Abstract
American divorce rates rose from the 1950s to the 1970s peaked around 1980, and have fallen ever since. The mean age at marriage also substantially increased after 1970. I explore the extent to which the rise in age at marriage can explain the decrease in divorce rates for cohorts marrying after 1980 using data from the Survey of Income and Program Participation, 1979 National Longitudinal Survey of Youth, and National Survey of Family Growth. Three different empirical approaches suggest that the increase in women’s age at marriage is the main proximate cause of the fall in divorce rates.
I. Introduction
Divorce rates more than doubled in the United States between 1950 and 1979. Only one-quarter of the marriages that started in the 1950s ended in divorce but half of all unions beginning in the 1970s eventually would dissolve. After 1980, divorce rates began to fall back to previous levels. American couples marrying in 2008 are projected to divorce about 40 percent less often than those who wed at the height of marital instability.1
Many researchers have tried to explain the initial rise in divorce rates. (See Stevenson and Wolfers 2007.) Work by Becker (1973, 1974, 1991); Johnson and Skinner (1986); and Ruggles (1997) suggests that rising female labor force participation may have caused the increase.2 Similarly, Weiss and Willis (1997) suggest that improvements in female earning capacity may be responsible.3 Friedberg (1998), Parkman (1992), Peters (1986), and Wolfers (2006) provide a range of estimates of the importance of changes in divorce law in explaining the increase in divorce rates. Decreases in marital stability have been further linked to decreases in occupational segregation by gender (McKinnish 2007); the rise of the welfare state (Moffitt 1997); household technological progress (Greenwood and Guner 2008; Greenwood, Seshadri, and Yorukoglu 2005; and Greenwood et al. 2012); and changing social attitudes toward divorce (Thornton 1989; Cherlin 2004).
There is a relative dearth of information on why marriages subsequently became more stable. Stevenson and Wolfers (2007) argue that a transition from marriages with gains due to production substitutability (that is, gains from specialization and exchange within the household) to those with gains due to consumption complementarity (that is, gains from joint consumption) could explain both the rise and fall in divorce. Formal evidence on this potential mechanism is limited to two studies: Andersen and Hansen (2012) simulate a sociological model validating this idea, and Isen and Stevenson (2010) demonstrate that trends in marriage and divorce by education are consistent with this explanation.
Work by others suggests that changes in divorce laws, female labor force participation, reproductive technology, and selection into marriage may be responsible for some of the decrease in divorce rates. Rasul (2006) and Mechoulan (2006) argue that changes in divorce laws could have led to both the rise and decline of marital instability though work by Wolfers (2006) suggests family law changes cannot account for a large portion of the decrease in divorce. Neeman, Newman, and Olivetti (2008) demonstrate that under certain conditions, increases in female labor force participation can lead to decreases in the divorce rate, potentially explaining the downward trend. This hypothesis is further supported in empirical work by Mencarini and Vignoli (2014) and Newman and Olivetti (2015). Goldin and Katz (2002) demonstrate that women with early access to oral contraceptives had lower divorce rates and posit that reproductive technology may play a role in the decline in divorce by improving average match quality. Finally, decreases in rates of marriage and increases in cohabitation have been linked to decreases in the divorce rate (see Brien, Lillard, and Stern 2006). But no single explanation can explain more than a small fraction of the decline in divorce.
I demonstrate that increases in age at marriage must be a part of any explanation for the decrease in divorce rates after 1980. The strong relationship between age at marriage and divorce was first established in Becker, Landes, and Michael (1977) and has further been explored in Teachman (2002), Lehrer (2008), and Lehrer and Chen (2013). Unions beginning when a bride is 18 are twice as likely to end in divorce as those starting when a woman is 22. Brides in their mid-30s have marriages four times as stable as teenage brides. If one applies these relationships to the increase in age at marriage that occurred from 1980 to 2003, age at marriage can explain 80 percent of the decrease in divorce over this period. Furthermore, the decline in teenage marriage alone can explain 45 percent of the decrease. Indeed, holding a bride’s age constant, one cannot reject the null hypothesis that the average divorce risk was constant for marriages beginning from 1980 to 2004. Thus, age at marriage can statistically explain the fall in divorce.
Increases in brides’ ages must cause decreases in divorce if age at marriage explains the change in divorce rates in a real sense. Many reasons could justify a causal relationship between these variables. Older brides have spent more time in the marriage market and thus are better informed about their options and optimal mate. Waiting to marry also may lessen a woman’s incentive to search for a new partner during marriage because a wife’s outside options could deteriorate with age. I use several econometric techniques to provide evidence that age at marriage and divorce are causally linked and thus that the rise in age at marriage caused the change in divorce rates. Taken together, my analyses suggest that age at marriage and divorce are robustly correlated. Under a variety of conditions, these results also suggest that the rise in age at marriage is the main proximal cause for the decline in divorce.
My causal analyses all rely on having a relatively constant pool of women who choose to marry. I thus first rule out selection into marriage as an alternative explanation for the decline in divorce. I demonstrate that the trend in divorce would occur even in the absence of any change in selection into marriage using a bounding exercise. I also show that the decrease in divorce since 1980 is similar when one accounts for the rise in cohabitation by examining the stability of both formal marriages and all coresidential unions. Given this stability of the trend in divorce, I proceed to explore the evidence that age at marriage caused the decline in divorce using three analyses of the ever-married population.
The first analysis controls for the driving forces behind family change at the state-year level to determine the extent to which omitted variables may lead to a spurious relationship between age at marriage and divorce.4 My results suggest that this bias is limited. Moreover, further analysis shows that if observable variables are at least one-quarter as important as unobservable variables are in predicting age at marriage, then increases in age at marriage cause decreases in the probability of divorce.
A second method uses the marital histories of sisters to determine if differences in family background bias estimates of age at marriage’s effect on divorce. If the family accounts for a large proportion of the variation in marital stability and a small proportion of the variation in age at marriage, these results suggest a causal relationship between age at marriage and divorce.
I finally use state laws on the minimum age at marriage as instrumental variables (IVs) to pin down the causal effect of early teenage marriage on divorce. The IV procedure will estimate a local average treatment effect (LATE) of early marriage if these laws are binding for some teens but changes in the laws do not otherwise influence divorce rates.
The above analyses point to the same conclusion: Uncorrected estimates do not largely overstate the causal relationship between age at marriage and divorce. Thus, my results demonstrate that increases in age at marriage caused a large portion of the fall in divorce rates from 1980 to 2004 under a variety of conditions: if estimates are mainly biased because of omitted (measurable) variables that caused changes in family formation, if a bride’s family accounts for a large proportion of the variation in divorce risk but a small proportion of the variation in age at marriage, or if my IV estimator is valid. Taken together, estimates suggest that age at marriage can explain 60 percent or more of the decline in divorce.
I conclude this paper with a discussion of the causal pathways between the gains to marriage, age at marriage, and divorce to develop a deeper understanding of both the rise and decline in divorce rates. I explore the mechanisms behind the causal relationship between age at marriage and divorce using two unique data sets; this analysis suggests that marriages starting later in life are more harmonious and egalitarian. Furthermore, a theoretical model helps me situate my results on the decline in divorce within the rich literature exploring the causes for the earlier rise in divorce rates. The model demonstrates that although factors such as female labor force participation and wages (see Ruggles 1997; Weiss and Willis 1997) and access to reproductive technology (as in Akerlof, Yellen, and Katz 1996; Goldin and Katz 2002; and Bailey 2010) were the root causes for changes in the family, these forces largely caused divorce rates to fall after 1980 via increases in age at marriage. Thus, I demonstrate that although many different economic and social changes can be considered important distal causes of the fall in divorce, the rise in age at marriage is the main proximate cause of divorce’s decline.
II. Age at Marriage and Divorce
This work is largely motivated by three key facts: (1) Average age at marriage rose from 1970 until the early 2000s. (2) Women marrying in the 1970s and 1980s experienced higher divorce rates than any other marriage cohort in the past 60 years, with marital stability being roughly equal for the cohorts who married in the 1960s and early 2000s. (3) When one holds age at marriage constant, divorce rates are roughly constant for cohorts marrying between 1980 and 2004.
