Hitsch et al, 2009, in press; Lee, 2008). Furthermore, we did not just expect that similarity in education would significantly increase the contact probability and the probability of replies between 184 I škopek, schulz and blossfeld users; we also expected educational homophily to vary across different levels of education as predicted by the above-mentioned model of exchange. Hence, we hypothesized (hypothesis c) that educational homophily would be higher for actors with higher educational status. Even though many empirical studies have supported the homophily and homogamy hypothesis, it is still quite common for men and women to at least partly follow traditional gender and family roles (see Blossfeld and Timm, 2003). This is particularly the case in a conservative welfare regime like Germany. Germany has a strongly gendered division of labour in society, reflected, for example, by segregated labour markets, gender-specific income and occupational structures or an unequal distribution of housework and childcare between men and women (see, for an overview, e.g. Blossfeld and Drobnic, 2001). The stronger the traditional gender roles in a society, the more men will invest in labour market skills (with a high income potential) and women in non-market skills (qualifying them for homemaking and care giving). If they seek such a traditional family model with a male breadwinner, women will search mainly for men with good socioeconomic resources such as a high income potential, which can, in turn, be approximated by good education (Oppenheimer, 1988). Traditionally oriented men, however, will primarily search for women with high competencies in the non-market sphere, for example, in homemaking. On the aggregate level, this leads to a negative sorting along market traits (Becker, 1974) that systematically matches men and women with quite different market resources; in the traditional case, for example, highly educated men also look for less educated women. Of course, this model is being challenged increasingly by the changing economic role of women in modern societies. During the course of their increasing educational and labour market participation, and the transition from male breadwinner to dual-earner societies (Blossfeld and Drobnic, 2001), women's high educational resources are also becoming distinctive attributes in the mate selection process. In this respect, traditional gender-related patterns of mating should lose their empirical relevance, although we do not expect them to disappear completely in Germany. The decline in traditional couples also implies increasing rates of homogamy and increasing downward marriage by women. However, women do not marry downwardly as we would expect from a structural perspective (Blossfeld and Timm, 2003). Especially from their standpoint, this is quite a relevant problem: as women participate increasingly in higher educational tracks and thus have higher educational attainment levels, whereas some qualified men still marry downwardly, the number of appropriate partners for educated women at the top of the educational distribution decreases sharply. The implications may well be that they have to lower educational standards for partners, engage in prolonged search, or stay single. Research still lacks explanations regarding how this structural phenomenon can be traced back to individual decisions, and, in particular, it is not known who—men, women, or both—systematically shy away from these constellations. We wish to contribute to solving this puzzle in our study of choosing behaviour in online dating. As a preliminary hypothesis, we suggest (hypothesis d) that both men and women systematically avoid a couple constellation in which the woman is better educated than her male partner. Data We obtained our data from the provider company of a German online dating site that allowed us to access its database. The data cover user activities over a randomly chosen time period of about half a year between January and June 2007. The site targets a broad audience in Germany and does not just address specific populations in terms of region or social groups. Registered users can create their own user profiles (an online equivalent of a personal ad), look for other people by filtering the database using search forms and contact other people through an internal messaging system. The user profiles contain self-descriptions filled with standardized information like age, height, weight, educational attainment level, gender, marital status as well as photographs and also free-text descriptions. The dataset used in the present analysis contained user profile data and time-related data on e-mail exchanges between profiles. We could use these to reconstruct who sent an initial contact e-mail, and whether the contacted user replied to this e-mail. In addition, we had information on which other users' profiles a given user had been browsing through. Since a profile