Is (Actual or Perceptual) Personality Similarity Associated With Attraction in Initial Romantic Encounters? A Dyadic Response Surface Analysis

Is (Actual and/or Perceptual) Similarity Appealing?

The attraction literature has long sought to establish similarity as a major attraction principle. In his seminal work on the “similarity-attracts” hypothesis, Byrne (1971) suggested that similarity might be appealing due to its adaptive functions, such as facilitating dyadic interaction, stimulating feelings of familiarity or safety, and validating people’s self-concepts. Sharing personality traits implies similarity in affective needs and ways of expressing affect, thereby easing communication (Izard, 1960; Prager, 1995). The similarity-attracts hypothesis thus implies that personality similarity should be relevant both (a) in the initial screening phase of partner selection and (b) in existing romantic relationships (dating phase or long-term relationships). In this article, we focus on the very early stages on getting-to-know each other, but note that the empirical evidence concerning similarity effects in existing relationships also remains equivocal.³

Empirical research on effects of actual similarity on initial romantic attraction is yet inconclusive. Positive effects of personality similarity have typically been found in studies involving hypothetical partners, akin to the bogus stranger paradigm (Byrne, 1961). When people are given explicit information about hypothetical romantic partners (e.g., in the form of written vignettes or bogus questionnaire results), they tend to prefer those who resemble their (self-concept about their) own personality (Klohnen & Luo, 2003; Montoya et al., 2008). By contrast, studies employing more naturalistic designs (e.g., speed-dating studies with real first-time encounters with potential romantic partners) have reported few and inconsistent effects of actual personality similarity on initial attraction (Asendorpf et al., 2011; Luo & Zhang, 2009; Tidwell et al., 2013).

Actual similarity refers to a comparison of two people’s personality trait levels, which implies that it can influence the persons’ attraction to each other only through the indirect path of behavioral expression and perception processes (Back, 2021). By contrast, perceptual similarity—the similarity of one person’s (self-concept about his/her) personality with his/her perception of the other person’s personality— is closer to an individual’s cognitions when evaluating the respective other person in terms of attraction. Therefore, perceptual similarity might act as a more proximate predictor of initial attraction than actual similarity (Klohnen & Mendelsohn, 1998; Selfhout et al., 2009). Empirical findings from the few studies that involved at least some amount of dyadic interaction indeed indicated mostly positive effects of perceptual personality similarity on romantic attraction and only a few effects of actual similarity (Montoya et al., 2008; Tidwell et al., 2013).

Statistical Challenges in Investigating Similarity Effects

Similarity effects in initial encounters are usually tested by computing absolute (or squared) difference scores between the dyad members’ trait levels and using these scores to predict romantic attraction in a (dyadic) regression model (e.g., Asendorpf et al., 2011). The main effects of trait variables are often included as predictors in the model, which avoids the risk of confounding linear main effects of individuals’ trait levels with similarity effects (Luo & Zhang, 2009). For example, if there is a linear actor effect in the sense that a person’s own trait level (e.g., his/her agreeableness) predicts how attracted he/she tends to be to potential mates, this effect can be detected only when the person’s trait level is included as a predictor that is separate from the difference score variable. If an actor effect is present in the data but not explicitly modeled, it can lead to spurious evidence of an effect of personality similarity, even if the other person’s trait level or the similarity between the two people is in fact irrelevant for attraction. An analogous situation occurs for partner effects (i.e., when the partner’s trait level is related to the actor’s feelings of attraction).

Using the difference score approach while controlling for linear actor and partner effects, however, is still systematically biased toward false-positive conclusions about similarity effects. The reason is that this approach falsely indicates evidence of a similarity effect when there are actor or partner effects that are curvilinear instead of linear or when attraction is maximized for a certain, nonzero discrepancy between actor and partner personality (Edwards, 2001).⁴ To overcome these limitations, we applied Dyadic Response Surface Analysis (DRSA; Nestler et al., 2015; Schönbrodt et al., 2018), a combination of the Actor Partner Interdependence Model (APIM; Kenny et al., 2006) with Response Surface Analysis (RSA; Edwards, 2002). (D)RSA is increasingly recommended and used in romantic relationship research (e.g., Eastwick, Finkel, & Simpson, 2019; Park et al., 2023; Weidmann et al., 2017) because it avoids the use of difference scores and can thereby detect effects of dyadic similarity and differentiate them from linear and curvilinear actor/partner main effects.

