Assortative Mating Patterns Based on Longitudinal Trajectories of Personality Traits and Relationship-Specific Characteristics

Contrary to the popular belief that “opposites attract,” extensive research has demonstrated that “birds of a feather flock together.” Research shows a matching pattern in romantic relationships for many socio-demographic and cognitive-behavioral variables, such that people tend to initiate and maintain relationships with partners who are similar to them. These effects range in size from strong for sociopolitical attitudes, educational attainment, and cognitive abilities to weak for Big Five traits such as Neuroticism, Agreeableness, and Extraversion (Horwitz et al., 2023). Although a range of constructs have been investigated, research on assortative mating can benefit from including more fine-grained constructs specifically related to relationship contexts and from examining longitudinal patterns of trait change over time. Thus, the current research used cross-sectional and longitudinal data from two separate samples of romantic couples at different stages of their relationship. In addition to the Big Five traits, we examined relationship-specific characteristics to increase our understanding of the function of psychological individual differences in the relationship context.

Assortative mating is “the tendency for two partners to be matched systematically on one or more characteristics” (Luo, 2017, p. 1). It has been investigated in two main ways: similarity and complementarity. However, the preponderance of current evidence primarily suggests a similarity effect (for a review, see Epstein & Guttman, 1984; Luo, 2017). As a result, the current research considers assortative mating to be a systematic matching pattern such that romantic partners are more similar to one another than would be expected by chance. Although the notion of non-random mate selection is well-known among both lay people and researchers across disciplines, debate remains regarding what specific selection strategies and which characteristics are considered (either implicitly or explicitly) and the potential consequences of selection for long-term relationship quality. The current research focused on intimate romantic relationships, while acknowledging that a rich body of research exists for matching in other social structures, such as platonic friends and social networks (e.g., Back et al., 2023; Harris & Vazire, 2016).

Prior research on assortative mating has proposed a hierarchical model to organize different variables that may be influential during mate selection (Luo, 2017; for older reviews see: Epstein & Guttman, 1984; Vandenberg, 1972). The most influential variables tend to be demographic variables and sociopolitical attitudes, with between-partner correlations upward of .60s for education and .70s for social attitudes. The second group includes values, intelligence, interests, and mental well-being, with correlations upward of .40s to .50s. The third group includes physical characteristics and various personality traits, with estimated correlations for Big Five traits rarely being above .30 in most studies (e.g., Botwin et al., 1997; Luo, 2009; Watson et al., 2004).

A recent meta-analysis of 22 complex traits across 30 studies and approximately 23,000 spousal pairs revealed that social attitudes, substance use behaviors, and cognitive abilities such as intelligence had the highest between-partner correlations, whereas personality traits assessed via surveys had much weaker although still positive and significant correlations: Extraversion = .08, Neuroticism = .10, Agreeableness = .11, Conscientiousness = .16, Openness = .21 (Horwitz et al., 2023). The fact that Openness showed the strongest effect among the Big Five is consistent with its strong association with intelligence, education, and sociopolitical attitudes (Kaufman et al., 2016; O’Connell & Marks, 2022).

In addition to the existence of assortative mating, prior research is also inconclusive with respect to whether assortative mating based on personality traits is beneficial to relationship quality. Most evidence has documented actor or partner effects rather than dyadic effects with respect to personality traits; that is, characteristics of individuals are generally more predictive of relationship quality than are the interactions of both partners’ characteristics. A comprehensive examination of large samples (> 20,000) across Australia, Germany, and the United Kingdom showed that actor effects of one’s personality traits on one’s own relationship satisfaction tends to be strongest, accounting for 6% of variance, with smaller but still significant partner effects of 1-3% of variance and no significant effects of between-partner similarity on either life or relationship satisfaction (Dyrenforth et al., 2010). This general pattern in which similarity effects on relationship quality are either null or much weaker than actor effects was also found in smaller studies of dating and married couples (e.g., Arránz Becker, 2013; Blum & Mehrabian, 1999; Gattis et al., 2004; Luo, 2009; Robins et al., 2000; Watson et al., 2004).

Some researchers have argued that the lack of existing evidence for the benefits of similarity in personality traits might be due to insufficient analytic methods. Instead of using simple between-partner correlations, Luo and Klohnen (2005) showed that profile similarity significantly predicted marital quality, even controlling for individual-level ratings of participants’ traits. However, Joel and colleagues (2020) applied machine learning techniques and expanded the scope beyond Big Five traits across 43 longitudinal studies to demonstrate that the most influential variables for relationship quality tended to be actor-reported variables rather than moderation effects or partner-reports.

In sum, assortative mating is a strong and consistent phenomenon in various domains, particularly with respect to demographic variables and sociopolitical and spiritual attitudes. Not only are couples generally much more similar than non-couples, but similarity in these domains is an important predictor of relationship quality. On the other hand, evidence for both the existence and benefits of assortative mating is much more limited in personality research. In particular, broad personality traits such as the Big Five show limited between-partner similarity, and most evidence favors either actor or partner effects over dyadic effects for relationship quality. Further, current evidence suggests that perceived similarity between partners, more so than actual similarity between partners, is a stronger indicator of assortative mating processes. However, it is important to emphasize that personality traits are not the whole of personality (DeYoung, 2015; McAdams & Pals, 2006) and that other more specific personality constructs, such as caregiving and support provision, attachment orientations, and interpersonal conflict resolution might contribute to assortative mating process.

Although personality is often treated in much of the literature as if it were synonymous with personality traits, traits alone are not sufficient to describe the whole person. Despite their utility in organizing broad cognitive and behavioral patterns and their predictive power in important life domains, personality traits largely remain “a psychology of the stranger” (McAdams, 1995). In other words, these broad descriptors are useful to differentiate people in a relatively general, decontextualized way, but they do not provide specific insights to understand individual people on an intimate level and in the context of their specific life situations. Although they have received less attention, personality constructs beyond traits are likely to be influential to relationship quality.

Characteristic adaptations, for example, are constructs at the second level of personality. They broadly include “motives, goals, plans, strivings, strategies, values, virtues, schemas, self-images” specified in relation to a person’s particular life context (McAdams & Pals, 2006, p. 208). Characteristic adaptations can be decomposed into goals, interpretations, and strategies, and they reflect the learned, habitual ways in which a person has adapted to their life experience (DeYoung, 2015). Unlike personality traits, characteristic adaptations can be highly specific to the situation in which they are conceptualized and, unlike traits, constructs at this level might not be applicable or measurable in all cultural or historical contexts. For example, being generally prevention-focused is a trait (which could be observed in any culture in human history) but checking the stove every time one leaves the house is a characteristic adaptation (which could be observed only in cultures with stoves and houses). In the interpersonal domain, consider three scale items: an Agreeableness item, I sympathizes with others’ feelings (DeYoung et al., 2007), an attachment Avoidance item, I feel comfortable sharing my private thoughts and feelings with partners (Brennan et al., 1998), and a Responsiveness item, My partner seems interested in what I am thinking and feeling (Reis, Reis et al., 2004). Although slightly different in content, the first item is the broadest, referring to a person’s tendency towards other people in general. The second item is then a bit more specific, referring to a type of target: romantic partners. Then, the last item is the most specific of the three, referring to one particular target: the current partner. It should not then come as a surprise that constructs at these different levels would have differential predictive powers and association patterns for relationship-specific outcomes.

The aim of the current research is not to delineate exactly which constructs should be categorized as a trait or a characteristic adaptation. Instead, we aim to bring further attention to personality constructs outside of the predominant Big Five trait taxonomy. As a result, our two samples included not only one-time measures of the Big Five traits, but also longitudinal measures of a variety of relationship-specific personality constructs that may be relevant to assortative mating patterns in romantic relationships, such as trust (Simpson, 2007), attachment orientations (Mikulincer & Shaver, 2017), perceptions of a partner’s responsiveness (Reis & Clark, 2013), caregiving tendencies (Kunce & Shaver, 1994), and conflict resolution strategies (Overall & McNulty, 2017). In the context of the hierarchical structure proposed by Luo (2017) and various previous reviews, because these constructs are much more contextualized than the global measures of Big Five traits, they would be more similar to the second category of habitual patterns and values. As a result, we would expect them to show higher level of assortment than the Big Five.

