The importance of financial considerations in divorce decisions.

Peters, H. Elizabeth

I. INTRODUCTION

One of the most dramatic and commonly cited consequences of divorce is the fall in family money income--especially for households headed by divorced women. Using the Panel Survey of Income Dynamics, Duncan and Hoffman |1985~ estimate that the family income of divorced and separated women fell by 19 percent during the year following divorce or separation. A study by Nestel et al. |1983~ using data from the Mature Women's cohort of the National Longitudinal Survey of Labor Market Experience indicates that the poverty rate for households headed by women who were divorced during the survey (and did not remarry) rose from 10 percent just prior to divorce to 25 percent immediately after dovorce. Similar results are reported by Mott and Moore |1982~ and Espenshade |1979~. These statistics, however, may overestimate the long-term negative consequences of divorce, because they ignore the possibility of an improvement in financial circumstances through remarriage.(1) Mott and Moore |1982~ find that even for women who do not remarry, economic status improves over time due primarily to better employment prospects.

Given the large financial costs of divorce, especially in the short run, it is reasonable to ask how important are financial considerations in divorce decisions. This paper explores the linkages between the financial consequences of divorce and the decision to become divorced. The paper extends previous work by measuring these consequences over a period of time rather than at one point in time. In particular, the paper addresses two issues: (1) Do the expected financial consequences of divorce affect the decision to become divorced? and (2) What is the relative importance of short- versus long-run considerations?

A standard choice-theoretic model predicts that the smaller are the financial costs of divorce, the greater is the probability of divorce. The financial opportunity cost of divorce is measured as the difference in the expected present values of the income streams within marriage and at divorce. The expected income stream at divorce includes the financial value attached to a possible remarriage, as well as income from employment, welfare, alimony, and child support payments.

The empirical analysis utilizes data from the National Longitudinal Survey of Work Experience of Young Women (NLS). A sample of those divorced between 1973 and 1978 is observed over time during marriage, after divorce, and, for some, during a remarriage. Thus the actual income changes resulting from divorce and remarriage are measured. Income changes for a sample of continuously married women are also measured over a comparable period. The expected changes in income for each subsample are estimated as a function of socio-economic variables, and the values are imputed for the entire sample to explore the relationship between these expected income streams and decisions about divorce. The result indicate that the expected short-term financial consequences of divorce are a better predictor of subsequent divorce than are the longer-term consequences. This evidence implies that individuals behave as if they face high discount rates.

II. ECONOMIC MODELS OF DIVORCE

One of the first economic models of divorce was developed and emphirically tested by Becker, Landes, and Michael |1977~. As in any choice-theoretic model, divorce occurs when the gains from that choice exceed the benefit from remaining married. Gains from marriage include both a pecuniary component, such as pooled household income from the market work of the husband and wife, and a non-pecuniary component, such as children, love, companionship, and household goods from home production. Gains from divorce include own labor market income, home production (which may differ between the divorced and married states) and the expected value of potential new relationship (e.g., remarriage). Marriage-specific capital, such as children, has a higher value within marriage and increases the opportunity cost of divorce. In their empirical work, Becker, Landes and Michael |1977~ focus primarily on estimating the effects of variables which relate to the value of the marriage--duration of marriage, age at marriage, family income, children, and similarity of social characteristics between the husband and wife.

A few studies have explored the other side of the coin--the value of opportunities after divorce, commonly called the independence effect. Most of this research has been limited to measuring the impact of the size of potential welfare benefits available to divorced women with children on the likelihood of divorce. The evidence of this effect has been mixed.(2) Ross and Sawhill |1975~ and Mott and Moore |1978~ also provided some evidence on the relationship between potential market earnings of the wife and the probability of divorce.

A stricter test of the choice-theoretic model would compare the discounted streams of future returns from each choice. A paper by Danziger et al. |1982~ took the first step in this direction. Their emphirical work estimates the probability that a woman is a household head (i.e. not married) as a function of the difference between the income she could expect if she were to marry versus that if she were to head her own household. One limitation of their study is that the analysis is cross-sectional and ignores the changes in income over time resulting from a particular choice. Our paper extends that previous work by examining the impact on divorce decisions of both the expected immediate financial consequences and the expected consequences over the long run.

