|Home | About | Journals | Submit | Contact Us | Français|
Previous research has explored possible origins of individual differences in subjective well-being, focusing largely on stable, internal characteristics of traits as predictors of life satisfaction (Diener & Lucas, 1999). Although past work has demonstrated that life satisfaction is largely stable over the life span, other evidence has also demonstrated the lasting impact of life events. In this study, we use married couples as a test of the impact of life circumstances on life satisfaction, focusing on similarity in life satisfaction before and after divorce. If life satisfaction is impacted by shared life circumstances, married couples (who share life circumstances) should show greater similarity in life satisfaction before divorce than after. We tested this possibility using a dyadic latent-state-trait model that examined cross-spouse similarity in the stable and changing components of life satisfaction. Using a nationally representative panel study from Germany (Wagner, Frick & Schupp, 2007), we showed that similarity declined substantially following divorce. This suggests that life satisfaction is related to shared life circumstances.
Subjective well-being (SWB) is a broad construct that reflects a global evaluation of the quality of a person’s life as a whole. A critical question for research in research in the field concerns the origins of individual differences in well-being. Researchers have typically focused on two broad classes of factors that might influence SWB. First, it is possible that stable, internal characteristics might promote a happy life for the people that possess them. In other words, differences in life satisfaction could be due to differences in individuals’ personalities. Second, individual differences in happiness might be at least partially determined by external factors, such as objective life circumstances and important life events. This debate, which mirrors the nature/nurture debate that is evident throughout much of the field of psychology, has important theoretical and applied implications. Research that can help clarify the relative importance of internal and external factors is important for the field.
Casual observation suggests that life satisfaction, a broad component of subjective well-being, has largely stable roots. Some people seem to be naturally happy, and this happiness appears to persist even in the face of less than ideal life circumstances. Indeed, research shows that even over periods as long as 20 years, some degree of trait-like stability in life satisfaction measures exists (Lucas & Donnellan, 2007). Furthermore, subjective well-being—particularly the stable component of well-being—is moderately to strongly heritable (Lykken & Tellegen, 1996). Finally, well-being measures appear to be more strongly correlated with personality variables than with external circumstance variables, supporting the idea that internal factors play an important role in the well-being that people experience (Diener & Lucas, 1999).
The stability and heritability of well-being judgments, along with their associations with personality traits and weak links to external circumstances, have led well-being theorists to develop set-point theory, or the idea that people adapt to the changes in life circumstances rather than being permanently affected by them (Brickman & Campbell, 1971). According to this view, well-being does change in response to life events, but the impact is not lasting. Instead, people eventually adapt back to stable set-points that are at least partially genetically determined. The idea of adaptation is often used to explain the relatively weak associations between external life circumstances and subjective well-being. For example, Brickman, Coates and Janoff-Bulman (1978) noted that people who experience even extreme events, such as winning the lottery, show little to no difference from people who did not experience such events. One potential explanation for the lack of differences between groups is that those who experienced these important events eventually adapted.
Although set-point theories dominated research on well-being for many years (see Diener, Lucas, and Scollon, 2006, for a review), much of the research used cross-sectional designs to examine the effects of external circumstances. These cross-sectional designs have known limitations that prevent strong conclusions about the role of life circumstances from being drawn. More recent studies, however, have taken advantage of existing, large-scale panel studies to examine within-person changes in life satisfaction that occur following major life events. These studies show that although adaptation does occur, some life events can indeed influence life satisfaction for very long periods (and perhaps even permanently; see Lucas, Clark, Georgellis, & Diener, 2004). Specifically, people do not seem to completely recover from the impact of unemployment (Lucas et al., 2004), divorce (Lucas, 2005), disability (Lucas, 2007) or widowhood (Lucas, Clark, Georgellis & Diener, 2003), although some researchers disagree with these conclusions (e.g., Oswald & Powdthavee, 2008). Therefore, although life satisfaction is quite stable, it can and does change in a more lasting way in response to life events.
