Is there Unmeasured Indication Bias in Radiation-Related Cancer Risk Estimates from Studies of Computed Tomography?

Abstract

Recently reported studies have associated radiation exposure from computed tomography (CT) scanning with small excess cancer risks. However, since existing medical records were used in these studies, they could not control for reasons for the CT scans and therefore, the results may have been confounded by indication. Here we conducted a study to estimate potential indication bias that could affect hazard ratios for colorectal, lung and female breast cancers by reasons for a CT scan. This involved a retrospective cohort study of electronic records from all patients aged 18-89 years without previous cancer diagnoses, who received at least one CT scan at Columbia University Medical Center in the period of 1994-2014. This investigation is not a study of CT-related cancer risks with adjustment for reasons, but an evaluation of the potential for confounding by indication in such studies. Among 75,968 patients, 212,487 CT scans were analyzed during a mean follow-up of 7.6 years. For colorectal and female breast cancers, no hazard ratio bias estimates for any of the CT reasons reached statistical significance. For lung cancer, significant biases occurred only in patients with unknown CT reasons and in patients with CTs for "abnormal findings" and in those with CTs for cancer- or nodule-related reasons. This retrospective cohort study among adults with ≥1 CT scan evaluates, for the first time, CT reason-specific indication biases of potential CT-related cancer risks. Overall, our data suggest that, in studies of adults who underwent CT scans, indication bias is likely to be of negligible importance for colorectal cancer and female breast cancer risk estimation; for lung cancer, indication bias is possible but would likely be associated with only a small modulation of the risk estimate. Radiat. Res.

FIG. 1

Flow chart of the study population.

FIG. 2

Computed tomography (CT) reason-specific hazard ratio bias estimates (symbols) and their 95% confidence intervals (error bars) for colorectal, lung and breast cancer. The different reasons for the CT scans are listed sequentially on the x-axis in the same order as in Tables 4–6 for colorectal cancer, lung cancer and breast cancer, respectively.