Recent work on the economic consequences of divorce by Duncan and Hoffman |1985~, Mott and Moore |1982~, and Peters |1992~ has documented a stylized pattern of income over time: family income for women falls precipitously at divorce and then begins a gradual recovery. The speed and level of that recovery vary across women and depend most importantly on whether the woman remarries and also on improved employment opportunities and changes in her labor market behavior. Duncan and Hoffman |1985~ find that by the fifth year following divorce remarried women have attained a level of family income that is comparable to continuously married women at the same point in their life cycle.

In this context we might view divorce as an investment decision. A woman would be willing to incur short-term costs in order to receive long-term benefits. In particular, the period immediately following divorce can be characterized as a transitional period during which search for a new spouse takes place. This assumption reflects the reality that a large majority of younger women will eventually remarry after a divorce. The value attached to a possible remarriage, as well as income from employment, welfare, alimony, and child support payments should properly be included in the measure of economic prospects after divorce.

The above discussion reveals a striking contrast between the short-term financial consequences and the long-term financial consequences of divorce. The importance of long-term versus short-term considerations in divorce decisions will depend on several factors. First, it will depend on the magnitude of the short-term loss and the magnitude and speed of the recovery. Secondly, it may depend on the degree of imperfection in capital markets. For example, many married women do not establish credit in their own names and at divorce have difficulty in borrowing to smooth consumption. If the marginal utility of money is not constant, individuals will place a greater weight on income losses than on income gains. Thirdly, prospects at divorce are uncertain, and it may be difficult to predict the timing of remarriage. The short-term consequences are more certain. This factor might also lead an individual to place a greater weight on short-term consequences. Whether financial considerations matter and the importance of short- versus long-term consequences are emphirical questions that we explore in the following sections of the paper.

One problem with focusing on opportunities after divorce is that these are measured for one individual, whereas a plausible theoretical model predicts that divorce is related to the sum of the gains for the husband and wife. This complication has often been avoided either by assuming that the wife is the only actor (see Danziger et al. |1982~) or that there is implicit bargaining or exchange between the husband and wife so that they eventually reach a common decision (see Ross and Sawhill |1975~; Becker, Landes, and Michaels |1977~; and Peters |1986~). An alternative assumption consist with the limitations imposed by the data and with a model of joint decision making is that the net financial consequences of divorce for the husband and wife are uncorrelated (or at least not negatively correlated). If this assumption holds, a ceteris paribus increase in the wife's net gains to divorce will increase the probability of divorce.

III. ESTIMATION ISSUES

We model the probability of divorce as a function of the expected pecuniary and non-pecuniary gains or losses from divorce. Assuming that utility is a linear function of pecuniary and non-pecuniary factors, a woman will choose to divorce if

(1) |delta~|E(P|V.sub.d~ - P|V.sub.m.~)~ |is greater than~ c

where P|V.sub.d~ and P|V.sub.m~ are the present values of the future income streams from the choices of divorce, d, and staying married, m; E is the expectations operator; delta is the weight on pecuniary factors in the utility function; and c is the difference between the expected non-pecuniary benefits to remaining married and becoming divorced.(3) The variable c can also be interpreted as the non-pecuniary opportunity cost of divorce.

Because c is unobservable, we assume that it can be represented as a linear function of a vector of variables, Z. In the theoretical literature discussed above, Z would include variables such as the presence of children, marriage duration, race, family structure when growing up, ethnicity, and age at marriage:(4)

(2) c = ||alpha~.sub.c~Z + ||epsilon~.sub.c~.

A random variable, ||epsilon~.sub.c~ is included to capture any unobserved individual-specific "tastes" for marriage. Thus a quasi-structural divorce equation can be written as

|Mathematical Expression Omitted~

I is unobserved, but we can observe an indicator variable |I.sup.*~ = 1 |is greater than~ 0 and |I.sup.*~ = 0 if I |is less than~ 0.

The next step in the estimation strategy is to characterize E(P|V.sub.d~ - P|V.sub.m~). If expectations are rational, the actual present values are unbiased estimates of the expected present values. With complete data on individual discount rates, r, and all future incomes for each choice, |Y.sub.d~(t) and |Y.sub.m~(t), the present values can be calculated. The data requirements to estimate this model, however, are severe, and no perfect data set is available. In particular, the empirical implementation of the model must address three problems: (1) we do not observe an individual's complete future income stream; (2) we do not observe an individual's discount rate; and (3) we only observe the income stream at divorce for those who choose to divorce; likewise we observe the income stream for staying married only for those who actually make that choice.

To solve the first problem we must make assumptions to impute the missing future income based on the available data. This imputation procedure is discussed in detail in the data section.

