Mayer SE, Edwards JK, Hallas J, Lund JL, Sturmer LT. Modeling cancer risk following drug initiation under various screening regimens: a discrete event simulation. Poster presented at the 2021 Virtual Society of Epidemiologic Research Annual Meeting; June 22, 2021.


Screening’s ability to advance cancer detection may result in the appearance of non-causal relationships between medication initiation and cancer risk when screening and treatment have shared predictors. We used discrete event simulation to model the lifetime risk of breast cancer initiation, growth and detection, and competing mortality in 100 cohorts of 100,000 women between the ages of 20-84. “High healthcare access” was positively associated with cancer screening between the ages of 50-75 and with treatment. Treatment initiation was modeled as a binary variable, with age at initiation distributed normally; true disease risk was independent of both treatment and healthcare access, leading to a true null treatment effect. Two-year screening coverage in the population varied between 0 and 100%. We used Aalen-Johansen risk estimation to calculate cumulative incidence (incidence) of breast cancer initiation and detection under each screening regimen and within treatment arms in the presence of competing mortality. The 10-year incidence of breast cancer increased from 2.48% (95% CI: 2.47, 2.49) to 2.83% (95% CI: 2.82, 2.84) with increasing screening coverage in the population; all observed risks of cancer incidence underestimated the incidence of cancer initiation (3.60%, 95% CI: 3.58, 3.61). In settings in which screening-eligible individuals were either always or never screened, minimal bias was observed in the estimated RD and RR throughout the 15-year follow-up period. However, when a proportion of the population received screening, screening occurred differentially across treatment groups, leading to bias in both the RD and RR across the 15 years of follow-up. These findings illustrate the sensitivity of both marginal and stratified risk estimates and their contrasts to the level of screening coverage in a population. Furthermore, they highlight the need for methods to account for differential outcome assessment in real-world analyses of longitudinal data.

Share on: