Prostate-Specific Antigen, Risk Factors, and Prostate Cancer: Confounders Nestled in an Enigma

Abstract

Prostate cancer is a curious disease. For almost three decades, a man's lifetime risk of death from the disease in the United States has been approximately 3% (1). With the development of prostate-specific antigen (PSA) testing, performed in approximately 50% of American men, the lifetime risk of a clinical diagnosis has almost doubled from approximately 9% to 16% (1–3). At autopsy, the risk of prostate cancer is approximately 40% for men aged 40–49 years, and this risk increases to approximately 70% for men aged 60–69 years (4). (We acknowledge that prostate cancer risk determined from autopsy results is a lower bound of risk at these ages because of the censoring of men who have been treated for prostate cancer.) Although not all tumors detected at autopsy (eg, small tumors) would have become clinically apparent, the results of the Prostate Cancer Prevention Trial (PCPT) showed that many prostate cancers can be detected by biopsy. The PCPT also showed that the risk of cancer if a biopsy were performed was 26.9% for a man with a PSA level of 3–4 ng/mL and 10.1% for a man with a PSA level of 0.6–1.0 ng/mL (5). The commonly stated adage that “if you live long enough, everyone will have prostate cancer but many fewer will die from it” is not too far from what these data portray. That a family history of prostate cancer increases a man's risk of a diagnosis of this disease is now entrenched in the medical literature. It is instructive, however, to learn when this observation was made. Perhaps a surrogate for our understanding of this relationship is the volume of published work on the association between family history and prostate neoplasms. On July 1, 2010, there were 788 such publications in PubMed. In the 10 years between 1977 and 1986, there were seven publications, for a rate of 0.7 publications per year. Between 1987 and 1996, there were 79 publications, for a rate of 7.9 publications per year. Between 1997 and 2006, there were 45.8 publications per year; and in approximately the past 4 years, there were 65 publications per year. The uptick occurred sometime after 1986. But, prostate cancer represented a major recognized disease for decades before then. Why did it take us until the mid-1980s to fully recognize or explore adequately the familial association? Certainly, the explosion in total numbers of prostate cancers has drawn our attention to the disease and provided the data to explore this association. The article by Bratt et al. (6) provides another explanation: in the PSA era, a man with a brother or father with prostate cancer was more likely to be screened for prostate cancer by PSA testing. Against the background of an enormous number of undiagnosed (but diagnosable) prostate cancers, this simple decision could lead to a prostate biopsy and exaggerate the link between family history and prostate cancer. The authors of that study (6) and the government of Sweden should be congratulated on rigorous gathering and assessment of data that speak directly to detection bias in studies of the hereditary nature of prostate cancer. Even more robustly designed studies might not be immune from this bias. For instance, despite a required end-of-study biopsy for all participants in the PCPT irrespective of risk factors, biopsy compliance was only approximately 60%; subjects who complied were more likely to have a positive family history of prostate cancer in a first-degree relative than those who did not (ie, there were 16.6% with a positive family history in the compliant group vs 13.3% in the noncompliant group, P < .001) (7). Other factors influencing compliance were older age, white race, and no history of a prostate biopsy (all P < .001). When the analysis was restricted to the subgroup who underwent biopsy (the definition of diagnostic activity in our discussion of the PCPT), positive family history of prostate cancer remained a statistically significant independent predictor of prostate cancer (odds ratio = 1.31, 95% confidence interval = 1.11 to 1.55, P = .002), with the other well-established clinical predictors of prostate cancer being PSA level, an abnormal digital rectal examination at time of diagnostic activity (ie, biopsy), and history of a negative prostatic biopsy. This analysis of prostate cancer risk is exempt from confounding from diagnostic activity because it is confined to the subgroup that has experienced diagnostic activity. Mathematically speaking, it has evaluated the probability of prostate cancer for a participant who has undergone diagnostic activity: P(PCA|DA, PSA, DRE, FamHist, PriorBiop), where PCA is prostate cancer, DA is diagnostic activity (in this case, a biopsy), DRE is digital rectal examination, FamHist is family history, and PriorBiop is previous biopsy status. To account for the fact that PCPT participants who complied with biopsy were not a random sample, but rather belonged to a group with higher rates of family history, more white participants, and older participants but without a biopsy history, a verification bias adjustment analysis was performed. It showed that qualitatively similar relationships could be obtained between risk factors and prostate cancer in the subgroup with biopsy results and could be inferred for the entire PCPT cohort (7,8). Bratt et al. make the claim that “The effects of increased diagnostic activity and hereditary genetic factors on the risk of cancer in a man with a family history of prostate cancer cannot be separated.” We do not agree. Certainly, in their study, the effects are confounded if the composite endpoint of diagnostic activity and prostate cancer is assessed. Such a population-based study of prostate cancer estimates the relationship between risk factors and cancers found via diagnostic activity [ie, P(DA and PCA|X), where X denotes all risk factors of interest that might affect both diagnostic activity and prostate cancer, including family history of disease and PSA]. The PCPT by design provided estimates of the relationship between risk factors and diagnostic activity [P(DA|X)] and between the risk of prostate cancer that is based on risk factors and whether there was diagnostic activity [P(PCA|X, DA)]. It should be noted, with a few caveats including a different set of risk factors and a chemoprevention agent in the mix, that the sample size of the PCPT (N = 18 882) approaches that in the Swedish study (N = 22 551). By Bayes rule, the connection between estimates from the Swedish and PCPT studies lies on either side of the following equality: P(PCA and DA|X) = P(PCA|DA, X)P(DA|X). In other words, the risk of undergoing diagnostic activity and being diagnosed with prostate cancer equals the risk of undergoing diagnostic activity multiplied by the risk of being diagnosed with prostate cancer having undergone diagnostic activity. Thus, if the Swedish study could be augmented with another study estimating diagnostic activity risk in Sweden [ie, P(DA|X)], then an unconfounded evaluation of prostate cancer risk associated with family history could be assembled by the ratio of the probability of prostate cancer identified by diagnostic activity given some known risk factors and the probability of diagnostic activity given these risk factors [ie, P(PCA and DA|X)/P(DA|X)]. Identification of epidemiological confounders is typically the greater challenge in dissecting risks. However, appropriately estimating and adjusting for confounders often requires ancillary studies, such as that for estimating diagnostic activity risk, and will raise new methodological challenges for combining information from potentially differing cohorts. Regardless of the conclusion, it does appear that assessments of variables that affect prostate cancer risk run the risk of several known and, more problematic, unknown biases in almost all studies, even those as rigorously conducted as the PCPT. How can we remedy this problem? Perhaps the best tactic would be to change our approach from seeking risk factors for prostate cancer (a disease that is ubiquitous, with many patients probably being better off if it were not detected by screening) to an assessment of factors related to biologically consequential prostate cancer (ie, metastatic disease or prostate cancer–specific death) as in the study by Bratt et al. We must remain vigilant regarding these confounds as we continue our investigations of this important disease.