UNOBSERVABLE INDIVIDUAL EFFECTS, MARRIAGE

AND THE EARNINGS OF YOUNG MEN

CHRISTOPHER CORNWELL and PETER RUPERT*

While there is compelling evidence that married men earn more than unmarried
men, the source of this premium remains unsettled. Using panel data from the
National Longitudinal Survey of Young Men, we show that much of the premium
normally attributed to marriage is associated with unobservable individual effects
that are correlated with marital status and wages. To the extent there is a gain,
it is purely an intercept shvt and no more than 5% to 7%. Ourfindings cast doubt
on the interpretation that marriage enhances productivity through specialization.
(JEL JO)

I. INTRODUCTION later than those who acquire less (Bergstrom

Married men earn more than unmarried and Schoeni [ 19921). An alternative explana-
men. This fact is unassailable and is robust tion for the marriage premium is that marriage
across data sets and over time. Two common signals certain unobservable individual char-
explanations of the married-male wage pre- acteristics, such as ability, honesty, loyalty,
mium are (i) the division of labor in a married dependability and determination that are val-
household which allocates more of the man’s ued in both the labor and marriage markets.’
time to the market and (ii) lower cost human Although (i) and (ii) are quite different in
capital acquisition for married men. Both of terms of their behavioral content, they do
these stories view marriage as productivity- share a common thread. Whether marriage per
enhancing. Kenny [ 19831 provides some evi- se increases productivity or high-productivity
dence that marriage facilitates human capital men are more likely to be married, marital
acquisition, however the potential endogene- status should not be treated as an exogenous
ity of marital status is ignored. Furthermore, determinant of the wage rate.
the argument that marriage makes it cheaper Nevertheless, in most wage regressions
to accumulate human capital is difficult to that control for marital status, it is assumed to
reconcile with the fact that individuals who be exogenous, resulting in an estimated mar-
acquire more formal education tend to marry riage premium that is biased upward. Two ex-
ceptions are Nakosteen and Zimmer [1987]
and Korenman and Neumark [1991]. Al-
* We wish to thank Eleanor Brown, Ben Craig, Eliza- though both papers base their empirics on
beth Powers, Mark Schweiker, two anonymous referees,
and seminar participants at the University of Rochester, samples of young men, they come to very dif-
University of Southern California, and the University of ferent conclusions regarding the size and
Waterloo. Opinions expressed here are not necessarily source of the marriage premium. Nakosteen
those of the Federal Reserve Bank of Cleveland or the
Federal Reserve System. and Zimmer find the marriage premium to be
Cornwell: Associate Professor, University of Georgia, insignificantly different from zero and claim
Athens, Phone 1-706-542-3670, Fax 1-706-542-3376 that marriage does not make men more pro-
E-mail [email protected] ductive. Korenman and Neumark (KN) pres-
Rupert: Economist, Federal Reserve Bank of Cleveland,
Ohio, Phone 1-216-579-2040, Fax 1-216-579-3050 ent evidence that the marriage premium is at
E-mail [email protected]
1. Reed and Harford [1989] provide another altema-
tive: the marriage premium represents a compensating dif- ABBREVIATIONS
ferential required to induce married men to accept unde- FGLS: Feasible generalized least squares
sirable working conditions. In their view, marriage is re- KN: Korenman and Neumark
lated to the purchase of costly “family goods” such as NLSYM: National Longitudinal Survey of
children. One other explanation (we thank Eleanor Brown Young Men
for this point), is that men “settle down” and spend more
effort in the market. OLS: Ordinary least squares

285
Economic Inquiry
(ISSN 0095-2583)
Vol. XXXV, April 1997, 285294 0 Western Economic Association International
286 ECONOMIC INQUIRY

