Male marital wage differentials: training, personal characteristics, and fixed effects.

Rodgers, William M., III ; Stratton, Leslie S.

I. INTRODUCTION

There is substantial evidence that married white men earn between 8% and 15% more than unmarried white men in the United States, even after controlling for education and informal job market experience (e.g., potential or actual experience). Cross-section analysis indicates that both the level and the growth rate of wages increase during marriage. These results suggest that men may become more productive after marriage, an effect often attributed to the increased specialization possible following the formation of a multiperson household. However, panel estimates indicate that much of the differential is explained by individual-specific fixed effects, which suggest that more productive men are more likely to marry. While these results and implications are well known, the precise nature of the implied productivity differentials has received little attention in the literature. Using the 1979 cohort of the National Longitudinal Survey of Youth (NLSY79), we investigate the structure or source of the observed differential.

We do so by first extending the analysis to a sample of African American men in the United States. To our knowledge, ours is the first study to conduct a detailed analysis of the marital wage gap among African American men that is fully comparable to that already available for white men. This extension is informative because African American households are known to exhibit less specialization along gender lines and because marital patterns are quite different by race.

Next, we explore formal job training as one avenue to productivity enhancement. We outline a simple story suggesting that married men may be more receptive to training opportunities than unmarried men and present some evidence suggesting that firms may be more willing to offer training to married as compared to unmarried men. Previous analysis has focused on labor market experience or years married as proxies for both informal and formal job training. We significantly improve upon this by using direct information on formal job training to see if wage changes at or following the time of marriage are driven by training-induced productivity differentials.

Third, we explore the nature of the selection effect or the possibility that it is more productive men who marry. The NLSY79 has an extensive array of information on personal characteristics that might be associated with both marriage and earnings. If certain personal characteristics are associated with both marriage and earnings, but not typically included in wage analyses, then the marital controls in standard wage equations may simply be proxying for these characteristics. Controlling for such characteristics directly should identify the nature of the selection effect and reduce the significance of the marital indicators in cross-section estimates. The attributes we examine include cognitive skills, parental background, and personality traits.

Our final contribution is to estimate panel models to eliminate all observable and unobservable individual-specific, time-invariant factors and to reconcile these results with those obtained from cross-section models. Specifically, we exploit the different identification conditions between panel and cross-section models, examining the role of respondents who are always married, never married, or always separated/divorced. Technically, such individuals contribute to the identification of the cross section but not the panel estimates of the marital wage differential. Thus, these individuals may explain the substantial differences often observed between panel and cross-section estimates of the white male marital wage differential.

II. LITERATURE REVIEW

Evidence of the marital wage differential for white men in the United States is extensive (see Ribar 2004, for a review). Some of this differential has been attributed to a selection effect, to the fact that men who marry earn more throughout their lives than men who do not marry. Some has been attributed to changes in the wage function that occur at or around the time of marriage (henceforth called the change effect). This change effect has been modeled both as a faster rate of wage growth and as a discrete jump in wages that occurs around the time of marriage. Most of the empirical literature focuses on identifying to what extent the differential is attributable to the selection and to what extent it is attributable to the change effect. Relatively little effort has been expended trying to identify the underlying reasons for these wage patterns. (1)

The most cited explanation offered for the change effect derives from Becker's (1991) work on the theory of the family. Becker noted that when two single-person households are joined to form a single, two-person household, there are gains to be had from specialization and the division of labor as well as economies of scale in household production. Women have historically specialized in home-based activities and men in market-based activities. Time may be thus allocated because of social norms or expectations regarding "acceptable" activities for men and women (see Shelton and John 1996, for a review of theoretical work including the social construction of gender) because of gender differences in comparative advantage or because of gender differentials in labor market opportunities that lead households to rationally allocate more market time to the partner with the highest earnings potential--typically the man. This specialization allows married men to devote more time and energy to market activities, and thus, men's productivity and wages rise following marriage. (2)

This specialization hypothesis relies on the assumption that following marriage men change their behavior. Evidence based on activity reports is mixed. Hersch and Stratton (2000) and South and Spitze (1994) found that men's total reported housework time does not differ by marital status but that married men report more time spent on home maintenance chores than on "routine" or "female type" housework such as cooking and cleaning. Gupta (1999) reported that this is not just a selection effect but that men actually reduce the time they spend on routine or female-type housework when they marry. Hersch and Stratton (2000) estimated wage equations controlling for time spent on housework as well as selection but found that while men who spend more time on housework earn less, controlling for housework time does not change the magnitude or significance of the male marital wage effect.

Few data sets contain information on how time is spent off the job and an alternate approach has been to relate men's behavioral change to spousal activity--the hypothesis being that men whose wives are not employed in the market are more able to specialize than men whose wives are employed. Daniel (1991), Gray (1997), and Chun and Lee (2001) reported that married men whose wives work fewer hours receive higher wages than married men whose wives work longer hours. Their results were robust to fixed effects and/or instrumental variables estimation to control for the possible endogeneity of the wife's employment decision with the husband's wage. Jacobsen and Rayack (1996), however, found a premium that was not robust to fixed effects or instrumental variables estimation, and Loh (1996) found the opposite effect, a positive return for men in dual earner households. In general, wives' working status may be related to men's wages, but the evidence is inconclusive. This may be because men's reported housework time is not sensitive to the employment status of their spouse as reported by Hersch and Stratton (2000) and South and Spitze (1994). It may be that housework is not reallocated in dual career households so much as it is outsourced.

Alternatively there exists some literature using time series data to test the specialization hypothesis. Blackburn and Korenman (1994) tried unsuccessfully to relate time series evidence of a declining gross white male marital wage differential to the increasing labor force participation rate of women in the United Sates. More promising are results by Gray and Vanderhart (2000), who presented evidence that the male marital wage differential declined in the 1970s in those states that instituted a unilateral divorce law. Such laws, by making divorce easier, should make intrahousehold specialization less attractive and so reduce the differential.

Cross-country comparisons have also been made. Evidence, as from Schoeni (1995), that married men in many developed countries receive higher wages than their unmarried counterparts suggests that the differential is not unique to the United States. More recent in-depth studies of the marital wage differential abroad have identified some significant cross-country differences. Given substantial cross-country differences in norms and in the institution of marriage, these studies hold some promise for helping to identify the nature of the differential. Bardasi and Taylor (2004) found evidence that married men earn more than never married men in Great Britain but that none of the differential takes the form of faster postmarriage wage growth and only a small fraction is not attributable to selection. Richardson (2003) used Swedish data and found that wages actually appear to decline with years married in Sweden. She offered no clear explanation for this effect, but Datta Gupta, Smith, and Stratton (2007) attribute a similar result observed using a recent cohort of Danish men to the substantially lower intrahousehold specialization observed in Denmark and in Scandinavia more generally. Thus, the intrahousehold specialization explanation for how marriage might influence the wage function may also receive support by comparing populations that demonstrate different intrahousehold specialization.

Research examining the selection component of the male marital wage differential has been limited at best. This component was first estimated by comparing fixed individual-specific effects estimates with pooled cross-section estimates of the marital wage effect in Korenman and Neumark (1991). An alternative approach has been to use data on twins and assume the selection effect is common to twins independent of their marital status (Krashinsky 2004). There is evidence from Gray (1997) that selection has become increasingly important in the United States and is of substantial importance abroad. Bardasi and Taylor (2004) provide evidence for Great Britain; Datta Gupta, Smith, and Stratton (2007) for Denmark; and Isacsson (2007) for Sweden.

While these articles present substantial evidence of the importance of selection effects, they provide little insight into the nature of the selection effect. Somehow men who marry are different from men who do not marry in a manner that is not otherwise controlled for in the wage equations but is linked to higher productivity on the job. In a rare article addressing the selection component, Krashinsky (2004) reported that ability differentials (and differences in the returns to ability) explain a substantial fraction of the total and the selection component of the U.S. marital wage differential. Other authors (e.g., Bardasi and Taylor 2004) mention attitude, self-esteem, congeniality, loyalty, honesty, dependability, leadership, industriousness, and even appearance. The literature on personality traits and how they influence earnings is a growing one. Bowles, Gintis, and Osborne (2001) provided a justification for including such measures in earnings equations and a partial review of the extant literature. Meuller and Plug (2006) are representative of more recent work in the field. It seems logical to suppose that many personality attributes that are likely to increase earnings may also attract marriage partners. This attraction may be driven by the higher earnings or by the attributes themselves. Thus, in wage equations that fail to control for personality traits but do control for marital status, the marital status dummy may be serving as a proxy for the personality traits.

This issue is complicated by the fact that even today most men marry. Many of the panel data sets used in the analysis of the male marital wage differential include a large fraction of men who marry before they enter the sample. This makes estimation of the marital wage effect dependent on the wages of these men in cross-section analysis. Panel estimates will, however, be dependent only upon those who change marital status during the sample period. If those initially married remain married, they do not contribute to panel-based estimates of the marital wage differential. This identifying condition may explain why panel and cross-section estimates of the marital wage differential are frequently quite different. Furthermore, if more productive men marry first then the years married measure, which is incorporated to capture differential wage growth, may instead be spuriously correlated with wages through its relation to marital timing. Of course, there are also men in any sample who never marry or are always divorced. They represent two additional groups of men whose information does not inform the panel data estimates. Their presence may also help to explain why panel and cross-section estimates differ.

