What is the current research on missing not at random data?

Question

I have a data set with N ~ 5000 and about 1/2 missing on at least one important variable. Comparing the people with missing data to those with complete data it is fairly clear that the data are not missing at random (nonignorable nonresponse). The missing data pattern is not monotone missing and at least some data is missing on about a dozen variables.

The main analytic method will be Cox proportional hazards.

I am using SAS which now offers an MNAR method for such data. If the pattern is monotone, then it offers options to estimate the missing data with either nearest neighbors or complete cases. If the pattern is non-monotone, then it offers only a method of 'adjusting' the parameters, which seems (from what I understand) like a pure sensitivity analysis.

SAS also offers a MCMC method of creating a monotone missing pattern from a non-monotone pattern.

My current plan is to first create a monotone missing pattern and then apply nearest neighbor, then analyze the multiply imputed data.

However, I am not sure this is best, nor what determines which options to choose in a scenario like this. Advice welcome as are references to the literature.

Answer 1

Whether an imputation method is the "best", depends in part, largely on the discipline.

If this is survey data with non-response because the question(s) were deemed sensitive by the responders, then imputing with the neighbour method is probably inappropriate.

If the data is missing because the recorder malfunctioned, then the nearest neighbor imputation is appropriate. What others have done in the same discipline may also be worth considering.

Whether any method is "best" is debatable. You can do the analysis under several scenarios (ignore missing data, exclude the variable with missing data altogether, nearest neighbor imputation, etc) and see if it you come to different conclusions. This is where experience and subject-matter expertise come in handy.