A model-based appproach to comparing imputation techniques for incomplete panel data.
by
Michael J. Cleveland
Iowa State University
Coauthors: Frederick X. Gibbons, Meg Gerrard, K.A.S. Wickrama, Jacques D. Lempers

Almost all areas of empirical research are prone to problems associated with missing data. One common method of dealing with missing data is to simply exclude any case that has missing data for any of the variables in the analysis (listwise deletion). However, a major concern with the use of listwise deletion is that it assumes that the data are missing completely at random. Under violations of this assumption, listwise deletion may result in incorrect results, leading to biased parameter estimates and inefficient standard errors.

Fortunately, new methods for dealing with missing data have recently been developed and are becoming more common in the social science literature. In this paper we have incorporated a model-based approach to missing data, which allows one to explore the missing data mechanism in longitudinal data and compensate for bias due to missing data. We then used a latent growth curve analysis to assess the bias and efficiency of parameter estimates obtained from listwise deletion and several different imputation techniques. These results suggested that the use of listwise deletion resulted in substantial differences in the magnitude of the growth curve parameters, compared to those obtained from imputed data. No clear differences were found between the different imputation techniques, however.

Date received: August 29, 2003