Spatial Cancer Epidemiology: A Memorial Session for David Wheeler

Methods for environmental mixtures ("mixtures analyses") consider the overall effect of multiple concurrent exposures instead of solely one exposure at a time. Mixtures analyses are particularly important for environmental exposures as overall effects can differ from effects of any single exposure alone. Crucially, these analyses are often conducted in small samples and mixtures analyses generally do not account for the potential spatial dependence of health outcomes, which may lead to poor characterization of overall effect estimates. To address this, we develop a novel flexible Bayesian quantile g-computation (BQGC) approach to mixtures. This methodologic innovation not only incorporates spatial dependence, but also leverages known groupings of exposures (e.g. grouping metals as essential or non-essential) through hierarchical modeling. Using simulations, we compare our approach to other mixtures methods, such as quantile g-computation and (Bayesian) weighted quantile sums regression. We compare bias, root mean square error (RMSE), and 95% coverage probability across approaches. Relative to mixtures methods that did not address spatial dependence, RMSE was improved in BQGC when strong spatial correlation was present, though precision benefits came at the cost of bias (and therefore under-coverage) in spatial models. We demonstrate our approach using the New England Bladder Cancer Study (NEBCS), a large, population-based case-control study. We estimate associations for a mixture of heavy metals and bladder cancer incidence in NEBCS. Our innovations improve the accuracy of statistical evidence of the mixture's effect on disease risk and expand the flexibility of mixtures analysis in public health.
 

Providing spatial explanations of observable phenomena – in a formal causal sense or an informal exploratory sense – is a fundamental objective in geography, the science of place and space. Geographers ask questions about how humans use space and interact with the environment, explore physical and social mechanisms behind differences in places, and seek to understand the Earth through its human and natural complexities. Many such spatially-oriented questions cannot be directly addressed using traditional spatial regression models, which are designed primarily for purposes of spatial prediction and smoothing as opposed to spatial explanation. As an alternative to spatial regression modeling, a data analytic technique known as geographically weighted regression (GWR) was introduced in the geography literature nearly 30 years ago. Despite not receiving much attention in the statistics literature and having some well-studied limitations, GWR remains a popular data analytic tool in the fields of geography and spatial epidemiology. In this presentation, I will review some of David Wheeler's important early contributions to GWR and its model-based counterpart, the spatially-varying coefficient model. I will also make connections between the motivation for and criticisms of GWR and themes in the emerging body of statistical research on spatial causal inference.  

In many studies examining environmental risk factors for disease, researchers often rely on the location at diagnosis as the geographic reference point for assessing environmental exposures. However, environmental pollutants typically exhibit continuous variation over space and time. The dynamic nature of these pollutants, combined with population mobility in the United States, suggests that for diseases with long latency periods, such as cancer, historical exposures may be more pertinent than exposures at the time of diagnosis.

In this study, we evaluated the extent to which the common assumption of no population mobility introduces bias into the estimates of the relationship between environmental exposures and long-latency health outcomes in case-control studies. We conducted a simulation study using the residential histories of a random sample from the National Institutes of Health-AARP (formerly American Association of Retired Persons) Diet and Health Study. Case-control status was simulated based on subject exposure and true exposure effects that varied over time. We then compared estimates derived from models using only the subject's location at diagnosis with estimates assuming that subjects had experienced mobility.

Our findings indicate that ignoring population mobility leads to underestimation of subject exposure, with the largest discrepancies observed at time points more distant from the study enrollment. Generally, the impact of population mobility on the bias in estimates of the exposure-outcome relationship was more pronounced when the exposures exhibited significant spatial and temporal variability. Based on our results, we recommend incorporating residential histories when environmental exposures and disease latency periods are sufficiently long for mobility to play a critical role.