There is a huge number of missing data in my study. and the data are all MNAR (Missing not at random), with little MCAR test (p
Your existing data do not contain information needed for overcoming MNAR. You need extra information (possibly in the form of a model) that gives information about how to overcome MNAR,
Danny -
You said "In clinical, those subjects with more severe syndromes will result in absent data in relevant variables." Perhaps you could stratify your data by 'severity' to see what can be said by such groups, where you have more missing data for the more 'severe' groups/strata. If you have too little data in the most 'severe' cases, you cannot make good inferences, but you may be able to conjecture/interpolate some ideas for further study based on any pattern as you move from stratum to stratum.
Also, you may look for auxiliary (i.e., regressor/independent variable) data for "predictions" for missing data, and consider estimated variances of prediction errors, again doing this stratum by stratum.
It sounds like you already know not to overinterpret any such tentative results, if even tentative results are reasonably possible.
Graphical depiction of your continuous data may be particularly helpful for gaining insight.
But when you are missing relevant data, there is no perfect solution, and often perhaps no reasonable solution at all.
Best wishes - Jim
At least for categorical measures, I prefer creating a 'Missing' category to imputation. It quantifies the MNAR factor rather than trying to adjust for it or pretend it isn't there. Missingness is generally pretty informative in EHR data, so this can be of use on its own. Obviously this does not work as well for continuous measures (though you can always categorize those by interval, which can also be informative if your effects are not truly linear).
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Statistical Software
- R Statistical Package