Although the use of panel studies has led to important new knowledge about the role that life events and life circumstances play in life satisfaction judgments, there is a critical limitation to this type of approach: Only the life events that were actually measured in the study (and occur with enough frequency to be investigated) can be used as predictors of life satisfaction judgments. This is problematic because researchers are often interested in the more general question of whether the environment matters, but they draw conclusions about this broader question from research on a relatively small sample of potential events that might occur in a person's life. Furthermore, this sort of approach cannot detect the effects of life events when these events affect different people in different ways. For instance, although marriage is generally considered a positive event, marriages vary in their quality. Thus, the event of marriage may have strong lasting effects on many who experience it, but the effects may vary from extremely negative to extremely positive, and the overall average effect may be close to zero (Lucas et al., 2003). Thus, even within-person studies are limited in their ability to determine the extent to which external circumstances affect life satisfaction ratings, and this means that complementary approaches can be useful.
In the current study, we take a different approach to understanding change in well-being. Specifically, we use a class of models designed to isolate stable trait-like variance from systematic, but less stable variance that changes over time. As we describe below, when used in the context of a dyadic design, these models provide a powerful way to isolate the effects of external circumstances on life satisfaction judgments.
As their name implies, state-trait models (Cole, Nolen-Hoeksema, Girgus, & Paul, 2006; Cole & Martin, 2005; Kenny & Zautra, 1995) decompose multiwave data into separable state and trait components. Although different variants of these models exist, in the current study, we use an extension of Kenny and Zautra's (2001) Stable Trait, Autoregressive Trait, State (STARTS) model. This model decomposes the variance in a set of assessments across multiple occasions into three components: A completely Stable Trait component that does not change over time, a systematically changing Autoregressive Trait component that has a stability coefficient that declines with increasing length of time, and a transient State component (which also incorporates measurement error) that is unique to a single occasion and is unrelated to variance at all other waves (see Figure 1). By modeling the precise pattern of correlations across differing lengths of time, the STARTS model can decompose the variance in multiwave data into these three components (for examples, Lucas & Donnellan, 2007, 2012). As long as at least four waves are available, the STARTS model can be estimated.
Note that there are alternatives to this type of model. In particular, researchers have used latent trajectory models, which focus changes in level over time (Bollen & Curran, 2004). In contrast, the STARTS model (and autoregressive models more generally) focus on modeling the correlations across waves, estimating how the stability of constructs varies with increasing lengths of time. Both approaches provide information on longitudinal change in a particular variable (here, life satisfaction). However, latent trajectory models emphasize individual differences in mean-level trajectories, which is not the focus of the current investigation.
The STARTS model and its variants provide a clear descriptive picture of the stable and changing components of constructs over time. This descriptive summary is useful in its own right (see, e.g., Lucas & Donnellan, 2007, 2012), but various extensions of the model can also be useful for testing more precise hypotheses about the underlying causes of these components. Specifically, Schimmack and Lucas (2009) showed how the STARTS model could be applied to dyadic longitudinal data from married couples to clarify the extent to which external life circumstances versus stable internal factors influence life satisfaction judgments. The logic of using the dyadic STARTS model in this way is borrowed from the logic of the twin design (and especially the separated twin design). Separated twin designs are powerful because similarity across genetically identical twins who have been reared apart can only be due to shared genes— separated twins do not share an environment. Thus, designs that allow for the decomposition of the variance across pairs of twins allow for strong statements about the role that genes play in that particular construct. The estimate of cross-twin shared variance is a direct estimate of genetic variance. This is true even though specific genes and their effects are not assessed.
The logic of the dyadic STARTS model is similar. One begins with the assumption that married partners are no more genetically similar than two strangers would be. This is a strong assumption that may be incorrect because of assortative mating. However, when considering the effects of assortative mating on the components that can be isolated by the STARTS model, such effects would be strongest for the stable trait component and perhaps the initial autoregressive variance at the start of the study. What is most interesting in this dyadic context is the new variance that is added to the autoregressive component at each wave. This is variance that was not there in the wave before (and indeed, is unrelated to anything that came before) but that is carried over into waves that come after. Because existing research using twin designs shows that this unstable variance in well-being is not genetically determined (Lykken & Tellegen, 1996) similarity across spouses in this “new” autoregressive component likely reflects the effects of external circumstances.