III. METHODOLOGY

In this article, we explore further the nature of the male marital wage differential. We do so first by replicating the analysis that has historically focused on white men with a sample of African American men. The structure of the marital wage gap may differ by race because there are racial differences in time use and marital choices.

For example, unemployment rates are higher and labor force participation rates are lower for African American men as compared to white men. There is also some evidence from Kamo and Cohen (1998) that African American men contribute a significantly greater share of household housework time than white men, even after controlling for household characteristics. If the marital wage differential is the result of increased productivity following marriage that is attributable to increased specialization following marriage then African American men's wages should rise less than white men's wages.

Distinct racial differences in marital patterns have been reported, too, by the Department of Health and Human Resources (2002). African American women are less likely to marry, more likely to divorce, and less likely to remarry than white women. This is the case even though the reported degree of attachment to marriage is similar by race. These racial differences in marital patterns have been attributed to the worse economic situation of African American males, as men who are more educated and employed are more likely to marry. In light of these racial differences in marriage, it will be of some interest to compare the relative selection component of the marital wage differential by race. If marriage among African Americans is more selective (as reflected in their lower marriage rates), there may be a greater selection effect for African American as compared to white men. Or, if the pool of marriage-eligible African American men is small, then there may be a smaller selection effect (Wilson 1987, 83). The higher divorce rate of African Americans as compared to whites is more consistent with there being less selection among African American men.

In general, replicating the analysis for African American men may help identify the nature of the marital wage differential for white men. Second, we explore the mechanism by which selection or marriage generally may change wages. In particular, we look at the possibility that married men may acquire job-related human capital faster than men who are not married. While controls for informal training, such as job experience and/or job tenure, are standard in the marital wage literature, additional controls for formal job training are not. If married men receive more job training than unmarried men with the same job experience and tenure then the observed marital wage differential may be due not to marital status but to training differentials. If this training is received before marriage then it could explain the selection effect--why men who marry earn more. If it is received after marriage then the training differential, not the marriage itself, may explain the faster growth rate of wages following marriage and the fact that the differential is often observed to persist following divorce (Hersch and Stratton 2000).

Why might married men receive more job training? On the one hand, marriage does often entail significantly greater financial responsibilities (particularly following the arrival of children) and may make men more receptive to job training opportunities. Job training may be offered without regard to marital status, but married men may be more likely to accept such opportunities. On the other hand, firms may be more inclined to offer firm-specific training to married men than to not married men. Optimally, both the firm and the recipient share the cost of firm-specific training. This up-front training cost is recouped after the training is completed. The worker then receives a higher wage than he/she could expect to receive at any other firm, while the firm receives a marginal value product of labor that exceeds the wage. Both parties benefit so long as the post-training employment spell endures long enough for each to recoup their portion of the training costs. The danger here is turnover. The greater is expected turnover, the less investment there will be in firm-specific training.

If married men are or are not perceived as being more stable employees as compared to unmarried men then firms may offer married men more training. While perceptions are difficult to measure, Table 1 presents evidence that married white men have higher tenure, lower turnover, and lower quit-to-fire ratios than unmarried white men. Data on tenure by marital status were obtained from the 2000 Tenure Supplement to the Current Population Survey (CPS). These data indicate that white, non-Hispanic men between the ages of 30 and 39 who are married have an average tenure of 6.4 yr, while similar men who have never married (are divorced or separated) have an average tenure of 5.0 (5.4) yr. These mean differences are statistically significant and are not due to the presence of outliers as median tenure is also higher (3) for married men than for never married or divorced/separated men. (4) Regression-based estimates controlling further for age within this sample yield similar highly statistically significant differences.

Data on turnover and quit-to-fire ratios were obtained from the 1989, 1991, and 1993 CPS files, a time period that spans that of the NLSY79 data used in this analysis. These data indicate that only 3.0% of married, white, non-Hispanic men aged 23-37 yr changed jobs while the comparable turnover rate was significantly higher for never married men (6.1%) and for separated/divorced men (6.9%). Employers will not want to invest in training individuals who are likely to leave their firm in short order. Employers have some control over hiring and firing but are less able to control voluntary quits. Thus, we also report the quit-to-fire ratio by marital status. For every 100 married men who are fired or laid off, approximately 15 quit. This ratio is similar for separated and divorced men (14) but significantly higher for never married men (24). Furthermore, Shaw (1987) presents evidence that the quit rate for married men is higher, the higher the earnings share of the wife. Thus, both the marital wage differential and the findings that the marital wage differential is lower for men whose wives work longer hours may be attributable to differences in job training that arise from differences in expected quit rates by marital status rather than to marital status itself. (5)

Next, we explore the selectivity story by examining the role of individual-specific factors. Men who marry may be higher earners throughout their lifetime. It may be simply good fortune that yields some men higher earnings and women may be more attracted to these men as better household providers. But it may be that some individual characteristics are associated with both higher earnings and higher marriage rates. The NLSY data contain information on a wide array of individual characteristics one can explore. We employ controls for cognitive skills, family background, and personality traits/self-esteem in our effort to identify the relevant individual attributes.

Standard wage regressions include controls for education, but there is a large literature, including Ferguson (1995), Neal and Johnson (1996), and Rodgers and Spriggs (2002), demonstrating that other measures of cognitive skills are positively correlated with earnings. If more able men (by these measures) are also more likely to marry then marriage dummies may be proxying for cognitive skills in cross-section analyses. Family background variables such as parental education and maternal employment may influence the degree to which men are attractive marriage partners and the degree to which they are likely to specialize following marriage (Cunningham 2001). These factors may also contribute to otherwise unobserved individual-specific ability or may be indicative of labor market connections that lead to better job matches. Personality traits may also be linked to both earnings and marital outcomes. Specifically, we focus on self-esteem. If the marital wage differential is attributable to selection and we correctly identify the individual-specific factors that link marriage and earnings then when we include these factors in our cross-section wage regressions, we should find the magnitude and significance of the estimated marital wage differential reduced. Indeed, we should find that these cross-section estimates are substantially the same as panel estimates of the male marital wage differential.

However, panel estimates identify the magnitude of the individual-specific component of the marital wage differential differently than cross-section estimates. Fixed-effect results use only information on those who change marital status, whereas cross-section results also compare those who are always married, never married, and always divorced/separated. In our final analyses, we examine the extent to which these individuals, henceforth called "stayers," drive the marital wage differentials observed in cross-sectional analysis to see if differences observed between panel and cross-sectional estimates are driven by differences in the nature of the identification process.

IV. DATA

The data used in this analysis are from the 1979 cohort of the NLSY79. The NLSY79 is a longitudinal data set of 10,000 civilian young adults who have been interviewed annually since 1979. (6) At the survey's beginning, the youth were 14-22 yr. We use data from the 1988 through the 1994 waves when respondents are between the ages of 23 and 37 and when information on formal training is available. The final sample is constructed subject to the following restrictions. Respondents are required to have completed their formal schooling as of the 1988 interview date, to not be in the military or self-employed, and to have provided full information on wages and all other variables. (7) Observations for which constructed hourly earnings are less than $1.00 or greater than $100 in 1982-1984 dollars (using the Consumer Price Index for Urban Consumers [CPI-U] to convert to real dollars) are treated as outliers and are excluded from the sample. After applying these restrictions, our sample contains observations on 2,333 white, non-Hispanic men (henceforth referred to as white men) and 911 African American men. Not all respondents provide complete information annually from 1988 to 1994. Some experience unemployment; others leave the labor force. Pooling the observations across time generates unbalanced panels of 11,581 white male-year observations and 4,642 African American male-year observations.

The data contain information on race, marital status, (8) number of children, education, occupation, industry, union status, and area of residence. Years married is inferred based on observed annual marital status. Data on the unemployment rate in the county of residence are gathered on the expectation that local labor market conditions will influence earnings. The number of employees is reported in recognition of the positive link established between wages and firm size. In addition, each year respondents are asked how many weeks they worked during the previous calendar year. These responses are used to construct a measure of actual work experience. Respondents are also asked how many weeks they have been at their current job, thus enabling the construction of a measure of job tenure.

Detailed firm-provided job training information has been available in the NLSY79 since 1988. At each interview, respondents are asked to describe whether they received any type of formal training since the last interview. (9) Training in this case includes both on-the-job programs such as company or apprenticeship training and off-the-job programs such as classes at a business school, vocational center, or correspondence school. Attendance at professional meetings and seminars on personal finance and lifestyle improvements are further examples of off-the-job training programs. If the respondent has received training then he is asked how long the training lasted. We test several alternative measures of training, reporting the result of a specification that controls for both the incidence of training and the log duration of training, as measured in hours.

