Simulation code to accompany my blog post on whether multiple imputation invents information. The description of the simulation study is in that blog.
Note: code by me is now under "Stata" because @Elessenne contributed R code under "R". He used 10,000 repetitions, which was wise: although I said I was comfortable with the Monte Carlo error after 800 reps, this was WRT the reassurance about MI, and I subsequently undermined myself by saying that I was not sure if differences were just Monte Carlo error.
A note on the Stata code: I simulated all the full data in one go and saved it as sim_data.dta. For each rep, I then used the relevant chunk of data. Initially I tried to do mi impute … , by(rep_id) on sim_data.dta, and was pleased with that idea, but Stata hung when it was trying to append all the imputed datasets – shame! So the full dataset is saved in the repo but the imputed datasets are not.
The community-contributed {simsum} Stata command is needed to run this code available here.
Caution: do not use the Stata code in this do-file as a model of a good simulation study. I did it offhand for a blog post and would be more structured and careful for a paper. See here for how to do it better.