Residual Spatial Autocorrelation Research Articles

Handedness predicts Conservative-Republican preference and eliminates relations of Big Five personality to political orientation using the 48 contiguous American states as analytical units

ABSTRACTThe two present nomothetic studies focused on the period from 1996 to 2012 to determine relations between handedness and political orientation using the 48 contiguous American states as analytical units. The estimated percentage of left-handers in each state operationally defined handedness. A composite measure of Conservative-Republican preference was created from CBS/New York Times/Gallup polls of state resident conservatism and the percent in each state voting Republican in each presidential election from 1996 to 2012. Study 1 showed that state levels of left-handedness correlated to an extremely high degree with Conservative-Republican preference (r = −.80). As well, with common demographic differences between states reflected in socioeconomic status, White population percent, and urban population percent controlled through multiple regression, handedness still accounted for an additional 37.2% of the variance in Conservative-Republican preference. Study 2 found that each of the Big Five personality variables correlated significantly with handedness and with Conservative-Republican preference, but in the opposite direction. Furthermore, Study 2 demonstrated quite surprisingly that all Big Five personality relations to Conservative-Republican preference were eliminated when handedness was controlled in multiple regression equations. For all regression equations, the global Moran’s I test specifically developed for detecting residual spatial autocorrelation indicated no significant spatial autocorrelation.

Landscape and anthropogenic factors affecting spatial patterns of water quality trends in a large river basin, South Korea

Understanding changes in water quality over time and landscape and anthropogenic factors affecting them are of paramount importance to human and ecosystem health. We analyzed the seasonal trends of total nitrogen, total phosphorus, chemical oxygen demand, and total suspended solid (SS) in the Han River Basin (HRB) of South Korea using the Mann-Kendall test. We explored the effects of anthropogenic (land cover and population) and natural factors (topography and soil) on the trends by using Moran’s Eigenvector based spatial filtering regressions at four different spatial scales. Water quality of the HRB generally improved from the early 1990s to 2016 with decreasing summer nutrient and winter SS concentrations. Water quality trends were spatially autocorrelated with distinct spatial variations within the basin. Some stations close to the Seoul metropolitan area, however, still exhibited poor water quality conditions. Approximately 20–70 percent of spatial variation of different water quality trends were explained by some combination of current agricultural land cover, forest land cover, % area covered by water, change in those land covers and slope variations. The 100 m buffer and one-kilometer upstream scale analyses generally showed higher explanatory power than the sub-watershed scale analyses, while the effect of seasons differed for different parameters. The significant factors in each regression model typically differed among different scales but not among different seasons of the same scale. The spatial filtering approach removed the residual spatial autocorrelation and thus significantly increased the explanatory power of water quality trend models.

A guide to Bayesian model checking for ecologists

AbstractChecking that models adequately represent data is an essential component of applied statistical inference. Ecologists increasingly use hierarchical Bayesian statistical models in their research. The appeal of this modeling paradigm is undeniable, as researchers can build and fit models that embody complex ecological processes while simultaneously accounting for observation error. However, ecologists tend to be less focused on checking model assumptions and assessing potential lack of fit when applying Bayesian methods than when applying more traditional modes of inference such as maximum likelihood. There are also multiple ways of assessing the fit of Bayesian models, each of which has strengths and weaknesses. For instance, BayesianPvalues are relatively easy to compute, but are well known to be conservative, producingPvalues biased toward 0.5. Alternatively, lesser known approaches to model checking, such as prior predictive checks, cross‐validation probability integral transforms, and pivot discrepancy measures may produce more accurate characterizations of goodness‐of‐fit but are not as well known to ecologists. In addition, a suite of visual and targeted diagnostics can be used to examine violations of different model assumptions and lack of fit at different levels of the modeling hierarchy, and to check for residual temporal or spatial autocorrelation. In this review, we synthesize existing literature to guide ecologists through the many available options for Bayesian model checking. We illustrate methods and procedures with several ecological case studies including (1) analysis of simulated spatiotemporal count data, (2) N‐mixture models for estimating abundance of sea otters from an aircraft, and (3) hidden Markov modeling to describe attendance patterns of California sea lion mothers on a rookery. We find that commonly used procedures based on posterior predictivePvalues detect extreme model inadequacy, but often do not detect more subtle cases of lack of fit. Tests based on cross‐validation and pivot discrepancy measures (including the “sampled predictivePvalue”) appear to be better suited to model checking and to have better overall statistical performance. We conclude that model checking is necessary to ensure that scientific inference is well founded. As an essential component of scientific discovery, it should accompany most Bayesian analyses presented in the literature.