Individual-specific attributes that might explain the selection component of the marital wage differential include cognitive skills, family background, and personality traits. The cognitive skills measures we employ are based on the Armed Forces Qualification Test (AFQT) administered to the NLSY79 sample in 1980. Several researchers have used the composite AFQT score to represent skills. (10) This score is assumed to be a direct measure of cognitive skills obtained via family, school, and early labor market experiences. Researchers have found this measure to be a significant determinant of wages; however, it is also influenced by age and education level (Hansen, Heckman, and Mullen 2004; Rodgers and Spriggs 1996). Since the AFQT score was administered to every respondent in the same year, some of the respondents were older and had more education than others when they took the test. To control for both age and education effects, we include not the AFQT score itself, but the residual from a regression of the AFQT score against the age and education level of the respondent at the time the test was administered. We focus on three components of the AFQT test: one representing word knowledge, one representing reading comprehension, and a third representing arithmetic reasoning.

The family background variables for which we control are parental education and occupation. Specifically, we construct a vector of dummy variables to identify those men whose fathers completed at least high school, whose mothers completed at least high school, as well as the occupation of each parent when the respondent was age 14. This approach effectively also controls for the employment status of the mother and hence may capture some information about the degree of specialization within the parental household.

Finally, the NLSY79 also includes some questions regarding personality type--with a particular focus on self-esteem. Waddell (2006) reported evidence that self-esteem measures such as those reported here have a significant impact on earnings and it seems likely that self-esteem may also be related to marriage prospects. There are a total of ten questions (see Table A1, for details) to which respondents are asked to "strongly agree," "agree," "disagree," or "strongly disagree." We test specifications that include three dummy variables for each question to allow for all nonlinearities, but present estimates from a more parsimonious specification in which responses are numbered one to four and summed up in such a way that larger numbers constitute higher self-confidence. Both specifications yield similar results. These questions were asked in both 1982 and 1987. As our wage observations begin in 1988, both sets of measures precede our wage estimates and we perform analysis with both.

Pertinent sample statistics by race and marital status are presented in Table 2. See Table A2, for information on the remaining variables: dummy variables to identify the interview year, the region and city size of residence, the industry and occupation of employment, union status, and public sector employment; continuous variables reflecting the number of children in the household, firm size, and the local county unemployment rate; and more detailed information regarding formal training, cognitive skills, family background, and self-esteem. These statistics are from the pooled cross-section sample.

The variables of particular interest here are the measures of job training, cognitive skills, family background, and self-esteem. The measure of job training reported in Table 2 is the log of hours spent in job training since 1988. As hypothesized, married men have received more training than men who have never married--30%-40% more. If this training raises wages then some of the observed marital wage differential may be attributable to differences in job training. In the empirical specification, cognitive skills are measured using three different components of the AFQT scores. Table 2 reports sample means only for the residual composite AFQT score. Detailed statistics for the three components are substantially similar (Table A2). What matters for this analysis is the value for married as opposed to not married men. Married men have higher residual test scores than all not married men, regardless of race. Again, some of the marital wage differential may be directly attributable to these measures of individual-specific cognitive skills. There is also evidence that family background and self-esteem differ by marital status. The evidence here suggests that separated/divorced men have less educated parents than all other men and that married men have more self-esteem (higher aggregate measures) than all not married men.

Regression analysis allows us to control for all these factors simultaneously. If differences in job experience, job tenure, job training, cognitive skills, family background, or self-esteem explain the marital wage differential then the dummy variable for marital status in a log wage regression that includes controls for all these factors will fall to zero.

V. EMPIRICAL SPECIFICATION

The wage analysis begins with pooled cross-section estimation separately for whites and African Americans of the simple marital wage differential with controls for years married. This specification will identify the gross male marital wage differential and any racial difference in this differential. Continuing to estimate models separately by race, we add the usual wage controls to see what fraction of the gross differential is explained by differences in education, experience, tenure, and occupation separately by race. Then, we control for formal job training to see to what extent these more precise measures of productivity-enhancing activities explain any of the estimated marital wage differential and add individual-specific factors into the cross-section model in hopes of explaining the nature of the selection effect, again separately by race. Next, we estimate a fixed-effects specification to determine the extent to which we have been successful at identifying the selection effect. Finally, we examine the role of "stayers" in driving the cross-section results and contributing to differences between the fixed-effects and the cross-section estimates.

In particular, the initial specification estimates the Gross or unadjusted marital wage differential:

(1) [LnWage.sub.it] = [[beta].sub.0] + [Mar.sub.it] [[beta].sub.1] + [[epsilon].sub.it],

where [Mar.sub.it] is a vector of marital status variables for individual i at time t. This vector has two components: dummy variables identifying those who are married and those who are not currently married but have been separated or divorced and continuous variables that capture years married. Never married men constitute the base case. The coefficient to the dummy variable identifying married men captures the gross wage differential associated with marriage. This differential may capture wage discrimination in favor of married men, productivity shifts following marriage, and/or reflect selectivity differentials between men who marry and men who do not. The coefficient to the measure of years married captures differential postmarriage wage growth. This differential wage growth is typically attributed to the benefits of increased specialization postmarriage that allow married men to spend more time and effort in the market than men who are not married but generally captures increased productivity growth following marriage that could be attributable to training.

The next specification, labeled the Basic specification, adds a vector [X.sub.it] of individual, job, and time-specific characteristics such as are typically incorporated in wage analyses. The third specification, labeled Training, incorporates a set of year-specific indicator variables to identify periods when the respondent was in training and a measure of log hours spent in training since 1986. We tested alternative specifications with linear and quadratic training duration measures, but the log specification provided the best fit. These purely cross-sectional analyses will tell us whether there exist racial differences in the marital wage differential and whether any of the observed cross-sectional differences in earnings, particularly those taking the form of faster wage growth for married men, are explained by differential productivity attributable to job training.

Then, we explore the degree to which the cross-section estimates can be explained by individual-specific, time-invariant factors. We do so first by estimating an Individual-Specific specification that includes such individual-Specific attributes as cognitive skills, family background, and self-esteem. To the extent that these attributes are associated with both marital status and earnings, they should capture that component of the marital wage differential that is attributable to selection, reducing the magnitude and significance of the Mar vector.

A common alternative approach to controlling for selectivity has been to estimate a Fixed-Effects specification, where individual dummy variables are included to capture all individual-specific, time-invariant factors that influence earnings ([[epsilon].sub.it] = [[mu].sub.i] + [[eta].sub.it], where [[mu].sub.i] is the individual-specific factor). We do so next. A comparison of our Fixed-Effect estimates and our Individual-Specific estimates provides us with a standard measure of how successful we have been in controlling for selection.

Estimates from such fixed-effects models have, however, a number of notable weaknesses. First, they are far more sensitive than cross-section estimates to measurement error. While we have eliminated those individuals whose marital history is clearly inconsistent, it is likely that our data are still imperfect, and as a result, our fixed-effects estimates of the marital wage differential will likely be biased toward zero. Second, even if fixed-effects estimates do identify the magnitude of the selection effect, they provide no information as to its nature. Finally, fixed-effects parameters are estimated only off those who change marital status within the sample period. Though "stayers" (those who are always married, never married, or always divorced/separated) within our sample contribute to our cross-section estimates of the marital wage differential, they contribute nothing to the fixed-effects estimates.

To gauge the importance of stayers in driving our cross-section estimates, we estimate a series of cross-section models that include dummy variables identifying these individuals (Stayer Models). Included here are modified Gross, Training, and Individual-Specific specifications, identified with a "+" character in the header to indicate the addition of the stayer dummies. Our goal was to see how including these dummy indicators alters our results.

VI. RESULTS

Selected coefficient estimates for the Gross, Basic, and Training specifications are reported in Table 3. (11) The first three columns present results for the white sample, the last three columns results for the African American sample. The Gross specification for whites indicates that married men earn about 14% (12) more than never married men, perhaps 20% more than separated/divorced men, and that wages rise faster following marriage by almost 1% annually. (13) There are substantial differences by race. The Gross specification for African Americans indicates that the marriage premium is almost 50% larger (at 19%) with no significant difference in the premium over separated/divorced men. The wage growth differential is about the same magnitude as that observed for white men. As do others in the literature, we allow wage growth to revert back to premarriage levels when a couple separates.

Including the basic controls for education; quadratics in actual experience and tenure; a measure of the number of children, the firm size, and the local unemployment rate; and dummies for one-digit industry and occupation, region of residence, city size, interview year, union status, and government employment reduces the observed white male marital wage differential by almost 60%-5.5% and the observed higher growth rate by almost 50%-0.5% annually. Introducing these controls also changes the wage differential between white separated/ divorced and never married men from a significant -6% to an insignificant +2%. These results are comparable to those reported elsewhere in the literature. The effect on African American men is quite similar with the coefficients to the marital dummy and the years married variables both declining by over half. One change for African American men is that after including controls for experience and tenure, there is no evidence of significantly faster wage growth following marriage. For both samples, the three marriage-related variables remain jointly highly statistically significant.

