Missing Data/Imputation Discussion > Significance level and MI
Hi Sam!
That is an excellent question! I don't think using a more conservative alpha (less than .05) is necessary when using MI. Even though you are doing multiple analyses (which I'm guessing is why you thought it would necessary), the results are pooled, so that should ameliorate any inflated risk for getting false positives when doing multiple analyses. If someone was applying MI incorrectly and just looking for significance in any one of their imputed datasets, then I would agree that an adjustment would be necessary, but as long as the results are properly pooled, then I think using .05 if fine. Does that make sense?
May 1, 2012 |
Jeremy Taylor
Jeremy Taylor
Hello,
I am responding to a published article that uses MI to fill in missing data from a cohort study. Missing data for lung function parameters are derived from 50 imputed models (47% cases have missing data). The subsequent analysis uses a p<0.05 significance level, however I feel like this should be more stringent given the amount of imputation needed, similar to adjustments for multiple testing.
Is this the case?
Thanks
Sam