least 15% and rises with years married, indi- II. MARRIAGE AND WAGES
cating that marriage may be productivity en- Virtually all cross-section wage regres-
hancing. sions that control for marital status report a
In this paper we reexamine the empirical large, statistically significant wage premium
relationship between marriage and wages for married men. Some of the more prominent
using a sample drawn from the 1971, 1976, examples are discussed by Reed and Harford
1978 and 1980 waves of the National Longi- [ 19891 and KN. Our view is that the true mar-
tudinal Survey of Young Men (NLSYM). Our riage premium is much smaller than that com-
data set is an extension of the one used by monly reported in cross-section wage regres-
KN, whose empirical work is based on the sions. We find that the wage premium can be
1976, 1978 and 1980 waves of the NLSYM.* explained largely in terms of unobservable in-
Our results may be summarized as follows. dividual characteristics which are positively
Estimates of the marriage premium derived correlated with marriage and wages. In other
from fixed effects models (which control for words, attributes leading to "good" (long and
unobserved heterogeneity) are uniformly stable) marriages are also important in obtain-
smaller than those obtained from random ef- ing "good" (long and stable) jobs and higher
fects models (and other cross-section ap- wages.
proaches that do not). Specification tests re- This view is supported by some of the work
ject the latter and suggest that the individual on the returns to tenure. Abraham and Farber
effects are positively correlated with both [ 19871 propose that workers in long-tenure
marriage and wages. Conditional on labor jobs earn more in every year on the job and
market experience, job tenure and year ef- that most of the cross-section return to tenure
fects, time spent in marriage has no signifi- is due to unobserved individual and job match
cant effect on wages. This stands as evidence effects. They test their proposition by estimat-
against any dynamic productivity enhancing ing wage equations conditional on predicted
effect of marriage. We conclude that whatever job duration. Interestingly, results from the es-
the source, the marriage premium is a pure timation of their job duration model indicate
intercept shift and no more than 5% to 7%. that marriage has a large, positive and statis-
Further support for this conclusion comes tically significant effect.
from the relationship between prospective Further evidence exists in the quit behavior
marriage and wages. Using our longer panel, of married men. Consistent with the positive
we construct a "to-be-married" variable which relationship between marriage and job dura-
identifies men who marry sometime during tion is the depressing effect marriage has on
the sample period; men who are not married quits. Shaw [1987] reports that the quit rate
in 1971, but who will be, earn at least as much for married men aged 25 to 54 is less than half
as those who are already married. Finally, an that of unmarried men, and marriage has its
interesting related result is the robust and sig- strongest deterrent effect on younger men.
nificant effect of children on wages; condi- Only Nakosteen and Zimmer [ 19871 argue
tional on marital status and years married, the explicitly that marriage does not significantly
presence of children raises wages about 5% influence the wages of young men. However,
regardless of the estimator employed. We in- the empirical support (which is based on a
terpret the effects of marriage and children on cross-section of young men from the 1977
wages similarly-as structural breaks in the wave of the Panel Study of Income Dynamics)
wage-generating process involving adjust- for their argument is weak. When they restrict
ments in market work and homework (among marital status to be exogenous and apply or-
other things). In any case, the findings regard- dinary least squares (OLS) to their wage equa-
ing prospective marriage and children are tion, Nakosteen and Zimmer obtain a statisti-
anomalous in a specialization story of mar- cally significant marital status coefficient es-
riage and wages. timate of 0.370. Relaxing the exogeneity re-
striction actually causes the marital status co-
efficient estimate to rise, although it is no
2. We are grateful to Sanders Korenman for providing longer statistically significant. Furthermore,
their data and regression programs to us. specification tests of exogeneity are inconclu-
 CORNWELL & RUPERT: MARITAL STATUS AND EARNINGS 287

