non-random attrition in the sample

[himani aggarwal]

I am running a field experiment with three treatment arms in a multiwave frame. Attrition is non-random for the final wave and hence I am concerned about how to deal with this in my regressions?

[Carlo Lazzaro]

Himani:
it's quite weird that your data are MNAR in the final wave of data only; usually (and regardless the missingness mechanism), missingness in longitudinal studies follows a monotonic pattern (that is, data missing in the last wave of data were, in all likelihhod, missing (at least in part) in the previous wave(s), too).
That said, missing non at random (MNAR) is, as always, a tricky issue.
Dealing with MNAR data boils down to a wide use of sensitivity analysis, aimed at proposing dfferent missingness scenarios.
However, whenever you suspect the presence of MNAR data, the literature (https://www.crcpress.com/Flexible-Im...9781439868249: page 182) advises to rule out once and for all that you cannot turn MNAR in "more MAR" data, by tracking down relationships between the missing data and observed variable, even though the latter were not included in your original model. When successful, the benefits of this attempt are clear: if data are MAR you can rely upon the superb -mi- routine in Stata, without banging your head against the wall in search of a good (usually bayesian) model to create different probability of missingness for MNAR data.
Another interesting literature sources for dealing with missing data in longitudinal dataset is: https://www.crcpress.com/Missing-Dat.../9781584886099.