Rejection Sensitivity Across Sex, Sexual Orientation, and Age: Measurement Invariance and Latent Mean Differences

Abstract

Intergroup differences in personality might be determined by systematic variation in social status and social experiences across groups. Because of its close association with social experiences, rejection sensitivity (RS)—a tendency toward anxious expectations of, and hypersensitivity to, interpersonal rejection—represents one such personality disposition that might differ across social groups, with implications for understanding mental health disparities. After first evaluating measurement invariance of the Adult Rejection Sensitivity Questionnaire (A-RSQ), the present research sought to assess whether latent mean differences in RS emerged across sex, sexual orientation, and age in a population-based sample of Swedish young adults (age 18–36; N = 1,679). Analyses revealed that the scale achieved full configural, metric, and scalar invariance across sex and sexual orientation and partial scalar invariance across age. As expected, tests of latent mean differences indicated that women, sexual minorities, and people 18–29 years old exhibited significantly higher RS levels than men, heterosexuals, and people 30–36 years old, respectively. Findings from the present research highlight the utility of attending to group differences in maladaptive personality dispositions and information processing styles and their potential role in contributing to persistent mental health hardships uniquely affecting women, sexual minorities, and younger people. Implications for scale administration and future research into the social causes and consequences of RS are discussed.

Interpersonal rejection is a common, yet highly aversive, social experience that threatens people’s need to belong (Baumeister & Leary, 1995) and influences their social behavior and mental health (Smart Richman & Leary, 2009; Williams, 2007). Research consistently shows that acute social rejection, relative to inclusion, reduces positive mood, self-esteem, and sense of control (Williams et al., 2000) and increases hostile emotions (Leary et al., 2006), antisocial behavior (Twenge et al., 2001), and suicidal thoughts (Chen et al., 2020). Yet, across individuals, responses to rejection are heterogeneous and differ markedly in intensity. Rejection sensitivity (RS) is one individual difference that might explain interindividual variability in perceptions of and reactions to rejection (Downey & Feldman, 1996). RS, the disposition to anxiously expect, readily perceive, and intensely respond to social rejection, is thought to be an interpersonal cognitive structure derived from early learning and attachment processes (Baldwin, 1992), such as rejection from attachment figures (e.g., caregivers, peers). Early interpersonal experiences are structurally organized in memory as a stable network of cognitions (e.g., rejection expectancies) and affects (e.g., rejection concerns) concerning the self in relation to others, guiding information processing and mediating social behavior in context (Baldwin, 1992; Mischel & Shoda, 1995). In interactional contexts, deeply entrenched interpersonal beliefs and affects are readily activated and influence the on-line selection and interpretation of social information (e.g., Higgins, 1996). As such, individuals high in RS, having learned to expect rejection from others based on prior experiences, interpret ambiguous social information as more threatening and distressing than people low in RS (Downey & Feldman, 1996). Such hypersensitivity to rejection, real or imaged, serves as an important interpersonal vulnerability that increases risk for poor mental health over the life course, as described below.

RS as a Cognitive–Affective Processing Dynamic Underlying Behavior

According to cognitive—affective processing system theory (CAPS; Mischel & Shoda, 1995), the personality system is conceptualized as a durable network of cognitions and affects, referred to as cognitive—affective units, that are differentially activated in response to particular situational cues, enabling stable or varying behavior depending on the stability or variation of situational contexts. Behavioral manifestations of the personality system are thought to be situationally specific and reflect interactions between the person (e.g., content and structural organization of cognitive—affective units) and the situation (e.g., social cues), such that behavioral tendencies are causally determined by activation of different cognitions and affects elicited in context. From this perspective, dispositional constructs, such as RS, are defined in terms of specific cognitive—affective units (e.g., expectations, concerns) that contribute to situation-specific behavioral patterns and can reliably distinguish prototypic exemplars of the personality type under study. Accordingly, the goal of social-cognitive personality assessment is to identify the content and accessibility of particular mental states (e.g., expectations, concerns) and the specificity of situations (e.g., interactional contexts) that routinely activate these states. Crucially, RS research adopts a CAPS approach to the conceptualization of RS, defining it as a personality type characterized by readily available and chronically accessible anxious expectations of interpersonal rejection that are stably activated upon exposure to certain social cues, namely those that signal rejection or the potential for it (e.g., Ayduk & Gyurak, 2008).

The distinctive social information processing style and situation-behavior relationships characteristic of RS are assumed to be highly stable across rejection-relevant contexts and chronically impact people’s social and psychological functioning. Prior research shows that RS is relatively stable over both short [3 weeks (r = .83), 4 months (r = .78), 6 months (r = .56); Downey & Feldman, 1996; London et al., 2007; Zimmer-Gembeck et al., 2016] and long [1 year (r = .64), 2 years (r = .65); Marston et al., 2010] periods of time, consistent with meta-analytic evidence showing that the average rank-order stability of personality traits is moderate (r ~ .50−.70; Roberts & DelVecchio, 2000). In further support of our conceptualization of RS as a stable processing dynamic underlying distinctive behavioral patterns, individuals high (vs. low) in RS exhibit increased attention to rejection-related social stimuli (Berenson et al., 2009; Downey et al., 2004), reduced extinction of a conditioned fear response to angry faces (Olsson et al., 2013), biased encoding and recall of rejection-related information (Mor & Inbar, 2009), increased emotional reactivity after exposure to ambiguous interpersonal information (Downey & Feldman, 1996), increased neural activity in brain regions involved in fear learning (Burklund et al., 2007) and emotional processing when exposed to disapproving faces (Kross et al., 2007), and more avoidant and aggressive behavioral responses following rejection (Ayduk, Gyurak, Luerssen, 2008; Romero-Canyas et al., 2010). Collectively, these data lend support to our conceptualization of RS as an enduring personality disposition that, once acquired, contributes to clinically relevant cognitive biases and maladaptive behavior in response to situations where rejection is possible.

