Caveats when modelling dropout

Modelling dropout is a difficult procedure for the simple reason that there is typically not a lot of information in the data to identify dropout parameters. As a result, some warnings need to be made.

Firstly, be aware that it can be extremely difficult to identify parameters in the dropout model, and likelihoods can easily be very flat and/or multimodal. Take care to ensure that the algorithm has indeed converged to the maximum likelihood, by experimenting with different initial estimates and changing the tolerance parameters of the maximising procedure.

Secondly, intermittent missing values (i.e. within-series missing values which are subsequently followed by a non-missing value) are handled in an unusual manner when the dropout process is modelled. When the dropout model includes a regression on previous observed values, intermittent missing values are ignored in the sense that the `nth previous' value is defined as the `nth previous non-missing value'. This has consequences in the interpretation of the dropout parameters $\theta_2, \theta_3, \ldots$ when intermittent missing values are present.