Schimmack and Lucas (2009) used this dyadic approach to assess similarity in spouses’ well-being over time. Specifically, they examined married couples from a large, nationally representative panel study from Germany, the German Socioeconomic Panel Study (GSOEP). Not surprisingly, Schimmack and Lucas's analyses showed that spouses are very similar in their trait level of life satisfaction (rs ranged from .62 to .77 across two 11-year periods). As noted above, this similarity could be due to the effect of their shared environment or to assortative mating. In addition, Schimmack and Lucas found a substantial correlation between the initial autoregressive trait component (rs ranging from .61 to .74). As with the similarity in the stable trait component, it is possible that similarity in the autoregressive component could reflect the effect of initial similarity that might result from assortative mating.
What is most important from Schimmack and Lucas's (2009) results are the similarity coefficients for the new autoregressive variance that emerges at each wave. As with the other components, similarity was quite strong, with rs ranging from .61 to .63. These high correlations mean that spouses are changing in the same way over time, an effect that is presumably due to their shared environment. It is also possible to determine how much of the variance at the end of the study is due to the accumulation of these environmental effects. Specifically, one can estimate the percentage of total variance that is accumulated “new” autoregressive variance that was not apparent at the beginning of the study and that is shared across spouses. Schimmack and Lucas noted that the estimate of this environmental effect is substantial: if the environmental effect for spouses is equal and independent, 60% of the changing variance could be attributed to environmental effects. Thus, the dyadic model provides strong support for environmental effects, even if the precise environmental factors are not assessed.
The dyadic STARTS model that Schimmack and Lucas (2009) developed provides compelling evidence about the role of the environment in well-being judgments, while also providing important information about the way that spouses change together over time. In the current study, we extend this model to incorporate changes that occur before and after divorce. If the similarity that has been found between spouses is truly due to an effect of the environment on both spouses (or to the effect of one spouse on the other), then once couples become divorced, their well-being should “decouple.” In other words, although there might be some trait-level similarity in life satisfaction that results from previous environmental effects and any assortative mating that led to the initial coupling, new environmental factors should lead to less and less similarity over time. The goal of the current research is to see whether similarity in spousal satisfaction is reduced when the couple ends their relationship.
This goal can be accomplished by modifying the dyadic STARTS model to incorporate multiple periods of time (in this case, pre- and post-divorce). Specifically, because the dataset we use has been conducted for such a long time, we can identify people who are married for part of the study and then get divorced. In the modified STARTS model, separate similarity coefficients can be estimated for the periods before and after the divorce. If the similarity in new autoregressive variance reflects the effects of shared environment on life satisfaction, then this similarity should decline when couples divorce. If, on the other hand, the similarity is due to the effects of genes on change in well-being combined with assortative mating effects, then similarity in autoregressive components should remain strong, even after the divorce.
Partially for pragmatic reasons having to do with model complexity, we focus only on first marriages. In addition, although similarity should decline following any marriage, we expect that the effect would be strongest for individuals’ first marriage. Thus, for this initial exploration, we will use couples in their first marriage only. In summary, applying the dyadic STARTS model to longitudinal data from a sample of married participants who later divorce provides strong evidence for the role of the environment in well-being judgments.
The German Socio-Economic Panel Study (SOEP) is an ongoing, nationally representative panel study of households in Germany. The first sample was collected in 1984, with approximately 6,000 households and 12,000 individual respondents (Wagner, Frick & Schupp, 2007). Initially, households were sampled using stratified random sampling, with refreshment samples added in subsequent years. The sample was originally designed to over-sample high-priority groups, such as immigrants. An East German sample was added in 1990. Refresher samples were added to better cover the high-income demographic in 2002, and more general refresher samples were added in 1998, 2000, and 2006. Data are collected yearly, and all adults in the household are asked to participate. This includes children who are older than 16, as well as individuals who marry into households who are already participating in the study. The total sample size now exceeds 50,000 participants.
