Table 4:

Real data experiment with n = 649 and p = 10001 under MAR. Approximately 20% features and 76% cases contain missing values. Linear RR, MICE, MIWAE and GAIN are not included due to running out of RAM. Detailed experiment setup information is in appendix.

Models	Style	Time(s)	Imp MSE	${\hat{β}}_{1}$	SE( ${\hat{β}}_{1}$ )
SoftImpute	SI	1008.6	0.0591	0.0213	0.0119
Sinkhorn	MI	843.3	0.0797	0.0223	0.0128
MI-NNGP1	MI	7.1	0.0637	0.0161	0.0104
MI-NNGP1-BS	MI	7.8	0.0685	0.0153	0.0112
MI-NNGP2	MI	11.8	0.0617	0.0171	0.0106
MI-NNGP2-BS	MI	21.3	0.0640	0.0166	0.0110
Complete data	-	-	-	0.0160	0.0085
Complete case	-	-	-	0.0221	0.0185
ColMean Imp	SI	-	0.1534	0.0188	0.0136