This section explores these facts and the statistical relationship between age at marriage and divorce. Section II.A establishes the stylized facts. Section II.B decomposes the decrease in divorce into changes explained by changes in age at marriage, other observable factors, and unobservable factors. Section II.C explores how younger and older brides differ in order to develop an understanding of factors other than a causal pathway that could imply a relationship between age at marriage and divorce. All analyses focus on age at marriage for women but using male age at first marriage yields largely similar results.5
A. Facts on Age at Marriage and Divorce
The first stylized fact motivating this study easily can be seen by looking at age at marriage over time, as in Figure 1. The average age of first-time brides increased by almost five years from 1970 until the early 2000s.6 First-time grooms also experienced a more than four-year increase in age.
Age at Marriage and Divorce: 1950 to 2004
Notes and sources: Average age of first-time brides and grooms by year of marriage. SIPP sample: Women (N = 62,588) and men (N = 56,109) from the 2001, 2004, and 2008 SIPP panels with complete information on first marriages, 1960–2003. Census sample: Women (N = 1,064,745) and men (N = 1,021,497) from the largest IPUMS Census sample directly following their first marriage, 1950–79 (for example, age at marriage in 1975 is calculated using the reported age at first marriage for those who married in 1975 in the 1980 Census 5 Percent Sample). Difference in SIPP and Census samples likely due to selective mortality. Divorce rates: 1950–95 from Carter, et al. (2006), 1996–2004 from United States Census Bureau (2007).
I examine the trend in divorce for marriages beginning from 1950 to 2004 to uncover my second stylized fact, using a Cox hazard regression of the form
(1)
Woman i’s hazard of divorce after t years of marriage (the risk of divorce in the tth year of marriage) is hi(t), δiy is a vector of variables indicating the year (in five-year groups) that a woman first married, and Xi is a vector of other control variables. The specification allows for a fully flexible hazard rate across the duration of a marriage but requires that covariates have the same proportional effect on the hazard of divorce for all t. I focus my analysis on divorce by year of marriage using the retrospective accounts of women’s first marriages commencing from 1950 to 2004 and reported in the 2001, 2004, and 2008 panels of the Survey of Income and Program Participation (SIPP).7
Figure 2 shows estimates of the relative hazard of divorce by marriage cohort, hy = exp (δy), from regressions controlling only for year of marriage (hy = 1 for marriages beginning from 1975 to 1979). This analysis reveals my second stylized fact: Couples marrying between 1970 and 1984 faced the highest divorce rates; marriages beginning both before and after this period were more stable, with unions beginning in the late 1990s and early 1960s dissolving at similar rates.
Hazard Rates of Divorce Across Marriage Cohorts
Notes and sources: Women’s first marriages from the 2001, 2004, and 2008 SIPP, 1950–2004 (N = 74,339). See Appendix 1.A for details. Effects are from a Cox hazard regression setting the hazard of divorce for marriages occurring from 1975–79 to one. Observations censored at time of interview or time of death of spouse. Age at marriage controls include indicators for marrying under 18, 18–19, 20–22, 23–26, 27–29, 30–34, 35–39, and 40+. Robust standard errors used to calculate 95 percent confidence intervals.
Trends in the relative hazard of divorce are different when one holds age at marriage constant, as shown by the second line in Figure 2. It is not surprising that the age-adjusted and unadjusted trends diverge, given the large increase in the age at marriage (see Figure 1) and previous findings demonstrating a negative relationship between bride’s age and divorce risk (for example, Becker, Landes, and Michael 1977; Teachman 2002; Lehrer 2008; and Lehrer and Chen 2013). There is no longer a subsequent increase in marital stability after 1980 once one controls for age effects although divorce propensities still increase for marriages beginning from 1950 to 1979. Thus, the increase in age at marriage statistically explains the decrease in divorce for women first marrying from 1980 to 2004, my third stylized fact.8
The rise and fall in divorce occurred throughout the United States and for many groups of women.9 Controlling for age at marriage mitigates a decline in divorce from 1980 to 2004 for white, black, and Hispanic women; women with a high school degree or less, some college, and those with a college degree; those living in liberal and conservative states; those living in the eastern, midwestern, southern, and western United States; and those in both urban and rural areas. The hazard rate of divorce conditional on age at marriage is roughly constant or increasing from 1980 to 2004 for many of these subgroups, including whites, noncollege graduates, those living in urban areas, and those from both liberal and conservative states. Age at marriage has a weaker effect on Hispanics and college graduates; the age-adjusted divorce rate in these groups decreased during the 1990s though the change was smaller than that suggested by the uncorrected trend.
A nonparametric function relating age at marriage to divorce further demonstrates the predictive power of a bride’s age for her marriage’s stability.10 Figure 3 shows the hazard of divorce by age at marriage (relative to age 22) in the SIPP, controlling for year of marriage fixed effects and other observable characteristics (education, race, premarital childbearing, urban location, and census division). The curve is both decreasing and convex, as initially suggested by Becker, Landes, and Michael (1977). Marriages beginning when a bride is 18 are twice as likely to end in divorce as those starting when a woman is 22. Brides in their mid-30s have marriages four times as stable as teenage brides.11
The Relationship Between Age at Marriage and Divorce
Notes and sources: Women’s first marriages from the 2001, 2004, and 2008 SIPP, 1950–2004 (N = 73,338). Marriage at age 22 is the baseline category. See Appendix 1.A for details. Observations censored at time of interview or time of death of spouse. Coefficients from Cox hazard regression also controlling for year of marriage fixed-effects, premarital childbearing, education at marriage (four groups), race (black, Hispanic, white, and other), census division, and urban location (both at interview). Robust standard errors used to calculate 95 percent confidence intervals.
B. Decomposing the Determinants of Divorce
Combining the coefficients from estimates of Equation 1 with the change in the vector of explanatory variables reveals the relative contribution of various factors to the decline in divorce rates. I use these estimates in Table 1 to decompose the change in divorce from 1980 to 2003 into components predicted by changes in age at marriage, predicted by changes in other observables, and not predicted by the included variables.12 Decompositions of changes in the divorce rate from 1980 to 1995 or 2000 yield largely similar results.
Decomposing the Change in Divorce Hazards
On average, first marriages starting in 1980 have a divorce hazard rate about 37 log points higher than those that began in 2003. Teens made up only 15 percent of first-time brides in 2003 but about 40 percent of brides in 1980. Because women who marry as teens divorce far more often than those who wait to marry, the change accounts for almost half of the fall in the average hazard of divorce. Other increases in age at marriage imply further declines in divorce, with changes in age explaining 80 percent of the total change in the hazard rate.
Brides were more educated in 2003 than in 1980 but the increase in education does not imply a large decline in divorce by itself. The doubling of the proportion of brides with a college degree only yields a decrease in the hazard rate of 5.4 log points. The hazard rates associated with other levels of education are approximately the same. Thus, the increased education of brides only accounts for about 15 percent of the change in divorce rates from 1980 to 2003.
The changing racial composition of married families also predicts a notable component of the change in the hazard rate. The proportion of Hispanic brides more than doubled from 1980 to 2003, implying a 6 log point decrease in overall divorce hazards. Additionally, women who enter a (first) marriage with a child have higher divorce rates than those who do not. Women with children increased as a percentage of first-time brides from 14 to 28 percent, and thus divorce hazard rates fell 4.5 log points less than they otherwise would have. Together, the observable variables predict a decrease in divorce from 1980 to 2003 approximately equal to the actual change.
This decomposition is robust to using male reports of (own) age at first marriage and divorce or alternative data sets, for example, the National Survey of Family Growth (NSFG).13 Moreover, using the NSFG to include demographic controls for both spouses leads to similar decompositions.14 Measurable demographic characteristics other than age can only explain about 30 percent of the change in divorce rates from 1980 to 2004. But age at marriage can account for at least three-quarters of the drop.
C. Younger Brides Differ from Older Brides
The decomposition in Table 1 may tempt one to conclude that changes in the age at marriage caused most of the decline in divorce. But the relationship between these variables cannot be taken as causal without further thought. Differences between women who wait to marry and those who do not could be driving the correlation.
The SIPP demonstrates that women who marry for the first time later in life are more educated; controlling for marriage cohort effects, the rate of college graduation increases by 1.5 percentage points if one considers a group of brides who are one year older. Later weddings are also more likely to involve blacks and women with children. Because college graduates have lower rates of divorce than other groups, this difference reinforces a positive association between age at marriage and marital stability. High rates of divorce among black women and women with children before marriage might temper this relationship.