The second issue, the choice of discount rate, reflects the weight that individuals place on short- versus long-term consequences. If the discount rate is high, then the immediate consequences are more important; if the discount rate is low, then the long-run consequences play a role as well. Studies by Hausman |1979~ and Hartman and Doane |1986~ of the purchase of consumer durables suggest that average consumer discount rates may be 30 percent or higher. If marginal utility of income is not linear, there is also reason to believe that the short-run loss in income after divorce is weighted more heavily than the long-run recovery. Furthermore, it may be difficult to predict factors such as remarriage, which lead to long-run changes in income. This, too, would lead to a greater weight on the short-run.(5)

In the empirical estimation, we utilize two different present value measures, one calculated with a discount rate of 10 percent and the other calculated with a discount rate of 30 percent, corresponding to the evidence on consumer durables. We then estimate separate structural divorce models which include these different present value measures as explanatory variables. Comparing the fit of these separate divorce models provides a natural way to test whether expected long-run changes in income play a role in divorce decisions or whether it is primarily short-run considerations that matter.

The third problem--that we only observe the consequences of the choice that is actually made--can be solved by using a standard switching regressions model.(6) First, we write P|V.sub.d~ and P|V.sub.m~ for all women as functions of exogenous variables which are observed at the time the decision about divorce is being made:

(4) P|V.sub.m~ = ||tau~.sub.m~X + ||epsilon~.sub.m~

|Mathematical Expression Omitted~

P|V.sub.m~ is determined by a vector of variables, X, which primarily affect the earnings capacity and labor supply of each spouse. The vector X includes race, education, age, location, current earnings and income. P|V.sub.d~ is determined by a vector of varibales, |Mathematical Expression Omitted~, which affect (1) the earnings capacity and labor supply of the woman and (2) how quickly (if at all) she remarries and the income of her potential new spouse. Many of the same variables will affect employment decisions and remarriage decisions. For example, the precence of children may raise the cost of being employed and lower the probability of remarriage.(7) Therefore we cannot identify the separate impact of employment and remarriage prospects on divorce. |Mathematical Expression Omitted~ includes the woman's education, earnings capacity, race, ethnicity, duration of marriage, the presence of children, and family income prior to divorce.

Because P|V.sub.d~ and P|V.sub.m~ are observable only for those individuals who make the particular choices, the equations that can be estimated are conditional on the choices being made:

|Mathematical Expression Omitted~

|Mathematical Expression Omitted~

It is now well known that if there is a correlation between ||epsilon~.sub.d~ or ||epsilon~.sub.m~ and the selection rule, the second term on the right-hand side of equations (6) and (7) is not zero. OLS regressions which omits this term produce biased estimates of the parameters ||tau~.sub.m~ and ||tau~.sub.d~. A standard solution to the selectivity problem developed by Lee |1978~ and Heckman |1979~ involves estimating a reduced-form probit equation for the likelihood of divorce which includes all the variables in X, |Mathematical Expression Omitted~, and Z. The inverse of the appropriate Mills ratio from that probit is then used as an instrument for the omitted term. Once equations (6) and (7) have been estimated and corrected for selectivity bias, the unconditional values of P|V.sub.d~ and P|V.sub.m~ can be calculated for the entire sample and are included as regressors in equation (3), the quasi-structural probability of divorce.

IV. THE NLS SAMPLE AND DATA CONSTRUCTION

The empirical analysis uses data from the NLS Young Women's cohort. In this survey a nationally representative sample of 5,159 women ages 14-24 were initially interviewed in 1968, and the interviews have continued every year or two up to the present. The data utilized in this paper include information from the eleven interviews up to 1982. At that time the women in the sample were between ages 28 and 38. The data contain detailed information about the characteristics of the respondent including age, race, family background, and, for each survey, education, fertility, employment, earnings, other family income, and household structure. In addition, marital histories obtained from the respondents provide information about the exact timing of marital transitions.

For the subsequent analysis, a divorced woman is defined as someone who was in her first marriage at the 1973 survey date and who subsequently got divorced before the 1978 survey date. The counterpart, a continously married woman, is someone who was in her first marriage at the 1973 survey date and who was still married as of the 1978 survey date. The sample is restricted to those who became divorced before 1978, because by 1982, the date of the last survey available, the change in income over a five-year period could be measured for everyone in the sample. In 1973 these women were 20 to 30 years old. Thus a significant fraction had still not been married and were not eligible to be included in the analysis. Given these restrictions, the final sample contains 1203 continuously married women and 123 divorced women.(8) One consequence of the age restrictions in the data set is that the empirical analysis will focus on the divorce experiences of younger women. The displaced homemaker phonemenon which is more relevant to older women will not be captured in this analysis.

