Statistical Issues in Evaluating Randomized Screening Trials

Screening examinations are routinely administered, due to the common belief that screening saves lives, reduces the need for agressive treatment, and improves quality of life. As with any medical protocol, benefits should be evaluated by a proper experiment. Randomized trials offer the most effective method for evaluating risks and benefits of screening (one arm offered screening, one arm follows usual medical care), yet nonetheless they pose special challenges for analysis: they are not blind, contamination can occur in both arms, and they are subject to biases that do not appear in ordinary clinical trials (e.g. lead time, length biased sampling, overdiagnosis). We describe these challenges and statistical methods to address them. The methods are illustrated on data from two highly cited screening trials: Health Insurance Plan (HIP) of New York, and the National Cancer Institute's PLCO trial for early detection of prostate, colorectal, lung, and ovarian cancers. The results suggest that the benefits of screening, measured by either reduction in mortality or extended lifetime, apply to some tests (e.g. flexible sigmoidoscopy for colon cancer) but not to others in common use (e.g., PSA for prostate cancer, CA-125 for ovarian cancer), due in part to false positives and overdiagnosis.