Participants for our analyses were selected from Waves 1 to 24 of the GSOEP. Participants were selected based on two criteria: (1) they participated in the study at some point during their first marriage and (2) they became divorced from their partner during the course of the study. Data from the marriage and divorce periods were included. If data were available from before the marriage began, these data were not included in the analyses. Data were included if the participant subsequently remarried someone else. Thus, the data set includes yearly life satisfaction ratings from both partners during the marriage and divorce period, and any subsequent marriages to another person. When marital status was unclear, due to either missing or conflicting data from a member of the couple, the couple was not included in the analysis. Note that this criterion was applied individually to both spouses; that is, if it was not the first marriage for either member of the couple, the couple was not included in the analysis. In total, there were 469 heterosexual, married couples selected. On average, people participated for 4.1 years before divorce (wife M = 4.19, SD = 1.27; husband M = 4.11, SD = 1.27) and for 4.4 years after their divorce (wife M = 4.45, SD = .98; husband M = 4.35, SD = 1.06).
Each year, participants were asked a single question, "How satisfied are you at present with your life as a whole?" and responded on a scale that ranged from 0 (totally unsatisfied) to 10 (totally satisfied). Lucas and Donnellan (2012) estimated the reliability of this measure using a multivariate model at approximately 74 percent reliable variance.
The model used in this paper is a modification of the dyadic STARTS model, which is itself a modification of the original STARTS mode by Kenny and Zautra (1995). Thus, to clarify our analyses, we first present the details of the STARTS model, followed by the modifications that are required to examine the basic dyadic model and the model that incorporates changes before and after divorce.
In the original STARTS model, each occasion is predicted from a single latent Stable Trait, and the path from this latent trait to each occasion is set to 1, reflecting a consistent effect of the Stable Trait throughout the study. In addition, life satisfaction at each occasion is predicted from a latent Autoregressive Trait component that is unique to that wave. However, the latent Autoregressive Trait component at any given wave is linked to the Autoregressive Trait component from the previous wave through a stability coefficient that is constrained to be the same throughout the study. In addition, the residual variances of the Autoregressive Trait component in each wave after the first wave are constrained to be equal, and nonlinear constraints are included to ensure that the total variance in the Autoregressive Trait component is stable across waves (this “stationarity constraint” is necessary for identification, and it reflects the theoretical assumption that the breakdown of Stable Trait, Autoregressive Trait, and State variance should be constant across waves). Finally, life satisfaction at each wave is predicted from a wave-specific State component that can also be thought of as a residual. It, too, is constrained to be equal across waves, and it consists of all variance that is unique to a wave.
The dyadic version of this model introduced by Schimmack and Lucas (2009) duplicates the STARTS model for each member of a couple. In other words, there are now twice as many observed variables, with one set for the husbands, and one for the wives. These observed variables are decomposed into separate Stable Trait, Autoregressive Trait, and State components for each member of the couple, and the correlations between the husbands' and wives' components are estimated. Specifically, there is a correlation between the husbands' Stable Trait and the wives' Stable Trait, and between the husbands' initial Autoregressive Trait and the wives' initial Autoregressive Trait. In addition, there is a correlation between the husbands' and wives' residual or “new” Autoregressive Trait variance that is added at each wave. This correlation is constrained to be equal across waves. Finally, there is a correlation between the husbands' and wives' State components that is constrained to be equal across waves.
Further adjustments are needed to determine whether a decoupling of life satisfaction ratings occurs once the divorce happens. We begin with the assumption that the Stable Trait component should reflect the long-term stable variance that is maintained even when individuals are exposed to major life changes. Therefore, as in the original STARTS model, we assume that a single latent Stable Trait component affects each wave of life satisfaction in the same way, regardless of whether that occasion occurs before or after the divorce. The disruption that divorce creates should mainly be seen in the Autoregressive and State components, so equality constraints from the basic Dyadic STARTS model are relaxed across the pre- and post-divorce periods. Specifically, we divide the years before and after divorce into three periods: A pre-divorce period consisting of five years before the divorce, the year of the divorce, and a post-divorce period consisting of the divorce and four years after. Although the length of time before and after divorce can vary across participants, we selected a pre-divorce period of five years, a year for the divorce, and a post-divorce period of four years. This balances having more waves of assessment (which is good for the estimates) with completeness of data coverage (fewer participants have many waves before and after the divorce). In addition, we have a single year period (the year of the divorce) which we refer to as the transition period. This single-year transition period of the divorce is estimated separately from the pre- and post-divorce periods.
