We begin by examining the trend in physical custody. Figure 1 shows significant change over time and demonstrates an acceleration in the trends documented in the earlier research. For example, Cancian and Meyer (1998) documented that between 1986 and 1993/1994, the share of cases awarded mother–sole custody fell from 80 % to 74 %. Here we show that by 2008, mother–sole custody declined further to 42 %. This decline is largely mirrored by a dramatic increase in shared custody: equal shared custody increased from 5 % to 27 % of all cases, and unequal shared custody increased from 3 % to 18 % of all cases. Most of the unequal shared custody cases—more than 80 %—have children staying with mothers the majority of the time (mother–primary shared custody). There is little change in the share of cases that are awarded father–sole custody: 11 % in 1988 and 9 % in 2008. Overall, the trend away from mother–sole custody and toward shared custody is dramatic, representing a substantial change in the living situations of children of divorce over a relatively short period. Figure 1 also shows that this change occurred throughout the period and does not seem to be greatly affected by the 2000 change in the custody statute explicitly stating a preference for placement with both parents.
In the analysis of the early period, Cancian and Meyer (1998) showed that shared custody is more common when both parents (rather than just the father) are employed, among parents with higher incomes, when the mother has a prior marriage, when the father does not have children from a previous relationship, when all the children are boys, and when only the father has a lawyer. They also showed an increase over time, controlling for these and other covariates. (As we show later, the analyses reported here for the early period, with one less year of data, largely show the same results.)
The distribution of some of these characteristics has changed substantially over time, but others have been more stable, as shown in Table 3 in the appendix. For example, mean total family income increased from about $65,000 in the early period to $80,000 in the recent period (in constant 2012 U.S. dollars), which was just prior to the height of the Great Recession. This is perhaps related to marriage being increasingly concentrated among higher-status individuals (see, e.g., McLanahan 2004). On the other hand, the share of that total income from the mother was fairly stable, increasing from 39 % to 41 %. Mothers are older in the more recent cohort (mean age of 36, compared with 33), and both parents are somewhat more likely to have had a prior marriage. The court process has also changed over the period: in 53 % of cases, both parents had a lawyer in the early period, declining to 40 % in the recent period.
Table 1 examines whether family characteristics are related to custody outcomes, showing the results of the multinomial logit models. Table 2 shows predicted probabilities of the three custody outcomes, given the coefficient estimates in Table 1.
Predicted probabilities that are derived from the estimates of nonlinear models, such as the multinomial logit, and that use mean values for independent variables need not match the observed probabilities (Borooah 2002). Thus, the first row of Table 2 need not match the observed data, which are shown in the bottom row of Table 2 (and derived from Fig. 1).Footnote 7 Nonetheless, the base estimates in the first row of Table 2 are similar to the observed custody outcomes: cases in the early period were very likely to be predicted to have mother–sole custody (87.1 %), this prediction declined to 52.3 % in the recent period, and the proportion of cases estimated to have shared custody increased dramatically from 7.5 % to 44.2 %. Father–sole custody is predicted to show a slight decline between the periods. Although the levels of predicted probabilities do not precisely match the observed custody outcomes, the trends are quite similar.
To examine the relationship between custody outcomes and income, in rows B–G, we show estimated probabilities for three levels of total income and vary each level by whether the father’s share of income is 80 % or 50 %. The coefficients on Table 1 and the estimated probabilities in Table 2 show that cases with higher total income are more likely to have shared custody in both periods; in the early period, the relationship was particularly strong, with the probability of shared custody about doubling when income rose from $30,000 to $75,000, and almost doubling again when it rose to $120,000. In the more recent period, shared custody remains more likely for higher income families, but the magnitude of the variation in outcomes by income is smaller. Turning to the mother’s share of total income, we see in Table 1 that when mothers have a higher share of income, all else equal, father custody is statistically less likely. However, the simulations show that estimated probability for father–sole custody is not greatly affected by the parents’ relative income in the most recent period; at each level of total income, father–sole custody declines by about 2 percentage points when we compare mothers whose incomes are one-half the total with those whose incomes are only one-fifth of the total.
In both periods, parents with previous marriages or previous children are less likely to receive sole custody. For example, when the mother has a previous marriage and previous children but the father does not (row H, Table 2), father custody is about twice as likely as in the base case. Similarly, if the father has a previous marriage and previous children but the mother does not (row I, Table 2), he is predicted to receive sole custody only about 1 % of the time.
Table 1 shows that child’s age and gender were important to custody outcomes in the early period: fathers were unlikely to get sole custody of young children, and shared custody was more likely than mother–sole custody when all children were boys. Most of these relationships are no longer statistically significant in the most recent period. The simulations in Table 2, rows J and K, show large differences in estimated custody outcomes in the early period between a case with one young girl and a case with two older boys; in the more recent period, the case with two older boys is more likely to receive father–sole custody, but the gender and ages do not matter as much in the later period.
The court process remains strongly related to custody outcomes. In cases in which only the mother has a lawyer and is the plaintiff, she was predicted in the early period to nearly always receive sole custody, as shown in row L of Table 2. In the recent period, the predicted probability of shared custody in this type of case is 22.7 %. In the cases in which only the father has a lawyer and is the plaintiff (row M), he was predicted in the early period to receive sole custody in about one-third of the cases, and the mother was predicted to receive custody in about one-half the cases; in the recent period, one-half of the cases are predicted to have shared custody. Although the probabilities would seem to suggest that the relationship between these variables and the outcomes differ between the two periods, the relationships are not measured precisely, and the coefficients on these variables in the early period are not statistically different from the comparable coefficients in the later period.
We have documented very large changes in the custody outcomes of the children of divorce over this 20-year period. These trends may be related to changes in the characteristics of cases that get divorced or to changes in the divorce process. To explore these issues, we conduct two additional simulations. In the first, we take the mean characteristics from the early period and apply the coefficients from the recent period. This results in an overall estimate of 51.9 % with mother–sole custody, 43.5 % with shared custody, and 4.6 % with father–sole custody. Comparing these estimates to the last columns of the top row of Table 2 highlights the role of changes in process: even if the characteristics had stayed constant, we would still see a dramatic change in custody outcomes—nearly the change that we observe.
In the second simulation, we explore a different counterfactual, examining the predicted outcomes if the characteristics of cases had changed to be what they are observed to be in the recent period but the coefficients were still those of the early period. This results in an estimate of 87.9 % with mother–sole custody, 7.8 % with shared custody, and 4.3 % with father–sole custody. These estimates are remarkably similar to the predicted outcomes in the early period. This again suggests that the trend in custody is less the result of changes in the characteristics of cases than the result of changes in the process of awarding custody.
Examining this issue more closely, what about the process has changed? One way to examine this question is to pool the data across the three periods (including the intermediate period not yet analyzed) and allow the relationships between selected variables and custody to differ in different periods. Estimates from a pooled model suggest that the share of income attributed to the mother does not have a markedly different relationship with custody outcomes in the three periods, nor do the employment variables or the variables reflecting the legal process. In contrast, the relationship between total income and custody changes across the periods, as can also be seen in Tables 1 and 2. Moreover, in the pooled model, the variable representing the period has a very large coefficient, suggesting that the time trend itself is a very important part of the changing process.
These results suggest that although the characteristics of divorcing cases are changing over time, the change in composition does not seem to explain the majority of the observed changes in custody outcomes. Although we find some evidence of changes in the importance of some characteristics for custody outcomes, these changes also do not seem to explain much of the transformation of child living situations. Instead, the explanation most consistent with the patterns is a change in norms and the custody determination process, with the growing adoption of the innovation of shared custody, across the distribution of characteristics.