For every complex problem there is an answer that is clear, simple, and wrong.
H. L. Mencken
Missing data in psychiatric trials is the rule, it is one of the key elements which makes the summary and analysis of psychiatric longitudinal data fraught with difficulty.
A common, quick fix solution is to apply the last observation carried forward (LOCF) algorithm. This method is computationally simple to impute and easy to understand. Yet the effect it introduces into the results is unpredictable, it is commonly thought to bias the results in a conservative direction thus safeguarding against false claims of efficacy. Extensive simulations have revealed this is not the case, LOCF can bias the result in either direction and it’s difficult to foretell what effect it will have.
Consider a patient who is lost to follow-up after their 3rd visit (at Week 2). The 4 panels above show what their 6 week outcome may have been. Patients tend to move in a given direction over the course of time but they zig and zag their way to the final visit; a combination of signal (the overall trend) and noise (the zig-zag pattern). LOCF assumes no change whatsoever, this has the effect of creating an artificially smooth journey by removing this “within-patient” variation. Any methods which destroy, ignore or smooth over the natural variation in psychiatric data are problematic and if possible should be avoided.
There is no cure for missing data. Make all efforts to keep patients in your trial, follow them up even if they stop taking treatment or switch treatment.
Report clearly the extent of missing data in your trial, give an appraisal of what effect this may have, consider more than one imputation and/or analysis strategy, try to gauge the sensitivity of your findings to different assumptions about the missing data. Avoid LOCF for the primary analysis or if you insist on using it, justify this choice in the Methods section.
This is a big topic…