The equality constraints from the basic Dyadic STARTS model are made within the pre-divorce period and within the post-divorce period, but the parameters are not constrained to be equal across these two periods. The stability coefficients linking the pre-divorce period to the transition period and linking the transition period to the post-divorce period are estimated. The Appendix provides syntax for the model, which was estimated using Mplus 7.11.
This model leads to a number of estimated parameters that can be examined to understand stability over time and similarity across spouses. First, for each spouse there is a stability coefficient prior to the divorce. This reflects the year-to-year stability during the period when spouses are still married. There is also a transition stability coefficient for the year of the divorce (i.e., from the last year prior to the divorce to the year of the divorce), along with a stability coefficient for the years after the divorce. These coefficients described average stability of life satisfaction, and the differences across spouses and periods clarifies whether people become more or less stable when major life events like divorce occur.
In addition to the stability coefficients, the model also includes a number of similarity coefficients for the different variance components. For instance, there is a correlation between husbands' and wives' stable trait component which reflects similarity in the long-term stable component that is consistent across all waves of the study. Again, this similarity likely reflects the influence of assortative mating or long-lasting environmental determinants that were experienced during the early years of marriage. In addition to this similarity coefficient, there is a correlation between husbands' and wives' initial autoregressive variance. This parameter reflects similarity in component of life satisfaction that slowly changes from wave to wave. Like the correlation between the stable trait component, this similarity coefficient could be due to assortative mating or early environmental effects.
The parameters that are of most interest are the similarity coefficients for the new autoregressive variance. As noted above, these reflect the shared effect of environment on two members of a couple, and it is here that we should expect to see decoupling after the couples divorce. Therefore, it is important to compare the correlation between spouses’ new autoregressive variance prior to the divorce to the similarity coefficients after the divorce occurs (the period surrounding the divorce is a transition period, and similarity may be high or low at this time). State similarity coefficients are also examined for the two periods, but because these include reliable state variance and all random measurement error for an occasion, similarity coefficients are expected to be low in both periods. In short, the critical comparison will be the similarity in new autoregressive variance before and after the divorce occurs.
The two-period dyadic STARTS model was fitted to the data from 469 couples. We evaluated model fit using the χ2 exact-fit test, the root mean square error of approximation (RMSEA), the Tucker-Lewis Index (TLI) and the Comparative Fit Index (CFI). The exact-fit test is likely overpowered here, due to the large sample size, and thus, we will focus on obtaining acceptable fit according to the RMSEA (lower than .06) and the TLI and CFI (which should ideally be above .95).
When specifying the model, we chose to constrain variance across the spouses whenever possible to create the most parsimonious model without decreasing fit. Beginning with a model that was totally free (i.e., no constraints across the spouses in terms of stable trait, autoregressive trait, or state variance), we then sequentially constrained the variances, and compared the models using a chi-square difference test. As there was no clear order for the constraints, each constrained was compared against the baseline model. The chi-square test for the baseline model with none of these constraints was χ2 = 302.47, df = 186, p < .001. Results of the comparisons showed that we could make a number of constraints without increasing model misfit. Trait variance was constrained to be equal across husbands and wives. In addition, state variance was constrained to be equal across spouses within time period (i.e., state variance was set equal between wives and husbands before divorce, and also after divorce), and state variance was constrained across time for wives. Note that we did not manually set state variance for husbands to be equal across time, but after making the first three constraints, that final constraint must be true (and estimates show that to be the case). Autoregressive variance was constrained to be equal across time periods for wives (that is, Autoregressive variance was set to be equal before and after divorce for wives only), and Autoregressive variance was constrained to be equal across spouses after divorce. These constraints (and the chi-square difference tests for them) are presented in Table 1.
For this last model (with all parsimony-related constraints), the exact-fit test was also significant, (χ2 = 314.96, df = 192, p <.001), however, the other fit indices indicated good fit: RMSEA = .04, CFI = .96, TLI = .96, SRMR = .077. This suggests that our overall model is a good fit for the data, or more specifically, that the dyadic, two-period STARTS model here does a good job of describing the changes and relationships among life satisfaction assessments.