sive. Nevertheless, Nakosteen and Zimmer 111. MODEL AND ESTIMATION

conclude that the true marriage premium is Our empirical analysis of the differences in
insignificantly different from zero. wages between married and single men is car-
Inference based primarily on an increase in ried out on models of the form:
a parameter estimate’s standard error, when
the parameter estimate remains large and pos- (1) yit = 6 + X i t p + M i ,y + ai + Eit ,
itive, is not very powerful. The imprecision of
Nakosteen and Zimmer’s results may reflect where i indexes individuals (i = l,...,N), t in-
the difficulty of specifying the process gener- dexes time ( t = 1,...,g,yit is the natural loga-
ating marital status. KN suggest that the prob-
lem may be collinearity of the selection term rithm of the real wage, 4, is a vector of ex-
with marital status. planatory variables and Mi, is a vector of mar-
KN offer a different view, claiming that the riage variables (which in the simplest form of
gains to marriage are large for young (white) the model contains only a status dummy). The
men, even in fixed effects regressions. From ai are intended to capture unobserved, indi-
their preferred specification, which includes vidual-specific attributes that are valued in
years married and its square in the wage equa- both the labor and marriage markets and
tion along with a marital status dummy, KN therefore may be positively correlated with
report a fixed effects marriage premium esti- both wages and marriage, while the sit reflect
mate of 15%. Eighty percent of the premium aspects of the wage-determining process
is attributable to time spent in marriage; only which may be represented as statistical noise.
3 percentage points are due to an (insignifi- Also, the ai and cif are taken to be i.i.d. ran-
cant) intercept shift. dom variables, uncorrelated with each other,
We are skeptical about the Nakosteen and with zero means and constant variances G:
Zimmer and KN results. The weaknesses of
the former study are outlined above, but are and CJ;, respectively.
not addressed here, owing to our belief that Estimation of (1) depends on whether or
marriage premium estimates derived from not inference is to be performed conditional
panel data techniques are more ~ r e d i b l eO
.u
~r on the unobserved individual attributes a,.If
skepticism regarding the latter is based on two not, then (1) is estimated by feasible general-
characteristics of their sample. First, marital ized least squares (FGLS). As long as the set
status changes represent a relatively small of explanatory variables (Xi.,Mit) is un-
proportion-less than 25%--+f their sample. correlated with ai, FGLS will be consistent,
More importantly, the vast majority-over as would be OLS. But given the composed
8 0 % - o f status changes are divorces and sec- disturbance, FGLS will be efficient relative to
ond marriages. The proportion and nature of o~s.4
the changes is important since identification If the explanatory variables (Xi.,, Mir) are
in fixed effects models is obtained through correlated with the individual-specific attri-
“within-group” variation (variation over time butes a ,then the individual effects should be
in the regressors). Thus, the main source of
the KN marriage premium is individuals who controlled for in estimation; that is, inference
divorce or marry for the second time during should be conditional on the effects. The sim-
the sample period. As discussed below, the plest means of accomplishing this is by treat-
ing the ai as fixed parameters and estimating
over-representation of divorces and remar-
riages among the status changes increases the (1) by applying OLS to data that has been
likelihood of an insignificant intercept shift, transformed into deviations from individual
along with a significant years-married coeffi- means. The fixed effects, or “within,” estima-
cient.
4. Note that KN estimate a slightly different version of
(1). The individual-specific component, a , is ignored in
3. Elsewhere, in Cornwell and Rupert [1993], we have
pursued the Nakosteen and Zimmer modelling strategy their cross-section estimation, but the T-vector E~ is as-
using the NLSYM, with similarly unconvincing results. sumed to have an arbitrary intertemporal covariance. Esti-
The main obstacle is the lack of good instruments for mar- mating (1) under both assumptions, we obtained essentially
nage. the same estimated coefficients and standard errors.
288 ECONOMIC INQUIRY

tor is consistent regardless of the relationship To the KN sample, we add observations on