The NSFG reveals further differences between younger and older brides.15 In this data set, waiting one additional year to marry is associated with a 1.4 percentage point increase in the probability of premarital cohabitation and a two-percentage point decrease in the probability of a shotgun wedding.16 Moreover, as a woman’s age at marriage increases, spouses’ ages move closer together but fewer husbands and wives have the same educational attainment. Younger brides are also more religious and more likely to be Catholic.17 These correlations further suggest the potential for bias in estimates of age at marriage’s effect on divorce.
III. Age at Marriage vs. Selection into Marriage
Changing selection into marriage also may lead to changes in divorce rates that may not be captured in a descriptive analysis. The number of divorces will mechanically decrease if fewer people marry. However, a more complex story is needed for selection into marriage to cause changes in the share of marriages that dissolve. In this section, I explore the potential for changing selection into marriage and the related rise in cohabitation to explain the decline in divorce. I am able to rule out these alternative explanations.18
A bounding exercise can be very useful in determining the effect of selection into marriage on divorce rates because only relatively few women choose to remain single throughout their entire lives. (See Manski et al. 1992.) One assumes that the probability of divorce for never-married women is one or zero and then calculates the resulting divorce rates. The true divorce rate that would result if these women married must lie within these bounds.
Figure 4 plots the trends derived from this exercise, focusing on the probability of divorce before one’s tenth anniversary. I assume anyone observed unmarried at age 45 will opt out of the marriage market.19 Both the lower and upper bound trends demonstrate a substantial rise in divorce before 1980 and a subsequent, nontrivial fall. Moreover, it is not possible to draw a trend for divorce rates within the generated bounds that does not include both an increase and subsequent decrease in divorce. That is, changes in selection into marriage cannot account for the decrease in divorce. Because the bounding exercise is atheoretical, this is true for changes in selection into marriage based on observable and unobservable characteristics.
Bounds on Divorce Rates Occurring if Never-Married Women Married
Notes and sources: Women’s first marriages and never married women age 45 and older from the 2001, 2004, and 2008 SIPP, 1950–2004 (N = 76,785). See Appendix 1.A for details. Bounds created by alternatively assuming that never-married women always or never divorce. I assume that the distribution of year of marriage is the same across those who marry versus those that do not by year of birth in allocating never-married women to dates of marriage.
Similarly, changes in the marriage market because of changes in selection into marriage might explain the decreases in divorce. I use the SIPP to estimate the characteristics (race, education, urban location, and presence of children) of men and women in the marriage market to address this potential threat. I control for the average characteristics of men and women choosing to marry for the first time in the year that a given woman married, defining a woman’s marriage market by her state of birth. This should capture the changing demographic profile of those choosing to wed. Additionally, I control for the average characteristics of never-married men and women in the year before a woman’s marriage and the sex ratio for single men and women.20 These variables allow me to condition on the types of men and women eligible to enter the marriage market for the first time and the relative balance of potential brides and grooms, respectively. Adding both sets of controls to Equation 1 explains only a small fraction of the fall in divorce from 1980 to 2004 (not shown).
Finally, one might worry that the rise in cohabitation influenced divorce rates via changes in selection (see Brien, Lillard, and Stern 2006). Couples may eschew marriage in favor of cohabitation when they are less certain of the quality of their match (Bumpass, Sweet, and Cherlin 1991; Manning, Cohen, and Smock 2011).21 Indeed, Lundberg and Pollak (2014) document that the share of women between 30 and 44 years of age that were currently cohabiting increased from 3 to 11 percent from 1982 to 2006–2008. The change fully accounts for the decrease in the probability of marriage among this group. This suggests that the rise in cohabitation and the rise in age at marriage may be conflated.
I turn to the 1988–2008 NSFG to assess whether the rise in cohabitation can explain the fall in divorce.22 These surveys include information on both dates of marriage and dates of cohabitation, though one must restrict analyses to relationships beginning after 1980. I use the NSFG to analyze overall trends in the dissolution of first unions. That is, I treat any coresidential union as a “marriage” and plot the resulting trend in “divorce.” If substitution from marriage to cohabitation drove the decline in divorce from 1980 to 2004, the trend in “divorce rates” for all unions should be far flatter than the one for formal marriages.
I examine the trends in relative divorce risk for all first unions, first unions lasting more than one year, and first (formal) marriages in Figure 5. The same fall in divorce rates may be seen for each categorization of unions. Moreover, controlling for the age when one entered a first union leads all of the considered trends to flatten. These findings thus demonstrate that growth in couples’ selection of cohabitation over marriage cannot explain the fall in divorce.23
Hazard Rates of Union Dissolution Across Cohorts
Notes and sources: Women’s first marriages (N = 11,933), unions (N = 15,882), or unions over one year (N = 14,381) from the 1988, 1995, 2002, and 2006–2008 NSFG, beginning 1980–2004. Unions defined as any formal marriage or cohabitation. See Appendix 1.C for details. Effects are from a Cox hazard regression setting the hazard of divorce for marriages occurring from 1980 to 1984 to one. Observations censored at time of interview or time of death of spouse. Regressions also control for survey date.
These analyses indicate that selection is not a likely driving force behind the fall in divorce rates. I therefore proceed to analyze the relationship between age at marriage and divorce within the ever-married population.
IV. The Impact of Age at Marriage on Divorce
Consider the following hypothetical experiment, which would allow one to estimate the causal relationship between a bride’s age and her marriage’s stability. Suppose that a woman selects her husband from a pool of men. If she chooses a spouse when she is older, she may make a better-informed decision. Alternatively, she may behave differently within her marriage if she waits to make her choice. (For example, older brides may have more limited outside options or greater ability to negotiate disagreements.) These effects imply that marital stability rises with age at marriage.
If one could manipulate when a woman selects her husband, one could then simply compare the stability of marriages randomly chosen to start at earlier or later ages. Such an experiment is, of course, infeasible. But comparing women who are very similar (and likely have similar requirements for a spouse) or those who marry at different ages for exogenous reasons allows me to approximate the estimates from this ideal experiment (for some group of women).
I conduct three analyses designed to build an understanding of how the correlation between age at marriage and divorce seen in Section II differs from the causal effect of age at marriage on divorce. Sections IV.A and IV.B explore whether omitted variable bias may lead to a spurious relationship between these variables. I add several variables to my hazard regression measuring more distal causes of changes in the family in Section IV.A. I add controls for family background to the regression in Section IV.B. Both analyses mitigate omitted variable bias but not endogeneity bias.24 That is, the variables added to my regression will decrease omitted variable bias but the utility of this addition is limited because age at marriage is likely a function of these variables. Neither of these analyses can yield a causal estimate of the relationship between age at marriage and divorce. Nevertheless, these sections allow me to conclude that omitted variable bias likely does not drive the estimated correlation. Moreover, a bounding exercise in Section IV.A allows me to conclude that unobservable characteristics would have to be far more important than observable characteristics in predicting age at marriage if there were no causal link between age at marriage and divorce.
In Section IV.C, I use IVs to estimate the causal relationship between marriage at young ages and divorce for a group of women who would like to marry as teenagers but are kept from doing so by marriage laws setting a minimum age. This analysis provides the most plausibly causal estimate of the relationship between age at marriage and divorce, and is able to eliminate both endogeneity and omitted variable biases. But this internal validity comes at the expense of external validity; the estimate is only valid for a limited group of women.
A. Controlling for Factors Influencing Family Structure
Past research has focused on measuring the impact of various forces on both age at marriage and divorce. Authors have linked changes in the labor market to trends in both variables.25 Increases in the availability of contraception and abortion also have been connected to different choices made by the family.26 Moreover, researchers have found relationships between age at marriage; divorce rates; and factors as diverse as welfare provision, household technological progress, family law, and social norms.27
These forces could potentially introduce a spurious correlation between age at marriage and divorce, biasing estimates of the causal effect of age at marriage on divorce. To reduce this bias, I estimate regressions of the form
(2)
where Cisy is a vector of the variables thought to influence family structure, measured at the state of birth (s) by year of marriage (y) level. The vector includes measures of access to abortion, access to oral contraceptives, rates of cohabitation, Comstock laws (limiting the distribution of contraceptives), female labor force participation, the gender gap in wages, occupational segregation by gender, unilateral divorce legislation, welfare generosity, and male wage inequality.28 θis is a vector of state indicators and Xi is a vector of individual-level controls.
