She Left, He Left: How Employment and Satisfaction Affect Men’s and Women’s Decisions to Leave Marriages

Abstract

Most past studies examining determinants of divorce have ignored differences between the factors that elevate wives’ and husbands’ initiation of divorce. We use three waves of the National Survey of Families and Households and a latent class model embedded in a competing-risks event history model to assess distinct predictors of wives and husbands leaving marriages. We assess who left using each ex-spouse’s answer to a question that asked who had wanted the breakup more. We find that when men are not employed, either husbands or wives are more likely to leave. When wives report better than average marital satisfaction, their employment affects neither their nor their husbands’ exits. However, when wives report below average marital satisfaction, their employment makes it more likely that they will leave. We compare findings to predictions from two theories: an institutional perspective that sees divorce to result from a violation of gender-specific norms, and exchange/bargaining theory, which posits that resources allow a spouse to leave if exchange and bargaining fail to provide a satisfactory marriage. To amend the theories to better fit the data, we foreground the asymmetric nature of gender change in recent decades, with women changing more than men.

In the U.S. legal and social system, it takes only one person’s decision to end a marriage. Indeed, in many divorces, one person initiates the breakup while the other still prefers that the couple remain together. Past research consistently shows that approximately two-thirds of divorces are initiated by women, whether “initiation” is assessed by asking who wanted the divorce more or by who filed legal papers (for the U.S. see Brinig and Allen 2000; Kitson 1992; Spanier and Thompson 1984; Wallerstein and Kelly 1980; see Hobcroft 2000 on the U.K.; Svedin 1994 on Sweden; Hewitt, Western, and Baxter 2006 on Australia; Kalmijn and Poortman 2006 for the Netherlands; and Boigeol and Commalle 1974 on France). Despite this consensus in the research that women initiate most divorces, and growing interest in studying how marriage is a gendered institution (Risman 1998), we know of no published studies that analyze determinants of divorce in the U.S. while distinguishing between cases in which he left and she left.

Because women are the ones who want and initiate about two-thirds of divorces, we may have been seriously misled by findings from the large body of research on determinants of divorce which combines wife- and husband-initiated divorces as one outcome. Factors presented as determining “divorce” will typically be those that predict women’s, but not necessarily men’s initiation. This follows from the fact that coefficients in models predicting divorce are, in effect, a weighted average of the variable’s effect on women’s initiation and its effect on men’s initiation, with the weight tilted heavily toward what causes women’s initiation because there are many more such cases. Consider, for example, the curious finding that wives’ assessment of how happy their marriage is predicts subsequent divorce, while husbands’ analogous assessment shows no significant effect (Heaton and Blake 1999; Nock 2001; Sayer and Bianchi 2000). Rather than implying that women leave over unhappiness but men don’t, the finding may be an artifactual result of using a (representative) sample in which most divorces were initiated by women and an analysis that fails to distinguish between the competing hazards of husband- and wife-initiated divorces. If male-initiated divorces had been separately measured and analyzed, perhaps we would have found that, as intuition suggests, men’s unhappiness predicts men’s exits just as women’s unhappiness predicts women’s exits. One contribution of our study is to isolate the determinants of the minority of divorces initiated by men that past research designs have obscured.

Some theories of divorce do imply which spouse leaves, but researchers have tested them with data and analyses that do not distinguish who leaves. An example is the “women’s independence effect,” the idea that wives unhappy with their marriages are more able to leave if they can support themselves and their children, and, thus, that women’s employment increases the odds of divorce, not by making marriage unhappy, but by allowing women’s exits from unhappy marriages. The clear implication of this theory is that women’s employment encourages women’s, not men’s, departures, but this has not been tested in a published paper on U.S. data. Another of our contributions will be to test this idea for the first time.

The increase in women’s employment has changed marital dynamics, but the institution of marriage is still substantially gendered (Bittman, England, Sayer, Folbre, and Matheson 2003; Risman 1998). This may imply different determinants of divorces initiated by wives and husbands, as well as the need to “gender” theories of divorce which have either made out-of-date assumptions about the gender system, or ignored it completely. An example of the latter is a perspective on divorce that comes from exchange theory in sociology or bargaining theory in economics, out of which the “women’s independence” idea discussed above comes. The theory emphasizes that the partner who has more resources (such as earnings from employment) can more successfully bargain for what they want in the marriage, or, failing that, can decide to exit with less economic loss. But this theory needs to be gendered to recognize that women may have more difficulty than men bargaining for what they want within marriage. Women may find it particularly difficult to bargain for husbands’ participation in household work; men may resist, less because of the work itself than because they find taking on “women’s” work stigmatizing (Brines 1994). This gendered obstacle to women’s successful bargaining suggests a larger effect of resources on exits for women than men.

The main contribution of our research is to uncover the separate determinants of men’s and women’s initiation of divorce. In any given period, a marriage may stay together or dissolve – and there are competing risks that the wife may leave or the husband may leave the marriage. We estimate a model predicting these outcomes using three waves of the National Survey of Households and Families (Sweet and Bumpass 2002). The innovation of our study is that, drawing on both ex-spouses’ reports of who wanted the breakup more, combined with a latent-class model embedded in a competing-risks event history model, we classify each divorce according to whether the wife or husband initiated it, and examine the distinct determinants of women leaving men and men leaving women. We focus on the effects of each spouse’s employment, and each spouse’s satisfaction with the marriage.

Theories Regarding Determinants of Divorce

Past theorizing on determinants of divorce has focused on economic factors and on marital satisfaction. We first review two theories with special attention to whether they generate predictions about how determinants of male-initiated and female-initiated divorces differ. We then discuss how the gendered nature of marriage has changed, its remaining gendered features, and ways this suggests that theories may need to be “gendered” to make accurate predictions. We then review what past research on the various factors has shown.

Marriage as a Gendered Institution Supported by Cultural Norms

One sociological view of divorce sees marriage as a social institution (Amato, Booth, Johnson, and Rogers 2007; Nock 1998; Wilcox and Dew 2009). In this view, there are widely shared cultural understandings of what a marriage is supposed to be, and when these norms are violated, either party is more likely to leave, because the marriage has less social support and/or spouses feel dissatisfied as a result of violating norms. Examples of taken-for-granted assumptions include that marriage is monogamous, entails co-residence, and is a long-term commitment between a man and woman who typically have children together. Some of these notions are codified into law (for example, in most states same-sex couples may not marry). These understandings are strongly gendered, and may still include the notion that husbands must be breadwinners. This is consistent with findings from research showing that men’s, but not women’s, earnings positively affect whether cohabiting couples move to marriage (Smock and Manning 1997). A corollary is that men’s nonemployment is reason enough for divorce. As for women’s employment, decades ago, Parsons (1949) argued that women’s employment, especially in a high-level job, was corrosive to marriage by creating competition rather than solidarity between the husband and wife. Today, when a majority of married women are employed, one has to question whether their employment can be controversial enough to cause divorce. Still, in its old-fashioned form, the prediction of the institutional view is that either men or women are more likely to leave if the marriage violates gender norms, which we will operationalize to be cases where men are not employed or women are employed. In this view, the effect of employment (where his deters and hers promotes divorce) is not present only if there is prior unhappiness in the marriage. Rather, the norm violation, and lack of external social support for this kind of a marriage, causes dissatisfaction for both spouses, and this leads to divorce. Or lack of social support could cause divorce without affecting marital satisfaction. Thus, some or all of the effect of employment might be mediated by satisfaction, but we would not predict an interaction between the two factors.

