Human-biting rates (HBR)
Since weak correlations were found for EIR and sporozoite rates, we focused on weather and biting rate relationships, and found stronger correlations between HBR and maximum and minimum temperatures for A gambiae. R =− 0.36(P < 0.001) and + 0.24 (P= 0.005), respectively. Using log of HBR slightly improved correlation for maximum temperature (r=− 0.41), as did one week lagging (r=− 0.40 and + 0.35 for maximum and minimum temperatures, respectively).
To study the effect of precipitation on the HBR, we first regressed the log of HBR on raw precipitation. There was high variability of the raw precipitation data, and lagged r2, which peaked at 4 weeks, reached only 0.13 and 0.05 for An. gambiae and An. funestus, respectively. Next, we applied the smoothed precipitation for the two anopheline species, and linear regressions were fitted. The correlation increased for An. gambiae (r2= 0.18) and for An. funestus (r2= 0.12), indicating some improvement in the relationship between smoothed precipitation and biting rate at the study site. Multiple regression including maximum and minimum temperatures with smoothed precipitation increased the correlation for both species, but especially for An. gambiaeTable 1).
Table 1. Coefficients of determination (r2) for the regression of smoothed log human biting rates (HBR) for An. gambiae with local weather variables and with modelled soil moisture of the Lake Victoria Basin (for both species), lagged from 0 to 4 weeks. A 5% statistical filter was also applied to precipitation values
Applying the hydrological WatBal model, essentially a more physically based type of precipitation smoothing that includes temperature and other weather and landcover parameters, showed improvement in predicting the biting rate of the two mosquito species. As an intermediate parameter of modelled soil moisture, river runoff also was examined. Regression of log HBR and river runoff (which should closely reflect smoothed precipitation) appropriately resulted in an r2 of 0.16 for An. gambiae and 0.12 for An. funestus. Unlagged regression of log HBR vs. modelled soil moisture increased the r2 value to 0.31 and 0.06, respectively, for the two species.
Lagging the precipitation values enhanced the robustness of the correlation with HBR. For An. gambiae, the peak soil moisture correlation occurred after 2 weeks (r2= 0.45). For An. funestus, the correlation progressively increased out to 4 weeks lagging (Table 1).
Figure 3 shows the time series of An. gambiae HBR and the modelled soil moisture of the Victoria Basin with two weeks lagging. To compare interspecies differences in response to soil moisture, a t-test was applied to the slopes of the regression lines of soil moisture vs. the log of HBR for An. gambiae and An. funestus. The two species' response to soil moisture at this study site (Figure 4) proved to be significantly different (P < 0.00001).
Figure 3. Time series of An. gambiae HBR and modelled soil moisture, with a 2-week lag. ——An. gambiae observed; ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅ Modelled soil moisture.
Download figure to PowerPoint
Figure 4. Regression of the log of An. gambiae and An. funestus HBR and modelled soil moisture. ○An. gambiae observed; ▪An. funestus observed; ——An. gambiae predicted; ⋅⋅⋅⋅⋅⋅⋅⋅⋅⋅An. funestus predicted.
Download figure to PowerPoint
Modelled soil moisture accounted for up to 56% variability of An. gambiae EIR. Lagged r2 values steadily rose after one week, peaking at six weeks, then declined at a similar rate thereafter Table 2). The normalized-difference Vegetation Index (NDVI) obtained from AVHRR satellite images correlated with An. gambiae HBR, yielding an r2 of 0.42.
Table 2. Coefficients of determination (r2) for the regression of smoothed log entomological inoculation rates (EIR) for An. gambiae with modelled soil moisture of the Lake Victoria Basin, lagged from 0 to 12 weeks
While malaria transmission is known to vary seasonally in areas of high endemicity, large interannual heterogeneity of malaria incidence also occurs. This variability has been attributed to entomological components of transmission (Fontenille et al. 1997). However, year-to-year predictability of malaria still remains difficult.