Effects of spatial autocorrelation and imperfect detection on species distribution models

Abstract Species distribution models (SDMs) are widely used in ecology and related fields. They are frequently adopted to predict the expected occurrence (presence/absence) or abundance over large spatial scales, that is, to produce a species distribution map. Two issues that almost universally affect these models are measurement errors (especially imperfect detection) and residual spatial autocorrelation. We explored the effects of imperfect detection and autocorrelation in abundance models by simulating datasets which did or did not contain these two effects and analysing them with four different models that did or did not accommodate them. Specifically, we used a Poisson GLM as a baseline, an N‐mixture model accounting only for imperfect detection and two N‐mixture models that accounted for imperfect detection, but differed in their specification of spatial autocorrelation (CAR random effects vs. two‐dimensional splines). In a case study, we then applied these models to Common Redstart Phoenicurus phoenicurus data from the second Swiss Breeding Bird Atlas (1993–1996) and validated them using an independent monitoring dataset. We found that both imperfect detection and autocorrelation strongly affected the quality and the uncertainty of abundance maps, especially when they occurred together. Spatial N‐mixture models were well able to estimate the true abundance maps. Explicit modelling of measurement error and spatial autocorrelation can thus greatly improve the quality of SDMs and should not be ignored when producing large‐scale abundance or occurrence maps.

Accounting for and Predicting the Influence of Spatial Autocorrelation in Water Quality Modeling

Several studies in the hydrology field have reported differences in outcomes between models in which spatial autocorrelation (SAC) is accounted for and those in which SAC is not. However, the capacity to predict the magnitude of such differences is still ambiguous. In this study, we hypothesized that SAC, inherently possessed by a response variable, influences spatial modeling outcomes. We selected ten watersheds in the USA and analyzed if water quality variables with higher Moran’s I values undergo greater increases in the coefficient of determination (R2) and greater decreases in residual SAC (rSAC). We compared non-spatial ordinary least squares to two spatial regression approaches, namely, spatial lag and error models. The predictors were the principal components of topographic, land cover, and soil group variables. The results revealed that water quality variables with higher inherent SAC showed more substantial increases in R2 and decreases in rSAC after performing spatial regressions. In this study, we found a generally linear relationship between the spatial model outcomes (R2 and rSAC) and the degree of SAC in each water quality variable. We suggest that the inherent level of SAC in response variables can predict improvements in models before spatial regression is performed.

Environmental uniformity, site quality and tree competition interact to determine stand productivity of clonal Eucalyptus

