According to the vulnerability-stress-adaptation model of marriage, marital quality:

Introduction

Because people commit to, and abandon, intimate partnerships with great regularity, hundreds of studies aim to explain how social bonds that were once fulfilling become sources of frustration and regret. Thoughtful application of observational methods characterizes much of this work, thereby illuminating how a wide range of complex interpersonal processes — including negative reinforcement and coercion [e.g., Jacobson and Margolin, 1979], lack of accommodation in response to relationship-threatening actions [e.g., Rusbult et al., 1991], insensitive responses to disclosures [e.g., Cutrona, 1996], and dysregulated emotion [e.g., Gottman and Levenson, 1986] — contribute to later relationship distress. Though unusually generative, at least two major blindspots can be identified in this literature. First, in the vast majority of studies, samples are comprised of couples who are white, well-educated, relatively affluent, and financially stable [Williamson et al., 2021], thus limiting the population to whom findings can be generalized and potentially constraining the effects of preventive and therapeutic interventions that are based on this work [e.g., Heyman et al., 2019]. Second, heavy emphasis on interpersonal processes has overshadowed consideration of plausible alternative explanations for how relationships change. Some argue, for example, that interpersonal processes are simply a manifestation of potent up-stream individual differences [e.g., personality traits, early experiences with caregivers, parental divorce] which are themselves the root causes of relationship dysfunction [e.g., Bowlby, 1979; Kelly and Conley, 1987; Amato, 1996]. Others look outside the relationship, proposing that normative life events [e.g., the transition to parenthood] and chronic and acute stress [e.g., discrimination, illness, work demands, financial concerns, parenting] undermine relationship maintenance and increase risk for distress and dissolution [e.g., Conger et al., 1990; Bodenmann, 2005]. The goal of this work is to address these two blindspots directly, by examining the independent effects of interpersonal processes, individual difference factors, and stress on the relationship satisfaction trajectories experienced by a diverse set of couples recruited from communities characterized by lower incomes.

With the Vulnerability-Stress-Adaptation [VSA] Model, Karney and Bradbury [1995] sought to integrate these three broad domains of influence, hypothesizing that relationship outcomes are governed by [a] adaptive processes, or dyad-level patterns of behavioral interdependence, [b] partners’ enduring vulnerabilities [i.e., the early experiences and stable traits that a person would bring to any relationship], and [c] stress [i.e., the chronic circumstances and acute events that demand partners’ regulatory efforts]. The VSA model further asserts that enduring personal vulnerabilities and stress will exert only indirect effects on relationship outcomes, as their influence is expected to be routed through the behavioral interdependencies and coordinated actions of the two partners. The VSA model therefore assigns interpersonal processes a central role in determining the course of couple relationships, while also assuming that couples will vary in their adaptive processes because of partners’ vulnerabilities and the stressful circumstances to which couples are exposed. According to this view, marriages become less rewarding over time not simply because of the adverse traits and experiences that partners bring to the relationship, and not simply because of the stresses and strains that partners encounter once in the relationship, but because vulnerabilities and stress outstrip partners’ capacity to create an adaptable and appropriately responsive behavioral repertoire that is necessary to support positive appraisals of their relationship.

Extensive research supports basic bivariate paths within the VSA framework. Consistent with model predictions, for example, meta-analyses and large-scale studies corroborate expected main effects of enduring vulnerabilities [e.g., negative emotionality, Malouff et al., 2010; parental divorce, Dronkers and Härkönen, 2008], stress [Randall and Bodenmann, 2009], and couple communication [Woodin, 2011] on relationship quality and longevity [for recent reviews and analyses, see McNulty et al., 2021; Righetti et al., 2022]. Other studies document associations among the three main model components. For example, partners with divorced parents and higher levels of impulsive and neurotic traits display more negative behaviors in couple interactions [e.g., Sanders et al., 1999; Weiss et al., 2018], and partners exposed to higher levels of stress are less effective and more negative when providing social support [Bodenmann et al., 2015] and resolving relationship differences [Williamson et al., 2013]. And, at least when considered separately, newlyweds’ enduring vulnerabilities [viz., neuroticism, trait anger, and low self-esteem], stress, and adaptive processes displayed during a relationship problem-solving task all predict relationship outcomes 4 and 10 years later [Lavner and Bradbury, 2010]. In principle, the VSA model applies to couples regardless of their ethnicity and socioeconomic standing, though in practice the above links are rarely tested with minoritized couples and couples living with low incomes.