There has been some research on assortment based on these relationship-specific constructs. For instance, having a high level of trust in one’s partner is the foundation upon which a person’s standing on all these variables is often based. Individuals who have high trust in their partners—those who can always count and depend on their partners for help and support, especially in times of need—are more likely to develop secure attachments with their partner, perceive them as more responsive, receive (and provide) better forms of care, and engage in more constructive conflict resolution strategies (see Simpson, 2007). Attachment orientations (i.e., attachment anxiety and avoidance) have shown between-partner similarity in early dating couples (Luo, 2009), although a different sample of newlywed couples found significant correlations only for attachment avoidance (Watson, 2004). Further, partners who both score high on either anxiety or avoidance tend to report lower martial quality (Ben-Ari & Lavee, 2005). Additional research on conflict resolution in romantic relationships by Zeidner & Kloda (2013) found strong between-partner correlations for several conflict strategies, such as constructive resolution (r = .51), mutual avoidance (r = .29), and demand/withdrawal (r = .32), but similarity was not related to marital quality. Interestingly, research designs using both self-reports and partner-reports have shown that perceived similarity in conflict styles was stronger in magnitude than actual similarity between partners’ self-reports, and it was also more strongly associated with relationship quality (Acitelli et al., 1993).

In addition to expanding constructs beyond standard personality traits, we can also expand personality research within the trait level by including longitudinal repeated measures. Longitudinal patterns can themselves be considered important individual differences because not everyone develops in the same way across their lives or changes in similar ways in response to the same life events or transitions (Schwaba & Bleidorn, 2018). As a result, it is worthwhile to consider longitudinal patterns and changes in addition to one-time trait measures when investigating assortative mating processes. Despite the fact that personality traits are traditionally defined to be relatively stable across time and situations, meta-analytic research has shown that they can and do change (Bleidorn et al., 2022; Roberts et al., 2006) and several mechanisms of change have been both proposed and empirically supported (Donnellan et al., 2007; Roberts et al., 2008). Importantly, robust longitudinal designs allow researchers to test the convergence hypothesis, which posits that long-term partners become more similar to one another over time, rather than just being similar at the time of selection; however, longitudinal findings for this hypothesis have been mixed (Caspi & Herbener, 1993; Gonzaga et al., 2007, 2010; Lewis & Yoneda, 2021). In particular, a recent longitudinal study on 1,180 cohabiting German couples found that similarity was driven by mate selection rather than convergence over time (van Scheppingen et al., in press).

An influential force in personality change is role expectations and demands (Roberts et al., 2008). Personality traits have been shown to change with major life events and transitions, both in work and relationship domains (Bleidorn & Denissen, 2021). Specifically, social investment principle posits that positive personality trait change can result during the transition to parenthood, which tends to be a tumultuous and highly influential time for personal development due to dramatic shifts in social investment and social networks (Lodi-Smith & Roberts, 2007). Nonetheless, results of research on transition to parenthood are mixed: Some studies have found no evidence of trait change across the transition (van Scheppingen et al., 2016), whereas others have found negative changes, including an increase in Neuroticism and decreases in Agreeableness and Conscientiousness (Hutteman et al., 2014; Jokela et al., 2009; Specht et al., 2011). Notably, the timing of change appears to be an important factor. Many studies did not have pre-birth data and they varied in the time between trait measurement and childbirth. Thus, the current research examined a sample of first-time parents across the first two years of the transition into parenthood, which included five waves of data starting shortly before childbirth. This more expansive, multi-wave data collection allowed us to capture variation that may occur during this important life transition, both at the individual (partner) level and at the dyadic level.

Surprisingly, little dyadic longitudinal research has investigated personality change across important life transitions. A three-wave research design from pregnancy to one-year postpartum revealed that new parents tend to have more similar personality development trajectories compared to childless dyads both for Big Five traits (Galdiolo & Roskam, 2014) and attachment orientations (Galdiolo & Roskam, 2017). In a more normative context of personality change, another study involving almost 4,000 couples indicated that couples generally did not show correlated or persistent patterns of personality change (except for Openness), but they did show evidence of correlated fluctuations (variability) in Neuroticism, Extraversion, and Openness over time (Lewis & Yoneda, 2021). Fluctuations in this study were operationalized as occasion-specific residuals; that is, on occasions when one partner reported higher levels of these three personality traits relative to their own mean level, their partner also reported an increase in their own mean level on these same traits. In addition, another study of 237 couples documented evidence for longitudinal similarity effects on satisfaction over two years, with similar levels of Neuroticism being associated with male partners’ relationship satisfaction and similar levels of Openness being associated with female partners’ relationship satisfaction (Weidmann et al., 2017). Nonetheless, this gender-specific finding requires replication, and additional research is needed to investigate correlated change or similarity in dyadic change patterns, both during normative time periods and across major life transitions.

Besides true similarity between partners (correlations between their two self-reports), it is also important to consider perceived similarity (correlations between partner A’s own self-report and their informant-report of partner B’s personality), perceptual accuracy (correlations between partner A’s self-report and partner B’s informant-report of partner A’s personality), and simple partner effects or matters of preference (which trait levels people want from their partners). Research has revealed that people tend to have both absolute preferences for potential partners as well as relative preferences based on their own characteristics (Buss & Barnes, 1986). In addition, research has consistently shown that partners’ perceived similarity tends to be stronger in magnitude than their actual similarity (e.g., Montoya et al., 2008; Sillars, 1985). For example, in a meta-analysis involving 460 effect sizes, perceived similarity had stronger associations with interpersonal attraction than actual similarity did, with actual similarity showing an effect only in designs with no- or short-interactions (Montoya et al., 2008). This differential effect of perceived and actual similarity is also present in studies of platonic friendship formation (e.g., Selfhout et al., 2009) as well as initial attraction in speed-dating paradigms (e.g., Tidwell et al., 2013). With the advent of fast-paced online dating, survey data on evaluations of 7,846 online profiles found that people tend to prefer potential partners who they perceived to have a personality similar to themselves judging from just photos and a short description (Neyt et al., 2020). Moreover, in a study of dating and married couples, relationship intimacy increased with perceptual accuracy both for negative and positive self-views (Swann et al., 1994). In other words, for individuals who possessed negative self-views, intimacy was higher even when their spouse perceived them negatively, indicating the importance of authenticity and self-verification in close relationships.

With a relatively mixed body of literature, it is important to not only conceptualize assortative mating as a general matching pattern, but also to be mindful of the importance of each specific construct that we examine. For instance, why are there such small effect sizes for between-partner correlations on the Big Five traits although they are so predictive of many life outcomes, and why do some of the Big Five traits (e.g., Openness) show more similarity patterns whereas others do not? We argue that the lack of synthesis in the personality literature on this domain is due partly to the predominance of trait research. Instead, assortative mating research would benefit from exploring personality beyond the one-time Big Five trait measures and include both short- and long-term longitudinal processes as well as finer-grained and more relevant constructs to the relationship process.

The current research investigated the existence of potential matching patterns not only in the Big Five personality traits but also in other relationship-specific characteristics, including attachment orientations, caregiving patterns, conflict resolution strategies, partner responsiveness, and trust. We analyzed data from two existing longitudinal studies: one involving early dating couples of between 3 and 12 months, and another involving married or cohabiting couples during the first two years of the transition to parenthood. We examined broad patterns of similarity, both cross-sectionally (at baseline assessment) and longitudinally (overtime). In addition to the evidence of assortative mating, we also examined the potential benefits of this matching pattern for relationship quality. The hypotheses were organized into two broad research questions described below:

H1. At baseline (i.e., the first assessment), romantic partners will be similar in their personality, such that their scale scores on personality traits and characteristic adaptations will be significantly and positively correlated.

H2. At baseline, romantic partners will be more similar in their relationship-specific characteristics than in their personality traits. There are no hypothesized differences among either traits or relationship constructs.

H3. Longitudinally, romantic partners will show a similar change trajectory in self-reported personality across the first two years of parenthood, such that their slopes will be significantly and positively correlated.

