Portal de Periódicos da CAPES

Spatial and temporal dynamics of Aedes aegypti larval sites in Bello, Colombia

2012; Wiley; Volume: 37; Issue: 1 Linguagem: Inglês

10.1111/j.1948-7134.2012.00198.x

1948-7134

Sair Arboleda, Nicolás Jaramillo, A. Townsend Peterson,

Mosquito-borne diseases and control

Counts of immature stages of the mosquito Aedes aegypti have been used to calculate several entomological indices of dengue vector abundance. Some studies have concluded that these indices can be used as indicators of dengue epidemic risk, while other studies have failed to find a predictive relationship. Ecological niche models have been able to predict distributional patterns in space and time, not only of vectors, but also of the diseases that they transmit. In this study, we used Landsat 7 ETM+ images and two niche-modeling algorithms to estimate the local-landscape ecological niche and the dynamics of Ae. aegypti larval habitats in Bello, Colombia, and to evaluate their potential spatial and temporal distribution. Our models showed low omission error with high confidence levels: about 13.4% of the area presents conditions consistently suitable for breeding across the entire study period (2002–2008). The proportion of neighborhoods predicted to be suitable showed a positive association with dengue case rates, whereas the vector-focused Bretau index had no relationship to case rates. As a consequence, niche models appear to offer a superior option for predictive evaluation of dengue transmission risk and anticipating the potential for outbreaks. Aedes aegypti is the main vector of dengue in the world and shows a close association with humans. As such, it is well adapted to urban environments with high human population density and low vegetation coverage (Braks et al. 2003, de Lima-Camara et al. 2006, Honório et al. 2009). In these areas, Ae. aegypti females lay eggs in human-made breeding sites such as used tires, metal drums, recycling containers, and domestic water storage containers (Christophers 1960, Maciel-de-Freitas et al. 2007a,b). The entomological indices commonly used in dengue vector surveillance efforts were developed originally for yellow fever control ("Stegomyia indices") and focus on immature populations. Examples include the Breteau Index (BI), Container Index (CI), and House Index (HI) (WHO 1997). BI is considered to be the best risk indicator of dengue outbreaks, as it combines information on containers and houses (Tun-Lin et al. 1996). However, several studies have shown that these indices may not always be useful in predicting dengue transmission risk in light of the complex and multifactorial nature of dengue transmission (Focks 2003). BI may be less successful than indices based on pupal stages, since the larval populations on which it focuses do not necessarily translate directly into adult populations (Focks and Barrera 2007). On the other hand, environmental conditions are very important for Ae. aegypti development, both in terms of micro-environments (Gubler 1997) and climate (Peterson et al. 2005), which determine breeding abundance and productivity (Maciel-de-Freitas et al. 2007b). As a result, risk maps based on environmental associations might hold promise. Information extracted from satellite images, and integrated and processed in geographic information systems (GIS), has proven useful in understanding spatial and temporal distributions of vector-borne diseases (Arboleda et al. 2009, Dister et al. 1997, Hayes et al. 1985, Linthicum et al. 1999, Peterson et al. 2005, Wood et al. 1992). One useful paradigm has been that of mapping occurrences, relating them to environmental conditions (the ecological niche), and projecting this "niche model" back onto geography to identify areas of potential distribution (Peterson et al. 2011), which has now been applied in numerous situations of disease transmission (González et al. 2010, Gorla 2002, Larson et al. 2010, Moffett et al. 2007). In the case of dengue, some previous studies have been based on human infections (Arboleda et al. 2009); here, we explore the spatial and temporal dynamics of dengue vector mosquito distributions (Peterson et al. 2005). The purpose of this study is to analyze and estimate the local-landscape ecological niche and dynamics of breeding sites for Aedes aegypti in the municipality of Bello, in the Aburrá Valley region of Antioquia, Colombia, during 2002–2008. After the invasion of Latin America by Ae. aegypti in the early 1970s, the first dengue epidemic in Colombia occurred in 1971. Epidemics in the country have occurred about every four years since then, as in other countries in the Americas. Among municipalities in Colombia, however, Bello has emerged as having environmental and social characteristics that allow near-constant dengue transmission. Our goal is to evaluate the degree to which ecological niche models are able to anticipate the dynamics of breeding by this mosquito species, both across the region and through time, and whether breeding suitability patterns of this vector