Although the effect of self harm among those without alcohol related problems may be estimated simply by excluding those with pre-existing harmful alcohol use, extending this approach to incorporate multiple imputed data is not trivial. Therefore utilising a series of appropriately parameterised interaction models we obtained the same estimates. Here we incorporated our flags for pre-existing problems as a series of moderator variables both within the imputation step and within the analysis that followed. By building interaction terms into the imputation routine the effect of self harm on a particular outcome is allowed to differ between cases with and without pre-existing problems. In addition, the ability to calculate the estimates of interest without excluding cases meant that the model estimated within each imputed dataset was based on the same sample size, irrespective of the flagged sample that would be expected to vary in magnitude owing to missing data in these measures. Given the number of interaction terms required, we derived a separate imputation model for each outcome in turn.