As observed in the sample statistics, married African American and white men have more actual experience and job tenure. Failing to control for these differences has a significant influence on estimates of the differential. While the point estimates of the marital effects in the Basic specification continue to differ by race, with married white men experiencing a higher rate of wage growth but a lower wage differential than their African American counterparts, these racial differences are not statistically significant. The finding that for both white and African American men, there is a significant marital wage differential of approximately the same type suggests that, at least within the United States, the factors contributing to the marital wage differential may be common across the population.

While we have evidence that married men earn more than their unmarried counterparts, however, the precise mechanism by which marriage influences wages still has not been identified. Thus, we turn to our analysis of formal training, including a measure of the log of total training hours as well as controls for participation in current training in the wage model. Current training is expected to lower earnings as training entails time that likely reduces productivity on the job today in return for higher productivity later. Employees are likely to bear some of the cost of this training by receiving lower wages. Our results (available upon request) indicate that this is generally the case except that those who report receiving training during the first-year training data are available experience a positive return to that training in cross-section models, likely reflecting the fact that they received productivity-enhancing training earlier. Accumulated training time, by contrast, should increase productivity and hence wages. Results in columns 3 and 6 of Table 3 confirm this. We find that a 1% increase in training time increases earnings by about 1% for both white and African American men. Controlling for training does not, however, substantially alter the estimated marital wage premium for either population. The parameter estimates of the marital wage differential fall, but the difference is neither significant nor substantial. Thus, while formal training clearly increases wages, it is not the mechanism driving the male marital wage differential.

The results of our analysis of the individual-specific or selection component of the male marital wage differential are presented in Table 4. Again, we report the results of three specifications separately for the white and African American samples. The first of these is the Training specification presented in Table 3. These results are repeated here to make cross-equation comparisons easier.

The second specification is the Individual Specific specification that includes controls for cognitive skills (as measured by three components of the AFQT), parental education and occupation, and own self-esteem that may plausibly be related to both marital status and earnings. Indeed, we find that cognitive skills and parental occupation are positively related to earnings for both white and African American men. Self-esteem is also significantly related to earnings for the white sample. However, incorporating these controls has only a small impact on the estimated marital wage differential. The coefficients to the dummy variables identifying married men fall by 6% for the white sample and 17% for the African American sample but remain statistically significant in both samples. The coefficient to the years married variable in the white sample falls by about 20%, remaining significant only at the 10% level.

We performed several sensitivity tests on this specification. The results reported include self-esteem measures taken from the 1980 NLSY interview. These substantially predate the first reported wage measures from 1988. The identical battery of questions was repeated in 1987. Given the high degree of serial correlation in wages and the likely relation between wages and self-esteem (with high wages increasing self-esteem), we felt that it was important to use personality measures from the earlier period; however, the only difference in the results is that this measure of self-esteem was significant for African Americans. In separate specifications (available upon request), we also included controls for individuals' gender role attitudes. It could be that men who marry perceive their role in the family differently than men who do not marry and that this is somehow correlated with earnings. Thus, we used responses to the questions, "A woman's place is in the home-not in the office or shop," "It is much better for everyone concerned if the man is the achiever outside the home and the woman takes care of the home and family," and "Men should share the work around the house with women, such as doing dishes, cleaning, and so forth," to control for gender role attitudes. The same attitudinal questions were asked in 1979, 1982, and 1987. In no case did we find these variables significantly correlated with earnings. In other specifications, we included measures of additional cognitive skills such as mathematical knowledge, mechanical ability, and science knowledge, but the marital wage effects were unchanged.

The third and sixth columns of Table 4 provide estimates of the male marital wage differential from a Fixed-Effects specification that controls for all individual-specific time-invariant factors. A comparison of the results from this specification with those from the Individual-Specific specification reveals substantial differences by race. The coefficient to the dummy variable identifying currently married individuals declines by more than 90% and becomes insignificant (both statistically and numerically) for white men. The same coefficient for African American men declines by less than 15% and remains significant. The effect of years married changes sign but becomes statistically insignificant for both samples. Since the cross-section estimates of a higher wage growth following marriage could simply be indicative of men with higher earnings marrying earlier in life, we estimated additional models to control for age at first marriage. In no case did we find that age at first marriage was significantly related to earnings. For white men, the marital variables are jointly statistically insignificant once individual-specific effects are incorporated. This suggests that for white men, selection effects may explain the entire differential. (14) Compared to white men, selection effects for African American men are not as important, but even so, the joint marital wage effect is only marginally significant. (15)

Thus, we are left with substantial differences between the panel and the cross-section estimates of the marital wage differential particularly for white men. To reconcile these differences, we examine the models' different identification conditions. Specifically, we concentrate on the well-known fact that fixed-effects estimates are identified only from individuals who change marital status during the survey period, while cross-section estimates are also influenced by the level of wages received by those who do not change marital status. Thus, cross-section estimates may indicate that married men earn higher wages not because wages increase for those who marry, but because the wages of those who are always married are higher than the wages of those who are not married.

Within our panel, over three-quarters of the respondents do not change marital status, so stayers, or nonswitchers, may have a significant impact. Such a high fraction of stayers is not unusual in the marital wage literature. Over three-quarters of the white men in Korenman and Neumark's (1991) seminal work and in Gray's (1997) work, as well as about two-thirds of those in Hersch and Stratton's (2000) sample, were married when first observed in the sample. However, while 6% of each of our samples is always separated/ divorced, the fraction always and never married differs considerably by race. Fully 45% of white men are always married as compared to 22% never married. These fractions are essentially reversed for African American men, with 21% always and 44% never married.

To gauge the degree to which the differences between the cross-section and the panel results are driven by these stayers, we re-estimate our Gross, Training, and Individual-Specific models including dummy variables to identify stayers of each type. These specifications remove the average contribution of stayers (or stayer-specific means) from the marital wage differentials. Movers, however, still contribute to the estimates in two ways: through wage changes experienced as marital status changes and through wage levels. Fixed-effects estimates further remove all wage-level effects and so will not quite be comparable. The expanded cross-section results are reported in Table 5. The base case to which all individuals are compared in these specifications is no longer all observations predating a first marriage, but rather only observations predating a first marriage for those who are observed marrying within the sample.

In the Gross+ estimates, the coefficient to the marital dummy falls substantially and becomes statistically insignificant for both whites and African Americans, suggesting that stayers play a major role. The impact of separation or divorce is, however, quite different by race. For white men, wages fall following divorce, but the difference is not statistically significant. Instead, it is white men who are always divorced who receive substantially and significantly lower wages. (16) By contrast, for African American men, being always divorced is associated with an insignificant positive wage effect, while wages fall significantly following divorce. Interestingly, the inclusion of the stayer controls also reduces the magnitude of the estimated wage growth differential for married men of both races.

As more controls are added, the magnitude and significance of many of the coefficients to the marital controls decline. For white men, the joint impact of both current and persistent marital status remains significant at even the 0.1% level, but a joint test of the significance of only the current marital status variables indicates that these are not significant at even the 10% level. Overall, controlling for the wage level of stayers appears to explain most of the difference between the individual-specific and the fixed-effects results for white men.

For African American men, there was less of a difference to explain, as the fixed-effects specification still showed a significant marital wage differential. The marital status variables do become individually statistically insignificant with the addition of the stayer variables, but joint tests indicate that the current status variables are more significant than the stayer status variables. The two marital variables that are most statistically significant in the expanded specification are those related to separation/divorce, and as mentioned above, these coefficient values indicate that it is African Americans who separate, not those who are always separated, who experience lower wages. A further analysis of the average wages of this group (results not shown) indicates that the wages of these men are lower no matter their marital status. Some low-wage African American men are marrying but unable to maintain that status. The low-wage level of these men likely depresses cross-section relative to panel estimates of the marital wage differential.

Overall, controlling for stayers and their demographic characteristics goes a long way toward reconciling the cross-section and panel estimates of the marital wage differential for white men, as the level effects across movers within the cross-section estimates do not appear to contribute substantially to the estimated marital wage differentials. By contrast, controlling for stayers aggravates the level effects across movers for African American men.

VII. CONCLUSIONS

In this analysis, we explore the nature of the white male marital wage differential. Evidence suggests that the differential is split between a wage change component possibly driven by differential productivity and a selection component, with selection becoming increasingly important for more recent cohorts in the United States. The wage change component has often been attributed to increased job skills acquired as a result of intrahousehold specialization, but explanations for the selection component have received little empirical attention. Our purpose here was to focus further attention upon the mechanisms underlying the observed differential.

We begin by extending what has traditionally been an analysis of white men in the United States to a sample of African American men. In one of the first analyses of African American marital wage differentials to control for job experience, job tenure, and years married, (17) we find little difference by race. The gross marital wage differential is larger for African American men than for white men, but much of this differential is explained by differences in other observable characteristics (like experience) that are typically included in a wage equation. Cross-section estimates of a standard wage equation suggest that African American men may experience a larger wage jump but less wage growth following marriage than white men, but the differences are not statistically significant. This result suggests that some factors common to both white and African American married men are driving the observed male marital wage differential.

