Previous studies have shown that the risk of divorce is low during the first months of marriage; it then increases, reaches a maximum, and thereafter begins to decline. Some researchers consider this pattern consistent with the notion of a “seven-year itch,” while others argue that the rising-falling pattern of divorce risk is a consequence of misspecification of longitudinal models because of omitted covariates or unobserved heterogeneity. The aim of this study is to investigate the causes of the rising-falling pattern of divorce risk. Using register data from Finland and applying multilevel hazard models, the analysis supports the rising-falling pattern of divorce by marriage duration: the risk of marital dissolution increases, reaches its peak, and then gradually declines. This pattern persists when I control for the sociodemographic characteristics of women and their partners. The inclusion of unobserved heterogeneity in the model leads to some changes in the shape of the baseline risk; however, the rising-falling pattern of the divorce risk persists.
An extensive amount of multidisciplinary literature exists on the trends and patterns of divorce and separation in industrialized countries (for recent reviews, see Amato 2010; Cherlin 2010; Lyngstad and Jalovaara 2010). An ingredient of longitudinal models on divorce is the marriage duration. Most studies show that the risk of separation is low during the first months of a marriage; it then increases, reaches a maximum, and thereafter begins to decrease. Various studies on industrialized countries have found this pattern (Andersson 1995; Diekmann and Engelhardt 1999; Hoem and Hoem 1992; Jalovaara 2013; Kiernan 1999; Kulu and Boyle 2010; Lyngstad 2011; Rootalu 2010; Schoen 1975; Thornton and Rodgers 1987). The reason for the rising-falling pattern of divorce risk over the marriage duration, however, has not been discussed much in demographic and sociological studies. Classical psychological literature and public discourse consider this pattern consistent with the notion of a “seven-year itch.” Most married couples experience a gradual but steady decline in marital quality after the first year of marriage, suggesting that the short honeymoon period of passion is followed by everyday routine, during which the differences between spouses’ attitudes, values, and behavior become visible and subject to discussion and arguing. This is a period when spouses encourage some behaviors of their partners and discourage others; at the same time, they try to adapt to those behaviors and characteristics of their partners that cannot be easily changed. Experiences are cumulated over time through a mutual learning process, and the couple determines whether staying together makes sense. If the mutual adaptation is successful, a longer period of stability follows in the marital relationship (Diekmann and Mitter 1984; Kurdek 1999; Levinger 1983; Sternberg 1986).
Although the premise of the seven-year itch is convincing, perhaps the rising-falling pattern of divorce risk is simply produced by misspecification of longitudinal models on divorce and separation. It is well-known that the baseline risk in hazard models is sensitive to model specification (Galler and Poetter 1990; Hoem 1990; Vaupel et al. 1979; Vaupel and Yashin 1985). Recent studies reporting the rising-falling pattern of divorce over the marriage duration have controlled for a set of demographic and socioeconomic characteristics of spouses. However, it is likely that some important characteristics of spouses have not been measured and included in the analysis; this is particularly the case with personality traits of spouses, their values, and their long-term plans. For example, the sample may contain individuals who are prone to divorce because of liberal values or because they are ambitious and never satisfied with their current life situation. Similarly, the sample may include individuals who are less prone to divorce because of traditional value beliefs or because they tend to avoid changes in their lives. Omitting important covariates from the analysis or unobserved heterogeneity may significantly distort the true shape of the baseline risk: the estimates on the risk of divorce over marriage duration would be downwardly biased. The logic is simple. The high-risk group leaves the risk population first; therefore, as time goes by, the share of the low-risk group increases, and the hazard of divorce for the population will approach their (low) risk levels (for details, see Online Resource 1).
The aim of this study is to investigate the causes of the rising-falling pattern of divorce risk over the marriage duration. Some literature supports the psychological interpretation (e.g., Diekmann and Engelhardt 1999), and other studies suggest that the unmeasured characteristics of individuals account for the rising-falling pattern of divorce risk (e.g., Vaupel and Yashin 1985). However, no study has explicitly controlled for unobserved heterogeneity when examining the hazard of divorce over the marriage duration. I use a large longitudinal data set from Finland and multilevel hazard regression models, which allow me to identify and control for unobserved heterogeneity.