species translate into variation in human dengue case frequency through time. As part of this aim, we compare the performance of niche models as predictors of human dengue case frequencies with that of the Breteau index. We focused on urban areas within the municipality of Bello, which is part of the greater Medellín metropolitan area, in Antioquia, Colombia, as a study region (Figure 1). This region is of particular interest in light of its high dengue endemicity, with co-circulation of the four dengue serotypes over the last ten years. Bello has a population of 371,973 inhabitants in 97,960 domiciles and covers 19.7 km2. It is divided into 82 neighborhoods grouped into ten communes. Bello is located on a broad, inclined slope that drops from 1,600 to 1,400 m elevation from northwest to southeast, with an average annual temperature of 22° C and annual total rainfall of ∼1,350 mm; no major variation in climatic characteristics exists among neighborhoods. The municipality apparently presents suitable climatic conditions for Ae. aegypti populations; however, at the same time, it has generally good public services, as 96.4% of houses have dedicated water service. The location of Bello municipality, in northern Medellín, Colombia. To make spatially stratified tests, the study area was divided into four quadrants based on the median longitude and latitude of occurrence points of human dengue cases (Arboleda et al. 2009). Points represent mosquito breeding from 2002 for illustration; in spatially stratified tests, points located in dashed quadrants (black circles) were used to train models, and points located in solid quadrants (dotted circles) were used to test them, and vice versa. This procedure was done for each year (2002–2008) and each algorithm (Maxent and GARP). Dengue vector surveillance and control is carried out by the Dirección Local de Salud de Bello (DLSB), which is charged with dengue control in the region. DLSB officials visit samples of 5,709–13,137 houses randomly two to four times yearly to examine water containers and evaluate and eliminate possible mosquito breeding sites. They re-visit houses previously reported as breeding sites to assure that mosquitoes were eliminated, visit houses of patients infected with dengue fever to eliminate point sources, and conduct spraying campaigns. These activities cover the entire municipality. Data regarding sites at which Ae. aegypti was documented as breeding during 2002–2008 were obtained from DLSB sampling of all neighborhoods of Bello. Within each neighborhood, geographic coordinates of sites positive for mosquito breeding were obtained from entomological surveillance by DLSB, which conducted larval surveys according to WHO guidelines (PAHO 1999). The basic unit of sampling was the house, where systematic searches found water containers with larvae, pupae, or exuviae. We employed the Breteau index (BI = number of positive containers per 100 houses inspected) as the infestation indicator in our analyses. It is considered to be the most informative as it establishes a relationship between positive containers and the set of houses inspected (Scott and Morrison 2003). The presence of Ae. aegypti in a water container was the criterion to designate a house as "positive" for breeding. Addresses of houses sampled in the seven years of activity were tabulated, and geographic coordinates assigned using the polygons in the "Plan de Organización Territorial (POT)" of Bello in ArcGIS (version 9.2). A total of 5,709 houses was visited and sampled for mosquitoes, and 2,300 yielded records of Ae. aegypti, of which 2,075 could be georeferenced satisfactorily (i.e., to a precision of ≤20 m) over the study period (Table 1). In 2002–2004, two sampling events were available yearly; in 2005–2008, three sampling events were available yearly (Table 1). Sampling in each year was not necessarily carried out on the same dates. Hence, we decided to condense information by year, both to make analyses more robust and to make possible comparisons among years. To complement information on mosquito breeding sites, we obtained information regarding human dengue cases occurring in Bello in the time frame of one to three weeks after the entomological surveys, and thus took into account the time lag of transmission of the virus. This information was obtained at the level of neighborhoods for each year in 2002–2008 (Table 1). Dengue case information was kindly provided by local health entities (see Acknowledgments), and corresponds to cases that met WHO definitions as at least "suspected" cases (WHO 1997). Symptoms accepted as indicative of probable dengue fever included acute illness with two or more of the following manifestations: headache, retro-orbital pain, myalgia, arthralgia, rash, hemorrhagic manifestations, leucopenia, and supportive serology. The latter evidence was based on a reciprocal hemagglutination-inhibition antibody titer of ≥1,280, a comparable IgG enzyme-linked immunosorbent assay (ELISA) titer, or a positive IgM antibody test on a