A quick question on sensitivity analysis

[GiaLee]

Dear experts:

I'm running a mixed model with a longitudinal dataset (3 time points) but have around 50% missingness at time 2 and time 3. I understand mixed models can handle missing data, but can they handle around 50% missingness?

I was suggested to conduct a sensitivity analysis with the complete dataset. The original population consists of around 1500 individuals, while only 200 of them have no missingness on the predictor, outcome, and covariates. Using the original population, the result is significant. However, in the sensitivity analysis, it is not significant. But I think these results are not comparable, since if the missing data are not missing completely at random (MCAR), the results of the complete dataset could be biased. Therefore, they should not be compared with the original findings, whether they robust or conflict with the original findings. Am I correct in my thinking?

If so, do you have any other recommendations on what else I should do for the large percentage of missing data? Thank you very much. 

[Ksharp]

1)You could try PROC GEE ,which also could handle missing value .
https://documentation.sas.com/doc/en/statug/15.2/statug_gee_examples03.htm


2)You could use PROC MI to impute these missing values and PROC MIANALYZE to pool result.
https://blogs.sas.com/content/iml/2020/12/02/score-external-logistic-model.html

[GiaLee]

Thank you Ksharp. I was requested to use mixed model and cannot use other models for the current analysis.
I know mixed models handle missing data using maxixmum likelihood estimation, which is another robust method commenly used as what PROC MI does. But same question is if multiple imputation or likelihood estimation could handle such large amount of missing data. And if yes, what kind of sensitivity analysis I could do to test the reliability of the findings.

[Ksharp]

" could handle such large amount of missing data."
Yes. I think so , why not try it yourself by PROC MI ?

"what kind of sensitivity analysis I could do to test the reliability of the findings."
Sorry. I have no idea about it. Maybe @StatDave @lvm @SteveDenham could give you a hand.

[GiaLee]

Thanks! I just tried Proc MI. But I met another problem. My outcome is score, and my predictor is the time since disease diagnosis. We have 3 time points. For each individual, the time should increase for the three visits, like 15 months, 24 months, and 40 months. However, the multiple imputation imputed the time as 14 months, 40 months, and 16 months. As a result, the mixed model reported an error and stopped working. May I know if it is possible to control the imputed time in an increasing trend? Thanks!