The marriage-as-gendered-institution view yields the same predictions as economist Gary Becker’s (1991) specialization view that either husbands or wives are more likely to leave when he is not employed or when she is employed. In Becker’s view, specialization (where men generally specialize in market and women in household work) provides efficiency gains. Marriages in which each spouse does both market and household work provide fewer efficiency gains and thus provide less incentive to enter or stay in a marriage. Despite the same prediction, the mechanism posited by the two theories differ. Whereas Becker saw less material incentive for either party to keep a marriage intact without the productivity gains from specialization, the institutional view sees marriages as crumbling when they violate gender norms of what husbands’ and wives’ roles should look like, through decreased social support and/or spouses’ themselves becoming dissatisfied because of the norm violation. Of course, the notion that one will leave a marriage because gender norms are being violated seems in contradiction with another institutionalized norm—that marriage is a lifelong commitment. We suggest, however, that many do not see the required commitment as absolute, making it likely that violation of other norms associated with marriage, such as those about gender, may predict divorce.

Exchange and Bargaining

We combine exchange theory in sociology and bargaining theory from economics because they make similar predictions (on sociologists’ application of exchange theory to marriage, see Bittman et al. 2003; Brines 1994; Cherlin 2000; Heer 1963; Molm and Cook 1995; on economists’ applications see Lundberg and Pollak 1994; on economists’ applications see Lundberg and Pollak 1996; on economists’ applications see Lundberg and Pollak 2000; Lundberg, Pollak, and Wales 1997; Manser and Brown 1980; McElroy 1990). These theories predict that a spouse who has more resources, such as earning from employment, is better able to exchange or bargain for what s/he wants when disagreements arise in the marriage, leading to greater marital satisfaction. However, the theory further predicts that, if one is unable to get what one wants in the marriage through bargaining, resources may lead to initiating a breakup. The idea behind the latter prediction—that a spouse is more likely to initiate divorce if she or he has more resources—is that resources that are portable if one leaves the marriage create better alternatives outside the current marriage, such as the ability to support oneself and one’s children. Thus, the exchange/bargaining perspective implies that resources allow either “voice” (bargaining) or “exit” (divorce). (See England and Kilbourne (1990) for an application of Hirschman’s (1970) notion of “exit, voice, and loyalty” to marriage.¹) The theories are named for the ability that resources give an actor to exchange or bargain, rather than by the “exit” that the theories also posit as permitted by resources. However, our interest here is in using the theories to predict the latter outcome, divorce, which the theory says will occur only when exchanging resources for (or “bargaining” for) what one wants does not work.² Since the theory says that “exit” will be used only if bargaining or exchange don’t work to keep the terms of the marriage satisfactory, we believe it is best operationalized to say that a spouse is more likely to leave if s/he is employed, but that the resources from employment only encourage exit if this spouses’s marital satisfaction is low. By contrast, in the institutionalist view, women’s employment or men’s nonemployment may cause divorce even where there was no prior dissatisfaction; the idea is that violating gendered norms will cause spouses’ dissatisfaction and/or a lack of social support that will in turn threaten the marriage. Marital satisfaction, thus, might mediate some of the employment/divorce relationship in the institutionalist view, but it should only interact with employment in the exchange/bargaining view.

The commonly posited “women’s independence” effect is an example of this exchange/bargaining logic (Cherlin 2000; Ruggles 1997; Schoen, Astone, Rothert, Standish, and Kim 2002). The idea is that women who are unhappy in their marriage are freer to divorce if they have earnings from employment. Although less discussed, we should expect a “men’s independence effect” as well; men who are economically dependent upon their wives but unhappy might be less likely to initiate divorce because of the effect leaving would have on their standard of living. The exchange/bargaining view further implies that having a spouse who is employed (especially with high earnings) makes one less likely to initiate a breakup, since the spouse is likely to share the resources as long as the marriage persists.

When each spouse engages in the calculus described above, considering alternatives inside as well as outside the marriage, then A’s resources encourage A to initiate divorce if A is dissatisfied, but A’s resources discourage B from initiating divorce, while B’s resources encourage B to initiate divorce if B is dissatisfied, but B’s resources discourage A from initiating divorce.

The Changing Gender System: Gendering Theoretical Predictions

The changing gender system has changed marriage. Yet the so-called “gender revolution” has been asymmetric, changing women’s activities much more than men’s (England 2010; England and Farkas 1986). The big change has been the dramatic increase in employment among married mothers (Blau 1998; Cohen and Bianchi 1999). Motherhood still reduces married women’s odds of employment, but much less so than previously (Cohen and Bianchi 1999). Most wives are employed, although when they have young children, some take time out of employment and many are employed in jobs with part-time hours (Cohen and Bianchi 1999). Thus, expectations for women are ambiguous—they are expected to be in charge of child rearing, and “intensive mothering” is increasingly expected (Bianchi, Robinson, and Milkie 2006), but norms about employment are ambivalent and inconsistent (Wilcox and Dew 2009). Mothers’ employment rose steadily until about 1990, after which it largely leveled off (Cotter, Hermsen, and Vanneman 2004).

Men’s activities have changed too, but much less than women’s. Husbands face increased expectations that they “help” with housework and child care and provide emotional companionship (Cherlin 2004; England and Shafer 2007). At the same time, men still feel stigma attached to doing “women’s work.” Time use studies show that men’s contribution to household work and child rearing has increased, but less than the increase in women’s hours of employment (Bianchi et al. 2006). Nor have these changes relaxed the expectation of full-time employment for men. In this sense, the male breadwinner model of decades ago seems firmly in place even while an acceptance that women may join men in breadwinning, and men should “help” somewhat at home, has grown up alongside it. The changes in the typical gender regime in marriage are one sense in which marriage has become “de-institutionalized,” to use Cherlin’s (2004) term. But the asymmetric nature of gender change implies that some understandings of marriage are still firmly institutionalized—particularly the notion that husbands should be breadwinners.

Noting the asymmetric nature of gender change does not change our prediction, from bargaining/exchange theory, that women’s employment will increase their odds of exiting a marriage if they are unhappy with the marriage. If anything, the asymmetric nature of gender change strengthens this prediction because the lack of change in the expectation that men will be fully employed, as well as the stigma some men face for participating in housework and day-to-day childcare, make it hard for women to successfully bargain for these changes, or even think of them as an option. Employed wives’ perception that their husbands aren’t doing enough around the house often ramps up marital dissatisfaction (Frisco and Williams 2003; Wilkie, Ferree, and Ratcliff 1998). But if women can’t or won’t bargain to change this because changes in men’s roles are not yet or only ambivalently accepted, these constraints on bargaining make exit, more than voice, the likely result of the dissatisfaction when the woman’s employment makes her able to support herself.

The exchange/bargaining model also predicts that men are less likely to leave if their wives are employed (since they are walking away from resources otherwise shared with them). Here the prediction needs to be gendered by recognizing the still large sex gap in pay. During most of the period covered by our data women working full-time earned between 60% and 75% what men do, and many married wives worked part-time (http://www.bsos.umd.edu/socy/vanneman/endofgr/cpsearn.html accessed 12-15-09; Cohen and Bianchi 1999). The smaller women’s earnings are, the less they give men an economic incentive to stay. It is an empirical question whether the incentive is large enough to create significant effects of women’s employment on men’s exits. Another way gender may undermine predictions from bargaining theory is if the norm that men should have power over their wives dampens any “men’s independence effect,” because it encourages men to prefer more marital power to a share of their wives’ earnings.

The dramatic increases in women’s employment casts doubt on the institutionalist prediction that either spouse is more likely to leave if women are employed. Even if we accept the institutionalist claim that marriages are scrutinized through the lens of social norms, we question the plausibility that women’s employment is sufficiently controversial to encourage divorce. In the period after 1988 when the divorces in the data we analyze took place, 60% or more of married mothers with preschool children were employed at least sometime during the previous year, with rates of 70–80% for wives with older children (Cotter, England, and Hermsen 2009:423). Oppenheimer (1997a; 1997b) goes further and argues that women’s employment helps marriage by reducing risks of specialization, such as the economic deprivation that can attend men’s job loss. Others have made an even stronger theoretical argument that companionship and emotional intimacy are enhanced when men’s and women’s roles are similar rather than different, where women’s employment is seen to make spouses’ roles more similar (Coontz 2005; England and Farkas 1986; Giddens 1992). Whether these advantages of women’s employment neutralize the possible disadvantages based on lingering traditional norms is an empirical question.

