Single stochastic imputation in Stata - how to?

[Pavel Roshanov]

Hello Statalist,

I have some analyses that would be computationally and practically challenging with multiple imputation and believe single stochastic imputation (conditional mean + random noise) will be a reasonable second-best alternative to MI. There seems to be a lot written about MI in Stata but I can't find a way to do single stochastic imputation (other than the retired -impute- command). Is single imputation possible with MI? Would it be equivalent to just doing MI with 1 iteration?

Thank you!
Pavel

[Stephen Jenkins]

Would it be equivalent to just doing MI with 1 iteration?

Strictly speaking: probably, yes. To me, the more important question is: what imputation model would you want to use? Use of mi means that you are buying into the imputation models that it provides (which may be fine in your case of course). Alternatively, you can simply roll your own. You refer to "conditional mean + random noise", so could use your regression model for the first bit and a random draw from the distribution of residuals (assuming they're normally distributed with s.d. equal to the estimated value). Ensure that you use set seed beforehand to ensure reproducibility. My recollection is that this approach is only one of many of this type. You could look at the survey literature on imputation: I think the (single) imputation methods used by data producers may help guide you -- see methods such as "predictive mean matching", and so on.

[Pavel Roshanov]

Thank you. Indeed I'm looking at semi-parametric methods like predictive mean matching for imputing continuous variables.