Evaluation of Breast Cancer Screening

Evaluation of Breast Cancer Screening

The efficacy of a screening test is best evaluated with a population-based, randomized clinical trial with a mortality end point. Randomized trials eliminate the potential biases in observational studies that may influence comparative survival among screen-detected and nonscreen-detected cases, most notably lead-time bias, length-bias sampling, and selection bias.

As Miller notes, in most instances screen-detected cases have better survival than cases diagnosed after the onset of clinical symptoms. It is therefore important to distinguish the actual improvements in survival from apparent improvements. As noted previously, the goal of screening is to gain lead time. If treatment before the onset of symptoms offers greater benefits, then improvements in survival should be associated with lead time gained, and mortality should be lower in cases diagnosed by screening.

On the other hand, if lead time only advances the time of diagnosis and life is not extended because death occurs at the same point in the natural history of the disease among screen-detected and nonscreen-detected cases, then there is only the appearance of a greater survival duration. Lead-time bias occurs when increased survival is a function only of the time gained before the point at which diagnosis would have occurred in the absence of screening.

Length-bias sampling refers to the tendency for screening to detect more slow-growing, less aggressive disease and to be less successful at detecting more aggressive, faster-growing disease. If screening selectively identifies cases at a lower risk of death, length-bias sampling may influence end results. Selection bias refers to the tendency for individuals who are healthier, or more health conscious, and with a different probability of developing and/or dying from cancer, to participate in screening.

In a population-based randomized trial, these biases are eliminated, because randomization should result in equal distributions of these confounding factors in the groups invited and not invited to several rounds of screening. Lead-time bias is eliminated, because disease-specific mortality in the group invited to screening is compared with mortality in the group not invited to screening at some future date after the starting point of the study, which is the same for both groups.

Length-bias sampling and selection bias are eliminated, because randomization should ensure the same distribution of individuals with underlying probabilities of developing cancer, with faster- and slower-growing tumors, and with similar overall health status into each group in the study.

Because the analysis is based on mortality differences in the two groups and not on the basis of the subgroups that were and were not screened, most known and unknown biases can be minimized.

Robert A. Smith and Carl J. D'Orsi