Our first aim is to test whether the three main VSA model components make unique and independent contributions to marital outcomes, using a diverse sample of newlywed couples. Although the above-noted bivariate associations appear to support this assumption of independent contributions, plausible alternative models include the possibility that one domain will override the predictive effects of all other domains, either because a domain encompasses developmental phenomena that are broadly influential in individuals’ lives [e.g., traits and vulnerabilities, as proposed by Kelly and Conley, 1987] or because a domain appears to be most causally proximal to judgments of relationship satisfaction [e.g., adaptive processes; as suggested by Jacobson and Margolin, 1979]. Examining the relative contributions of VSA model components has important implications for social policies aimed at helping families: e.g., if associations between stress and relationship outcomes are eliminated after controlling for adaptive processes, then a stronger case can be made for couple communication as the primary target of large-scale intervention efforts. In contrast, if stress predicts relationship outcomes independent of adaptive processes, then outside demands on couples’ resources also emerge as a viable target for strengthening committed relationships—particulary if the expected association between adaptive processes and relationship satisfaction falls to nonsignificance.

While enduring vulnerabilities and stress appear to contribute to between-couple variance in relationship processes and outcomes, fewer studies test the critical VSA assumption that relationship processes will mediate these associations. Evidence from smaller studies supports this mediational role of adaptive processes [e.g., Story et al., 2004; Fisher and McNulty, 2008] as does a synthesis of ten longitudinal studies [representing 1104 couples] conducted by McNulty et al. [2021], which showed that higher levels of attachment anxiety predicted slower declines in satisfaction via positive effects of attachment anxiety on observed behavioral engagement. Building from these results, our second aim is to test the assumption that adaptive processes mediate effects of enduring vulnerabilities and stress on relationship outcomes, a possibility tested against the alternative view that partners’ satisfaction reports are responsive to all three predictor domains, with no mediation by adaptive processes. The latter finding would be consistent with the idea that stress directly alters perceptions of relationship quality by elevating expectations of what constitutes a fulfilling relationship or that early experiences with parental divorce or caregiving lead partners to appraise their relationships in fundamentally different ways.

A third source of ambiguity within the VSA model concerns its generalizability across couples with varying social and economic resources. VSA model paths are assumed to be broadly applicable across the economic spectrum [Karney and Bradbury, 1995], yet the model makes no special provisions for couples’ socioedemographic risk status, and the intersection between such risk factors and VSA model predictions remains undeveloped. This may be a significant oversight, insofar as relationship distress and dissolution are known to be more common among couples with fewer years of formal education, lower incomes, greater use of government services, and greater social isolation [Bramlett and Mosher, 2002; Copen et al., 2012; Karney, 2021; Pietromonaco and Overall, 2022]. How might the VSA model accommodate couples’ social and economic disparities? One possibility is that, regardless of sociodemographic risk, all couple relationships are governed by enduring vulnerabilities, stress, and adaptive processes in essentially the same way. While couples living with lower incomes, fewer years of formal education, or fewer close social ties might enter their relationships with more enduring vulnerabilities or might encounter more stress, for example, the magnitude of the paths linking vulnerabilities, stress, and adaptive processes to couple outcomes might be uniform across levels of risk. However, such a prediction overlooks a second possibility, that couples with differing levels of sociodeomographic risk may be qualitatively distinct from one another in how their relationships develop. Indeed, abuse in childhood, an enduring vulnerability, exerts stronger effects on relationship distress among women with less social capital [Liang et al., 2006], and the transition to parenthood, a normative stressor, exerts stronger effects on satisfaction among couples with lower incomes [Doss et al., 2009]. If the effects of model components on satisfaction are moderated by sociodemographic risk, as these studies suggest, then the VSA model could be revised in ways that capture the experiences of a wider spectrum of couples. Our third aim is to distinguish between these two possibilities, by testing the moderating role of sociodemographic risk on paths linking enduring vulnerabilities, chronic stress, and adaptive processes to relationship satisfaction.

