On informative and random dropouts in longitudinal studies.

Shih (1992) correctly points out that my discussion of the distinction between informative, random (MAR) and completely random (MCAR) dropouts fails to consider explicitly issues of model parameterisation (Diggle, 1991). However, I did use a notation which was intended to make clear that I was assuming separate parameterisations of the dropout process (parameter ,B) and the measurement process (parameter 0), thus satisfying Shih's DP (distinct parameters) condition. The implications of the violation of DP are discussed explicitly in Diggle and Kenward (1994). Shih goes on to suggest: "It may be possible that in certain situations when MAR holds without DP, ignoring the missing data process still leads to consistent but inefficient estimates." In fact, ignoring the missing data process in such cases always leads to consistent estimates. Under MAR and DP, the log-likelihood partitions as L(f3, 0) = LI(3) + L2(0), where Ll(.) involves only the data on dropouts and L2(A) involves only the measurement data. '1 ~~~~~~~~~~~~~~~~~~~~~0 ^ S ' 3 . . . ~~~~I U * ** * O-~ ~ ~ * . ,, O o ~ ~~~~~~~~ o i l a 2 4 6 1 Figure~~~~~~~~~~~~~~~~ 1. Siuae aa(os n miia en tec aue ft(once iesget) Corlto paaee p . 9. * S B~ ~ 94