The first thing to consider when evaluating state-trait models like the STARTS is the decomposition of within-person variance into separable Stable Trait, Autoregressive Trait, and State variance. Because the metric for variances is somewhat arbitrary, this decomposition is usually reported as percentage of variance accounted for by each of the three components. Specifically, this can be interpreted as the percentage of variance at any one occasion that can be accounted for by each of the three components. In our two-period dyadic extension, this decomposition can be made separately for each member of the couple in both periods. This decomposition is presented in Table 2.
The stable trait component accounted for approximately one quarter of the variance in life satisfaction. This was relatively consistent across spouses and across the both periods. The percentage of variance accounted for by the autoregressive component varied more, both across spouses and across periods. Before and during the divorce, the autoregressive component accounted for between 30% and 40% of the variance in life satisfaction ratings. After the divorce, the autoregressive component accounted for slightly higher amounts of variance for both husbands and wives than it did before the divorce.
The second set of estimates to consider are the stability coefficients. Again, these coefficients are attached only to the autoregressive component, as the stable trait component is perfectly stable from one wave to the next, whereas the state component is completely unstable—no state variance carries over from one wave to the next. The stability coefficients are presented separately by spouse in Table 2. This table shows that even before the divorce occurs, stability of this autoregressive component is somewhat lower than what has been found in previous analyses. For instance, Schimmack and Lucas (2009) found stability estimates closer to .88. Stability does rise somewhat following the divorce, and was substantially lower during the transition period, as we would expect. Notably, the stability of life satisfaction for husbands drops to essentially zero for during the transition, but not so for wives. This suggests that husbands might be experiencing greater upheaval following the divorce than wives, although it is unclear here the possible reason for that low stability here. It is also important to note, however, that much of the year-to-year stability is incorporated in the stable-trait component, so zero-order stability coefficients would not be predicted to drop to zero for individuals. The weak stability coefficient only shows that stability in the autoregressive component decreases to zero in this transition year.
The key estimates to consider are the similarity coefficients, which are presented in Table 2. The first line shows the correlation between the husbands' and wives' stable trait component, which is substantially lower than that found in prior work (r =.77 found in Schimmack & Lucas, 2009). This suggests that among couples who will divorce in the next 7 years, trait similarity is lower than couples who will not divorce. This may also be due to the fact that the stable trait in these models is extracted from data that was collected over a period during which participants experienced an extreme event, and thus, there may be less trait variance than in consistently married individuals.
Table 1 also shows that the initial autoregressive correlation (the correlation between husbands' and wives' autoregressive variance six years before the divorce—the first wave included in our analysis) was strong and similar to what has been found previously. Note that there is only a single initial autoregressive correlation (for the pre-divorce period), as there is no comparable value at subsequent periods.
The most important estimate for the purposes of this study is the correlation between the husbands' and wives' new autoregressive components before and after the divorce. In addition, we are also interested in the degree of state similarity following divorce. Compared to previous results using only married couples, the to-be-divorced couples in our sample were as similar as couples who will not divorce (r = .64 versus r = .62 in Schimmack and Lucas, 2009). However, in line with predictions, the correlation dropped substantially from before to after the divorce. While still married, the correlation between husbands' and wives' new autoregressive components was a very high .64, whereas after divorce, the results of the unconstrained model show that it dropped to a much smaller .20. This shows the predicted decoupling effect following divorce. Once the divorce occurs, members of the former couple do not change in similar ways across subsequent waves.
Following the parsimony-related constraints, in order to test the strength of the similarity in the post-divorce period, we estimated three models: one in which autoregressive similarity and state similarity after the divorce was freely estimated, one in which state similarity after the divorce was set to zero, and finally, one in which the state and the autoregressive similarity are constrained to zero (the results of which are shown in Table 1). Constraining the state similarity to zero in the post-divorce period did not significantly increase misfit, χ2 = 316.61, df = 193, p <.001, with a χ2difference = 1.65, df = 1, p = .19. This demonstrates that the similarity in the state component declines to essentially zero following the divorce. Following this model, we then tested a model in which state similarity and autoregressive similarity are constrained to zero. This model did show a significant decrement in fit, χ2 = 320.94, df = 193, p <.001, with a χ2difference = 5.146, df = 1, p = .01. This suggests that, although autoregressive similarity declines following divorce, it does not decline to zero. Nonetheless, all other fit indices remained above thresholds, suggesting that autoregressive similarity after divorce is quite low, but not zero.
Table 2 also shows the correlations between the spouses' state components. Although previous research has found moderate cross-spouse correlations in this component, they should be small and may not be sensitive to changes in status, as the state component includes all measurement error in addition to any reliable occasion-specific variance. Indeed, correlations were weak during both periods (though they were somewhat stronger before the divorce occurred). Indeed, a likelihood ratio test showed that constraining the state similarity to zero post-divorce did not substantially increase model misfit, as shown in Table 1.
In general, these results suggest both that similarity in life satisfaction does decline following divorce (suggesting the influence of shared life events during marriage), despite the fact that couples who are going to divorce begin just as similar as those who do not divorce (see Schimmack & Lucas, 2009).
The goal of this study was to examine spousal similarity in life satisfaction before, during, and after divorce using an extension of the dyadic STARTS model developed by Schimmack and Lucas (2009). Previous work used this approach to demonstrate similarity in the changes in spouses' life satisfaction over time, providing evidence of similarity that is likely above and beyond genetic similarity, but is due to shared life circumstance. By demonstrating that there is a reduction in similarity in life satisfaction in spouses following divorce, we can strengthen this evidence further, because this suggests that similarity over time is not due to similar developmental trajectories that might be influenced by internal factors. The results here suggest that life circumstances do influence life satisfaction. Combined with prior results, these data suggest that shared life circumstances in spouses influence their life satisfaction. There is a substantial decline in similarity in life satisfaction when couples fail to share life events.
The results of the model showed that we were able to replicate the basic decomposition of life satisfaction into Stable Trait, Autoregressive Trait, and State components, as the model did fit the data well. However, the percentage of variance that is associated with each of these components did vary slightly across spouse and across time. The amount of variance associated with the AR Trait increased from approximately 30% before divorce and during the transition period to about 40% after the divorce. Nonetheless, the percentage of variance roughly replicated previous work showing the decomposition of variance among the components.
In addition to some changes in the percentage of variance over time, there are differences in stability across the three different periods. The results of the model revealed that the stability in life satisfaction increases slightly after divorce, and decreases substantially during the divorce transition period. Note that this refers to stability in the Autoregressive Trait component of life satisfaction. This indicates the auto-regressive component of life satisfaction is accounting for slightly more variance after divorce and that variance becomes more stable.
Theoretically, the most important aspects of the model concern the similarity in life satisfaction during both periods. In this paper, we examined similarity in three different components of life satisfaction: Stable Trait similarity, Autoregressive Trait similarity, and State similarity. Schimmack and Lucas (2009) previously provided a strong argument that similarity in the new variance added yearly to the AR component is influential and does have a substantial effect on well-being. The model here demonstrates several interesting results. First, similarity between spouses' stable trait and new AR during the marriage period was weaker than in past analyses. Couples who are going to divorce show lower levels of similarity to couples who will not divorce as compared to past studies (Schimmack & Lucas, 2009) in the Stable Trait component, and equivalent similarity in their initial AR variance than those couples. In general, we replicated the earlier findings that couples appear to be similar in their life satisfaction prior to divorce.
Second, we were able to show here that similarity does decrease substantially following divorce. The similarity in the AR component of life satisfaction drops substantially following divorce, although it does not decline to zero. This suggests that spouses are changing in similar ways during their marriage (as shown by positive correlations between new AR variance at each wave), but that they stop doing so once they divorce. This finding suggests that life circumstances, and specifically, shared life circumstances across spouses, have an impact on the autoregressive component of life satisfaction. They are influencing spouses to change in similar ways in their level of life circumstances across time. More generally, this supports the assertion that individual characteristics (i.e., traits) are not the sole explanation for individual differences in happiness; life circumstances matter as well.
