Quantifying uncertainty in climate-driven disease risk predictions

This thesis considers the uncertainty in forecasts of climate-driven disease risk, focusing on seasonal and decadal timescales. An analysis of the skill of decadal climate predictions is carried out, looking at the first multi-model decadal hindcast set produced as part of the ENSEMBLES project. Some skill in the prediction of global average temperature trends over the forthcoming decade is shown, with no skill evident for precipitation. Focusing on smaller areas shows limited skill in predicting temperature trends and no skill for precipitation trends, suggesting that decadal climate models cannot currently make useful predictions of disease risk. Seasonal climate forecasting skill is then considered. Seasonal hindcasts produced by two research projects, DEMETER and ENSEMBLES, are compared with the most recent version of the European Centre for Medium-Range Weather Forecast’s seasonal forecast model, System 4. The models are validated over Africa and the Indian subcontinent, and it is shown that in general System 4 forecasts are an improvement over the DEMETER and ENSEMBLES multimodel ensembles, particularly for West Africa. A more in depth study of System 4 is subsequently carried out, comparing the variation in skill between forecast start dates. Forecast value is demonstrated at multiple lead times, with most skill found for West African regions and Botswana and limited skill for India; indicating when and where forecasts can potentially be issued to users. Forecasting malaria is then studied by using Liverpool Malaria Model (LMM) driven by System 4. Skill is demonstrated over Botswana, particularly for forecasts issued in November, validating against laboratory confirmed cases of malaria. This is an improvement on previous work where the LMM was driven with the DEMETER seasonal hindcasts. Where malaria data is not available, System 4-driven LMM hindcasts are compared to LMM driven by ERA-Interim in a tier-2 validation context. Skill is demonstrated at the epidemic fringe of the Sahel and in north west Malawi, whilst the Gulf of Guinea shows no skill. This is consistent with previous work suggesting the LMM performs better in epidemic than in endemic regions. A method for interpreting hindcast validation results as uncertainty quantification is then presented. Finally, the uncertainty in the relationship between seasonal average climate and malaria risk is analysed, using the LMM driven by the 20th century reanalysis dataset. The relationship parameters describing seasonal average climate and malaria risk is explored and impact surfaces are created, relating seasonal average temperature and precipitation to average seasonal malaria incidence. The robustness of these impact surfaces is investigated by comparing the surfaces associated with different LMM survival schemes. A method of combining impact surfaces based on tercile categories is described and implemented and it is demonstrated how the resulting graphic could be integrated with a seasonal ensemble forecast system. Such a tool is potentially useful for decision-makers, allowing an intuitive visual communication of the quantified uncertainty in predicting climate-driven disease risk at seasonal timescales.