I first study the hazard of divorce over the marriage duration, controlling for a set of demographic and socioeconomic characteristics of women and their partners. I then identify and control for possible unobserved heterogeneity to detect any changes in the shape of baseline risk. In further analysis, I examine the hazard of divorce over the marriage duration separately across periods to identify any changes in the shape of baseline risk over years.
Data, Methods, and Modeling Strategy
The data come from the Finnish Longitudinal Fertility Register, a database developed by Statistics Finland that contains linked individual-level information from different administrative registers (see Vikat 2004). I had access to full marital records for 10 % of the (randomly selected) Finnish women born between 1945 and 1983. The data also include full educational and fertility histories of women and their demographic characteristics (date and place of birth, and native language). In total, 74,441 Finnish women were included in the analysis, excluding foreign-born women (3 %). I linked to women’s data records of their husbands; I had information on the men’s educational histories and their demographic characteristics. I included in the analysis all marriages for women that were formed between 1967 and 2000. The database at hand contains no information on divorces before that period. In total, there were 81,025 marriages: 74,441 first, 6,106 second, and 478 third or higher-order marriages (see Table S1 in Online Resource 1). The number of divorces was 18,839: 17,201 for first, 1,498 for second, and 140 for third and subsequent marriages. I controlled for basic demographic and socioeconomic characteristics of women when investigating the risk of divorce over marital duration (Table 1).
The identification of the model was attained through within-person replication. Some women had experienced more than one marriage episode; this was sufficient for the identification of the model with a woman-level random effect or “shared frailty” for marriage episodes (Aalen 1994; Hoem 1990; Hougaard 1995). More specifically, the information provided by repeated episodes was used to estimate the variance of the random effect in the data (rather than individual random effects); the parameters of the model were estimated conditional on (or along with) the variance of the random effect (Lillard 1993; Lillard and Panis 2003).1 I thus used a multilevel event-history model to control for unobserved heterogeneity—a strategy that has become common in demographic studies (see Brien et al. 1999; Kulu 2005; Kulu and Steele 2013; Lillard 1993; Lillard et al. 1995; Steele et al. 2006).2
In Model 1, I included marriage duration (baseline), marriage order, and period. As shown in Fig. 1 and Table 2 (Model 1), the hazard of divorce significantly increased during the first three years of marriage. In the following two years, the risk further increased and reached its peak at the fifth year of marriage. In the next five years of marriage, the risk of separation significantly decreased, and a decline also continued later. The analysis thus supports the results of previous studies: a rising-falling pattern of divorce risk over the marriage duration, with its peak around the fifth year of marriage. The effects of the two covariates were as expected; the hazard of divorce increased over time (Andersson 1995), and women in their second and third marriages were more likely to separate than women in their first marriages (Andersson 1995; Hoem and Hoem 1992).
In Model 2, I included a set of sociodemographic characteristics of women and their partners. The likelihood ratio (LR) test showed that the model fit improved significantly (LR = 7,829.1, df = 26, p < .001). The effects of covariates were as expected: as shown in Table 2, the hazard of divorce increased with a decrease in the woman’s age at marriage (cf. Tzeng and Mare 1995); this risk was also higher for couples with a low educational level (cf. Hoem 1997), for women who spoke Finnish rather than Swedish (cf. Finnäs 1997), and for couples in which the woman was older than the man (cf. Chan and Halpin 2003). As expected, the risk of divorce decreased with an increasing number of children in the family. However, the hazard of divorce significantly varied with the age of the youngest child. The risk of separation decreased substantially when a woman became pregnant; the risk remained low during the pregnancy, around the birth of a child, and during the first two years of the child’s life; the hazard of divorce gradually increased as the child became older (cf. Erlangsen and Andersson 2001). Couples who experienced a premarital conception or childbirth had a significantly higher risk of marital disruption than those whose children had all been conceived and born in marriage, which was also expected (cf. Hoem 1997). Most importantly, however, the shape of the baseline hazard changed somewhat after controlling for basic sociodemographic characteristics of couples (see Fig. 1; Table 2, Model 2). The rising-falling pattern became even more pronounced, again consistent with the results of previous studies.