cannot be contacted before it has been browsed, we could distinguish between those profiles she or he was looking at both with and without a subsequent contact trial. Although the whole database was completely anonymized, we were able to use sociodemographic descriptors to characterize users. We focused on educational resources as the main educational homophily in online mate selection | 185 independent variable while controlling for other attributes like gender, age and physical appearance. We interpreted a first contact trial as a sign of the user's willingness to engage with the addressed user. Therefore, we call the former user the initiator and the latter the receiver. A given user can be either initiator or receiver, but not both in the same dyad. We applied the label reciprocal contact when a first contact trial by an initiator was answered by the receiver, thereby reflecting some kind of first consensual decision.5 Our empirical analysis is divided into two parts: The first analyses first-contact and first-reply behaviour in male and female users. It considers (a) the probability of a contact when a user has been browsed and (b) the probability of a reply when a user has been contacted. The second part narrows the focus of the analysis to educational homophily by estimating the probability of contacting other users displaying the same education in their profile. To disentangle intentional homophily from the structurally induced opportunity to meet, we incorporated the gender-specific educational opportunity structure of site users in our model. The model was also estimated for reciprocal contacts, assessing whether homophily is reinforced when the consen-suality of selection is taken into account. Sample Our sample consisted of users who sent at least one message to another user regardless of whether it was a first contact or a reply in our observation window (13,573 users). We excluded users who declared themselves 'unfaithful' or said they were looking for 'a mate for leisure and sporting activities.' The major part of the remaining sample (over 80 per cent of the original database dump) indicated that they were explicitly looking for a 'serious relationship.' In addition, we excluded a small number of users stating homosexual preferences in their profile and removed all same-sex interactions as well as self-directed messages. In a final step, we restricted the sample to first contact messages by cutting all subsequent messages in a dyad, retaining only the information on whether the receiver replied to a first contact. Our final sample contained 12,608 users (59 per cent male, 41 per cent female) and 116,138 first contacts. The average age of users was 36 years and did not differ significantly between men and women. Well over 80 per cent of both female and male users were aged between 20 and 50 years. This was about twice as high as the proportion of the overall population of Germany in this age range. Slightly more than 9 out of 10 users were unmarried or separated/divorced from a former partner. Whereas there were more men in the first group, women were overrepresented in the second one. The average man browsed about 138 profiles, sent about 12 first contact messages, and received about 4 answers. In contrast, women were less active on the platform in terms of browsing and sending e-mails. However, with almost the same average number of answers, they were evidently more successful in receiving replies to their contact offers. Table 1 summarizes the distributions for important profile characteristics as well as basic contact statistics.6 Table 1 Sociodemographic characteristics of users and contact statistics (column percentages) Men Women Total Educational level Not specified 16.57 17.94 17.13 Basic secondary 5.36 4.98 5.20 Vocational 28.88 33.01 30.58 secondary/ apprenticeship University entrance 19.81 20.93 20.27 University degree 29.38 23.14 26.82 Age (years) Not specified 0.05 0.04 0.05 <20 1.45 3.92 2.47 20-29 28.38 31.42 29.63 30-39 34.23 23.77 29.93 40-49 24.25 26.59 25.21 50-59 9.14 11.86 10.26 >60 2.49 2.39 2.45 Marital status Not specified 0.85 0.31 0.63 Single 70.98 56.95 65.22 Married 3.31 2.55 3.00 Separated/divorced 23.47 36.44 28.80 Widowed 1.39 3.75 2.36 Desired relationship Not specified 11.28 6.86 9.46 Chat/e-mail 6.15 10.80 8.06 friendship Serious relationship 82.57 82.35 82.48 Body mass index (mean) 24.47 22.88 23.87 Height (mean) 180.95 167.94 175.61 Contact statistics (mean) Browsed profiles 138.49 73.01 111.73 First contacts sent 11.72 5.61 9.21 Replies to first 3.64 4.39 3.95 contacts sent Individuals (N) 7,430 5,178 12,608 Calculations based on the sample of active users. Source: database dump of a German dating site, first half-year of 2007. 