Re-examining environmental correlates of Plasmodium falciparum malaria endemicity: a data-intensive variable selection approach

Daniel J. Weiss, Bonnie Mappin, Ursula Dalrymple, Samir Bhatt, Ewan Cameron, Simon I Hay, Peter W. GethingMalaria 2015 · DOI

Abstract

This analysis represents the first effort to systematically audit covariate utility for malaria mapping and then derive an objective, empirically based set of environmental covariates for modeling Plasmodium falciparum parasite rate (PfPR). The new covariates produce more reliable representations of malaria risk patterns and how they are changing through time. These covariates will be used to characterize spatially and temporally varying environmental conditions affecting PfPR within a geostatistical-modeling framework, building upon previous research by the Malaria Atlas Project (MAP) that produced global malaria maps for 2007 and 2010.

Introduction: The Challenge of Covariate Selection

To fight malaria effectively, health organizations require high-fidelity risk maps. These maps are constructed using "covariates"—gridded datasets of environmental factors like temperature, rainfall, and vegetation that correlate with malaria transmission. Historically, scientists selected these covariates based on "expert opinion" or data availability, often leading to models with limited quantitative justification.

Historical frequency of covariate categories used in malaria risk mapping, highlighting the traditional reliance on temperature and rainfall data.

Furthermore, many legacy models relied on "synoptic" data (long-term averages) which fail to capture how malaria risk fluctuates month-to-month or year-to-year. This research addresses the need for an objective, data-intensive methodology to select the most predictive environmental variables for mapping malaria across Africa, moving beyond subjective selection to an algorithmic foundation.

Methodology: A Six-Phase Selection Pipeline

The researchers at MAP developed a rigorous six-phase variable selection pipeline to overhaul the predictive model:

  1. Covariate Leveraging: Starting with 33 base datasets, the team "leveraged" them into over 50 million possible terms by applying various spatial aggregations (e.g., mean vs. max), temporal lags (1–4 months), non-linear transformations (e.g., log, square root), and pairwise interactions.
  2. Strength of Relationship: Generalized Linear Models (GLMs) were used to rank every term based on the Akaike Information Criterion (AIC), retaining only the top 10% of performers.
  3. Redundancy & Stability: Highly correlated variables were thinned out. Bootstrapping—repeatedly testing on random subsets of data—ensured the variables were robust and not the result of statistical noise.
  4. Spatial Stationarity: The team tested whether variables performed consistently across different geographical regions of Africa to ensure broad applicability.
  5. Collinearity: Variables that were too similar to one another were removed to prevent model over-fitting.
  6. Multivariate Analysis: A probabilistic approach was used to select a final "dream team" of 20 interaction terms.

Data Sources

The study utilized a massive repository of survey and environmental data:

  • Response Variable: Plasmodium falciparum parasite rate (PfPR) data from 13,680 survey points across Africa spanning 2000–2012.
  • Environmental Covariates:
    • MODIS Satellite Data: Land Surface Temperature (LST), Enhanced Vegetation Index (EVI), and Tasseled Cap Wetness/Brightness.
  • Climate/Topography: WorldClim precipitation data, SRTM elevation and slope models, and various aridity indices.
  • Socio-economic Factors: AfriPop population density, Nighttime lights (as a proxy for development), and accessibility metrics (travel time to urban centers).

Results: The Power of Interaction and Dynamics

The automated selection process substantially outperformed previous expert-led models. The pseudo-R² value increased from 0.43 to 0.52, representing a major leap in predictive power for geospatial disease modeling. Key findings include:

  • Interactions are Essential: Every one of the final 20 variables was an interaction term (the product of two different variables). This suggests that isolated factors (e.g., temperature) are far less predictive than how factors collide (e.g., the relationship between heat and vegetation density).
  • The Dynamic-Synoptic Interplay: The most powerful predictors combined "synoptic" data (local norms) with "dynamic" data (current conditions). This allows the model to capture anomalies—deviations from the norm that often trigger malaria outbreaks.
  • The "V4" Variable: The single most influential variable was an interaction between mean annual precipitation and the monthly EVI anomaly. Essentially, malaria thrives where high baseline rainfall meets a sudden spike in vegetation "greenness."

Discussion and Future Directions

While the new model is significantly more accurate, it introduces a trade-off in interpretability. Complex transformations, such as a square-root transformed monthly EVI anomaly multiplied by log-base-10 annual precipitation, are harder to explain to policy-makers than simple rainfall metrics. Additionally, processing 50 million variables requires massive computational resources, necessitating a cap on the final model size.

Moving forward, the authors plan to integrate these optimized covariates into Bayesian geostatistical frameworks. This will facilitate the creation of the next generation of high-resolution, time-varying malaria risk maps, providing a more dynamic tool for global health intervention.

Summary updated

Abstract

This analysis represents the first effort to systematically audit covariate utility for malaria mapping and then derive an objective, empirically based set of environmental covariates for modeling PfPR. The new covariates produce more reliable representations of malaria risk patterns and how they are changing through time, and these covariates will be used to characterize spatially and temporally varying environmental conditions affecting PfPR within a geostatistical-modeling framework, thus building upon previous research by MAP that produced global malaria maps for 2007 and 2010.
BibTeX
@inproceedings{weiss2015re,
  author = {Daniel J. Weiss and Bonnie Mappin and Ursula Dalrymple and Samir Bhatt and Ewan Cameron and Simon I Hay and Peter W. Gething},
  title = {{Re-examining environmental correlates of Plasmodium falciparum malaria endemicity: a data-intensive variable selection approach}},
  booktitle = {Malaria},
  year = {2015},
  doi = {10.1186/s12936-015-0574-x},
}
Figure