Mental Health Significance of RS

RS is an important, yet understudied, personality disposition that might render individuals vulnerable to a range of mental health problems by biasing information processing and increasing active avoidance and fight-or-flight behavior in interactional contexts (see Gao et al., 2017). In fact, a growing body of literature indicates that RS is concurrently and prospectively associated with depression and anxiety symptoms (Ayduk et al., 2001; Chango et al., 2012; Gardner et al., 2020; Marston et al., 2010), borderline personality disorder symptoms (Ayduk, Zayas, et al., 2008), self-directed hostile cognitions and suicidal thoughts (Breines & Ayduk, 2015), systematically negative interpretation styles (Normansell & Wisco, 2017), defensive psychophysiological reactions to interpersonal rejection (Downey et al., 2004), impaired emotion regulation (Gardner et al., 2020; Silvers et al., 2012), poor relationship quality (Norona et al., 2018), and social withdrawal and avoidance (Gardner et al., 2020; Zimmer-Gembeck et al., 2016). Converging evidence further suggests that RS might be a key social-cognitive mechanism linking early exposures to adverse social events (e.g., parental abuse, peer victimization) to maladaptive attachment styles in adulthood (Downey et al., 1997; Feldman & Downey, 1994), self-regulatory deficits (Gardner et al., 2020; Silvers et al., 2012; Zimmer-Gembeck et al., 2016), and adverse mental health later in life (Gardner et al., 2020; Zimmer-Gembeck et al., 2014, 2016), lending support to the role of RS in the development and maintenance of mental health problems.

Among young adults, individual differences in RS are measured with a 9-item self-report scale, the Adult Rejection Sensitivity Questionnaire (A-RSQ; Berenson et al., 2009, Study 2), and the construct is operationalized in terms of the cognitive (rejection expectancies) and affective (rejection concerns/anxiety) components underlying the distinctive RS processing dynamic. Participants are presented with nine hypothetical interpersonal scenarios involving interactions with important others that demonstrate trigger situations hypothesized to activate anxious expectations of rejection (e.g., “You approach a close friend to talk after doing or saying something that seriously upset him/her”). For each scenario, participants indicate the extent to which they would be concerned/anxious about the possibility of rejection on a 6-point scale ranging from 1 (very unconcerned) to 6 (very concerned) and report their subjective estimates of the likelihood of rejection on a 6-point scale ranging from 1 (very unlikely) to 6 (very likely). Drawing from expectancy-value models (Bandura, 1986), RS scores are computed by first weighting participant’s rejection expectancies by their rejection concerns for each of the nine scenarios and then taking the average of the nine weighted scores. Such a scoring method is consistent with RS being conceptualized as a cognitive—affective disposition defined by anxious expectations of rejection, rather than either only expectations of rejection or only anxiety about rejection. In this way, individuals high in RS do not only expect rejection but are also highly anxious about and feel threatened by the possibility of rejection. Prior research suggests that the factor structure of the A-RSQ is best represented by Downey and Feldman (1996) theoretically derived one-factor measurement model (Preti et al., 2020). Moreover, the A-RSQ generates reliable composite scores and shows good criterion and incremental validity when predicting theoretically relevant social (e.g., aggressive behavior) and psychological outcomes (e.g., rejection attributions), even over and above related constructs, such as neuroticism, introversion, social anxiety, self-esteem, and attachment style (Downey & Feldman, 1996; Preti et al., 2020).

Intergroup Differences in RS

Meaningful intergroup differences in personality are well established and show, for example, robust differences in Big Five personality traits (e.g., neuroticism) across sex (Schmitt et al., 2008), sexual orientation (Allen & Robson, 2020), and age (Soto et al., 2011). However, little prior research has investigated intergroup differences in specific lower-order personality dispositions, such as RS, representing a noteworthy gap in the literature and obscuring potentially important insights about key mental health disparity populations. Given that RS originates from early interpersonal rejection, specific populations who are disproportionally burdened by such adverse social experiences might exhibit higher overall RS. These populations might include, for instance, women, sexual minorities (e.g., gay, lesbian, bisexual), and younger people. For example, women, relative to men, are at increased risk for childhood sexual abuse (Cutler & Nolen-Hoeksema, 1991), adolescent sexual harassment (Hand & Sanchez, 2000), relational aggression (Crick & Grotpeter, 1995), indirect forms of bullying (e.g., cyberbullying; Connell et al., 2014), and depression (Nolen-Hoeksema & Girgus, 1994). Relatedly, sexual minorities exhibit increased rates of sexual abuse, parental physical abuse, peer victimization, and stress sensitive mental health problems compared to their heterosexual counterparts (Flores et al., 2020; Friedman et al., 2011; King et al., 2008). Further, developmental research shows that younger people are disproportionately exposed to relatively novel forms of interpersonal rejection (e.g., cyberbullying) during critical periods of development (e.g., adolescence, emerging adulthood; Arnett, 2000) and report higher levels of cyberbullying (Kennedy, 2021), more frequent interpersonal tensions (Birditt et al., 2005), lower relationship quality and satisfaction (Birditt & Fingerman, 2003), more problematic and ambivalent relationships (Fingerman et al., 2004), and greater sensitivity to social evaluation (Somerville, 2013) than older people. Cross-temporal meta-analyses further demonstrate that self-reported levels of depression and anxiety symptoms are increasing across age groups, with younger age groups reporting higher levels than older age groups (Twenge et al., 2010).

Because the social antecedents of RS, namely early exposures to interpersonal rejection, are unequally distributed across social groups, group-level RS estimates might follow a similar trend. More specifically, intergroup differences in social experiences across sex, sexual orientation, and age might determine, in part, the content and valence of interpersonal information represented in memory, translating into socially patterned, trait-level differences in RS between groups, with important implications for population mental health. Prior research shows, for example, that maladaptive interpersonal cognitions and affects guide how people interpret and experience negative interpersonal events, increasing risk for emotional and behavioral maladjustment in the face of real or imagined interpersonal rejection (Ayduk et al., 2001; Gardner et al., 2020; Silvers et al., 2012; Zimmer-Gembeck et al., 2016). In this way, population-level differences in RS might contribute to mental health disparities in internalizing psychopathology among key mental health populations by sensitizing women, sexual minorities, and younger people to interpersonal rejection and amplifying its negative effects on mental health. Thus, by investigating intergroup differences in RS, the present research aims to identify whether population-level differences might emerge in one important individual difference factor that shapes both the onset and course of psychopathology.

