Implications of climate change for malaria transmission across Africa have been the subject of numerous commentaries [10-12] and a few attempted analyses [13-18]. One analysis [18] developed a global picture of climate change effects on malaria transmission, but relied on country-level classifications of risk, and did not discern the fine details of mosquito-species-specific distributional shifts. To date, then, specific climate change scenarios have yet to be applied to derivation of detailed projections of changing patterns of potential for malaria transmission based on mosquito vector species' distributions across Africa.

Ecological niche models were generated using the Genetic Algorithm for Rule-Set Prediction (GARP) [34], a niche-modeling approach that has been the basis for most previous niche modeling applications to disease geography [9]. GARP is an evolutionary-computing method that builds ENMs based on non-random associations between known occurrence points for species and sets of GIS coverages describing the ecological landscape. Occurrence data are used as follows: 50% of occurrence data points are set aside for an independent test of model quality (extrinsic testing data), 25% are used for developing models (training data), and 25% are used for tests of model quality internal to GARP (intrinsic testing data). Distributional data are converted to raster layers by the GARP program; then, by random sampling from areas of known presence (training and intrinsic test data) and areas of 'pseudoabsence' (areas lacking known presences), two data sets are created, each of 1250 points; these data sets are used for rule generation and model testing, respectively.

The first rule is created by applying a method chosen randomly from a set of inferential tools (i.e., atomic rules that specify particular environmental value combinations as suitable, logistic regression, bioclimatic range rules, negated bioclimatic range rules [34]). The genetic algorithm consists of specially defined operators (e.g., crossover, mutation) that modify the initial rules, and thus the result are models that have "evolved" – after each modification, the quality of the rule is tested (to maximize both significance and predictive accuracy) and a size-limited set of best rules is retained. Because rules are tested based on independent data (intrinsic test data), performance values reflect the expected performance of the rule, an independent verification that gives a more reliable estimate of true rule performance. The final result is a set of rules that can be projected onto a map to produce a potential geographic distribution for the species under investigation.

Following recent recommendations [35], for each species, I developed 100 replicate random-walk GARP models, and filtered out 90% based on consideration of error statistics, as follows. The 'best subsets' methodology consists of an initial filter removing models that omit (omission error = predicting absence in areas of known presence) heavily based on the extrinsic testing data, and a second filter based on an index of commission error (= predicting presence in areas of known absence), in which models predicting very large and very small areas are removed from consideration. Specifically, in DesktopGARP, I used a "soft" omission threshold of 20%, and 50% retention based on commission considerations; the result was 10 'best subsets' models (binary raster data layers) that were summed to produce a best estimate of geographic prediction. I took as a prediction of suitable conditions those areas for which ≥8 of the 10 replicate models predicted potential for presence, and a prediction of unsuitable conditions those areas for which ≤4 of the 10 replicate models predicted potential for presence.

Future climate conditions were reconstructed based on two general circulation models (GCMs) that have been used to simulate future climates: those of the Hadley Centre (HadCM3) [36] and the Canadian Center (CGCM1) [37]. From each GCM, I analyzed two greenhouse gas emissions scenarios: the B2 scenario, which is a relatively conservative estimate of climate change, and the A2 scenario, which is more extreme in the climates reconstructed. As they are based on a 30 yr average around 2055, these models do not take into account potential effects of increased climate variability (El Niño events, in particular) on species' distributions. Because these future climate data are provided at a very coarse native spatial resolution (2.5 × 3.75°), I calculated expected changes in temperature (°C) and precipitation (mm) under each scenario from the relatively coarse raw model results; these expected changes were applied to the original IPCC current climate data layers to provide a final pixel resolution of ~30 × 30 km for future-climate data layers. Models developed based on present-day occurrence patterns and environmental variation were projected to these 4 views of likely future climate conditions.

In the previous global analysis [18], as in my more detailed analysis, increasing malaria risk was projected in East Africa and decreasing malaria risk was projected for West Africa, so the general pictures presented by the two analyses are parallel, but the detail offered in the present study is considerably greater. Nonetheless, my results require careful examination and exploration, because overall reductions in potential for human exposure nonetheless are achieved by reductions in some regions and increases in others.

Ecological niche models reconstruct patterns of association between species' occurrences and environmental variation across space and time [40], and offer a predictive understanding of distributional patterns that can extend over changing environmental conditions. They require testing and validation prior to use [41], however, which in the present case has been the subject of a detailed, published analysis demonstrating excellent predictivity of distributional patterns of the vector species even across broad, unsampled regions [28]. These tests, as well as more general validations of the analytical framework [7,42-44], increase confidence that projections of niche dimensions across periods of environmental change will also have predictive power. That mosquito species are likely to respond to changing climates is quite probable, particularly in light of recent studies demonstrating that mosquito distributions in time and space respond to the same sets of stimuli [26].

The climate change picture regarding shifting patterns of malaria vector distributions across Africa is complex. In general, future expectations are that, while fewer people live in areas that will be suitable for malaria vectors across West Africa, more people will be living in exposed areas in East Africa and southern Africa. Because West Africa is more densely populated than parts of eastern and southern Africa, a net decrease in potential for human exposure to these vectors results. For A. gambiae, 78–112 million fewer people are projected to be in suitable areas for the vector species, as compared with 33–34 million people living in areas becoming suitable; for A. arabiensis, reductions in potential for exposure are more dramatic: 136–171 million in areas no longer suitable, but only 14 million in areas newly suitable. The spatial distribution of these shifts, however, is perhaps the most intriguing result: malaria transmission risk is likely to shift eastward and southward, establishing new foci in areas presently lacking intense malaria transmission.