What is the effect of predicting in the period after 1975, using the B-W coefficients fitted on the years 1947-75, a base of 28 years for forecasting?
This underprediction in the period since 1974 appears to be due in part to a failure of the predicted values to "bottom out" until approximately three years after the actual series had done so.(10) This is a common feature of the B-W model using its original data (note the lags mentioned in the B-W quote above), although it is more marked in the 1975-78 period.(11) An analysis of the model's residuals demonstrates a strong pattern of serial correlation (as indicated, too, by its low Durbin-Watson scores).
In order to obtain more accurate measures of the variables used by B-W in their model, time series were prepared using the March Current Population Survey (CPS) micro-data.
B-W were forced to approximate the patterns for married persons using data for all persons, and to approximate the income of husbands of women in a particular age group by taking weighted averages of the incomes of all men in related age groups.(12) The methods used to derive the CPS data are described in the Appendix.
Because there were no available time series on the female hourly wage, B-W were forced to make a number of approximations regarding annual earnings, weeks worked per year, and hours worked per week, in order to estimate one.