186 I škopek, schulz and blossfeld Variables Our central variable was educational attainment level. Users of the dating platform could choose between the standardized options 'basic secondary school', 'vocational secondary school', 'apprenticeship', 'university entrance qualification' and 'university degree'.' Because we had no information on educational attainment level for 17 per cent of the women and 18 per cent of the men in our sample, we excluded these cases from our analysis. We combined 'apprenticeship' (only few cases) and 'vocational secondary school' into one category. Men had a slightly higher educational level than women. Compared to the German population, more highly educated people were overrepresented on the platform and less well-educated people were underrepresented. This reflects a result known from other studies regarding the digital divide in using online dating websites (see Schulz et al., 2008; Sautter et al, 2009). We classified two users stating the same educational level as having educational equality. To enter education into our regression analyses in a parsimonious way, we treated it as a metric variable ranging from 1 to 4.8 Research has identified age and the relative age of partners as crucial factors for mate selection, and age homogamy is a particularly significant outcome in marriage markets (see Van Poppel et al., 2001). Therefore we controlled for age based on the dates of birth users reported in their profiles. Two users were classified as having age equality when their age difference did not exceed 2 years. We also controlled for physical attractiveness, which is regarded as another crucial factor in mate selection research (for online dating, see Hitsch et al., 2009, in press; Lee, 2008). As a proxy variable (see Tovee et al., 1998), we calculated the body mass index (BMI) from users' weight and height information and classified users according to the recommendation of the World Health Organization (WHO) into eight discrete body types: severe, moderate, and mild underweight; normal weight; overweight; and three degrees of obesity.9 We interpreted deviations from normal weight as an indicator of being less attractive.10 Two users sharing the same class of body type were classified as having similar physical attractiveness. Finally, we controlled for body height (in cm) in an attempt to correct the BMI to make it an even more meaningful proxy of physical attractiveness. Two users were classified as having height equality when their difference in height did not exceed 2 cm. Empirical Results Effect of Attribute Similarity on Contacting and Replying We first estimated the probability of sending a first contact message. Note that the probability of user A contacting user B always depended on A having visited B's profile beforehand. Because the data were hierarchical, that is, browsing events were nested in initiating users,11 and observations correlated significantly within users, we estimated multilevel models. The dependent variable was binary, taking the value one if the browsed profile was contacted and otherwise zero. Explanatory variables were the relation between initiator (i) and receiver (r) in terms of educational level, controlling for age and physical appearance (xir). Moreover, on the level of the choosing individual, we controlled for fixed effects on contacting probability (fli) and accounted for interindividual heterogeneity by introducing a subject-specific random effect (h;). This split the total variance of the model into a residual variance term and an intercept variance term (see, e.g. Rabe-Hesketh and Skrondal, 2005). Assuming logisti-cally distributed error terms (eir), this resulted in the following model logit {Pr(y;r = 1 \Xjr, fl;, M;)} = a + Xyfi + a[y + M; + Eir- We estimated the probability of replying to a given contact analogously by simply reversing sender and receiver. In that case, observations represented first contact events. Table 2 presents the results of the probability estimation for initial contacts and their replies. Models la and 2a show the effects of attribute similarity between users on the probability of contacting. When browsing profiles, both male and female initiators contacted other people with a higher probability when these people were similar in terms of educational level. For instance, the odds of a male user contacting a woman increased by a factor of about 1.1 in the case of educational similarity (Model la); for a female user, there was a factor change in the odds of about 1.3 (Model 2a). This supported our hypothesis that educational homophily is basically symmetrical across gender. Models lb and 2b report the effects of educational dissimilarity on the probability of an initial contact when browsing a profile. For example, when holding everything else constant, the coefficient of 'Educational level: ri -0.01 0.02 -0.36*** -0.24*** Age: ri —0.89*** -0.16*** -0.12* -0.24*** Phys. attractiveness: ri 0.20*** -0.06 -0.27*** -0.22*** Height: ri -0.66*** 0.40*** 0.42** -0.33* Attributes of individual3 Educational level -0.06* -0.04 —0.19*** -0.05 0.04 0.14*** 0.08* 0.12*** Age -0.03** -0.06*** -0.10*** -0.10*** 0.04 0.02 0.06*** 0.06*** Age2 0.00 0.00** 0.00*** 0.00*** -0.00 -0.00 -0.00 -0.00 Phys. attractiveness 0.01 0.24*** -0.10 -0.04 -0.02 0.17* -0.15*** 0.04 Height -0.01 -0.01** 0.02** 0.02*** 0.01 0.00 -0.00 -0.00 Intercept -0.69 -0.120 -2.47* -3.75** -3.20** -3.22** -3.15*** -4.63*** Random effects variance: In al 0.44*** 0.47*** 0.90*** 0.91*** -0.04 -0.02 0.49*** 0.51*** Log-likelihood -41,728 -41,126 -16,792 -16,625 -5,834 -5,796 -15,679 -15,642 Intraclass correlation (p) 0.32 0.33 0.43 0.43 0.23 0.23 0.33 0.34 Observations'1 133,947 133,947 64,845 64,845 10,922 10,922 39,552 39,552 Individuals 5,239 5,239 2,421 2,421 4,592 4,592 6,542 6,542 Source: Database dump of a German dating site, first half-year of 2007; own calculations, logit coefficients and levels of significance are reported. aIn Models la-2b, individuals are initiators; in Models 3a-4b, receivers of first contacts. Coding: 1-5 according to WHO weight classes, with 5 being normal weight and 1 being third-grade obesity as the largest deviation from normal weight. c Models la—2b: For users revealing more than 50 browsing events, we took a random sample of 50 events. Models 3a-4b: In order to avoid an inflation of the receiver sample, we excluded first contacts from the 1 percent of initiators making the most contacts for each sex, declaring them to be 'mass' senders. Significance: *P<0.05, **P<0.01, ***P<0.001. Z\0Z '61 -raqtuaictag uo saiprqg jbioos jo jooipg - äitsj3atuq 3[£res^p\[ ye /Sjo-s|^ujnofpjojxo-js3//:diiij uioxj paprajiiMOQ 188 I škopek, schulz and blossfeld browsed user had a lower educational attainment level, the results revealed a negative effect on contacting probability for both men and women. Thus, compared to educational similarity, both male and female users avoided contacting people with lower educational attainment levels. Note that selecting a partner with higher education did not differ significantly from similarity constellations, showing that users had a strong preference for a partner with the same or higher level of education. Given the rather traditional male breadwinner family model in Germany, we had not expected this result for men, although still for women. Homophily also seemed to be a dominant mechanism for replies to initial contacts. Models 3a and 4a show that similarity in educational level significantly increased the logits for a reply. Thus, receivers replied to users' first contacts more often when they resembled them in terms of educational level. In sum, we can conclude that the probability of getting in contact was higher for educational similarity than for educational dissimilarity. This mechanism even seemed to be reinforced when it came to replies. The effects of dissimilarity revealed analogous results to our similarity models. Contacts stemming from people with a lower educational level had a significantly lower probability of receiving a reply. Surprisingly, there was a positive effect for male receivers who were contacted by better educated women, and even a small negative (albeit not significant) effect for female receivers. Theoretically, we would have expected the effects (if there were any at all) to take the opposite direction. However, it is important to see that 'Educational level: ri: 28.8 per cent). This educational constellation had the lowest proportion of females contacting males (18.7 per cent compared to r = i: 38.2 per cent and r>i: 43.1 per cent). One explanation might be that contacts with better educated women represented particularly valuable opportunities for these men. Men in online dating seemed to have fewer reservations than women regarding all couple constellations in which women were better educated than men. This puzzle could, nonetheless, be a good starting point for further research. Turning to the control variables, our models showed that similarity in age and physical attractiveness significantly increased the probability of both contacting and replying. There was also evidence for the importance of age homophily here (see the negative effects for dissimilarity). For physical attractiveness, our models indicated that contacting and replying to somebody who was less physically attractive was less likely than contacting and replying to somebody who was similar. In addition, more attractive users were contacted or replied to with a higher probability than similar ones. Thus, at least for physical attractiveness, the principle of 'the more the better' seemed to apply. With regard to height, we found a quite gender-specific choice mechanism, with women (men) more likely to contact or reply when the man (women) was taller (shorter).12 Effects on Educational Homophily Based on the models in Table 2, we analysed how far attribute constellations facilitated first contacts or replies. This revealed that attribute similarity, especially in terms of education, favoured (mutual) contacting. As homophily