Data Manajemen | Fakultas Ekonomi Universitas Maritim Raja Ali Haji 698.full

rect for small sample bias in different ways. Third, unlike BDM, I use variation across groups defined both by the type of husband and the type of wife and so I exploit vari- ation that arises because conditional on characteristics of the wife, there are pre- dictable differences in wage growth of husbands that are explained by husband characteristics such as age and education. This addresses the possibility that declin- ing wages of men occurring alongside increasing hours of women may be purely a demand-side phenomenon as suggested by Juhn and Kim 1999 rather than resulting from compensatory labor supply behavior. Fourth, I use regional variation in relative wage changes and labor supply behavior to relax the assumption that any changes in unobserved factors over time are the same for all education and age groups.

III. Data

To analyze wage and labor supply outcomes at the regional level for detailed demographic groups, I use data from the Public Use Micro Samples PUMS of the Census of Population in 1980 and 1990. 1 I use the 5 percent samples in each of these years. Such detailed analysis would not be possible using alternative sources of data such as the Current Population Survey CPS because they do not have nearly as many observations. I restrict the sample to individuals who are aged between 21 and 60. Persons in school or the military during the survey week are omitted as are persons who are liv- ing in group quarters. I also exclude all cases where the husband or wife has self- employment income. The wage measure used is average hourly earnings where earnings include wage and salary earnings only. Earnings are topcoded at 75,000 in 1980 and 140,000 in 1990. Fewer than 1 percent of observations are topcoded in 1980 and the same is true in 1990. I replace the topcoded value by the topcode times 1.33 in both years. I deflate earnings using the personal consumption expenditures deflator. 2 I omit couples in which the husband does not work at all in the previous cal- endar year. With the exception of older men, participation rates are very high for mar- ried men in 1980 and 1990 so this is not a huge restriction. Both participating and nonparticipating women are included in the sample. I discuss the selection issues aris- ing from nonparticipation in the next section. The labor supply measures refer to the hours worked in the previous calendar year and the wage measure is the log of average hourly earnings in the previous calendar year. I include three labor supply measures. The first, Hours, is annual hours worked. The second, Full Time ⎪Full Year FTFY, is an indicator for individuals who worked at least 50 weeks and usually worked 35 or more hours per week. The third, Annual Devereux 699 1. I do not use the 1970 census because there is no information on usual weekly hours last year, and weeks worked is bracketed. Hence, some strong assumptions are required to calculate average hourly earnings and hours worked. See Pencavel 1997 for more detail on the drawbacks of making these assumptions. Furthermore, the available sample sizes are inadequate for the approach taken in this paper. 2. I do not include cohabiting couples as these can not be distinguished from roommates in the 1980 Census. Bumpass and Sweet 1989 report that cohabitation is a short-lived state and has a median duration of 1.3 years. Thus, cohabiting couples are a small proportion of all couples. Also, one would expect the work behavior of cohabiting couples to be less interdependent because of the short expected duration of these unions. Participation , is an indicator for people who worked for even one hour in the year. Like Juhn 1992, Welch 1997, and Devereux 2003, I make no distinction between periods of nonemployment that are classified as unemployment and periods spent out of the labor force.

IV. A Specification for Family Labor Supply