Population, geographic, and environmental predictors of malaria incidence.

As 2010 data were not separated into active and passive case detection, only 2011–2012 passive P. vivax and P. falciparum cases were included in the regression analysis in an effort to remove bias from the analysis. Because of temporal correlation between incidences each month and the desire to focus on spatial prediction at the time when burden was highest, cases were aggregated into an average for “high season” from August to November for each year. A generalized estimating equation (GEE) model (R function geeglm, package geepack) was used to perform this analysis of 2011 and 2012 “high season” malaria case incidence.^20–22 This model was used to assess the effects of multiple independent variables on P. vivax and P. falciparum incidence. The variables included population density, the designation of urban versus peri-urban versus rural, longitude, latitude, ward-wise literacy rate, difference between male and female ward-wise literacy rate, percentage of the ward classified as urban in the dry and wet season in 2011, the distance of the ward centroid from the Sabarmati River, the shortest distance of the ward centroid from the Sabarmati River or one of the urban lakes, the percentage of the ward classified as green vegetation in the dry and wet season in 2011, the percentage of the ward classified as water during the dry season in 2011 and the wet season in 2010, the land surface temperature during monsoon season in 2011, and the normalized difference vegetation index (NDVI) during monsoon season in 2011.^14,16,17,23 Variables were first evaluated with univariate Poisson regression to identify variables for inclusion in the GEE model. Then a stepwise backward elimination based on t-statistics was used to identify the optimal GEE model. Finally, the most appropriate correlation structure was determined based on the quasi-likelihood under the independence model criterion of different models.