Our portrait of the asymmetric nature of gender change, focusing on how little change has occurred in the expectation that married men are to be employed, suggests that the marriage-as-gendered-institution view may still have predictive power; men’s nonemployment may still encourage either spouse to leave because norms have been violated. Having reviewed theory, we turn now to a review of past empirical research.

PAST RESEARCH ON DETERMINANTS OF DIVORCE

Most studies take whether or not the couple divorced as the dependent variable, failing to distinguish between divorces initiated by women and by men. Divorce has generally been found to be more likely when men’s earnings are lower (Hoffman and Duncan 1995; South and Lloyd 1995) or declining (Weiss and Willis 1997). In contrast, findings on the effects of women’s earnings are inconsistent. Some studies find that women’s earnings are positively related to divorce (Cherlin 1979; Heckert, Nowak, and Snyder 1998; Hiedemann, Suhomlinova, and O’Rand 1998; Moore and Waite 1981; Ono 1998; Rogers 2004; Ross and Sawhill 1975; Spitze and South 1985), especially when men’s earnings are lower (Heckert et al. 1998; Ono 1998), but others find no effect of women’s earnings (Greenstein 1995; Hoffman and Duncan 1995; Mott and Moore 1979; Sayer and Bianchi 2000; South and Lloyd 1995; Tzeng and Mare 1995), and a few suggest that women’s earnings, like men’s, stabilize marriage (Greenstein 1990; Hoffman and Duncan 1995; and for changes in earnings, Weiss and Willis 1997). Cooke’s (2006) analysis finds that, in the U.S., relative to dual earner couples in which the husband earns as much or more than the wife, the probability of divorce is elevated by either a traditional arrangement where the man but not the woman is employed and he does little housework, or a nontraditional arrangement in which the wife earns more than the husband.

Women’s employment has been theorized to produce unhappy marriages, and data gathered before 1970 offered some support for this assumption, but more recent studies do not find that women’s employment, per se, reduces marital happiness (Rogers 1999; Spitze 1988). Indeed, a recent study by Schoen, Rogers, and Amato (2006) that uses longitudinal data finds that couples have a lower risk of divorce when the wife remains in or transitions into full-time employment between survey waves. Moreover, they found that wives’ full-time employment had no association with changes in either spouse’s assessments of marital satisfaction. Men’s nonemployment, by contrast, is associated with lower satisfaction for both partners (Conger, Rueter, and Elder 1999).

Schoen et al. (2002) reasoned that we would expect women’s earnings to encourage divorce only for women significantly dissatisfied with their marriages. To test this, they interacted women’s full-time employment status with their reported marital happiness. The interaction was significant; women’s employment does, indeed, encourage divorce, but only in marriages where the wife reported substantial unhappiness. This is consistent with the “women’s independence effect” predicted from bargaining/exchange theory, but the theory suggests that the effect would come from effects on women’s, not men’s, exits and Schoen et al. (2002) did not distinguish between female- and male-initiated divorces. Not surprisingly, many studies show that divorce is more common when individuals previously reported marital dissatisfaction (Heaton and Albrecht 1991; Nock 1995; Sanchez and Gager 2000; Sayer and Bianchi 2000; Schoen, Rogers, and Amato 2006).³

In addition to effects of economic variables and marital satisfaction on divorce, research has shown associations of other socio-demographic factors with divorce. Net of economic factors, being African American, not having young children, cohabiting prior to marriage, early age at marriage, a premarital conception, having parents who divorced, or having previously been divorced are associated with higher likelihoods of divorce (Amato and Rogers 1997; Bumpass and Lu 2000; Bumpass, Martin, and Sweet 1991; Martin and Bumpass 1989). The effect of children varies by age, with younger children deterring and older children increasing divorce (Waite and Lillard 1991), while stepchildren are not as protective as a couple’s own children (Martin and Bumpass 1989). Some of these characteristics may increase risks of dissolution because they lower marital satisfaction via influences on intracouple dynamics. Specifically, entering marriage at younger ages, low levels of education and income, and experiences of parental divorce are associated with less positive patterns of communication in marriage (Bradbury, Fincham, and Beach 2000).

The longer marriages have lasted so far, the less likely they are to end (Martin and Bumpass 1989; Sayer and Bianchi 2000). White and Booth (1991) suggest that this may be because there are more shared social networks and assets over time, the benefits of which are much greater if the marriage persists.

There is no previous published research examining the effect of spouses’ employment, earnings, or marital satisfaction on the separate hazards of men’s and women’s initiation of divorce in the United States. There is one unpublished U.S. study, Godecker’s (2002) dissertation modeled separate determinants of wives and husbands leaving, focusing on effects of men’s and women’s perceptions of marital quality (satisfaction), with little attention to economic factors. The two published papers on determinants of wife- and husband-initiated divorces use Australian and Dutch data, respectively (Hewitt et al. 2006; Kalmijn and Poortman 2006). The data used for these two analyses, however, had interviews of only one spouse, and thus had only that spouse’s report of who left, and also lacked many characteristics of the other spouse that are available in the data we use. For example, Hewitt et al. (2006) modeled wife and husband initiated marital separations using Australian data, but did not include measures of employment or earnings. In a model not containing earnings, they found that women with higher levels of education had elevated hazards of initiating divorce (relative to staying married), but because education increases employment and earnings, one could interpret this as an effect of resources. In an analysis of Dutch married individuals, Kalmijn and Poortman (2006) included measures of wife’s employment hours but did not have the data to include either spouse’s earnings. They find that a wife’s higher employment hours increase her risks of initiating divorce, but are not associated with the husband’s initiation risks, consistent with the women’s independence effect. Their models do not include measures of marital quality, leaving unanswered the question of whether wife’s employment increases risks of divorce in all marriages or only unhappy ones.

Our analyses are directed at testing the three theoretical views explained above. Table 1 summarizes the predictions of each of the theories regarding effects of male and female employment, and provides a preview of our empirical findings.

Table 1.

Predictions of Theories Regarding Sign of Effects of His and Her Employment on Husbands and Wives Initiating Divorces

Theories:	Effects of Wife’s Employment on		Effects Of Husband’s Employment on
	Her Exit	His Exit	Her Exit	His Exit
Marriage as Gendered Institution1	+	+	−	−
Exchange/Bargaining, (Wo)men’s Independence	+2	−	−	+2
Findings (Table 4)	+2	0	−	−

¹The effect may be partly mediated by marital satisfaction, so the prediction pertains to a model not controlling for marital satisfaction. The predictions do not involve an interaction between employment and marital satisfaction. Becker’s specialization theory yields the same predictions via a different mechanism.

²The effect is predicted or found only when the marital satisfaction of the spouse whose exit is predicted is low; employment interacts with marital satisfaction.

Data, Methods, and Hypotheses

Data

We use Waves 1, 2 and 3 of the National Survey of Families and Households (NSFH) (Sweet and Bumpass 2002). The NSFH is a national probability sample survey of 13,007 adults age 19 and older interviewed in 1987–88. In married couple households, one adult was randomly selected as the primary respondents, and the spouse also completed a self-administered questionnaire. Wave 2 was conducted in 1992–1994, again obtaining data from both spouses. In cases of divorce between waves, both ex-spouses were interviewed individually, and asked about the divorce. Analogously, Wave 3 was collected in 2001–2002.⁴ The NSFH is the only longitudinal U.S. data set on a national probability sample that tracks marital dissolution and includes a measure of which spouse wanted the relationship to end. Our analysis sample (N=3,622) includes couples married at NSFH1 in which the spouse completed the questionnaire at NSFH1; at least one spouse was reinterviewed at NSFH2 or NSFH3; and neither spouse was age 55 or older at NSFH1. We exclude couples in which one or both are older than age 56 because of low risk of divorce among this group. We transformed the data into a form where couple/months are the units of analysis, appropriate to the event history models described below. Each couple has a record for every month from the NSFH1 interview survey month to the month separation occurred, if the couple divorces, or to the NSFH3 survey month if the couple remains together.