Abstract The major determinants of productivity in Eucalyptus plantations include the site productive capacity, local environmental uniformity, and the tree genetics, which impact their growth potential and competitiveness. While these factors have received a lot of attention individually, less attention has been given to the relative importance of these factors and the interactions between them. We examined these factors and their interactions using eight experiments located along a gradient in site productivity, consisting of 54 clonal genotypes (Eucalyptus urophylla × E. grandis hybrids) at two plant spacings (3 × 3.33 and 3 × 2 m). Stand productivity was quantified as the mean annual increment at age 6 years (m3 ha−1 year−1). Site productive capacity was represented using a site index based on mean dominant height at age 6 years. Environmental uniformity was calculated from within-plot variability in growth of very young stands using residual spatial autocorrelation. The competitive ability of a given phenotype was calculated as the percent loss in potential growth due to neighbor competition. Productivity increased with all variables but site quality had the greatest effect, followed by environmental uniformity and then competitive ability, with corresponding changes in productivity of about 30, 20 and 10 m3 ha−1 year−1, respectively, across the range of the given variable. With higher environmental heterogeneity, competition among genotypes increased, especially at dense spacing, showing that in such situations, highly competitive genotypes may display productivities similar to the less competitive genotypes in uniform environments. Experimental plots distributed across site quality and uniformity gradients provide opportunity for greater genetic differentiation than plots covering more homogeneous environments, reinforcing the recommendation of establishing genetic tests in locations with extreme site qualities. Our work suggests that the G × E interaction is linked to site quality and, to a lesser degree, to the local environmental uniformity. This work shows the value of adopting environmental gradient-based approaches in tree genetic testing and clone recommendation as a way to more accurately match genotypes to specific sites.

Mapping of soil properties at high resolution in Switzerland using boosted geoadditive models

Abstract. High-resolution maps of soil properties are a prerequisite for assessing soil threats and soil functions and for fostering the sustainable use of soil resources. For many regions in the world, accurate maps of soil properties are missing, but often sparsely sampled (legacy) soil data are available. Soil property data (response) can then be related by digital soil mapping (DSM) to spatially exhaustive environmental data that describe soil-forming factors (covariates) to create spatially continuous maps. With airborne and space-borne remote sensing and multi-scale terrain analysis, large sets of covariates have become common. Building parsimonious models amenable to pedological interpretation is then a challenging task. We propose a new boosted geoadditive modelling framework (geoGAM) for DSM. The geoGAM models smooth non-linear relations between responses and single covariates and combines these model terms additively. Residual spatial autocorrelation is captured by a smooth function of spatial coordinates, and non-stationary effects are included through interactions between covariates and smooth spatial functions. The core of fully automated model building for geoGAM is component-wise gradient boosting. We illustrate the application of the geoGAM framework by using soil data from the Canton of Zurich, Switzerland. We modelled effective cation exchange capacity (ECEC) in forest topsoils as a continuous response. For agricultural land we predicted the presence of waterlogged horizons in given soil depths as binary and drainage classes as ordinal responses. For the latter we used proportional odds geoGAM, taking the ordering of the response properly into account. Fitted geoGAM contained only a few covariates (7 to 17) selected from large sets (333 covariates for forests, 498 for agricultural land). Model sparsity allowed for covariate interpretation through partial effects plots. Prediction intervals were computed by model-based bootstrapping for ECEC. The predictive performance of the fitted geoGAM, tested with independent validation data and specific skill scores for continuous, binary and ordinal responses, compared well with other studies that modelled similar soil properties. Skill score (SS) values of 0.23 to 0.53 (with SS = 1 for perfect predictions and SS = 0 for zero explained variance) were achieved depending on the response and type of score. GeoGAM combines efficient model building from large sets of covariates with effects that are easy to interpret and therefore likely raises the acceptance of DSM products by end-users.

The challenge of estimating a residual spatial autocorrelation from forest inventory data

Estimates of stand averages are needed by forest management for planning purposes. In forest enterprise inventories supported by remotely sensed auxiliary data, these estimates are typically derived exclusively from a model that does not consider stand effects in the study variable. Variance estimators for these means may seriously underestimate uncertainty, and confidence intervals may be too narrow when a model used for computing a stand mean omits a nontrivial stand effect in one or more of the model parameters, a nontrivial spatial distance dependent autocorrelation in the model residuals, or both. In simulated sampling from 36 populations with stands of different sizes and differing with respect to (i) the correlation between a study variable (Y) and two auxiliary variables (X), (ii) the magnitude of stand effects in the intercept of a linear population model linking X to Y, and (iii) a first-order autoregression in Y and X, we learned that none of the tested designs provided reliable estimates of the within-stand autocorrelation among model residuals. More-reliable estimates were possible from stand-wide predictions of Y. The anticipated bias in an estimated autoregression parameter had a modest influence on estimates of variance and coverage of nominal 95% confidence intervals for a synthetic stand mean.

