Missing in space: an evaluation of imputation methods for missing data in spatial analysis of risk factors for type II diabetes

Table 2 Comparison of imputation methods by root mean squared error (RMSE) and bias from cross-validation

Covariate	N missing	RMSE,mean (sd)					Average bias		Average width of CI		% of CIs including zero bias
Covariate	N missing	Mean imputation	MVN		CAR prior		Average bias		Average width of CI		% of CIs including zero bias
			Poisson	Binomial	Poisson	Binomial	MVN	CAR prior	MVN	CAR prior	MVN	CAR prior
% over 45yrs of age	3	49.7	108.7 (21.2)	109.6 (19.5)	46.3 (17.8)	46.2 (17.9)	0.012	0.041	0.897	0.345	100%	100%
% Overweight/obese	28	26.3	73.7 (22.3)	73.1 (20.6)	45.4 (18.5)	45.4 (18.4)	0.016	0.081	0.413	0.200	100%	75%
% Daily smokers	28	25.8	53.2 (18.1)	54.4 (19.3)	49.8 (21.1)	49.8 (21.0)	0.153	0.271	1.144	0.640	100%	68%
% Insufficient physical activity	28	36.7	90.7 (37.5)	91.4 (37.1)	67.0 (41.3)	67.1 (41.3)	0.048	0.047	0.535	0.246	100%	93%
% Adequate fruit intake	28	34.4	67.4 (14.3)	68.6 (14.7)	37.5 (22.7)	37.6 (22.7)	0.069	0.052	0.382	0.221	100%	91%
% Adequate vegetable intake	32	21.9	39.1 (18.5)	39.2 (18.4)	30.4 (19.6)	30.6 (19.9)	0.157	0.185	1.144	0.973	100%	94%
Overall	-	32.5	71.1 (11.4)	72.7 (11.1)	46.1 (11.7)	46.1 (11.7)	0.076	0.113	0.752	0.438	100%	87%

RMSE = root mean squared error, sd = standard deviation, MVN = Multivariate normal imputation, CAR prior = conditional autoregressive prior imputation; CI = 95% credible interval.

ISSN: 1476-072X