TWO DECADES OF FAMILY CHANGE 137 not asked in every interview year.Thus the 1980s that are available from other surveys analytical sample for the early cohort is lim- were primarily obtained retrospectively ited to men who responded in 1976,and to (Smock 2000),although little is known about women who responded in either 1978 or the quality of retrospective reporting on co- 1983,when more complete marital histories habitation.The lack of information on co- were collected.I retain the cross-sectional habitation is a drawback of using the NLS sample and the oversample of blacks in all data,but the generally brief duration of co- surveys.Because of data limitations,the habitation for most couples should lead to samples are restricted to blacks and whites, relatively small differences in estimations of with Hispanics coded by race rather than the timing of first marriages versus first ethnicity.4 To minimize problems from left- unions of any type.Indeed,fully half of co- censoring of important economic covariates habitations end within 16 months,and only associated with marriage formation,I further one in ten couples is still cohabiting (and un- limit the analysis to individuals under age 19 married)after five years (Bumpass and at first interview.The analytical samples in- Sweet 1989).It is important to keep in mind, clude a total of 3,631 women and men in the however,that the present study sheds light early cohort and 4,920 women and men in on just one piece of the broader union forma- the late cohort.All descriptive statistics are tion process-entry into a legal marriage. weighted to adjust for design issues,such as the oversampling of blacks,and for differ- VARIABLES AND METHODS ential nonresponse to the interviews(Center for Human Resource Research 1997). I use logistic regression analysis to estimate It is desirable to have information col- discrete-time hazard models of the effects of lected at as many time points as possible for economic prospects on entry into first mar- an event-history analysis of marriage.While riage.This approach permits the estimation the Youth sample (NLSY)was interviewed of effects of fixed and time-varying annually from 1979 through 1994,the Young covariates on entry into marriage,and avoids Men's (NLSM)and Young Women's the assumption of proportional hazards (NLSW)samples were interviewed some- (Allison 1995).The dependent variable in what less regularly.Wherever possible,the the analysis is a dichotomous indicator of annual measures are constructed from retro- whether a marriage occurred in the interval spective questions about activities during the between two given years,with time-varying noninterview year.This approximation is independent variables fixed at the beginning generally feasible for measures of school en- of the interval.Data are organized into per- rollment,for example.In other cases,where son-year records,with one record for each these retrospective questions were not asked, annual interval in which respondents were at missing data are generally imputed from the risk of first marriage,including intervals in previous interview year.5 A variable indicat- which a first marriage occurred.For this ing a noninterview year for the early baby- analysis,the risk of marriage is assumed to boom cohort (NLSM and NLSW)is in- begin at age 17.Because the focus of this cluded in all models. analysis is on the comparison of coefficients Unfortunately,it is not possible to exam- between cohort models,the models them- ine cohabitation using the NLS data because, selves are kept relatively simple,with few like most surveys conducted prior to the mid- complex variable or model specifications. 1980s,the relevant questions were not asked Table 1 shows the mean values of the in- consistently.Data on cohabitation prior to the dependent variables,separately by race and sex.This analysis uses three well-estab- lished measures of standing in the labor mar- 4 It is not possible to identify Hispanic eth- ket to reflect economic prospects:earnings, nicity in the NLSM sample,and ethnicity is not educational attainment,and employment sta- ascertained until the 1993 interview in the NLSW sample. tus.Both a continuous measure of logged 5A variety of approaches to dealing with miss- earnings (wage and business income)and a ing data were considered,with substantive results dummy variable for zero earnings in the robust to choice of method prior year are constructed,based on the as- This content downloaded from 128.103.149.52 on Mon,12 Aug 2013 19:18:11 PM All use subject to JSTOR Terms and Conditions
