Simulating screening bias in case-control risk-factor studies

It is well known that bias such as lead-time and length distort studies of screening efficacy whether survival or incidence is of interest. A third bias, called overdiagnosis bias, can be defined as an extreme form of these biases. Only recently has it been recognized that screening bias can arise in cancer risk-factor studies where screening rates vary by stratum. Forms of screening bias conventional analyses cannot address are described, and a model to quantify them was developed and used to evaluate them in two case-control studies nested within the Prostate, Lung, Colorectal, and Ovarian (PLCO) randomized trial. Surveillance, Epidemiology, and End Results registry data were used to estimate age-specific preclinical incidence of lung cancer, and to assess the sensitivity of the bias to the preclinical duration distribution, the parameters of a log-normal were varied over a plausible range. The model used literature-based chest x-ray sensitivity estimates, weights to represent the study age structure, and screening patterns (proportions and frequencies) developed from PLCO study data. Within simulations, a joint null hypothesis of no effect of smoking or screening on the preclinical incidence of lung cancer was assumed; bias was indicated by deviation of the risk ratio from the null. As the difference in screening pattern during the PLCO study increased between smokers and nonsmokers, the bias increased from 1% to 64%. The simulations illustrate the range of potential bias affecting the smoking-lung cancer risk estimates in modern case-control studies and demonstrate the importance of quantitatively evaluating it in susceptible studies.