A nonparametric multiple imputation approach for missing categorical data

Table 4 Simulation results from probability estimation for Y, where Y is generated using a logit link function with five covariates, δ is generated using a probit link function with not extreme missingness probabilities (M1) based on five covariates, N = 400

	Pr(Y=1)=0.386				Pr(Y=2)=0.288
Method	Est	SD	SE	CR	Est	SD	SE	CR
FO	0.386	0.023	0.024	0.960	0.286	0.023	0.023	0.934
CC	0.456	0.033	0.035	0.512	0.357	0.033	0.034	0.472
	Working models for Y:				Five covariates with logit link
	Working models for δ:				Five covariates with logit link
					(misspecified scenario 4)
CE	0.386	0.056	0.051	0.944	0.287	0.060	0.051	0.910
PMI	0.388	0.033	0.034	0.950	0.288	0.035	0.034	0.926
NNMI _MLR(5,0.4,0.4;0.2)	0.391	0.036	0.038	0.954	0.294	0.040	0.039	0.942
NNMI _MLR(5,0.1,0.7;0.2)	0.397	0.038	0.041	0.948	0.291	0.039	0.038	0.928
NNMI _MLR(5,0.7,0.1;0.2)	0.388	0.035	0.037	0.966	0.299	0.042	0.041	0.928
NNMI _CLR(5,0.4,0.4;0.2)	0.387	0.035	0.036	0.948	0.303	0.040	0.041	0.928
NNMI _CLR(5,0.1,0.7;0.2)	0.379	0.035	0.036	0.938	0.304	0.040	0.041	0.930
NNMI _CLR(5,0.7,0.1;0.2)	0.395	0.036	0.037	0.956	0.302	0.041	0.040	0.924

Est: Estimates of probabilities; SD: Empirical standard deviation; SE: Estimate of standard error; CR: Coverage rate of 95% confidence intervals; FO: fully observed; CC: Complete Cases; CE: Calibration estimator; PMI: Parametric Multiple Imputation; NNMI _MLR(NN,ω ₁,ω ₂;ω ₃): the NNMI method using Multinomial Logistic Regressions, NN is the number of nearest neighbors and weights are ω ₁,ω ₂, and ω ₃; NNMI _CLR: the NNMI method using Cumulative Logistic Regressions; K = 10 imputed datasets are used for PMI and NNMI methods

ISSN: 1471-2288