Accounting for the temporal variation of spatial effect improves inference and projection of population dynamics models

Population dynamics models incorporating density dependence and habitat heterogeneity are useful tools to explain and project the spatiotemporal variation of wildlife abundance. Despite their wide application in ecology and conservation biology, the inference and projection of these models may be problematic when residual spatial autocorrelation (SAC) is found. We aimed to improve the inference and projection of population dynamics models by accounting for residual SAC. We considered three Gompertz models that incorporated density dependence and the effect of wetland habitat to explain and project the abundance of Mallard (Anas platyrhynchos). We compared a conventional model that did not account for residual SAC (ENV) with two novel models accounting for residual SAC, one incorporating a spatial effect (a spatially autocorrelated process error) that did not vary over time (STA) and the other incorporating a spatial effect that varied over time (DYN). We evaluated model inference using data from 1974 to 1998 and projection using data from 1999 to 2010. We then forecasted Mallard abundance from 2011 to 2100 under different levels of wetland habitat loss. The DYN model eliminated residual SAC and had better model fit than the ENV and STA models (ΔD¯=2498.3 and 1988.8, respectively). The projection coverage rate of the DYN model was the closest to the nominal value among the three models. The DYN model forecasted smaller areas with decrease in Mallard abundance under future wetland habitat loss than the ENV and STA models. The novel and conventional population dynamics models we considered in this study, combined with the practical model evaluation approach, can provide reliable inference and projection of wildlife abundance, and thus have wide application in ecological studies and conservation practices that aim to understand and project the spatiotemporal variation of wildlife abundance under environmental changes. In particular, when conservation decision-making is based on model projections, the DYN may be used to minimize the risk of reducing conservation effort in areas that still have high conservation value, due to its favorable projection performance.

Spatial autocorrelation reduces model precision and predictive power in deforestation analyses

AbstractGeneralized linear models are often used to identify covariates of landscape processes and to model land‐use change. Generalized linear models however, overlook the spatial component of land‐use data, and its effects on statistical inference. Spatial autocorrelation may artificially reduce variance in observations, and inflate the effect size of covariates. To uncover the consequences of overlooking this spatial component, we tested both spatially explicit and non‐spatial models of deforestation for Colombia. Parameter estimates, analyses of residual spatial autocorrelation, and Bayesian posterior predictive checks were used to compare model performance. Significant residual correlation showed that non‐spatial models failed to adequately explain the spatial structure of the data. Posterior predictive checks revealed that spatially explicit models had strong predictive power for the entire range of the response variable and only failed to predict outliers, in contrast with non‐spatial models, which lacked predictive power for all response values. The predictive power of non‐spatial models was especially low in regions away from Colombia's center, where about half the observations were clustered. While all analyses consistently identified a core of important covariates of deforestation rates, predictive modeling requires parameter estimates informed by the spatial structure of the data. To inform increasingly important forest and carbon sequestration policy, land‐use models must account for spatial autocorrelation.

Choose your neighborhood wisely: implications of subsampling and autocorrelation structure in simultaneous autoregression models for landscape ecology

Context Large datasets that exhibit residual spatial autocorrelation are common in landscape ecology, introducing issues with model inference. Computationally intensive statistical techniques such as simultaneous autoregression (SAR) are used to provide credible inference, yet landscape studies make choices about autocorrelation structure and data reduction techniques without adequate understanding of the consequences for model estimation and inference.