I estimate the effect of age at marriage on the log hazard rate of divorce (α) using both the entire SIPP sample and the limited number of state-year variables available from 1950 to 2004 (Table 2A) as well as the period in which all the Cisy variables can be matched to the SIPP (1968–2004; Table 2B). All specifications indicate that waiting one extra year to marry is associated with a 8 to 9 percent lower hazard rate of divorce.29 Adding individual-level controls decreases the estimate of α (in absolute value) by about one log point. But the inclusion of the Cisy terms does not change the coefficient on age at marriage by more than 0.3 log points. Similar results hold when one uses several dummy variables for age at marriage instead of a single, continuous variable.
The variables in Cisy predict both age at marriage and divorce, even though adding these variables to the regression does not lessen the relationship between these factors. One can reject a hypothesis of γ = 0 at the 5 percent level in either sample.30 The full set of state-year variables also predicts almost one-third of the 4.9-year change in women’s age at marriage from 1968 to 2003. The more limited set of variables can explain 10 percent of the 5.7-year change in bridal age from 1950 to 2003.
Four potential factors could lead the impact of age at marriage on divorce to remain constant across the specifications. First, omitting some of the Cisy variables could have biased α upward, whereas the omission of others biased α downward. Together, the effects cancelled each other out. Second, even if there was a net bias, both γ and the effect of Cisy on age at marriage are small and could be easily swamped by a large unbiased value of α. Third, variables measured at the state-year level simply may not pick up much of the important variation in the factors influencing age at marriage and divorce. Even when one looks within state-year pairs, controlling for all factors varying at this level, α does not substantially change. (See Table 2, Column 4.) Finally, including Cisy in the regression may have introduced additional endogeneity bias. There is no way to rule out this issue.
Divorce Risk Controlling for the Determinants of Family Structure
Nevertheless, the estimates suggest that the omission of observable variables does not lead to incorrect conclusions about the effect of age at marriage on divorce. But there may be important unobservable variables. I thus also use the method proposed by Altonji, Elder, and Taber (2005) to determine how large the bias from omitting unobservables would have to be to explain the estimated coefficient on age at marriage.
I simplify the problem of omitted variable bias by focusing on individual estimates of the coefficients associated with indicators for marrying before a certain age (18, 22, or 28) in regressions predicting divorce before certain points in a couple’s marriage (the fifth, tenth, 15th, or 20th anniversary). I then estimate the effect of age using the 1968–2004 SIPP sample and including either no other covariates or the vector of observable controls in the regression. (See Table 3.) One can then assume that the true effect of age at marriage on divorce is zero (α = 0) and back out the implied extent of selection on unobservables (relative to observables) using these estimates.
The Relative Importance of Selection on Observables and Unobservables
Unobservables would have to influence age at marriage strongly for selection to explain the entire estimated value of α. (See Table 3.) Selection on unobservable characteristics would have to be about five times as strong as selection on observable characteristics to conclude that age has no causal effect on ten-year divorce rates. Alternatively, selection on unobservables would have to be more than 1.7 times as important as selection on observables if one allows early marriage to have an effect on divorce but imposes that there is no difference in divorce rates between those marrying before and after age 28. These levels of relative selection are unlikely because the observed variables I use include important determinants of both age at marriage and divorce (and indicators for year of marriage and state of birth).31 This suggests that age at marriage, and not some other combination of observable or unobservable factors, is the main proximate cause of the decline in divorce from 1980 to 2004.
B. Controlling for Family Background
Family and personal background also could affect both age at marriage and marital stability.32 For example, many religions advocate early marriage, limited divorce, or both. Young women who grew up in intact families may view marriage and divorce differently from those who experienced their parents’ separation. I therefore use the 1979 National Longitudinal Survey of Youth (NLSY) to control for factors such as religion and childhood family structure.33 The survey tracks a single cohort of women (age 14 to 22 in 1979), who on average married in 1984 at age 23.
The most straightforward approach to using the data adds controls for family background to the hazard regression, as in
(3)
where Fi is a vector of background variables (controls for religion; religious participation; family structure; media access; and mother’s and father’s labor force participation, education, and occupation).34 The NLSY also includes a sample of almost 900 sisters that I use to estimate regressions with family fixed effects, as in
(4)
Ideally, the inclusion of family effects or family background variables allows one to estimate the effect of age at marriage on divorce, holding some of the determinants of the gains to marriage constant. However, adding fixed effects to Equation 4 may increase the bias in α. In particular, fixed effects will only lessen the bias if the fraction of variability in divorce that the family explains exceeds the fraction of variability in age at marriage that the family explains.35
Similar to the results found when adding controls at the state-year level, including controls for factors other than education does not change the coefficient on age at marriage by more than one log point, as demonstrated by the estimates in Table 4. In addition, including family effects does little to change the value of α. (See Table 5.) Specifications that replace the linear age at marriage term in Equations 3 and 4 with a set of dummy variables also produce qualitatively similar results. Overall, the estimates from the NLSY further suggest that increases in age at marriage, and not changes in family structure, religion, or other background variables, explain most of the drop in divorce rates from 1980 to 2004.
Divorce Risk, Age at Marriage, and Family Background
Within-Family Estimates of the Effect of Age at Marriage on Divorce
C. IV Using State Age Restrictions on Marriage
The analyses in Sections IV.A and IV.B allow one to understand potential threats to a causal interpretation of the relationship between age at marriage and divorce. But they do not provide point estimates of the causal impact of age at marriage. I use an IV procedure to do this. The approach exploits variation in marriage age caused by laws limiting the earliest age that a woman can marry, with and without parental consent.36
The analysis is conducted on a sample consisting of ever-married women in the SIPP born from 1920 to 1974, regardless of year of marriage (because the regressions use instruments defined at the birth cohort level).
In total, 39 (22) states changed the minimum age at marriage with (without) parental permission, on average 2.00 (1.95) times. These changes then identify systems of equations such as
(5)
(6)
where Yit is a variable indicating if a couple divorces within t years of marriage, X and θ are defined as before (individual-level controls and state-specific indicator variables), δ is a vector of birth cohort fixed effects, (Agei < 18) is an indicator for a girl marrying prior to her 18th birthday, and Aisc is a vector of indicators for the legal status of marriage (with and without parental consent) for girls of different ages.
Age at marriage laws force many teenagers to wait to marry even if they have found a desirable spouse. The laws, however, are not binding for all teens seeking to wed. Many young women go across state lines or misrepresent their age to obtain an illegal marriage license.37 Moreover, an analysis of legal records indicates that an underage girl could sometimes receive judicial permission to marry if she could present good reason (for example, her pregnancy) to the court. My IV estimates of α are therefore LATEs specific to teens who do not circumvent these laws but would otherwise choose to marry.38
This analysis is limited to estimating the impact of age at marriage on divorce for women who would like to marry at young ages. Such marriages are rare for recent cohorts, with only 3 percent of marriages in 2003 including a woman under age 18 and 13 percent of marriages involving a woman age 18 or 19. (See Table 1.) But in the cohorts where divorce was at its peak, teenage brides made up a large share of women getting married. In 1980, 40 percent of brides were under age 20 and 10 percent were under 18.39 Indeed, changes in the proportion of brides under age 18 statistically can explain about one-fifth of the fall in divorce from 1980 to 2003, and decreases in teenage marriage altogether can account for almost half the change. (See Table 1.) Understanding the effects of marriage on teens who can be persuaded to wait to marry is therefore important for explaining divorce trends.