seemed to be the dominant mechanism, the second step in our analysis was to change our perspective and explain homophily in contact relationships directly by estimating the probability of an attribute constellation (here: educational similarity). In more technical terms, we now took the constellation as the dependent variable to be explained by covariates. This tested the hypothesis that homophily would vary with the level of own resources. A major advantage of our research setting was that we could analyse choices while controlling for opportunity structure in the online dating environment. Thus, we could disentangle the effects on homophily due to distributional chances from those that were based on the preferences of the actor. To analyse educational homophily in online dating, we needed to compare actual choices with the choices to be anticipated theoretically under conditions of statistical independence (see Verbrugge, 1977). In other words, we drew on a statistical reference model providing information on the probability of contacting somebody from the other sex category with a certain educational level if choosing were to take place randomly (see, for similar approaches, e.g. Blossfeld and Timm, 2003; Fiore and Donath, 2005). The reference model is reported in Table 3. If theoretically expected and empirically observed choices were to prove congruent, we would have to assume that educational matching is based mainly on the distribution of education in the population of users. The more empirical observations differ from the theoretical expectations of the statistical independence model, the stronger is the empirical evidence for intentionally choosing individuals. Since users differed strongly in the number of initiated or replied contacts, and as these events educational homophily in online mate selection | 189 might not be distributed identically and independently within users, we weighted the single events by all events of each user. Otherwise, the analysis of homophily might be biased through highly active users who communicated a lot. This problem is quite common in studies of sociometric choice (see Signorile and O'Shea 1965). Events nested in users were weighted by taking the inverse value of the total number of events per user; all weights therefore summed to one for a given user. In this vein, we equated users who generated more information by being more active senders with users who might have sent only one contact and therefore left behind only sparse information. Technically speaking, we assigned low weights to the message events of the first group, and high weights to the latter. Thus, we connected Table 3 Distribution of educational levels in the active user population (column percentages) Educational class Men Women Total Basic secondary school 6.4 6.1 6.3 Vocational secondary school 34.6 40.2 36.9 University entrance 23.8 25.5 24.5 University degree 35.2 28.2 32.4 Total 100.0 100.0 100.0 N 6,199 4,249 10,448 Source: Database dump of a German dating site, first half-year of 2007; own calculations. the interpretation intuitively to the average user, however, averaging out all variation created within individuals. Navigating our way between the Scylla of overrepresenting idiosyncrasies and the Charybdis of potentially losing valuable within variance, we decided that, as we were analysing intentions in terms of educational homophily, an interpretation based on users was preferable to one based on message events.13 For descriptive purposes, we first calculated the mean proportion of contact relations per user by educational constellation. Table 4 shows the results for the initiator by gender as well as by the observed proportion of contacts and the proportion expected by the independence model. We also calculated a factor expressing the amount of 'overchoosing' when it exceeded one and the amount of 'underchoosing' when it fell below one. Factor values close to one suggest a random choice behaviour with regard to education. On average, more than one third (35 per cent) of an average user's contacts fell into the same educational class, and about the same proportion (36.7 per cent) was found for reciprocal contacts. We observed a higher overall homophily factor for women. When looking at contacts characterized by dissimilarity in education, women showed strong underchoosing of educationally downward contacts (18.5 per cent for first contacts) and marginal overchoosing of upward contacts (40.7 per cent). Men, in contrast, though behaving by and large as expected, slightly Table 4 Observed and expected proportion (in percent) of educational patterns in first and reciprocal contacts averaged by user Male initiators Female initiators and female receivers and male receivers Educational First Reciprocal First Reciprocal constellation contacts contacts contacts contacts i = r. 'Similarity' Observed Expected3 Factor i>r. 'Downwards' Observed Expected3 Factor i