The Present Study

We aimed at a comprehensive investigation of whether personality similarity matters for initial romantic attraction. Our data stemmed from a real-life study in which 940 real speed-dates between 397 participants took place, capturing the initial screening stage of the partner selection process. We considered actual similarity and perceptual similarity with regard to the Big Five traits. For each of the 10 combinations of similarity perspective (actual, perceptual) and personality trait (Big Five), we used DRSA to test whether (actual/perceptual) personality similarity was related to romantic attraction in first encounters. We tested the robustness of our findings across different instruments to measure the Big Five (BFI vs. adjective list scales) and across outcome measures (romantic attraction vs. date quality) and tested for gender differences in the detected effects.

Transparency, Openness, and Reproducibility

A full codebook of the “Date me for Science” study can be found at https://osf.io/n7dw9/. The codebook contains detailed descriptions of the procedures and all applied instruments, and thereby allows independent replication of the study. The processed data and the R code that can be used to reproduce our results are provided in the folder “data_and_code” in the Supplementary Materials. The “Date me for Science” data was used in other studies (enlisted at https://osf.io/hmr6j/). Whereas some variables used here were also used in those previous analyses, the research questions do not overlap; the present paper presents the first analysis of similarity effects with these data.

We preregistered the study background, our expectations, and the analytical procedures and interpretation rules, after the data was assessed but before it was analyzed (see Supplementary Materials). The test for similarity effects was exploratory. If similarity effects occurred, we expected perceptual similarity to have stronger effects than actual similarity. In analyses that showed no evidence of similarity effects, we explored linear actor and partner main effects. We expected rather weak and nonsignificant actor and partner main effects on romantic attraction; if anything, on the basis of previous research, we expected positive partner effects of males’ self-reported extraversion and openness on female attraction (Asendorpf et al., 2011; Back, Penke, et al., 2011). All analyses were conducted as described in the preregistration; we only corrected one mistake that concerned the equality constraints in additional analyses on gender differences (see the description in the Supplementary Materials for details), and we conducted one further robustness check that was not preregistered (see below).

Participants

We recruited participants who were between the ages of 18 and 28, heterosexual, single, and currently looking for a romantic partner. We obtained 400 participants within our time frame for data collection (April–December 2015). Three participants objected to the use of their data afterwards, so the final sample consisted of 397 participants (200 women; M_age = 22.87, SD_age = 2.62) who formed 940 speed-dating dyads. No other data was excluded from the analyses. Most participants were students (91%) from a variety of different majors; the remaining participants were about to acquire their university entrance diploma. No compensation was provided apart from the opportunity to meet a romantic partner.

Procedures

Participation in the study began with a pre-event online questionnaire on demographics and self-ratings. Participants were then invited to one of 42 evening speed-dating events in the laboratory. For each event, we invited five women and five men. Most speed-dating events involved 10 participants (26 events with 10 people, 11 with nine, 3 with eight, 2 with seven). The female and male participants were kept separate during the welcoming, instructional, and waiting periods so that no interaction with opposite-sex participants was possible before the speed-dates. Each participant dated each opposite-sex participant who was present at the event. One speed-date took place at a time. The respective female and male participants were led separately (the woman first) into the speed-dating room, where they were seated on a couch at a 90° angle. The speed-date took 3 min, during which the daters were alone in the room and free to choose what to say or do. The dates were videotaped. After the speed-date, the daters were separated and answered a post-date questionnaire. The “Date me for Science” study was approved by the local ethical committee.

Measures

Table 1 shows the applied instruments. Descriptive statistics, item-, and scale-intercorrelations for the measures can be found in the Supplement 3 (see Supplementary Materials).

Analytical Strategy

We investigated similarity effects in 2 (actual vs. perceptual similarity) x 5 (Big Five traits) = 10 separate analyses, using DRSA for dyads that are distinguishable by gender (Schönbrodt et al., 2018; see also Nestler et al. 2015). In analyses referring to actual similarity, we used the BFI-S measures of the respective trait. In analyses referring to perceptual similarity, we used the adjective-list assessments of the actor’s self-reported personality and the actor’s perception of the partner’s personality to achieve commensurability of the predictor variables as required for DRSA (Schönbrodt et al., 2018).