The income measure used in the present value calculations is family income adjusted for family size. This measure includes own earnings, income from a new spouse if remarriage occurs, welfare, child support, alimony, and other non-earned income. The adjustment for family size is based on a household equivalence scale used in calculating poverty thresholds as reported in the U.S. Department of Commerce, Current Population Reports |1986~. In subsequent discussion we call this measure of income "equivalent family income."(9)

Respondents in the Young Women's cohort of the NLS were only interviewed during three out of every five years. Therefore the first step in constructing the present value variables was to impute income for the missing years for t |is less than~ 5. This was done by taking a simple average of income in the two surrounding years for each individual.(10) The second step was to impute income for years t |is greater than~ 5. Because there is little information in the literature on which to base such long-run projections, especially for divorced women, we assume that income remains constant over time, i.e. if t |is greater than~ 5, then Y(t) = Y(5).(11,12) Incomes up to t = 30 are then used to calculate present values.

To test for the sensitivity of the results to the assumptions that were used for the present value calculations, we also calculate alternative indicators of the short- and longer-term financial consequences of divorce: the average annual growth (or percentage change) in income measured over a two-year period and the average annual growth in income measured over a five-year period. These two variables are calculated as follows:

|Mathematical Expression Omitted~

|Mathematical Expression Omitted~

where Y(0) is income in the year before divorce for women who divorce and income during a comparable period for women who stay married.(13) Y(2) and Y(5) are incomes two and five years later, respectively.(14,15)

Due to data limitations, this paper defines long-term consequence as changes in income over only five-year period. Is this period long enough to reflect the long-run? Alternatively, we can ask, are the changes in income over five years substantially different than what we observe in the year following divorce? To answer these questions it is necessary to examine the two events or behaviors that are primarily responsible for improvements in the financial circumstances over time: remarriage and employment. By year five, 47 percent of the divorced sample had remarried. This represents more than 60 percent of those who will eventually remarry.(16) In contrast, less than 20 percent had remarried within one to two years of divorce. Incomes of those who did not remarry by year five had also substantially improved by this period.

V. EMPIRICAL RESULTS

Table I presents means by subsequent marital status. The time-dependent variables (except future incomes) are measured during the survey prior to divorce (or the comparable survey for continuously married women; see footnote 13). Consistent with other studies, divorced women are on average younger, have marriages of shorter duration, higher labor force participation rates, and are less likely to have children.(17)

Table I illustrates the difference in outcomes for divorced women by whether remarriage had occured. The equivalent family income of divorced women who had not remarried was substantially lower during the first year following divorce; by the end of the five-year period, equivalent family income was still 9 percent lower than before divorce. In contrast, after five years remarried women are slightly better off than they were just prior to divorce.(18) In contrast to many other studies, Table I TABULAR DATA OMITTED indicates that pre-divorce equivalent family income for the divorced sample is slightly higher than that for the married sample.(19)

Table II presents the results of the present value (P|V.sub.m~) and growth rate (|g.sub.m~) regressions for married women. As discussed above, the financial circumstances for continuously married women should depend primarily on factors affecting a woman's labor supply and earnings and her husband's labor supply and earnings.(20) TABULAR DATA OMITTED The wife's education is positively related to P|V.sub.m~. This relationship may be capturing the usual correlation between education and the slope of the age-earnings profile that is found in much of the labor literature. Becker |1981~ provides evidence on assortative mating that shows a positive correlation between the education of the wife and the education of the husband. Therefore wife's education may also act as a proxy for husband's education and may reflect a steeper age-earnings profile for the husband as well. Not surprisingly, women with high current family income, women who live in urban areas, white women, and those who are currently working in the labor force also have a higher present value of future income. The result in Table II indicate that the level of the wife's earnings and the husband's age do not significantly affect the present value of future income in the married state.

Except for the constant term, the estimates in column 1 (discount rate = 10 percent) are very similar to the estimates in column 2 (discount rate = 30 percent). Thus the effects of the explanatory variables on long-run income for married women are qualitatively similar to the effects of those same variables on short-run income.