late-acute or convalescent-phase serum specimen. An additional criterion was occurrence at the same place and time as other confirmed cases of dengue fever (WHO 1997). For geospatial data, we used 23 remotely-sensed images from the ETM+ sensor onboard the Landsat 7 satellite (path/row: 9/56). These images have a spatial resolution of 30 m, and images are in theory available for each site every 16 days, although cloud cover frequently made images unusable. We used bands 1–7 as environmental variables in niche modeling exercises (band 6.1 has a spatial resolution of 60 m, but was resampled to 30 m for consistency with other bands). We derived a vegetation index using bands 3 and 4. For each year, we chose the images closest temporally to the mosquito sampling dates that showed low cloud contamination. Numbers of images available and suitable ranged 2–5 for individual years (Table 1). All images from a given year were averaged for analyses. All layers were masked to the extent of Bello only (Figure 1). Data on elevation were obtained from the Shuttle Radar Topographic Mission (SRTM) digital elevation model (http://srtm.csi.cgiar.org/). This data set has a native spatial resolution of 90 m, but was resampled to 30 m for consistency with the Landsat data sets. We derived data layers from this original elevation dataset that summarized aspect and slope using the Spatial Analyst tool kit in ArcGIS 9.2. In sum, then, we used the following raster datasets to develop niche models for each year of sampling: Landsat bands 1–7, vegetation index, elevation, aspect, and slope. Recent comparative evaluations of niche modeling algorithm performance have attempted to identify the "best" algorithm (Guisan et al. 2007, Wisz et al. 2008). While a few approaches certainly perform less well than others, and should probably not see further use, performance of a suite of the "better" approaches is surprisingly consistent (Peterson et al. 2007). Furthermore, as current conceptual frameworks (Peterson et al. 2011) and accepted approaches to evaluating model predictions (Lobo et al. 2008, Peterson et al. 2008) are both rather generally flawed, comparative model evaluation exercises have rarely presented conclusions that genuinely speak to the question of predictive ability of niche models. As a consequence, in our analyses, we used two contrasting inferential algorithms to estimate the ecological niche of Ae. aegypti in the Bello region and sought areas of agreement between the two as a best estimate of suitable conditions and sites (Araújo and New 2007). Maxent is a method developed to estimate ecological niches of species based only on presence data, although the broader "background" of conditions across the study area is used in the analysis. The geospatial information available generally takes the form of a set of real-number-valued environmental variables, called "features," and distributions are fitted under the constraint that expected values of each feature should match the empirical average (average value for a set of sample points taken from the target distribution). Maxent thus attempts to estimate the probability distribution for the occurrence of species as the "maximum entropy" distribution—an approximation to a uniform probability distribution, subject to the constraints imposed by environmental conditions associated with known occurrences of the species in question (Phillips et al. 2006). All Maxent models were run with version 3.3.3, using 50% of data to train models and default values for the other parameters. Maxent is relatively robust to small sample sizes: if sites sampled represent well the environmental diversity of the species' distributional area and the study area, models will generally have good predictive ability (Pearson et al. 2006, Wisz et al. 2008). A real-number suitability value is assigned to each pixel, which can vary from 0 (no suitability) to 1 (complete suitability). However, one facet of focusing on estimating ecological niches, emphasizing predictive ability of models, and minimizing overfitting (i.e., avoiding "good" local solutions that have little generality), is that raw, continuous predictions should generally be converted to binary formats by means of a thresholding step (Pearson et al. 2006, Peterson et al. 2007). Input occurrence data are not without error, which should be taken into account in the design of such analyses. In the present case, since immature stages of Ae. aegypti can be confused with those of other mosquito species (e.g., Culex pipiens) that occur in the study area, and some georeferences may be erroneous or not representative of where the person was infected, we estimated the percentage of input occurrence sites that may be misleading, such that a "presence" is placed under mistaken conditions falling outside of the true niche, termed E by Peterson et al. (2008). To convert continuous raw Maxent predictions into binary formats, we first multiplied the raw grids by 10,000 and converted them to integer format, and then followed a modification of the