Effects of different sampling densities on geographically weighted regression kriging for predicting soil organic carbon

Geographically weighted regression kriging (GWRK) is a popular interpolation method, considering not only spatial parametric non-stationarity and relationship between target and explanatory variables, but also spatial autocorrelation of residuals. However, little attention has been paid to the effects of different sampling densities on GWRK technique for estimating soil properties. Objectives of this study were: (i) comparing the GWRK predictions with those obtained from multiple linear regression kriging (MLRK) and ordinary kriging (OK), and (ii) examining how different sampling densities affect the performance of GWRK for predicting soil organic carbon (SOC). Soil samples were simulated with four sampling densities, including 0.010, 0.020, 0.041, and 0.082 sites/km2. The results showed that GWRK made less prediction errors and outperformed MLRK and OK in the case of a high sampling density, with the root mean squared errors of GWRK<MLRK<OK and coefficient of determination of GWRK>MLRK>OK. However, in the case of a low sampling density, GWRK generated larger prediction errors, exhibiting a poorer performance than MLRK and OK. Accordingly, we conclude that GWRK can be considered as the best approach for predicting SOC in these three approaches with sufficient data points, but it has a poorer performance than the other methods with sparse data points.

A Community Needs Index for Adolescent Pregnancy Prevention Program Planning: Application of Spatial Generalized Linear Mixed Models.

Objective The objective is to estimate community needs with respect to risky adolescent sexual behavior in a way that is risk-adjusted for multiple community factors. Methods Generalized linear mixed modeling was applied for estimating teen pregnancy and sexually transmitted disease (STD) incidence by postal ZIP code in New York State, in a way that adjusts for other community covariables and residual spatial autocorrelation. A community needs index was then obtained by summing the risk-adjusted estimates of pregnancy and STD cases. Results Poisson regression with a spatial random effect was chosen among competing modeling approaches. Both the risk-adjusted caseloads and rates were computed for ZIP codes, which allowed risk-based prioritization to help guide funding decisions for a comprehensive adolescent pregnancy prevention program. Conclusions This approach provides quantitative evidence of community needs with respect to risky adolescent sexual behavior, while adjusting for other community-level variables and stabilizing estimates in areas with small populations. Therefore, it was well accepted by the affected groups and proved valuable for program planning. This methodology may also prove valuable for follow up program evaluation. Current research is directed towards further improving the statistical modeling approach and applying to different health and behavioral outcomes, along with different predictor variables.

Avaliação da dependência espacial na modelagem do desempenho da segurança viária em zonas de tráfego

Uma técnica comumente usada no processo de modelagem do Desempenho da Segurança Viária (DSV), no nível de planejamento, são os Modelos Lineares Generalizados (MLG) assumindo a distribuição binomial negativa para os erros. Limitações dessa técnica, por não considerar os efeitos espaciais, têm sido contornadas com a utilização de modelos espaciais locais a partir de técnicas de regressão espacial como a Regressão de Poisson Geograficamente Ponderada (RPGP). Este trabalho tem por objetivo apresentar uma análise comparativa entre modelos de previsão de acidentes globais não espaciais e locais espaciais para a estimação do DSV agregado ao nível de zonas de tráfego em Fortaleza/CE. Foram calibrados modelos para a variável dependente acidentes totais e acidentes com vítimas e os resultados mostraram que os modelos RPGP apre-sentaram melhor desempenho que os MLG nas medidas de ajustes e na redução da autocorrelação espacial dos resíduos, sendo capazes de captar a heterogeneidade espacial da frequência dos acidentes.

