Social scientists have been writing about the divorce cycle, the propensity to end one’s own marriage as a result of growing up in a divorced family, at least since the 1930s (for an overview, see Wolfinger 2005). At least 25 studies have demonstrated that marital instability runs in families, evidence that divorce transmission is of concern to social scientists. The possibility that some of the negative consequences of growing up in a divorced family have abated adds a new level of interest. For several years around the end of the twentieth century, there was considerable support for rolling back the clock on easy divorce laws in order to preserve two-parent families (Nock et al. 2008). It would undercut the critics of no-fault divorce if ending a marriage no longer hurt children as much as it used to.

Twelve years ago, I used data from the 1973–1996 General Social Surveys (GSS) to show that the intergenerational transmission of divorce had weakened substantially over time (Wolfinger 1999; see also Wolfinger 2005). Li and Wu (2006, 2008) contested this finding on methodological grounds: the ostensible decline in the probability of divorce transmission, they averred, is nothing more than a methodological artifact resulting from the failure to properly model right censoring. Using data from the National Survey of Families and Households, they found no evidence of a trend in the divorce cycle after accounting for right censoring via event history analysis.

In this article, I develop a strategy for studying divorce using GSS data on completed marriage cohorts that obviates concerns about right censoring. This approach reveals a trend in the divorce cycle that is consistent with my earlier findings (Wolfinger 1999) but contrary to Li and Wu’s (2006, 2008), thus contributing new evidence to an ongoing debate about whether the adverse consequences of growing up in a disrupted family abated during the years that divorce became more common in America and other Western nations.

Right Censoring, the General Social Survey, and Completed Cohorts

Li and Wu (2006, 2008) contended that limitations of the GSS produced a spurious decline in the divorce cycle reported in my earlier work (Wolfinger 1999). The GSS lacks adequate information on marriage duration to conduct event history analysis, the preferred statistical technique for right-censored phenomena like divorce (Allison 1984, 1995). I therefore analyzed divorce transmission via logistic regression. Li and Wu (2008) claimed that

. . . if divorce rates are identical across marriage cohorts, more respondents in earlier marriage cohorts would be observed to divorce relative to respondents in later marriage cohorts simply by virtue of longer exposures to the risk of divorce. If Wolfinger’s controls for exposure to risk are inadequate, his conclusion that divorce transmission has declined could be a methodological artifact (p. 875).

It may well be true that divorce is more fully observed for earlier marriage cohorts, but this by itself would not produce a spurious decline in the intergenerational transmission of divorce because people from both divorced and intact families of origin would be affected. For my finding about a decline in divorce transmission to be incorrect, the probability of right censoring would had to have disproportionately increased for the children of divorce. Li and Wu did not consider this issue. Instead, they reproduced my finding using data from the National Survey of Families and Households—albeit without my controls for duration dependence—and then explained it away as an artifact of improper adjustment for right censoring. The lesson to be learned from Li and Wu is that right censoring, if not properly accounted for, can induce bias (though demonstrating right-censoring bias using the National Survey of Families and Households does not prove that it exists for the GSS data I analyzed).

Li and Wu did establish the need for an analytic strategy that can rule out any suspicion of bias. The solution is an analysis of completed marriage cohorts. After about four years of marriage, the hazard rate for divorce declines monotonically (Diekmann and Mitter 1984; Goldstein 1999). After 30 years, the hazard is negligible.Footnote 1 By this point, most couples inclined to divorce will have done so. Because the GSS measures age at first marriage, it is possible to identify respondents for whom 30 years have passed since the time they first wed. At this point, a marriage cohort is essentially completed—either first marriages have long since dissolved or their participants are probably together for life—allowing researchers to study trends in the intergenerational transmission of divorce without worrying about whether right censoring is affecting the results.

Methods

Data

This research uses data from the General Social Survey (GSS) (Davis and Smith 2007). The GSS, a national probability sample of English-speaking households within the continental United States, has been conducted annually or biennially since 1972. Within each household, an adult aged 18–89 is randomly selected as the respondent. I use data for the years 1973–1994, excluding the black oversamples in 1982 and 1987. After 1994, the GSS ceased inquiring about age at first marriage and did not do so again until 2006. The 1972 survey did not adequately measure family structure of origin.