Analytical Procedure for Investigating Similarity Effects

Each of the 10 analyses comprised several steps; this strategy was inspired by the analytical procedure in Weidmann et al. (2017) but adapted to more recent recommendations on (dyadic) RSA (Humberg et al., 2019; Schönbrodt et al., 2018). As the first step, for each speed-date, the female participants’ rating Z_f of her attraction to her male dating partner and the man’s rating Z_m of the woman were predicted from their respective personality variables in the DRSA model (Schönbrodt et al., 2018):

$Z_f=\mathrm{ }{b}_{0f}+b_{1f}{X}_f+b_{2f}{Y}_f+b_{3f}{X}_f^2+b_{4f}{X}_f{Y}_f+b_{5f}{Y}_f^2+e_f$

1

$Z_m=\mathrm{ }{b}_{0m}+b_{1m}{X}_m+b_{2m}{Y}_m+b_{3m}{X}_m^2+b_{4m}{X}_m{Y}_m+b_{5m}{Y}_m^2+e_m$

$e_f~~e_m$

Here, X_f denotes the woman’s self-reported personality trait level, Y_f denotes the man’s trait level (self-reported for the analyses on actual similarity, perceived by the woman for analyses on perceptual similarity), X_m denotes the woman’s (self-reported/male-perceived) trait level, and Y_m denotes the man’s self-reported trait level (see Table 1 in the preregistration (see Supplementary Materials) for a systematic overview of the variables). The correlation between the error terms e_f and e_m was freely estimated to account for dependencies within the dyads. The DRSA model reflects how the female participant’s attraction ratings (Z_f) and the men’s ratings (Z_m) relate to all possible combinations of the two people’s personality (e.g., X_f and Y_f). Owing to the second-order terms (X² etc.), the DRSA model can represent curvilinear relationships that would be present, for example, in the case of a similarity effect (see Figure 1; see below for explanations of the figure).

Click to enlarge

Figure 1

Prototypical Similarity Effect

Note. Response surface indicating a similarity effect (with additional linear main effects of male and female personality, reflected in the rising nature of the ridge line). Surface of the estimated regression equation $Z_f= 1+0.1 X_f+0.1{ Y}_f-0.08 X_f^2+0.16{ X}_f{Y}_f-0.08{ Y}_f^2$

Second, we estimated a corresponding APIM model that included only the main effects of the two participants’ trait levels:

$Z_f=\mathrm{ }{b}_{0f}+b_{1f}{X}_f+b_{2f}{Y}_f+e_f$

2

$Z_m=\mathrm{ }{b}_{0m}+b_{1m}{X}_m+b_{2m}{Y}_m+e_m$

$e_f~~e_m$

The APIM model is nested in the DRSA model but includes only linear terms, so that it represents only linear, no curvilinear, actor effects (b_1f and b_2m) and partner effects (b_2f and b_1m).

Third, we compared the fit of the DRSA model with the fit of the APIM with a chi-square likelihood ratio test. If the DRSA model did not fit the data significantly better than the APIM, the similarity-attracts hypothesis was rejected. This is admissible because in such a situation, curvilinear effects (which would have to be present for a similarity effect) do not explain a meaningful amount of variance beyond linear actor and partner effects.

Fourth, if the DRSA model had a significantly better fit than the APIM, we used the coefficient estimates from the DRSA model to calculate auxiliary parameters a_1f to a_5f for the female part of the model and a_1m to a_5m for the male part (Edwards, 2002; Schönbrodt et al., 2018). Evidence of a similarity effect for predicting female attraction would be found if a_4f was significantly negative, a_3f was nonsignificant, and a_5f was nonsignificant; and for male attraction accordingly (Schönbrodt et al., 2018). We imposed no conditions on a₁ and a₂ to allow for the possibility that women’s and men’s personality trait levels had (linear or curvilinear) main effects on attraction in addition to a similarity effect (e.g., Edwards, 2002; Humberg et al., 2019).