TWO DECADES OF FAMILY CHANGE 137 not asked in every interview year. Thus the analytical sample for the early cohort is limited to men who responded in 1976, and to women who responded in either 1978 or 1983, when more complete marital histories were collected. I retain the cross-sectional sample and the oversample of blacks in all surveys. Because of data limitations, the samples are restricted to blacks and whites, with Hispanics coded by race rather than ethnicity.4 To minimize problems from leftcensoring of important economic covariates associated with marriage formation, I further limit the analysis to individuals under age 19 at first interview. The analytical samples include a total of 3,631 women and men in the early cohort and 4,920 women and men in the late cohort. All descriptive statistics are weighted to adjust for design issues, such as the oversampling of blacks, and for differential nonresponse to the interviews (Center for Human Resource Research 1997). It is desirable to have information collected at as many time points as possible for an event-history analysis of marriage. While the Youth sample (NLSY) was interviewed annually from 1979 through 1994, the Young Men's (NLSM) and Young Women's (NLSW) samples were interviewed somewhat less regularly. Wherever possible, the annual measures are constructed from retrospective questions about activities during the noninterview year. This approximation is generally feasible for measures of school enrollment, for example. In other cases, where these retrospective questions were not asked, missing data are generally imputed from the previous interview year.5 A variable indicating a noninterview year for the early babyboom cohort (NLSM and NLSW) is included in all models. Unfortunately, it is not possible to examine cohabitation using the NLS data because, like most surveys conducted prior to the mid- 1980s, the relevant questions were not asked consistently. Data on cohabitation prior to the 1980s that are available from other surveys were primarily obtained retrospectively (Smock 2000), although little is known about the quality of retrospective reporting on cohabitation. The lack of information on cohabitation is a drawback of using the NLS data, but the generally brief duration of cohabitation for most couples should lead to relatively small differences in estimations of the timing of first marriages versus first unions of any type. Indeed, fully half of cohabitations end within 16 months, and only one in ten couples is still cohabiting (and unmarried) after five years (Bumpass and Sweet 1989). It is important to keep in mind, however, that the present study sheds light on just one piece of the broader union formation process-entry into a legal marriage. VARIABLES AND METHODS I use logistic regression analysis to estimate discrete-time hazard models of the effects of economic prospects on entry into first marriage. This approach permits the estimation of effects of fixed and time-varying covariates on entry into marriage, and avoids the assumption of proportional hazards (Allison 1995). The dependent variable in the analysis is a dichotomous indicator of whether a marriage occurred in the interval between two given years, with time-varying independent variables fixed at the beginning of the interval. Data are organized into person-year records, with one record for each annual interval in which respondents were at risk of first marriage, including intervals in which a first marriage occurred. For this analysis, the risk of marriage is assumed to begin at age 17. Because the focus of this analysis is on the comparison of coefficients between cohort models, the models themselves are kept relatively simple, with few complex variable or model specifications. Table 1 shows the mean values of the independent variables, separately by race and sex. This analysis uses three well-established measures of standing in the labor market to reflect economic prospects: earnings, educational attainment, and employment status. Both a continuous measure of logged earnings (wage and business income) and a dummy variable for zero earnings in the prior year are constructed, based on the asI It is not possible to identify Hispanic ethnicity in the NLSM sample, and ethnicity is not ascertained until the 1993 interview in the NLSW sample. 5 A variety of approaches to dealing with missing data were considered, with substantive results robust to choice of method. This content downloaded from 128.103.149.52 on Mon, 12 Aug 2013 19:18:11 PM All use subject to JSTOR Terms and Conditions