between explanatory variables and the (xi. The each of these variables from the 1971 wave
issue of which set of estimates to use in con- of the NLSYM. In addition, we include the
structing the marriage premium can be re- variable job tenure.6 Variable means by year
solved by performing a Wu-Hausman test of and marital state are presented in Table I. The
the difference in the within and FGLS esti- addition of the 1971 data allows us to observe
mates. these young men over a period (between 1971
and 1976) when marriage market activity is
IV. RESULTS FROM ESTIMATION relatively greater. The result is a higher per-
centage (34) of the sample changing status
Data and a lower percentage (52) of the status
Our marriage premium estimates are de- changes coming from divorces and second
rived from a sample of young men drawn from marriages. At the same time, the ratio of never
the 1971, 1976, 1978 and 1980 waves of the marrieds to always marrieds is roughly equiv-
NLSYM. The young men who comprise the alent in both samples.
NLSYM were first surveyed in 1966, at which Extending the KN sample to 1971, while
time they were all between the ages of 14 and maintaining a balanced panel, forces a reduc-
24. Thus, in 1971 we begin with a cohort that tion in the cross-section size (N) to 666 young
is aged 19 to 29. As noted above, our sample men7 Despite the smaller cross section, in
is an extension of KN’s, which uses only the terms of non-wage characteristics, the average
1976, 1978, and 1980 waves of the NLSYM. young man in our sample looks very much
Since their sample does not start until 1976, like the average young man in theirs. The only
the youngest men are already 24 years old, notable cross-sectional difference is that the
restricting the number and type of their status average man in our sample is about one year
changes. Fewer than 25% of their young men older, due in large part to the elimination of
change status, and over 80% of the changes some young men who are still in school in
are divorces and second marriages. 1971.
The details of the construction of the data
set are given in KN, so we will not repeat
Random Eflects (FGLS) Results
them here. However, we will briefly review
the variables contained in the data set and in- First we estimate (1) by FGLS, restricting
dicate how our sample differs from theirs. the effect of marriage on wages to an intercept
After attrition and imposing sample selection shift. These results, which are consistent only
criteria (for example, restricting the analysis if the explanatory variables are uncorrelated
to white men who have completed their with the individual effects, are reported in col-
schooling and report a positive wage), they umn (1) of Table 11. As indicated in Table 11,
end up with 1228 individuals. For each young the FGLS marriage premium is estimated con-
man in each time period, KN were able to ditional on labor market experience and its
observe their wage rate, marital status square, five education dummies identifying
(whether married with spouse present, di- highest grade completed, dummy variables for
vorced or never married),’ years married and residence in the south and in an urban area, a
divorced, experience (which is actual post- dummy variable for union status of the job
schooling work experience, not the usual age and a dummy variable for the presence of non-
- schooling - 6), residence (whether he lives spouse dependents. Although the coefficient
in the south or in an urban setting), union sta- estimates are not presented, we also included
tus (whether the wage is set by collective bar- eight occupation, eleven industry and ten age
gaining), non-spouse dependents (whether dummies, as well as dummy variables to cap-
there are any), age, education (in terms of
highest degree obtained) and one-digit occu- 6. KN used tenure as a control in regressions that they
pational and industry affiliation. did not report (see p. 296 of their paper), but the variable
was not provided in the sample they sent us.
7. Missing wages reduces N to 966. Another 48 men
5. The few men who are mamed but whose spouse is are lost due to missing years married. Finally, restricting
absent, or are widowed, are included in the divorced cate- the sample to only those who have completed their school-
ing leaves us with N = 666.
gory.
 CORNWELL & RUPERT MARITAL STATUS AND EARNINGS 289

TABLE I
Summary Statistics by Marital Status and Year ( N = 666)
1971 1976
Never Never
Married Divorced Married Married Divorced Married
Percent of Sample 68 4 28 85 4 11
Wage 6.00 5.77 5.74 6.46 6.39 6.28
Years Mamed 5.14 4.61 8.61 5.47
Years Divorced 0.09 1.33 0.23 3.45
Experience 6.18 7.65 3.21 8.24 7.17 6.07
Tenure 4.19 3.77 2.28 6.87 5.79 6.14
South 0.32 0.35 0.25 0.3 1 0.3 1 0.28
Urban 0.66 0.73 0.78 0.67 0.83 0.78
Union 0.37 0.35 0.28 0.35 0.4 1 0.28
Dependents 0.73 0.62 0.03 0.85 0.48 0.06
Age 25.1 25.3 22.3 29.5 29.4 28.1
No High School 0.04 0.04 0.02 0.04 0.03 0.01
Some High School 0.12 0.3 1 0.09 0.13 0.07 0.07
Some College 0.15 0.1 1 0.3 1 0.18 0.24 0.27
College Grad 0.15 0.04 0.1 1 0.14 0.07 0.14
Post-College 0.10 0.04 0.10 0.10 0.07 0.13
1978 1980
Percent of Sample 88 3 9 86 6 8
Wage 6.65 6.33 6.47 6.84 6.67 6.67
Years Married 10.5 4.10 12.4 7.78
Years Divorced 0.30 5.10 0.36 3.71
Experience 10.0 8.36 8.13 11.8 10.5 10.3
Tenure 7.76 5.15 7.43 8.92 7.58 8.79
South 0.32 0.30 0.28 0.32 0.32 0.3 1
Urban 0.68 0.65 0.75 0.67 0.7 1 0.77
Union 0.38 0.25 0.3 1 0.37 0.37 0.29
Dependents 0.89 0.40 0.1 1 0.92 0.58 0.09
Age 31.5 31.2 30.2 33.5 32.7 32.3
No High School 0.04 0.02 0.03 0.05 0.02
Some High School 0.12 0.20 0.07 0.12 0.16 0.07
Some College 0.18 0.35 0.26 0.18 0.26 0.23
College Grad 0.14 0.16 0.14 0.16
Post-College 0.10 0.05 0.11 0.10 0.05 0.12