The effects of the explanatory variables on the present value of income are expected to be somewhat diffferent than the effects of those same variables on the income growth rates, because P|V.sub.m~ is measured as a level, but |g.sub.m~ reflects a change in income. As in the present value regressions, highly educated women and white women have higher income growth rates (although the coefficient on race does not quite reach standard levels of significance). In contrast to the present value results, family income has a negative impact on growth rates. Holding education constant, higher earnings may represent a temporary deviation from expected earnings, and thus lead to lower growth rates over time.(21) The effect of the wife's prior labor force participation on growth rates is also negative and significant in the growth rate regression using a five-year change. This result could reflect the "added worker" effect where the wife enters the labor market in response to a fall in the husband's earnings caused by unemployment or other unanticipated events.(22) The sign on husbands's age is unexpected. The usual concave age-earnings profile should lead to a negative coefficient. Perhaps the limited age range of the sample is the cause of this result.

Table III presents the results of the present (P|V.sub.d~) and growth rate (|g.sub.d~) regressions for divorced women. The interpretation of the coefficients in this table represents the combined effect on remarriage opportunities and labor market opportunities. The effect of wife's education is positive, and in each case it is larger than in the corresponding regression for the married sample. One explanation for this result is that own labor market earnings are a larger component of total family income for divorced women than for married women. Most studies have not found a strong relationship between education and remarriage. Therefore it may be reasonable to conclude that education operates TABULAR DATA OMITTED primarily through its effect on the woman's labor market earnings.

The effect of current labor force participation on P|V.sub.d~ and |g.sub.d~ is positive and substantially lager than its effect in the married women's regressions. If this variable is a signal for labor market commitment, it may reflect higher future labor supply or the choice of an occupation with a steeper age-earnings profile. In addition, the labor force variable may operate through its effect on remarriage. Peters |1985~ provides evidence that labor force participation increases the probability of remarriage, perhaps because the opportunities for finding a marriage partner are greater for women in the labor force. Wife's earnings exert a negative effect on income growth rates, although the effect is only significant in the five-year income growth rate equations.

It is interesting to note that the presence of children has no significant effect on financial circumstances at divorce. This result is surprising, because we might expect a negative effect to operate both through a reduced probability of remarriage and through the costs which children impose on labor market activities. Duration of the previous marriage is another variable which has been found to increase the probability of remarriage. This variable also has no effect on present values or growth rates in this sample of divorced women.

The coefficients on the selection terms are, in general, negative for both married and divorced women (and are statistically significant in the present value regressions). The value of the inverse Mills ratio (lambda) is always positive for divorced women and always negative for continuously married women.(23) The pattern of coefficients on lambda reported in Tables II and III implies that the correlation between the unobserved ability to produce income in the divorced and the married states is, in general, positive, and that divorced women have a lower level of that ability than do married women. In other words, divorced women have unobserved characteristics that lead to lower than average income in both the married and divorced states, whereas married women have unobserved characteristics that would lead to higher than average income in both the married and divorced states. For example, if divorced women have a characteristic which raises search costs in both the marriage market and the job market, they may choose a spouse with a lower than average quality and may accept a job with lower than average wages. The results imply that divorced women have lower income than average in both states, but are relatively better off in the divorced state.(24)

Table IV presents the estimates of the parameters of the structural divorce probit. Four specifications are shown, each one using a different measure of the financial consequences of divorce. The results indicate that financial considerations do significantly affect divorce decisions, but it is the short-term consequences that matter. When the discount rate is increased from 10 percent (column 1) to 30 percent (column 2), the coefficient on the difference in the present value of the two income streams nearly doubles in size and the precision of the estimate increases substantially.(25) Similarly, when income TABULAR DATA OMITTED growth rates are calculated using a five-year change in income, which ignores the substantial drop in income immediately following divorce, the coefficient on married income is the only one that comes close to significance. The coefficient on divorce income growth is essentially zero economically and statistically.(26) In column 4 only the short-term change in income over a two-year period is utilized in calculating growth rates. In this specification the coefficient on the difference in the growth rates has the expected sign and is strongly significant.