To investigate intergroup differences in RS, one necessary prerequisite is to determine whether the most commonly used measures of this construct—the A-RSQ—is measuring the same construct across social groups, referred to as measurement invariance (Meredith, 1993). Measurement invariance testing is a process whereby increasingly constrained confirmatory factor analytic (CFA) models are fitted to determine the extent to which measurement parameters (e.g., factor loadings, item intercepts) are invariant across two or more groups (e.g., women vs. men). Configural, metric, and scalar invariance is achieved when the scale’s factor structure, factor structure and loadings, and factor structure, loadings, and intercepts are equal across groups, respectively (Brown, 2015). Critically, the validity of intergroup comparisons hinges on the assumption that the same construct is being measured across groups of interest, yet empirical support for such an assumption is rarely assessed. In terms of the A-RSQ, no extant research has empirically assessed whether the scale’s measurement properties are equivalent across sex, sexual orientation, or age. Without establishing measurement invariance, the extent to which trait-level differences across these groups reflect true population differences in RS or biased psychometric properties at the item level is unknown (Vandenberg & Lance, 2000). Consider, for example, items from the A-RSQ concerning interactions with parents (e.g., “You ask your parents or other family members to come to an occasion important to you”) or significant others (e.g., “After a bitter argument, you call or approach your significant other because you want to make up”). It is possible that older individuals, for whom parental relationships might be less important than marital relationships (Umberson et al., 2010), will interpret some items differently than younger individuals who affix more importance to their parental relationships. Relatedly, well-established differences in romantic selectivity across sex (Schacter et al., 2009) and in the availability and formation of romantic relationships across sexual orientation (Gates, 2014) might contribute to biased interpretations of items focusing on interactions with romantic partners among women and sexual minorities, leading to systematic differences in the measurement properties of these items across sex and sexual orientation and biased test scores.

The present study involves two primary aims. First, we sought to assess measurement invariance of the A-RSQ using a multigroup CFA framework across sex, sexual orientation, and age. To do this, we first assessed the factor structure of the A-RSQ by comparing the previously identified one-factor measurement model (Downey & Feldman, 1996, Study 1) to two alternative models: a correlated two-factor model and a bifactor model. Next, focusing on one demographic category at a time (e.g., sex), we examined a series of CFA models with increasingly stringent equality constraints concerning invariance, namely invariance of factor structure across groups (configural invariance); invariance of factor loadings across groups (metric invariance); and invariance of item intercepts across groups (scalar invariance). Second, we assessed intergroup differences in latent means across sex, sexual orientation, and age after establishing full or partial scalar invariance. Because some social groups appear to be less burdened by the social causes of RS (e.g., early and prolonged exposures to interpersonal rejection), we hypothesized that women (vs. men), sexual minorities (vs. heterosexuals), and younger people (vs. older people) would exhibit higher mean levels of RS, given that challenging interpersonal events are more likely to affect these groups. In addition, the present research aimed to assess measurement invariance of the A-RSQ and latent mean differences in RS across three different sociodemographic groups known to be uniquely associated with differential risk for mental health problems.

Method

Participants

Data were drawn from the Pathways to Longitudinally Understanding Stress (PLUS) study, which is an ongoing population-based, prospective cohort study examining longitudinal associations between stress and mental health among young adults (i.e., 18–35 years old at baseline) in Sweden. The study’s primary aim is to evaluate whether psychosocial mechanisms (e.g., social stress) account for the sexual orientation disparity in internalizing psychopathology among young adults. The baseline sample was recruited in 2019. Specifically, all respondents reporting a minority sexual identity in the 2015, 2016, and 2018 Swedish National Public Health Survey (n = 181,937) were invited to participate in the present study, and a matched heterosexual cohort was recruited by inviting a random sample of heterosexual respondents also from the 2015, 2016, and 2018 Swedish National Public Health Survey to participate in the present study. In total, 5,885 individuals were invited to participate in the PLUS study in 2019, with 2,548 providing informed consent and 2,226 providing adequate data. In the second wave of data collection in 2020, the A-RSQ was administered; therefore, the present sample consists of participants who completed the second wave of data collection. Of the 2,226 participants contacted to complete the second survey, 1,887 participants opened the survey. Of these, 15 failed to provide informed consent, two were ineligible for the study, and 90 completed less than 10% of the survey items. In addition, 101 repeat completers were detected and excluded from data analysis. In total, 208 participants were excluded from data analysis, and the final analytic sample was composed of the remaining 1,679 participants (Table 1). Table 1 reports sample characteristics. As a population-based sample, study participants are representative of Swedish residents aged 18–35 years old at Wave 1. Yale University’s institutional review board approved this study.

Table 1.

Sample Characteristics, N = 1,679

Sociodemographic variable	n or M (% or SD)
Age (n = 1,663)	27.12 (5.03)
Sex (n = 1,660)
Male	496 (29.9)
Female	1,164 (70.1)
Gender (n = 1,660)
Male	489 (29.5)
Female	1,128 (68.0)
Transgender male/man	10 (0.6)
Transgender female/woman	4 (0.2)
Genderqueer/gender nonconforming	20 (1.2)
Different identity	9 (0.5)
Sexual orientation (n = 1,679)
Lesbian or gay	179 (10.7)
Straight	951 (56.6)
Bisexual	446 (26.6)
Something else	51 (3.0)
I don’t know	52 (3.1)
Employment (n = 1,679)
Full-time	740 (44.1)
Part-time	377 (22.5)
Permanently or temporarily disabled	60 (3.6)
Unemployed	502 (29.9)
Education (n = 1,675)
Some high school	89 (5.3)
High school diploma or Graduate Equivalency Degree	699 (41.7)
Some college or Associate’ s degree	58 (3.5)
4-year college degree	166 (9.9)
Some graduate school	455 (27.2)
Master’ s degree or higher	208 (12.4)

