1
$\begingroup$

I know that this question has been asked quite a lot - but I did want to see what people's opinions currently were on how applicable Multiple Imputation (MI) is to perform on a dataset with a high missing rate.

The dataset I have this issue with contains ~15,000 cases (patients in a database). Each case has 5-6 variables of interest that I will be inputting as covariates in a Cox-PH model. Two of these variables have a high missing rate, reaching roughly 75% each. The pattern of missingness has been identified as MAR. I plan to use 100 imputations for the MI procedure.

From what I have read - there is no hard-n-fast rule as to how high a missing rate is tolerable. However, I was hoping that someone may be able to guide me on some, perhaps, diagnostic statistics (or references) I could interpret that may help justify performing MI on this dataset (or not!)...

Improve this question
$\endgroup$
1
  • $\begingroup$ How would you know the missingness is MAR?Presumably just an assumption? 75% certainly feels very high, but what would tolerable mean? There is never going to be a threshold below which we just impute and do not worry, while above it we do not even bother to impute and do not analyze the data. The higher the proportion the more I worry and get tempted to do sensitivity analyses. $\endgroup$
    – Björn
    Commented Jul 19, 2016 at 4:45

1 Answer 1

Reset to default
1
$\begingroup$

I've had this question myself and agree that there isn't much, if anything, in the literature to help guide a decision. You haven't told us what type of information is missing: is it lesser priority information such as demographics or information that is hypothetically critical to your analysis? If the former then no worries about a high level of missingness, if the latter, then special care should be taken in monitoring the results.

There are a few critical outcome metrics to monitor in any MI. Probably the most important metric is the extent to which the plugged information diverges from actual values in a holdout sample, however generated. Next, are the average plugged results consistent with the pre-plugged marginals across all data? A third useful metric is to compare the coefficient of variation pre and post. Together, this information informs the extent to which the plugged data is behaving more or less like the pre-plugged data.

One thing Rubin discusses that has been helpful is to order the MI models by the amount of missing information. This means that you plug those features with the least missing data first based on the full information features -- including the dependent variable and any additional information besides the key covariates. As each feature is plugged, add the newly plugged feature, step by step, as a predictor in the subsequent MIs.

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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