The method applied here allows examination of the impact of life circumstances on life satisfaction in couples. Here we have good evidence that similarity in life satisfaction decreases following a divorce, suggesting that there may be an influence of shared life circumstances on spousal similarity during the marriage. However, there is also an increase in overall autoregressive variance following divorce, and for husbands, the variance there is more stable. This might indicate the process of adaptation beginning to occur following divorce, with state variance therefore accounting for more of the variance as individuals adapt to the changes in their lives.
Despite some of the questions that remain concerning the changes in stability following divorce, this study provides an important step in demonstrating the impact of life events on life satisfaction. Not only do major life events appear to have lasting impacts on life satisfaction (e.g., Lucas, 2007), sharing life circumstances within a marriage results in greater similarity in life satisfaction than exists after divorce. This suggests that life events are more important to an individual’s happiness than previously thought by set-point theorists. However, the autoregressive similarity in life satisfaction between spouses does not decline to zero following divorce. This suggests that, although couples are becoming less similar, that similarity is not zero even after they marriage ends. There are a variety of possible reasons for this. In particular, it could be that couples had children together, and the presence of this continuing shared life experience (through parenting) is impacting their life satisfaction in such a way that it remains similar even after couples divorce. Future researchers might examine children as a potential moderator of similarity in life satisfaction following divorce.
In addition, the data here are from a single culture (Germany). When examining any question within a single culture, there is always the possibility of cultural difference. Here, particularly, the degree of similarity in life satisfaction in married couples might differ depending on cultural differences such as mate selection processes. Future researchers might examine these questions of cultural variability. In addition, this study helped to demonstrate the utility of the multi-period STARTS model. By creating a model that spans one or two different periods of time, it is possible to examine change processes based on any number of events and their impact on any number of different constructs. The utility of this model is not limited only to examining changes in life satisfaction, but can be applied to changes in other constructs such as personality or health. It provides a useful method for understanding how changes in related events might impact an individual in a variety of ways, and thus, can be extremely useful in understanding how people adapt.
However, this method does have some important limitations. In particular, it can only show that the environment is having an effect. It does not provide information about the kinds of life circumstances or events that might be causing that effect. From a practical standpoint, as well, this model requires assessment of the outcome of interest at repeated time points. Longitudinal data are a necessity to understand the environmental effects that might be occurring using the multi-period STARTs model. Finally, it requires examination of change that occurs in dyads; in order to examine change over time to demonstrate the influence of life circumstances, it is important to first demonstrate that two people are changing in the same way. In the future, it would also be interesting to examine couple- or individual-level differences in similarity over time. That is, it is possible that there are important couple differences in the degree of similarity in life satisfaction or its decrease following divorce. Unfortunately, the STARTS model cannot examine this possibility (nor does any existing model, to our knowledge). Despite these qualifications, this method might prove very useful as a tool to examine change processes in dyads.
Overall, this study helps clarify the impact of life circumstances on happiness. Although lasting individual differences such as personality do explain a large portion of the stable individual differences in happiness or life satisfaction, it appears from the results here that life circumstances matter as well. Previous work has suggested that life circumstances do seem to impact life satisfaction (Schimmack & Lucas, 2009). This impact can be seen by the fact that spouses show similarity in their life satisfaction, above and beyond the effects of assortative mating (which is reflected in stable trait similarity in life satisfaction). Here, we were able to demonstrate that this similarity in life satisfaction is not only demonstrated in spouses, but that the similarity declines to almost zero after divorce, when spouses are no longer sharing life circumstances.
Such a decline in similarity in life satisfaction suggests that life circumstances do seem to have an important impact on how satisfied people are with their lives. Changes in life circumstances, then, can have important and lasting effects on a person’s level of happiness. Individual differences in life satisfaction are due not only to individual differences in traits, but also differences in life circumstances and situations. Although this paper can only suggest these sorts of potential causal relationships, it suggests people might be able to exercise some control over the circumstances of their lives and therefore, on their satisfaction with them.
The work presented in this article was supported by a grant from the National Institute on Aging (R01AG040715). The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.