MI for outcome

[Amanda Persson]

Hi! I have not found information about MI for an outcome,

My data consists on a survey of aprox 3800 individuals and 75% did not reply the questionnaire.

I am wondering if there is any way for inputting missing values in my outcome, is a continuous outcome.

Then I would like to apply a multiple regression, I am also wondering if I focus on the 25% of non missing values for my outcome, then apply a regression and finally, using the predictions, I will predict for my missing sample, and after that, run the model again with these missing values converted to "predicted values" is a possible or not?

tips?

Thank for your time,

[Felix Bittmann]

Unless you have strong predictors of your outcome available, imputing it will not help much. These must be variables without missing information that correlate strongly with the outcome (but need not to be part of your later analytical model). And 75% missing is a really high share. This seems highly problematic. Technically, you can impute an outcome like any other variable in Stata.

[Rich Goldstein]

first, if your predictors/covariates have no missing data, and if your outcome is either missing completely at random (MCAR) or missing at random (MAR), then imputation will give you essentially the same answer as a complete case analysis (but, depending on your N, the SE's might differ just because of difference in N); so, in this specific case, there is no reason (except possibly for power considerations) to impute

second, if you estimate without imputation, and there are no missing values for your predictors/covariates, then -predict- will give you predictions for those with missing outcomes as well as those with complete data

third, if your outcomes are missing not at random (MNAR), then you have a much more complicated situation and you will need to supply more information to get specific help (but with 75% missing you are probably in trouble in this situation and will need to use sensitivity analysis on your estimates