A test of the missing data mechanism for repeated categorical data.

Due to the occurrence of missing observations, longitudinal data are rarely balanced and complete. Weighted least squares analyses described by Grizzle, Starmer, and Koch (1969, Biometrics 25, 489-504) have been developed for the analysis of incomplete longitudinal categorical data [Stanish, Gillings, and Koch (1978, Biometrics 34, 305-317); Woolson and Clarke (1984, Journal of the Royal Statistical Society, Series A 147, 87-99)]. However, all these analyses have assumed that missing observations are missing completely at random in the sense of Rubin (1976, Biometrika 63, 581-592). When the occurrence of missing observations is related to the unobserved response values, these analyses may result in biased results. In this paper, we develop a simple and practical test of the missing mechanism in incomplete repeated categorical data. The proposed test is an extension of the test of Little (1988, Journal of the American Statistical Association 83, 1198-1202) and uses a test criterion given in general form by Wald. The test is illustrated using data from a longitudinal investigation of obesity in school-age children.