Minimum age at marriage laws also might discourage women from ever getting married. However, I find no evidence that more restrictive age at marriage laws during a woman’s teenage years decreased the probability that she appears in my sample of marriages. Further, more restrictive minimum age at marriage laws are not associated with higher rates of likely cohabitation in the Current Population Survey (both overall and for women under 25), suggesting these laws do not encourage nonmarital unions.40 Changes in the laws also are not preceded by trends in teenage marriage, young divorce, or single motherhood.41
The legal variables are relevant instruments, as demonstrated in Table 6 by the first stage of the IV procedure. Logically, states with more permissive laws have higher rates of early teenage marriage. Together, the variables have a joint F-statistic near 12 and a probit specification demonstrates that the strength of the instruments does not rely on the specific functional form chosen. Thus, weakness of these instruments is likely not a problem.42
First-Stage Estimates Using Minimum Age at Marriage Laws as Instruments
I use Maximum Likelihood to calculate the marginal effect of early teenage marriage from Equations 5 and 6, assuming e and ε are jointly normal. The coefficients from the IV regressions are generally larger than the non-IV estimates, as depicted in Figure 6. But the two coefficient vectors are statistically indistinguishable. Non-IV probit regressions imply that marriage before age 18 is associated with a 12-percentage point or 50 percent increase in the probability of divorce before one’s tenth anniversary (a 10 percentage point or 25 percent increase in the probability of divorce before the 20th anniversary). At the tenth anniversary, the IV and non-IV estimates are nearly identical; at the 20th anniversary, the IV estimates are about 50 percent larger than those produced by a standard probit model. The IV estimates significantly differ from zero at most anniversaries, despite their large standard errors. Additionally, these results are robust to analyzing only more recent cohorts of women; using only laws for age at marriage with parental permission as instruments; using minimum age at marriage laws for both men and women as instruments; or considering the effect of laws on subsets of women who marry by age 20, 22, or 25.43
IV Estimates of the Effect of Marriage Before Age 18 on Divorce
Notes and sources: Women’s first marriages from the 2001, 2004, and 2008 SIPP (born 1920–1974, N = 60,914). See Appendix 1.A for details. IV regression estimated using Maximum Likelihood, assuming regression stages have jointly normal errors. Observations included in estimating the effect of teenage marriage on the probability of divorce before the nth anniversary if the possible length of marriage (the difference between the date of spousal death or date of interview and date of marriage) is n or more years. Regressions also include year of marriage and state of birth fixed-effects and controls for having children prior to marriage, race (black, Hispanic, white, and other), census division and urban location (both at interview), and education (four categories). Robust standard errors clustered by state of birth used to create 95 percent confidence intervals. First stage using minimum age at marriage laws to predict age at marriage less than 18 available in Table 6.
The LATEs estimated using minimum age at marriage laws as instruments are well defined but limited because of the inability to extrapolate the results to noncompliers. Nevertheless, the analysis suggests that age at marriage can be a very important predictor of divorce, affects some group of women, and has the potential to explain at least a portion of the fall in the divorce rate for couples marrying between 1980 and 2004. Moreover, similarity between the plausibly causal estimates produced using IV and the estimates calculated using state-year fixed effects (see Table 2A) suggests that estimates using the latter method are not upward-biased measures of the effect of age at marriage on divorce.
Together, my three analyses indicate that the increase in age at marriage is a major proximate cause of the decrease in divorce rates from 1980 to 2004. Comparing the uncorrected and corrected coefficients on age at marriage further suggests that increases in age account for at least 60 percent, and potentially more, of the decline in divorce.
V. Causal Pathways Between the Gains to Marriage, Age at Marriage, and Divorce Rates
Building a deeper understanding of the relationships between the gains to marriage, age at marriage, and divorce rates can help one to further understand the causal relationship established in Section IV. This section takes two different approaches to explore why both age at marriage and divorce changed since 1950. Section V.A outlines a model wherein decreases in the gains to marriage can lead to increases in divorce rates for current marriages but decreases in divorce rates for future marriages, with the latter change occurring in part because of increases in age at marriage. In Section V.B, I explore the mechanisms behind the causal relationship between age at marriage and divorce. This analysis suggests that marriages starting later in life are more harmonious and egalitarian.
A. Accounting for Both the Rise and Fall of Divorce Rates
Past work on the family suggests that various factors relating to decreases in the gains to marriage (for example, increased access to birth control or women’s growing role in the labor market) led to increases in age at marriage.44 The empirical evidence presented in the previous sections indicates that these increases in age drove down divorce rates. My results therefore imply, when viewed within the scope of the literature, that decreases in the gains to marriage led to lower rates of divorce via increases in age at marriage. However, this hypothesis is potentially at odds with trends earlier in the century. Although age at marriage did decrease slightly from 1940 to 1960, it rose thereafter. (See Elliott et al. 2012.)45 Furthermore, most of the variables associated with the gains to marriage evolved monotonically from 1960 to 2004. But the earlier part of this period is characterized by increasing divorce rates. Similar changes in the gains to marriage thus appear to imply different changes in the divorce rate before and after 1980.
I rationalize these facts and further situate my results within the scope of the literature by using a nonstationary, one-sided search model of the marriage market, detailed in an online appendix. The key features of this model include the following:
Each period after completing her education, a woman can choose to stay single and search for a spouse, get married but continue searching for a better spouse, or get married and stop searching.
Searching for a spouse is costly both when single and when married but costs are greater during marriage.
Men are differentiated in the model by their wage rates; higher-earning men are more desirable spouses.
A woman benefits from marriage because of household economies of scale and the additional income of her spouse.
Divorce occurs when either a woman searches for and finds a better spouse while she is married or when the parameters of the model shift to make being single more desirable than the current marriage.
The model produces two decision rules: A woman should get married if she finds a potential husband with wage w>w1 and stop searching in the marriage market if she finds a potential husband with wage w>w2. The nonstationarity of the model implies that w1 and w2 vary over time and produces the negative relationship between age at marriage and divorce.
The key prediction of this model is that changes in the gains to marriage impact current and future marital stability asymmetrically. Consider an increase in the wages a woman can earn in the market, holding the male wage distribution constant. The change will lead some married women to divorce their current spouses in favor of being single. But the change will differently impact women who were single when it occurred. These women become pickier about the man they will marry and are less likely to choose to enter a marriage but continue to search for a better spouse. This leads single women to have lower eventual divorce rates for two reasons. First, the marriages they enter are intrinsically more stable. Second, they marry at later ages, which further drives down their divorce rates. These effects imply that an increase in female earning capacity first leads to higher divorce rates but eventually causes divorce rates to fall—both directly and via the rise in age at marriage. My model produces similar patterns after expansions in reproductive technology (modeled as a decrease in the relative cost of marital search while single), increases in female labor force participation and decreases in household specialization (modeled as reduced household economies of scale), and increases in the returns to female education.
The asymmetry allows this simple model to predict that monotone changes in the gains to marriage can produce a rise in age at marriage and an inverted U-shaped trend in divorce rates. That is, the model connects my empirical work on the decrease in divorce to the vast literature on the causes for the rise in divorce rates; the well-established causes for the increase in divorce also can be the distal causes for the decrease in divorce, with increases in age at marriage acting as one key mechanism for the decline.
Finally, this model captures that marriages formed under one regime will not necessarily persist under another (see Isen and Stevenson 2010, Stevenson and Wolfers 2007). The matches made in the 1950s, 1960s, and 1970s were optimal given the conditions then; however, such marriages dissolved as society changed, raising the divorce rate.46 Marriages formed under the new regime were stronger, partly because of the rise in age at marriage. That is, underlying changes in marriage led existing marriages to dissolve while making future marriages stronger.
B. How Are Earlier and Later Marriages Different?
The above model highlights one mechanism behind the causal relationship between age at marriage and divorce: women who marry later are less likely to look for a new spouse while married. But outside the abstraction of the model, there are likely several mechanisms behind the causal relationship. This section explores these mechanisms by discussing how the intact marriages of women who married earlier and later in life differ, using data from the National Survey of Families and Households (NSFH) and Marital Instability over the Life Course Panel Study (MILC).47 I regress several different variables capturing the characteristics of marriages on woman’s age at marriage, controlling for year of marriage.48
Table 7 reveals notables differences in the marriages of those who wed earlier and later in life. Results from the NSFH indicate that although increases in a woman’s age at marriage do not increase her or her husband’s reported happiness with marriage, marriages occurring later in life exhibit characteristics usually attributed to better marriages. Couples who marry later are more likely to spend quality time together (though a four-year increase in age at marriage is associated with a couple having sex about one time less per month). Furthermore, wives report arguing with their husbands less as a woman’s age at marriage increases, in particular about money, spending time together, sex, and in-laws. Women who marry at later ages are less likely to report engagement in domestic violence (both as the victim and the aggressor).