The present study

To test our three main aims, we collected self-report data on three prominent enduring vulnerabilities [abuse history, experiences in the family-of-origin, and symptoms of depression]; self-report data on several common forms of chronic stress [e.g., stress arising from close friendships, family relationships, work, finances, and health]; and observational data from three 8-minute, in-home couple interaction tasks that were later coded for positivity, negativity, and communication effectiveness. Although observational data on communication behaviors may not provide a fully comprehensive operationalization of adaptive processes, we rely on them here because communication behaviors are a robust correlate of relationship satisfaction [Woodin, 2011], and because observational data reduce shared method variance when compared to self-report assessments of comparable constructs. To assess the potential moderating role of sociodemographic risk, we employed a 10-item index developed by Amato [2014] that defines risk in terms of partners’ age at marriage, levels of formal education, employment status, income, use of public assistance, and access to others during an emergency. To address concerns about the lack of racial and economic diversity in studies of marriage, we used census data to sample couples living in communities with low incomes, drawing from public marriage licenses to recruit first-time newlywed couples; doing so also helps to ensure sufficient variability on the risk index, which is critical for effective testing of our generalizability hypothesis [Aim 3]. Finally, 4 self-report assessments of relationship satisfaction, collected at 9-month intervals, were conducted to enable longitudinal analysis of the first 27 months of marriage. Multiple waves of satisfaction data allow us to separate between-person levels of satisfaction from within-person changes in satisfaction, and because most findings in the literature are between-person effects [e.g., Lavner et al., 2014], we expect level [intercept] effects to predominate here as well.

Materials and methods

Participants

The 431 newlywed couples [862 spouses] recruited to participate had been married an average of 4.8 months; 38.5% entered marriage with children. Men and women averaged 27.9 [SD = 5.8] and 26.3 [SD = 5.0] years of age. The mean household income for the sample at baseline in 2009 was $51,290 [SD = $27,711], while the median household income was $47,500, with incomes ranging from $2,500 to $100,000+ per year. Note that the median household income for Los Angeles County residents in 2009 was $54,375 [U.S. Census Bureau], suggesting that our sampling procedure successfully included a range of couples with varying socioeconomic statuses. Couples were Black [12%], White [12%], and Hispanic [76%], in proportions roughly comparable to those of people living in poverty in Los Angeles County [12.9%, 14.7%, and 60.5%, respectively; United States Census Bureau, 2002]. Missing and incomplete data, in the interaction task and/or the self-report questionnaires, reduced the initial sample of 431 couples by 43 couples, leaving 388 couples [776 spouses] available for analysis, 36.3% of whom entered marriage with children. For comparions of couples entering marriage with and without children, see Lavner et al. [2020].

Procedure

Recruitment

Sampling was undertaken to yield a group of first-married newlywed couples of the same ethnicity, living in low-income neighborhoods. Recently married couples were identified through names and addresses on marriage license applications. Addresses were matched with census data to identify applicants living in low-income communities, defined as census block groups wherein the median household income was no more than 160% of the 1999 federal poverty level for a four-person family. Next, names on the licenses were weighted using data from a Bayesian Census Surname Combination [BCSC], which integrates census and surname information to produce a multinomial probability of membership in each of four categories [Hispanic, Black, Asian, White/Other]. Couples were selected from the population of recently married couples using probabilities proportionate to the ratio of target prevalence to the population prevalence, weighted by the couple’s average estimated probability of being Hispanic, Black, or White, which are the three largest groups among people living in poverty in Los Angeles County [United States Census Bureau, 2002]. These couples were telephoned and screened to ensure that they had married, that neither partner had been previously married, and that both spouses identified as Hispanic, Black, or White.

Assessments

Couples were visited in their homes four times, once every 9 months, by two trained interviewers who described the IRB-approved study and obtained written informed consent from each participant. Interviewers took spouses to separate areas in their home to orally administer self-report measures at baseline [T1], and identical procedures were used 9 months [T2], 18 months [T3], and 27 months [T4] after baseline. Couples who reported that they had divorced or separated did not complete the interview. Following each interview, couples were debriefed and paid $75 for T1, $100 for T2, $125 for T3, and $150 for T4. Data collection occurred between 2009 and 2013.