ture year effects. The list of right-hand-side Like KN, our results indicate a positive and
variables in our FGLS regression corresponds statistically significant cross-section effect of
exactly to the KN specification (1) (see their non-spouse dependents on wages. When the
Table 11), from which they obtain an estimated marriage premium is computed conditionally
cross-section marriage premium of 10.6%. ‘on the presence of non-spouse dependents, it
FGLS applied to the dummy variable spec- rises to 14.4%, which is essentially the result
ification of (1) yields a marital status coeffi- obtained by KN for married men with non-
cient estimate of 0.083 with a standard error spouse dependents. Thus married men with
of only 0.022, which implies significance at children earn, on average, 5.3% more than
any reasonable test size. This suggests a mar- married men without children. As we will see
riage premium of 8.7%, which is about 2 per- below, in our sample the dependents effect is
centage points smaller than the cross-section positive, statistically significant and typically
premium reported by KN. We also find a sta- about 5%, whether the individual effects are
tistically significant wage premium of about treated as random or fixed, and even when we
6.6% for divorced men. control for marriage duration. K ” s depen-
290 ECONOMIC INQUIRY

TABLE I1
Estimated Wage Regressions
(Standard Errors in Parentheses)
(1) (2) (31 (4)
Variable FGLS Within Within Within
______

Married 0.083 0.056 0.05 1 0.033

(0.022) (0.026) (0.026) (0.028)
Divorced 0.064 0.062 0.057 0.040
(0.033) (0.036) (0.036) (0.038)
Years Married -0.005
(0.006)
Years Married’ -0.0003
(0.0003)
Years Divorced -0.014
(0.008)
Experience 0.027 0.027 0.024 0.021
(0.004) (0.004) (0.004) (0.005)
Experience’ -0.001 -0.001 -0.001 -0.001
(0.0001) (0.0002) (0.0002) (0.0002)
Tenure 0.013 0.011
(0.004) (0.004)
Tenure2 -0.0006 -0.0005
(0.0002) (0.0002)
South 4.091 -0.121 -0.117 -0.118
(0.019) (0.034) (0.034) (0.034)
Urban 0.137 0.057 0.059 0.059
(0.017) (0.024) (0.024) (0.024)
Union 0.109 0.106 0.102 0.103
(0.015) (0.018) (0.018) (0.018)
Dependents 0.052 0.052 0.048 0.047
(0.017) (0.019) (0.019) (0.020)
No High School -0.325
(0.057)
Some High School -0.148
(0.032)
Some College 0.091
(0.028)
College Grad 0.278
(0.034)
Post-College 0.322
(0.041)
Standard error 0.215 0.212 0.212 0.2 11
2
x2 7
111.9

dents effect is much less robust, becoming eral, despite the smaller cross-section dimen-
smaller and insignificant in their fixed effects sion of our sample, the basic qualitative and
regressions. quantitative results obtained from the larger,
With regard to the other coefficient esti- but shorter KN panel have been preserved. We
mates, they are generally typical of those emphasize that the only significant difference
found in the literature and very similar to between our results and those of KN lies in
those obtained by KN in their sample. In gen- the returns to marriage.
 CORNWELL & RUPERT MARITAL STATUS AND EARNINGS 29 1