Women’s Age at Marriage and Characteristics of Marriages in the NSFH and MILC
Both the MILC and NSFH also indicate that marriages occurring when a woman is older are characterized by less traditional values; women work more in the market, spend less time doing housework, and are less likely to report a moral issue with divorce as a bride’s age rises.49 Increases in a woman’s age at marriage are also associated with different reasons for her working or wanting to work. In 1980 (when the MILC began), women who married later in life were more likely to want to work to be around people, have a career, or feel like they had accomplished something. Moreover, they were less likely to work for purely financial reasons, suggesting that couples were either more financially sound, less likely to view their jobs as simple means to an end, or both.
Although none of these relationships is necessarily causal, these results provide a rationale for the plausibly causal relationship between age at marriage and marital stability shown in Section IV. They also imply that waiting to marry may produce gains within marriage and not simply affect relationships on the cusp of divorce.
VI. Conclusion
During the past several decades, women moved into the labor force, experienced wage gains, and gained greater control over their fertility. Divorce rates first rapidly rose but then began to fall as these changes occurred. Although much is known about the initial rise in divorce, little had been previously said about its subsequent decline.
This paper demonstrates that once one controls for bride’s age, couples marrying from 1980 to 2004 face similar average risks of divorce. I use three different techniques that mitigate bias in estimates of the effect of bride’s age on marital stability to determine if age at marriage is a major proximate cause of the decline in divorce. Controlling for the major causes of family change (for example, female labor force participation, access to birth control, and divorce laws); controlling for family background (including family fixed effects); and instrumenting for early teenage marriage using state laws governing the minimum age at marriage, I provide evidence suggesting that the true, causal relationship between a woman’s age at marriage and her future probability of divorce cannot be substantially weaker than suggested by uncorrected estimates. The estimates suggest that the hazard of divorce falls by at least 6 percent when a bride waits one additional year to marry, implying that age at marriage can explain at least 60 percent of the fall in the divorce rate for cohorts marrying from 1980 to 2004. I am further able to argue that age at marriage, rather than selection into marriage, is the force behind the fall in divorce rates.
When viewed within the literature on the family and the scope of a search model, these results indicate that after first causing the divorce rate to rise, decreases in the relative value of marriage caused an increase in age at marriage, which in turn caused the divorce rate to decrease from 1980 to 2004. The underlying, distal causes of the drop in divorce are likely multifaceted and complex. But I demonstrate that the rise in age at marriage is the mechanism through which these factors caused divorce rates to fall.
Appendix 1 Data Sources
A. Survey of Income and Program Participation
The bulk of my analysis utilizes the Survey of Income and Program Participation (SIPP) panels beginning in 2001, 2004, and 2008. These data sets provide retrospective information about a respondent’s first marriage, including the year of marriage and the date and way a marriage ended, if applicable.50 I focus my analysis on the 74,339 women in these SIPP waves with complete marital records who began their first marriages between 1950 and 2004.51 I also can match 62,572 of these women to their state of birth, which I use to incorporate additional data.
B. National Longitudinal Survey of Youth (1979)
I also utilize the 1979 National Longitudinal Survey of Youth (NLSY), which follows young men and women as they marry and divorce.52 The NLSY is smaller than the SIPP (containing 3,831 women with adequate data) and may not be used to study trends over time because the data set includes a single cohort of individuals (age 14 to 22 in 1979) but it includes variables omitted from the SIPP.53 On average, women in the sample marry at age 23 in 1984. This is younger than the average age at marriage reported by the same cohort in the SIPP. The difference is likely due to attrition from the NLSY over time.
These data include valuable family background variables such as religion and the presence of a father figure within the home, making the NLSY relevant to my study. In this sample, 2,827 women also report detailed parental characteristics (mother’s and father’s labor force participation, standardized Duncan socioeconomic indices, and years of education).54 Finally, the NLSY contains 894 sisters (from 422 families), whom I use to estimate within-family regressions.
C. National Survey of Family Growth
Additional information on marriages and relationships comes from the National Survey of Family Growth (NSFG), which has surveyed ever-married women every few years since 1973. The NSFG contains a large sample of women ages 15 to 44 and rich data on fertility; some waves also contain information about a woman’s first spouse. All variables in these data sets are reported by these women; male characteristics are reported by a man’s wife or ex-wife. I select all women from the NSFG with complete marital histories for first marriages beginning from 1950 to 2004, producing a sample of 34,124 women surveyed in 1973, 1976, 1982, 1988, 1995, 2002, or 2006–2008.
I also consider women surveyed in 1988, 1995, 2002, or 2006–2008 who have ever married or cohabited to construct alternative union histories, considering any coresidential union instead of only formal marriages. This yields a sample of 15,882 unions. I classify a nonmarital union as dissolved if a woman reports she stopped living at the same address as her partner. Cohabitations that transition to marriage are considered a single union with length equal to the length of cohabitation plus the length of marriage.
D. Data on Marital Attributes
I finally utilize data from the initial waves of the National Survey of Families and Households (NSFH) and the Marital Instability over the Life Course Panel Study (MILC), which provide personal details about the marriages of younger and older brides. I select marriages beginning from 1950 to 1984 in the NSFH and those beginning from 1950 to 1979 in the MILC (the former survey was conducted in 1988 and the latter in 1980), yielding a total of 3,465 marriages in the NSFH and 1,610 marriages in the MILC. Because these surveys contain information about current marriages, I use only observations for families with a woman married to her first spouse at the date of the interview. I use data reported by wives in the NSFH, although reports from husbands demonstrate similar patterns. Data from the MILC come from a designated survey respondent, regardless of gender.
Analysis of the NSFH data required that I manipulate and combine several variables. I calculated variables for times per month a couple engaged in an activity by taking less than once per month = 0.5, once per month = 1, several times per month = 3, once per week = 4, several times per week = 10, and almost every day = 20. Disagreement variables are indicators for if a couple deals with disagreements in a particular way “sometimes” or more often. Finally, I estimate hours of housework as the sum of usual weekly hours in nine activities.
Rates of Divorce by Specific Anniversaries
Notes and sources: Women’s first marriages from the 2001, 2004, and 2008 SIPP, 1950–2004. Marriages excluded if possible duration is less than five or ten years based on dates of marriage, survey, and/or spousal death (N = 67,054 for Panel A and 58,981 for Panel B). See Appendix 1.A for details. Adjusted estimates are means predicted from OLS regression with controls for age at marriage being under 18, 18–19, 20–22, 23–26, 27–29, 30–34, 35–39, and 40+.
E. State-Level Data
This section describes the state-year variables that I use in the analysis of Section IV.A.55 Variables are measured using the average value for the five years prior to a woman’s marriage and matched to individuals using state of birth and year of marriage.56
Abortion Access: the number of abortions per woman age 15–44. Numbers for 1970–72 come from the Center for Disease Control (1971, 1972, 1974) and those for subsequent years are from the Guttmacher Institute (Jones and Kooistra 2011). I use information on abortion by state of occurrence to create the longest sample possible and assume a rate of zero prior to 1970 (when abortion was first legalized on demand in five states). This measure, rather than access to legal abortion, is used because, although access may be better for establishing a causal relationship, the number of abortions captures a broader measure of variation in prevalence.
Cohabitation: I use the March Current Population Survey (CPS) for 1963–2004 to identify those households involving “likely cohabitation” using Manning’s (1995) definition: two unmarried, unrelated, opposite-sex adults over age 15 living together with no other people above age 15. Manning shows that this definition gives aggregate estimates of cohabitation close to actual rates.
Comstock Laws: an indicator for a state sales ban on contraceptives, using Bailey’s (2010) preferred classification.
Reproductive Technology: an indicator for the earliest year an unmarried, childless woman could consent to contraception, following the coding recommendation of Bailey et al. (2011).
Unilateral Divorce Law: Wolfers’ (2006) preferred classification of state unilateral divorce laws and reforms.
Welfare Benefits: The measures available for welfare vary over time. Benefit levels for early years are taken from the Statistical Abstracts of the United States (1941–95). For 1940–44 and 1946–52, I use total benefits paid out/total families receiving benefits in June; the numbers for 1945 are from December. Benefits for 1951 and 1952 include reimbursement for medical services. For 1953–65, the data sources provide average payments per family in December including medical benefits and such numbers without medical benefits during 1966–73. For 1974–95, average monthly benefits per family are taken over all months, excluding Medicaid payments. Starting in 1995, intermittent reports from the Administration for Children and Families (1995, 1997, 1999, 2001–2004) became available, allowing one to calculate the average monthly benefit. I log-linearly interpolate the missing years. Finally, the Integrated Public Use Microdata Series (IPUMS) of the censuses provide estimates of the proportion of unmarried mothers who have one, two, three, or four or more children at the state-by-year level. I create my final measure of log welfare benefits using the errors from a regression of the log benefit level on these proportions and year fixed effects.