Analysis is limited to respondents for whom at least 30 years have passed since they first married (N = 7,226). Cases with missing data are deleted listwise except for parental education (for which an additional dummy variable is coded for missing data) and occupational status (for which missing data are set to the sample mean, and a dummy variable for missing data is included). More sophisticated means of handling missing data, such as multiple imputation, do not perform appreciably better (Paul et al. 2008).

Variables

The dependent measure in all analyses is whether a respondent reports ever having been divorced (summary statistics appear in Table 1). A single dichotomous measure is formed by merging information from two questions: one inquiring whether respondents have ever been divorced, and the other querying respondents about current marital status. Never-married respondents are excluded from the analysis; previous research suggests that differential selection into marriage cannot explain trends in the divorce cycle (Wolfinger 2005: Appendix B). Unfortunately the GSS does not have adequate data for event history analysis of divorce (including event history analysis with time-varying covariates).

Table 1 Summary statistics
Full size table

The GSS includes two items that measure the structure of respondents’ families of origin. Respondents were first queried about household composition at age 16. If respondents were not living with both biological parents, a second item ascertained the reason. Following my earlier work (Wolfinger 1999), my analysis here is based on GSS respondents who reported the three most common varieties of family structure: intact two-parent families, mother-only families resulting from divorce or separation, and mother/stepfather families resulting from divorce or separation. Respondents reporting other living arrangements are omitted from the sample, as are those whose living arrangements at age 16 were the product of parental military service, parental incarceration, or parental death. These exclusions have little appreciable affect on rates of divorce transmission (Wolfinger 2005). The family structure items are recoded as a single dummy variable measuring whether a respondent hailed from a divorced family (including stepfamilies).

Analyses include continuous variables measuring three dimensions of time: marriage cohort, birth cohort, and survey year. My earlier work (Wolfinger 1999, 2005) used survey year as the temporal index for studying trends in the divorce cycle; Li and Wu (2006, 2008) used marriage cohort. I present regression results based on different combinations of these three variables. None of the three are mean-centered (centering them on their means produces virtually identical results). No model contains both birth cohort and marriage cohort given the high correlation of these two variables (r = .92). Alternate model specifications are discussed in greater detail in Online Resource 1.

On average, adults reared in nonintact households complete fewer years of schooling (McLanahan and Sandefur 1994) and fare less well vocationally (Biblarz and Raftery 1993). To ascertain whether trends in the divorce cycle are the result of diminished socioeconomic status, I use three variables: occupational prestige for respondents, and education for both respondents and their parents. For respondents reared in intact families and stepfamilies, the higher level of education between the two parents is used. For people from mother-only families I use mothers’ education. Measures of income or occupational status for respondents’ parents would be helpful but are not available. An item that asks respondents to recall their families’ economic well-being almost certainly fails to provide accurate recollections.

Researchers have shown that various other factors may affect the relationship between parental divorce and respondent divorce. I ascertain whether the following affect trends in the probability of divorce transmission: race (Bumpass et al. 1991; Glenn and Kramer 1987; McLanahan and Bumpass 1988), presence of siblings (Mueller and Pope 1977), Catholicism (McLanahan and Bumpass 1988), rural origins (Pope and Mueller 1976), age at marriage (Wolfinger 2003a, 2005), and gender (Amato 1996; Glenn and Kramer 1987; Kulka and Weingarten 1979). Controlling for gender is especially important because men often fail to report their own divorces (Bumpass et al. 1991; Mitchell 2010).

Analysis

I estimate logistic regression models assuming the following general form:

(1)

where p is the probability of respondent divorce; DIV is the dummy variable measuring whether respondents hail from divorced families; TIME is survey year, marriage cohort, and/or birth cohort; and CONTROL represents miscellaneous control variables. The interaction between family background and survey year, marriage cohort, or birth cohort allows for exploration of trends in the divorce cycle; my previous research indicates that the functional form of the decline in divorce transmission is linear (Wolfinger 1999, 2005). Robust standard errors based on primary sampling units are reported to account for the cluster-sample design of the GSS.

