Adjusting for Attrition in Event-History Analysis

Daniel H. Hill *

The sensitivity of parameter estimates of event-history models to alternative methods of correcting for panel attrition is not well understood. This paper will investigate the issue of weighting for panel attrition in event-history models by comparing alternative treatments of sampling weights in a divorce model for members of the 1986 Survey of Income and Program Participation (SIPP). Three distinct weighting procedures are compared. These procedures are based, on (1) the initial selection probability weights; (2) the 1986 SIPP panel weights; and (3) the monthly attrition-adjusted weights. The paper also compares these weighted estimates with the estimates of a structural model in which attrition is treated as an error-correlated competing alternative to divorce. Although it is impossible to identify a “best” procedure without accurate external data, significant differences in the estimates for the various procedures are indicative of significant attrition related problems in event-history models. None of the weighting adjustments are found to have any appreciable effect on the estimates of the divorce hazard model examined. The reason is that all of the weighting procedures are based on the assumption of independent censoring. The competing hazards structural model relaxes this assumption and finds evidence of significant correlated unmeasured heterogeneity. Once corrections for this are made, the net divorce hazards are seen to increase by more than one-half. This suggests that in many instances divorces in the SIPP end up being recorded as attrition.

* University of Michigan