Several variables come from samples of workers in the 1964–2005 CPS. I drop anyone living in group quarters or with incomplete demographic information (on education, marital status, and state of residence) to calculate these measures. The sample includes only those age 18–35 because my analysis largely deals with people earlier in life.
7. Female Labor Force Participation: the proportion of married women working any hours for some number of weeks last year and the proportion of women who worked 30 or more hours in the past week and 50 or more weeks in the past year (full-time, full-year).57
8. Segregation by Gender at Work: I use CPS data on those working any hours to compute the segregation index by occupation-industry cells. Both occupations and industries are classified into 16 groups (available upon request). I then calculate the segregation index for a given state at a given time as
where i indexes industry, o indexes occupation, Nf(Nm) is the total number of working females (males) in the state at a given time, and Niof(Niom) is the number of females (males) primarily working in industry i and occupation o. I also create variables indicating the proportion of women working in traditionally male and traditionally female jobs, where traditionally male (female) jobs are defined as containing more than 95 percent male (75 percent female) workers in the 1950 Census.
Additionally, I use wage data from the CPS to control for both male wage inequality and the gender gap in wages. I use the sample of full-time, full-year workers from the CPS and drop any workers in the armed forces, agricultural sector, or private household sector from the sample. I also omit any observations with allocated or missing wage and salary income and multiply any top-coded income measures by 1.5. The hourly wage is calculated by taking yearly wage income and dividing it by 50 times the number of hours worked last week. Any observations with nominal hourly wages below the minimum wage or a wage rate that would be top-coded if a person worked 30 hours per week for 52 weeks of the year also are removed from the sample.
9. Gender Gap in Wages: the difference in the log median wages of men and women, calculated using the above sample.
10. Wage Inequality: the difference in the 90th and 50th, as well as 50th and tenth, percentiles of the log wage distribution in the above sample.
F. State Age at Marriage Laws
The majority of data on minimum age at marriage laws comes from the 1933–2001 editions of the World Almanac and Book of Facts. The almanac stopped reporting these laws in 2001, and thus I use the database of the Cornell Legal Information Institute for the 2001–2004 laws. When the two sources for these laws do not match, information from state legislative archives resolves the conflict. If a law allows for marriage before the age of majority, but no age limit is specified, I set the minimum age of marriage with consent to 12, the common law minimum age for girls.58
Many states changed their laws throughout time but some changes reported in the almanac may be erroneous.59 If a law changes for one or two periods only, then switches back, I remove the change. If a law changes for one period and then changes again, and the changes do not move in the same direction, the first change is set to the original value. Massachusetts and Montana are omitted from the analysis because there are many changes in the recorded laws that may or may not coincide with actual legislation. Appendix Figure 2 shows the evolution of minimum age at marriage laws over time for women, with (Panel A) and without (Panel B) parental permission. The figures show an increase in the allowed age at marriage with parental consent but a decrease in the age without such consent.
Minimum Age at Marriage Laws Over Time
Notes: See Appendix 1.F for details.
I also looked at trends leading up to (first) changes in laws (shown in Appendix Figure 3) to determine whether law changes were driven by rates of young marriage, childbearing, or divorce. One sees no clear trend in teenage marriage; in the proportion of children under 11 living with young, unmarried mothers; or in divorce among those under 25 before an increase or decrease in the minimum age, making these laws plausible instruments.
Young Marriage, Single Motherhood, and Divorce Before Changes in Marriage Laws
Notes and sources: See Appendix 1.F for details on law changes. Divorce and teenage marriage rates from CPS 1963–2004. Child living conditions from largest IPUMS Censuses (log-linearly interpolated for intercensal years). Residuals from regressions of given values on state and year fixed effects plotted.
Footnotes
Dana Rotz is a Senior Researcher at Mathematica Policy Research. The author would like to thank Roland Fryer, Claudia Goldin, and Larry Katz for continued guidance and support on this project as well as Timothy Bond; Richard Freeman; Jeff Liebman; Claudia Olivetti; Laszlo Sandor; Justin Wolfers; several anonymous referees; and participants at several seminars for helpful comments and discussions. She is also grateful to Phillip Levine for providing data on abortion rates. This research has been supported by the NSF-IGERT program, “Multi-disciplinary Program in Inequality and Social Policy” at Harvard University (Grant No. 0333403). The views expressed herein are those of the author and not necessarily those of Mathematica Policy Research. The data used in this article can be obtained beginning six months after publication through three years hence from Dana Rotz, 955 Massachusetts Avenue, Suite 801, Cambridge MA, 02139, drotz{at}mathematica-mpr.com.
Supplementary materials are freely available online at: http://uwpress.wisc.edu/journals/journals/jhr-supplementary.html
↵1. See Goldstein (1999); Kreider and Ellis (2011); and Stevenson and Wolfers (2007, 2011).
↵2. See Oppenheimer (1997) for an alternative view or Özcan and Breen (2012) for a survey of this literature.
↵3. See also recent related work by Bertrand, Pan, and Kamenica (2015) and Guvenen and Rendall (2015).
↵4. See Stevenson and Wolfers (2007) for a summary of these factors.
↵5. This paper does not attempt to disentangle the differential impacts of male and female age at marriage.
↵6. The changes in age may be overstated in the Survey of Income and Program Participation (SIPP) before 1960 because of selective mortality. However, most of the age change is concentrated after 1960, when selective mortality is relatively unimportant. Estimates of average age at marriage using data from the 1960–1980 censuses demonstrate this effect. (See Figure 1).
Changes in age at marriage occurred across all age quantiles, causing a shift in the variable’s distribution. See Stevenson and Wolfers (2007) for a discussion of the causes of these changes.
Key papers providing explanations for trends in age at marriage using a myriad of factors include Akerlof, Yellen, and Katz (1996); Bitler, et al. (2004); Becker (1973, 1974, 1991); Brien, Lillard, and Stern (2006); Ellwood and Bane (1985); Goldin and Katz (2002); Gould and Paserman (2003); Loughran (2002); and Mechoulan (2006).
↵7. See Appendix 1.A for details.
↵8. Trends both with and without age controls are robust to different specifications of the hazard function. Analyzing divorce rates for couples in the nth year of marriage also yields similar conditional and unconditional trends. See Appendix Figure 1 for trends in divorce by the fifth and tenth anniversaries.
↵9. Americans are not alone in experiencing increasing age at marriage and marital stability after 1980. The United Nations (2009) provides measures of the singulate mean age at marriage (calculated using the marital status of a country’s population by age to estimate the average number of years a member of the population was single) and divorce rate around 1980 and 2004 for 28 OECD countries. Age at marriage rose over the given period in each country by between 0.7 years (Norway) and 7.4 years (Belgium). Divorce rates also decreased in many of these countries throughout this time (for example, Canada, Germany, and the United Kingdom). However, no statistically significant relationship exists between the change in age at marriage and the change in divorce rates cross-nationally.
↵10. See Lehrer (2008) and Lehrer and Chen (2011) for more detailed discussions of this relationship.
↵11. The convexity of the relationship also suggests caution when interpreting a coefficient on a linear variable for age at marriage. Further analysis of the relationship between brides’ ages and divorce rates indicates that age at marriage has a roughly constant effect across marriage cohorts when one holds constant the distribution of age at marriage (as in Dinardo, Fortin, and Lemieux 1996). Moreover, one can show that marriages beginning earlier in life have higher divorce rates early in marriage. But after a couple’s tenth anniversary, age at marriage becomes less predictive of divorce.
↵12. Age at marriage jumps from 2003 to 2004 in my SIPP panels although it is not clear if this change is real or due to sampling error or changes in survey methodology. Thus, I consider the change from 1980 to 2003 instead of from 1980 to 2004.
↵13. See Appendix 1.C for details.
↵14. One cannot observe the characteristics of both husbands and wives in the SIPP.