Results

The logistic regression analysis of completed marriage cohorts appears in Table 2. Model 1 follows Wolfinger’s (1999) lead by using survey year as the temporal index measuring trends in the divorce cycle. All variables in this model are statistically significant, most notably the interaction between parental divorce and survey year. As in Wolfinger (1999), the negative coefficient for this interaction indicates that divorce transmission has declined over time for GSS respondents in completed marriage cohorts. The magnitude of the decline can be obtained by substituting values for the year variable into the following equation, derived from the parameter estimates shown for Model 1:

Table 2 Logit analysis of respondent divorce on parental divorce, survey year, and marriage cohort for completed cohorts
Full size table
(2)

For 1973, the equation yields an odds ratio of 4.03, indicating that GSS respondents from divorced families interviewed in 1973 were about four times more likely to report a personal divorce than were respondents who lived with both biological parents at age 16. By 1994, this ratio had declined to 1.89. These figures represent a larger decline in the intergenerational transmission of divorce than I previously reported (Wolfinger 1999, 2005). Based on completed marriage cohorts, this result cannot be an artifact of the right-censoring bias alleged by Li and Wu (2006, 2008). Furthermore, the data span many years: birth cohorts from 1884 to 1948, marriage cohorts from 1901 to 1964, and GSS waves from 1973 to 1994.

Survey year is not the best temporal index for trends in divorce transmission in an analysis of completed cohorts, given that the period of high divorce risk in a marriage has long since passed by the time the data are collected (Diekmann and Mitter 1984; Goldstein 1999). Model 2 offers results based on the interaction between marriage cohort and parental divorce. Survey year is also included in order to account for survey-specific change. Model 3 omits survey year, while Model 4 shows results based on the interaction between birth cohort and parental divorce. In all three cases, the interaction term measuring trends in the intergenerational transmission of divorce is negative and statistically significant at the p < .10 level. These results show that the trend in the divorce cycle is robust to alternative model specifications.

It might be argued that statistical significance at .10 is less than impressive, but the point of the analysis of completed cohorts is not to produce the definitive assessment of trends in divorce transmission based on the GSS; this has already been accomplished in my earlier work (Wolfinger 1999, 2005). Instead, my objective is to rule out the possibility that the trend in the divorce cycle is a product of improper controls for right censoring, as alleged by Li and Wu (2006, 2008). This is amply demonstrated by the significance tests in Models 1–4: these tests should have been nowhere near significance if Li and Wu (2006, 2008) had been correct in their criticism of Wolfinger (1999). Moreover, it could be argued that one-tailed tests are appropriate for my analysis given my directional hypothesis: divorce transmission has declined over time. If one-tailed tests are used, the results in Models 1–4 are all significant at the .05 level.

Model 5 introduces social and demographic variables (education, parental education, occupational prestige, sex, race, Catholicism, age at first marriage, urban origins, and presence of siblings) into the analysis. These variables attenuate the effect of parental divorce on offspring marital stability, as the interaction between marriage cohort and parental divorce becomes nonsignificant. This finding is consistent with the contention that the etiology of divorce transmission is partially attributable to social differences between respondents (Amato 1996; Wolfinger 2005).

Discussion

Although my analysis of completed cohorts provides strong evidence that the divorce cycle abated, it is important to acknowledge what others have found. Four U.S. studies (Amato and Cheadle 2005; Li and Wu 2008; McLanahan and Bumpass 1988; Teachman 2002) showed that divorce transmission has remained stable over time.Footnote 2 Teachman provided the strongest test, analyzing multiple waves of the National Survey of Family Growth; McLanahan and Bumpass employed a single wave of this survey. Amato and Cheadle used a sample of parents and children from the Marital Instability Over the Life Course survey. However, Engelhardt et al. (2002) presented equivocal evidence of a decline in divorce transmission in Germany. Diekmann and Engelhardt (1999) provided somewhat stronger evidence of a decline, also in Germany. Kulka and Weingarten (1979) found that the divorce cycle was weaker for survey respondents interviewed in 1976 than for a comparable sample from 1957.

Additional evidence of the weakening of the negative consequences of parental divorce exists for outcomes besides its transmission between generations. Amato and Keith’s (1991) meta-analysis of almost 100 studies found that the average negative effect of parental divorce on offspring well-being has weakened over time.Footnote 3 More recently, I showed that rates of teenage marriage for the children of divorce declined disproportionately faster than they did for offspring from two-parent families (Wolfinger 2003b, 2005). This by itself should prolong the marriages of people from divorced families. Teenage marriage is strongly correlated with divorce; in turn, age at marriage can itself account for a portion of the divorce cycle (Wolfinger 2003a, 2005). The final piece of evidence is indirect. Dronkers and Härkönen (2008) found that the probability of divorce transmission varied inversely by the level of parental divorce across 18 European countries. In other words, the divorce cycle was strongest when parental divorce was least common. This is analogous to the finding presented here and in my earlier work (Wolfinger 1999, 2005), with the prevalence of parental divorce varying by country instead of historical time.