Measures

Adult Rejection Sensitivity Questionnaire

The A-RSQ is a 9-item self-report measure developed to assess expectations of rejection and concerns of or anxiety about rejection among young adults (Berenson et al., 2009, Study 2). To complete the scale, participants read nine hypothetical interpersonal scenarios where rejection is a possible outcome (e.g., “You call a friend when there is something on your mind that you feel you really need to talk about”). For each scenario, participants are asked to use a 6-point scale to indicate their level of anxiety about the possibility of being rejected from 1 (very unconcerned) to 6 (very concerned) and to report their perceived expectations of being accepted from 1 (very unlikely) to 6 (very likely). Consistent with previous RS research (e.g., Ayduk, Gyurak, & Luerssen, 2008), acceptance expectancies were reverse scored to index rejection expectancies. To calculate item scores, a product term was created by multiplying rejection expectancies by rejection concerns to index anxious rejection expectancies, which served as the item score. To calculate total scores, we calculated the mean of the nine product terms across items. Total and item scores range from 1 to 36, with higher scores indicating more RS.

Demographic Information

In the present study, participants self-reported their sex assigned at birth, sexual orientation, and age. Sex assigned at birth was measured with a single item (“What sex were you assigned at birth?”). Participants responded to this question by selecting “Male” or “Female.” Sexual orientation was measured by a single item capturing self-identification (“Which of the following best represents how you think of yourself?”). Response options included: “Lesbian or gay,” “Straight, that is, not lesbian or gay,” “Bisexual,” “Something else,” and “I don’t know.” Age was measured by having participants select their age in years from a drop-down menu.

According to Meade et al. (2008), sample sizes of 400 per group are necessary to achieve adequate power to assess measurement invariance using multigroup CFA. As such, we specified, prior to analysis, the goal to create grouping variables for sexual orientation and age that resulted in sample sizes of at least 400 per group to ensure adequate power to detect lack of invariance, described in greater detail below.

Data Analysis

Analyses proceeded in several steps. First, missing data, descriptive statistics, and univariate and multivariate normality were examined for all study variables, as appropriate. Second, based on the work of Downey and Feldman (1996), a pooled, single-group CFA was conducted to examine whether the model fit of the previously identified, one-factor measurement model (Figure 1, Panel A) was appropriate for measurement invariance testing. Before proceeding to measurement invariance testing, model fit of the hypothesized one-factor model was compared to model fit of a correlated two-factor model (Figure S1, Top Panel) and a bifactor model (Figure S1, Bottom Panel).

Figure 1. Graphical Depiction of the Previously Identified Factor Structure of the Adult Rejection Sensitivity Questionnaire (A-RSQ) and Standardized Factor Loadings From the Present Study.

Note. The hypothesized one-factor model, depicted above, fit the data well and served as the baseline model for subsequent measurement invariance testing: Y–Bχ²(27) = 100.60, p < .001, CFI = .973, TLI = .964, RMSEA = .045, 90% CI [.040, .051], SRMR = .057. Y–Bχ² = Yuan–Bentler scaled chi-square statistic; CFI = robust comparative fit index; SRMR = standardized root mean square residual; TLI = robust Tucker–Lewis index; RMSEA = robust root mean square error of approximation.

Third, a series of increasingly constrained multigroup CFA models were fitted sequentially to evaluate measurement invariance across sex, sexual orientation, and age. For each demographic group, at least three multigroup CFA models were fitted and included the following equality constraints: factor structure (Model 1; configural invariance), factor loadings (Model 2; metric invariance), and item intercepts (Model 3; scalar invariance). Reductions in comparative fit index (CFI) of .010 or greater or increases in root·mean square error of approximation (RMSEA) of .015 or greater between nested models were used as criteria to identify lack of invariance, as recommended (Chen, 2007). Chi-square difference tests are reported for completeness but were not used as criteria for assessing measurement invariance, given their sensitivity to sample size and limited ability to differentiate invariant from noninvariant models reliably (Cheung & Rensvold, 2002).

Fourth, after establishing scalar invariance, latent factor means were freely estimated and statistically compared across groups via z-tests at an a level of .05, two-tailed. Latent means were computed by fixing one group’s latent mean to zero (reference group) and freely estimating the remaining group’s latent mean(s). Because the latent RS factor was scaled by fixing its variance to one in all CFA models, estimated differences in latent means are standardized by design and can be interpreted as effect size estimates (i.e., Cohen’s d).

For models not achieving full invariance at the metric or scalar levels, several steps were taken to examine partial invariance. First, noninvariant items were identified one at a time by inspecting univariate score tests to determine which item-level equality constraints should be freed. If univariate score tests revealed measurement parameters differing at p < .05 between groups, the parameter corresponding to the largest change in chi-square difference was selected to be freely estimated between groups. Second, partially invariant models were estimated by constraining all measurement parameters to be equal across groups, except for the parameter identified as noninvariant in the previous step. Notably, research suggests that partially invariant measurement models generate accurate structural estimates that can be compared between groups when variant parameters are freely estimated (Dimitrov, 2010). Thus, differences in latent means were assessed across groups for models achieving either full or partial scalar invariance.

All analyses were conducted in R, using the lavaan package (Rosseel, 2012). Given the ordered nature of the data, CFA models were fitted using polychoric correlations and robust diagonally least squares estimation with delta parameterization (Jöreskog, 1990). For all models, we report the following model fit indices: CFI, Tucker–Lewis index (TLI), RMSEA and its 90% confidence interval (CI), and standardized root mean square residual (SRMR). Conventional cutoff criteria for good model fit (e.g., CFI ≥ .950, TLI ≥ .950, RMSEA < .06, and SRMR < .08; Hu & Bentler, 1999) were developed using normal-theory maximum likelihood estimation and therefore do not apply as strict cutoff criteria for models using different estimation methods, such as robust diagonally least squares estimation. For completeness, we also report the Yuan–Bentler scaled chi-square test statistic (Y–Bχ²) and its associated p value for all CFA models.

