1
$\begingroup$

I want to do predictions of missing data MAR. I do the simulations by generating normal data with 1 y and some x. after that I remove the data: Proportion of missing value from 5%, 10% to 50%. I use imputation mice. I would to ask, how much imputation should I do for each of the missing data proportions. I have not found literature about it. I just get the MCAR, the amount of imputation is as much as a lot of missing data.

$\endgroup$

1 Answer 1

1
$\begingroup$

If you are wondering how many records to impute per imputation sample, then the answer would be as many as there would have been without missing data. If the question is on the number of imputed datasets then the answer is more is better. Just a few (3-5 is often quouted) may be enough to achieve coverage probabilities at or usually above the nominal level etc., but with more you get higher efficiency (narrower CIs, more power) and less dependence of the results of the random number seed. Unless it's computationally a problem, I normally like 1000 or more imputed datasets for the types off problems I work on (for which simulation studies suggest 250 are often good and 1000 a tiny not better and the computational cost of going to 1000 is pretty negligible). With a lot of missingness (50% is certainly a lot), you may want to go even higher, if feasible.

$\endgroup$
8
  • 1
    $\begingroup$ I read in the literature "Multiple imputation using chained equations: Issues and guidance for practice" written by white et al, 2010. They said that for MCAR, the number of imputations will be equal to the number of missing data. Could you give me the literature for the MAR mecanism as you said above? thank you $\endgroup$ Oct 22, 2017 at 9:38
  • $\begingroup$ Imputed records or imputed complete datasets? $\endgroup$
    – Björn
    Oct 22, 2017 at 9:40
  • $\begingroup$ I work now with incomplete datasets, not complete datasets. $\endgroup$ Oct 22, 2017 at 9:55
  • $\begingroup$ Not sure how you mean, I took that as a given. What number are you worrying about? $\endgroup$
    – Björn
    Oct 22, 2017 at 9:57
  • $\begingroup$ I search the literature which say how many the number of imputation (m) in mice algorithm for MAR mecanism. is that the same with MCAR mecanism that the number of imputation (m) is be equal to the number of missing data. for example, I do simulate the proportion of missing data 40% in my datasets, and I use m =40 for the method MICE. and now I have not found teh literature that how many the number imputation (m) for mice algorithm in MAR mecanism. $\endgroup$ Oct 22, 2017 at 10:04

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.