Depending on why and how you want to use the results of the ITT analysis, you could just replace each missing value by the "worst", i.e. most conservative possible value, whatever that is in your specific trial. Somewhat more sophisticated, you could model the missing values taking the reasons for their missingness into account. To find out which model works best, one can introduce artifical missingness into the complete cases according to the assumed MNAR mechanism, and then predict the missings using different models and compare the predictions to the original values. Finally, it seems that multiple imputation works OK in many cases even if the missingness pattern is MNAR. The following quite recent article points into this direction and might offer additional insights: Article Imputation of data Missing Not at Random: Artificial generat...