Of the 1,679 participants included in the final analytic sample, 1,621 had complete data on the A-RSQ (~3% missing data). As participants did not consent to the sharing of individual-level data, these data cannot be made publicly available. The raw data that support the findings of this study, study materials, and analysis code are available on reasonable request from the corresponding author. This study was not preregistered.

Results

Descriptive Statistics

Table 1 reports sample characteristics, and Table 2 presents means (M), standard deviations (SD), skewness, and kurtosis of item and total scores. Item and total score skewness and kurtosis fell within ranges indicative of univariate normality (skewness: ±3, kurtosis: ±7; Kline, 2005), except for Item 3 whose kurtosis was elevated at 8.61. Mardia’s multivariate normality test indicated that the data were not multivariate normal, lending support to our use of a robust estimator. Total RS scores were adequately reliable: α = .79, ω = .80.

Table 2.

Descriptive Statistics for the 9-Item Adult Rejection Sensitivity Questionnaire (A-RSQ)

Item number	Mean (M)	Standard deviation (SD)	Skewness	Kurtosis
1. “You ask your parents or another family member for a loan to help you through a difficult financial time.”	4.86	6.72	2.59	7.00
2. “You approach a close friend to talk after doing or saying something that seriously upset him/her.”	11.38	8.69	0.80	−0.07
3. “You bring up the issue of sexual protection with your significant other and tell him/her how important you think it is.”	4.11	5.27	2.65	8.61
4. “You ask your supervisor for help with a problem you have been having at work.”	7.85	7.88	1.44	1.70
5. “After a bitter argument, you call or approach your significant other because you want to make up.”	8.16	8.08	1.47	1.83
6. “You ask your parents or other family members to come to an occasion important to you.”	5.23	6.86	2.36	5.98
7. “At a party, you notice someone on the other side of the room that you’d like to get to know, and you approach him or her to try to start a conversation.”	15.05	10.58	0.53	−0.81
8. “Lately you’ve been noticing some distance between yourself and your significant other, and you ask him/her if there is something wrong.”	10.98	8.40	0.97	0.43
9. “You call a friend when there is something on your mind that you feel you really need to talk about.”	6.27	7.26	1.91	3.64
Total score	8.21	4.84	0.96	0.96

Note. Item scores are Anxiety × Expectation product scores. Total scores are the average of the nine Anxiety × Expectation product scores.

Baseline Model

A single-group, one-factor CFA model was fitted and compared to two alternative models. Consistent with previous RS research (e.g., Downey & Feldman, 1996), item-level product terms were computed by weighting participant’s rejection expectations by their rejection concerns, which then served as congeneric indicators of the latent RS factor (Figure 1). For the two alternative models, rejection expectancies and rejection concerns were modeled as separate latent factors, and raw, item-level scores loaded onto their respective factor (Figure S1). Because the hypothesized one-factor model (Figure 1) utilizes a different set of indicators than the two alternative models (i.e., product scores vs. raw scores), these models are nonnested and cannot be compared using traditional inferential methods (e.g., chi-square difference tests). Thus, the hypothesized one-factor model was compared to the two alternative models descriptively.

The hypothesized one-factor model fit the data well: Y–Bχ²(27) = 100.60, p < .001, CFI = .973, TLI = .964, RMSEA = .045, 90% CI [.040, .051], SRMR = .057 (Figure 1). By contrast, the correlated two-factor model (Figure S1, Top Panel), Y–Bχ² (134) = 1694.85, p < .001, CFI = .908, TLI = .895, RMSEA = .088, 90% CI [.085, .090], SRMR = .088, and the bifactor model (Figure S1, Bottom Panel), Y–Bχ² (117) = 1527.77, p < .001, CFI = .916, TLI = .891, RMSEA = .089, 90% CI [.087, .092], SRMR = .084, fit the data poorly. As expected, the previously identified one-factor measurement model provided better fit to the data than the two alternative models and therefore served as the basis for subsequent tests of measurement invariance using multigroup factor analytic models (see Figure 1 and Figure S1 for standardized factor loadings).

Measurement Invariance and Latent Mean Differences

Table 3 presents goodness-of-fit indices for all multigroup CFA models testing measurement invariance across sex, sexual orientation, and age. When full invariance was not achieved at any of the assessed levels, partial invariance was investigated by identifying variant measurement parameters and allowing them to be freely estimated. After establishing full or partial scalar invariance, differences in latent means across groups were estimated. More detailed information is presented below, separated by group.

Table 3.

Goodness-of-Fit Statistics Associated With Tests of Measurement Invariance Across Sex, Sexual Orientation, and Age

Model number	Y-Bχ2 (df)	CFI	TLI	RMSEA	90% CI	SRMR	Comparison	Δχ2 (df)	ΔCFI	ΔRMSEA
Sex
1. Configural	117.21 (54)	.971	.961	.047	[.041, .053]	.060	—	—	—	—
2. Metric	165.05 (63)	.964	.959	.048	[.041, .056]	.067	2 vs. 1	17.65* (9)	−.007	.001
3. Scalar	199.89 (71)	.955	.954	.051	[.045, .058]	.070	3 vs. 2	49.56** (8)	−.009	.003
Sexual orientation
1. Configural	117.24 (54)	.970	.960	.047	[.041, .053]	.058	—	—	—	—
2. Metric	134.58 (63)	.969	.964	.044	[.038, .051]	.062	2 vs. 1	9.50 (9)	−.001	−.003
3. Scalar	156.73 (71)	.964	.963	.045	[.039, .052]	.063	3 vs. 2	29.14** (8)	−.005	.001
Age
1. Configural	128.62 (81)	.972	.962	.046	[.040, .052]	.061	—	—	—	—
2. Metric	182.08 (99)	.969	.967	.043	[.035, .052]	.070	2 vs. 1	20.83 (18)	−.003	−.003
3. Scalar	232.45 (115)	.957	.960	.048	[.041, .055]	.075	3 vs. 2	71.53** (16)	−.012	.005
4. Partial scalara	212.38 (113)	.963	.965	.045	[.037, .052]	.073	4 vs. 2	43.51** (14)	−.006	.002

Note. Y-Bχ² = Yuan-Bentler scaled chi-square statistic; df = degrees of freedom; CFI = robust comparative fit index; TLI = robust Tucker-Lewis index; RMSEA = robust root mean square error of approximation; 90% CI = RMSEA 90% confidence interval; SRMR = standardized root mean square residual.

