Sources of bias in women’s wage equations

Wage equations estimated for samples of women are used frequently in labor economics. Often the wage equation estimates are the direct objects of interest. Examples include: estimating wage discrimination via “decomposition” techniques; documenting and explaining the rise inwomen’s earnings relative tomen’s; testing theories explaining wage differentials between men and women; and estimating effects of demographic decisions or outcomes on wages.1 In addition, estimates of women’s wage equations are used as initial inputs in studying other questions. Examples include: constructing instrumental variables forwages to estimate labor supply parameters (e.g. Nakamura andNakamura, 1985a; Mroz, 1987); computation of wage values in microanalytic simulation models (e.g. Orcutt and Glazer, 1980); and legal applications regarding pay discrimination or computations of potential earnings (e.g. Bloom and Killingsworth, 1982). Much of this research uses standard Ordinary Least Square (OLS) estimates of

variants of the human capital earnings function. But much of this research also recognizes the potential for biases from endogeneity of the regressors, and from unobserved heterogeneity associated with the regressors. Different researchers address alternative sources of bias, using varying identifying assumptions. At the same time, some research papers in this area continue to useOLS estimates ofwage equations for women (e.g. O’Neill and Polachek, 1993). In general, our reading of the literature on women’s wage equations, summarized in the section on Past research: an overview, suggests that there is no consensus regarding the empirical importance of these various sources of bias, nor the validity of the assumptions used to attempt to correct for them; as a consequence, there is no consensus regarding the treatment of these alternative sources of bias. In this chapter, we utilize data on sisters in an attempt to provide a more uni-

fied analysis of sources of bias in women’s wage equations. Data on sisters offer advantages for estimation of wage equations correcting for endogeneity and heterogeneity bias; these advantages are spelled out in the section on Advantages of using sibling data. The chapter shows how the estimated effects of marital status, number of children, labor market experience, and schooling differ depending on the source of bias considered, and the identifying assumptions used to correct

for the bias. It also exploits the sibling data to test a variety of overidentifying assumptions. Our goal is to contribute toward building a much-needed consensus in the statistical approaches to estimating women’s wage equations.