The European Randomized study of Screening for Prostate Cancer (ERSPC) has recently reported a 20% reduction in death from prostate cancer in a population-based prostate cancer screening (core age group: 55–69 years of age). The effect of screening may be diluted by noncompliance in the screening arm and contamination by PSA testing in the control arm. The purpose is to analyze the effect of prostate cancer screening on the incidence of metastatic prostate cancer, both with and without adjustment for noncompliance and contamination. We analyzed the occurrence of metastases in 42,376 men aged 55–75 years who were randomized in the Rotterdam section of the ERSPC between 1993 and 1999. Contamination adjustment was based on follow-up findings and questionnaire data from all men in the control group who developed prostate cancer and from a random sample of 291 men without cancer who had a PSA test. Prostate cancer screening significantly reduced the occurrence of metastatic prostate cancer in the intention-to-screen analysis [RR 0.75, 95% CI 0.59–0.95, p = 0.02] and more so in adjusted analyses; contamination adjusted RR 0.73, 95% CI 0.56–0.96; noncompliance adjusted RR 0.72, 95% CI 0.55–0.95 and fully adjusted analysis RR 0.68, 95% CI 0.49–0.94, p = 0.02. In the population of ERSPC Rotterdam (N = 42,376 men), screening reduces the risk to be diagnosed with metastatic prostate cancer considerably on population level, an effect which is even more pronounced in men who are in fact screened.
The European Randomized study of Screening for Prostate Cancer (ERSPC) evaluates the effect of prostate cancer screening on prostate cancer mortality and morbidity in 8 European Countries.1 The study has been ongoing since 1993 and has recently reported a 20% overall reduction in death from prostate cancer in subjects between the age of 55 and 69 years and a 15% reduction in men between the age of 55 and 74 years (RR: 0.85 [95% CI: 0.73–1.00]).2 Although the conventional intention-to-screen analysis estimates the effect of prostate cancer screening offered to the study population, the effect size estimated from an intention-to-screen analysis for an individual choosing to be screened can be diluted by contamination in the control group and by noncompliance (those invited to screening choosing not to avail themselves for this activity).3, 4
Cuzick et al.5 developed a statistical model, adjusting for noncompliance and contamination, which estimates the effect size among those participants who are managed according to their random assignment. This method has been previously applied to breast cancer screening trials.6 Such an adjustment is difficult because noncompliers and contaminators may differ in their underlying risk in ways which are not related to screening. The Cuzick model fully adjusts for this possibility whilst preserving the benefits of randomization. Use of this model requires a proper definition of contamination, in our case the actual reason for performing a PSA test in the control group.
Moreover, the impact of noncompliance and contamination on the results depends on the time of occurrence. For example, a participant randomized to screening who never attended to a screening visit is likely to have a larger diluting effect than a participant who dropped out after 2 rounds of screening. This may also be the case for contamination, where most of the PSA tests in the control arm were performed in the first year after randomization.7
We aimed to compare the effect of prostate cancer screening on the incidence of metastatic prostate cancer, both with and without adjustment for noncompliance and contamination.
Metastatic disease is specific, clinically relevant and is more powerful in detecting important effects, because events occur earlier than prostate cancer death.8 Additionally, it reflects morbidity from prostate cancer, because some men with metastatic disease will nevertheless die from some other cause, but most men with metastatic disease will suffer pain and experience the psychological impact of advanced disease.
Material and Methods
In the period of December 1993–December 1999, all men aged 55–74 years living in the Rotterdam area were invited to participate. After filling in a questionnaire and signing an informed consent, men were randomized to an intervention arm (N = 21,210) or to a control arm (N = 21,166). Prevalent prostate cancer was assessed by means of the questionnaire, and they were excluded from randomization. After linkage with the cancer registry, an additional small number of men were found to have prostate cancer and were excluded after randomization.
Reliable information on noncompliance and the use of screening in the control arm (contamination) is available for the Rotterdam part of the ERSPC.9 Here, 42,376 men were randomized using population registries from December 1993 through December 1999. We excluded 39 men in the screen arm and 29 in the control arm because they were found to have had prostate cancer diagnosed or had died before randomization. Follow-up was complete till January 1, 2007.
The primary outcome was metastatic prostate cancer. Prostate cancer mortality was used as an additional endpoint to validate the relevance of metastatic disease. Men with prostate cancer were identified after cross matching with the national cancer registries, and follow-up information is retrieved from their medical records. Metastases could be identified either at diagnosis or during follow-up and were defined according to the 1992 version of the TNM classification if described in the medical records, whereas the M classification is defined as distant metastases or as a PSA level ≥100 ng/ml when data on bone scans were missing. Prostate cancer mortality was based on the cause of death as reviewed by the ERSPC CODC committee.10
The effect of prostate cancer screening on metastatic prostate cancer was first determined by a proportional hazards analysis using all randomized patients (intention-to-screen analysis). For the adjustment for noncompliance and contamination, the efficacy analysis of Cuzick et al.5, 11 was applied using a Poisson model. We developed a specific model variant to adjust for the time and sampling of potential contaminators and noncompliers. Three analytic variants were performed including a binary analysis (outcome yes/no), a Poisson analysis using the time of noncompliance, contamination and metastases taken into account and a semiparametric Cox proportional hazards model assuming contamination and noncompliance occurred at randomization. Here, we focus on the Poisson model results, because the 3 variants gave very similar results.