↵15. See Appendix 1.C for details.
↵16. Defined as a couple marrying zero to eight months prior to a woman’s first birth.
↵17. Religiosity from the National Longitudinal Survey of Youth (see Appendix 1.B), measured using frequency of worship.
↵18. Note that my IV technique also employs variation unrelated to eventual marriage rates and thus is not biased by selection effects.
↵19. I use the distribution of year of marriage by year of birth for those who marry to allocate never-married women to dates of marriage. Similar results hold if one focuses on different anniversaries or different cutoff ages.
↵20. The sex ratio captures the number of men per woman in the population or marriage market. (See Angrist 2002, Grossbard-Shechtman 1993, or Guttentag and Secord 1983 for details.) I construct the sex ratio in a variety of different ways, defining a woman’s marriage market alternatively based on age and race. Different measures were constructed based on both the population as a whole and the population of single individuals.
↵21. Conversely, see Manning and Cohen (2012) and Jose, O’Leary, and Moyer (2010).
↵22. See Appendix 1.C.
↵23. Additionally, controlling for premarital cohabitation in regressions predicting divorce risk does not influence the shape of divorce trends conditional or unconditional on age at marriage.
↵24. See Section V for a model exploring age at marriage and divorce as results of a more complex decision-making process.
↵25. See Becker (1973, 1974, 1991); Blau and Van der Klaauw (2013); Gould and Paserman (2003); Johnson and Skinner (1986); Loughran (2002); Oppenheimer (1997); Ruggles (1997); Santos and Weiss (2015); Watson and McLanahan (2011); and Weiss and Willis (1997).
↵26. See Akerlof, Yellen, and Katz (1996) and Goldin and Katz (2002).
↵27. See Bitler, et al. (2004), Ellwood and Bane (1985), Hoynes (1996), and Moffitt (1997) on welfare; Greenwood, Seshardi, and Yorukoglu (2005); Greenwood and Guner (2008); and Greenwood et al. (2012) on household technological progress; Friedberg (1998), Mechoulan (2006), Parkman (1992), Peters (1986), Rasul (2006), and Wolfers (2006) on family law; and Cherlin (2004) and Thornton (1989) on social norms.
↵28. Appendix 1.E contains details on these variables and their measurement.
↵29. The average yearly hazard rate of divorce is 2.0 percent at t = 10, 1.4 percent at t = 20, and 0.6 percent at t = 30.
↵30. When individually included in the regression, most of the Cisy variables do not have coefficients significantly different from zero. However, the significant γ coefficients are consistent with the model presented in Section V.A and an online appendix. Further, when individually added to the regression, none of the variables meaningfully changes α.
↵31. One could alternatively estimate the extent of selection on unobservables and observables that would be required to reduce the estimate of the relationship between age at marriage and divorce to that implying that age at marriage explains some fraction of the decrease in divorce. As shown by Oster (2013), this also requires one to set the maximum extent to which unobservable variables (at the point of marriage) also predict divorce. A variety of tests suggest that if observables are about as important as unobservables in predicting age at marriage but that observables are 50 percent more predictive than unobservables for divorce, age at marriage can still explain about half the decline in divorce rates from 1980 to 2004.
↵32. See Gruber (2004) and Trent and South (1989).
↵33. These variables are omitted from the SIPP. See Appendix 1.B for details on the NLSY.
↵34. Because the NLSY contains a cohort of women born within a relatively short time span, the regression omits controls for year of birth or year of marriage.
↵35. Griliches (1979) shows that this is a necessary and sufficient condition for the addition of fixed effects to reduce bias if there is no measurement error. Stronger assumptions may be needed if age at marriage is measured with error. One cannot test this assumption, but the inherent randomness of the marriage market makes it more likely to hold.
↵36. Each girl is matched to a vector of indicators for the laws prevailing in her birth state when she is age 16. Dahl (2010) previously used this instrument to determine the relationship between early marriage and welfare receipt. See Appendix 1.F for details on these laws.
↵37. See Blank, Charles, and Sallee (2009).
↵38. This LATE also could reflect the effect of a culture that encourages such early marriages or a combination of such culture and the act of marrying at a young age.
↵39. Note that one should not think of teenage marriage and teenage pregnancy as concurrent events. Over the period analyzed in the NSFG (see Appendix 1.C), 35 percent of brides under 18, 20 percent of brides age 18 to 19, and 10 percent of brides age 20 and older give birth fewer than nine months after marriage.
↵40. See Appendix 1.E for details on the measure of likely cohabitation. One might also worry that limits on teenage marriage decreased age at marriage for those with ages just above a cutoff, violating the monotonicity assumption of instrumental variables. However, there is no evidence that raising the minimum age at marriage leads women above the new minimum to wed earlier.
↵42. Estimates using limited information maximum likelihood, which Stock and Yogo (2002) show to be more robust to weak instruments, confirm my results.
↵43. However, the estimates are somewhat sensitive to the specific form of the regressions used to calculate the effects in Equation 6. One could instead model both stages of the regression as linear or use a linear first stage and a hazard function for the second stage. Many combinations may be of interest. The specification shown has the most conservative point estimates of the attempted combinations, suggesting the true effect of age at marriage on divorce may be larger. Other specifications yielded less precise point estimates though effects were often statistically different from zero.
↵44. For example, see Becker (1973, 1974, 1991) on female labor force participation and Goldin and Katz (2002) on the pill.
↵45. Note that it is feasible that this decrease in age at marriage is responsible for some portion of the rise in divorce as well.
↵46. Note that although marital instability started to rise with marriages beginning in the 1950s, Figure 1 demonstrates that the number of divorces per marriage did not begin to increase until the late 1960s, when age at marriage also began to change.
↵47. See Appendix 1.D for details on these surveys. Questions in the NSFH are answered by the wife; a designated survey respondent answered all questions in the MILC. Using a husband’s responses in the NSFH yields similar results.
↵48. Regressions control for year of marriage indicators in five-year groups and the gender of the respondent in the MILC. In the regressions analyzing characteristics of current marriages, one could control for duration of marriage (age at interview). However, the NSFH and MILC conducted all wave I interviews at roughly the same time. Therefore, duration of marriage (current age) and year of marriage (given age at marriage) are collinear and one cannot control for both in the same regression. Using either alternative set of variables leaves the qualitative results mostly unchanged.
↵49. Although individuals with these characteristics simply may be more likely to marry at later ages.
↵50. My analysis treats length of marriage as a censored variable if a given marriage has not ended because of divorce.
↵51. Average age of marriage jumps in 2004, although such a change may be the result of differences in coding across SIPP panels or noise. Except when looking directly at trends in age at marriage, I include 2004 in my analysis. Results are robust to the inclusion or exclusion of this year.
↵52. Marriages occurring before a young adult enters the panel also are recorded retrospectively.
↵53. Other NLS cohorts exist but cannot be used because of limited variation in age at marriage (the 1968 young women’s cohort) or an insufficient time horizon (the 1997 cohort).
↵54. I assign a Duncan score to those who do not work, as detailed in Dworkin (1981).
↵55. Both Stevenson and Wolfers (2007) and Greenwood and Guner (2008) also discuss the possibility that household technological progress influenced trends in marriage and divorce. These factors are likely important but one cannot measure them at the state-year level, and thus I must omit them from my analysis.
↵56. Other choices for the form of Cisy yield similar results. Sophisticated methods of lag selection are computationally infeasible because of the large number of possible combinations of my 14 variables of interest.
↵57. These are the only continuously available measures of work hours and weeks in the CPS from 1963 to 2005.
↵58. The age at marriage without parental consent was recorded by the Almanac to be the age of majority if no law provided for early marriage.
↵59. See Dahl (2010). Dahl’s work uses laws for a somewhat different period (1935–69 versus 1936–90 in my main analysis). I use the same data sources and process the raw data in a similar manner.
- Received February 2014.
- Accepted May 2015.
References
In this issue
Jump to section
Related Articles
Cited By...
- Longer Time for School: The 1981 Marriage Law and Rural Female Education in China
- Making Big Decisions: The Impact of Moves on Marriage among U.S. Army Personnel
- Risk Factors and Pivotal Periods: Marital Status over 28 Years for Parents of Individuals with Autism
- Does Access to Family Planning Increase Childrens Opportunities?: Evidence from the War on Poverty and the Early Years of Title X