Female income and the divorce decision: evidence from micro data

Abstract

Escalating divorce rates during the 1960s and 1970s led to large numbers of academic investigations into the causes of divorce. Most of these studies concentrated on a significant increase in female income that resulted from rising female labour force participation rates. The difficulty with quantifying these arguments is that it is possible to observe the income of married females or it is possible to observe the income of divorced females, but it is not possible to observe both outcomes, simultaneously. This research attempts to resolve these difficulties by using sample selection techniques to correct for possible bias from simple observation of the income of married and divorced females.

Acknowledgements

The corresponding author would like to express appreciation to the Amsterdam Institute for Advanced Labor Studies (AIAS) and its able staff for support of this research.

Notes

¹ For a very good explanation of this technique see Greene (2000).

²Vella (1998) provides an excellent discussion of these issues.

³ Full results will be supplied on request.

⁴ Full results will be supplied on request.

⁵Stanley and Jarrell (1998, p. 963) in their meta-regression on gender wage discrimination point to the importance of job classification variables that specify governmental employment in the estimation of female earnings.

⁶ Full results will be supplied on request.

⁷ Full results will be supplied on request.

⁸ Additional insight on this issue is provided by Ermisch and Wright (1994).

⁹ An anonymous referee requested that in addition to the series of state dummies an additional series of variables be investigated – interaction variables between year and state. This created an unacceptably large number of variables for the estimation procedures. So, two alternatives were tried. First, a probit procedure with divorce as the dependent variable and just the year and state dummies was estimated. Then, a series of interaction variables between one year and the states was added. Neither probit was able to correctly predict even a single divorce (0.5 was used as the threshold for prediction). In addition, the Hosmer–Lemeshow statistic recommended against both specifications. The same results were produced for the final two series of interaction variables. Finally, a similar procedure was attempted for the structural probit that included the state dummies. The interaction variables (added separately, one year at a time) had no effect on the signs or the levels of significance for the difference variables. They also had almost no effect on the Estrella statistic and were recommended against by the Hosmer–Lemeshow statistic.

¹⁰ The coefficient estimates and t-scores for the state dummy variables are omitted from Table 6 because of concerns for space. Full results will be supplied on request.

¹¹ Washington, D.C. was included in the estimations as a state and Kansas was excluded. Thus, 50 coefficients were estimated.

¹² In addition, several other goodness of fit measures all indicate that either of the “structural” probits provides superior statistical results to the estimates produced by the reduced form equation.