^aInspection of modification indices revealed that the intercept of Item 6 (“You ask your parents or other family members to come to an occasion important to you”) differed significantly across age groups such that the intercept was systematically lower for people aged 30–36 (τ = 4.77) and 24–29 (τ = 5.99), relative to people aged 18–23 (τ = 6.70). In Model 4, the intercept of Item 6 was freely estimated for each age group.

^*p < .05.

^**p < .01.

Sex

The configural model (Model 1) provided adequate fit to the data, suggesting that the factor structure of the A-RSQ is equal across males (n = 496) and females (n = 1,164). The addition of equality constraints to the factor loadings did not result in appreciable changes in CFI or RMSEA, indicating that the meaning of the items is the same across sex (Model 2). Moreover, the inclusion of equality constraints to item intercepts provided good fit to the data and did not result in substantial changes in CFI or RMSEA (Model 3), demonstrating full scalar invariance across sex. Together, the A-RSQ achieved configural, metric, and scalar invariance across sex.

Having established full scalar invariance (Model 3), latent factor means were compared between males and females. Males served as the reference group, with their latent mean fixed to zero. As expected, results revelated that, compared to males, RS levels were significantly higher among females, τ = 0.290, 95% CI [0.171, 0.409], SE = 0.061, z = 4.778, p < .001.

Recognizing that some individuals’ gender identity differs from their sex assigned at birth (e.g., transgender people), a sensitivity analysis was conducted by recategorizing respondents on the basis of their self-reported gender identity, as opposed to their self-reported sex assigned at birth, and refitting the same set of models. Results remained virtually unchanged (Table S1).

Sexual Orientation

A two-level grouping variable was created for sexual orientation: heterosexuals (n = 951) and sexual minorities (i.e., those self-reporting their sexual identity as gay or lesbian, bisexual, or something else; n = 676). Participants who responded to the item assessing sexual orientation with “I don’t know the answer” were not included in the present analyses (n = 52, 3.1%).

The configural model (Model 1) provided adequate fit to the data, suggesting that the factor structure of the A-RSQ is equal across sexual orientation groups. The addition of equality constraints to the factor loadings did not result in appreciable changes in CFI or RMSEA, indicating that the meaning of the items is equal across sexual orientation groups (Model 2). The inclusion of equality constraints to item intercepts provided good fit to the data and did not result in substantial changes in CFI or RMSEA (Model 3), demonstrating full scalar invariance across sexual orientation groups. Overall, the A-RSQ achieved configural, metric, and scalar invariance across sexual orientation.

Having established full scalar invariance across sexual orientation (Model 3), latent factor means were compared between heterosexual and sexual minority participants. Heterosexuals served as the reference group, with their latent mean fixed to zero. As expected, results revelated that, compared to heterosexuals, RS levels were significantly higher among sexual minorities, τ = 0.532, 95% CI [0.407, 0.656], SE = 0.063, z = 8.377, p < .001.

Age

Based on the distribution of the age variable, a three-level grouping variable was created to split age into three nonoverlapping age groups: 18–23 (n = 477), 24–29 (n = 557), and 30–36 (n = 629) years old. These age groups are also broadly consistent with developmental theories identifying the age periods from 18 to 23, 24 to 29, and 30 to 45 years old as three differentiated developmental periods, namely emerging (18–23 years old), young (24–29 years old), and established (30–45 years old) adulthood, respectively (Arnett, 2000; Mehta et al., 2020).

The configural model (Model 1) provided adequate fit to the data, suggesting that the factor structure of the A-RSQ is equal across people aged 18–23, 24–29, and 30–36 years old. The addition of equality constraints to the factor loadings did not result in appreciable changes in CFI or RMSEA, indicating that the meaning of the items is equal across the three different age groups (Model 2). However, the addition of equality constraints to item intercepts did result in large reductions in CFI (Model 3; ΔCFI = −.012). This suggests that participants who belong to different age groups differ systematically in their responses to certain items of the A-RSQ. Univariate score tests revealed that releasing the equality constraint on the intercept of Item 6 (“You ask your parents or other family members to come to an occasion important to you”) would significantly improve model fit. A partially invariant scalar model (Model 4), with the intercept of Item 6 freely estimated and all other measurement parameters constrained to be equal, provided adequate fit to the data and did not result in large changes in CFI or RMSEA, relative to the metric model (Model 2). The partially invariant scalar model revealed that the intercept of Item 6 was higher among people aged 18–23 (τ = 6.70) and 24–29 (τ = 5.99) than among people aged 30–36 (τ = 4.77). Overall, the A-RSQ exhibited equal factor structure and factor loadings across age groups. The intercept of Item 6, however, differed significantly across age groups and therefore had to be freed for the scale to achieve partial scalar invariance.

Having established partial scalar invariance across age (Model 4), latent factor means were compared between participants in the 30–36 age group and participants in the 18–23 and 24–29 age groups. The 30–36 age group served as the reference group, with their latent mean fixed to zero. As expected, results revelated that, compared to those in the 30–36 age group, RS levels were significantly higher among those in the 18–23 age group, τ = 0.357, 95% CI [0.217, 0.497], SE = 0.071, z = 4.997, p < .001, and among those in the 24–29 age group, τ = 0.246, 95% CI [0.103, 0.388], SE = 0.073, z = 3.382, p = .001.