Measures

Our dependent variable is whether a breakup occurred and, if so, who initiated it. For couples separating or divorcing after NSFH1, the month at which the breakup occurred is ascertained at the next interview. Consistent with most past research, we consider marriages dissolved at the point of separation, even if there is no legal divorce. In the wave after the divorce each ex-spouse was asked which person wanted the breakup more. In NSFH2, husbands and wives who had experienced a marital separation or divorce between NSFH1 and NSFH2 completed a self-administered module on the experience of relationship dissolution, including a question that ascertained which spouse most wanted the divorce. The question was worded as follows: “Sometimes both partners equally want a marriage to end, other times one partner wants it to end much more than the other. Circle the number of the answer that best describes how it was in your case.” Response categories included 1) I wanted the marriage to end but my husband/wife did not; 2) I wanted it to end more than my husband/wife did; 3) We both wanted it to end; 4) My husband/wife wanted the relationship to end more than I did; or 5) My husband/wife wanted the marriage to end but I did not. NSFH3 asked about divorce initiation for those who separated or divorced between NSFH2 and NSFH3. Whereas in NSFH2 the question was part of a self-administered paper module, in NSFH3 the module was administered verbally by the interviewer and the question wording was changed slightly to collapse and simplify response categories. The NSFH3 question was: “Some partners disagree about how much they want their marriage to end. In your case, who most wanted your marriage to end? Would you say that you wanted it most, you both wanted it equally, or that your (husband/wife) wanted it to end most?” Respondents were then queried as to whether or not either partner had not wanted the marriage to end at all, and if so, which spouse this was. To make the analysis tractable, and because the assumption seems reasonable, we make the simplifying assumption that one person wanted the breakup more, and use both ex-partners’ reports, together with the latent-class analysis described below, to analyze who initiated the breakup and the determinants of such initiation.

Table 2 shows how spouses who had broken up characterized initiation of the breakup, and the extent to which ex-spouses agree. To simplify, for this table (but not our eventual analysis), we collapse responses from five to three categories (so, for example, a woman saying that she wanted it when her partner didn’t and a woman saying that she wanted it more than her partner are combined into “she wanted more”). Table 2 describes the 747 divorces in the data (out of 3,622 eligible married couples at NSFH1). By women’s report, she wanted the divorce more in 59% of cases, both wanted it equally in 16%, and he wanted it more in 25% of cases. The analogous percentages for men are 46% saying she wanted it more, 25% saying they wanted it equally, and 30% saying he wanted it more. So, in the aggregate, both men and women attribute more initiation to women than men.

Table 2.

Percent of Divorced Husbands Reporting that She, He, or Both Wanted Divorce More Within Each Category of Divorced Wives Reports to Who Wanted Divorce More

Husband’s Report	Wife’s Report
	She Wanted		He Wanted		Both Wanted		Row Total		No Answer/No Interview
	%	N	%	N	%	N	%	N	N
She Wanted	67%	143	9%	8	25%	14	46%	165	57
He Wanted	9%	18	79%	73	29%	16	30%	107	32
Both Wanted	24%	52	12%	11	46%	26	25%	89	37
Column Total	100%	213	100%	92	100%	56	100%	361
		59%		25%		16%		100%
No Answer/No Interview		126		50		42			42
									168

Note: N = 747 Divorces

Table 2 also shows that individual couples agree more often than they disagree on which partner wanted the divorce. When the wife reports that she initiated, 67% of husbands agree, an additional 24% report that both wanted to break up, and in only 9% of cases is there the extreme disagreement where she says she initiated whereas he says he did. Similarly, when the wife reports that he initiated, 79% of husbands agree that they were the initiators, 12% say both wanted the divorce equally, and only 9% had the extreme disagreement where she said he initiated while he said she did. The considerable agreement between ex-spouses who answered the question on who wanted the divorce suggests that the item is a meaningful indicator of who initiated the divorce. At the same time, there is a problem of substantial missing data for the divorce initiation question: 35% of ex-husbands and 22% of ex-wives either were not interviewed at the survey wave after their breakup, or did not respond to the question about which spouse wanted the divorce. Item nonresponse declined at NSFH3, perhaps because respondents were more reluctant to refuse to answer the question when asked by an interviewer. Attrition of one spouse is a concern at NSFH2 and NSFH3: 26% of divorced men whose ex-wives were interviewed were lost to follow-up and 14% of divorced women whose ex-husbands were interviewed were lost to follow-up. However, similar levels of agreement about which spouse wanted divorce are found for Dutch data with minimal nonresponse (Kalmijn and Poortman 2006), bolstering our confidence in the validity of NSFH results.⁵

A major focus of our analysis is how husbands’ and wives’ employment affect either spouse’s propensity to leave. We use employment as an indicator of economic resources rather than annual earnings. The primary reason for this is that, at NSFH2 and NSFH3, respondents were asked about their employment history between waves, so we know for every month between waves whether each spouse was employed, making this variable a changing covariate measured as frequently as union outcome. By contrast, annual earnings of each spouse were measured only for the year preceding each wave, and, since waves are as much as a decade apart, the covariate available for analysis is sometimes sorely out of date.⁶

We created a measure for whether each spouse was employed in each person month of the marriage, based on the employment history between NSFH1, NSFH2 and NSFH3. The measure is the number of months each spouse was employed out of the 24 month period that ended 12 months before the month in question. Lagging the measure 12 months is undertaken to minimize endogeneity of employment to divorce or a clear decision to divorce. Johnson and Skinner’s (1986) analysis of the Panel Study of Income Dynamics showed that women often moved into the labor market in the year prior to a divorce, and their interpretation was that women who have decided to divorce may seek jobs first so they are more ready to support themselves upon exit. We do not want to incorrectly conclude that employment caused divorce when the intent to divorce soon is what caused the employment, so lagging employment a year seems prudent.

To measure the wife’s and husband’s marital dissatisfaction, we constructed a “Better-off-divorced” (BOD) scale from five questions that ask respondents to think about how, hypothetically, getting a divorce would affect their standard of living, social life, job opportunities, overall happiness, and sex life. Responses range from 1 = much worse, to 5 = much better (alphas for wives’ BOD scales are .78 at NSFH1 and .75 at NSFH2; alphas for husbands are .80 at NSFH1 and .78 at NSFH2). These questions tap perceptions of the costs and benefits of remaining in the relationship, compared to the alternative life respondents think they could have outside of this marriage. Theory and empirical research indicate marital quality is a multidimensional concept that encompasses subjective evaluations of numerous relationship domains (Amato et al. 2007). The BOD measure closely corresponds to key relationship domains found to be associated with marital quality and dissolution and hence offers an appropriate composite assessment of overall marital quality. For ease of interpretation, we rescaled this measure to have a mean of zero and a standard deviation of 1; standardization was done separately for each wave. Wife’s and husband’s scores on the Better off Divorced scale were measured at NSFH1 and NSFH2 (for couples who did not divorce prior to NSFH2); each person month contains the most recent measure available.

We use the BOD scale measuring marital dissatisfaction two ways. First, we assess how much each spouse’s marital satisfaction affects their own hazard of initiating divorce. Second, we interact the employment measures for each spouse with his or her own BOD score to see if, as predicted by bargaining/exchange theory, any “independence effect” of employment obtains only when a spouse is unhappy with the relationship.