H4. At baseline, perceived similarity in personality traits and characteristic adaptations will be stronger than actual similarity. That is, the correlation between each partner’s self-perception and perception of their partner will be stronger than the correlation between the partners’ self-perceptions.

H5. At baseline, there will be a positive association between partner similarity in self-reported characteristics and relationship quality.

H6. Longitudinally, there will be a positive association between partner similarity in change trajectories and average relationship quality.

In sum, the current paper combined multiple levels of personality constructs both within and outside of the Big Five framework, multiple perspectives through self and informant reports, and multiple timescales from one-time measures to longitudinal change trajectories. By including both sets of variables in both samples, we were able to directly compare effect sizes between scores on the Big Five traits versus relationship-specific characteristics. Furthermore, the dyadic self-perception and partner-perception designs allowed us to investigate differences between actual and perceived similarities. In doing so, we aimed to investigate the added utility of expanding personality assessments beyond one-time, self-reported measures of Big Five traits which may provide the additional granularity needed for the romantic relationship context.

The current research employed two samples of romantic couples for all analyses. The first consisted of early dating couples, whereas the second consisted of established couples across their first two years of being new parents. These two samples were complementary in that they both measured personality traits and various characteristic adaptations using a longitudinal dyadic design, but they drew from different populations of romantic couples in different stages of their relationship and spanned different time ranges. Because the sets of relationship-specific characteristics differed across the two samples, differing result patterns should not be attributed directly to their demographic differences. Instead, these two samples altogether simply provided a broader examination of these processes. Notably, the only variables that were present across both samples were the Big Five traits and two attachment orientations of Avoidance and Anxiety; however, they were all measured by different scales.

Descriptive statistics and bivariate correlations for all measured variables at baseline are shown in Tables 1 and 2. We have indicated the variables for which there was a significant difference between male and female participants, using two-tailed paired-samples t-tests.

This sample included 184 heterosexual couples who were recruited online (on a university department website) and via flyers distributed on the University of Minnesota – Twin Cities campus and in nearby neighborhoods (Weisberg, 2011). There were 184 dyads at Time 1 for which both partners provided data, 118 at Time 2, and 67 at Time 3. Handling of missing data due to attrition is addressed in the Results section. To be eligible, both members of each couple needed to be at least 18 years of age and have been dating for between 3 and 12 months. Each participant was compensated with up to $25 or 10 extra-credit points for their psychology class. The surveys were administered online using the University survey platform. The study had a dyadic longitudinal design with three waves of data collection eight weeks apart, spanning four months in total. Participants were invited to complete a fourth wave of data collection, but the last wave was not reported due to high attrition and our pre-registration specified a three-wave design. Surveys were completed independently by each partner. Participants were 21.30 years old on average (SD = 4.33), ranging from 18 to 58 years old, with an average relationship duration of 11.7 months. Female participants were slightly younger (p < .05) on average (M = 20.72, SD = 3.99) than male participants (M = 21.88, SD = 4.59). The majority of participants were white (79.1%), followed by Asian (11.7%), Hispanic (3.0%), and African American (2.2%). 6.8% of participants identified with a non-listed racial/ethnic group. All longitudinal variables showed good test-retest reliabilities, ICC(1,k) > 0.75 (Table 3).

Big Five Aspect Scale (BFAS; DeYoung et al., 2007) is a 100-item measure of the Big Five domains and ten underlying aspects (two per trait). The current sample used the full 100-item version for self-reports at baseline and a shortened 40-item version with four items per aspect for the remaining assessment waves. Participants answered two versions: one about their own personality, and one about their partner’s perceived personality. All items were answered on a five-point Likert-type scale, ranging from 1 (strongly disagree) to 5 (strongly agree). The scale at baseline instructed participants to answer about their personality in general, whereas subsequent scales instructed them to answer about their personality during the past two months since the last assessment wave. Cronbach’s alphas at baseline are: Agreeableness α = .81, Conscientiousness α = .82, Extraversion α = .87, Neuroticism α = .88, and Openness α = .81.

Adult Attachment Questionnaire (AAQ; Simpson et al., 1992) is a 17-item measure of romantic attachment orientations to partners in general. All items were answered on a seven-point Likert-type scale, ranging from 1 (I strongly disagree) to 7 (I strongly agree). The measure contains two subscales: Avoidance – eight items (baseline α = .80) and Anxiety – nine items (baseline α = .76). Each subscale was computed as the mean across the items.

Responsiveness Scale (Reis, 2004) is an 18-item measure of partner’s perceived responsiveness, including the degree to which they are perceived as understanding, validating, and caring. All items were answered on a nine-point Likert-type scale, ranging from 1 (not at all true) to 9 (completely true). A responsiveness score was computed as the mean across all 18 items, with baseline α = .95.

Trust Scale(Rempel et al., 1985) is a 17-item measure of general trust in romantic partners/relationships. The items inquire about romantic partners in general instead of a person’s current partner. All items were answered on a seven-point Likert-type scale, ranging from 1 (strongly disagree) to 7 (strongly agree). A trust score was computed as the mean across all 17 items, with baseline α = .87.

Perceived Relationship Quality Component Scale (PRQC; Fletcher et al., 2000) is an 18-item measure of current relationship quality. All items were answered on a seven-point Likert-type scale, ranging from 1 (not at all) to 7 (extremely). There are six subscales, each computed as the mean across three items: Satisfaction, Commitment, Intimacy, Trust, Passion, and Love. We computed a total relationship quality score as the mean across the most representative items from each subscale, as specified by the scale authors (Fletcher et al., 2000, p. 351), with baseline α = .88.

This sample consisted of 168 heterosexual couples who were recruited from childbirth preparation classes and flyers distributed at a local hospital (Rholes et al., 2011). In order to be eligible, participants had to be married or cohabiting and expecting their first child. Participants were 27.67 years old on average (SD = 4.26), with a range from 19 to 45 years old, with an average relationship duration of 40.9 months or 3.4 years. Female partners were slightly younger (p < .05) on average (M = 26.88, SD = 4.03) than male partners (M = 28.46, SD = 4.35). The majority of participants (274 individuals or 81.55%) identified as white, followed by Hispanic (28 individuals or 8.3%), Asian (25 individuals or 7.4%), and Black (only 1 individual).

There were 168 complete dyads at Time 1 in which both partners provided data, 153 at Time 2, 144 at Time 3, 142 at Time 4, and 129 at Time 5. Handling of missing data due to attrition is addressed in the Results section. Surveys were mailed separately to participants’ homes and returned separately by each partner, except for Time 2, during which participants completed the surveys in-person at the research lab. The surveys, in other words, were completed independently by both partners. The first wave of data was collected approximately six weeks prior to childbirth. The first two waves were approximately 7.5 months apart and the last four waves were six months apart, spanning a total of two years and six weeks. All longitudinal variables showed good or excellent test-retest reliabilities ICC(1,k) > 0.75 or 0.90 (Table 4).

Big Five Inventory (BFI; John et al., 1991; John & Srivastava, 1999) is a short-form version that includes 35 items that measure each of the five personality traits: Agreeableness (α = .68), Conscientiousness (α = .73), Extraversion (α = .86), Neuroticism (α = .85), and Openness (α = .70). All items were answered on a five-point Likert scale, ranging from 1 (disagree strongly) to 5 (agree strongly). This measure was administered only at baseline.

Experiences in Close Relationships (ECR; Brennan et al., 1998) is a 36-item measure of adult attachment Avoidance (18 items, baseline α = .91) and Anxiety (18 items, baseline α = .90). The seven-point Likert-type items range from 1 (disagree strongly) to 7 (agree strongly).

Social Support Questionnaire (SSQ; Sarason et al., 1987) is a seven-item measure of social support from romantic partners. Participants answered two versions of the questionnaire: one about their perceived social support from their partner (baseline α = .90), and one about the social support they provide to their partner (baseline α = .88). For instance, a similar question stem was adapted into two versions: “How much can you count on your partner/spouse to console you when you are very upset?” and “How much can your partner/spouse count on you to console him/her when he/she is very upset?” In addition, the self-reported version of perceived support from partner includes an additional general satisfaction question: “Overall, how satisfied are you with the support you receive from your partner/spouse?” All items were answered on a seven-point Likert-type scale, ranging from 1 (not at all) to 7 (very much).