In Model 3, I also included a woman-level random residual to control for the effect of her (time-invariant) unmeasured characteristics on the hazard of divorce. I thus moved beyond the dominant setup of the analysis of divorce. The model fit improved significantly (LR = 83.6, df = 1, p < .001); the estimated standard deviation of the woman-level residuals was 0.77. This suggests that there were woman-specific unobserved characteristics that influenced the risk of separation for any of her marriages. A woman at 1 standard deviation above the average had more than two times higher hazard of separation (exp(0.77) = 2.2), and a woman at 1 standard deviation below the average had a 54 % lower risk of divorce (exp(–0.77) = 0.46). Most importantly, the shape of the baseline hazard changed when I controlled for unobserved heterogeneity. However (and perhaps surprisingly), the change was relatively small, and the rising-falling pattern of divorce persisted: the risk of divorce significantly decreased five years after marriage formation, and the decline continued later (Fig. 1; Table 2, Model 3). The results thus challenge the idea of the critical role of unmeasured characteristics: after I controlled for demographic and socioeconomic characteristics of women and their partners, the rising-falling pattern of the hazard of divorce persisted even when I also controlled for unmeasured women’s characteristics. The rising-falling pattern persisted for two reasons. First, the decline of the hazard after its peak was steep in the model in which I controlled for all available (observed) characteristics of women and their partners (particularly the age of the youngest child). Second, the estimated variance in the heterogeneity term was not large enough to substantially modify the baseline risk.3
Next, I extended the analysis by investigating the hazard of divorce over the marriage duration across periods to identify any changes in the shape of baseline risk over years. I estimated a model with separate baselines for five periods (until 1979, 1980–1984, 1985–1989, 1990–1994, and 1995 and later). The model fit improved significantly (LR = 146.1, df = 22, p < .001), suggesting significant differences in the hazard of divorce over the marriage duration by period. On closer inspection, however, the differences were not fundamental. The shape of the baseline hazard was rather similar for the three periods (Fig. 2; Table S2 in Online Resource 1, Model 4). Interestingly, however, the hazard of divorce risk reached its peak sooner in the 1980s and in the beginning of the 1990s than in the 1970s; the relative risk of divorce at longer marriage durations was also slightly higher in the later periods than in earlier periods.
Summary and Discussion
The aim of this study was to investigate the causes of the rising-falling pattern of divorce levels. I used a large longitudinal dataset and applied multilevel event-history models to control for unobserved heterogeneity when examining the hazard of divorce by marriage duration. The initial analysis supported the rising-falling pattern of divorce by marriage duration: the risk of marital dissolution increased, reached its peak at around the fifth year of marriage, and then gradually declined. This pattern persisted when I controlled for the sociodemographic characteristics of women and their partners and the number of children. The inclusion of unobserved heterogeneity in the model led to changes in the shape of the baseline risk; however, the rising-falling pattern of the divorce risk persisted.
Some issues related to specification of the models and robustness of the results are worthy of discussion. One issue is whether to include in the divorce models marriage cohort or calendar year to measure the effect of social context. Although both model specifications are correct, there is a clear reason to favor calendar year over marriage cohort if one wishes to distinguish various duration effects from one another. Further analysis revealed that a model with marriage cohort (rather than with calendar year) would have led to the overestimation of the divorce risk at longer marriage durations because of increased divorce rates over years. This suggests that using the calendar year is a better variable than marriage cohort to measure changes over time in the context where the risk of divorce steadily increases.