One possibility is that marriage is associated with greater job training for men of all races. We use detailed direct information on job training to explore this possibility and find that married men receive more informal and formal job training than men who are not married and that this training does significantly increase wages. However, controlling for training does not substantially alter the estimated marital wage differential for either the white or the African American sample. Given our use of more direct training measures, this finding casts some doubt on the hypothesis that the marital wage differential is attributable to a productivity differential.

Next, we examine the selection component of the male marital wage differential. We do so first by using some of the rich data available in the NLSY on cognitive skills, parental background, and personality traits all of which have previously been linked to higher productivity and could credibly be associated with marital status. While cognitive skills, parental occupation, and self-esteem were significantly associated with earnings, none of these variables appear to explain more than 15% of the selection component of the male marital wage differential. By contrast, fixed-effects estimates that difference out all individual-specific and time-invariant effects reveal that the vast majority of the differential for white men (as reported elsewhere in the literature), but relatively little of the differential for African American men, is attributable to individual-specific, time-invariant factors. These results suggest that selection into marriage is critical for white but not for African American men and that marriage actually has an impact upon wages for African American men. The small selection effect for African American men is consistent with a story that the pool of marriage-eligible African American men is small and the fact that marital disruption is higher for African Americans than for whites.

To reconcile the differences between our expanded cross-section model and the fixed-effects models, particularly for white men, we examine the identification conditions underlying these two approaches. Fixed-effects estimates of the marital wage differential are identified only off those individuals who are observed changing marital status within the sample. By controlling separately for these stayers in our cross-section models, we are, in fact, able to substantially reconcile our results. Much of the difference between the cross-section and the fixed-effects estimates of the marital status effects for white men is attributable to those men who never marry, who are always divorced, or who are always married within the sample, explaining the fact that fixed-effects estimates for white men show no evidence of a marital wage differential. By contrast, these stayers explain little of the robust marital wage effect found for African American men, and, indeed, controlling for African American stayers highlights the relative importance of men who separate within this sample who always receive lower wages. This finding, too, is consistent with the notion that the marriage-eligible pool of African Americans is smaller and hence that more "mistakes" are made in the marriage market.

Future work examining the male marital wage differential should be cognizant of these cross-section/panel differentials. Different racial and ethnic samples should be further studied to capture differences in marriage markets. Additional analysis of personality traits should be explored to develop additional insight into the nature of the individual-specific component of the marital wage differential. For example, men who marry may have a more agreeable temperament than men who do not marry. More agreeable men may be easier to work with on the job and hence be more productive and earn higher wages, particularly given the shift from a manufacturing to a service-based economy, where a variety of "soft" skills such as teamwork have been found to be important. More agreeable men also may be better marriage partners. More work to identify how marriage affects men and women's time use may help explain differences in wage growth following marriage.

ABBREVIATIONS

AFQT: Armed Forces Qualification Test

NLSY79: National Longitudinal Survey of Youth 1979

doi: 10.1111/j.1465-7295.2008.00209.x

APPENDIX

TABLE A1
Questions Regarding Personality Traits

WORTH I feel that I'm a person of worth, at least on an
 equal basis with others
GOODQUAL I feel that I have a number of good qualities
FAILURE All in all, I am inclined to feel that I am a failure
CAPABLE I am able to do things as well as most other people
PROUD I feel I do not have much to be proud of
POSATT I take a positive attitude toward myself
SATMYSELF On the whole, I am satisfied with myself
SELFRESPECT I wish I could have more respect for myself
USELESS I certainly feel useless at times
NOGOOD At times I think I am no good at all

TABLE A2
Sample Means by Race and Marital Status

 White Men

 Never Separated/
Variable Married Married Divorced

Never married 0.00 0.78 0.00
Always married as observed 0.75 0.00 0.00
Always divorced 0.00 0.00 0.42
Log of firm size (number of
 employees) 4.05 3.72 3.60
County unemployment rate x 10 68.44 66.06 67.77
Number of children 1.45 0.10 0.32
Regional dummies
 Midwest 0.30 0.29 0.24
 South 0.34 0.27 0.42
 West 0.20 0.26 0.22
City/metro dummies
 SMSA, not central city 0.38 0.36 0.32
 SMSA, central city unknown 0.27 0.31 0.33
 SMSA, central city 0.08 0.15 0.09
Job characteristics
 Union member 0.15 0.11 0.11
 Employed in public sector 0.15 0.10 0.15
 Industry
 Agriculture, forestry, and
 fisheries 0.03 0.04 0.05
 Mining 0.02 0.01 0.01
 Construction 0.11 0.13 0.17
 Manufacturing 0.29 0.22 0.24
 Transportation, communication,
 public utilities 0.09 0.07 0.08
 Wholesale and retail trade 0.17 0.24 0.19
 Finance, insurance, and real
 estate 0.04 0.06 0.03
 Business and repair services 0.07 0.08 0.08
 Personal services 0.01 0.04 0.01
 Entertainment and recreation
 services 0.01 0.02 0.01
 Professional and related services 0.09 0.06 0.07
 Public administration 0.06 0.04 0.04
 Occupation
 Professional, technical 0.14 0.12 0.08
 Managers 0.16 0.12 0.09
 Sales workers 0.06 0.06 0.03
 Clerical workers 0.06 0.09 0.05
 Craftsmen and foremen 0.22 0.21 0.29
 Operatives 0.20 0.17 0.20
 Laborers, except farm 0.07 0.10 0.12
 Farmers and farm managers 0.00 0.00 0.00
 Farm laborers and foremen 0.01 0.02 0.01
 Service workers 0.07 0.12 0.11
 Private household workers 0.00 0.00 0.00
Year dummies
 1988 0.15 0.23 0.15
 1989 0.16 0.20 0.16
 1990 0.16 0.17 0.17
 1991 0.13 0.12 0.13
 1992 0.13 0.11 0.13
 1993 0.13 0.09 0.13
 1994 0.13 0.08 0.14
Training variables
 Received training in 1989 0.01 0.01 0.01
 Received training in 1990 0.00 0.00 0.00
 Received training in 1991 0.01 0.01 0.02
 Received training in 1992 0.01 0.00 0.00
 Received training in 1993 0.04 0.02 0.01
 Received training in 1994 0.01 0.01 0.01
 Time in training (in h) 62.30 71.12 49.67
 Ever received training 0.18 0.15 0.12
AFQT variables
 Word knowledge 1.74 2.14 1.70
 Paragraph comprehension 0.45 0.34 0.43
 Arithmetic reasoning 2.76 1.99 1.83
Family background variables
 Father's education 10.41 10.72 9.90
 Father's education missing 0.07 0.08 0.09
 Mother's education 10.45 10.46 10.23
 Mother's education missing 0.06 0.06 0.06
 Father's occupation
 Missing/not employed 0.20 0.22 0.24
 Professional, technical 0.11 0.09 0.05
 Managers 0.12 0.12 0.08
 Sales workers 0.04 0.05 0.05
 Clerical workers 0.04 0.03 0.05
 Craftsmen and foremen 0.20 0.18 0.21
 Operatives 0.01 0.01 0.01
 Laborers, except farm 0.16 0.17 0.14
 Farmers and farm managers 0.03 0.04 0.05
 Farm laborers and foremen 0.03 0.01 0.02
 Service workers 0.02 0.02 0.03
 Private household workers 0.05 0.05 0.07
 Mother's occupation
 Missing/not employed 0.43 0.46 0.43
 Professional, technical 0.08 0.09 0.09
 Managers 0.04 0.02 0.04
 Sales workers 0.03 0.03 0.04
 Clerical workers 0.16 0.16 0.13
 Craftsmen and foremen 0.01 0.02 0.01
 Operatives 0.00 0.00 0.00
 Laborers, except farm 0.11 0.09 0.10
 Farmers and farm managers 0.01 0.01 0.00
 Farm laborers and foremen 0.00 0.00 0.00
 Service workers 0.02 0.01 0.01
 Private household workers 0.11 0.12 0.14
Attitude (a) 6.95 6.92 7.07
Self-esteem measures
 Summary measure from 1980 22.63 21.90 22.01
 Summary measure from 1987 24.10 23.24 23.00
Sample size 6,959 3,209 1,413