Fixed Effects (within) Results with marital status, then excluding it would
Next, we treat the individual effects in (1) likely bias upward the marital status coeffi-
as fixed and re-estimate the model with the cient. The results in column (3), where we
same set of explanatory variables as in the have added tenure and its square to the base-
FGLS regression. These fixed effects, or line fixed effects regression, suggest that this
within, estimates are given in column (2) of is the case. First, both tenure and tenure-
Table 11. Note that all time-invariant variables squared enter statistically significantly, and
(the education and age dummies) are elimi- their estimated coefficients imply roughly a
nated along with the effects by the data trans- 1.3% return per year in the early years on the
formation implied by the within estimator. j o b . S e c o n d , in c lu d in g te n u re f u r t h e r
The fixed effects regression that produced depresses the marital status coefficient esti-
the estimates in column (2) corresponds to the mate to 0.051, which translates into a mar-
specification (1‘) of KN. Based on our sample, riage premium of 5.2%.
the within estimate of the marital status coef- This result might be expected given the ev-
ficient is 0.056, and it is statistically signifi- idence cited from Abraham and Farber [ 19871
cant at the 5% level. This translates into a on the relationship between jo b duration
marriage premium of 5.8%, which is close to (completed tenure) and marital status. We find
the KN result from their model (1’) of 6.1%. the Abraham and Farber result in our sample
The sharp diminution of the marriage pre- as well. In an estimated Weibull hazard (not
mium in our fixed effects regression suggests reported) ofjob duration of the 1971 job, cor-
there are important omitted individual-spe- recting for right-censoring, conditional on
cific characteristics which are positively cor- union status, education, occupational and in-
related with both marriage and the wage rate. dustry affiliation and pre-job experience, mar-
A question not addressed by KN is whether riage has the largest statistically significant
the difference between the cross-section and positive impact.
fixed effects results are statistically signifi- If the effect of marriage on wages varies
cant. The question is important because a sta- over time, the dummy variable specification
tistically significant difference between the is incorrect. Furthermore, evidence of a time-
two regressions would cast doubt on the cross- varying wage premium would tend to support
section results. Since the within estimates are the view that marriage is productivity enhanc-
consistent even if the explanatory variables ing; for example, via ongoing specialization.
are correlated with the individual effects, Thus, we. re-estimate our fixed effects regres-
while cross-section based estimates (like sion, adding years married, its square and
those produced by OLS and FGLS) are not, years divorced.8 The results from this specifi-
the null hypothesis of no correlation between cation, which corresponds to KN’s model (2’)
the regressors and the effects is easily tested with the addition of tenure and tenure-
through a Wu-Hausman test of the estimates squared, are presented in column (4).
in the first two columns of Table 11. The test First, notice that the estimated coefficients
statistic for this contrast is asymptotically dis- of the years-married variables are insignifi-
tributed as xi7 and has a value of 111.9, which cantly different from zero. In neither case do
the coefficient estimates even exceed their
is almost three times greater than the critical standard errors. In contrast, KN report statis-
value. Hence, the null hypothesis is rejected tically significant years married and years
and cross-section based marriage premia married-squared coefficients of 0.022 and
along with it. Interestingly, aside from the co- -0.00096. However, as noted above, divorces
efficient of the urban dummy, it is the marital and remarriages are over-represented among
status coefficient estimate that is the most af- the status changes in the KN sample. Second,
fected by controlling for unobserved hetero- including the years-married and years-di-
geneity. vorced variables reduces the status coefficient
The regressions in the first two columns do
estimate from 0.051 to 0.033 and its associ-
not control fo r j o b ten u re, which is a
misspecification to the extent firms make spe-
cific investments in workers. If wages rise 8. Years married and years divorced are measured as
with tenure, and tenure is positively correlated the total time spent in each state.
292 ECONOMIC INQUIRY