The coefficients on the variables representing non-pecuniary factors are generally consistent with other studies and are reasonably stable across the various specification in Table IV. The Intact Family variable represents possible intergenerational transmission of marital stability. As expected, women who come from stable families have lower probabilities of divorce, although this effect is not measured precisely. Women who live in urban areas have higher divorce rates than those living in rural areas. Duration of Marriage, a proxy for marriage-specific capital, significantly reduces the probability of divorce.(27) The presence of children, another form of marriage-specific capital, also reduces the probability of divorce.(28,29) Finally women who marry at older ages have lower probabilities of divorce. This variable could represent the outcome of differences in the cost of search or differences in the rate of discount.(30)

VI. CONCLUSIONS

This paper explores the link between the expected financial consequences of divorce and the decision to become divorced. A standard choice-theoretic model predicts that, holding other factors constant, the probability of divorce should be negatively related to the financial opportunity cost of divorce. This opportunity cost is measured as the difference in the present values of the future income streams that a married woman might expect if she were to remain married versus getting divorced. Because the stylized pattern of income for divorced women shows an sharp drop in income immediately after divorce followed by a gradual improvement in income over time, the short-term financial consequences of divorce are often quite different than the longer-term financial consequences.

Results utilizing data from the NLS Young Women's cohort indicate that women do take the financial consequences into consideration when making decisions TABULAR DATA OMITTED about divorce. However, it is the short-term consequences that matter more. This may be the result of imperfect capital markets (for example, many married women have not establish credit in their own name), the large uncertainty about outcomes at divorce, or if the marginal utility of income is not constant, a greater weight is put on the short-term loss compared to the longer-term recovery. This evidence that individuals behave as if they have very high discount rates is also consistent with a number of studies estimating high discount rates for the purchase of consumer durables.

Assistant Professor of Economics and Research Associate in the Population Program, University of Colorado. This researedurch was funded by NICHD grant #5R23HD21882-02. The author would like to thank Bettina Herr and Laura Argys for excellent research assistance and George Jakubson, Tom Mroz, and William Schulze for comments.

1. See Duncan and Hoffman |1985~.

2. See Moffitt |1992~ for a review of these studies.

3. The weight on non-pecuniary factors in the utility function is normalized to one.

4. C may also include pecuniary costs of divorce such as lawyer's fees. Since these are not obdervable in the data, the effect of this kind of financial cost on the probability of divorce cannot be distinguished from the effects of non-pecuniary costs of divorce.

5. See Leigh |1985~ for a model of divorce which explores the tradeoff between the mean and variance of future income.

6. See Willis and Rosen |1979~, Borjas and Rosen |1980~, and Lee |1978~ for similar applications of this estimation technique.

7. See Peters |1988~.

8. A few other restrictions were also imposed: (1) the woman had to have been interviewed during all surveys--this reduced the initial sample from 5159 to 3650; (2) information about the relevant personal characteristics, income, and dates of divorce could not be missing; and (3) two women who reported an age at marriage of less than 13 were eliminated from the analysis.

9. There is some debate among economists about the appropriateness of adjusting income for differences in family size, if fertility is a choice variable. The empirical analysis in this paper was also done using unadjusted family income. The results were qualitatively similar to the ones reported in Tables II-IV.

10. To test for the reasonableness of this imputation procedure, we compare the mean of actual income with the mean of imputed income separately for each year and for each marital status. In general, the difference in the means of the two groups is not significant.

11. Although women who divorce earlier would have longer period over which we can observe post-divorce income, we do not use information on income for t |is greater than~ 5, because this sample is biased towards women whose marriages were of very short duration or who married at a very young age.

12. The assumption that income growth rates are zero in all years t |is greater than~ 5 is obviously unrealistic. Because the structural divorce equation is a function of the difference in P|V.sub.m~ and P|V.sub.d~, however, this assumption is equivalent to the less restrictive assumption that the age-income profiles in the two states are parallel after year five.

13. For women who separated before divorce, we define the "survey before divorce" (i.e. t = 0) to be the last survey during which a husband is present in the household. For those women Y(0) and the other time-dependent explanatory variables (except future income) are measured during the survey before separation. For continuously married women there is no natural starting date for |Y.sub.m~(0) comparable to the survey prior to divorce for those who become divorced. To circumvent this problem, each continuously married woman is assigned at random one of the three possible surveys--1973, 1975, and 1977--to be comparable to the "survey before divorce" (note that respondents were not interviewed during either 1974 or 1976). The assignment is done so that the distribution of the comparable survey for continuously married women is the same as the distribution of the survey before divorce for divorced women.

14. We do not use information on family income in the year in which divorce occurred (t = 1), because the measure of family income in that year is confounded by the change in household structure that occurred.

15. Note that because of the time patterns of interviews there are some individuals for whom this two-year change in income is not available. To estimate g(2) we use the sample of 100 divorced women and 1001 married women who were interviewed during the appropriate survey year.

