I conducted a between-subjects experiment with one 3-level factor (high group vs low group vs control group). Because of dropouts from the treatments, the final distribution is
control group: 85 subjects
low group: 58 subjects
high group: 70 subjects
What conditions are required to recover an unbiased effect estimate? What methods or solutions can I adopt?
Dropouts are exceedingly difficult to handle correctly, having the possibility of creating a bias so large that results are uninterpretable. A common mistake is to assume that you can enroll more subjects to make up for the dropouts. This is only true if the dropouts are completely random events. They seldom are.
Perhaps the only simple interpretation comes from treating dropout as a bad outcome. This can sometimes be elegantly handled by using an ordinal outcome where the need for dropout has its own category at one end of the scale. Ordinal regression will not need to assume that the spacing of the categories are correct, only that the ordering of categories is.
An alternative is to have frequently-measured longitudinal response variables such that for dropouts the last-measured response captures the worsening trajectory before the dropout. This would make the dropouts creating “missings at random” so that you could analyze all the observed serial measurements using standard full-likelihood longitudinal models (serial correlation-based generalized least squares, Markov models, mixed effects models).