We estimated the relative incidence rates of metastatic disease using the person-years at risk contributed to each subgroup (compliers, noncompliers, contaminators and noncontaminators). The person-years of follow-up for a noncomplier are counted from randomization or from 4 years after the time of the last screening visit in partial noncompliers. The person-years for the contaminator group were counted from the time they had their first PSA test until the end of their follow-up. No adjustment was made for individuals who failed to have further investigation, typically a biopsy following a positive screening PSA test. The model entails analyses based on the identification of the person-years and associated endpoints of noncompliers in the screening group. These person-years and endpoints (adjusted for the precise randomization ratio) are subtracted from the controls to take into account of the potential noncompliers randomized to control (Fig. 2). For the adjustment of contamination, we subtract the contaminators' person-years and associated endpoints from the screening arm. The impact of screening on metastatic prostate cancer is estimated in men who would accept their allocated procedure, whether it had been screening or control.
Definition of noncompliers and contaminators
A participant was classified as a noncomplier if he refused a PSA test after the screening invitation. Such a noncomplier was not invited for PSA testing in the following screenings rounds. The effects of applying other definitions of noncompliance were also studied. A sensitivity analysis revealed that the precise definition of noncompliance (e.g., noncompliance immediately after randomization or noncompliance in at least one of the screenings rounds) had minimal effect on the results.
We used different methods to estimate asymptomatic contamination in men with versus men without prostate cancer. Patients diagnosed with prostate cancer were traced and followed through medical records, which stated the reason for referring the patient to the urologist. A patient allocated to the control arm of the study was considered as a contaminator when his prostate cancer was detected following a PSA test clearly carried out for screening purpose and not for clinical needs. To determine how often asymptomatic PSA tests occurred in controls without prostate cancer, a random sample of 345 men was selected. The sample was taken from the control group who had a PSA test, which was determined after linkage with the general practitioners laboratory. The reason for performing the PSA tests was determined by sending questionnaires to the general practitioner of the 345 men. The general practitioners of 291 men returned a questionnaire about the reasons for performing the PSA test after inclusion in the study (response rate 84%). This additional information was essential to classify PSA tests as diagnostic or contaminating in men without prostate cancer. A PSA test was considered contaminating when the general practitioner or the participant himself wanted the test in the absence of specific symptoms possibly related to prostate conditions, or when the PSA test was part of a general blood check. A diagnostic PSA test was performed on clinical indication, i.e., urinary complaints, other relevant urological complaints or a suspicious rectal examination. The decision to biopsy was up to the patient and his doctor. The number of contaminators and years of follow-up postcontamination found in the random sample of controls was applied to both arms of the trial to estimate the amount of actual contaminators in the control arm and the potential contaminators in the intervention arm.
A Taylor series expansion5 was used for calculating the confidence intervals.
From December 1993 through January 2007, 2,983 prostate cancers were diagnosed in the cohort with a mean follow-up of 8.9 years (Table 1). Of these, 36 cases were metastatic at the time of diagnosis in the intervention arm and 90 cases in the control arm, a total of 244 developed into metastatic prostate cancer during follow-up (121 in the intervention arm and 159 in the control arm). Of them, there were 6 cases in the control arm with a PSA > 100 ng/ml without an additional bone scan and 9 cases in the screening arm. Death from prostate cancer occurred in 150 of the 280 metastatic cases, with 65 and 85 deaths in the intervention and control arm, respectively. Mortality not related to prostate cancer occurred in 26 of 159 cases in the control arm and in 24 of 121 in the intervention arm.
Table 1. Population characteristics of the ERSPC Rotterdam (N = 42,308)
Up to January 1, 2006, 5,097 (24%) participants of the control arm had tested their PSA after randomization, and 321 prostate cancers were diagnosed in this group. Based on the information on reason for referral in their medical records, 122 of these 321 prostate cancers could be classified as diagnosed following an asymptomatic PSA test, including 15 metastatic cancers (Fig. 1). Of 291 PSA tests among random control participants, 146 (50%) were classified as contaminating (Table 2). Extrapolation of the information from the smaller questionnaire study resulted in 2,520 estimated contaminators in the control arm contributing 9,730 person-years and an estimated 15 metastatic cancers (Fig. 2).
Table 2. Reason for the use of the PSA test based on 291 questionnaires from a random sample of 345 control participants
We found that 5,581 participants randomized to the intervention arm became noncompliant during the study period.
Effect estimates of screening
The intention-to-screen analysis resulted in a HR of 0.75 [95% CI: 0.59–0.95] for the development of metastatic prostate cancer in favor of screening. Based on the Poisson model, the noncompliance and contamination-adjusted estimate was a RR of 0.68 (95% CI: 0.49–0.94, p =0.02). Adjusting for noncompliance or contamination separately resulted in RRs of 0.72 (95% CI: 0.55–0.95, p = 0.02) and 0.73 (95% CI: 0.56–0.96, p = 0.02), respectively.