Social Provisions Scale (SPS; Cutrona, 1989) is a 14-item measure based on Weiss’ social provisions theory which outlined six categories of social provisions that are necessary in healthy relationships (Weiss, 1974). For instance, a similar question stem was adapted into two versions: “Does your partner/spouse enjoy the same social activities that you do?” and “Do you enjoy the same social activities as your partner/spouse?” Participants answered two versions: one about their perceived social provisions from their current partner (baseline α = .59), and one about the social provisions they provide to their partner (baseline α = .69). For instance, a similar question stem was adapted into two versions: “Does your partner/spouse enjoy the same social activities that you do?” and “Do you enjoy the same social activities as your partner/spouse?” All items were answered on a three-point scale with response options being no, sometimes/not sure, and yes.

Caregiving Questionnaire (CQ; Kunce & Shaver, 1994) is a 32-item measure of individual differences in caregiving and care-seeking behaviors. Participants answered two versions: one about their own caregiving behaviors toward their partner, and one about how they perceive their partner’s caregiving behaviors toward them. There are four subscales, each with eight items: (1) Sensitivity vs. Insensitivity measures one’s ability to perceive a partner’s care-seeking signals (baseline self α = .85, partner α = .86); (2) Proximity vs. Distance measures one’s willingness to respond and provide care (baseline self α = .90, partner α = .92); (3) Cooperation vs. Control measures one’s support of their partner’s own effort to solve problems (baseline self α = .88, partner α = .90); and (4) Compulsive Caregiving measures to tendency to overly intrude in a partner’s problems (baseline self α = .83, partner α = .76). All items were answered on a seven-point Likert-type scale, ranging from 1 (disagree strongly) to 7 (agree strongly).

Conflicts and Problem-solving Scales (CPS; Kerig, 1996) is a 32-item measure of different conflict resolution strategies. Participants answered two versions: one about their partner’s conflict strategies, and one about their own conflict strategies. The scale assesses four dimensions: (1) Verbal Aggression with items such as “Make accusations” or “Become sarcastic” (baseline self α = .86, partner α = .89); (2) Collaboration with items such as “Try to understand what the other is really feeling” or “Accept the blame, apologize” (baseline self α = .90, partner α = .89); (3) Stalemate with items such as “Threaten to end relationship” or “Complain, bicker without really getting anywhere’ (baseline self α = .72, partner α = .72); and (4) Avoidance-Capitulation with items such as”Leave the house” or “Clam up, hold in feelings” (baseline self α = .82, partner α = .72). The original 44-item scale (Kerig, 1996) included two more subscales that were not assessed in the current sample: Physical Aggression and Child Involvement. All items were answered on a seven-point scale indicating the frequency of each strategy, ranging from 1 (once a month or less) to 7 (just about every day).

Dyadic Adjustment Scale (DAS; Spanier, 1976) is a 32-item measure of romantic relationship quality. The sample answered 2 of the 4 original subscales on 6-point scales: (1) Dyadic Satisfaction, with 10 items measuring general satisfaction with one’s partner (baseline α =.81), and (2) Dyadic Cohesion, with 5 items measuring the degree to which each couple participated in different activities together (baseline α = .72). Scale scores were computed as the sum across all items and could range from 0-50 for Satisfaction and from 0-25 for Cohesion.

The research program was preregistered as part of a dissertation (https://osf.io/hzvuj). However, the current manuscript presented an updated and more concise version of the original preregistered plan. Specifically, several original hypotheses and analytic plans were removed for a more coherent manuscript. In addition, the authors subsequently learned about more stringent and appropriate analytic methods to replace original ones (Bayesian models allowing simultaneous estimations of partners’ slopes and dyadic response surface analysis for similarity effects). All preregistered analytic code in the R programming language (R Core Team, 2023) and associated results are retained in the online supplemental materials¹. The datasets had also been previously analyzed by members of the author team for separate research projects.

Another substantial change to the preregistered plan included changing the inference criteria to control for false discovery rate, which led to a more conservative inference threshold than the common .05. In addition, exploratory un-preregistered analyses of gender differences were conducted to better understand the patterns of responses at baseline. Although evidence of longitudinal change was not the focus of the current research, this can be extracted from our longitudinal models and is reported in the Results section to clarify patterns of responses across assessment waves (also see Tables 3 and 4).

The current research included three research questions with ten total hypotheses. Each hypothesis was tested separately on two samples and again separately across various personality traits and relationship-specific constructs. As a result, we conducted a large number of significance tests. Our pre-registered decision to maintain an alpha level of .05 despite the numerous tests was due to our concern with statistical power. Even though we had repeated-measures data that increases statistical power in longitudinal analyses, our sample sizes were moderate for both studies at baseline, and we did not want to further reduce power with a conservative alpha threshold. However, as a non-preregistered post-hoc decision, we chose to control the False Discovery Rate (FDR), which is the proportion of statistically significant tests that are truly null. Methods to control FDR are less vulnerable to type II error compared to those used to control for Family-Wise Error Rate, such as Bonferroni or employing a more stringent alpha based on the number of comparisons and, therefore, are better equipped to preserve statistical power (Murray & Blume, 2021). Specifically, we used the Benjamini-Hochberg approach, which controls FDR using the extracted p-values from all independent or semi-independent significance tests (Benjamini & Hochberg, 1995). Unlike the .05 alpha threshold for p-values, there is no established standard FDR threshold, although many researchers use values between 10% and 30%, depending on the research context (James et al., 2013). We chose the conservative 10% in the current research.

Analyses for research question three were not all independent due to multiple tests within the Actor-Partner Interdependence Models. Moreover, some additional tests as well as profile correlation analyses associated with hypothesis H1 did not strictly rely on p-values alone. As a result, this adjusted inference criterion was applied to only the first two research questions and excluded the profile correlation analyses, resulting in a total count of 249 p-values. A sizeable portion of this total count resulted from hypothesis H2, which involved comparisons of between-partner correlations among all 172 pairs of baseline variables across the two samples. All p-values were extracted and transformed into q-values using the p.adjust function in the stats R package (R Core Team, 2023), Values smaller than the desired FDR of .10 were considered significant. For the current research with combined hypothesis tests from both samples, this Benjamini-Hochberg procedure produced an adjusted significance threshold of p-values at or below .029. In other words, a p-value of .029 was the highest raw p-value that was considered significant while controlling FDR at 10%, and any p-value higher than .029 was considered not significant. In the result section, raw p-values are still reported to conform to typical reporting practices, but they were interpreted using this adjusted threshold.

The current paper involved many separate analyses, each with their own assumptions and specifications that may have different implications for power calculations. Further, we used two longitudinal datasets for which data collection had been completed prior to this paper preregistration and analyses. As a result, we provided here power considerations rather than a priori power calculations for the two primary analyses: (1) correlations between partners’ characteristics and (2) dyadic response surface analysis for similarity effects. These two analyses covered the basis of our research questions: the existence of assortative mating and its association with relationship quality.

For the simple bivariate correlation between partners’ characteristics, using a typical alpha level at .05, a power level of 80%, and a one-sided test as we are hypothesizing a positive association to support assortative mating, we were able to find the smallest effect size that could be detected with our given samples. This yielded an effect size of r = .18 for sample one of 184 couples and r = .19 for sample two of 168 couples. For references, an extensive meta-analysis of 23,000 romantic couples have provided the following estimates for bivariate correlations: Extraversion = .08, Neuroticism = .10, Agreeableness = .11, Conscientiousness = .16, Openness = .21, with an average of r = .11 (Horwitz et al., 2023). However, the current study proposed that more contextualized variables, such as relationship-specific characteristics and longitudinal change trajectories, would show stronger evidence for assortative mating compared to the broad one-time personality trait measures. As a result, our analyses of the existence of assortative mating using bivariate correlations should be adequately powered to detect these effects.

