Malaria Risk Map for Mutasa District, Zimbabwe

Published Date: 09 Oct 2017

Mufaro Kanyangarara

Objectives

The main aim of this study was to describe the spatial patterns of malaria in Mutasa District, an area characterized by epidemic and highly seasonal malaria. The objectives of the study were to identify environmental risk factors explaining the spatial distribution of malaria and develop a risk-map by making predictions at non-sampled locations.

Methods

Study area

Mutasa District is situated in the north-eastern part of Zimbabwe, bordering with Mozambique and covering an area 622 km² (Figure 1). Mutasa District stretches from 18.20° to 18.58°S latitude and from 32.71° to 33.06° E longitude. The average monthly temperature is 21.5°C. November is the hottest month with an average monthly temperature of 24.5°C and July is the coolest month with an average monthly temperature of 16.3°C. The rainy season commences from November to April, followed by a dry season. Elevation in the study area rises from 600 meters in the valleys to 2500 meters in the mountains.

Parasitological data

The malaria data was derived from ongoing community surveys in Mutasa District as a part of Southern Africa International Centers of Excellence for Malaria Research (ICEMR). The sampling and study procedures have been described elsewhere [1]. In brief, we generated a sampling frame using a high-resolution satellite image of the study area obtained from DigitalGlobe Services, Inc. (Denver, Colorado). Using the sampling frame, we then generated a random sample of households to visit from each of the selected grids. All household members were informed of the study purpose and procedures, and invited to participate. Demographic and socio-economic information and data on the use of malaria preventative measures at the individual and household level was obtained by means of standardized questionnaires administered to household members and heads of households respectively. The latitude and longitude of participating households were recorded using a handheld Global Positioning System (GPS). As part of the survey, participants were asked to give a finger prick blood sample which was tested for malaria parasites using rapid diagnostic tests (RDTs).

Environmental data

The environment is an important driver in the development and survival of the malaria parasite and mosquito vector. Using a variety of sources, we compiled a set of environmental variables corresponding to the selected households: elevation, slope, aspect, vegetation cover, land use, distances to streams of different categories, distance to the main road, distance to the nearest health facility, distance to the Mozambique border and house density. Elevation was extracted from a 90 meter high-resolution Shuttle Radar Topography Mission (SRTM) digital elevation model (DEM) [2]. DEM-derived raster maps were used to obtain slope and aspect in degrees for each participating household. The Normalized Difference Vegetation Index (NDVI) was used as a proxy for vegetation cover. Using a multispectral Landsat 8 ™ image from July 2014, available from the United States Geological Survey Land Processes Distributed Active Archive Center, an NDVI raster layer was created. Using Band 4 and Band 5 corresponding to the red and near-infrared spectral bands respectively, NDVI was calculated as: . Using the same Landsat 8 ™ image, a land use raster layer was created by performing unsupervised land use classification. Land use classes included: water, impervious, bare land, grass, crop, and forest.

Hydrologic analysis was performed on the DEM to create a stream network layer for Mutasa District containing attribute information expressing the classifications of streams using Strahler’s method. When two streams of order 1 join, a stream of order 2 is formed. Stream classifications ranged from 1 indicating low water volume streams that may only exist during the rainy season, to 4 indicating high volume streams usually found at low elevations. The two major rivers in Mutasa District, Pungwe and Honde had a stream order of 4. Using the Near tool in ArcGIS, the Euclidean (simple) distance from each household to nearest stream in each of the 4 classes was calculated. Similarly, the Near tool was used to identify the distance from each participating household to the nearest road and to the Mozambique border. Using the geographic coordinates of all households in the study area, we computed the house density as the number of houses within 250 meters, 500 meters and 1,000 meters surrounding each selected household, respectively. All images and features were projected into the World Geodetic System (WGS) 84 / Universal Transverse Mercator (UTM) Zone 36 S to allow the calculation of distances in meters.

Statistical analysis

A household containing one or more RDT positive individuals was considered a positive household. We conducted exploratory data analysis comparing environmental variables between positive households and negative households using chi-square tests for categorical variables and t-test for continuous variables. We also constructed maps of environmental variables to allow an initial inspection of spatial patterns in the environmental features in the study area.

We then undertook the following analytical steps. First, we used binomial logistic regression to estimate the probability of a positive household as a function of each of our environmental covariates. An initial multivariate logistic regression model included all independent environmental variables found to be significant (p<0.1) in the univariate analyses. Possible effect modification by season of the relationship between environmental variables and household RDT status was evaluated by considering interaction terms. A manual stepwise variable selection procedure was used and the Akaike Information Criterion (AIC) was used to examine overall model fit. The best combination of covariates that were most predictive of malaria positive households were those with the lowest value of the AIC statistic.

Our initial modeling approach used logistic regression which assumes uncorrelated or independent residuals. However, assuming that RDT status of a household is independent of that of surrounding households may not be a valid assumption as nearby households are more similar than households further apart. The presence of residual spatial variation if overlooked can lead to underestimated coefficient standard errors resulting in the spurious significant inclusion of covariates. To detect residual spatial variation, semivariograms of the standardized residuals from the logistic regression model residuals were used (Cressie 1991). Based on the appearance of the semivariogram, we selected a correlation function to characterize the correlation between pairs of households.

We performed internal and external validation of the selected model. Internal validity of the model was assessed using Monte Carlo cross-validation with 1000 iterations and 30% leave out. The sample (N=398) was randomly split; one third of households sampled in Mutasa District between October 2012 and March 2015 were assigned to the test set (n=133) and the remainder to the training set (n=265). The selected multivariate logistic regression model was then fit to the training set, and then predictions made over the test set. The observed values and model predictions of the probability of a positive household at ‘test’ locations were compared using the root mean squared error (RMSE) of prediction. After implementing 1000 iterations of this process, the RMSE was averaged. Smaller values of RMSE indicate better prediction. The external validity of the selected model was assessed by fitting the model to the sample of households enrolled on or before March 31, 2014 (n=234) and predicting the remaining households sampled from April 1, 2014 (n=164). The RMSE between the observed and predicted presence of a malaria case in a household was calculated as the index of accuracy. Sensitivity, specificity and the area under the curve (AUC) of the Receiver Operating Characteristic (ROC) were also used as additional tools to assess model performance. The greater the AUC, the more discriminatory the model is and the closer the predictions are to the observed data.

Lastly, we used geostatistical techniques to predict household RDT status at unsampled locations. A fine grid of 100x100m cells covering the whole study area was created using the Fishnet tool in ArcGIS. The values of underlying environmental variables were extracted to the centroid of each grid cell using the ArcGIS 10 Spatial Analyst Tool and our final selected model used to predict household RDT status and obtain the corresponding standard errors at the centroid locations. Inverse distance weighting was used to produce smoothed maps of household malaria risk in the rainy season and in the dry season as well as maps of corresponding uncertainty in predictions.

All spatial data manipulations, processing of environmental data and distance calculations were performed in ArcGIS 10.1 (Redlands, California). Statistical analyses including logistic regression analyses and variogram analyses were carried out in R statistical software, while inverse distance weighting and mapping were implemented in ArcGIS 10.2

Ethical considerations

The Johns Hopkins Bloomberg School of Public Health Institutional Review Board and the Medical Research Council of Zimbabwe approved this research. Permission was sought from local chiefs and written informed consent from all participants. For minors, consent was obtained from caregivers or legal guardians. Individuals with malaria based on the RDT result were treated following national treatment guidelines.

Page 1