1985: Compensating for Wave Nonresponse in the 1979 Isdp Research Panel

The choice between weighting adjustments and imputation for handling missing survey data is generally straightforward: as a rule, weighting adjustments are used for total nonresponse and imputation is used for item nonresponses. There are, however, several situations where the choice is debatable. In general, these are situations of what might be termed partial nonresponse, where some data are collected for a sampled unit but a substantial amount of the data is missing. These situations include cases where the respondent terminates the interview prematurely, where data are not obtained for one or more members of an otherwise cooperating household (for household level analysis), and where an individual provides data for some but not all waves of a panel survey. If weighting is used for partial nonresponse, the available responses for that unit may be employed in the determination of the weights, but the unit itself is discarded, resulting in a loss of data. On the other hand, if imputation is used, a sizeable number of responses for a partially nonresponding unit will need to be imputed, giving rise to concerns about the fabrication of much of the data and the effect of this fabrication on the relationships between variables. This paper examines the choice between weighting and imputation for handling the partial nonresponse that occurs when a respondent fails to provide data on one or more waves of a panel survey. Kalton (1985) provides further discussion of the issues involved in choosing between weighting and imputation t~ handle wave nonresponse, and Cox and Cohen (1985) report the results of an experimental investigation of these alternatives in the National Medical Care Expenditure Survey. The objective of this study is to provide evidence on the choice between weighting and imputation for handling wave nonresponse in the Survey of Income and Program Participation (SIPP). The SIPP is a panel survey in which households are interviewed every four months over a period of about two-and-a-half years (Herriot and Kasprzyk, 1984). One major product of the SIPP will be an annual file combining three waves of data, and the focus of the present study is on this annual file. Since a longitudinal file for the first three waves of the first SIPP panel is not yet available, the empirical investigation reported here is based on the first three waves of the 1979 Income Survey Development Program (ISDP)