Multiple imputation

[Adeola Kolade]

Dear all,

I will please like to know how to carry out Instrumental variable post estimation tests (e.g. test of endogeneity , test of weak instrument) when data is imputed.
Also can the results from IV be trusted when data is imputed?
I will appreciate your replies. Many thanks.

[Fei Wang]

Adeola, this is technically easy but conceptually difficult. You may just add a post option to mi estimate to conduct a post-estimation analysis, but results may be nonsense as many such analyses are even not well defined with imputation.

To my understanding, mi can be used when missing is random or at least random conditional on observables -- a condition that is hard to test but is easily violated in practice. So myself never impute data. If observations are missing at random, nothing is wrong to use the non-missing sample as long as the sample size is large enough. If observations are non-randomly missing, mi may bring about larger biases; in this case, I would instead think about methods of correcting endogenous sample selection biases when necessary.