Because emerging adulthood is occasionally operationalized as the age period from 18 to 29 years old, a sensitivity analysis was conducted by collapsing the emerging (18–23 years old) and young (24–29 years old) adulthood groups into a single age group (18–29 years old) and refitting the same set of models. Results remained virtually unchanged (Table S1).

Discussion

According to social structural theories of personality development (Eagly & Wood, 1999), intergroup differences in personality might be socially determined and mediated by systematic variation in social status, socialization processes, and social experiences across groups. Complementing sociological perspectives of this type, social-cognitive conceptualizations of the personality system (e.g., CAPS theory; Mischel & Shoda, 1995) locate the source of personality differences across social groups in linkages among cognitions and affects stored in memory that might be common among individuals who share membership in a particular social group, reflecting within-group similarities in social experiences (e.g., exposure to interpersonal rejection). Yet, despite theoretical links among social group membership, shared social experiences, and the personality system, little is known about whether populations routinely shown to be repeatedly and disproportionally exposed to interpersonal rejection might be more likely to share a common network of maladaptive interpersonal cognitions and affects, as indicated by higher trait-level RS, than populations less burdened by such events. In seeking to address this issue, the present research sought to assess whether latent mean differences in RS, a personality type consistently shown to confer risk for poor mental health, emerged across sex, sexual orientation, and age after establishing measurement invariance in a large, population-based sample of Swedish young adults. Preliminary, pooled analyses indicated that the one-factor measurement model previously identified by seminal RS research (Downey & Feldman, 1996) provided good fit to the data and outperformed two alternative measurement models, which modeled the cognitive and affective components underlying the RS processing dynamic as two separate latent factors. Thus, in replicating prior work (Downey & Feldman, 1996), the one-factor model served as the baseline model for subsequent measurement invariance testing.

Across sex, sexual orientation, and age, the configural model was supported, indicating that the same latent factor structure was present in all three groups. Upon imposing equality constraints on the factor loadings, metric invariance was achieved across sex, sexual orientation, and age, confirming that the unit of measurement for RS is comparable for males and females, sexual minorities and heterosexuals, and people 18–23, 24–29, and 30–36 years old. Regarding scalar invariance, the A-RSQ demonstrated full scalar invariance across sex and sexual orientation, establishing that latent factor means reflect unbiased estimates of RS and are statistically comparable between these groups. Across age, however, full scalar invariance was not observed. The intercept of Item 6 (“You ask your parents or other family members to come to an occasion important to you”) differed systematically across age groups such that intercept values decreased linearly, with the youngest and oldest age group having the largest and smallest intercept value, respectively. Because the importance people affix to different social relationships varies across the life course (Umberson et al., 2010), systematic age differences in responses to Item 6 likely stem from the fact that those in the youngest age group (18–23 years old) might rely heavily on their parents to maintain their belonging needs, whereas those in the older age groups (24–29, 30–36 years old) might depend on different social relationships (e.g., spouse, child) to meet their belonging needs. Although full scalar invariance was not achieved across age, partially invariant models with fewer than 20% of scale items freely estimated between groups are robust and can be used to anchor mean-level comparisons (Dimitrov, 2010). Thus, results of measurement invariance testing confirm that intergroup differences in latent means across sex, sexual orientation, and age reflect true trait-level differences in RS between groups, rather than differences in measurement at the item level.

Establishing full or partial scalar invariance of the A-RSQ across sex, sexual orientation, and age laid the foundation for the primary aim of the present research—to examine differences in latent factor means across social groups. Consistent with study hypotheses, women, sexual minorities, and individuals 18–23 and 24–29 years old exhibited significantly higher RS levels than men, heterosexuals, and individuals 30–36 years old, respectively. As the first empirical tests of intergroup differences in RS across sex, sexual orientation, and age, these findings advance research and theory on the social patterning of RS and its role in understanding differential risk for mental health problems among vulnerable populations. First, given that RS is conceptually related to neuroticism (Downey & Feldman, 1996), findings from the present research complement prior research showing robust differences in neuroticism across sex (Schmitt et al., 2008), sexual orientation (Allen & Robson, 2020), and age (Soto et al., 2011). Our findings extend this work by establishing that intergroup differences in RS follow the same population-level trends as neuroticism and by demonstrating, via measurement invariance testing and latent variable modeling, that these differences cannot be explained by measurement bias or error.

Second, influential models of the personality–psychopathology link identify neuroticism as a key contributor to the development of negative cognitive biases and to the onset and course of internalizing psychopathology (e.g., depression; Klein et al., 2011). Because previous work shows that RS is a lower-order facet of neuroticism (Downey & Feldman, 1996), results suggest that individual differences in RS might at least partially explain previously identified trait-level differences in neuroticism across sex, sexual orientation, and age. However, one limitation of trait theory perspectives on the relationship between personality and mental health is that commonly used trait models (e.g., Big Five, Big Three; Goldberg, 1990; Markon et al., 2005) are mostly atheoretical and nonexplanatory, failing to identify the source of traits and how they influence mental health. By focusing on RS, a personality type consistently shown to derive from the social environment and to bias social information processing (Ayduk & Gyurak, 2008), results provide indirect evidence for the role of environmental factors (e.g., social learning) and social-cognitive processes in the emergence of population-level differences in personality types and mental health. One avenue for future research is to utilize bifactor modeling to assess the relative contribution of RS (i.e., chronic, anxious expectations about rejection) versus trait neuroticism (i.e., a global tendency toward negative affect) to predicting cognitive biases in various aspects of information processing (e.g., memory, interpretation) and risk for the onset and maintenance of mental health disorders within and between social groups.