ated t statistic from 1.96 to 1.17. Finally, the in the marriage market-are at least as pro-
estimated coefficients and standard errors of ductive as men who are already married.
the tenure variables are essentially unaffected
by the addition of years married and divorced. The Eflect of Children
In sum, when we control for unobserved
heterogeneity and job tenure, our sample does As noted above, we find the positive effect
not provide evidence of a statistically signif- of children on wages to be robust across a
icant time-varying marriage premium. If mar- wide variety of specifications. Although con-
riage influences wages at all, the premium is ditioning on the individual effects reduces the
a pure intercept shift and only about 5%, overall marriage premium, the positive and
which is much smaller than standard cross- significant effect of non-spouse dependents is
section results suggest. To the extent it exists, unchanged. The results in column (2) of Table
such a shift in the wage-generating process I1 still suggest that young men with non-
might be regarded as the effect of “settling spouse dependents earn 5.3% more those
down”+ kind of structural break involving without. Thus, the premium for married men
adjustments in market work and homework in with children is over 11%. Controlling for ten-
the move from single to married life. It is pos- ure diminishes the “dependents premium”
sible that years married captures the same slightly, but it remains nearly 5% and statis-
thing in the KN sample. Since the marriage tically significant.
premium in their study is identified largely by With regards to the marriage premium, the
men who divorce and remarry (not by men robust effect of children on the wages of
who marry for the first time), years married young men raises several questions. Since
(rather than the status variable) may indicate married couples with children usually have
whether a man ever “settled down.’’ It is rea- been married longer than childless couples,
sonable to suppose that such adjustments are could the dependents premium be masking the
more likely to have taken place in marriages effect of years married on wages? If not, how
that lasted longer. are the wages of young men affected by the
That marriage also signals certain individ- presence of children in the household? Does
ual attributes that are valued in the labor mar- the effect of marriage on wages depend on the
ket is further supported by the relationship be- presence of children?
tween prospective marriage and current First, it does not appear to be the case that
wages. Using the 1971 cross section, we de- the dependents variable and years married are
fine the variable, “to-be-married,” as a measuring the same thing. When we omit the
dummy equal to one if a (single) young man dependents variable from the specification
marries sometime during the sample period given in column (4), the estimated coefficients
(after 1971). Then, in a separate regression (and standard errors) of years married and its
(not reported), we estimated the effects of to- square are, respectively, 0.00008 (0.006) and
be-married on the 1971 wages of the young -0.0005 (0.00027). Thus, the years-married
men in our sample. Conditional on the 1971 coefficient estimate is clearly insignificant,
values of the explanatory variables corre- while the estimated coefficient of the qua-
sponding to column (1) in Table 11, the wages dratic term has a marginal significance level
of those who are not yet but will be married of only 0.06. Dropping dependents does affect
are about 4% higher than those who are mar- the marital status coefficient estimate; it rises
ried in 197l. The to-be-married coefficient es- to 0.41, although it remains insignificant even
timate is 0.148 and the marital status coefi- at the 10% level.
cient estimate is 0.106; the former is statisti- Second, the robust effect of children on
cally significant at the 5% level, while the wages seems to be like that of marriage de-
latter is significant at only the 10% level.g Ap- picted in column (3)-a change in the process
parently, those young men who are not mar- generating wages associated with the reallo-
ried, but for whom marriage is likely-i.e., cation of market, housework and leisure time
they have characteristics that lead to success to finance the added costs of children. At the
simplest level, young men with children will
9. The large return to marriage is due to the fact that
allocate more time to market work and less to
it is obtained from a cross section regression. leisure; for example, they may amve to work
 CORNWELL & RUPERT MARITAL STATUS AND EARNINGS 293

. TABLEIII
Dependents and the Returns to Marriage
(Standard Errors in Parentheses)
(1) (2)
Variable Within Within

Married 0.071 0.027

(0.026) (0.038)
Divorced 0.008 - 0.033
(0.049) (0.058)
Years Married 0.011
(0.0 12)
Years Married’ -0.001
(0.0007)
Years Divorced -0.013
(0.01 1)
Experience 0.024 0.021
(0.004) (0.005)
Experience’ -0.001 -0.001
(0.002) (0.0002)
Tenure 0.013 0.011
(0.004) (0.004)
Tenure’ -0.0006 -0.0005
(0.0002) (0.0002)
south -0.113 -0.109
(0.034) (0.034)
Urban 0.061 0.061
(0.023) (0.024)
Union 0.101 0.110
(0.018) (0,018)
Dependents 0.292 0.281
(0.076) (0.076)
Married x Dependents -0.266 -0.232
(0.077) (0.087)
Divorced x Dependents -0.156 -0.124
(0.095) (0.103)
Years Married x Dependents -0.013
(0.012)
Years Mamed’ x Dependents 0.001
(0.0008)
Years Divorced x Dependents -0.001
(0.0 12)
Standard error 0.213 0.211