Taken together, existing studies reveal profound dissensus about the possibility that divorce transmission has abated over time. At the same time, there are enough studies suggesting a decline in the negative consequences of parental divorce that it is hard to contend that no such decline has taken place. My disagreement with Li and Wu (2006, 2008) should be read in this context. Based on the present study, there is little doubt that the GSS is showing a decline in the negative consequences of growing up with divorced parents. My earlier studies using GSS data also provide evidence of a decline (Wolfinger 1999, 2003b, 2005). However, many studies based on other data sets suggest that the divorce cycle has not abated.

How can this discrepancy be explained? Although the nature of the differences between data sets cannot be known with certainty, one possibility concerns the extremely long time frame covered by the GSS. As noted earlier, my sample contains 63 years of marriage cohorts; the sample used in my earlier work spanned 92 years of marriages (Wolfinger 1999). Many nineteenth century births are represented in these data. Perhaps this is the difference from other data sets that fail to show the decline in divorce transmission reported in this article and my other work based on the GSS (Wolfinger 1999, 2005).

The literature on the consequences of parental divorce is replete with conflicts, controversies, and unresolved issues. For instance, more than 25 studies have produced conflicting evidence as to whether parental divorce leads to earlier or later marriage among offspring, or has no effect at all (Wolfinger 2003b, 2005). Fifteen studies have suggested that parental divorce leads to earlier marriage, while seven others have found that parental divorce delays marriage; still others showed no relationship. The question of trends in the divorce cycle is similarly conflicted.

The ideal solution is more research with new data that combines the best features of existing data sets. A repeated survey, like the General Social Survey, allows analysts to better separate age, period, and cohort effects in the intergenerational transmission of divorce; retrospective data based on a single cross-section, like the sample analyzed by Li and Wu (2006, 2008), allow researchers less insight into the temporal dynamics of divorce transmission (Mason and Wolfinger 2001). The ideal data set should have sufficient information for event history analysis. It should also offer information on family structure backgrounds of married couples, not just individual respondents. The family histories of both spouses contribute equally to the likelihood of divorce transmission (Amato 1996; Wolfinger 2003a, 2005). Furthermore, the divorce cycle is strongest for people who experience multiple family structure transitions in their families of origin (Wolfinger 2000, 2005). An optimal exploration of trends in the intergenerational transmission of divorce would take all these factors into account.

Conclusion

The extent to which the consequences of parental divorce have diminished over time remains unknown. Nonetheless, using a widely respected data source—the General Social Survey—and a strategy that controls for period of exposure—allowing all couples 30 years to divorce—I find clear evidence of a decline in divorce transmission. Based on completed cohorts, this analysis establishes that trends in the divorce cycle cannot be attributed to the absence of proper event history data (pace Li and Wu 2006, 2008). In addition, the GSS shows that the effect of parental divorce on offspring marriage timing has weakened (Wolfinger 2003b, 2005).

This article also demonstrates a completed cohorts approach to studying marital stability. In the absence of event history data, this technique can be used for exploring how individual characteristics affect the probability of divorce—and potentially other right-censored phenomena—without having to worry about censoring bias. The completed cohorts approach indeed has a long-standing precedent in the demographic literature, the study of completed fertility (e.g., Bumpass and Westoff 1969). It may also be applicable to studies of marriage timing, given that the probability of getting married for the first time asymptotically approaches zero past a certain age (Goldstein and Kenney 2001).

There are at least two obvious shortcomings to the completed cohort approach to divorce transmission. First, it offers no insight into how trends in the divorce cycle have changed in recent years. Second, the results may be influenced by selective mortality. The work of Linda Waite and others (e.g., Waite and Gallagher 2000) demonstrates that divorced people die younger; it has also been shown that parental divorce decreases life expectancy (Schwartz et al. 1995). Needless to say, selective mortality is a generic problem that affects much social research.