Least Training Presence Thresholding approach of Pearson et al. (2006): we incorporated the potential for error by seeking the threshold that included (100 –E)% of training occurrence points. We used E= 0.1; that is, we assumed that our data include meaningful errors at rates as high as 10%. As our final model predictions were dichotomous (i.e., predicted presence or absence), we used a binomial test to evaluate model predictivity, using as trials the independent subsets of the breeding occurrences used to train the model; as successes, the breeding sites that fell in area as predicted as suitable by models; and as the success probability, the proportion of area predicted as suitable for breeding. The omission error index was measured as the proportion of independent breeding sites that did not fall in the area predicted as suitable. The Genetic Algorithm for Rule-set Prediction (GARP) relates ecological characteristics of occurrence points to those of points sampled randomly from the rest of the study region, developing a series of decision rules that best summarize factors associated with presence, by means of a genetic algorithm (Stockwell and Peters 1999). Within GARP processing, occurrence points are divided evenly into calibration and evaluation datasets. GARP works in an iterative process of rule selection, evaluation, testing, and incorporation or rejection. A method is chosen from a set of possibilities (e.g., logistic regression, bioclimatic rules) and applied to the training data to develop or evolve a rule. Predictive accuracy is evaluated on the basis of the testing data. Rules may evolve in ways that mimic DNA evolution (e.g., point mutations, deletions). Change in predictive accuracy between iterations is used to evaluate whether particular rules should be incorporated into the model; the algorithm runs 1,000 iterations or until convergence. To optimize GARP model performance, we developed 100 replicate models for each round of modeling, based on independent random subsamples from among available occurrences, each of which was evaluated using an independent testing dataset (50% of available occurrence data) set aside prior to model development. We chose a "best subset" of the replicate models on the basis of optimal error distributions for individual replicate models (Anderson et al. 2002): we chose the 20 models with the lowest omission error, and then retained the ten of these models with predicted area closest to the median area predicted among the twenty low-omission models. The geographic predictions of these ten models were summed on a pixel-by-pixel basis to provide a summary of potential geographic distributions; we then converted maps to binary using the modified Least Training Presence Thresholding approach described above, also incorporating an intrinsic error rate of E= 0.1, and calculated the cumulative binomial probability and omission error index, also as above. Models fitted in highly dimensional environmental spaces tend to be overfitted to the input data (Peterson et al. 2007). This point thereby prioritizes steps that maximize the signal-to-number of dimensions ratio in environmental data sets. For this reason, we assessed the importance of each element of the overall suite of 11 variables using a jackknife manipulation (Peterson and Cohoon 1999). We ran models with both algorithms (Maxent and GARP) based on different sets of combinations of variables: all variables, each variable alone, and all possible combinations of N– 1 variables. We chose variables for inclusion in final model development based on omission error rates and probability values from binomial tests. After identifying combinations of variables for each year that best predict mosquito breeding, we evaluated the capacity of each model to predict into broad unsampled areas by dividing the study area into four quadrants based on the median longitude and latitude of the occurrence points available to us (Figure 1). In these analyses, we used points located in the northwestern (NW) and southeastern (SE) quadrants ("area 1") to train models, and points in the northeastern (NE) and southwestern (SW) quadrants ("area 2") to test them; the converse predictions and tests were also developed. Once we obtained binary (thresholded) predictions for Maxent and GARP models as described above, we calculated the cumulative binomial probability that the observed degree of coincidence between predictions and independent test points (i.e., separate spatial subsets, described above) could occur at random as a final evaluation of model predictions over space. Best models based on optimized variable combinations and built without spatial subsetting were used to generate a consensus model for each year by combining the final binary model from Maxent with the corresponding final model from GARP. We identified as potential distributional areas those sites where both models agreed in predicting suitability for Ae. aegypti breeding. The percentage of area overlapping between Maxent and GARP models was calculated