Our models also include other socio-demographic variables that past research and our hypotheses suggest should be relevant to one or the other spouse initiating divorce. Several of these control variables are time-varying to the month, representing characteristics of the couple or one of the spouses for which we know changes to the month. We include a continuous measure of the length of the marriage in months. We include time varying measures of the number and ages of children at NSFH1 and NSFH2 grouped into 4 measures: the number of children ages 0 to 1, the number ages 2 to 5, the number ages 6 to 12, and the number ages 13 to 18. Since we know birthdays of children to the month, these measures are accurate to the month.

Some other control variables, like the BOD scale discussed above, are time-varying, but we only have data that measures them at Wave 1 and at Wave 2. (Since all reported marital dissolutions occurred before Wave 3, and since we lag all covariates behind the month at risk, Wave 3 scores on independent variables were not used.) These include two dummy variables: 1) whether children are present who are not the biological children of the husband, and 2) whether children are present who are not the biological children of the wife. Measures of wife’s and husband’s education are coded into four categories: less than high school education (the reference group in the regressions), high school graduate, some college, and college graduate. These measures are also updated at each wave.

Other control variables are unchanging characteristics of the couple or one of the spouses. Wife’s age at marriage is measured with a continuous variable. (We did not also enter husband’s age at marriage because of strong collinearity.) We also include a dummy variable indicating whether the wife is 3 or more years older than the husband. Measures of whether each spouse grew up in a disrupted family are dummy variables coded 1 if the wife (husband) did not live with both their biological or adoptive parents until age 19. The measure of previous cohabitation is a dummy variable coded 1 if husband (wife) had ever cohabited, regardless of whether it was with the person to whom they are married. The measure of whether the couple’s first child was conceived before marriage is a dummy variable coded 1 if a birth occurred in the window of 6 months prior to 7 months after the date of the marriage. Wife’s and husband’s measures of a previous divorce are dummy variables coded 1 if the wife (husband) had ever been previously divorced. The measure of race is a dummy variable coded 1 when either the husband or the wife is Black. (The vast majority of these marriages contain two African American partners. Due to the small number of inter-racial marriages, we did not have sufficient power to distinguish effects of being in a marriage with both, just husband, and just wife who are Black.)

Means and standard deviations of independent variables are presented in Table 3.

Table 3.

Means and Standard Deviations of Independent Variables

	Mean	Standard Deviation
Variables that are Time-Constant or Updated only at Wave 2
Wife 3 or more Years Older than Husband (0,1)	0.073	0.260
Wife’s Age at Marriage in Years	23.486	5.779
Children present that are not husband’s (0,1)	0.073	0.260
Children present that are not wife’s (0,1)	0.186	0.389
Premarital conception of first birth to couple (0,1)	0.142	0.349
Wife’s Education (< high school = reference)
High School Graduate	0.391	0.488
Some College	0.231	0.421
College Degree	0.250	0.433
Husband’s Education (< high school = reference)
High School Graduate	0.347	0.476
Some College	0.226	0.419
College Degree	0.250	0.433
Either wife or husband Black (0,1)	0.091	0.288
Husband from disrupted household (0,1)	0.203	0.402
Wife from disrupted household (0,1)	0.225	0.418
Wife previously divorced (0,1)	0.196	0.397
Husband previously divorced (0,1)	0.204	0.403
Wife ever cohabited (0,1)	0.300	0.458
Husband ever cohabited (0,1)	0.320	0.467
Wife’s Better Off Divorced Scale	−0.050	1.047
Husband’s Better Off Divorced Scale	−0.010	1.047
Variables that are Time-Varying to the Month
Marriage Duration in century months	213.060	122.157
Number and Age of Children
Number of children 0 to 1	0.133	0.362
Number of children 2 to 5	0.286	0.557
Number of children 6 to 12	0.525	0.797
Number of children 13 to 18	0.451	0.747
# months wife employed in previous 2 years, lagged 1 year	16.341	10.343
# months husband employed in previous 2 years, lagged 1 year	19.831	8.530

Note: Units of analysis are couple/years.

Models and Estimation

Our goal is to estimate a model that combines a latent class model (Clogg and Shihadeh 1994; Clogg and Goodman 1984; Goodman 1974) with a discrete-time event history model (Allison 1982; 1984). In this model, husbands’ and wives’ reports of divorce initiation are regarded as fallible indicators of the true initiation status. In turn, the true initiation status is modeled in a competing risks framework as a multinomial logistic regression.

We begin with the latent class portion of the model. Data are arrayed with couple-months as the unit of analysis. In each couple month, we assume that couples can be classified into one of three states: wife initiated a divorce/separation, husband initiated a divorce/separation, or neither initiated. Let C be a variable with values of 1, 2, and 0, corresponding to these three states. We do not directly observe C. Instead, we observe the husband’s report H and the wife’s report W. Both of these variables are coded as 1=husband wanted the divorce more, 2=wife wanted the divorce more, 3=both wanted equally, and 0=neither partner initiated a divorce. For simplicity, the latent class variable does not allow for the possibility that both partners wanted the divorce equally, even though respondents may have chosen this option. Thus, we make the not unreasonable assumption that one spouse always wanted the divorce more, however slightly.

The relationship between the latent class variable and each spouse’s report can be described by a set of conditional probabilities (the probability of falling into one of the observed classes given membership in one of the latent classes). The model specifies the joint probability of the two observed variables and one latent variable as:

$$Pr(C=j,W=k,H=l)=Pr(C=j)Pr(W=k\mid C=j)Pr(H=l\mid C=j)$$

This equation expresses the usual “local independence” assumption that H and W are statistically independent, conditional on the value of C.

The conditional probabilities are subject to the following constraints, which merely express the fact that if no separation occurred in a given month, there is no discrepancy between the husband’s report, the wife’s report, and the value of the latent variable:

$$\begin{array}{ll}Pr(W=0\mid C=0)=Pr(H=0\mid C=0)=1\hfill & \hfill \\ Pr(W=j\mid C=0)=Pr(H=j\mid C=0)=0,\hfill & j\ne 0\hfill \\ Pr(W=0\mid C=j)=Pr(H=0\mid C=j)=0,\hfill & j\ne 0\hfill \end{array}$$

The next step is to let the probability distribution of C depend on predictor variables in a multinomial logistic regression model, as described by Yamaguchi (2000). Let p_ijt be the conditional probability that couple i at time t is in state C=j, given that the couple has not already separated. Dependence on predictor variables is specified as

$$log\left(\frac{p_{\mathit{ijt}}}{p_{i0t}}\right)={\alpha }_j(t)+{\beta }_j{x}_{it},j=1,2,$$

where x_it is a column vector of possibly time-varying covariates, β_j is a row vector of coefficients, and α_j(t) is some function of t. This multinomial logit model can be viewed as two simultaneous, binary logit models, one predicting initiation by the husband versus no separation, the other predicting initiation by the wife versus no separation.

The risk of separation/divorce begins with the first month after the Wave 1 interview, although the dependence on duration α_j(t) will be specified as the number of months since the couple was married. Couples that experience the death of a spouse, or whose marriage remains intact from Wave 1 to Wave 3 are treated as right-censored observations.

The combination of the latent class and multinomial logit models can be viewed as a discrete analog of the “multiple indicator-multiple cause” model for continuous variables (Hauser and Goldberger 1971). We simultaneously estimate the two components of the model by maximum likelihood using Mplus software (Muthén and Muthén 2007). One of the attractions of this implementation of maximum likelihood is that it accommodates missing data on the dependent variable, under the assumption that the data are missing at random. This capability is especially important because, as noted earlier, a substantial fraction of couples are missing the report of either the husband or the wife. Multiple imputation was used to handle missing data on the predictor variables (Allison 2001). Missing data were imputed under a multivariate normal model and the missing-at-random assumption using PROC MI in SAS. All analyses are unweighted.⁷

Results from our latent class/event history models are presented in Table 4. We present results from three models. All models contain the covariates discussed above as controls that have been found to affect divorce in other research. The first model also contains each spouse’s employment history—how many of the previous 24 months they were employed, lagged one year behind the month at risk. Thus, it answers the question posed by institutionalist theory—whether “gender nonconformity,” as assessed by men’s nonemployment or women’s employment, makes it more likely that either partner leaves. The second model adds the scale assessing spouses’ marital satisfaction relative to alternatives—their perception of how much they would be better off divorced (BOD). This allows us to see if each partner’s own satisfaction affects their own exit most strongly, and whether any employment effects are mediated by one spouse’s employment affecting that or the other partner’s satisfaction, a possible implication of either theory. The third model interacts each spouse’s BOD with his or her own employment history. This allows us to test the prediction from exchange/bargaining theory that one partner’s employment elevates his or her odds of exiting the marriage, but only under conditions of unhappiness with the marriage, which may be an indication that bargaining was unsuccessful. For ease of interpretation, columns have been reordered so that successive columns from left to right show, first, odds ratios from Models 1, 2, and 3 predicting wife-initiated divorces (relative to spouses staying together), then odds ratios from the three models predicting husband-initiated divorces (relative to spouses staying together). The far right three columns of the table present, for each model, probabilities from a Wald test for the statistical significance of the difference between the effects of each variable’s effect on husband-initiated and wife-initiated divorces.

Table 4.

Odds Ratios from Latent 3-Class Model Predicting Whether Married Couple Stayed Together, Divorced with Wife Wanting Divorce More, or Divorced with Husband Wanting Divorce More

	Divorce that Wife Wanted More Relative to Stayed Together1			Divorce that Husband Wanted More Relative to Stayed Together1			Wald Test that Coefficient is Significantly Different for Divorce that Wives v. Husbands Wanted
	M1	M2	M3	M1	M2	M3	M1	M2	M3
# Months Wife Employed previous 2 years, lagged 1 year	1.010	1.009	1.005	1.010	1.007	1.007	−0.017	−0.210	0.268
# Months Husband Employed previous 2 years, lagged 1 year	0.976*	0.979*	0.978*	0.964*	0.966*	0.962*	−1.219	−1.316	−1.651
Wife’s Better Off Divorced Scale		1.626*	1.349*		1.212*	0.946		−2.988*	−1.677
Husband’s Better Off Divorced Scale		1.155*	0.999		1.492*	1.093		2.560*	0.459
Wife’s Better Off Divorced *
Employed (interaction effect)			1.012*			1.015			0.372
Husband’s Better Off Divorced *
Employed (interaction effect)			1.008			1.018*			1.006
Wife’s Education (< high school = reference)
High School Graduate	0.786	0.839	0.839	1.039	0.934	0.939	0.775	0.296	0.313
Some College	0.817	0.899	0.904	0.918	0.862	0.901	0.306	−0.115	−0.008
College Degree	0.735	0.798	0.807	1.230	1.130	1.170	1.250	0.861	0.917
Husband’s Education (< high school = reference)
High School Graduate	1.175	1.239	1.212	1.112	1.123	1.117	−0.159	−0.282	−0.234
Some College	1.065	1.117	1.088	1.035	1.027	1.030	−0.079	−0.233	−0.152
College Degree	0.914	1.019	0.987	0.921	0.976	0.992	0.018	−0.106	0.011
Wife 3 or more Years Older than Husband (0,1)	0.950	0.875	0.882	2.106*	1.939*	2.012*	1.876	1.835	1.898
Wife’s Age at Marriage in Years	0.921*	0.921*	0.919*	0.984	0.982	0.976	2.432*	2.378*	2.251*
Marriage Duration in century months	0.993*	0.993*	0.993*	0.996*	0.996*	0.996*	2.365*	2.257*	2.099*
Children present that are not husband’s (0,1)	0.705	0.717	0.697	1.560	1.738*	1.685*	2.097*	2.330*	2.343*
Children present that are not wife’s (0,1)	0.934	0.976	0.948	0.853	0.802	0.787	−0.263	−0.573	−0.534
Premarital conception of first birth to couple (0,1)	1.105	1.003	1.000	0.875	0.850	0.868	−0.746	−0.518	−0.441
Either wife or husband Black (0,1)	1.553*	1.078	1.064	0.782	0.557	0.534*	−2.049*	−1.848	−1.947
Husband from disrupted household (0,1)	1.123	1.158	1.131	1.273	1.239	1.215	0.526	0.286	0.311
Wife from disrupted household (0,1)	1.305*	1.250	1.235	1.087	1.085	1.101	−0.802	−0.608	−0.501
Wife previously divorced (0,1)	1.696*	1.797*	1.844*	0.893	0.908	0.919	−1.985*	−2.217*	−2.224*
Husband previously divorced (0,1)	1.537*	1.501*	1.499*	1.420	1.391	1.456	−0.300	−0.290	−0.111
Wife ever cohabited (0,1)	1.350*	1.310	1.328	1.250	1.223	1.212	−0.245	−0.225	−0.298
Husband ever cohabited (0,1)	1.115	1.089	1.091	0.969	0.975	0.976	−0.437	−0.359	−0.359
Number and Age of Children
Number of children 0 to 1	0.843	0.856	0.854	0.847	0.822	0.815	0.020	−0.163	−0.182
Number of children 2 to 5	1.104	1.064	1.088	1.106	1.068	1.082	0.007	0.024	−0.034
Number of children 6 to 12	1.040	1.062	1.073	1.052	1.095	1.107	0.091	0.240	0.256
Number of children 13 to 18	0.950	0.934	0.935	1.008	0.998	1.009	0.423	0.490	0.559

¹Odds Ratios predicting divorce that wife or husband wanted more, relative to reference category of couple staying together.

^*p<.05

RESULTS

The effects of each spouse’s recent employment history are shown in Model 1, which includes all covariates except the Better-off-divorced scale, so as to estimate the effect of employment including any indirect effect mediated through satisfaction. (Recall that the institutionalist theory predicts an additive effect of employment that may be partially mediated through women’s employment or men’s nonemployment causing dissatisfaction.) The basic pattern is that wives’ employment has no significant additive effect on whether either women or men leave, but husbands’ employment reduces the risk of either spouse leaving.⁸ Each additional month of his employment reduces the odds that she will leave by 2.4% (risk ratio=.976=1−.024) and reduces the odds he will leave by 3.6% (risk ratio=.964=1−.036).⁹ The Wald test shows that the effects of men’s employment on his versus her leaving are not significantly different; his nonemployment should thus be considered to elevate her and his leaving to the same extent.¹⁰

Comparing our findings to predictions of the theories in Table 1, we see that the institutionalist notion that marriages survive better when they follow social norms (as well as Becker’s specialization model) predicts that men’s employment discourages divorce initiated by either spouse, and we find this effect. But this model (as well as Becker’s specialization model) also predicts that women’s employment encourages either spouse to leave, but we don’t find any additive effect on either spouses’s initiation. As we discussed, while an old-fashioned version of the marriage-as-gendered-institution model predicts that women’s employment encourages divorce, we suspect that the dramatic increase in women’s employment has rendered norms at least neutral on women’s employment. Consistent with this, we find no effect of women’s employment on men’s exits.¹¹ (Effects on women’s exits are discussed later, since they involve interactions.)

Model 2 in Table 4 adds the better-off-divorced scale. Recall that the scale measures the extent to which a spouse agreed, at the most recent wave, that he or she would be better off divorced on a number of dimensions including standard of living, social life, job opportunities, overall happiness, and sex life. We take it to be a measure of satisfaction with the marriage. Since the variable is normalized,¹² an increase in perceiving she would be better off divorced of one standard deviation increases the odds that a wife will leave by 63% (odds ratio=1.626). An analogous increase in husband’s BOD increases the odds that he will leave by 49%. Results are as predicted, and consistent with the exchange/bargaining notion that those who assess themselves better off divorced are more likely to be the initiators. However, there are also significant cross-spouse effects. A standard deviation increase in his sense that he is better off divorced increases her odds of leaving by 16%—much smaller than the 63% elevation of risk from her own BOD, but still nontrivial. The analogous cross-spouse effect for husbands from her BOD is 21%, again nontrivial but smaller than the 49% from his own BOD.¹³ The cross-spouse effects suggest either that spouses will sometimes leave over the negative assessment of the marriage of the other spouse, or that either spouse’s BOD can be seen as an imperfect indicator of dimensions of the quality of the marriage that affect the odds that either will leave.

Model 3 in Table 4 adds interactions between each spouse’s employment and their own assessment of whether they are better off divorced. The idea behind these interactions is that the independence effects posited by exchange/bargaining theories are expected to operate only when one is unhappy; a person won’t leave a happy marriage just because it is economically feasible. Thus, we predict a positive interaction effect, such that the effect of one’s own employment on leaving is more positive as one takes a dimmer view of the marriage relative to outside options, which is measured by the BOD. (A positive interaction also implies that the effect of believing one is better off divorced on actually leaving is greater if one is employed.) As predicted by exchange/bargaining theory, whether amended by the view of asymmetric gender change or not, the interaction between wife’s BOD and wife’s employment is positive. Figure 1, computed from Model 3 in Table 4, with the interaction, shows the direction and magnitude of the effect of months of women’s employment on her leaving for various levels of believing she is better off divorced. The solid line in Figure 1 (for women) shows that in most of the range where BOD is below the mean (indicating a more positive appraisal of the marriage relative to divorce than the average woman has), women’s employment has no positive effect on leaving (indeed it is negative). The nonsignificant coefficient on women’s employment in the model where her employment is interacted with her BOD tells us that her employment has no effect on her leaving when her appraisal of the marriage relative to alternatives is average (0). Thus, women at average or higher levels of appraisal of their marriage are not more likely to leave if they have been employed more; for women in this range the findings contradict expectations for the institutionalist view, at least in its old-fashioned view that sees women’s employment as violating gendered notions of appropriate behavior for wives. But as BOD gets higher (i.e. the appraisal of the marriage relative to divorce gets worse), the effect of women’s employment on her odds of exiting moves towards and into the positive range. As Figure 1 shows, the effect of women’s employment on their own exits becomes positive when BOD is a bit below its mean. In the entire range above the mean, with worse than average appraisals of marriages, women’s employment encourages them to exit, and more so the dimmer her appraisal of the marriage. Calculations from the coefficients underlying Model 3 show that if she was employed 24 (as opposed to 0) months, the odds of her exit would be increased by 50% when her perception that she is better off divorced is a standard deviation above the mean. If she is even more negative on the marriage (a higher BOD score), her employment encourages her to leave even more.¹⁴ This is consistent with bargaining/exchange theory’s “women’s independence” effect.¹⁵

Figure 1.

Odds Ratio for Effect of Months of Employment on Leaving One’s Partner by Score on the Better Off Divorced (BOD) Scale.

Note: The Y axis shows the effect of employment; the X axis shows the BOD scale. The BOD scale is standardized with a mean of 0 and standard deviation of 1. A higher score on BOD indicates more dissatisfaction. The solid line (marked Female) shows the effect of women’s employment on women’s leaving (expressed as an odds ratio) at various levels of women’s BOD, calculated from results of Model 3 in Table 4. The dashed line (marked Male) shows the effect of men’s employment on men’s leaving at various levels of men’s BOD. When the Y axis is above 1, this means a positive effect of employment.

To test whether there is a men’s independence analogous to that found for women, we examine the interaction of men’s marital satisfaction (BOD) and men’s employment affecting men’s initiation of divorce. The sign of the interaction is the same as for women, and the magnitude even larger. But Figure 1, which uses regression results to plot how the effect of each spouse’s employment varies by levels of their own satisfaction, shows that men’s employment discourages their exit from marriage if they are anywhere up to 2 standard deviations above the mean on dissatisfaction. The sign of the interaction tells us that the worse his appraisal of the marriage, the less his employment deters him from leaving the marriage, and implies that there is some level of feeling better off divorced at which his employment encourages him to leave, but calculations show that this only holds in the extraordinarily rare situation where his BOD is at least 3 standard deviations from the mean (results not shown). By contrast, women’s employment elevates their odds of leaving if they are at all above the mean level of dissatisfaction. The fact that men’s employment does not encourage them to leave, even when they are quite unhappy, is inconsistent with exchange/bargaining predictions of a “men’s independence effect.” In sum, contrary to prediction from exchange/bargaining theory, men’s employment does not increase their own likelihood of leaving unless they are extraordinarily unhappy in the marriage. The overall negative effect of his employment on his exits in most of the range is consistent with the institutionalist claim that norm-violating marriages are more likely to be ended by either party.

The fact that even extremely unhappy men do not have their odds of leaving elevated by their employment should not be taken to mean that men’s unhappiness cannot induce them to leave. The sign of the interaction effect implies that the effect of dissatisfaction in encouraging exits is stronger when men are employed, but supplementary calculations show that it is present even for men not employed at all the previous two years.

Turning to effects of control variables, consistent with past literature, we find women who were older when they married have reduced odds of leaving, but men’s wives’ age at marriage has no effect on whether the men leave (Table 4). An interesting finding is that wives that are three or more years older than their husbands have approximately doubled odds of being left (Table 4, Model 1 or 2). In contrast, relative age has no effect on women’s exits (Table 4, Model 1 or 2). This suggests a double standard of aging in which men devalue women as they age, but women do not judge men by age. This may be because women are judged more on looks than men, and men employ a standard of beauty in which youth is seen as more attractive (England and McClintock 2009).

Past research often shows the higher risk of divorce among Black couples. In this analysis we show that this effect comes entirely from women being more likely to leave in couples where either partner is Black, and not from more male exits in such couples. In Table 4’s Model 1, which does not enter the subjective Better-off-divorced scale, couples in which one or both (the more usual case) are Black are 55% more likely to have the wife exit, but not significantly more likely to have the husband exit, and the Wald test shows that these effects are significantly different. After controlling for each spouse’s BOD, the elevated risk that the woman will leave in a Black couple drops to nonsignificance, showing that much of the differential comes, in a proximate sense, from a lower subjective evaluation of the marriage among Black couples. Ascertaining the situational or behavioral sources of the lower quality perceived by women in Black couples is beyond the scope of our analysis, but we note that these effects are net of income, employment, and our other socio-demographic controls. Our results are consistent with one recent study indicating that higher risks of dissolution experienced by Black couples are explained by greater heterogamy among black couples on a number of sociodemographic and attitudinal characteristics (Clarkwest 2007).

Table 4 also shows that wives who did not grow up with both parents are more likely to leave a spouse, but are no more likely to be left.¹⁶ This may be because having parents who separated elevates the chance that women think it is acceptable to divorce. Yet this doesn’t hold for men; whether the husband grew up with both parents does not have a significant effect on his (or her) exits.

If the wife was previously divorced, her risk of leaving is elevated by about 70%, and none of this is mediated by her subjective appraisal of the marriage (the effect is larger after this control). Her previous divorce has no effect on his chance of exiting. If he is previously divorced, it elevates the likelihood that either spouse leaves by about half, and again, virtually none of this is mediated by subjective perception (BOD). The fact that each spouse’s own previous divorce history “affects” his or her own propensity to leave, with little of the effect mediated by subjective perceptions of quality of the marriage (relative to exiting), suggests that people who think it is more acceptable to leave are more likely to have divorced in the past, and this same selectivity is affecting the present divorce. Men’s divorce history affecting women’s divorce is less easy to interpret.

As found in previous studies, we find that the longer the duration of the marriage, the less the odds of breakup. What other studies haven’t explored is whether the diminution of risk applies to his exits, her exits, or both. We find that it applies to each, although the effect on women’s exits is slightly but significantly stronger. (This is seen by the slightly smaller odds ratios for women and the significant Wald test in Table 4.)

DISCUSSION AND CONCLUSION

In this section, we summarize our findings, and discuss the leverage they provide in adjudicating between theories we have considered. We consider whether theories need to be “gendered” to take into account the asymmetric nature of recent gender change. We acknowledge limitations of our analysis. In some cases our empirical analysis could not adjudicate between theories; our contribution has been to provide the first analysis on U.S. data considering how each spouse’s employment and marital satisfaction affects the separate hazard of a wife-initiated and a husband-initiated divorce.

One of the biggest changes in marriage over recent decades has been the rise of women’s employment. The theories considered all imply that women’s employment makes it more likely that women will leave marriages (Table 1), though the mechanisms differ. In the view of marriage as a gendered institution, the idea is that there is less social support for staying in the marriage when it transgresses the cultural norm that women will be homemakers. (The same prediction flows from Becker’s perspective, but not because women’s employment violates norms, but because it implies fewer specialization gains in production.) The institutionalist view does not see the elevation of odds of leaving from employment to be limited to cases of marital dissatisfaction; rather women’s employment undermines the rationale for marriage even if emotional quality is high. In contrast, the “women’s independence effect” posited by the bargaining/exchange view posits that resources from employment make an exit from marriage possible with less fall in economic status, but we have no reason to think that being able to support oneself well outside the marriage will motivate divorce unless marital satisfaction is low. Our finding about the effects of women’s employment on their own exits and its interaction with marital dissatisfaction fits the prediction made by bargaining/exchange theory: women’s employment doesn’t encourage exit for those more satisfied with their marriage than average, but among those with a dimmer assessment of the marriage, employment ups the odds of leaving (Figure 1). Why do these women use their resources to leave rather than to exchange or bargain to change the marriage to be more to their liking? We suggested that the asymmetric nature of gender change is such that bargaining (“voice”) is particularly difficult if women want men to take on some of the traditionally female tasks such as housework or routine child care. The devaluation of everything associated with women makes it stigmatizing for men to make these changes and leads to their resistance. If this is part of what women would otherwise bargain for, the exit alternative is more likely because the bargain is unlikely to be struck. Overall, the asymmetry of gender change leads women’s resources to be more usable as “exit” than “voice.”

What about the effects of women’s employment on men’s exit? The institutionalist perspective predicts that men (or women) are more likely to leave if women are employed because the marriage violates gender norms and thus receives less social support. (The economic specialization model has the same prediction but for a different reason: because women’s employment entails an absence of specialization that is seen to lower the productivity gains of marriage and hence the material motive for staying married.) Bargaining/exchange predicts the opposite—that men are less likely to leave employed wives because they will lose their share of her income if they leave (see Table 1). In fact, contrary to either of these predictions, we find that their wives’ employment has no effect, either positive or negative, on men’s exits. Why is this? It is possible that both described mechanisms operate, so that the predicted positive and negative effects just cancel. Alternatively, considering today’s forms of gender inequality helps us see both predictions as less feasible. The bargaining prediction is weakened by recognizing that, even though most wives are employed, women’s earnings are low, absolutely and relative to men’s, reducing their power to provide an economic motivation for husbands to stay. Moreover, men may still believe that having power over a woman is to be preferred to being able to share her earnings. As for the institutionalist prediction, while marriage is still a gendered institution, women’s employment has become so standard that it is hard to imagine that it violates norms or makes couples and their networks question whether this “looks like a marriage,” particularly when wives continue to uphold normative expectations for doing most household labor. As for why nonspecialization does not lessen the gains for marriage sufficiently to encourage men’s exit, as Becker claims, we suggest that modern marriages are simply not held to an “efficiency in production” standard. All this helps explain why men with employed wives are no more likely to leave than are men married to homemakers.

Men’s employment is predicted by all theories to encourage women to stay. In the institutionalist view, his nonemployment is a serious violation of the gendered norm of male breadwinning, and we argued that gender change has been so asymmetric that, even as women’s employment has grown, the norm mandating male employment is fully in force. In the bargaining/exchange view, men’s employment encourages women to stay because they get to share his earnings if they do. Our finding that women are less likely to leave if husbands are employed is thus consistent with all these; we suspect it reflects both the bargaining/exchange and institutionalist processes. That is, women leave more when men are not employed both because men aren’t providing an economic incentive for them to stay, and because it seriously violates norms of what a marriage is supposed to be when the man does not have a job.

Theories differ on how men’s employment will affect their own exits from marriage. For reasons discussed above, specialization and institutionalized norms suggest that men (as well as women) will be more likely to leave if men are not employed. In contrast, the “men’s independence” version of exchange/bargaining theory suggests that, in the presence of low marital satisfaction, men who are employed will leave marriages at elevated rates, while the unemployed won’t be able to leave and support themselves. We find no “men’s independence” effect; on the contrary, husbands who aren’t employed are more apt to initiate divorce, despite their economic dependence, and even when they are quite happy in their marriages (Figure 1). Despite the fact that their own economic dependence creates an incentive to stay when they are not employed, nonemployed men are more likely to leave, suggesting that norms are at play, as posited by the institutionalist view. This effect is equally consistent with the specialization or institutionalist view, but, we find it more plausible that, in cases where men leave in response to their own nonemployment, it is more likely because the marriage doesn’t look like they think a marriage is supposed to, than that they leave because of insufficient efficiency gains owing to nonspecialization.

Returning to Table 1, comparing findings to theoretical predictions to see how well each theory performed, we see mixed results for both. The marriage-as-gendered-institution view correctly predicts the positive effects of men’s employment in minimizing risks of divorce for either partner’s exit. We noted the asymmetric nature of gender change, such that, despite increases in women’s employment, there is little toleration for men not remaining employed breadwinners. A deviation from this norm appears to make either partner more likely to leave. In contrast, institutionalist theory gets it wrong as regards the effects of women’s employment (Figure 1); it suggests that men will leave their wives more if the wives are employed, but we find no such effect. It predicts that women are more likely to leave if they are employed, and that this effect is not only present when marital satisfaction is already low. What we find, instead, is that women’s employment elevates their exits only when their marital satisfaction is low. This interactive effect is what is predicted only by bargaining theory—when their marital satisfaction is low, the independence (allowing self-support) of employment encourages women to leave, but if satisfaction is good women will not leave just because they can support themselves.

Why did the institutionalist prediction work for men’s employment but not women’s? Deriving predictions from institutionalist theory is only possible when we know what the cultural understandings of marriage are, and these have changed. We suggested that, while the expectation that men be breadwinners is still fully in force, women’s employment has become well accepted, and now does not challenge whether a couple’s arrangement “looks like a marriage.” This is a reflection of the asymmetric nature of gender change. Thus, we believe the institutionalist view is correct to claim that conformity with taken-for-granted understandings about marriage affects marital satisfaction and/or the social support upon which marriages depend. The problem with our initial predictions from institutionalist theory was that they did not acknowledge the change in norms about women’s employment. Institutions are defined by staying power, but they are not unchangeable.

We believe that acknowledging the asymmetric nature of gender change is necessary to understand our findings. Consider this broad-brush interpretation of our findings: men’s nonemployment increases divorce because it violates norms, while women’s employment increases divorce by providing a way to support oneself outside marriage for women deeply unsatisfied with their marriages, not because it violates norms. Both of these effects probably emanate from the greater change in women’s than men’s roles; women’s employment has increased and is accepted, men’s nonemployment is unacceptable to many, and there is cultural ambivalence and lack of institutional support for men taking on “feminized” roles such as household work and emotional support. Women’s employment is translated into exit rather than voice in many cases because the changes that would most increase women’s marital satisfaction would entail men “feminizing” their roles in a way that many are still ambivalent about and institutions don’t support. Men’s breadwinning is still so culturally mandated that when it is absent, both men and women are more likely to find that the marital partnership doesn’t deserve to continue.