To estimate preliminary PSA growth parameters, we analyzed PCPT data from 1022 cancer cases (414 interim cases and 608 diagnosed by end-of-study biopsy) and 7058 subjects who did not have prostate cancer by the end of the follow-up period. When the resulting PSA growth estimates were combined with preliminary disease progression estimates, we found that model-projected age-adjusted test-positive and cancer detection rates (25% and 31%) did not validate well with corresponding values reported in the initial screening round of the PLCO (8% and 44%) (Andriole and others, 2005
). Investigating this lack-of-fit, we found that several combinations of μ0
led to equally good incidence projections but that some more closely approximated reported PLCO results. Our final selected values of these parameters yield model-projected test-positive and cancer detection rates of 11% and 57%. Note that we expect projected values to be at least modestly higher than corresponding trial values because we are simulating de novo
screening, while many trial participants had in fact undergone previous screens (Andriole and others, 2005
presents preliminary and final PSA growth estimates. The estimated mean intercept (i.e. PSA at age 35) across subjects is 0.2 ng/mL, smaller than prior median estimates of 0.6 for men aged 40–49 (Anderson and others, 1995
), (Lein and others, 1998
), and the estimated mean preonset slope corresponds to an average increase in PSA of approximately 2% per year, which is consistent with several studies (Oesterling and others, 1993
), (Whittemore and others, 1995
), (Ellis and others, 2001
), (Inoue and others, 2004
). After disease onset, mean PSA growth accelerates to an annual percent change of approximately 14%. This is considerably lower than the PSA growth rates among cases in previously published stored-serum studies (Whittemore and others, 1995
), (Inoue and others, 2004
), but the cohorts of cases in these studies included many pre-PSA-era patients; in contrast, the PCPT case cohort was identified under a screening and end-of-study biopsy program. Indeed, the majority of the PCPT cases were diagnosed in the absence of elevated PSA levels. The between-subject variability is considerably lower than the within-subject variability, which is similar for cancer cases and noncancer cases.
Table 1. Log PSA growth parameter estimates (upper panel) obtained via separate linear fits to PCPT cancer and noncancer cases. Preliminary preonset intercept and slope means were fine-tuned based on test-positive and cancer detection rates in the initial round (more ...)
We assessed incidence projections of model variants based on how well they captured broad patterns in observed incidence. We paid particularly close attention to the peak in local-regional stage incidence and the decline in distant stage incidence. This assessment indicated that a clinical detection hazard that depends on stage outperforms the baseline specification, a linear onset hazard outperforms an exponential onset hazard, and a diagnostic testing interval of 2 years is reasonably consistent with observed incidence. We integrated these findings into a final model and reestimated the disease progression parameters across 20 random seeds. Parameter estimates are reported in the lower panel of , where we see that the greatest uncertainty is associated with the multiplier for the clinical detection hazard for distant stage cancers (θc). This parameter is difficult to estimate precisely since a small increase when θc is already large advances clinical diagnosis of a distant stage cancer a matter of days and so corresponds to only a small impact on the likelihood.
illustrates observed and model-projected age-adjusted incidence curves averaged across random seeds by disease stage in the presence of PSA screening. Confidence interval (CI) estimates are shown in each year based on uncertainty due to the random seed. The figure also shows model-projected age-adjusted incidence had there been no PSA screening, that is, the model's best estimate of the secular trend in stage-specific disease incidence in the absence of PSA screening. Projected incidence in the absence of PSA is more or less constant at about the level observed in 1985 for both local-regional and distant stage disease. In the presence of PSA screening, model projections match the general shape of observed stage-specific incidence trends fairly closely, though the model overprojects in the late 1970s for both stages and underprojects (overprojects) in the late 1990s for local-regional (distant) stage. The difficulty with capturing the distant stage incidence decline observed since 1990 was observed by Etzioni and others (2008)
under a completely different model, leading them to suggest that alternative explanations (e.g. changes in public awareness of prostate cancer during the PSA era or use of PSA as a diagnostic test in symptomatic patients) likely contributed to producing a decline beyond that due solely to PSA screening.