Our study used innovative interdisciplinary methods to analyse this heterogeneity and documents the weather's influence on An. gambiae biting rates and EIR, found previously to account for up to 74% of attack rates in children (Beier et al. 1994). In particular, modelled soil moisture correlations found in our study are biologically plausible, and these results could improve predictions of malaria transmission based on weather conditions. Hydrological modelling can incorporate temperature, precipitation, landcover, soil type and other environmental factors highly relevant to mosquito ecology.
Our analysis demonstrated the added benefit of hydrological modelling compared to reliance on conventional weather parameters often used by infectious disease scientists. For An. gambiae, the soil moisture model predicted up to 45% and 56% of the variability of HBR and EIR, respectively. The peak in correlation between modelled soil moisture and HBR occurring at 2 weeks is not inconsistent with larval development times. Likewise, peak correlation for EIR at 6 weeks is not inconsistent with development time of sporozoites (approximately 12–14 days) plus mosquito survival, averaging 3–4 weeks.
Probably the most important malaria vector in Africa, An. gambiae is known to breed in swamps and temporary puddles. High soil moisture conditions and puddles can remain well after precipitation events depending on the outflow of water from the watershed via runoff and evapotranspiration. The outflow is a function of watershed characteristics and potential evapotranspiration (itself a function of temperature, wind speed, humidity, and solar radiation). The modelled soil moisture therefore is likely a better explanatory variable of mosquito breeding sites than raw weather variables such as temperature and precipitation. This methodology would be especially useful under conditions of extreme rainfall variability, such as those driven by El Niño/La Niña events. In addition, climatologists are projecting a more extreme hydrologic cycle to accompany long-term global warming and climate change; hydrologic modelling for predicting malaria transmission may therefore become that much more important.
Considering the difference in geographical scales, soil moisture modelling could likely be expected to explain even more variability in entomological transmission factors given local hydrological parameters for a study site. Modelled soil moisture for the entire Lake Victoria Basin predicted An. gambiae biting rates slightly better than the combination of local temperature, precipitation variables and NDVI. These methods may be particularly useful in regions of Africa where the malaria incidence is highly variable, and where predictive models can optimize control measures.
Diverse topography and soil composition worldwide require better methods for determining surface water availability for mosquito breeding. For example, the effect of rainfall over steep rocky terrain would substantially differ from that of a flat silty area. Soil moisture modelling can account for landuse changes, such as draining of swamps or water development projects for agriculture. Raw or smoothed precipitation data can not account for such changes that could substantially influence mosquito ecology and subsequent malaria transmission. Studies in multiple sites would further test the utility of this type of modelling for use in malaria prediction.
This study also showed statistically significant interspecies differences in response to soil moisture. An. gambiae and An. funestus are known to have different ecological niches; An. gambiae breeds in surplus surface water, whereas An. funestus breeds in stagnant water at the edge of rivers (Evans & Garnham 1936). One would therefore expect An. gambiae to be more sensitive to soil moisture values and An. funestus populations to be less so, excepting very wet conditions that can wash away eggs and larvae from stream edges (Oliver & Grobler 1992). For example, in Senegal An. funestus abundance was found to peak during the dry season (Fontenille et al. 1997). Furthermore, the change in correlation shown by the different lagging profiles of the two species (maximum strength at 2 weeks for An. gambiae and at 4 weeks or longer for An. funestus) is consistent with larval development rates observed in the laboratory (J. Beier, unpublished data).
Soil moisture modelling and satellite NDVI nearly equally predicted HBR. Remote sensing has proved useful in the prediction of disease distributions and abundance when their distribution in space and time depends largely on climate and landscape features (Washino & Wood 1994; Hay et al. 1997, 1998a). Plant composition and activity in a region reflects and can modify local temperature, precipitation and humidity. Meterological satellite sensors can measure such climate and vegetation variables directly (Hay et al. 1996) and the NDVI has been strongly related to the incidence of severe malaria in three sites in Kenya (Hay et al. 1998b) and in The Gambia (Thompson et al. 1996; Thompson et al. 1997).