Nonetheless, we conducted analyses that were far more complex than bivariate correlations. For instance, dyadic response surface analyses involved multiple paths and associated parameters which made power calculations much more difficult. In these cases, a simulation study is recommended; however, this requires clear definitions of the numerous (co)variances and path parameters in the simulation configuration (Schönbrodt et al., 2018). An alternative rule of thumb suggested multiplying the necessary sample size for linear main effects by 2 or 3 times for higher order terms (Aiken et al., 1991). A meta-analysis on the effects of personality traits on relationship satisfaction provided the following estimates: Openness = .03, Extraversion = .06, Conscientiousness = .12, Agreeableness = .15, and Neuroticism = -.22 (Malouff et al., 2010). To detect this effect even for the most relevant trait of Neuroticism at 80% power with alpha of .05, we would need 160 couples. Following this rule of thumb would result in a total of 320-480 couples, far exceeding our available sample. As a result, readers should interpret these model results with caution and regard them as exploratory and hypothesis-generating rather than confirmatory. In other words, the lack of significant findings here should not be treated as definitive confirmation that there are no similarity effects on relationship satisfaction but potentially a result of an underpowered study design.

Attrition is an unavoidable component of nearly all longitudinal studies, and it was evident in both of our samples. Although many of our research questions focused on baseline data, several hypotheses involved analysis of change over time and thus required multiple waves of data collection. As a result, we conducted a missingness analysis to determine whether missing data could be predicted by any of our variables. Specifically, our variables were treated as predictors in a logistic model predicting the presence or absence of any missing data wave for each couple. We did not see any evidence of non-random missingness in sample one; none of the demographic variables, relationship-specific characteristics, or Big Five personality traits at baseline were significant predictors of missingness. In addition, relationship duration and overall quality reported at baseline did not predict missingness.

We saw more evidence of non-random missingness in sample two. This is not surprising given that the data collection was more demanding for participants. Whereas the early dating couples in sample one provided only three waves of data eight weeks apart, the couples in sample two started the study right before the birth of their first child and then completed five data collection waves at roughly six-month intervals across the first two years of parenthood. Age was the only demographic variable that significantly predicted missingness. Couples with older than average male partners had a lower missingness rate, such that a 1-year increase in age was associated with a 9% decrease in missingness likelihood (z = -2.41, p = .02). Higher female self-reported Extraversion (z = 2.11, p = .03) and Agreeableness (z = 2.43, p = .01) were also associated with higher missingness likelihood. However, there was no association between self-reported relationship-specific characteristics or relationship quality reported at baseline and missingness likelihood. Taken together, predictors of missingness were few and of generally small magnitude. As a result, data from all participants at all available time points were used in the longitudinal models.

At baseline, we examined the bivariate Pearson r correlations between dyadic member’s self-reported scale scores at baseline, with all results presented in Tables 3 and 4 and depicted in Figures 1 and 2. All dyads were distinguishable (i.e., heterosexual couples with one female and one male partner). For sample one, using our FDR-adjusted significance threshold of p ≤ .029, we found that romantic couples were significantly and positively correlated on Agreeableness (r = .19), Openness (r = .25), Responsiveness (r = .46), and Trust (r = .24). There was no evidence of assortative mating in either attachment orientations or the other Big Five traits. For sample two, we found that many relationship-specific variables showed significant and positive assortative mating patterns for both self-reports and partner-reports. This included both attachment Avoidance (r = .18) and Anxiety (r = .20), Social Provision (r = .31), caregiving Sensitivity (r = .27), and all conflict strategies (rs ranging from .21 to .42). Notably, the other three caregiving styles did not reveal significant between-partner correlations (Proximity, Cooperation, and Compulsive Caregiving; rs ranging from .04 to .07). On the other hand, none of the Big Five traits showed significant between-partner similarity, with bivariate correlations ranging from .00 for Extraversion and Neuroticism to .07 for Conscientiousness and .12 and .16 for Agreeableness and Openness, respectively.

In addition, profile correlations (Humbad et al., 2013) were computed for all scales for which there were more than two subscales because these analyses entailed bivariate correlations between partner’s scores on the subscales. As a results, these analyses included the self-reported scale scores of the Big Five Aspect Scale BFAS in sample one and the Big Five Inventory, Caregiving Questionnaire, and Conflict and Problem-solving Scale in sample two as indices of multivariate similarity in personality traits, caregiving styles, and conflict strategies. The unit of analysis was at the dyad level. For each of these scales, a bivariate Pearson r correlation was computed between the two partners’ score vectors. For example, for the Big Five traits, profile correlation for each couple was computed as the bivariate correlation between two vectors of five scores each for each partner. For each scale, three sets of models were run with profiles correlations calculated by raw, gender-mean-centered, and standardized scores. Centering and standardizing scores prior to computing profile correlations alleviate the effects of a “normative” personality profile when assessing similarity. For instance, we might expect Big Five trait profiles of strangers to be correlated simply due to common patterns of personality traits in the population: their personality profiles might both be similar to the average person, or the socially desirable person, and removing this normative component is important to understanding their true similarity (Furr, 2008).

For profiles using Big Five traits, roughly 15.1% of early dating couples and 14.4% of first-time parents had significant raw-profile correlations. This percentage dropped to 5% and 5.4% for gender-mean-centered profiles and dropped to 5% and 4.2% for standardized profiles. In contrast, for profiles of relationship-specific characteristics, we found that 76.8% of first-time parents showed significant profile correlations in conflict strategies and 19.2% in caregiving styles using raw scores. These percentages dropped to 21.4% and 7.1% for centered scores and dropped to 19% and 8.9% for standardized scores. These steep drops indicate that much of the profile correlations are due to normativeness and not distinctiveness in participants’ reports.

We next compared assortative mating in personality traits and relationship-specific characteristics by conducting a z-difference test using Fisher’s z-transformed bivariate correlations to test for differences between each trait correlation and each CA correlation Zou et al., Zou, 2007. A significant difference was found for several pair-wise comparisons across both samples. Namely, among early dating couples in sample one, partner similarity in all Big Five traits was significantly lower than similarity in Responsiveness (but not significantly so for Trust).

In sample two, partner similarity in all Big Five traits was also lower than many of the relationship-specific variables measured in first-time parents. For example, similarity in Verbal Aggression was significantly higher than similarity in all Big Five traits and in both attachment orientations; similarity in Social Provision was significantly higher than that in Conscientiousness, Extraversion, and Neuroticism; and similarity in caregiving Sensitivity was significantly higher than that in Extraversion and Neuroticism, especially for Verbal Aggression. Notably, attachment orientations behaved similarly to the Big Five traits in sample one, such that similarity in these variables was lower than similarity in many other relationship-specific variables. There were also a few differences among the non-attachment relationship variables, but this was primarily due to the high between-partner correlation in Verbal Aggression (r = .42). All of the comparison tests are presented in Tables S1 and S2.

Longitudinally, we fit Bayesian multivariate multilevel models with time as the Level 1 predictor. Two separate models were specified and combined for two partners. Residual correlations for the two dependent variables represent the correlated change between the two partners. This analytic approach allowed for simultaneous estimation of all parameters in the same system of equations (DiGiovanni et al., 2021). Because these are Bayesian models, results were interpreted using 95% credible intervals instead of traditional p-values.

For early dating couples, we found positive and significant slopes for self-reported attachment Avoidance (r = .17, 95% CI [.00, .34]) and Responsiveness (r = .28, 95% CI [.12, .42]. There were no significant longitudinal correlations involving any of the Big Five variables, attachment Anxiety, or Trust. Notably, participants also did not show changes in most of these variables across the short four-month duration of sample one. Although evidence of change was not the focal point of the current research, all fixed effects estimates are shown in Tables 3 and 4.

