0
$\begingroup$

So far I read that with multi-imputation, I put the new values of each set into a different data, run the training and than vote for a final results, like in the following schema: enter image description here

However, what should I do if I want to further engineer the imputed data? It sound unlikely to run each line of code multiple times for each dataset.

$\endgroup$
  • $\begingroup$ I do not understand what you mean by feature engineering. Could you elaborate? Additionally, the pooling of multiple imputation datasets is not done by voting into a 'single' completed dataset. On the contrary, as unlikely as it may sound, the power of imputation is obtained by running the analysis of interest within each imputation set and consequently pooling the results of the analysis according to predefined pooling rules (i.e. Rubin's Rules). $\endgroup$ – IWS Sep 12 '17 at 14:46
  • $\begingroup$ Under "feature engineering" I mean editing the existing variables and/or creating new ones based on them, for example: multiplying all the imputed variables by 100. $\endgroup$ – Riddle-Master Sep 12 '17 at 15:05
1
$\begingroup$

It might "sound unlikely to run each line of code multiple times for each dataset," but that would be the way to proceed.

Multiple imputation provides an important advantage over single imputation, as it directly incorporates the uncertainty in the imputation process. So if you want to take advantage of that benefit of multiple imputation, then you need to go through your further "engineering" of each imputed data set independently. That might not end up being a big practical problem, as depending on your application a dozen or so imputed sets can sometimes be adequate.

The above assumes that the further "engineering" depends on values that you impute; if "engineering" can be done on the incomplete data, then that would be OK before imputation. But if you hold off on "engineering" until after results are pooled, the step from "analysis results" to "pooled results" would not incorporate the uncertainty in the engineered values.

$\endgroup$

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.