Contemporary prostate cancer screening is based on testing total PSA in blood. Total PSA comprises the sum of the unbound (“free”) PSA and PSA bound to inhibitors (“complexed PSA”) (5
), with each form of PSA having different characteristics (6
). There is an extensive literature on screening for prostate cancer. Studies include those of case-control design, comparing PSA screening history in men who died of prostate cancer with controls (7
); cohort studies, comparing outcomes of patients presenting for treatment with and without PSA screening (8
); ecologic studies, comparing prostate mortality in geographic areas with high versus low prevalence of screening (9
); simulation studies, simulating the effects of screening based on computer models of prostate cancer development (10
); randomized trials (11
); and meta-analyses of randomized trials (13
Much of this literature is unconvincing. Case-control studies of prostate cancer screening are prey to problems in ascertaining the exposure. It is widely accepted that prostate cancer is a slow-growing disease, and that the effects of screening will take a considerable period of time to become apparent. What matters, therefore, is screening behavior 10 or 15 years previously, which can be difficult to ascertain.
Cohort studies comparing outcomes of patients presenting clinically with those presenting after screening have a denominator problem. For example, imagine that, for every 1,000 in the population, 80 had an indolent cancer detectable by screening and 20 had an invariably fatal cancer that would eventually lead to clinical detection. In this scenario, screening would obviously have no possible benefit, but a cohort study would report survival rates of 80% for screen-detected cancers (for every 100 cancers detected, 80 are indolent and 20 are fatal) compared to 0% survival for clinically detected cancers, which are only detected at a late stage.
Ecologic studies are well known to be subject to important confounding. Two geographical areas that differ in rates of PSA screening will also likely differ in terms of treatment and may differ in genetic risk. Moreover, ecologic studies rely on administrative data, and there can be important variations in death certificate practices in different countries.
Simulation studies are based on assumptions about cancer behavior that are often open to interpretation and debate. As one example, computer models require a distribution for lead time; this was generally assumed to be exponential, until an empirical study recently demonstrated that prostate cancer lead time has close to a Gaussian distribution (15
Randomized trials are widely thought to provide the best evidence as to the benefits and harms of interventions. Four randomized trials of prostate cancer screening have been reported to date, and we believe that two of these should not be considered in detail. The “Quebec” trial involved close to 50,000 men randomly selected to be invited to PSA screening or to no-screening control, but it did not include several key methodologic safeguards. In particular, prostate cancer mortality data were obtained from administrative data rather than being adjudicated by a committee blinded to treatment assignment (16
). The “Norrköping” trial is relatively small—fewer than 1,200 men in the screening arm—and used only digital rectal examination (DRE) as the screening tool during the initial two rounds (six years) of screening, with PSA testing added halfway through the trial (17
For these reasons, we focus on two large, methodologically rigorous trials: the Prostate, Lung, Colorectal and Ovarian Cancer Screening Trial (PLCO trial), which took place at 10 separate sites in the United States (11
), and the European Randomized Trial of Screening for Prostate Cancer (ERSPC trial), which included participants from the Netherlands, Sweden, Finland, Belgium, Spain, Italy, Switzerland, and France (12
). Notably, the screening study in Göteborg, Sweden was originally an independent trial that later joined ERSPC.