Developing a spatial-statistical model and map of historical malaria prevalence in Botswana using a staged variable selection procedure

Table 2 Results of bootstrap step-wise procedures. Variables included in the candidate lists of Stage 3 and Stage 5, and their selection frequency (fq), in four separate automated stepwise backward variable exclusion procedures, each time against 1000 bootstrap samples of the malaria prevalence data.

Theme	Stage 3		Stage 5
	Candidate variable list	fq	Candidate variable list 1‡	fq	Candidate variable list 2	fq	Candidate variable list 3	fq
Rainfall
	annual maximum *	904	annual maximum	560	annual maximum	533	annual maximum	914
			summer total †	821
			number of months >80 mm	760
			SD	726
			total in months >80 mm	716
			annual total	612
	winter total	749
	proportional SD	642
Temperature
	winter mean *	885	winter mean	993	winter mean	878	winter mean	665
					annual mean †	914
					summer mean	885
					number of months >16°C	681
					mean in months >16°C	670
					annual maximum	665
					winter minimum	627
					effective	615
					annual minimum	558
	proportional SD *	754	proportional SD	897	proportional SD	544	proportional SD	624
							SD	786
							annual range	537
	annual maximum	660
Vapour pressure
	SD	495
	summer mean	441
NDVI	annual maximum	567
	SD	469
Elevation *†		874	elevation	988	elevation	819	elevation	994
Log distance to perennial water		616
Land cover *†		988	land cover	996	land cover	997	land cover	996
Month of survey		527

NDVI – normalized difference vegetation index; SD – standard deviation
* Variables selected into Stage 4 model
† Variables selected into Stage 5 model
‡ Example: Five alternative rainfall indicators, listed in candidate list 1 under Stage 5, were strongly correlated with – and had been excluded in favour of – the annual maximum in Stage 2. In Stage 5, all six competing rainfall indicators were included in the candidate list, along with the other variables of the Stage 4 model. Of the six competitors the most frequently selected was summer total. In Stage 5 summer total therefore replaced annual maximum rainfall.

ISSN: 1476-072X