We saw more evidence of change and correlated change for our established couples in sample two, across the first two years of parenthood. Although we did not have longitudinal assessments of the Big Five traits, we saw individual change across the five assessment waves for many relationship-specific variables, such as decreases for both partners in attachment Anxiety (but not Avoidance), Collaborative conflict strategies, and Social Support and Provision. Interestingly, there was a decrease in Proximity caregiving and increase in Compulsive caregiving for female partners, but no significant change for male partners. Further, consistent with our hypothesis, we found evidence of significant positive correlated change in Social Provision (r = .16, 95% CI [.08, .25]) but not Social Support (slope r = .04). Correlated change was also found for Sensitivity caregiving (r = .09, 95% CI = [.00, .17]) as well as Stalemate (r = .13, 95% CI [.05, .22]) and Verbal Aggression (r = .14, 95% CI [.06, .23]) conflict strategies. We did not find any effect for self-reported attachment orientations or the remaining 5 conflict strategies and caregiving behaviors. Further, interestingly, many of these variables revealed significant correlated changes in partners’ perceptions rather than self-reports. For example, partners were more likely to report that each other had changed in the same way, but their self-perceptions did not reflect this similarity. In fact, the only two partner-reported variables that did not show significant correlated change were Compulsive Caregiving (slope r = .09, p = .24) and Collaboration conflict style (slope r = .15, p = .06). All of the other caregiving styles and conflict resolution styles showed significant partner-reported correlated change.

In addition to self-reports of personality traits and relationship-specific characteristics, our two samples also included partner-reports of several variables, which included the Big Five traits for early dating couples, and Caregiving styles and Conflict resolution strategies for first-time parents. As a result, we were able to compare perceived and actual similarity effects. That is, are romantic partners more similar to one another, and/or do they perceive themselves as more similar to one another than they actually are? Further, is perceived similarity or actual similarity more strongly associated with higher relationship quality?

In order to compare the effect sizes of actual and perceived similarity, we conducted z-difference tests using Fisher’s z-transformed bivariate correlations between: (a) self-reports provided by each partner, (b) the female partner’s self-perception and perception of their male partner, and (c) the male partner’s self-perception and perception of their female partner. All bivariate correlations are presented in Table 5 and 6 and depicted in Figures 3 and 4 for the two samples. For early dating couples, the only difference in trait perception involved Openness, such that male partners reported significantly higher similarity in Openness (r = .50, p < .001) than their actual similarity (r = .25, p = .001; z difference score = -2.78). For first-time parents, for all ten measured variables, participants reported higher similarity on these variables than the actual correlations between their self-reports. There was no difference in perceived similarity between female and male partners; all perceived similarities were positive and significant (ps ≤ .001).

We first examined whether there was an association between personality variables and relationship quality at baseline. To do so, we conducted dyadic response surface analysis (DRSA; Schönbrodt et al., 2018) which allow for direct test of the congruence effects between two independent variables (i.e., two partners’ scores on a personality characteristic) on an external outcome (i.e., relationship quality). DRSA involved fitting a polynomial regression model in which both partners’ scores along with their squared values and two-way interactions are used to predict each partner’s relationship quality. The general structure of this model framework is provided in Figure 5. First, model comparison was performed between a full model that allows for gender-differences in path estimates and a reduced model with equality constraints on gender-specific paths. If model fit was not significantly improved by releasing the equality constraints, the more parsimonious model would be preferred. For inferences on the congruence effects, a series of auxiliary parameters were created from these regression estimates:

where $b1$ and $b2$ were the actor and partner paths between one’s own personality characteristics and one’s own or one’s partner’s relationship quality; $b3$ and $b5$ were the squared actor and partner paths; and $b4$ was the interaction term. We specifically conducted 3 tests for the presence of (1) actor or partner effects: if $a1$ and $a2$ were both significant; (2) broad congruence: if $a4$ was significantly negative and $a3$ and $a5$ were both significant; and (3) strict congruence: if broad congruence was achieved but there was no actor or partner effects such that if $a1$ and $a2$ were both non-significant. Broad congruence was sufficient to confirm our hypothesis of the benefits of personality similarity, which still allowed for general actor/partner effects.

For sample one of early dating couples, the only predictors for which a full model performed better than a gender-constrained model were Agreeableness and Responsiveness. The resulting 11 models (2 models each for Agreeableness and Responsiveness) all showed evidence for actor/partner effects, with significant $a1$ and $a2$ terms. As a result, it was impossible to support a strict congruence pattern. Further, none of these models supported a broad congruence pattern. In other words, there was no relationship between partner similarity and relationship quality with actor and partner paths in the same model. Detailed model results can be found in Table S3.

For sample two of first-time parents, the only predictor for which a full model performed better than a gender-constrained model was caregiving Sensitivity v. Insensitivity. Similar to sample one, there was no evidence for either broad or strict congruence effects. However, not all models supported actor/partner effects either, particularly so for the Big Five traits. Specifically, there was no actor/partner effect for Extraversion and Agreeableness on either Dyadic Satisfaction or Cohesion. Dyadic Cohesion was also not predicted by actor/partner paths for Conscientiousness, Neuroticism, or Verbal Aggression. Detailed model results for this sample can be found in Table S4.

Alternatively, for a multivariate perspective of similarity, for the Big Five Aspect Scale in sample one and for the Big Five Inventory, Caregiving Questionnaire, and Conflict and Problem-solving Scale in sample two, we ran a simple linear regression model in which the Fisher’s z-transformed score of the profile correlations was used to predict relationship quality. After centering and standardizing scores to remove normativity effects, this similarity approach also did not indicate any significant association with relationship quality, either for early dating couples or for first-time parents.

In addition to baseline benefits, we also examined potential benefits of longitudinal similarity on relationship quality. In other words, we tested whether similarity in longitudinal trajectory of different personality characteristic would be associated with higher relationship quality. To this end, we conducted analogous DRSA models to those for benefits of baseline similarity but substituted baseline self-reported personality scores with extracted individual slopes of personality trajectories over time. In addition, the outcome variables of relationship quality were operationalized as average scores across timepoints for ease of analysis. Similar to the previous section, we tested for the presence of actor or partner effects, broad congruence, and strict congruence effects on relationship quality.

For sample one of early dating couples, the only predictors for which a full model performed better than a gender-constrained model were changes in Neuroticism and attachment Avoidance. All but 2 of the resulting 11 models showed evidence for actor/partner effects, with significant $a1$ and $a2$ terms. The exceptions of no actor/partner effects were for changes in Conscientiousness ($a1$ = -.04, p = .76; $a2$ = -.04, p = .74) and changes in attachment Avoidance for female partner’s relationship quality ($a1f$ = .10, p = .49; $a2f$ = -.03, p = .89). Further, there was no evidence of either broad or strict congruence patterns for any of the 11 models. Detailed model results can be found in Table S5.

For sample two of first-time parents, there were a total of 30 models, 12 of which resulted from gender-specific model pairs for the following slope predictors: Social Provision (for Dyadic Satisfaction), caregiving Sensitivity (for Dyadic Cohesion), and conflict resolution strategies of Capitulation and Verbal Aggression (for Dyadic Satisfaction) as well as Stalemate (for both Dyadic Satisfaction and Cohesion). Interestingly, actor/partner effects were not present across the board, with no evidence of effects for Dyadic Cohesion with attachment Anxiety, caregiving Sensitivity (female), Cooperation, Compulsiveness, and Stalemate conflict strategy, as well as no effects for Dyadic Satisfaction with caregiving Sensitivity or Compulsiveness. There was evidence for broad congruence for only one model: longitudinal similarity in caregiving Proximity v. Distance predicting Dyadic Cohesion. A visual representation of this model is presented in Figure 6. The Line of Congruence (LOC) represents perfect combinations of male and female partners slopes. Due to the slightly rising edge, we can see that this was not a strict congruence pattern, because the level at which partners were matching also mattered for the outcome variable. Specifically, when both partners showed a strong decrease in this caregiving style over time, they reported very low average Dyadic Cohesion. Nonetheless, on average, congruent couples reported higher Dyadic Cohesion than non-congruent couples for this predictor. Detailed model results can be found in Table S6.