Figure 1 shows the graph (the response surface) of the female part of an estimated DRSA model. The predictions of this prototypical model are in line with a similarity effect, as the highest female attraction is predicted for male-female personality combinations that lie on the blue line, which consists of dyads with congruent actor versus partner personality. The auxiliary parameters a_4f, a_3f, and a_5f are statistical descriptions of properties of the graph. They satisfy the three conditions mentioned above if and only if the corresponding graph is shaped roughly like the surface in Figure 1; that is, if the model predictions are in line with a similarity effect (for detailed explanations, see Schönbrodt et al., 2018; see also Edwards, 2002; Humberg et al., 2019). More specifically, the parameters a₄ and a₃ reflect the shape of the response surface along the line of incongruence (red line in Figure 1; formally defined as Y = -X), which describes the predicted attraction rating for dating partners whose personality scores are equal in magnitude but opposite in sign; in particular, this line ranges from dyads where the man’s personality score is much higher than the woman’s (left corner of the coordinate cube) to dyads with similar personalities, to dyads where the woman’s score is much higher than the male’s (right corner). This line can be expressed as the quadratic equation $Z_f= b_{0f}+a_{3f}{X}_f+a_{4f}{X}_f^2$. Therefore, a negative a₄ and a non-significant a₃ imply that the line of incongruence has an inverted U-shape with the highest attraction predicted for dyads with the congruent personality constellation (0,0). When this is the case and when the parameter a₅ (which has no direct graphical representation) is non-significant, the ridge of the surface matches the blue line of congruence, so that highest attraction is predicted for congruent dyads for all possible personality constellations.

Effects of Similarity on Attraction

Table 2 shows the results of the model comparison in the first analytical step. In the analyses on effects of actual similarity, the DRSA model did not fit the data significantly better than the APIM (at α = .01) for any of the Big Five traits. The similarity-attracts hypothesis had to be rejected in all cases; actual similarity did not seem to be related to attraction.

Concerning the analyses for perceptual similarity, the model comparison was again nonsignificant for the traits neuroticism, openness, agreeableness, and conscientiousness, so the similarity-attracts hypothesis was rejected for these traits. For extraversion, the DRSA model had a significantly better fit than the APIM. However, neither the women’s nor the men’s auxiliary RSA parameters satisfied the conditions for a similarity effect because, for example, the a₄ parameter was not significantly negative (female attraction: a_4f = .06, 99% CI [-.08, .21]; male attraction: a_4m = -.02, 99% CI [-.11, .07]; see Supplement 7 for all results). That is, there was no evidence of an association between perceptual similarity and attraction.

In sum, none of the 10 analyses supported the idea that actual or perceptual personality similarity between a male and female dating partner was related to their attraction to each other. We further inspected the results in an exploratory way and revealed a nuanced pattern of actor and partner effects.

Exploratory Inspection of Main Effects of Self-Reported Personality

Concerning actor and partner effects of self-reported personality, the coefficients from the APIM models (which were sufficient for describing the data, see Table 2) are provided in the first two rows of Table 3.

For neuroticism, there was a positive actor effect and a negative partner effect on female attraction: Women felt more attracted to men the higher their own level of neuroticism and the lower the men’s level of neuroticism. There were significant partner effects of extraversion for both genders: Men and women preferred more extraverted partners. Concerning openness, there was a positive actor effect for women, indicating that women with higher levels of openness felt more attracted to potential partners. Results also postulated a positive actor effect of men’s agreeableness, indicating that more agreeable men were more attracted to their dating partners. Finally, women considered men to be more attractive the more conscientious the men were (i.e., a positive partner effect).

Exploratory Inspection of Main Effects of Perceived Personality

Concerning perceived personality, the partner effects for neuroticism, openness, agreeableness, and conscientiousness estimated in the APIM are displayed in the last row of Table 3. The coefficients for self-reported personality are not shown because the association that they reflect (effect of actors’ self-reported personality while controlling for their perceptions of the partner) is not meaningful in this context (see Supplement 8 for all coefficients). The partner effects for all four traits were strong and significant. Women and men preferred romantic partners whom they perceived to be less neurotic, more open, more agreeable, and more conscientious.