Modelling the influence of biotic factors on species distribution patterns

Biotic interactions can have large effects on species distributions yet their role in shaping species ranges is seldom explored due to historical difficulties in incorporating biotic factors into models without a priori knowledge on interspecific interactions. Improved SDMs, which account for biotic factors and do not require a priori knowledge on species interactions, are needed to fully understand species distributions. Here, we model the influence of abiotic and biotic factors on species distribution patterns and explore the robustness of distributions under future climate change. We fit hierarchical spatial models using Integrated Nested Laplace Approximation (INLA) for lagomorph species throughout Europe and test the predictive ability of models containing only abiotic factors against models containing abiotic and biotic factors. We account for residual spatial autocorrelation using a conditional autoregressive (CAR) model. Model outputs are used to estimate areas in which abiotic and biotic factors determine species’ ranges. INLA models containing both abiotic and biotic factors had substantially better predictive ability than models containing abiotic factors only, for all but one of the four species. In models containing abiotic and biotic factors, both appeared equally important as determinants of lagomorph ranges, but the influences were spatially heterogeneous. Parts of widespread lagomorph ranges highly influenced by biotic factors will be less robust to future changes in climate, whereas parts of more localised species ranges highly influenced by the environment may be less robust to future climate. SDMs that do not explicitly include biotic factors are potentially misleading and omit a very important source of variation. For the field of species distribution modelling to advance, biotic factors must be taken into account in order to improve the reliability of predicting species distribution patterns both presently and under future climate change.

Crime Risk Estimation with a Commuter-Harmonized Ambient Population

Residential population data are frequently employed to link the crime incidence of an area with the number of residents to estimate the underlying risk. Human mobility patterns cause shifts in the baseline population, however, that can potentially influence the crime statistics. This study therefore employed an ambient population that combined residential population data with data depicting the commuting activity in small administrative areas. The effects of the commuter-harmonized ambient population on crime were then evaluated in a series of negative binomial regression models. The models also controlled for criminogenic factors and incorporated eigenvector spatial filtering to adjust for spatial effects. The results show significant effects of commuting patterns on crime outcomes. For certain crimes, such as violence, theft, and disorder, the inbound commuters are significantly associated with high risk. It was further discovered that an offset variable comprising the commuter-harmonized ambient population data models the crime outcomes more reliably than when residential population data are used. Spatial filtering was found to effectively eradicate residual spatial autocorrelation after accounting for effects of the predictor variables. We conclude that calculating crime rates using the residential population does not constitute an accurate risk measure and that the ambient population has crucial implications for realistic and reliable target representation and crime modeling.

Exploring non-stationary and scale-dependent relationships between walleye (Sander vitreus) distribution and habitat variables in Lake Erie

Global regression techniques that assume spatial stationarity are usually used to study the interaction between aquatic species distribution and habitat variables. In the present study, a local regression model, named geographically weighted regression (GWR), was used to question the spatial stationarity assumption in exploring the relationships between walleye (Stizostedion vitreum) distribution and habitat variables in Lake Erie. The GWR model resulted in a significant improvement of model performance over the two global linear and non-linear regression methods (a generalised least-squares (GLS) model and a generalised additive mixed model (GAMM)), accounting for residual spatial autocorrelation using the same response and explanatory variables as in the GWR model. The values of local regression coefficients from the GWR model changed among spatial locations significantly, implying spatially varying and scale-dependent relationships between walleye distribution and habitat variables. The k-means cluster analyses based on the t-values of local regression coefficients of GWR model characterised special zones of species–environment relationships of walleye in Lake Erie. In conclusion, spatial stationarity needs to be questioned in studying the relationships between aquatic species distribution and habitat variables and a non-stationary approach, such as GWR, is recommended as a complementary tool.

Predicting the Influence of Multi-Scale Spatial Autocorrelation on Soil-Landform Modeling

Although numerous soil–landform modeling investigations have documented the effects and importance of spatial autocorrelation (SAC), little is known about how to predict the magnitude of those effects from the degree of SAC in the model variables. In this study, we quantified the SAC inherent in soil and landform variables of four widely divergent pedogeomorphological systems around the world to examine general relationships between SAC and spatial regression model results. Spatial regressions were performed by incorporating spatial filters, extracted by spatial eigenvector mapping, into non-spatial models as additional predictor variables. Results indicated that incorporation of spatial filters improved the performance of the non-spatial regressions—increases in R² and decreases in both Akaike Information Criterion (AIC) and residual SAC were observed. More remarkable was that the degree of improvement was strongly and linearly related (i.e., proportional) to the level of SAC inherently possessed by each soil variable. Our findings show that spatial modeling outcomes are sensitive to the degree of SAC possessed by a soil property when treated as a response variable. Thus, the level of SAC present in a soil variable can serve as a direct indicator for how much improvement a non-spatial model will undergo if that SAC is appropriately taken into account.

