Recap: Spatial Confounding and Interference in Environmental Health Studies
Thank you to all who joined CAFE’s Spatial Confounding and Interference in Environmental Health Studies webinar with Sophie Woodward and Salvador Balkus, fourth-year PhD students in Biostatistics at Harvard T.H. Chan School of Public Health.
This webinar discussed two methodological challenges in environmental epidemiology: spatial confounding and interference. If left unaddressed, spatial confounding and interference can bias effect estimates and lead to invalid confidence intervals. Drawing on recent work, the session introduced easy-to-implement models for addressing these challenges and presented a case study demonstrating their application in practice.
Why your data might reflect bias
Tobler’s First Law of Geography states that “Everything is related to everything else, but near things are more related than distant things.” Variables like air pollution, greenspace, and socioeconomic conditions are not scattered randomly across a map, they are clustered in space. This clustering can have consequences for spatial environmental health research.
If unmeasured confounders in your study are spatially patterned, that structure can bias your estimates in ways standard regression models won’t account for. Additionally, if outcomes in one geographic unit are shaped by what’s happening in neighboring units, this bias can be also be missed in standard regression models. These are the two problems this webinar aimed to tackle: spatial confounding and spatial interference.
Spatial Confounding
Using a California air pollution dataset, Woodward demonstrated how bias from unmeasured confounders—variables that are correlated with both exposure and outcome—can be mitigated if the unmeasured confounder exhibits spatial structure. . At its core, if a missing confounder varies smoothly across space, spatial methods can use geographic information to partially account for it. Woodward introduced a spatial method using a thin plate regression spline of latitude and longitude, to adjust for this confounding. To test how robust your estimates are to potential unmeasured spatial confounders, Woodward suggested running the spatial method as a sensitivity analysis alongside your main model.
Spatial Interference
Balkus tackled a related distinct problem: What if a unit’s outcome is affected not just by its own exposure, but by neighboring unit’s exposures too? Using commuting patterns as an example he showed how ignoring this network structure biases estimates.
To address this issue, Balkus used an exposure mapping, which is a weighted summary of neighbor's exposures included as an additional variable in a model. Balkus walked through a four-step linear model procedure using Census Bureau commuting data that recovered the true effect where a standard model did not. Importantly, accounting for interference can reduce variance by explaining more of the outcome.
Related Links:
