Methods/Statistics
Using distributed lag models to account for nonlinearity of the lag-response association a study of air pollution and incident breast cancer Judy Ou* Judy Ou Jennifer Ish Alison Rector Alexandra White
Background: Understanding the lag-response association may be critical for outcomes with long latency periods. As common statistical models may assume a linear lag-response function, we used distributed lag models to determine if non-linear modeling of the lag-response function improved model fit in a study of air pollution and breast cancer.
Methods: Using data from the Sister Study, we included 34,616 women (ncases=2,704) with 15 years of continuous air pollution exposure history available before diagnosis or end of follow-up. Spatiotemporal models estimated fine particulate matter (PM2.5) and nitrogen dioxide (NO2) for participant residences from 1990 to 2017. The lag-response function was modeled linearly (standard Cox model), stratified every 3 years, and as a quadratic B-spline with a knot at 7.5 years. Using a linear exposure-response function, we developed Cox models with a cross-basis for the exposure-lag-response function to compute Akaike information criterion (AIC) for tumors occurring before and after menopause. We adjusted for race/ethnicity, education, household income, and census region.
Results: For PM2.5 and tumors occurring before menopause, AICS were lower for the B-spline (7578.7) and time-stratified lag-response functions (7582.9) than the linear function (7593.5). For tumors occurring after menopause, AICs for the time-stratified lag-response function (39773.9) were lower than the linear (39855.8) and B-spline functions (39794.0). For NO2 and tumors occurring before menopause, AICs were lowest for the linear lag-response function (6918.1) compared to the B-spline (6821.6) and time-stratified functions (6922.0). For tumors occurring after menopause, AIC for the linear lag-response function (38630.7) was comparable to the B-spline function (38632.3).
Conclusions. Future studies with historic air pollutant exposure estimates exploring the impact of PM2.5 exposure on cancer should account for nonlinearity of the lag-response function.