138 AMERICAN SOCIOLOGICAL REVIEW Table 1.Sample Means by Race and Sex for Variables Used in Analysis of Transition to First Marriage:Men and Women,Age 22,from the National Longitudinal Surveys of Young Women(NLSW),Young Men (NLSM),and Youth (NLSY) White Women Black Women White Men Black Men Early Late Early Late Early Late Early Late Independent Variable Cohort Cohort Cohort Cohort Cohort Cohort Cohort Cohort Earnings (log) .74 .75 .71 .72 .80 .79 .79 .75 No income .16 .11 .26 28 .03 .05 .06 .11 Educational attainment: Less than 12 years .10 .08 32 .17 .10 .13 .28 .22 12 years .48 .45 .43 38 .45 ,49 13 to 15 years .29 28 .19 .28 .17 .22 16 or more years .12 16 03 05 ,12 .12 .03 .04 Currently enrolled in ,13 19 2 .2 .11 .13 school Currently employed .65 .70 72 .69 .65 .55 Service in military .12 07 .13 .10 Family Background Variablesa Two-parent family at age 14 87 .80 .60 .86 .79 .63 .50 Family head's job is 29 .29 04 7 26 .28 04 .06 professional or managerial Mother's education: Less than 12 years .31 .25 .66 .49 .36 22 68 .45 12 years .45 .48 .14 31 .43 .51 .11 33 13 or more years .20 .24 .05 .12 .16 .22 04 .14 Residence Lives in South .27 .29 .55 .57 .24 .27 .51 .49 Lives in SMSA .70 .69 .73 .73 .61 .63 51 .66 Number of cases 1,203 1.474 517 877 1,440 1,608 471961 Note:Sample means are weighted a Measured at the time of the first interview. sumption that having some earnings is quali- from time spent in school,I constructed an tatively different from having no earnings. additional measure indicating current school Consistent with other national data,the enrollment.A dummy variable indicates em- NLS cohorts show some growth over time in ployment at the time of interview.For men, women's earnings and some decline in men's an additional measure indicated whether re- earnings,particularly among black men. spondents were on active duty in the military Consistent with national trends in education, (these men are coded 0 on the employment some decline over time is seen in the com- variable),as military service has been found pleted education of young white men,likely to impact marital transitions in prior re- reflecting the opportunity to postpone service search.Given the timing of the Vietnam War, in the Vietnam War by extending one's time it is not surprising that substantially more in school (Bernhardt et al.1999:Mare 1995) early-cohort than late-cohort men were en- and the related higher rates of military par- listed in the military at age 22. ticipation among less educated men.To dis- As prior research has found that family tinguish effects of accumulated education background characteristics are related to This content downloaded from 128.103.149.52 on Mon,12 Aug 2013 19:18:11 PM All use subject to JSTOR Terms and Conditions
138 AMERICAN SOCIOLOGICAL REVIEW Table 1. Sample Means by Race and Sex for Variables Used in Analysis of Transition to First Marriage: Men and Women, Age 22, from the National Longitudinal Surveys of Young Women (NLSW), Young Men (NLSM), and Youth (NLSY) White Women Black Women White Men Black Men Early Late Early Late Early Late Early Late Independent Variable Cohort Cohort Cohort Cohort Cohort Cohort Cohort Cohort Earnings (log) .74 .75 .71 .72 .80 .79 .79 .75 No income .16 .11 .26 .28 .03 .05 .06 .11 Educational attainment: Less than 12 years .10 .08 .32 .17 .10 .13 .28 .22 12 years .48 .45 .43 .47 .38 .45 .42 .49 13 to 15 years .29 .28 .19 .30 .31 .28 .17 .22 16 or more years .12 .16 .03 .05 .12 .12 .03 .04 Currently enrolled in .13 .19 .12 .16 .20 .21 .11 .13 school Currently employed .65 .70 .52 .51 .72 .69 .65 .55 Service in military .12 .07 .13 .10 Family Background Variables a Two-parent family at age 14 .87 .80 .60 .45 .86 .79 .63 .50 Family head's job is .29 .29 .04 .07 .26 .28 .04 .06 professional or managerial Mother's education: Less than 12 years .31 .25 .66 .49 .36 .22 .68 .45 12 years .45 .48 .14 .31 .43 .51 .11 .33 13 or more years .20 .24 .05 .12 .16 .22 .04 .14 Residence Lives in South .27 .29 .55 .57 .24 .27 .51 .49 Lives in SMSA .70 .69 .73 .73 .61 .63 .51 .66 Number of cases 1,203 1,474 517 877 1,440 1,608 471 961 Note: Sample means are weighted. a Measured at the time of the first interview. sumption that having some earnings is qualitatively different from having no earnings. Consistent with other national data, the NLS cohorts show some growth over time in women's earnings and some decline in men's earnings, particularly among black men. Consistent with national trends in education, some decline over time is seen in the completed education of young white men, likely reflecting the opportunity to postpone service in the Vietnam War by extending one's time in school (Bernhardt et al. 1999; Mare 1995) and the related higher rates of military participation among less educated men. To distinguish effects of accumulated education from time spent in school, I constructed an additional measure indicating current school enrollment. A dummy variable indicates employment at the time of interview. For men, an additional measure indicated whether respondents were on active duty in the military (these men are coded 0 on the employment variable), as military service has been found to impact marital transitions in prior research. Given the timing of the Vietnam War, it is not surprising that substantially more early-cohort than late-cohort men were enlisted in the military at age 22. As prior research has found that family background characteristics are related to This content downloaded from 128.103.149.52 on Mon, 12 Aug 2013 19:18:11 PM All use subject to JSTOR Terms and Conditions