by dividing the coincident area predicted as suitable by the total study area. Final consensus models for each year were in turn combined to identify areas with conditions suitable for breeding continuously through the study period (2002–2008). Breeding records from each year were superimposed on this final map to calculate omission error rates for each year with respect to these long-term stable areas. To assess whether omission error rates were related to the size of the area predicted as suitable, the additional area predicted as suitable in each year model with respect to the final map was compared with these omission error rates using linear regression. Of particular interest was the predictive ability of our models over time. Hence, we used each yearly model to anticipate the spatial distribution of Ae. aegypti breeding sites in subsequent years by transferring the niche model from each year with mosquito occurrence onto the environmental conditions of later years. We calculated cumulative binomial probabilities for each model prediction from both Maxent and GARP. Noteworthy was the degree to which temporal dynamics of conditions across landscapes shifted suitable areas for dengue vector mosquito breeding from area to area among years. Finally, we compared the area predicted as suitable across space for Ae. aegypti breeding with human case rates. The proportional area suitable in each neighborhood in each year was calculated by dividing the area predicted as suitable by the total area of the neighborhood. For comparison, dengue case rates were calculated as the number of dengue cases occurring within one to three weeks after the sampling period in the neighborhood divided by the neighborhood population (data provided by DLSB). We used linear regression to test for relationships between human case rates and proportional area predicted by the models as suitable for mosquito breeding. In parallel, we evaluated relationships of the Breteau Index with dengue case rates by neighborhood and the proportional area predicted as suitable, also using linear regression. Analyses of Maxent and GARP models based on different variable combinations showed that some variables do not contribute significantly to model quality, as reflected in lower significance or non-significance (P > 0.05) of models based on them: topographic variables (elevation, aspect, and slope), band 3, band 7, and NDVI (Table 2). Hence, we conclude that bands 1, 2, 4, and 6 are useful predictors. We found significant associations between numbers of georeferenced breeding sites and area predicted as suitable for Ae. aegypti breeding (P < 0.001), with omission error rates for independent random 50% subsets of occurrence data ranging 1–10%, in predictions covering 28–41% and 34–66% of the study area for Maxent and GARP, respectively (Table 2). GARP models generally showed lower omission rates than Maxent models, but Maxent models tended to identify more restricted areas (Figure 2); these observations are consistent with the idea that GARP models often identify areas that are broader and more inclusive than Maxent models (Ortega-Huerta and Peterson 2008). Ecological niche model summaries of suitable areas (in gray) for Aedes aegypti breeding in each year as predicted using Maxent and GARP, and based on the combination of the two methods. The spatially stratified tests of the capacity of models in each year to predict into unsampled areas indicated significant predictive ability. In fact, in all cases except two, model predictions of independent testing data in unsampled regions were significantly better than random expectations (P < 0.05; Table 2). The exceptions were as follows: in 2004, the Maxent model (area 1 predicting area 2) was unable to predict independent testing points better than random, and in 2005, the GARP model (area 2 predicting area 1) was unable to predict independent testing points better than random (Table 2). For Maxent, significant models tested in this manner had omission error rates of 9–11% and predictions that covered 46–83% of the study area, while significant GARP models had omission error rates of 9–11% and predictions that covered 23–61% of the area. Combination of the best Maxent and GARP models for each year revealed that the area of overlap between the two covered from 44.3% (2008) to 76.1% (2006; Table 2, Figure 2) of the municipality. The yearly final consensus models for each year themselves coincided in identifying 13.4% of the total area of Bello as consistently suitable for Ae. aegypti breeding across all years (Figure 3). Additional area predicted by models for individual years ranged 2.1–3.0-fold greater than the core 13.4%, indicating that one-half to one-third of yearly suitable area is consistently suitable for Ae. aegypti breeding. When breeding sites were overlaid on the final among-year consensus model, omission error rates were high (Table 3), indicating frequent breeding in areas outside of the core area. However, omission rates showed no association with additional area