Hydrological modelling has some advantages over NDVI, however, in the application to malaria predictions. Firstly, while NDVI is good for observed historical analysis, it cannot be used easily for long-term modelled forecasts, such as general circulation model simulations of climate change. Secondly, daily and weekly modelled soil moisture can be calculated, whereas satellite NDVI is most robust at a monthly timescale and cannot capture weekly variability in biting rates. The most critical periods to assess surface water for mosquito breeding sites often occur during the rainy season; clouds during these periods can impede acquisition of uninterrupted longitudinal data. A soil moisture model does not have this limitation. Finally, hydrological modelling is relatively inexpensive and may be more practical for use by resident public health scientists who have access to local streamflow and weather data in areas with disease risk.
NDVI has utility in assessing disease risk over remote regions where meteorological data may be unavailable. In such locations, satellite remote sensing can be utilized to determine soil moisture (Washino & Wood 1994); vegetative index and surface temperatures provided by satellite can be used to estimate evapotranspiration. Diurnal temperature difference obtained from satellites has been used as a surrogate for soil moisture to predict prevalence of bancroftian filariasis in the Nile delta (Thompson et al. 1996). Remote sensing has been used to predict malaria transmission in several endemic regions (Beck et al. 1997).
Warmer sea surface temperatures and variable rainfall patterns accompanying El Niño events also have been correlated with malaria epidemics in many regions of the world (Bouma & van der Kaay 1996), and predictable temporal cycles of malaria incidence have resulted. For example, malaria in Surinam and Venezuela recurred so regularly (on a 5-year cycle) that the term ‘paraquinquennial’ was coined (Gabaldon 1949). Recent findings by Bouma et al. (1997) show such epidemic periodicity to be linked to the cyclical phenomenon of El Niño. This further demonstrates the strong relationship between weather variability and heterogeneity of malaria transmission, and underscores the need for more in-depth studies of climatic factors influencing malaria.
In our study, the strength of the correlation between soil moisture and HBR dropped after the most extreme dry period during the study, and HBR was slow to recover (Figure 4). Possible changes in reporting accuracy were looked for, but none were found. When HBR finally did increase, it sharply rose to the highest biting rate for the entire study period. This finding is consistent with past observations of large epidemics following on the heels of droughts (Gilles 1993). Given the projection of ‘more extremes in the hydrologic cycle’ as a result of global warming (Karl et al. 1995), our findings are especially relevant to the assessment of malaria risk under such conditions. Shorter-term El Niño events offer a natural experiment in extreme climate variability to apply our methods for gaining insight into such long-range climate scenarios.
In the laboratory, ookinete development shortens as temperatures increase from 21 °C to 27 °C, whereas higher temperatures (30 °C to 32 °C) interfere with the developmental processes (Noden et al. 1995). Unfortunately, there was little temperature variability in Kisian and so temperature/SR analysis was limited. In addition, mosquitoes seek microclimates most suitable for their survival (Service 1993). This species-specific natural behaviour may require greater variability in temperatures to see any relationship between SR and ambient temperatures in the field. In African highland areas, temperatures fluctuate more widely. In the typically nonendemic highlands of Kenya (Garnham 1948), Rwanda (Loevinsohn 1994), and Zimbabwe (Freeman & Bradley 1996) increases in ambient temperature and rainfall have been linked to malaria epidemics. Also in Zimbabwe, summer/fall temperatures partially determine the severity of malaria in the following rainy season from December through March (Freeman & Bradley 1996). In the highlands of Ethiopia, the increase in falciparum malaria has strongly correlated with a steady rise in minimum temperatures over a 25-year period (Tulu et al. unpublished observation), the study controlled for drug-resistance and population migration. Mathematical models show malaria vectorial capacity increasing, as global temperatures rise from the accumulation of greenhouse gases in the atmosphere (Martens et al. 1995; Martin & Lefebvre 1995; Matsuoka & Kai 1995; Bryan et al. 1996; Jetten et al. 1996).