 African American Men

 Never Separated/
Variable Married Married Divorced

Never married 0.00 0.88 0.00
Always married as observed 0.58 0.00 0.00
Always divorced 0.00 0.00 0.47
Log of firm size (number of
 employees) 4.65 4.01 4.24
County unemployment rate x 10 61.51 61.23 61.22
Number of children 1.67 0.26 0.32
Regional dummies
 Midwest 0.15 0.18 0.16
 South 0.68 0.59 0.65
 West 0.07 0.07 0.08
City/metro dummies
 SMSA, not central city 0.25 0.18 0.18
 SMSA, central city unknown 0.36 0.33 0.39
 SMSA, central city 0.18 0.28 0.25
Job characteristics
 Union member 0.22 0.16 0.15
 Employed in public sector 0.25 0.17 0.15
 Industry
 Agriculture, forestry, and
 fisheries 0.01 0.04 0.01
 Mining 0.01 0.00 0.01
 Construction 0.12 0.14 0.13
 Manufacturing 0.26 0.19 0.21
 Transportation, communication,
 public utilities 0.13 0.08 0.12
 Wholesale and retail trade 0.15 0.21 0.19
 Finance, insurance, and real
 estate 0.03 0.03 0.04
 Business and repair services 0.07 0.09 0.12
 Personal services 0.02 0.04 0.04
 Entertainment and recreation
 services 0.01 0.01 0.00
 Professional and related services 0.09 0.11 0.07
 Public administration 0.10 0.06 0.05
 Occupation
 Professional, technical 0.09 0.06 0.05
 Managers 0.09 0.05 0.06
 Sales workers 0.02 0.02 0.03
 Clerical workers 0.10 0.10 0.06
 Craftsmen and foremen 0.16 0.15 0.13
 Operatives 0.27 0.21 0.27
 Laborers, except farm 0.12 0.19 0.17
 Farmers and farm managers 0.00 0.00 0.00
 Farm laborers and foremen 0.00 0.01 0.00
 Service workers 0.15 0.21 0.22
 Private household workers 0.00 0.00 0.00
Year dummies
 1988 0.14 0.20 0.12
 1989 0.14 0.18 0.14
 1990 0.14 0.16 0.15
 1991 0.15 0.14 0.15
 1992 0.15 0.13 0.14
 1993 0.14 0.11 0.15
 1994 0.15 0.09 0.14
Training variables
 Received training in 1989 0.01 0.01 0.00
 Received training in 1990 0.01 0.01 0.00
 Received training in 1991 0.01 0.01 0.03
 Received training in 1992 0.01 0.00 0.00
 Received training in 1993 0.02 0.00 0.01
 Received training in 1994 0.01 0.01 0.01
 Time in training (in h) 71.47 60.31 82.78
 Ever received training 0.15 0.10 0.11
AFQT variables
 Word knowledge -5.18 -5.66 -6.03
 Paragraph comprehension -2.06 -2.34 -2.46
 Arithmetic reasoning -3.44 -3.70 -3.94
Family background variables
 Father's education 7.42 7.13 6.96
 Father's education missing 0.26 0.29 0.28
 Mother's education 9.85 9.71 9.28
 Mother's education missing 0.09 0.10 0.11
 Father's occupation
 Missing/not employed 0.37 0.46 0.41
 Professional, technical 0.02 0.01 0.01
 Managers 0.02 0.03 0.03
 Sales workers 0.01 0.00 0.00
 Clerical workers 0.02 0.03 0.03
 Craftsmen and foremen 0.19 0.13 0.13
 Operatives 0.02 0.01 0.01
 Laborers, except farm 0.18 0.18 0.18
 Farmers and farm managers 0.09 0.07 0.11
 Farm laborers and foremen 0.01 0.01 0.01
 Service workers 0.01 0.01 0.01
 Private household workers 0.07 0.05 0.07
 Mother's occupation
 Missing/not employed 0.41 0.44 0.43
 Professional, technical 0.14 0.14 0.10
 Managers 0.01 0.01 0.02
 Sales workers 0.00 0.01 0.01
 Clerical workers 0.09 0.09 0.10
 Craftsmen and foremen 0.02 0.01 0.02
 Operatives 0.00 0.00 0.00
 Laborers, except farm 0.14 0.11 0.10
 Farmers and farm managers 0.00 0.00 0.00
 Farm laborers and foremen 0.00 0.00 0.00
 Service workers 0.00 0.02 0.01
 Private household workers 0.18 0.17 0.22
Attitude (a) 6.91 6.88 7.17
Self-esteem measures
 Summary measure from 1980 22.94 21.70 21.96
 Summary measure from 1987 23.77 22.93 22.73
Sample size 1,682 2,338 622

SMSA, standard metropolitan statistical area.

(a) Attitude is calculated off the 1979 responses to 3 questions
when available, else the 1982 responses, else the 1987 responses.

REFERENCES

Bardasi, E., and M. Taylor. "Investigating the Marriage Wage Premium for Men in Britain." Paper presented at ESPE meetings, Bergen, Norway, 2004.

Becker, G. S. A Treatise on the Family. Cambridge, MA: Harvard University Press, 1991.

Blackburn, M., and S. Korenman. "The Declining Marital-Status Earnings Differential." Journal of Population Economics, 7, 1994, 247-70.

Bowles, S., H. Gintis, and M. Osborne. "The Determinants of Earnings: A Behavioral Approach." Journal of Economic Literature, 39, 2001, 1137-76.

Chun, H., and I. Lee. "Why Do Married Men Earn More: Productivity or Marriage Selection?" Economic Inquiry, 39, 2001, 307-19.

Cunningham, M. "Parental Influences on the Gendered Division of Labor." American Sociological Review, 66, 2001, 184-203.

Daniel, K. "Does Marriage Make Men More Productive?" Unpublished manuscript, University of Chicago, 1991.

Datta Gupta, N., N. Smith, and L. S. Stratton. "Is Marriage Poisonous? Are Relationships Taxing? An Analysis of the Male Marital Wage Differential in Denmark." Southern Economic Journal, 74, 2007, 412-33.

Department of Health and Human Resources. "Cohabitation, Marriage, Divorce, and Remarriage in the United States." Vital and Health Statistics, Series 23, No. 22, 2002.

Ferguson, R. F. "Shifting Challenges: Fifty Years of Economic Change Toward Black-White Earnings Equality." Daedalus, 124(1), 1995, 37-76.

Gray, J. S. "The Fall in Men's Return to Marriage: Declining Productivity Effects or Changing Selection?" Journal of Human Resources, 32, 1997, 481-504.

Gray, J. S., and M. J. Vanderhart. "On the Determination of Wages: Does Marriage Matter?" in The Ties That Bind. Perspectives on Marriage and Cohabitation, edited by L. J. Waite. New York: Aldine de Gruyter, 2000, 356-67.

Gupta, S., "The Effects of Transitions in Marital Status on Men's Performance of Housework." Journal of Marriage and the Family, 61, 1999, 700-11.

Hansen, K., J. J. Heckman, and K. J. Mullen. "The Effect of Schooling and Ability on Achievement Test Scores." Journal of Econometrics, 121, 2004, 39-98.

Hersch, J., and L. S. Stratton. "Household Specialization and the Male Marriage Wage Premium." Industrial and Labor Relations Review, 54, 2000, 78-94.

Hill, M. S. "The Wage Effects of Marital Status and Children." Journal of Human Resources, 14, 1979, 579-94.

Isacsson, G. "Twin Data vs. Longitudinal Data to Control for Unobserved Variables in Earnings Functions Which Are the Differences?" Oxford Bulletin of Economics and Statistics, 69, 2007, 339-62.

Jacobsen, J. P., and W. L. Rayack. "Do Men Whose Wives Work Really Earn Less?" American Economic Review, 86, 1996, 268-73.

Kamo, Y., and E. L. Cohen. "Division of Household Work between Partners: A Comparison of Black and White Couples." Journal of Comparative Family Studies, 29(1), 1998, 131-5-45.

Korenman, S., and D. Neumark. "Does Marriage Really Make Men More Productive?" Journal of Human Resources, 26, 1991, 282-307.

Krashinsky, H. A. "Do Marital Status and Computer Usage Really Change the Wage Structure?" Journal of Human Resources, 39, 2004, 774-91.

Loh, E. S. "Productivity Differences and the Marriage Wage Premium for White Males." Journal of Human Resources, 31, 1996, 566-89.

Mamun, A. "Is There a Cohabitation Premium in Men's Earnings?" Working Paper No. 2004-02, Center for Research on Families, University of Washington, 2004.

Meuller, G., and E. Plug. "Estimating the Effect of Personality on Male and Female Earnings." Industrial and Labor Relations Review, 60, 2006, 3-22.

Neal, D. A., and W. R. Johnson. "The Role of Pre-Market Factors in Black-White Wage Differences." Journal of Political Economy, 104, 1996, 869-95.

Reed, R. W., and K. Harford. "The Marriage Premium and Compensating Wage Differentials." Journal of Population Economies, 2, 1989, 237-65.

Ribar, D. "What Do Social Scientists Know About the Benefits of Marriage? A Review of Quantitative Methodologies." IZA Discussion Paper No. 998, 2004.

Richardson, K. "The Evolution of the Marriage Premium in the Swedish Labor Market 1968-1991." Unpublished manuscript, Swedish Institute for Social Research, Stockholm University, 2003.

Rodgers, W. M. Ill, and W. E. Spriggs. "What Does the AFQT Really Measure: Race, Wages, Schooling and the AFQT Score." Review of Black Political Economy, 24(4), 1996, 13-46.

--. "Accounting for the Racial Gap in AFQT Scores: Comment on Nan L. Maxwell, 'The Effect on Black-White Wage Differences of Differences in the Quality and Quantity of Education'." Industrial and Labor Relations Review, 55, 2002, 533-41.

