Simulation Study when Data are MNAR Sample Clauses

Simulation Study when Data are MNAR. Here we describe a simulation study to evaluate the PPMI-RW method for data that are MNAR. We set n = 1000 and X1, X2, Y are generated from X1 0 0.2 0.1 0 X2 ∼ N 0, 0.1 0.1 0  s  0  0 0 1   Y = X1 + X2 + s Missing values are created in X1 using the logit model: logit(Pr(R = 0)) = 1 + Y + X1 + 0.1X2, resulting in approximately 37% of subjects having missing values. We perform 200 simulations to generate 200 MC data sets and averaged the resulting estimates of ▇▇, ▇▇, ▇▇ (▇▇▇▇ values: 0, 1, 1). We use M = 100 imputations as suggested in ▇▇▇▇▇▇▇▇▇ et al. (2007) and also not burdensome in our simulation setting. We compare PPMI-RW method with GS, CC, MI-na¨ıve, MI-pooled, and MI-pooled-RW. Note that MI-na¨ıve and MI-pooled assume the data are MAR and MI-pooled-RW method re-weight the imputations from MI-pooled assuming the data can be pooled and do not have any privacy constrains. Table 3.4 summarizes the results. MI-pooled-RW and PPMI-RW remove substantially more of the bias than MI-pooled. Moreover, the MSE decreases dramatically after using the re-weighting technique.