predicted (P= 0.916), indicating that the size of the area predicted does not determine omission rates. Consensus map summarizing areas as consistently suitable (in gray) for Aedes aegypti breeding across the municipality of Bello by both Maxent and GARP models across all years. All yearly models showed good capacity to anticipate breeding mosquito distributions in other years significantly better than random expectations (all P < 0.05), with omission error rates ranging 13–23% for sum models. In general, omission error rates increased as the evaluation year became more distant in time from the training year (Table 4). Models appeared more predictive when projected onto the environmental conditions of the year in question. However, calculating omission rates for models projected to the environmental conditions of the "other" years (10–31%), as compared with omission rates for models not projected but applied to the occurrence data of other years, we see that in 11 of 21 cases omission rates were lower with the projected models, which is not significantly different from random expectations (cumulative binomial probability, P= 0.33; Table 4). All omission error rates were ≤0.23; in other words, all among-year predictions could predict correctly >77% of independent test points. The prediction of mosquito breeding sites can be used to anticipate concentrations of human dengue cases. Our results showed a statistically significant positive correlation between area predicted as suitable by models and dengue case rates (P < 0.05; Table 5) in all years except 2005 and 2008. However, BI was not related significantly to dengue case rates in any year (all P > 0.05; Table 5). Proportional areas in each neighborhood predicted as suitable were also not related to BI in any year (all P > 0.05; Table 5). Monitoring mosquito breeding is one of the most common tools used to measure risk of dengue outbreaks, yet the effectiveness of these indicators as regards transmission risk has been debated (Focks and Chadee 1997, Tun-Lin et al. 1996). However, these indicators have long been used to reflect the situation of the disease (Brown 1977, Connor and Monroe 1923, Soper 1967). The municipality of Bello is not an exception: entomological indices are used to monitor dengue risk, even though the causal association between entomological indices and dengue case rates is not well established. In this study, we tested for such relationships and found none. We evaluated whether models based on remotely sensed information can predict the spatial and temporal dynamics of mosquito breeding. Although mosquito breeding is also influenced by microenvironmental conditions and human behavior, macroenvironmental conditions play fundamental roles in determining mosquito breeding activity, mosquito development, and virus development within mosquitoes. Our results showed that these macroenvironment-occurrence models indeed succeed in anticipating breeding distributions in space and time consistently better than expected at random. The methodology that we employed tests the ability of models to anticipate distributions of breeding mosquitoes across broad areas from which no sampling is available; as such, it speaks directly to the ability to use existing sampling to understand and anticipate breeding distributions in other areas. The approach has been used with success in other spatial distribution studies conducted by Peterson et al. (2005) for dengue vectors in Mexico and by Arboleda et al. (2009) with human dengue cases in Bello, Colombia. Our ability to predict presence of breeding vector populations in unsampled areas is ideal, as it allows us to reduce effort and money and optimize their use in prevention and control programs. Although both niche modeling algorithms performed well in these initial tests, comparative studies of Maxent and GARP have shown that results differ at some level (Peterson et al. 2007); however, areas predicted as suitable by the two algorithms coincided closely, at least in broad outlines. In 2008, the two models matched only by 44.3%; however, for the other years, the match ranged 67.2–76.1% (Table 2). This coincident area from two different methods could indicate suitable sites for Ae. aegypti breeding with greater certainty. The consensus map based on predictions from all years showed that only 13.4% of Bello was consistently suitable for Ae. aegypti breeding through time. The picture that emerges is that breeding suitability is in a dynamic state that depends on both human behavior and environmental conditions. Because the additional area predicted as suitable in a given year beyond this core area was not correlated with omission error rates (P > 0.05), we rule out the possibility that area predicted depends on sample sizes used in each year to build the models. As mentioned above, an important aspect of niche modeling is its predictive abilities through time and over space; the spatial tests performed showed that our models fulfilled the spatial challenge. In subsequent tests, we evaluated