16. This assertion is based on independent projections by Schoen et al. |1985~ that 77 percent of young divorced women will eventually remarry. The five-year remarriage rate in this NLS sample is highly similar to national estimates of five-year remarriage rates.

17. See Ross and Sawhill |1975~, Becker, Landes, and Michael |1977~, Peters |1986~, and Johnson and Skinner |1986~.

18. The mean equivalent family income for continuously married women actually fell (in real dollars) by 1 percent over the five-year period. This surprising finding is due primarily to two factors. First, because the period over which income is measured includes the recession of the late 1970s and early 1980s, real average total family income for continuously married women rose only slowly at an annual rate of 1.6 percent. The women in this sample are of prime child-bearing age, and the growth in family size over time for married women tended to offset the small increase in total family income, resulting in a decline in average equivalent family income. Second, the average ratio of Y(5) to Y(0) is higher than the ratio of average incomes in the two periods and actually shows a small average increase in equivalent family income over the period.

19. One reason why the income of the divorced sample might be biased upward relates to differential attrition of divorced women. If divorced women are more likely to be eliminated from the sample because they were not interviewed in subsequent years, and if non-interview status is also more likely for lower-income women, the income of divorced women who remain in the sample will be higher than the average for all divorced women. In addition, the income measures reported in Table I are adjusted for family size. Because women who divorce have fewer children than women who do not, adjusting for family size makes the divorced sample appear to be relatively better off in the year prior to divorce than the married sample. Pre-divorce family income unadjusted for family size is also slightly higher for the divorced sample ($16,585) than for the married sample ($15,769), although the differences are much smaller than observed for equivalent family income and are not statistically significant.

20. These women were continuously married through 1978. However, 1.2 percent of the sample became divorced after 1978 but before 1982. An analysis which eliminates this group from the married sample does not produce substantially different results from the ones reported in the paper.

21. See Borjas and Rosen |1980~ for similar reasoning and results.

22. Johnson and Skinner |1986~ show that women who eventually divorce have higher labor force participation rates during marriage. Thus women in the married subsample who particiapte in the labor force might be more likely to get divorced after 1978, and their lower income growth would partly be a reflection of later divorce. However, this hypothesis is not likely to be a significant part of the explanation, because there are so few women in the married subsample who became divorced after 1978 and before 1982 (see note 20).

23. Lambda = |phi~(X|beta~)/|phi~(X|beta~) for the divorce sample and lambda = -|phi~(X|beta~)/|1-|phi~(X|beta~)~ for the married sample, where the |beta~'s are obtained from the reduced-form probit in Appendix A.

24. Note that in the occupation choice model of Roy |1951~, incomes are the only variables that matter. In the divorce decision, however, non-pecuniary factors will play an important role as well. Therefore it is possible in this model for the difference between married and divorced financial prospects to be positive for those who choose divorce as long as the non-pecuniary gains to divorce are large enough.

25. In results from additional specifications not presented here, the size of coefficient on P|V.sub.d~ - P|V.sub.m~ and the precision of the estimate increased monotonically as the discount rate used in calculating the present values increased from 5 percent to 40 percent.

26. The theory indicates that it is the difference in income prospects that should matter in the decision to become divorced. We test the restriction that coefficient on P|V.sub.d~ (|g.sub.d~) and P|V.sub.m~ (|g.sub.m~ are equal but have opposite signs. The restriction was not rejected in three of the four specifications, and in those cases only the coefficient on the difference is reported. In column 3 the restriction was rejected and the separate coefficients are reported.

27. This effect could also be due to unobserved heterogeneity where "divorce-prone" individuals become divorced more quickly.

28. A recent paper by Waite and Lillard |1991~ explores the relationship between children and divorce in detail. They find a very complex pattern which depends on the age distribution of children. However, they find that the impact of other variables on divorce does not change when a sampler measure of children is used. Similarly, the results in our paper are not substantially different when a variable measuring the number of children younger than six is included. Because we focus primarily on the impact of the expected financial consequences, we report the results from the more parsimonious specification for children.

29. One referee suggested that the effect of expected financial consequences on divorce decisions might differ for women with and without children. An additional model which included an interaction between Any Children and P|V.sub.d~ -P|V.sub.m~ was run using both the 10 percent and 30 percent discount rate measures. The coefficients on the interaction terms were not significantly different from zero, and the estimates of the other parameters were virtually identical to those reported in Table IV. These results would imply that in making decisions about divorce, both women with and without children respond to the expected financial consequences of divorce in similar ways.

