^{1}

^{*}

^{2}

^{1}

^{1}

^{2}

Conceived and designed the experiments: LFC JMC MLW. Analyzed the data: LFC. Contributed reagents/materials/analysis tools: LFC JMC. Wrote the paper: LFC JMC MP MLW.

The authors have declared that no competing interests exist.

The emergence of American Cutaneous Leishmaniasis (ACL) has been associated with changes in the relationship between people and forests, leading to the view that forest ecosystems increase infection risk and subsequent proposal that deforestation could reduce re-emergence of this disease.

We analyzed county-level incidence rates of ACL in Costa Rica (1996–2000) as a function of social and environmental variables relevant to transmission ecology with statistical models that incorporate breakpoints. Once social marginality was taken into account, the effect of living close to a forest on infection risk was small, and diminished exponentially above a breakpoint. Forest cover was associated with the modulation of temporal effects of El Niño Southern Oscillation (ENSO) at small spatial scales, revealing an additional complex interplay of environmental forces and disease patterns.

Social factors, which previously have not been evaluated rigorously together with environmental and climatic factors, appear to play a critical role that may ultimately determine disease risk.

American Cutaneous Leishmaniasis emergence has been associated with changes in the interaction between people and forests. The association between outbreaks and forest clearance, higher risk for populations living close to forests, and the absence of this disease from urban settings has led to the proposal that it will disappear with the destruction of primary forests. This view ignores the complex nature of deforestation as a product of socioeconomic inequities. Our study shows that such inequities, as measured by a marginalization index, may ultimately determine risk within the country, with socially excluded populations most affected by the disease. Contrary to the established view, living close to the forest edge can diminish the risk provided other factors are taken into account. Additionally, differences in vulnerability to climatic variability appear to interact with forest cover to influence risk across counties where the disease has its largest burden. Incidence exacerbation associated with El Niño Southern Oscillation is observed in counties with larger proportions of deforestation. Our study calls for control efforts targeted to socially excluded populations and for more localized ecological studies of transmission in vectors and reservoirs in order to understand the role of biodiversity changes in driving the emergence of this disease.

American cutaneous leishmaniasis (ACL), a neglected infectious disease

Here we examined county-level ACL case data from 1996 through 2000 for Costa Rica, a country that proportionally has had the largest rate of landscape transformation in the New World

The monthly number of cases of American Cutaneous Leishmaniasis (ACL) from January 1996 through December 2000 was obtained from the epidemic surveillance service of Costa Rica “^{2}) resolution was obtained from the United States Geologic Survey (USGS) [

Data on species and locations of sand fly captures were obtained from systematic reviews on human biting species from Costa Rica ^{2} grid cells and then used to define the ecological type in which each of the sand fly species was located.

This method finds spatio-temporal clusters by detecting the excess of cases in a given region under the assumption that cases are generated by an inhomogeneous Poisson point process with an intensity, μ, proportional to the population at risk. The method is implemented by moving a circular window systematically through the study area, starting at the centroid of each location in the dataset

We used this technique to analyze the patterns of clustering in potential risk factors for the disease. LISA, a local adaptation of Moran's I, compares the value of the variable of interest in a given county with those in neighboring counties. The degree of similarity between neighboring counties was compared to that expected by chance to determine where clusters of high or low values occur

We introduced breakpoints in predictors by transforming the predictor using a breakpoint basis function of the form:_{L}(x) and B_{R}(x) join each other, and are used to separate the relationship between the response and the predictors to the left and the right of the break point respectively. This technique is known as hockey stick regression

Models have the same predictors described for the GAM presented in

To make comparisons reliable, the variance over-dispersion parameter of the negative binomial response was fixed to 1, and not estimated independently for each model ^{2} test with degrees of freedom (df) defined as n-p-1, where n is the number of observations, p the number of parameters estimated in the model, and the additional df accounts for the dispersion parameter of the negative binomial. Diagnostics for spatial autocorrelation were carried out by regressing residuals on the centroids of each county. The error (ε) was assumed to be identically and independently normally distributed for the linear predictor of the NB-GLM

Parameters have a linear relationship with the response variable and were computed using ordinary least squares

In the process of model building, autoregressive components were tested but they were not significant. However, for the sake of comparison, the fitting of model (5) only included the data from 1997 through 2000. Diagnostics for spatial autocorrelation were carried out by regressing residuals on the centroids of each county.

We began by exploring whether the spatial distribution of disease incidence was heterogeneous across the country, a pattern that might be expected from the considerable heterogeneity of ecosystems in Costa Rica.