more promptly. Or, as indicated by Reed and (as in Table 11, column (3)), but includes an
Harford [19891, young men with children also interaction of marital and divorced status and
may accept jobs that offer fewer amenities and dependents. Column (2) adds the years mar-
benefits, but greater wage compensation. ried variables (as in Table 11, column (4))and
Finally, as indicated in Table 111, the de- interactions thereof with dependents to test
pendents (marital status) effect varies signifi- whether the effect of children (marriage) var-
cantly with the marital status (dependents) ef- ies with years married (children in the house-
fect. Column (1) reports results from a regres- hold).
sion which omits years married and divorced
294 ECONOMIC INQUIRY

In the first case, the interaction between from the NLSYM) is shorter and the status
marital status and dependents is large, nega- changes contained therein are dominated by
tive and statistically significant. With the in- divorces and remarriages. As an alternative to
teraction, wages of married men with children the specialization story, perhaps marriage in-
are still about 10% higher than childless sin- duces men to “settle down” and substitute lei-
gle men (which is similar to the premium im- sure for wages.
plied in column (3) of Table II), but the esti- Additional support for this view is pro-
mated premium for married men without chil- vided by two findings which are anomalous
dren rises from 5.2% to 7.3%. In the second in the specialization story of marriage. First,
case, the only term involving a marriage vari- men who are not married in 1971, but who
able that enters statistically significantly is the eventually will be, earn at least as much as
interaction between status dummy and depen- those who are already married. Second, con-
dents. Like in column (l), its coefficient esti- ditional on marital status and years married,
mate is large and negative. Although they are the presence of children raises wages about
not statistically significant, the estimated co- 5%, regardless of the estimator employed.
efficients of the years-married variables and Parenthood seems to affect wages similarly to
their interactions with dependents suggest that marriage-as a shift in the wage-generating
the marriage premium for young men with process involving adjustments in market work
children declines during the early years of and homework.
marriage.
V. CONCLUSIONS REFERENCES
Abraham, K. G., and H. S. Farber. “Job Duration, Seniority,
While there is little or no debate about the and Earnings.” American Economic Review, 77(3),
claim that married men earn more than unmar- 1987,278-97.
ried men, the source of the wage premium re- Bergstrom, T., and R. Schoeni. “Income Prospects and Age
mains unsettled. Using a sample of young men at Marriage.” University o f Michigan CREST Work-
ing Paper 92-10, 1992.
from the NLSYM, observed between 1971
Comwell, C., and P. Rupert. “Unobservable Individual Ef-
and 1980, we show that much of the return is fects, Mamage and the Earnings of Young Men.”
associated with unobservable individual ef- University of Georgia Economics Department Work-
fects that are correlated with marital status ing Paper, 1993.
and wages. Fixed effect regressions, which Kenny, L. W. “The Accumulation of Human Capital During
condition on the unobservables, substantially Marriage by Males.” Economic Inquiry, 21(2), 1983,
223-3 1.
attenuate the estimated wage premium nor- Korenman, S., and D. Neumark. “Does Marriage Really
mally attributable to marriage. Moreover, to Make Men More Productive.” Journal of Human Re-
the extent that there is a gain to marriage, it sources, 26(2), 1991, 282-307.
is purely an intercept shift (does not vary with Nakosteen, R. A., and M. A. Zimmer. “Marital Status and
years married) and is no greater than 5% to Earnings of Young Men: A Model with Endogenous
Selection.” Journal of Human Resource, 22(2), 1987,
7%. We regard this as evidence against mar- 248-68.
riage enhancing productivity through special- Reed, W., and K. Harford. “The Marriage Premium and
ization. In contrast, KN [1991] report a large Compensating Wage Differentials.” Journal of Pop-
marriage premium that increases with years ulation Economics, 2(4), 1989, 237-65.
married; however, their panel (also drawn Shaw, K. “The Quit Propensity of Married Men.” Journal
of Labor Economics, 5(4), 1987, 533-60.