In the real-world, we are exposed to multiple pollutants which may contribute to disease susceptibility in combination or as mixtures. In contrast, individual pollutants may have relatively small effects. Our study supports this notion that only a few pollutants were significantly associated with serum lipids levels while many individual pollutants had relatively weak associations (Figure 2 and Figure S3). The ERS as a multi-pollutant approach allows us to integrate those relatively small effects from multiple pollutants and provides a better opportunity to identify subpopulations that are at higher risk for diseases. We used multi-pollutant information at different steps of our process. Our discovery approach is different from Patel’s [5], [6] as we performed analysis with single pollutant models and then evaluated additional pollutants conditional on the identified pollutants. We then formed ERS using the set of all pollutants identified via this process using the weights from assessing them one at a time (ERS1) and jointly (ERS2). It appears that in terms of overall prediction, ERS1 and ERS2 were very similar in performance (Table 3), however, ERS2 was often slightly better