Following their individual interviews, partners were reunited for three 8-min discussions. Discussions took place in a location of the couples’ choosing that would enable them to talk privately and without interruption. The first two discussions used procedures designed to assess social support behaviors [Pasch and Bradbury, 1998]. One randomly chosen spouse was asked to “talk about something you would like to change about yourself” while the partner was instructed to “be involved in the discussion and respond in whatever way you wish.” Spouses were asked to avoid sources of tension within the relationship. After a short break, a second discussion was held that was identical to the first discussion, with the roles reversed. Common topics included losing weight, making a career change, and dealing with stress. For the third interaction, partners were asked to identify a topic of disagreement in their relationship and to then devote 8 min working toward a mutually satisfying resolution of that topic. Common topics included management of money, chores, communication, and spending time together as a couple.

Observational coding

Videotapes were scored by 16 trained coders using the Iowa Family Interaction Rating Scales [IFIRS; Melby et al., 1998]. Coders—five of whom were native Spanish speakers—coded only in their native language. The IFIRS gives each participant a single score for each behavioral code at the end of the discussion. This score is determined by the coder based on the frequency and intensity with which the participant exhibits the verbal and nonverbal behavior described in the code. The scores range from 1 to 9, with a score of 1 indicating that the behavior did not occur. In general, a score of 3 indicates that “the behavior almost never occurs or occurs just once and is of low intensity,” a score of 5 means “the behavior sometimes occurs and is at a low or moderate level of intensity,” a score of 7 means that “the behavior occurs fairly consistently or is of elevated intensity,” and a score of 9 means “the behavior occurs frequently or with significant intensity” [Melby et al., 1998, pp. 7–8.]

Coders were trained for 10 hours per week for 3 months and were required to pass written and viewing tests with 80% accuracy level before coding, using criterion scores determined by the developers of IFIRS. Coders participated in 2 h of continuing training each week [e.g., coding a tape as a group and watching examples of specific codes] to minimize drift and ensure continued fidelity to IFIRS codes. Coders viewed each interaction three or four times using the Noldus Observer XT coding software, using the built-in capabilities to note behaviors of both spouses. After viewing an interaction, coders used their recorded notations to tabulate the frequency and intensity of each type of behavior and used this information to assign a score for each spouse for each code, using criteria from the IFIRS coding manual [Melby et al., 1998].

To assess reliability, 20% of videos were coded by two randomly chosen coders. Their scores were compared, and any discrepancies greater than one point were resolved by coders working together. Thus the final set of scores used in analyses for the reliability tapes included scores that matched across the two coders during their initial individual coding [when codes were off by 1 point, the score from the randomly designated “primary coder” was used]; discrepant scores were replaced by the scores from the second joint coding.

Measures

Enduring vulnerabilities

Each spouse’s vulnerabilities were defined as a latent variable with three indicators: abuse history, environment in the family-of-origin, and negative emotionality. Abuse history was conceptualized as spouses’ experience with maltreatment, which included sexual abuse and physical abuse. Individuals were identified as having a history of physical abuse with three self-report items assessing victimization in different contexts. E.g., “Before you turned 18, were you ever hit, beaten up, burned, assaulted with a weapon, or had your life threatened by a member of your family?” Individuals were identified as having a history of sexual abuse from items assessing victimization before as well as after the ages of 18 [i.e., “did anyone—a stranger, friend, acquaintance, date, or relative—ever try or succeed in doing something sexual to you or make you do something sexual to them against your wishes?”]. Type of abuse was summed into one global measure of abuse history with scores ranging from 0 to 3. Family environment in childhood was assessed with three items measuring the closeness and happiness of the family prior to age 14 [e.g., “My parents’ relationship would be a good example to follow for any married couple,” “The members of my family were always very close to each other”], with 0 = true and 1 = false, as well as a fourth item assessing whether participants’ parents were divorced/separated versus intact. Scores for the four items were summed to form the family environment score for each participant, with an alpha of 0.67 for husbands and 0.66 for wives. Symptoms of depression, an indicator of negative emotionality, were assessed by summing 9 items from the CESD [Radloff, 1977] measuring dysphoria, sadness, and depression over the past 30 days [e.g., “During the last 30 days, about how often did you feel so sad that nothing could cheer you up?”], with an alpha of 0.75 for husbands and 0.73 for wives.