The current research can be summarized under two central aims: to examine (1) the evidence for and (2) the benefits of assortative mating patterns in romantic relationships. To further examine the evidence for assortative mating patterns, we compared assortment based on perceived versus actual similarities. To further examine its benefits, we compared the effects of assortment to the effects of each partner’s scores on relationship quality (i.e., actor and partner effects). The discussion below is organized under these two aims. We found strong support for one of our central expectations—that there would be evidence of stronger assortative mating for relationship-specific constructs than for the broader Big Five personality traits as measured at baseline. We further found evidence for correlated change such that partners showed similar change trajectories, mainly in relationship-specific characteristics, across both samples. However, there was limited evidence for the benefits of assortment for relationship quality, particularly after controlling for the effects of each partner’s characteristics, and perceived similarity was often much stronger than actual similarity as examined using self-reports.

In both samples, between-partner correlations for the Big Five traits were all below .25 and often not statistically significant. Openness showed the highest assortment in both samples, which is consistent with past meta-analytic results (Horwitz et al., 2023) and unsurprising given its association with cognitive abilities (Kaufman et al., 2016). In contrast, effect sizes for assortment on relationship characteristics such as Responsiveness, Trust, and specific conflict resolution strategies were not only moderate in magnitude and statistically significant but were also often significantly stronger than those of the Big Five traits. Responsiveness showed the strongest between-partner correlation across both samples, at r = .46. This variable assessed each person’s perception of how understanding, validating, and caring their current partner is to their personal needs. It includes items such as “My partner usually sees the same virtues and faults in me as I see in myself” and “My partner usually seems interested in doing things with me.” Unlike the Big Five traits, which describe general personality tendencies across time, different people, and different situations, or attachment orientations, which inquire about the nature of romantic partners and relationships in general, Responsiveness is contextualized to measure what has transpired with one’s current partner and in the current relationship. This finding pattern supported our central argument, that Big Five traits may lack the granularity needed to demonstrate assortative mating patterns and that more conceptualized variables like the relationship-specific characteristics captured in these samples would be more appropriate for this investigation.

Interestingly, the effects for attachment orientations were more trait-like in that they tend to be smaller than the relationship-specific variables, even though similarity in attachment was larger in magnitude among established couples in sample two than among the early dating couples in sample one. This may be related to the fact that both samples used assessments of global attachment orientations, rather than with regard to a specific partner, which qualify as personality traits under some definitions (DeYoung, 2015). As a result, they might behave more similarly to the Big Five traits than the other more contextualized relationship constructs. On the other hand, nearly all of the relationship-specific characteristics, such as partner responsiveness, conflict resolution strategies, social support/provision, and caregiving styles, are outcomes of dyadic phenomena that inherently assess how individuals interact with and relate to their current partners. As a result, an extensive body of literature has demonstrated the associations between these variables at the individual levels with relationship outcomes (e.g., Berli et al., 2021; Kim et al., 2015; Stanton et al., 2019), and it was thus reasonable to also expect higher levels of correlations for these variables than the global attachment orientations or Big Five traits.

This basic pattern also emerged in the longitudinal trajectories of traits and relationship-specific characteristics in both samples. Specifically, Responsiveness and Trust in sample one showed strong correlated changes across the duration of that study. Sample two provided additional important insights into self-reports and partner-reports of longitudinal change. Across the first two years of parenthood, self-reports provided by each partner did not show much correlated change, except for Social Support and Provision, yet dyad members reported that they both changed in the same direction on several variables. It is possible that, during this major life transition with the added cognitive demand of rearing one’s first child, partners might be answering questions about their partners in more normative ways or relying on heuristics with their new and very salient social roles as parents, which may lead to higher correlations between partner-reports, but not self-reports.

Furthermore, a sizeable portion of dyads in both samples showed significant profile similarity. This is consistent with some past research that has documented different patterns of assortative mating, depending on the analytic approach used and how assortment is operationalized, especially with evidence of profile correlations (Luo & Klohnen, 2005). Nonetheless, in the current research, we found that the majority of profile similarity was due to effects of normative profiles instead of distinctive features in partners’ profiles, because the proportion of significant profile similarity dropped markedly between profiles computed using raw versus centered versus standardized scores. Specifically, conflict resolution profiles showed a striking 76.8% significance rate in sample two using raw scores, which dropped to just 19% using standardized scores. This highlights that similarity in conflict resolution profiles across different subscales (e.g., Collaboration and Stalemate) may simply be due to normative patterns in the population rather than anything specific to each dyad.

We also found more support for perceived over actual similarity effects in both samples using different analytic approaches, which is highly consistent with the literature (e.g., Montoya et al., 2008). This poses further questions about the potential mechanisms underlying mating assortment. If perceived similarity is consistently stronger than actual similarity in romantic dyads, are people choosing mates based on their explicit perceptions rather than on who may truly be more similar to them? The fact that this pattern was found for both early dating and established couples suggests that it exists beyond initial attraction and relationship initiation and may extend to relationship maintenance processes. Further, it may be the case that romantic partners possess intimate and unique knowledge of one another above and beyond what can be captured via self-reports, and it might be misleading to consider correlations between self-reports to be indicative of actual similarity. Indeed, the Self-Other Knowledge Asymmetry model suggests that self-reports need not be the “golden standard” but are instead just one of many perceptions of a person, each with its own unique perspectives and biases (Vazire, 2010). Meta-analyses have demonstrated that informant-reports were strong predictors of both work and academic performances (Connelly & Ones, 2010; Oh et al., 2011), and another primary study have shown a predictive advantage of informant-reports over self-reports in predicting work performance from acquaintances across different contexts (Connelly & Hülsheger, 2012). Informant-reports from well-known raters also provided incremental validity in other evaluative constructs such as prosocial and moral behaviors (Thielmann et al., 2017).

Although we found evidence in support of assortative mating for many variables, there was little to no evidence that assortment is beneficial to relationship outcomes. We conducted dyadic response surface analysis which accounts for the dependence across all actor and partner effects. Consistent with existing research (Dyrenforth et al., 2010; Joel et al., 2020), we found overwhelming support for the predominance of actor effects over partner and similarity effects. It was often each person’s score on a relationship-specific characteristic, more so than their partner’s score or their similarity, that significantly predicted their relationship quality. Importantly, any benefits of assortment did not remain after controlling for each partner’s individual scores in the same model. Interested readers may refer to the supplemental materials for simple regression models, for which there were significant associations between difference scores and relationship quality. Nonetheless, these models indicated that these effects were simply statistical artifacts. For instance, similarity on Responsiveness predicted higher relationship quality for both male and female partners in sample one; however, this might be due to the fact that Responsiveness was positively skewed (M = 7.47 out of 9), and thus what was captured in similarity was in most cases two partners who were both high in Responsiveness. As a result, after controlling for two partners’ scores, there was no longer a benefit for assortment. This lack of benefits for assortment was corroborated by a recent extensive longitudinal research program which followed 1,180 German couples from the start of cohabitation to dissolution (van Scheppingen et al., in press). Using the same approach with dyadic response surface analysis, they did not find dissimilarity to be associated with increased risk of dissolution. Nonetheless, their samples relied on Big Five traits alone and did not include additional individual differences in relationship-specific characteristics.

Lastly, we employed two samples from two different stages: early dating couples and established cohabiting/married couples. However, it is important to note that our research did not capture the initial stage of assortative mating through attraction alone, because even our early-dating couples were required to have been together for at least three months. Some research programs have directly examined assortative attraction when individuals preferred people who seemed similar to them after either no interaction or very short interactions (Montoya et al., 2008) and after viewing an online profile (Neyt et al., 2020). The processes underlying these superficial preferences are likely distinct to those underlying assortment in both of our samples. As a result, although we did not find an association between assortment and reported relationship quality, we cannot say definitively that there are no benefits to assortment because we have pre-selected couples who have made it at all to the dating or marriage/cohabitation stages. In other words, the fact that there was sufficient initial attraction and compatibility to form an established relationship might already be a testament to the benefits of assortment.

In sum, this research study highlighted the importance of incorporating multiple constructs, perspectives, and time scales in the investigation of assortative mating in romantic relationships. First, by expanding to personality constructs beyond the Big Five traits, we demonstrated that many relationship-specific constructs, such as Responsiveness, Trust, and conflict resolution strategies, showed much stronger evidence for assortment than Big Five traits. This further highlighted the bandwidth-fidelity trade-off: although Big Five traits are predictive of many different outcomes, their broad conceptualization means that they may simply lack the granularity needed to be strongly predictive, resulting in effects that are in magnitude, particularly in specific domains such as relationship initiation and maintenance.

