Intention to treat estimates

[Annie Mueller]

Hello everyone,
I was wondering if anyone has ever constructed intention to treat estimates with Stata? Is there a certain command or what is the easiest way to construct them?
Thank you in advance!

[Rich Goldstein]

I'm confused: for me, intention to treat (ITT) comes from a randomized study in which the ITT analysis keeps all subjects in the group they were originally randomized to; you appear to mean something different; please either clarify what you mean or provide a citation

[Clyde Schechter]

In some subfields of medical research, the term "intention to treat" has taken on a new meaning, tangentially related to the original one that Rich refers to. Whereas the original meaning of ITT refers to the analysis of trials in which some people don't follow their randomly assigned treatment, the new meaning refers to situations where people drop out of a longitudinal study and no end-of-study outcome data are available. In these situations, some analysts impute some kind of "no improvement" or "worst case scenario" outcome to these people and analyze them that way. It really has almost nothing to do with intention to treat in the original sense, but the phrase seems to be catching on.

I'm not sure what I think about this kind of analysis (although I do think it deserves a more descriptive name!) Single imputations have real problems with regard to variance estimation. On the other hand, the kinds of studies in which this is typically applied are ones in which you can't say the data are missing at random with a straight face, but you also can't really specify a missingness model. I sometimes do this kind of analysis myself, and usually just refer to it as a sensitivity analysis for the effects of missing data. When I do this, I also do other types of single imputations for the missing values, such as a "best case scenario" or a "last observation carried forward" scenario. I don't know of any truly satisfactory solutions to this rather common problem.

In any case, I don't think there are any canned commands for doing this. You have to first decide on what rule you want to use to impute the missing end of study values. Then write the appropriate -replace- commands, and run your analysis.