Stress

Chronic stress was assessed by items identifying typical conditions in each of eight domains: close friendships, intimate relationship, extended family relationships, relationships with children, work, finances, health of self, and health of family members, which were summed to create an overall chronic stress variable [Hammen et al., 2012]. Husbands’ and wives’ scores were then summed to form the couple level chronic stress score. The majority of couples were not administered the chronic stress measure at Time 1 [n = 330]; the chronic stress measure was administered to Black and White couples at Time 1 but only to Hispanic couples starting at Time 2. For couples who did provide chronic stress data at Time 1 [n = 101], their reports of chronic stress at Time 1 and Time 2 were correlated [husbands’ r = 0.56, p < 0.001, wives’ r = 0.66, p < 0.001], suggesting that this measure of chronic stress is indeed capturing a relatively stable phenomenon, consistent with our conceptualization of stress as chronic. Given the reasonably stable nature of the perceived stress measured by this variable, for couples with missing data, Time 2 reports of chronic stress were imputed into the their Time 1 data.

Adaptive processes

Couple adaptive processes were measured by six observed indicators of communication during an interactional task at Time 1: husband and wife positivity, husband and wife effectiveness, and husband and wife negativity [reverse-coded]. A composite positivity scale was created by averaging an individual’s scores on the group enjoyment, positive mood, warmth/support, physical affection, humor/laugh, endearment, and listener responsiveness codes in each task, and then averaging across the three tasks; ICC = 0.83 for husbands, 0.81 for wives. A composite effectiveness scale was created by averaging an individual’s scores on the assertiveness, communication, effective process, solution quality, and solution quantity in each task, and then averaging across the three tasks; ICC = 0.74 for husbands, 0.80 for wives. A composite negativity scale was created by averaging an individual’s scores on the hostility, disruptive process, contempt, denial, angry coercion, dominance, verbal attack, interrogation, and externalized negative codes in each task, and then averaging across the three tasks; ICC = 0.73 for husbands, 0.74 for wives. The negativity score was reverse-coded in order to be included in the latent variable along with positivity and effectiveness.

Relationship satisfaction

Relationship satisfaction, conceptualized as spouses’ global sentiment toward the relationship, was assessed by summing responses on an eight-item questionnaire. The measure was adapted from Rauer et al. [2008] and included items from the General Social Survey [Davis et al., 2006]. It has been used in large surveys with low-income couples [e.g., Rauer et al., 2008] and has been shown to covary systematically with observed communication, thus lending support to its validity as an indicator of relationship functioning [Williamson et al., 2013]. Five items asked how satisfied the respondent was with various areas of their relationship [e.g., time spent together] and were scored on a 5-point scale [1 = Very dissatisfied, 5 = Very satisfied]. Three items asked to what degree the participant agreed with a statement about their relationship [e.g., “how much do you trust your partner?”] and were scored on a 4-point scale [1 = Not at all, 4 = Completely]. Scores could range from 8 [very dissatisfied] to 37 [very satisfied]. Coefficient α exceeded .70 for husbands and wives at all assessments.

Sociodemographic risk

Sociodemographic risk, our hypothesized moderator variable, was assessed using a 10-item index developed by Amato [2014]. Couples were given 1 point for the presence of each of the following items: [a] either partner was under the age of 23, [b] husband had less than a high school education, [c] wife had less than a high school education, [d] husband was unemployed, [e] wife was unemployed, [f] couple’s income was below the poverty line, [g] husband was receiving public assistance, [h] wife was receiving public assistance, [i] husband reported no one to help in an emergency, and [j] wife reported no one to help in an emergency. Prior work shows that this index moderates the effects of skills training on observed communication 18 months later [Williamson et al., 2016].