Incorporating spatial regression model into cellular automata for simulating land use change

Cellular automata (CA) model has become one of the most widely used spatial explicit methods for simulating land use change in recent decades. There are many CA models that apply ordinary least squares (OLS) regression model to estimate the weights of land-use driving factors. In these models, however, the dependent variable, which is normally the amount of change in land use, may be correlated with each other spatially. This would result in spatial dependency in residuals causing higher standard deviations of estimations that might accordingly influence the overall performance of the model. In order to solve the problem, this paper applies a spatial regression model to explain the amount of land use change as a function of land-use factors in the CA model. The results are further compared with that of an OLS-based CA model. The Local Indicators of Spatial Association (LISA) analysis showed that there was significant spatial autocorrelation (P < 0.05) in residuals of the OLS estimation, while the spatial autocorrelation in residuals had been significantly reduced when spatial regression was applied. Based on the coefficients of each land use driving factor in both models, weights of factors were determined and the transition potential maps were generated. Cramer's V indicators suggested that map of transition potentials created using spatial regression had better explanatory power. Finally, both maps were implemented into CA simulations, and the validation of each model indicated that the CA simulation based on spatial regression was more accurate according to Kappa statistics. The results show that the integration of spatial regression into the CA model improved the performance of land-use modeling for this region and thus could be further applied in related studies.

Spatially variable risk factors for malaria in a geographically heterogeneous landscape, western Kenya: an explorative study

BackgroundLarge reductions in malaria transmission and mortality have been achieved over the last decade, and this has mainly been attributed to the scale-up of long-lasting insecticidal bed nets and indoor residual spraying with insecticides. Despite these gains considerable residual, spatially heterogeneous, transmission remains. To reduce transmission in these foci, researchers need to consider the local demographical, environmental and social context, and design an appropriate set of interventions. Exploring spatially variable risk factors for malaria can give insight into which human and environmental characteristics play important roles in sustaining malaria transmission.MethodsOn Rusinga Island, western Kenya, malaria infection was tested by rapid diagnostic tests during two cross-sectional surveys conducted 3 months apart in 3632 individuals from 790 households. For all households demographic data were collected by means of questionnaires. Environmental variables were derived using Quickbird satellite images. Analyses were performed on 81 project clusters constructed by a traveling salesman algorithm, each containing 50–51 households. A standard linear regression model was fitted containing multiple variables to determine how much of the spatial variation in malaria prevalence could be explained by the demographic and environmental data. Subsequently, a geographically-weighted regression (GWR) was performed assuming non-stationarity of risk factors. Special attention was taken to investigate the effect of residual spatial autocorrelation and local multicollinearity.ResultsCombining the data from both surveys, overall malaria prevalence was 24 %. Scan statistics revealed two clusters which had significantly elevated numbers of malaria cases compared to the background prevalence across the rest of the study area. A multivariable linear model including environmental and household factors revealed that higher socioeconomic status, outdoor occupation and population density were associated with increased malaria risk. The local GWR model improved the model fit considerably and the relationship of malaria with risk factors was found to vary spatially over the island; in different areas of the island socio-economic status, outdoor occupation and population density were found to be positively or negatively associated with malaria prevalence.DiscussionIdentification of risk factors for malaria that vary geographically can provide insight into the local epidemiology of malaria. Examining spatially variable relationships can be a helpful tool in exploring which set of targeted interventions could locally be implemented. Supplementary malaria control may be directed at areas, which are identified as at risk. For instance, areas with many people that work outdoors at night may need more focus in terms of vector control.Trial registration: Trialregister.nl NTR3496—SolarMal, registered on 20 June 2012