Another issue is that the data used in this study contain no information on premarital cohabitation. However, research shows that cohabitation significantly increased in Finland during the study period: most women born before 1945 started their (first) union by contracting a marriage, but the majority of those born in the 1960s cohabited before marriage (Lindgren et al. 1992). Therefore, it is expected that most marriages in the 1980s and 1990s followed a spell of cohabitation with an average duration of two to three years (Andersson and Philipov 2002:218; Lindgren et al. 1992). As cohabitation gradually spread during the study period, this may also explain why marriage durations slightly declined in the 1980s in comparison with the 1970s. At first glance, this suggests that the risk of separation for relationships reaches its peak a few years later than observed in the divorce data used. However, the fact that divorce process lasts some time may cancel out the time that premarital cohabitation adds to the length of the relationship. In Finland, the process of divorce usually lasts a year because of a required period of reconsideration after the submission of an application for divorce. Further, before 1987, the partners who had considered divorce first had to live separately for a year before their application for divorce was approved. There is thus a period of a year or so between the timing of a decision to divorce and the date of divorce recorded in statistics.
This study used register data from Finland to investigate the causes of the rising-falling pattern of the divorce risk over the marriage duration. The analysis found no support to the idea that the rising-falling pattern is attributed to misspecification of the divorce models—that is, to the omission of unmeasured (time-invariant) characteristics of individuals from a model on marital divorce. Future research should study the role that the aging of individuals plays in declining divorce rates over the marriage duration.
The author is grateful to three anonymous referees and former Editor Stewart Tolnay for valuable comments and suggestions on a previous version of this article. The author also thanks Statistics Finland for providing the register data used in this study, as well as Mrs. Marianne Johnson for valuable suggestions when preparing the data order. The analyses made in this study are based on the Statistics Finland Register Data at the Max Planck Institute for Demographic Research (TK-53-1662-05).
I conducted experiments to investigate the sensitivity of residual variance to the share of population with repeated episodes. As expected, the estimate for the residual variance depended little on how large the share of population with repeated episodes was (100 %, 50 %, 25 %, 10 %, and 5 %) in the context where the intraclass correlation was moderate to strong (i.e., the durations of an individual were positively correlated).
The identification of the model with unobserved heterogeneity was thus based on the existence of multiple marriages (and divorces) for some women. It is likely that there were more disruption-prone women (or women with unmeasured characteristics that made them disruption-prone) among this group than among women who had been married only once. Therefore, some assumptions of my statistical model might not be fully met (e.g., normality of the residuals), and I might overestimate or underestimate the true amount of heterogeneity in the population (e.g., unmeasured individual values or personality traits that made some women more and others less likely of experiencing divorce). However, the proposed approach was the only way of estimating unobserved heterogeneity from the data without imposing strong assumptions on data. Alternatively, one could have used all women to identify unobserved heterogeneity; however, this approach would have required strong assumptions about distribution of both the residuals and the baseline risk (i.e., the risk of divorce by marriage duration), the shape of which is the main interest of this study. My approach was thus well justified; I also tested sensitivity of the results to underestimation and overestimation of the amount of unobserved heterogeneity in the population (see footnote 3).
I conducted additional analysis to explore how sensitive the shape of the baseline was to the estimates for the residual variance/standard deviation. The shape of the baseline became less pronounced with an increase in the variance/standard deviation, as expected (see Fig. S1 in Online Resource 1). However, the value of 3.0 for the standard deviation (9.0 for the variance) was needed to substantially modify the shape of the baseline. This suggests that a woman with unmeasured characteristics that place her at 1 standard deviation above the average had 20 times higher hazard of separation (exp(3.0) = 20.1) than a woman with average unobserved characteristics (e.g., in the middle in the liberal-conservative scale), while a woman at 1 standard deviation below the average had a 95 % lower risk of divorce (exp(–3.0) = 0.05). Empirical analysis gave no support to (or even no indication of) the existence of such enormous unobserved heterogeneity in the data.