Comments on “Causal versus spurious spatial exposure-response associations in health risk analysis” and “A cautionary tale: The characteristics of two-dimensional distributions and their effects on epidemiological studies employing an ecological design”

We are writing in response to the two review articles by Berman, Cox and Popken on associations in epidemiological studies. The authors have misinterpreted the conclusions from our study (Pan et al., 2005), created inappropriate comparison models to discredit the study, and demonstrated a lack of understanding of the public health process for addressing environmental hazards in general, and asbestos, in particular.

The article by Cox et al. (2013) focuses on the flaws of ecological studies (those examining associations between risk factors and outcomes measured at the level of a county, state, country, or other geographical unit), and shows that county-level analyses of potential risk factors for cancer incidence are subject to considerable potential bias from confounding factors. This criticism fails to acknowledge that our investigation was a case-control study at the individual level to examine the relationship between residential proximity to naturally occurring asbestos (NOA) and mesothelioma risk, after adjusting for occupational exposure to asbestos and other major confounders. Our study was not a spatially ecological study, as the outcome, the risk factor (residential proximity to NOA), and the confounders were all measured at the individual level. Most of the individual data in our study were obtained from the California Cancer Registry (CCR). Cox’s investigation did not have access to CCR’s individual-level data, and it was not able to measure individual-level exposure or adjust for individual-level confounders. Thus it is not surprising that their analyses demonstrate the pitfalls of ecological studies at the county level for relationships that depend critically on within-county variation in exposures, outcomes and confounders. Cox describes the problem as “spatial autocorrelation” but an epidemiologist would attribute the autocorrelation to confounding variables, measured or unmeasured. Cox also fails to recognize the importance of the control cases, pancreatic cancer, in our study. We are aware of the flaws of ecologic studies at the county level, which is why we did a case-control study with individual level data.

Berman’s et al.’s (2013) article raises concern by using simulations to create artificial person-level data. The simulations demonstrated that confounding can still lead to incorrect conclusions in analyses using artificial data and crude (unadjusted) regression models. The artificial data and the analyses in the Berman paper differ from our investigation, however, in several critical ways. First, the simulations used county-level cancer incidence data, and assumed that the risk was homogeneous across census-tract-level subsets of the population. Second, the simulations did not generate individual-level confounder data, nor did they allow for actual residential proximity, just census-tract centroid. Third, the analyses in the simulations were not regression-adjusted for potential confounding variables. Berman’s simulations demonstrated both at the county level and at the level of simulated individuals that unadjusted associations could still be subject to the effects of confounding, even though the simulation used county-level cancer rates for specific demographic groups. Berman’s simulations also are not directly comparable to our case-control approach as we obtained all data for both cases and controls at the individual level, and we adjusted via regression for confounding.

We concur with both Cox and Berman that ecological studies at the county level run a high risk of confounding, and that even individual-level associations can be misleading if the analysis does not adequately address person-level confounders. We acknowledged in our paper that there may be unmeasured confounding variables. We further examined the possibility that our conclusions might reflect our a priori choice of distance metric or our inclusion criteria by carrying out secondary analyses using alternate measures and excluding data from regions without NOA, and our results were very similar. Within the limits of a single study, we tried to ensure that our findings were not an artifact of data handling or analysis choices.

Berman and Cox incorrectly claim that we stated that our study showed a “causal” association of NOA and mesothelioma. We acknowledged the potential limitations of our study and the need for additional research to address those issues, but on balance we believe that our findings support an association of NOA and mesothelioma, paralleling the widely accepted relationship between mesothelioma and occupational asbestos exposure.

Unfortunately, public health officials often do not have the luxury of obtaining perfect data before acting to reduce health risks. Policies must be implemented based on existing data and reasonable conclusions.

Declaration of Interest

The authors claim full ownership for the content of this letter and declare no other special interests that may have influenced its composition. The authors and their respective employment affiliations are listed on the cover page of this letter.