Schoeni, R. F. "Marital Status and Earnings in Developed Countries." Journal of Population Economics, 8, 1995, 351-59.

Shaw, K. L. "The Quit Propensity of Married Men." Journal of Labor Economics, 5, Part 1, 1987, 533-60.

Shelton, B. A., and D. John. "The Division of Household Labor." Annual Review of Sociology, 22, 1996, 299-322.

South, S. J., and G. Spitze. "Housework in Marital and Nonmarital Households." American Sociological Review, 59, 1994, 327-47.

Stratton, L. S. "Examining the Wage Differential for Married and Cohabiting Men." Economic Inquiry, 40, 2002, 199-212.

Waddell, G. R. "Labor-market Consequences of Poor Attitude and Low Self-esteem in Youth." Economic Inquiry, 44, 2006, 69-97.

Wilson, C. M., and A. J. Oswald. "How Does Marriage Affect Physical and Psychological Health? A Survey of the Longitudinal Evidence." IZA Discussion Paper No. 1619, 2005.

WILLIAM M. RODGERS III and LESLIE S. STRATFON *

* We are grateful to seminar participants at the Danish National Institute of Social Research, Arhus University, Uppsala University, Amsterdam's Institute for Labor Studies, IZA, and the workshop on Family Decisions and Family Policy held in Arhus, Denmark, as well as to Jacobsen and referees for valuable comments. Responsibility for the contents rests solely with the authors.

Rodgers III: Professor and Chief Economist, John J. Heldrich Center for Workforce Development, Edward J. Bloustein School of Planning and Public Policy, Rutgers, The State University of New Jersey, 33 Livingston Avenue 5th Floor, New Brunswick, NJ 08901. Phone 1-732-932-4100, ext 783, Fax 1-732-932-3454, E-mail [email protected]

Stratton: Associate Professor, Department of Economics, Virginia Commonwealth University, 301 West Main Street, PO Box 844000, Richmond, VA 23284-4000. Phone 1-804-828-7141, Fax 1-804-828-9103, E-mail [email protected]; IZA, Bonn, Germany; and CIM, Aarhus, Denmark

(1.) Wilson and Oswald (2005) report that the same is true for the literature relating marital status to health.

(2.) There are a number of other explanations for a male marital wage differential. For example, married men may choose different types of jobs, jobs that offer a compensating wage differential, as compared with unmarried men (Reed and Harford 1989), the wife may positively affect the husband's productivity by serving as a sounding board or other job aid (Loh 1996), or employers may simply discriminate in favor of married men (Hill 1979).

(3.) The test is conducted in Stata using a binomial method for calculating confidence intervals that imposes no distributional assumptions upon the underlying variable.

(4.) Similar differences are evident for blacks and Hispanics, though only the difference between married and never married men is statistically significant for nonwhites.

(5.) The relation may be considerably more complicated. Married men may quit less because of the higher wage they receive being married. This lower expected quit rate can make them more suitable candidates for training and so further increase their wages. Our regression analysis will allow us to differentiate between such training-and marriage-driven effects.

(6.) We exclude the NLSY's original military sample as there is an explicit marriage premium in the military. Doing this reduces the sample by 1,280. Along with a stratified random sample, the NLSY over-samples low-income whites, blacks, and Hispanics. In 1990, the low-income whites were dropped from the sample. Respondents used in this study are from the stratified random sample or the low-income black sample.

(7.) A small number of men who were widowed are also excluded from the sample on the grounds that such a change in marital status is, at this age, typically unforeseen. Also excluded are a small number of men whose marital history is incomplete or inconsistent.

(8.) We focus only on marriage, not on cohabitation in this analysis. The evidence regarding wage differentials for cohabiting men in the United States is mixed at best (Stratton 2002, e.g., finds a cross-sectional but not a panel effect) and cohabitation itself is difficult to identify.

(9.) The interval in the 1988 interview covers 1986-1988. The intervals for the 1989-1994 interviews cover the preceding year.

(10.) See, for example, Rodgers and Spriggs (1996, 2002), Neal and Johnson (1996), and Ferguson (1995).

(11.) Full results are available from the authors upon request. All standard errors are robust and corrected for clustering at the individual level.

(12.) All returns are calculated using the conversion exp([beta])--1, where [beta] is the coefficient estimate.

(13.) In other specifications not reported here, we included a quadratic in years married, but this variable was not statistically significant at even the 10% level.

(14.) Following Krashinsky (2004), we compared the wages of men who married in a given year with the wages of men who were not married in that year. The wages of the men who married were higher than the wages of the men who did not marry even several years before the marriage date, confirming that these results are indeed likely attributable to selection.

(15.) This racial difference is all the more striking as it does not appear to be driven by the higher standard errors inherent in fixed-effects models that have so many fewer degrees of freedom. The effect remains significant for the smaller, not the larger, sample and is driven at least as much by changes in the point estimate as by changes in the standard errors. Further, while fixed-effects estimates are more sensitive to measurement errors, there is no reason to suppose that these are larger for the white sample than for the African American sample.

(16.) Those for whom always divorced equals one also have a value of one for separated/divorced. In general, the stayer variables identify marginal effects above and beyond the marital status effects.

(17.) Mamun (2004) controls for age rather than job experience or tenure and continues to find a faster rate of wage growth following marriage for African American men. It is well known that age is a poor proxy for experience. As shown here, unmarried men have less experience and tenure than married men of the same age. In addition, Mamun does not include the rich set of individual-specific controls included here. When we mimic Mamun's specification, we also find a faster rate of wage growth for married African American men.

TABLE 1
Data on Tenure, Turnover Rates, and Quit-to-Fire Ratios by
Marital Status

 Marital Status

 Separated,
White, Non-Hispanic Men Married Never Married Divorced

Mean tenure (b) 6.37 5.03 5.44
Median tenure 5.0 4.0 4.0
Turnover rate (c) 2.98% 6.07% 6.89%
Quit-to-fire ratio (c) 15.3 23.6 14.3

 Marital Status

 Statistical
White, Non-Hispanic Men Significance (a)

Mean tenure (b) a, b
Median tenure a, b
Turnover rate (c) a, b, c
Quit-to-fire ratio (c) a, c

(a) The lowercase online letters "a," "b," and "c" indicate that
married and never married means are significantly dif ferent at
the 5% level, married and separated/divorced means are
significantly different at the 5% level, and never married and
separated/divorced means are significantly different at the 5'1,,
level, respectively.

(b) Calculated using CPS 2000 Tenure Supplement for white,
non-Hispanic men aged 30-39 yr. Measured in years.

(c) Calculated using the data from the January, May, and
September 1989, 1991, and 1993 CPSs for white, non-Hispanic men
aged 23-37 yr who were not enrolled in school; married, spouse
absent; widowed; or in the military. Sample consists only of
employed and unemployed persons, excluding those between
temporary jobs. Turnover rate is calculated as percent who lost,
quit, or were laid off from theirjob. Quit-to-fire ratio is
calculated as 100 x number of quit job/(number of lost + number
of laid off).

TABLE 2
Sample Means by Race and Marital Status

 White Men

 Never Separated
Variable Married Married Divorced

Log hourly wage 2.11 1.92 1.90
Wage (1982-1984$) 9.34 7.95 7.58
Education 12.82 12.95 11.87
Age 30.03 28.53 29.94
Experience (in wk) 521.53 422.54 490.97
Tenure (in wk) 241.81 156.87 170.96
Years married 6.70 0.00 4.10
Log of training time (in h) 1.36 1.05 1.14
Residual composite
AFQT score 7.61 6.27 4.26
Father's education (a) 11.23 11.70 10.92
Mother's education (a) 11.07 11.14 10.86
Mother's occupation missing 0.43 0.46 0.43
Self-esteem (b) 22.63 21.90 22.01
Number of observations 6,959 3,209 1,413

 African American Men

 Never Separated
Variable Married Married Divorced

Log hourly wage 1.92 1.69 1.73
Wage (1982-1984$) 7.52 6.15 6.35
Education 12.70 12.21 12.02
Age 29.97 28.84 30.23
Experience (in wk) 480.90 371.20 461.00
Tenure (in wk) 209.03 129.71 126.90
Years married 5.62 0.00 3.32
Log of training time (in h) 1.41 0.95 1.36
Residual composite
AFQT score -16.88 -17.29 -20.73
Father's education (a) 9.98 10.09 9.64
Mother's education (a) 10.80 10.84 10.45
Mother's occupation missing 0.41 0.44 0.43
Self-esteem (b) 22.94 21.70 21.96
Number of observations 1,682 2,338 622

Note: Sample source--1988-1994 pooled cross sections of employed
men from the NLSY, who have completed school and are not in the
military or self-employed.

(a) Sample restricted to those with nonmissing values. Less than
10% of any white sample reported missing parental education but
almost 30% of separated/divorced African American men fail to
report their father's education level. Missing values are
identified with a dummy variable.

(b) The sum of ten categorical variables, each with values
running from 1 to 4, designed to measure self-esteem in 1980.
Variables are redefined to ensure that values indicate less
self-esteem.