Analytic plan

All three aims were tested using latent variable structural equation models [SEM], as SEM allows for the creation of latent variables using multiple measured variables as indicators [thereby accounting for the measurement error in each of the observed variables and yielding more precise regression coefficients]. In addition, SEM accounts for the dependency between spouses and allows for all independent and dependent variables to be tested simultaneously [thereby estimating the effects of each variable over and above all others]. We obtained maximum likelihood estimates of the model coefficients using MPlus Version 7.0.

We evaluated three structural equation models. In the first, three latent factors [husband vulnerabilities, wife vulnerabilities, and couple adaptive processes], and one observed variable [couple chronic stress], were used to predict the intercepts and slopes of relationship satisfaction, simultaneously for husbands and for wives. In the second model, we examined whether adaptive processes would mediate effects of vulnerabilities and stress on satisfaction. The third model was similar to the first but added the index of sociodemographic risk, as an observed variable, to determine whether associations between the three predictor variables [vulnerabilities, stress, and adaptive processes] and the two dependent variables [satisfaction intercepts and slopes] varied systematically as a function of risk.

Results

Descriptive statistics and preliminary analyses

Descriptive statistics for all study variables, and their estimated loadings on our three latent factors, are shown in Table 1, and correlations among our primary study variables are shown in Table 2. Correlations among the three latent variables in the model were generally small but in expected directions [r = |0.17–0.26|]. Adaptive processes were unrelated to chronic stress, a couple-level measured variable [r = −0.04, ns], whereas enduring vulnerabilities did covary with chronic stress, for husbands [r = 0.53, p < 0.001] and for wives [r = 0.54, p < 0.001]. When considering VSA model components in relation to the index of sociodemographic risk, Table 2 shows that people with fewer social and economic resources enter marriage with more enduring vulnerabilities [r = 0.36 for husbands, p < 0.001; r = 0.16 for wives, p < 0.05] and poorer adaptive processes [r = −0.39, p < 0.001] and more chronic stress [r = 0.12, p < 0.01]. It should be noted that variability in sociodemographic risk in our sample [SD = 2.12] was significantly higher [F[387, 7923] = 1.37, p < 0.0001] than the variability originally reported in Amato’s [2014] sample [SD = 1.81].

Table 1. Descriptive statistics and factor loading estimates for study variables.

Table 2. Correlation matrix for latent variables and observed variables in final model.

Two important points can be drawn from these analyses. First, evidence that chronic stress is unrelated to couple communication [r = −0.03, ns] is at odds with VSA model predictions. However, sociodemograpic risk and couple communication do covary reliably [r = −0.38, p < 0.001], indicating that there may be contextual constraints on adaptive processes [or that poorly communicating couples select themselves into difficult circumstances]. Second, evidence that chronic stress and sociodemographic risk are weakly associated [r = 0.12, p < 0.010] indicates that stress as assessed here is distributed evenly across the sociodemographic spectrum; that is, there are couples with high levels of chronic stress who also possess high levels of social and economic capital, just as there are couples with low levels of chronic stress who confront high levels of social and economic vulnerability. Both findings corroborate the separability of chronic stress and sociodemographic risk, supporting our plan to examine sociodemographic risk as a moderator of paths linking VSA model components to satisfaction trajectories [i.e., Aim 3].

Turning to correlations involving our satisfaction outcomes, Table 2 shows that spouses with more enduring vulnerabilities had lower satisfaction intercepts [r = −0.55, p < 0.001 for husbands; r = −0.51, p < 0.001 for wives], as did couples who experienced more chronic stress [r = −0.43, p < 0.001 for husbands; r = −0.37, p < 0.001 for wives]. Couples displaying higher-quality adaptive processes also had higher satisfaction intercepts [r = 0.41, p < 0.001 for husbands; r = 0.53, p < 0.001 for wives]. VSA predictor variables were unrelated to satisfaction slopes, for husbands and for wives. Overall, these bivariate associations corroborate VSA model predictions and replicate established effects in research on couples. At the same time these results leave open questions about whether these three factors independently predict outcomes [Aim 1] and whether adaptive processes function as a mediator [Aim 2].

For the baseline model [with and without mediation], with no moderation, the model fit the data well, exceeding the minimum value of.95 for the comparative fit index [CFI] and achieving good fit [