Our study has numerous limitations. The individual pollutants used to construct the ERS were identified in linear regression models with log-transformation due to skewed distributions, which assumes linear (in fact, log-linear) exposure-outcome relationships for all individual pollutants. However, not all pollutants are linearly associated with health outcomes, for example, some pesticides and/or other endocrine disrupting chemicals may have thresholds or non-monotonic dose-responses [50], [51]. Pollutants whose dose-responses were misspecified may not be selected and not contribute to the ERS. Examining non-linearity in each of the single pollutant models may identify new pollutants but construction of a simple weighted risk score like ERS would no longer be possible, which led us to a linear regression based screening strategy in this initial paper. Moreover the ERS itself may have a non-linear association with the outcome when treated as a single predictor. We used quintiles of ERS to somewhat address this issue in the association models but a completely flexible generalized additive model will be more appropriate from a statistical point of view. We tried to retain simplicity in our approach for usability and thus compromised on some finer points that may be expanded upon in the future.
