Figure 1

Change in mean-squared-error resulting from the addition of topographic (white boxes) and land-cover/land-use (gray boxes) variables to predictive models with only a person-time offset. The results of 300 split-sample repetitions are pictured (the box represents the 25^th to 75^th percentile, while whiskers show range). Negative values indicate that addition of the variable improved the accuracy of the model in predicting malaria risk; boxes crossing the zero line indicate prediction was improved in some random subsets of houses but not in others.