For extraversion, the DRSA model had a significantly better fit than the APIM. For female attraction, the model coefficients revealed a monotonously positive but nonlinear partner effect of perceived male extraversion (indicated by b_2f = .72, 99% CI [.63, .81] and b_5f = .08 [.03, .13], whereas the other coefficients were nonsignificant: b_1f = -.07 [-.16, .03], b_3f = -.03 [-.10, .05], b_4f = 0 [-.09, .07]; see Supplement 7 for all results). Women were more attracted to men the more extraverted they perceived the men to be, and this effect was stronger for higher levels of perceived male extraversion than for lower levels (see also Figure 2, left). An analogous pattern was found for men’s attraction to women (Figure 2, right): Men were more attracted to women the more extraverted they perceived the women to be, and this positive association was stronger for higher extraversion levels than for lower levels (indicated by b_1m = .63 [.56, .69] and b_3m = .05 [.01, .10], whereby b_2m = 0 [-.07, .08], b_5m = -.05 [-.11, .00], b_4m = .02 [-.04, .08]).

Click to enlarge

Figure 2

Response Surfaces of the DRSA Model for Self-Reported and Perceived Extraversion

Note. The figure shows the response surfaces for female attraction (left) and male attraction (right) predicted by the actor’s self-reported extraversion level and his/her perception of the dating partner’s extraversion level. The points on the surfaces represent the model-predicted attraction values for the predictor combinations in the data. The lines on the surface represent the bagplot of the data, where the outer black line demarcates the region of the surface that can be interpreted.

Robustness Analyses

Consistent with the main analyses, the robustness analyses with different choices of the self-report measures (adjective list in the analyses on actual similarity; BFI-S in the analyses on perceptual similarity), of the adjective list scales, and/or date quality as the outcome variable yielded no evidence of similarity effects (see Supplements 6 to 8 and Supplement 10 for all results). In addition, the patterns of actor and partner effects reported above were largely replicated (see Supplement 5 for a summary of results and more information). The only substantial exceptions implied that (a) the positive actor effect of women’s self-reported neuroticism should be considered tentative (as it was replicated in only one analysis) and (b) the two-item openness scale seemed to yield systematically different effects than the established BFI scale (see Supplement 5 for details). The latter observation concerns only the partner effects of perceived openness (the only result that was based on the two-item measure). When we repeated the analyses with an alternative scale for perceived openness, this partner effect replicated (see Supplement 10), but the effect nonetheless awaits replication with a more established instrument.

Gender Differences

Table 3 shows that the analyses involving only self-reported personality revealed a differentiated pattern of actor and partner effects, where some effects seemed to differ between genders and others did not. A formal test of gender differences (see Supplement 9 for all results) supported the respective pattern for neuroticism (significant difference between genders, χ²(2, N = 940) = 20.19, p < .001; i.e., more positive actor effects for women than for men and more negative partner effects of men’s neuroticism than of women’s), for extraversion (no significant difference, χ²(2, N = 940) = 2.92, p = .232), and tentatively for openness (marginally significant difference, χ²(2, N = 940) = 7.86, p = .020; i.e., more positive actor effects for women). The main analyses also suggested gender-specific effects for agreeableness (stronger actor effects for men) and conscientiousness (stronger partner effects for men). The differences between the equality-constrained versus the gender-specific APIMs were, however, nonsignificant (agreeableness: χ²(2, N = 940) = 1.00, p = .608; conscientiousness: χ²(2, N = 940) = 1.84, p = .399). The respective equality-constrained APIMs showed a positive actor effect of agreeableness (b = .08, 99% CI [.02, .14]) and a positive partner effect of conscientiousness (b = .09, 99% CI [.02, .15]). This pattern was replicated in most robustness analyses and tentatively indicates that the two effects might also occur for women, a conjecture that awaits future replication. For the analyses involving perceived personality, none of the detected effects differed significantly between genders (see Supplement 9).

