Scenario 1c: Additive model (misspecified)    θ(0)=0.35 Bias CMLE 0.099 −0.248 −0.078 0.096 MLE 0.100 −0.250 −0.078 0.095 PCL 0.001 −0.137 0.000 0.103 RMSE CMLE 0.179 0.287 0.168 0.184 MLE 0.157 0.277 0.145 0.158 PCL 0.139 0.188 0.131 0.168 Scenario 1c: Model with second-order contrasts    θ(0)=0.35 Bias CMLE −0.090 0.023 0.022 0.371 MLE −0.080 0.016 0.02 0.363 PCL 0.015 −0.012 0.004 −0.001 RMSE CMLE 0.224 0.256 0.25 0.447 MLE 0.191 0.219 0.217 0.419 PCL 0.211 0.266 0.272 0.248

Table 3: Simulation results for the complete-case MLE, the MLE, and the pseudoconditional likelihood method. Here RMSE represents root mean squared error. The results were based on 2,000 runs. There were 2×2×2 = 8 disease subtypes. The model for the intercepts was misspecified. The missingness probabilities depend on the covariate. This is Scenario 1c, where the true values of some of the second-order contrasts of the log-odds ratio parameters were not zero.