Accurate prediction of Plasmodium transmission risk is essential in heterogeneous environments to permit focal control measures and heightened surveillance in the regions that require them the most . Such prediction requires simultaneous consideration of factors related to vector distribution, human-vector contact, human practices, and the environmental context in which they occur . Results of this investigation corroborate previous studies demonstrating associations between malaria risk and both topographic characteristics, specifically TWI and elevation, and human-influenced environmental variables including land-cover and land-use. Importantly, however, they demonstrate that factors associated with malaria are not necessarily predictive of it, most likely because strong correlations between environmental factors can lead to confounded relationships.
The predictive models tested here indicate that land-cover/land-use variables failed to unambiguously improve prediction of household malaria risk based solely on topographic factors like the distance between a house and the highest TWI locations in the community, despite the fact that these variables were statistically associated with malaria risk. Given the cost and time required to classify land-cover/land-use data at high resolution, the superior accuracy of TWI in predicting high-risk foci suggests that it may be both a simpler and more useful remotely-sensed tool for risk prediction. These findings indicate that control programs operating in such rugged terrain may consider application of the TWI to derive risk maps in such settings. While the most predictive combination of topographic variables changed as different houses were randomly selected for model fitting, a subset of variables - including distance to the wettest and driest locations around the community and the distance to the nearest stream channel - were reliably included in a majority of these combinations, implying important commonalities.
Individual environmental variables, including distance to natural swamps, community of residence, and nearby farmland, each improved predictions of malaria risk over models considering person-time alone. However, none of the examined land-cover/land-use variables improved more than 75% upon the best-fitting topographic models (Figure 2), nor did combinations of variables perform better. In other words, little effect of land-cover/land-use was evident after accounting for general topographic patterns. Given that these human-modified environmental variables are strongly correlated with topographic variables (Table 1), they may be too highly interrelated to observe individual effects. Alternatively, the strong, overarching patterns of malaria risk in this specific region may be largely determined by the uneven topography and varied elevation, making it difficult to identify localized, nuancing effects of small-scale environmental modifications such as those examined here. It seems probable that repeating this analysis in a geographic setting with less topographic variation would produce different results. Thus, analyses similar to these in lowland regions with less extreme topographies are needed to evaluate whether these results are applicable elsewhere.
The community in which a house was located was, unsurprisingly, a good predictor of malaria risk since overall incidence was three times higher in Kapsisiywa than in Kipsamoite. However, after adding the best-predicting topographic variables (Figure 2), consideration of community rarely improved the models. This finding suggests that a great deal of the difference in malaria rates between these two neighbouring communities can be attributed to differences in their topographies.
Distance from households to swamp edges was a better predictor than distance to the stream channels that generally run through the center of swamps, corroborating studies of larval habitat that indicate swampy margins are more suitable than deeper waters for vector breeding . However, distance to any swamp generally did not enhance prediction of malaria case-counts when added to models with the best-fitting topographic variables. This result indicates that the presence of "swampy" land-cover, marked by papyrus and other characteristic "natural" plants, as well as channeled cropland in its reclaimed form, was a less specific measure than the topographic variables describing the shape of the land.
More extensive agriculture surrounding houses, which primarily included cultivation of maize, beans, and other crops, was associated with reduced malaria risk. It is possible that crops decreased the suitability of land for mosquito breeding, either by channeling away standing water or by preventing direct sunlight from reaching water pooling under tall crops, but such an effect would run counter to the conclusions of other investigations . Alternatively, large amounts of farmland may be indicative of greater SES, and the observed association accordingly may reflect such unmeasured variables . These and other possibilities including specific crop types and cultivation methods should be more carefully evaluated.
Misclassification of environmental variables may have affected these results. Land-cover measures were derived from a single IKONOS image taken in 2002, and any changes in land-use or land-cover occurring after that time could have introduced error. Transects conducted at the study site in 2005 that georeferenced the location and type of fields, pastures, trees, and other land-cover/land-use matched closely with the classified satellite-image (data not shown), indicating that features had not changed greatly over the time period between the image and the study. Although farmers reported that crop types can change from month-to-month or from year-to-year, the location of fields remained stable over the study period. However, if human modification of the local environment causes only subtle changes in malaria risk from general topographic trends, even a small amount of error could compound the difficulty of observing such effects. Additionally, the TWI data described here was based on a single algorithm ; consideration of other methods for deriving TWI may yield different results.
Classification of satellite imagery is challenging, and methods for doing so have not improved greatly in recent years . Elevation data, however, is easily processed into useful metrics like the topographic wetness indices computed here with the use of freely available GIS programs. The June 2009 release by NASA and Japan's Ministry of Economy, Trade, and Industry of a global 30 m digital elevation model derived from Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) imagery http://asterweb.jpl.nasa.gov/gdem.asp permits free access to global elevation data. These findings suggest that such data resources may be used in identifying other foci at high risk for malaria, at least in highly topographically variable regions. Although the IKONOS image used here was selected for its fine spatial resolution of 1 m2, replication of this analysis using other types of remotely-acquired imagery, such as Landsat or ASTER, would allow consideration of land-cover and land-use variables derived from alternative sources; the greater temporal resolution and use of different sensors would allow examination of other types of data that have previously been demonstrated to be important for malaria risk prediction [4, 29].
The results of this study indicate that elevation and predicted accumulation of water are highly predictive of malaria patterns in this small region. People living in areas with a high TWI appeared to be at significantly greater risk of malaria than those living in areas of lower TWI. Other variables related to land-cover/land-use and human modification of the environment demonstrated associations with household malaria and improved prediction of models over that of person-time alone. However, these variables generally failed to produce unambiguous improvements in models controlling for important topographic factors, suggesting that their importance for determining patterns of malaria in this region was slight when accounting for the varied shape of the land around these communities. Future studies should assess the utility of TWI for malaria risk prediction across larger and more geographically diverse areas, especially at scales useful for Ministries of Health to target interventions; replication of this work in regions of both similarly rugged terrain and flatter, more arid areas will be required before conclusions can be drawn about the utility of these methods elsewhere. However, these results suggest that malaria control programs in similar highland regions might use topographic and geographic variance rather than land-cover/land-use to efficiently identify locations that are highly suitable for transmission and which may benefit from enhanced vigilance.