Third, in terms of age differences, our findings align with the maturity principle (Bleidorn et al., 2013) which contends that, as people age, socially desirable personality traits (e.g., agreeableness) increase and socially undesirable traits (e.g., neuroticism) decrease continuously. Because RS constrains prosocial behavior and self-regulatory capacity (e.g., Ayduka, Gyurak, & Luerssen, 2008), patterns of mean-level changes in RS across differentiated developmental periods might mirror those found for other socially undesirable traits (e.g., neuroticism) by decreasing as people mature and acquire increasingly more responsibilities and social skills (Soto et al., 2011). At a population level, this hypothesis is further supported by the work of Twenge et al. (2010, 2012), which finds that age differences in personality might reflect adaptations to time-varying social norms, pressures, and restrictions. Consequently, people of the same or similar age, having developed within the same larger sociocultural environment, might have significant overlap in their social learning histories and resultant knowledge structures (e.g., network of cognitions and affects; Mischel & Shoda, 1995), given that exposure to some social experiences (e.g., cyberbullying) differs markedly by developmental period. Thus, age differences in RS might be determined, in part, by social and cultural shifts that facilitate differential early learning processes and exposure to adverse interpersonal events across age groups, consistent with findings from the present study.

Fourth, although women, sexual minorities, and young adults represent three populations disproportionally burdened by mental health problems, the origins of such population-level differences in mental health remain somewhat unclear. To date, a number of different mechanisms have been proposed to explain why some social groups might be more prone to poor mental health than others, such as intergroup differences in exposure to social stress (Klonoff et al., 2000; Meyer, 2003), genetic and biological factors (Diamond et al., 2021; Jacobson & Rowe, 1999), emotional reactivity (Hyde et al., 2008; Nolen-Hoeksema, 2012), and social isolation (Hidaka, 2012). Relatively less attention, however, has been paid to specific personality dispositions and associated social-cognitive factors. This lack of attention is noteworthy, given that influential theories of depression (Beck, 1967), anxiety (Bar-Haim et al., 2007), and suicidality (Baumeister, 1990) identify individual differences in information processing as a core cognitive vulnerability to psychological problems. By showing that key mental health disparity populations tend to exhibit higher anxious expectations of rejection than populations less affected by mental health problems, results provide preliminary evidence that information processing dynamics might differ stably between groups due to differences in the availability and accessibility of rejection-related beliefs and affects in memory (Mischel & Shoda, 1995). It is possible that elevated RS among women, sexual minorities, and younger people—perhaps itself a function of disproportionately accumulated stress exposure experienced by these populations (Stroud et al., 2018)—might contribute to population-level differences in mental health by exacerbating the link between exposure to negative interpersonal events and internalizing psychopathology on an ongoing basis (e.g., Ayduk et al., 2001). Yet, whether and how systematic differences in the on-line processing and interpretation of social information might contribute to well-established disparities in mental health among women, sexual minorities, and young adults awaits future research.

Despite the methodological strengths of this large population-based study, results should be interpreted in light of several limitations. First, this study was conducted in Sweden, one of the wealthiest and most equitable countries in the world, which might constrain the transferability of our findings to other countries or cultures. Recognizing that the majority of RS research has been and continues to be conducted in the United States, it is also worth noting that the psychometric properties of the A-RSQ might differ slightly across countries. Second, alternative modeling techniques exist to assess measurement invariance, such as item response models, and might generate different results. Future research might consider formally testing these possibilities by examining the item-level measurement properties of the A-RSQ across countries and by replicating the current findings via item response models in a different sample. Third, limitations stemming from data-coding decisions regarding the categorization of sexual orientation and age warrant mention. Grouping variables were created with practical constraints in mind, namely that groups with a sample size smaller than 400 reduce power to detect noninvariant models substantially (Meade et al., 2008). Therefore, to achieve sample sizes of 400 or more per group, a two-level grouping variable was created to categorize sexual orientation (i.e., heterosexual and sexual minority), and age was split into three nonoverlapping categories (i.e., 18–23, 24–29, and 30–36 years old), in line with influential theories of human development (Arnett, 2000; Mehta et al., 2020). Even though more granular divisions would have resulted in reduced power to detect lack of invariance and sacrificed the validity of the results, we acknowledge that the two- and three-level categorizations used in the present study might limit the results by obscuring important differences among sexual minority subpopulations (e.g., bisexual vs. lesbian) and across more narrowly defined developmental periods. Sample size permitting, future research should further assess the measurement and structural properties of the A-RSQ across sexual orientation and age using more fine-grained group divisions.

In terms of implications for future RS research, results from the present research suggest that the majority of A-RSQ items perform equally across sex, sexual orientation, and age, with the exception of Item 6. In light of this noninvariant item, we advise researchers to consider utilizing latent measurement models to correct for measurement error and to account for the noninvariance of items, especially when their samples are diverse in terms of age. This recommendation stems from the fact that noninvariant items bias composite test scores (Brown, 2015). Therefore, the use of latent variable models would result in more precise assessments of RS by statistically differentiating between the structural (e.g., latent mean) and measurement (e.g., factor loadings) properties of the assessed construct. An additional advantage of latent measurement models is that they can be incorporated into more complex structural models, specifying directional relationships among RS and external constructs. By adopting these recommendations, intergroup differences in RS and items exhibiting biased psychometric properties, such as those identified in the present research, are less likely to influence structural parameter estimates, increasing the verity of results about the impact of RS on social behavior and mental health.

In conclusion, with the goal of identifying whether RS might vary across socially salient groups, we first assessed measurement invariance of the A-RSQ and then examined latent mean differences in RS across sex, sexual orientation, and age. Results revealed that the measurement properties of the A-RSQ are fully invariant across sex and sexual orientation and partially invariant across age. Tests of differences in latent means indicated that lower status social groups disproportionally exposed to the social causes of RS across the life course (e.g., parental rejection, peer victimization) exhibit higher levels of RS than higher status social groups, as expected. Findings from the present research highlight the utility of attending to group differences in maladaptive personality dispositions and information processing styles and their potential role in contributing to persistent mental health hardships uniquely affecting women, sexual minorities, and younger people.

Public Significance Statement.

The Adult Rejection Sensitivity Questionnaire measures rejection sensitivity (RS), a personality type marked by highly anxious expectations about rejection, in the same way across sex and sexual orientation but slightly differently across age groups. RS is higher among women, sexual minorities, and younger people than among men, heterosexuals, and older people.