TABLE 3
Estimate of the Marital Log Wage Differential Racial Differences
and Training Effects

 White Men

 Basic (a)
Variable Gross Specification Specification

Married 0.1273 *** (0.0254) 0.0540 *** (0.0188)
Separated/divorced -0.0579 ** (0.0291) 0.0188 (0.0203)
Years married 0.0097 *** (0.0028) 0.0052 ** (0.0026)
Log of training time
F test on marital status 46.92 *** 7.96 ***
Number of observations 11,581 11,581
[R.sup.2] 0.0396 0.4350

 White Men African American Men

 Training (b)
Variable Specification Gross Specification

Married 0.0503 *** (0.0187) 0.1781 *** (0.0363)
Separated/divorced 0.0170 (0.0202) 0.0107 (0.0399)
Years married 0.0051 ** (0.0026) 0.0095 * (0.0051)
Log of training time 0.0128 *** (0.0033)
F test on marital status 7.49 *** 27.81 ***
Number of observations 11,581 4,642
[R.sup.2] 0.4401 0.0550

 African American Men

 Basic (a) Training (b)
Variable Specification Specification

Married 0.0790 *** (0.0249) 0.0782 *** (0.0247)
Separated/divorced 0.0076 (0.0280) 0.0048 (0.0276)
Years married 0.0035 (0.0041) 0.0025 (0.0040)
Log of training time 0.0119 ***
F test on marital status 7.95 *** 7.41 ***
Number of observations 4,642 4,642
[R.sup.2] 0.4456 0.4520

Note: Robust standard errors corrected also of clustering by
individual are reported in parentheses below coefficient
estimates.

(a) This specification includes controls for years of education;
quadratics in experience and tenure; dummy variables to identify
11 industries, 10 occupations, 3 regions, 3 city sizes, and 5
interview years; a measure of the number of children, firm size,
and the local unemployment rate; and dummy variables to identify
union members and government employees.

(b) This specification includes all the controls listed in the
footnote "a" as well as dummy variables to indicate receipt of
training in the particular interview year.

*** indicates statistical significance at the 1% level,
two-sided test; ** indicates statistical significance at the 5%
level, two-sided test; and * indicates statistical significance at
the 10% level, two-sided test.

TABLE 4
Estimates of the Marital Log Wage Differential
Individual-Specific Factors

 White Men

 Training Individual-Specific
Variable Specifications Specification

Married 0.0503 *** (0.0187) 0.0472 *** (0.0182)
Separated/divorced 0.0170 (0.0202) 0.0130 (0.0198)
Years married 0.0051 *** (0.0026) 0.0043 * (0.0025)
Log of training time 0.0128 *** (0.0033) 0.0108 *** (0.0033)
Including
 3 AFQT residual scores Yes ***
 Parents' education Yes
 Parents' occupation Yes *
 Self-esteem Yes ***
F test on marital status 7.49 *** 6.3 ***
Number of observations 11,581 11,581
Number of individuals 2,333 2,333
[R.sup.2] 0.4401 0.4588

 White Men African American Men

 Fixed-Effects Training
Variable Specification Specification

Married 0.0046 (0.0155) 0.0782 *** (0.0247)
Separated/divorced -0.0186 (0.0220) 0.0048 (0.0276)
Years married -0.0056 (0.0041) 0.0025 (0.0040)
Log of training time 0.0137 *** (0.0038) 0.0119 *** (0.0042)
Including
 3 AFQT residual scores
 Parents' education
 Parents' occupation
 Self-esteem
F test on marital status 1.18 7.41 ***
Number of observations 11,581 4,642
Number of individuals 2,333 911
[R.sup.2] 0.8100 0.4520

 African American Men

 Individual-Specific Fixed-Effects
Variable Specification Specification

Married 0.0649 *** (0.0248) 0.0573 ** (0.0285)
Separated/divorced -0.0016 (0.0276) 0.0484 (0.0358)
Years married 0.0034 (0.0039) -0.0090 (0.0061)
Log of training time 0.0110 *** (0.0040) 0.0028 (0.0062)
Including
 3 AFQT residual scores Yes ***
 Parents' education Yes
 Parents' occupation Yes ***
 Self-esteem Yes
F test on marital status 6.19 *** 1.89 *
Number of observations 4,642 4,642
Number of individuals 911 911
[R.sup.2] 0.4727 0.7295

Notes: Robust standard errors corrected also for clustering by
individual are reported in parentheses below coefficient
estimates. All specifications include control for years of
education; quadratics in experience and tenure; dummy variable to
identify 11 industries, 10 occupations, 3 regions, 3 city sizes,
and 5 interview years; a measure of the number of children, firm
size, and the local unemployment rate; dummy variables to
identify union members and government employees; and dummy
variables to indicate receipt of training in the particular
interview year.

*** indicates statistical significance at the 1% level, two-sided
test; ** indicates statistical significance at the 5% level,
two-sided test; and * indicates statistical significance at the
10% level, two-sided test.

TABLE 5
Estimates of the Log Wage Differential Controlling for Stayers

 White Men

 Fixed-Effects Gross+
Variable Specification Specification

Married 0.0046 (0.0155) 0 0258 (0.0308)
Separated/divorced -0.0186 (0.0220) -0.0653 (0.0405)
Years married -0.0056 (0.0041) 0.0067 ** (0.0030)
Never married -0.1025 *** (0.0387)
Always married 0.0562 ** (0.0270)
Always divorced -0.1428 *** (0.0428)
Log of training time 0.0137 *** (0.0038)
Including
 3 AFQT residual scores
 Parents' education
 Parents' occupation
Self-esteem
F test on marital status 1.18 6.33 ***
F test on stayer status 9.61 ***
Number of observations 11,581 11,581
Number of individuals 2,333 2,333
[R.sup.2] 0.8100 0.0450

 White Men

 Training+ Individual-Specific+
Variable Specification Specification

Married 0.0105 (0.0215) 0.0116 (0.0209)
Separated/divorced -0.0088 (0.0284) -0.0092 (0.0279)
Years married 0.0031 (0.0028) 0.0023 (0.0028)
Never married -0.0327 (0.0270) -0.0283 (0.0260)
Always married 0.0373 * (0.0201) 0.0367 * (0.0200)
Always divorced -0.0258 (0.0296) -0.0278 (0.0294)
Log of training time 0.0129 *** (0.0033) 0.0108 *** (0.0033)
Including
 3 AFQT residual scores Yes ***
 Parents' education Yes
 Parents' occupation Yes *
Self-esteem Yes ***
F test on marital status 0.53 0.36
F test on stayer status 2.31 * 2.21 *
Number of observations 11,581 11,581
Number of individuals 2,333 2,333
[R.sup.2] 0.4409 0.4595

 African American Men

 Fixed-Effects Gross+
Variable Specification Specification

Married 0.0573 ** (0.0285) 0.0108 (0.0521)
Separated/divorced -0.0484 (0.0358) -0.1514 ** (0.0660)
Years married -0.0090 (0.0061) 0.0078 (0.0056)
Never married -0.1750 *** (0.0554)
Always married 0.0397 (0.0448)
Always divorced 0.0298 (0.0632)
Log of training time 0.0028 (0.0062)
Including
 3 AFQT residual scores
 Parents' education
 Parents' occupation
Self-esteem
F test on marital status 1.89 6.53 ***
F test on stayer status 4.05 ***
Number of observations 4,642 4,642
Number of individuals 911 911
[R.sup.2] 0.7295 0.0633

 African American Men

 Training+ Individual-Specific+
Variable Specification Specification

Married 0.0312 (0.0360) 0.0185 (0.0338)
Separated/divorced -0625 (0.0453) -0.0733 * (0.0429)
Years married 0.0013 (0.0045) 0.0028 (0.0044)
Never married -0.0395 (0.0365) -0.0424 (0.0332)
Always married 0.0379 (0.0291) 0.0264 (0.0293)
Always divorced 0.0749 * (0.0434) 0.0753 * (0.0440)
Log of training time 0.0122 *** (0.0042) 0.0113 *** (0.0040)
Including
 3 AFQT residual scores Yes ***
 Parents' education Yes
 Parents' occupation Yes ***
Self-esteem Yes
F test on marital status 3.39 ** 3.15 **
F test on stayer status 2.02 1.88
Number of observations 4,642 4,642
Number of individuals 911 911
[R.sup.2] 0.4537 0.4741

Notes: Robust standard errors corrected also for clustering by
individual are reported in parentheses below coefficient
estimates. Each specification includes control for years of
education; a quadratics in experience and tenure; dummy variable
to identify 11 industries, 10 occupations, 3 regions, 3 city
sizes, and 5 interview years; a measure of the number of
children, firm size, and the local unemployment rate; dummy
variables to identify union members and government employees; and
dummy variables to indicate receipt of training in the particular
interview year.

*** indicates statistical significance at the 1% level, two-sided
test; **indicates statistical significance at the 5% level,
two-sided test; *indicates statistical significance at the 10%
level, two-sided test.