the capacity of models to predict through time, and our results showed that models trained in one year had good predictive value for other years. That is, conditions predicted as suitable in a particular year can anticipate Ae. aegypti breeding sites in other years better than expected at random, particularly if the time span involved is not large. The predictions were good within initial years (i.e., from one year to the next, or to a second-third year); for predictions that were more distant in time, we obtained higher omission values, suggesting some degree of overfitting, particularly for Maxent models. This generally good predictivity opens the possibility of anticipating spatial patterns of dengue vector breeding from one year to the next by means of ecological niche modeling. While it is important to be able to predict mosquito breeding in space and time, epidemiologically, such predictions are much more meaningful if they correlate with patterns of occurrence of human dengue cases in space and time, effectively providing an indicator of transmission risk. The customary indicator, the BI, showed no association either to number of human dengue cases or to proportional area identified as suitable in our models within neighborhoods. In contrast, predictions from our models indeed were associated significantly with eventual numbers of human dengue cases by neighborhood, as was found in a previous study at broader spatial extents in Mexico (Peterson et al. 2005). Although our models explain in relatively low percentage the dengue case occurrences, we found it acceptable since the area predicted as suitable for breeding does not relate directly to the distribution of adult mosquito populations in our study area (Fox et al. 2001). These findings reinforce the ideas of Focks and Barrera (2007) that occurrence of immature stages of mosquitoes does not reflect the situation of dengue transmission risk at all accurately. However, some care with this reasoning is indicated, because other factors such as immunity of human populations can play an important role in dengue incidence, which could potentially mask predictive relationships with entomological indices (Pham et al. 2011). Some previous studies have developed spatial or temporal regression models based on mosquito breeding, using Landsat images to predict larval indices of Ae. aegypti (Estallo et al. 2008). Others have developed models for mosquito breeding based on distance to places with human populations, elevation, and slope, using niche modeling algorithms to produce risk classifications that appeared to be better than random (Khatchikian et al. 2010). However, they did not test the capacity of models to predict into unsampled areas, and used receiver operating characteristic (ROC) curves to evaluate predictive ability of the models, which have several undesirable qualities (Lobo et al. 2008, Peterson et al. 2008). Honório et al. (2009b) developed GAM models attempting to relate dengue case rates to mosquito densities in three neighborhoods of Rio de Janeiro (Brazil), but found no clear relationship, which they attributed to the possibility that infections had occurred at sites other than the residence. Studies based solely on entomological data have been able to predict mosquito breeding productivity (Aldstadt et al. 2011). Although these studies identified areas for targeted larval control, they did not relate results to dengue case rates to evaluate their epidemiological importance. De Castro-Medeiros et al. (2011) developed a stochastic cellular automata model to simulate spread of dengue in areas of Recife, Brazil. They incorporated parameters such as dengue infection rates in humans, vector behavior, and transmission cycle and concluded that human movement is an important factor in dengue transmission. Unlike these previous studies, our results could predict mosquito larval habitats accurately in space and time, and showed some correlation with dengue case rates, which speaks to their applicability. This study follows on to a previous, broader-extent application of the same methods (Peterson et al. 2005) and arrives at similar conclusions. Given that mosquito larvae can remain submerged for long periods (Clements 1992), sampling activities can be inaccurate, making Stegomyia indices inaccurate. These indices, therefore, may not be an adequate way of meeting vector surveillance needs (Clements 1992, Regis et al. 2008). We thank to Henry Pulido and Ruth Alzate of the Bello Department of Health, for their generous provision of dengue cases and Ae. aegypti breeding data. Thanks also to Jorge Delgado of the Bello Department of Cartography for generous provision of digital cartography for the municipality. Funding for this research was provided by Instituto Colombiano para el Desarrollo de la Ciencia y la Tecnología Francisco José de Caldas, COLCIENCIAS (Project: 1115–343–19131 contrato No. 360–2006, and Programa de Doctorados Nacionales-2007), and Proyecto de Sostenibilidad 2010–2011, Universidad de Antioquia, Colombia.