4

I'm starting to prefer visualizations of my regression models as opposed to tabular output (OR's, beta-coefficients, 95%CIs). However, I struggle to find a good way to do this when I am undertaking multiple imputation by chained equations (mice). The output of mice (in R) is usually one data frame containing m complete datasets after m imputations. With this, I can run the same model on each of the m complete datasets and then pool the results according to Rubin's Rules. In R there are nice functions to do this with the mice package. But no really good way to plot this output.

My current strategy is just to pick a random imputed dataset out of the m complete datasets and then do the visualization on that one dataset (and model).

Does anyone out there have ways they typically check these models (with plots) or visualize the data (in exploration)? I'm open to creative suggestions. They need not be in R, however, then maybe you could show me the types of plots you make?

RNB
  • 506
  • 5
  • 12
  • 2
    I am very interested in your question and posted a similar one about calibration plotting some weeks ago (http://stats.stackexchange.com/questions/265985/pooling-calibration-plots-after-multiple-imputation; alas, no answer as of yet). As to your question, randomly picking a dataset to plot is indeed considered a viable option, but I guess you'd also like to account for the uncertainty of imputation. Which plots are you interested in? – IWS Apr 05 '17 at 14:54
  • This might be a case in which plotting, as a data graphic per se, the output from analysis of multiple imputed datasets is less desirable than diagramming the process. – rolando2 Apr 05 '17 at 15:40
  • @IWS, I guess I'm just wondering if randomly selecting one of the datasets to create plots for is valid? As you point out, i'm worried because a single plot doesn't show the uncertainty. I've also thought about collecting all the B-coefficients from the models and then plotting them somehow? – RNB Apr 06 '17 at 07:23
  • 2
    I could not point you to references which would back the following, but here goes: If all is right with the imputation technique, all imputation datasets created are completed based on the information inherent in your data while adding/subtracting some randomness. So, if you have to pick a dataset to plot data, and *truly* randomly pick an imputation dataset, dataset number x and its plots are as good as any other. However, if you start cherrypicking (knowingly or unknowingly), this is a bad idea as you might just pick the one dataset with the most favourable results. Does that help? – IWS Apr 06 '17 at 08:30

0 Answers0