The data are based on measurements of SBP in the HERS study. The data set allows us to compare methods of analysis with complete data and under simulated missing data. In the data sets are missing data indicators (miss mar for bpmisslong and miss mar1 for bpmisswide) which have value 1 to flag SBP values which should be dropped to simulate data which displays MAR missingness. In particular, year 1 values from patients with higher baseline SBP are flagged more frequently and hence will be simulated as missing. You can consult the course website for the data sets and more complete documentation and details on Stata code.
(a) Attempt to mimic the effects of multiple imputation by performing imputation to fill in SBP values flagged as missing in the simulated scenario. You may choose the imputation model but it should include baseline SBP, BMI at baseline and year 1 as well as diabetes. In Stata, it will be simplest to perform multivariate normal-based imputations.
(b) Fit a GEE model (as in Problem 11.7) with multiple imputation. How do the results compare to the results in Problem 11.7? Note, to fit the GEE model you will need to convert the data from a wide to long format. In Stata, this can be done with the mi convert command.
(c) Fit a mixed model (as in Problem 11.8) with multiple imputation. How do the results compare to the results in Problem 11.8?
Problem 11.8
Using bpmisslong, fit a mixed linear regression model with SBP as the outcome and visit (visit) as the predictor. In Stata, the command would be xtmixed sbp visit || pptid:. Compare the mixed model which uses the full data to one restricted to nonmissing data (miss_mar equals 0). Compare the results with the GEE results in Problem 11.7. How do you explain the difference in results between the GEE and a linear mixed model with MAR missing data?
Problem 11.7
Using bpmisslong, fit a GEE model with SBP as the outcome and visit (visit) as the predictor. In Stata, the command would be xtgee sbp visit, i(pptid) corr(exch). Compare a GEE model which uses the full data to one restricted to nonmissing data (miss_year equal to 0). What do you conclude about GEE with MAR missingness?