Concentration Response Functions in ELAPSE, Medicare, and MAPLE Cohorts
Tanya Christidis, on behalf of the MAPLE, ELAPSE, and Harvard (Medicare) teams; Michael Brauer1,2; Bert Brunekreef3; Francesca Dominici4
1University of British Columbia, Vancouver, British Columbia, Canada
2Institute for Health Metrics and Evaluation, University of Washington, Seattle, Washington, USA
3Institute for Risk Assessment Sciences, Utrecht University, Utrecht, Netherlands
4Harvard T.H. Chan School of Public Health, Boston, Massachusetts, USA
Background. To better understand the relationship between air pollution and mortality at low levels, we explored how this association varies across the range of exposures by modelling the shape of the concentration response function (CRF). We focus on three Health Effects Institute funded studies assessing health effects of long-term exposure to low levels of ambient air pollution, PM2.5: the MAPLE, ELAPSE, and Harvard Medicare studies. As each study encompasses differences in exposure distributions, cohort representation, and geographical context, each took a different approach to modelling. We provide the rationale behind these methodological choices and summarize findings. We also share the results of the linear models from all three projects, for the full cohorts, and sub-populations that have only been exposed to low levels of exposure.
Methods. All three studies produced non-linear concentration response relationships. ELAPSE used a 2-knot natural spline to model the relationship between PM2.5 and natural-cause mortality in their pooled cohort, and MAPLE used a 9-knot restricted cubic spline to model the relationship between PM2.5 and non-accidental mortality. To control curvature and produce output that is usable for benefits analysis, these splines were complemented with a Shape Constrained Impact Function (SCHIF). The Harvard Medicare project examined the relationship between PM2.5 and all-cause mortality within a causal inference framework and controlled curvature with smoothing parameters. Linear models were run for each of the three studies, and for a subset of person-years with annual PM2.5 exposures below 12µg/m3.
Results. The ELAPSE spline had no low-level threshold and was supralinear. The ELAPSE SCHIF was also supralinear with a lower confidence interval (CI) that was above 1 at 3µg/m3. The MAPLE restricted cubic spline was wiggly between 5 and 12µg/m3 of PM2.5 with a lower CI above 1 starting at 2.9µg/m3. The MAPLE extended SCHIF (eSCHIF) was supralinear. The Harvard Medicare CRF had gradual changes in curvature, with slight increases from 0-5µg/m3 and 10-15µg/m3. For linear models the ELAPSE administrative cohorts had a lower hazard ratio (HR: 1.11 95%CI 1.04-1.18 per 10µg/m3) in the full population than the subpopulation with exposures under 12µg/m3 (HR: 1.20 95%CI 1.00-1.43 per 10µg/m3). The MAPLE full cohort had a higher hazard ratio (HR: 1.20 95%CI 1.00-1.43 per 10µg/m3) than the under 12µg/m3 group (HR: 1.06 95%CI 1.05-1.08 per 10µg/m3). The Harvard Medicare full cohort had a lower hazard ratio (HR: 1.07 95%CI 1.06-1.07 per 10µg/m3) than the under 12µg/m3 group (HR: 1.37 95%CI 1.34-1.40 per 10µg/m3).
Conclusions. The ELAPSE SCHIF and MAPLE eSCHIF were supralinear, indicating that with increasing concentrations of PM2.5 the marginal changes in risk appear to decline. The Medicare function was near-linear and indicated aggravated effects at all levels of PM2.5. Together, the CRFs from all three studies provide evidence for the PM2.5 -mortality association at low levels. The results of the linear models for low-exposure groups indicate that risk remains, and in some cases is elevated, even for those who have only had exposures below 12µg/m3. In future joint work, each team will model the CRF with the eSCHIF with harmonized models across each project.