If we want to look at the impact of a given exposure on depression outcomes, we commonly require the assumption that the missingness in our depression indicator is "at random". "At random" is something of a misnomer here, and means "random conditional on measured covariates". In the instance that missingness is in fact not at random, i.e. depression itself affects participants likelihood to respond - then common analytical procedures to investigate the impact of exposures on depression may be biased.

We will develop a method to attempt to address this gap, using the effect of smoking on depression at 18 as a question to demonstrate our proposed approach, which will seek to provide testable conditions under which we can falsify the "missing not at random" assumption.