NIRG Constrained estimation for the association between health and environmental exposures

Environment-health associations of critical public health importance, in particular relating to impacts of weather and climate, are often complex. This complexity is addressed by advanced statistical methods, used by researchers to estimate various shapes of associations, and then derive various measures of health impacts. However, in contexts with lower mortality baseline, these statistical methods may lack power and lead to unrealistic estimates of environment-health associations, which in turn leads to unaccurate health impacts estimates. This project proposes to add constraints to the statistical methods used by researchers, to include information about assumptions on the shape of environment-health associations and obtain more accurate estimates. Such constraints can impose that the risk of pollution on health increases with the level, or that the risk increase fo both heat and cold compared to the average comfortable temperatures.
The project will develop statistical method to include these constraints in two steps. The first step is to develop a general algorithm to fit models estimating association with constraints, along with uncertainty assessment. The second step is to focus on models more specifically suited to environmental epidemiology, including nonlinear associations and delayed response after exposition. Open source software code will be provided to allow other researcher easy application and replication of the developed methods. The final part of the project is to use the newly developed methods in two case studies using an international dataset of mortality and environmental exposures: i) assessment of global variations in the minimum mortality temperature, i.e. the comfortable temperature to understand differences in adaptation, and ii) assessment of the shape of the association between mortality and fine particulate matter.

Flexible regression methods used environmental exposures and health studies, such as distributed lag non-linear models, allow the estimation of complex associations. However, their flexibility becomes a curse in applications with low mortality baselines, such as small or rural areas, or younger age groups. This project proposes to add constraints to the estimation of environment-health associations, encoding prior assumptions on the shape of the association, and reducing uncertainty of estimates. Such constraints can take various forms such as imposing non-decreasing exposure-response associations between air pollutant and mortality or imposity convexity constraints for temperature-related risks on mortality. Methods to specify and estimate shape constraints will be proposed in this project, along with inference procedures. The methods will then be used to estimate the spatial variations in minimum mortality temperature, as well as in the shape of the exposure-respone function between PM2.5 and mortality. The proposed methods have several advatanges including, i) more accurate estimates of exposure-response functions and derived measures of impacts, ii) a general framework allowing a wide range of constraints as well as study designs, from time series of mortality counts to case-control studies, and iii) efficient estimation from a computational point of view.
A detailed work plan is provided, starting with the development of a fitting algorithm and inference procedures for generalized linear models with constraints on coefficients, then a shape-constrained fremwork for the distributed lag nonlinear models based on B-splines, publicly available software implementing the methods, to finish with case studies on an international database of mortality and environmental exposure.