Observed and mean model-projected age-adjusted local-regional (a) and distant (b) stage prostate cancer incidence in the presence and in the absence of PSA screening. Superimposed are 95% CIs reflecting uncertainty due to the random seed.
reports mean times between natural history and clinical or screen detection events based on the parameter estimates in . The measures are generally consistent with previously published studies. For example, the model indicates that the age-adjusted mean sojourn time is approximately 13.5 years, which is similar to the estimate in Etzioni and others (1998)
. Mean lead times are longer for younger men than for older men; this is a consequence of our definition of lead time, which applies to the subset of cases with clinical diagnosis in their lifetimes. Naturally, as men age and their remaining life expectancy declines, the range of plausible intervals until any event such as clinical diagnosis narrows. The mean lead times are slightly higher than some previously published studies (Gann and others, 1995
), (Telesca and others, 2008
) but are lower than others (Draisma and others, 2003
), (Tornblom and others, 2004
). Note, however, that estimates based on data from the European screening trial differ in important ways from the US population setting, and these cannot, strictly speaking, be compared (Draisma and others, 2009
Table 2. Mean years from screen detection to clinical detection by age at screen detection among men who would have been clinically diagnosed in their lifetimes (lead), mean years from onset to clinical detection by age at onset among men who would have been clinically (more ...)
presents comparisons of outcomes for the 4 candidate PSA screening policies. It can be clearly seen in Panel (a) that, irrespective of the screening age cap, lowering the PSA test-positive threshold from 4.0 to 2.5 ng/mL incurs a large number of additional overdiagnoses while moving only a handful of potentially fatal cases to a presumably more curable stage. Specifically, on average, lowering the test cutoff when screening men aged 50–84 generates 150 (95% CI due to the random seed [95–199]) additional overdiagnoses for each additional early detection and when screening men aged 50–74 generates 101 (95% CI:68–154) additional overdiagnoses for each additional early detection.
Fig. 2. Comparisons of model-projected benefits and harms across random seeds corresponding to PSA test-positive cutoffs 2.5 or 4.0 ng/mL for screening ages 50–74 or 50–84 and for years 1990–2000 in the SEER 9 population. In both panels, (more ...)
Complementing this tradeoff, Panel (b) presents overdiagnosis rates (as percentages of PSA detections) versus mean lead times (in years) projected for the 4 PSA screening policies. From this perspective, we see that, irrespective of the screening age cap, lowering the PSA test-positive threshold from 4.0 to 2.5 ng/mL increases both the overdiagnosis rate and the mean lead time. Specifically, on average, lowering the test cutoff when screening men aged 50–84 increases the overdiagnosis rate by 5.0% (95% CI [4.7–5.3]) while extending the mean lead time 0.75 (95% CI [0.72–0.77]) years and when screening men aged 50–74 increases the overdiagnosis rate by 6.8% (95% CI [6.4–7.4]) while extending the mean lead time 0.93 (95% CI [0.90–0.96]) years. The value of this tradeoff will depend on the benefit associated with early detection of disease within the same broad SEER stage. To quantify this would require modeling of disease-specific survival under population treatment patterns, which is beyond the scope of the present study.
Finally, to demonstrate model rankings, we consider total overdiagnoses per total early detections. The model ranks the 4 policies as follows: screening ages 50–74 with cutoff 4.0 ng/mL (4.3, 95% CI [3.9–4.7]), screening ages 50–84 with cutoff 4.0 ng/mL (5.0, 95% CI [4.5–5.4]), screening ages 50–74 with cutoff 2.5 ng/mL (5.8, 95% CI [5.4–6.3]), and screening ages 50–84 with cutoff 2.5 ng/mL (6.8, 95% CI [6.3–7.4]). Interestingly, while the interpretation is less straightforward, considering overdiagnosis percentages per year of expected lead time yields the same ranking. While these metrics represent a small number of the great many possible systems for developing policy rankings, it is worth noting that these rankings for these policies are consistent with updated recommendations by the US Preventive Service Task Force (2008)