Climate impacts

Exploring how physical science influence statistical climate impact models

Statistical climate impact models estimate how climatic factors like temperature and precipitation influence a human outcome of interest like economic activity (GDP) or agricultural output. In this project we explore how considerations from the physical Earth sciences bias these models and how those biases could potentially be addressed.

I am also working on including this type of statistical impact assessment in coupled human-climate models. The objective of this type of work is to provide data-driven coupled models that can capture human-environmental dynamics by drawing on observations of how human systems have responded historically.

In our 2025 Climatic Change paper we interigate these statistical models models; we consider the role of:

  • measurement error
  • correlation between independent variables
  • mean climate state

Measurement error in this context often occurs because of interpolation used for gridded data products. Correlation between temperature (T) and precipitation (P) comes from physical coupling through mechanisms like evapotranspiration, but can lead to bias. Means are often subtracted as part of the statistical analysis (through the use of spatial fixed effects), but this may obscure a substantial part of the longer term climate signal.

The length scales of temperature and precipitation are different and measurement systems are heterogeneous. (a-b) Measurement station locations for temperature (T) and precipitation (P) in the Global Historical Climatology Network (GHCN) database and their most recent year of reported data in color. (c) The distance [km] from every location on Earth to the nearest 2021 P measurement (note >100km is grouped). (d-i) Correlations between monthly local P and T anomalies measured in three regions. The pair-wise correlations between station anomalies are shown with an exponential fit to estimate the e-folding distance, or characteristic spatial scale. Horizontal axes differ. Unless otherwise specified, the rest of the study uses UDEL gridded T and P, which are largely based on the GHCN.