No Evidence of Similarity Effects on Initial Romantic Attraction

The lack of evidence of similarity effects was consistent across all Big Five traits, both perspectives on similarity (actual vs. perceptual), and all robustness analyses. Along with the observation that earlier studies found only a little and inconsistent evidence of similarity effects (though the applied statistical approaches were even biased toward false-positive results), we consider it increasingly unlikely that the similarity-attracts hypothesis is empirically supported in the personality domain. Our findings could also help to explain why—with regard to existing relationships—romantic partners do not seem to be more similar in their personalities, whereas they notably resemble each other on a number of other characteristics, such as attitudes, religiosity, and physical attractiveness (Luo, 2017). Moreover, our findings are consistent with recent evidence that, although the largest share of variance in romantic attraction is unique to the specific dyad in question, it is difficult to predict this unique attraction from distal person variables, such as personality traits (Joel et al., 2017). Consequently, understanding how people decide who is a “good match” for them, and understanding the role that similarity plays in this context, may require higher resolution investigations (Joel et al., 2017). Specifically, researchers might focus on (the degree of congruence between) dating partners’ expressed behaviors (e.g., how dominantly someone acts on a date) and actors’ affective preferences for respective behaviors (e.g., the extent to which dominant behavior evokes positive affect in the dating partner; see also Back, 2021; Back, Schmukle, et al., 2011).

Actor and Partner Effects of Actual and Perceived Personality

Considering actual (self-reported) personality, we found that women preferred men who were less neurotic, more extraverted and more conscientious, and that men preferred more extraverted women. These partner effects are largely consistent with effects attained in prior studies (Asendorpf et al., 2011; Humbad, 2012; Jauk et al., 2016; Luo & Zhang, 2009). The actor effects (more open women and more agreeable men felt more attracted to potential partners) that we found in our main and robustness analyses were of small sizes and have not been frequently reported in previous studies on initial attraction. They thus await replication.

Personality perceptions were consistently related to attraction. Men and women were more attracted to dating partners the less neurotic and the more extraverted, open, agreeable, and conscientious they perceived the partners to be. In other words, people were more attracted to partners whom they perceived to have more normative (e.g., Furr, 2008; Wood & Wortman, 2012), healthier (e.g., Bleidorn et al., 2020), and more mature personality trait levels (e.g., Roberts & Wood, 2006). Future research should explore whether this pattern results from people preferring partners whose personality profiles promise high psychological adjustment (Ozer & Benet-Martínez, 2006), from people generalizing their attraction to someone to favorable evaluations of that person’s personality (Thorndike, 1920), and/or from individual differences in the tendency to evaluate others more or less positively (i.e., perceiver effects; Rau et al., 2021).

Limitations and Future Directions

The present study was the first to test for similarity effects in initial encounters using DRSA. Despite the relatively high power for detecting similarity effects, our results await direct replications in further large-scale and similarly naturalistic studies (e.g., via multilaboratory collaborations).

One limitation of our study concerns the Big Five scales for perceived personality that were constructed from the adjective list items. While it is standard to collect personality perceptions with adjective lists in speed dating (or round robin) designs to achieve a rapid and broad coverage of interpersonal judgments, the items cannot represent the Big Five as comprehensively as specialized Big Five instruments. The robustness analyses with different construction strategies and accordingly different scales indicate that the relative narrowness of the scales and their partial overlap did not essentially influence the results (see Supplement 10). Nevertheless, future studies should aim at assessments of personality perceptions and respective self-reports that, besides the necessary briefness, capture the Big Five more broadly.

The participants of our study were young adults (18–28) who were mostly students and who were heterosexual. Whereas we expect our results to generalize to other age groups and people without an academic background, we do not necessarily expect generalization to homosexual individuals. Furthermore, the present research question and study procedure referred to the initial screening phase of two unacquainted individuals; a direct replication of our study would therefore use a similarly short time frame (e.g., 3 minutes) that the two dating partners have to get an impression of the other person and form an initial attraction evaluation. Future research might explore whether the present results generalize across design variants, for example, across speed-dating designs with and without videotaping of the dates. However, we expect our findings to not be specific to an “organized” encounter in a speed-dating event, but to generalize to real-life initial encounters (e.g., in a pub). We have no reason to believe that the results depend on other characteristics of the participants, materials, or context.

Finally, future studies could extend our perspective by (a) zooming in on the dynamic processes (e.g., the interplay of expressed behaviors, perceptions, cognitions) involved in real-life first encounters and (b) taking a closer look at alternative forms of the initial screening phase (e.g., online dating; e.g., Bruch & Newman, 2018) and subsequent steps of the dating phase (e.g., relationship formation and development; Campbell & Stanton, 2014). We are convinced that these avenues will lead to a deeper understanding of how people make one of the most relevant decisions in their lives—who is a “good match” for them.