30. See Becker, Landes, and Michael |1977~ for a more detailed argument about the relationships between these variables and divorce.

REFERENCES

Becker, Gary S. A Treatise on the Family. Cambridge, MA: Harvard University Press, 1981.

Becker, Gary S., Elisabeth M. Landes, and Robert T. Michael. "An Economic Analysis of Marital Instability." Journal of Political Economy, December 1977, 1141-87.

Borjas, George J., and Sherwin Rosen. "Income Prospects and Job Mobility of Younger Men," in Research in Labor Economics, Ronald G. Ehrenberg (ed.) vol 3. Greenwich, CN: JAI Press, 1980, 159-81.

Danziger, Sheldon, George Jakubson, Saul Schwartz, and Eugene Smolensky. "Work and Welfare as Determinants of Female Poverty and Household Hardship." Quarterly Journal of Economics, August 1982, 519-34.

Duncan, Greg J., and Saul D. Hoffman. "A Reconsideration of the Economic Consequences of Marital Dissolution." Demography, November 1985, 485-97.

Espenshade, Thomas J. "The Economic Consequences of Divorce." Journal of Marriage and the Family, August 1979, 615-26.

Hartman, Ray S. and Michael J. Doane. "Household Discount Rates Revisited." The Energy Journal. 7(1), 1986, 139-48.

Hausman, Jerry A. "Individual Discount Rates and the Purchase and Utilization of Energy-Using Durables." The Bell Journal of Economics, Spring 1979, 33-54.

Heckman, James J. "Sample Selection Bias as a Specification Error." Econometrica, 47(1), January 1979, 153-62.

Johnson, William, R., and Jonathan Skinner. "Labor Supply and Marital Separation." American Economic Review, June 1986, 455-69.

Lee, Lun Fei. "Unionisn and Wage Rates: A Simultaneous Equations Model with Qualitative and Limited Dependent Variables." International Economic Review, June 1978, 415-33.

Leigh, J. P. "Divorce as a Risky Prospect." Applied Economics 17, 1985, 309-20.

Moffitt, Robert. "Incentive Effects of the U.S. Welfare System: A Review." Journal of Economic Literature, March 1992, 1-61.

Mott, Frank L., and Sylvia F. Moore. "Marital Transitions and Employment," in The Employment Revolution, edited by Frank L. Mott. Cambridge, MA: MIT Press, 1982, 120-45.

-----. "The Causes and Consequences of Marital Breakdown," in Women, Work and Family, edited by Frank L. Mott. Lexington, MA: Lexington Books, 1978, 113-35.

Nestel, Gilbert, J. Mercier, and Lois B. Shaw. "Economic Consequences of Midlife Change in Marital Status," in Unplanned Careers: The Working Lives of Middle Aged Women, edited by Lois B. Shaw. Lexington, MA: Lexington Books, 1983, 109-25.

Peters, H. Elizabeth. "Responses to Divorce: Remarriage and Employment Among Women in the U.S." Population Program Working Paper 85-6, University of Colorado, December 1985.

-----. "Marriage and Divorce: Informational Constraints and Private Contracting." American Economic Review, June 1986, 437-54.

-----. "Retrospective Versus Panel Data in Analyzing Life-Cycle Events." Journal of Human Resources, Fall 1988, 488-513.

-----. "Economic Changes Over Time," in Dividing the Child: Legal and Social Dilemmas of Custody, by Eleanor E. Maccoby and Robert H. Mnookin with Charlene E. Depner and H. Elizabeth Peters. Cambridge, MA: Harvard University Press, 1992, Chapter 10.

Ross, Heather L., and Isabel V. Sawhill. Time of Transition. Washington D.C.: The Urban Institute, 1975.

Roy, A. D. "Some Thoughts on the Distribution of Earnings." Oxford Economic Papers, June 1951, 135-46.

Schoen, Robert, William Urton, Karen Woodrow, and John Baj. "Marriage and Divorce in Twentieth Century American Cohorts." Demography, February 1985, 101-14.

U.S. Department of Commerce. Current Population Reports: "Characteristics of the Population Below the Poverty Level: 1984." Series P-60, no. 152, U.S. Department of Commerce, Bureau of the Census, June 1986.

Waite, Linda J., and Lee A. Lillard. "Children and Marital Disruption." American Journal of Sociology, January 1991, 930-53.

Willis, Robert J., and Sherwin Rosen. "Education and Self-selection." Journal of Political Economy, October 1979, S7-S36.