Second, by incorporating both partners’ perceptions of themselves and of each other, we replicated existing research and demonstrated a consistent pattern that romantic partners perceived one another significantly more similarly than they actually were, highlighting the importance of accounting for multiple sources of truth beyond traditional self-reports. This pattern provided us with further insights into the mechanisms of assortative mating and suggests that it may involve at least some amount of conscious judgment. However, because all studied couples were already in established relationships, it may very well be the case that over time, these individuals had grown to perceive their close others as a more integral part of themselves and thus more similar to their own self perceptions. To further clarify this mechanism, a more extensive and careful consideration of the timescale is needed. Nonetheless, we would not have gained any of these insights if relying on self-reported personality alone.

These multiple perspectives also allowed us to examine the differential effects of actor and partner, as well as their similarity effects on relationship quality. Although there was little evidence for assortative benefits after controlling for the effects of each individual’s characteristics, all couples were in an established relationship rather than at the initial attraction stage, so it is therefore not appropriate to state that similarity was not at all beneficial. Nonetheless, this lack of finding indicated that at least there was no dosage effect to assortment: although some similarity may be beneficial, more is not necessarily.

Finally, we highlighted the importance of context by including two longitudinal samples across two very different relationship stages. As a result, we found evidence not only of simple bivariate similarity but also some correlated change over time, even with limited evidence of change from fixed effects estimates. This pattern is indicative of individual differences in change: although the sample did not show a consistent pattern of change on average, individuals may change in unique ways and often in similar ways to their partners. It is also important to consider the significance of the studied time period. For instance, the sample of first-time parents underwent dramatic changes to their personal and interpersonal lives during the duration of the study. In contrast, our sample of early dating couples were captured during a normative period and understandably did not show much longitudinal change patterns throughout the shorter duration of the study. Another way to provide additional context to the investigation was to use multivariate profile indicators instead of examining all traits in isolation. Unfortunately, this approach is not common in the literature and does not follow traditional significance testing frameworks. As a result, findings are more descriptive and difficult to interpret and synthesize. Nonetheless, we were able to provide evidence for assortment based on profiles of Big Five traits, conflict resolution strategies, and caregiving behaviors. We were further able to demonstrate that most of these effects resulted from a normative effect, that is, how two partners’ profiles are similar to a typical person, rather than distinctive features of the couple themselves.

An important limitation of the current research is that its numerous significance tests may have led to occasional false positives, despite our efforts to control False Discovery Rate and our detailed preregistration plan. Our statistical power was also somewhat limited in each study. Although a bivariate correlation of .25 only requires 97 couples to detect with 80% power, we conducted analyses more complex than bivariate baseline correlations. Our numerous hypothesized effects, coupled with the use of previously collected data, made power calculations difficult. We had to conduct our analyses with the data that were available to us and instead provided a discussion on power considerations to guide readers in interpreting our result patterns. All analytic codes and results are shared in the OSF repository, even those that failed to converge due to limited or missing data, so that future researchers can review them as desired. Our results, therefore, should not be interpreted as definitive, but rather as fodder for more directed research in the future. Future research should use our primary findings as a starting point to decide which variables and hypotheses merit more in-depth investigation. Additionally, some of our findings were not hypothesized, such as some gender differences in the effects of assortment (e.g., assortment on some variables was beneficial only for female or male partners).

Further, even though one of our samples was experiencing a distinct life transition (the transition to parenthood), our research focused on documenting more general assortative mating patterns. Thus, although we used this sample to examine individual differences in change trajectories across different variables, our hypotheses were not formalized with this specific life transition in mind. Not only is first-time parenthood a very unique period for the dyad, it is also an important individual milestone. As a result, it is highly likely that the individual and dyadic processes that transpire during this period may be quite different from those for couples in other relationship stages or for couples who would never undergo parenthood together at all. Future researchers interested in this relationship stage should develop pre-specified justifications and adequately powered designs to target these questions.

It is also important to note that although our main motivation was to highlight the additional information provided by more contextualized measures of individual differences, there is much more to explore. For instance, although we included self- and partner-reports, all perspectives come with their own set of unique insights and biases. Each partner’s perception may be influenced not only by their relationship dynamics but also greater societal context and expectations as well as personal biases. Further, we may argue that even a full 360 view of a target as reported by many informants would still be insufficient as they would all fall under quantitative measures of personality. Future research on dyadic assortment would benefit from expanding to personality constructs at higher levels by incorporating behavioral assessments and life narratives (McAdams & Pals, 2006), which offer a picture of the target in context and as a story-telling agent. In addition, some of our quantitative measures for relationship-specific characteristics shared some similarity between the set of predictors and outcomes. In particular, the trust construct was included as both a subscale for relationship quality (an outcome measure) and as an individual difference (a predictor for evidence or and benefit of assortment). This raised an important measurement consideration for contextualized constructs: at which point is a construct no longer just an individual difference but instead have evaluative qualities and should instead be used as an outcome measure? This limitation has indeed also been found for the Big Five traits as Neuroticism is often used as a predictor but share similar qualities to many psychotherapy outcome measures (e.g., Nguyen et al., 2020).

Another way to expand on this line of research is to capture the time period prior to or during mate selection. All romantic dyads in our samples have been together for at least three months or were committed enough to be cohabiting/married and expecting their first child. As a result, it was difficult to fully examine the benefits of assortative mating, because we did not capture the counterfactual: dyads who did not choose to pursue or commit to the relationship. A complete examination of this effect would necessitate expanding the time duration even further back to include the exploration stage. Admittedly, it would be challenging to identify and follow dyad members from when they were single individuals exploring different options in the dating market, but the advent and rising popularity of online dating environments may provide researchers with exciting new opportunities and research designs in this area.

The current research investigated assortative mating patterns, using both cross-sectional and longitudinal measures of personality traits and relationship-specific characteristics in two distinct samples of romantic couples. We found evidence for assortative mating across various relationship-specific characteristics, which were often stronger in magnitude than assortment based on Big Five traits. However, couples often perceived each other to be more similar than their actual similarity indicated. Further, we found evidence for longitudinal assortment such that both early dating couples and first-time parents had similar change trajectories for several relationship-specific characteristics over time. Nonetheless, although there was evidence for assortment, there was little evidence to support its benefits. After controlling for actor and partner effects, similarity no longer predicted relationship quality in either early dating couples or first time parents. In sum, although romantic couples across stages showed similarity and perceived each other similarly, assortment was not explicitly beneficial to measures of relationship quality.

This research highlighted the importance of incorporating multiple constructs, perspectives, and time scales in this investigation. Specifically, the inclusion of relationship-specific individual differences such as trust, attachment, perceived partner responsiveness, caregiving, and conflict resolution strategies revealed several unique and novel patterns of assortative mating beyond the Big Five traits. Further, employing both self-reports and partner-reports allowed for multiple interesting perspectives in assessing similarity and emphasized that perceived similarity is more important than actual similarity in assortment. Lastly, we wish to highlight the importance of within-person variability and of examining individuals not as static but flexible to change either normatively during their relationship or due to specific challenges during life transitions.

Contributed to conception and design: PLLN, MS

Contributed to acquisition of data: YJW, WSR, JAS, CGD

Contributed to analysis and interpretation of data: PLLN, MS

Drafted and/or revised the article: PLLN, MS, JAS, CGD, YJW

Approved the submitted version for publication: PLLN, YJW, CGD, WSR, JAS, MS

The authors declare no conflicts of interest.

All analyses are based on secondary data sources and the first author is not authorized for data sharing. Anonymized data may be available upon request.

The preregistration and analytic code are available on the OSF repository: https://osf.io/52nzf/.

Full descriptive summary: https://rpubs.com/nguyenllpsych/longassort-descr-01 (Sample One) and https://rpubs.com/nguyenllpsych/longassort-descr-02 (Sample Two)

Full analytic reports: https://rpubs.com/nguyenllpsych/longassort-results-01 (Sample One) and https://rpubs.com/nguyenllpsych/longassort-results-ttp (Sample Two)