(A) Quinquennial (1996–2000) cutaneous leishmaniasis case rates (cases/population) in Costa Rica at the county level. Colors indicate clustering in monthly rates per 10,000 inhabitants obtained using the Scan method: blue corresponds to the most likely cluster, comprised of the Talamanca county, with a monthly rate 308 per 10,000 from January 1999 to December 2000 (loglikelihood ratio = 3020.06, P<0.001); green depicts the second most likely cluster, comprised of the counties of Osa, Buenos Aires, Aguirre, Perez Zeledon, Golfito, Coto Brus, Aguirre y Corredores, with a rate of 7 per 10,000 from June 1996 to November 1999 (loglikelihood ratio = 515, P<0.001); and red corresponds to the third most likely cluster, comprised of the county of Limon with a rate of 12 per 10,000 from April 1997 to May 2000 (loglikelihood ratio = 265, P<0.001). (B) The county marginalization index (See

To examine further and more quantitatively the factors determining observed spatial patterns of ACL, we fitted GAMs to the five-year ACL incidence rate (total cases during 1996 through 2000 divided by the 2000 population) as a function of several variables (see statistical methods in

A schematic representation of the breakpoint in marginalization (MI) and people living close to the forest (%close), when minimum elevation (ME) is set to 500 m and rainfall (log(MinRfll)) is set at its breakpoint. The surface illustrates major qualitative differences in disease risk as a function of the covariates. Specifically, risk increases exponentially as the proportion of people living close to the forest decreases above the breakpoint. The change has the opposite sign and decreases in magnitude for smaller values below the breakpoint. Marginality exacerbates this difference above its own breakpoint. Parameters are those of the model selected as best. This model has 7 parameters (AIC = 5768.7) and fits the data satisfactorily (Residual deviance = 79.718, df = 72, P>0.24), explains 71.34% of the deviance (null deviance = 278.108) and is not different from the more complex models presented in

Model | Log(Min Rainfall) | Margin Index | % Close | No. Parameters | AIC |

I | 7.78 | 4.13 | 49.40 | 8 | 567.7 |

II | 7.78 | 4.13 | 49.98 | 9 | 569.5 |

III | 7.77 | 4.13 | 49.99 | 9 | 569.1 |

IV | 7.77 | 4.13 | 49.99 | 10 | 570.7 |

Smooth | 7.79 | Poly 2 | Poly 2 | 8 | 574.2 |

Null | — | — | — | 4 | 595.4 |

The number of parameters does not include the dispersion parameter for the negative binomial generalized linear models, which was set to 1 (see

To address effects of hierarchically nested geopolitical units (e.g., counties belonging to provinces) and of interannual climatic variability (El Niño Southern Oscillation (ENSO)), we fitted Linear Mixed Effects Models (LMEM). These models incorporated geopolitical subdivisions of the country as nested random factors, and ENSO as a continuous predictor (details in ^{2} = 85%). The effects of ENSO are variable, with some counties showing an increase and others a decrease in incidence during a cycle of the oscillation (

(A) Local Effects of ENSO. Linear model results for a model testing for localized effects of ENSO in the counties where the disease was clustered. Color indicates clusters found with the spatio-temporal scan analysis of ^{2} = 0.85) that outperforms a similar model with the same number of parameters but that uses a first order autoregressive structure (R^{2} = 0.26) instead of ENSO. (B) Differences in forest cover for counties where the incidence diminishes or increases with ENSO. In the boxplot, 1 stands for the counties where the annual rate decreases with ENSO (Talamanca, Limón, Golfito, Buenos Aires & Coto Brus) and 2 for those where the incidence increases with ENSO (Aguirre, Corredores, Osa & Pérez-Zeledón). The difference is statistically significant as shown by a one tail Welch's t-test (a test robust to differences in variance) in which the alternative hypothesis is that the difference in forest cover between 1 and 2 is larger than 0 (

The finding that ACL tended to afflict socially marginal populations more heavily is common to other infectious diseases, and has been historically documented in public health studies particularly at small spatial scales

Risk of ACL infection in rural Costa Rica has been especially associated with the exposure to forests close to agricultural environments

Changes in landscape quality are also likely to affect composition of the arthropod vector community

Future work should examine the role of local climate variability encompassing multiple ENSO events over a longer time span, as was previously done at the coarse scale of the whole country

(0.03 MB DOC)

(0.03 MB DOC)

(0.03 MB DOC)

(0.03 MB DOC)

(0.03 MB DOC)

(0.03 MB DOC)

(0.03 MB DOC)

Intercept and ENSO are respectively the intercept and slope for Talamanca County, the reference county. For all other counties, intercept and slopes are found by adding the values in the table to the values for the reference county.

(0.04 MB DOC)

(0.12 MB PDF)

(A) Weather stations and interpolated values. Clusters of deforestation: (B) Queen contiguity. (C) 4 nearest neighbors. (D) Ecosystems of Costa Rica and number of sand fly species for each locality (see references

(0.83 MB TIF)

(A) Marginalization index. (B) % of People living within 5 km to the border of the forest. (C) Minimum elevation. (D) Log(minimum rainfall).

(0.71 MB TIF)

Top panels include the first three components from the PCA analysis presented in

(0.65 MB TIF)

We thank the personnel at Vigilancia de la Salud y Centro Centroamericano de Población of Costa Rica for sharing epidemiological and demographic data, Dr. Ivette Perfecto and Dr. John Vandermeer for their comments on the manuscript, Dr. Edward Ionides for his help with algorithms to estimate breakpoints, Dr. Martin Kulldorff for his comments on the statistical analyses, and Dr. Rodrigo Zeledón and Prof. SergioVelásquez Mazariegos for help in accessing Geospatial data.