The ERSPC trial has shown that the PSA-based prostate cancer screening significantly lowers the disease-specific mortality by 20% after an average follow-up of 9 years.2 This benefit of screening is restricted to the core age group of subjects who were between the age of 55 and 69 years, in men 50–74 years the relative risk is 0.85 [95% CI: 0.73–1.00]. Subsequently, Roobol et al. performed an analysis to adjust for noncompliance and contamination within the ERSPC trial.12 The effect of prostate cancer screening on prostate cancer mortality in the intention-to-screen analysis gives a relative risk of 0.80 [95% CI: 0.68–0.96]. After adjustment for noncompliance and contamination, the relative risk turns out to be 0.71 [95% CI: 0.55–0.93].
In this analysis, based on the data of the Rotterdam section of the ERSPC trial, we found that PSA-based screening reduced the risk of metastatic prostate cancer. In an intention-to-screen analysis, relative risk estimate is 0.75, and to a slightly larger degree in a compliance-adjusted analysis, relative risk estimate is 0.68.
It is important to realize that the presence of metastases is not actively sought by standard bone scans, creating the possibility that metastases are missed.
A prior publication reporting on metastatic disease in the Swedish arm of the ERSPC was limited to events at the time of diagnosis. No adjustments were applied for noncompliance and contamination.13
The apparent synergistic effect when taking into account both noncompliance and contamination, which gave a relative risk of 0.68 compared to 0.72, respectively, 0.73 when adjusting for noncompliance and contamination separately, is a feature of the model we have applied. In the model, the correction for noncompliance and contamination takes out similar proportions in both arms, so the effect on the ratio is more than additive (or multiplicative). A simple example might help to illustrate the phenomenon. Suppose a situation that the denominators in both arms are exactly equal. Suppose also that there are 100 events in the intervention arm and 125 in the control arm (ratio 4/5). In the situation where there are 25 events among noncompliers and also 25 events among contaminators, the ratio for adjusting for either is 75:100 (i.e., 3/4), whereas the ratio for adjusting for both is 50:75 (i.e., 2/3).
Adjusting for noncompliance and contamination gives a better estimate of the effect of screening than the ITT analysis, those who chose to accept an invitation, but it is still not perfect, especially when contamination takes place some time after randomization or when screening is a multistage process and noncompliance can take place at intermediate stages.
In this analysis, contamination was defined as truly contaminating PSA testing carried out in asymptomatic men. Another way to look at contamination could consider opportunistic PSA testing, which is followed by a biopsy as effective contamination.7 This is not an appropriate definition for our purposes because the majority of screened men will not have a biopsy, and there is no reason why the PSA testing that led to a biopsy could not have been done to investigate symptoms.
In the sample of participants in whom the reason for the PSA test was evaluated, the response rate of 80% (291/345) is considered to be adequate to allow the differentiation between diagnostic and screening application of the PSA tests.
In determining contamination and noncompliance, we have focused only on the initial PSA test and not the attendance for any subsequent recommended biopsy. Compliance with biopsy may differ in contaminators compared to those invited for screening, but additional adjustment for this is beyond current methodology.
To test whether there is any effect of time in this compliance-adjusted analysis, the binary model with only dichotomous (yes/no) variables was extended to an analysis adjusting for the moment of contamination and noncompliance as well as the time to metastases in a survival analysis. Results of this analysis showed that there was minimal effect of these adjustment compared to the binary analysis.
The main purpose of a compliance-adjusted analysis is the ability to estimate the effect of screening if a man undergoes screening as intended, as the intention-to-screen analysis is diluted by data of the noncompliers and contaminators. Another approach is to exclude the noncompliers and contaminators in a per-protocol analysis. However, with that approach, the problem of self-selection arises, which undermines the original reason for randomization. In the compliance-adjusted analysis used in this article, the randomization is not violated. It may hence be considered a reasonable approach for estimating the effect of screening at the individual level.
To apply this method, we assume that the study group consists of 3 types of subjects; the insisters of screening, refusers and ambivalents. Furthermore, we assume that the probability of insisters and refusers is the same for each randomization group. When a control participant switches to screening directly after randomization, the screening protocol is identical to that in the intervention group, and when a participant randomized to screening immediately refuses screening, the effect of the lack of screening is identical to that in the control group. This assumption is fulfilled for the adjustment for noncompliance, as the potential noncompliers in the control arm and noncompliers in the intervention arm are both not screened. However, in the analysis adjusting for contamination, this assumption is probably not completely fulfilled, as asymptomatic PSA testing in the control arm will not have exactly the same effect as protocol-based screening in the intervention arm. The baseline risk of the contaminators is the same, but there may be some difference in the effect of screening. In this adjusted analyses, designed to estimate the individual effect of screening, the adjustment of noncompliance is essential and contributes the most to the difference with respect to the intention-to-screen effect.
There are some substantial differences between the incidence rates and risks of metastatic prostate cancer in the subgroups (Fig. 2). Note for instance that the proportion of noncompliers with metastases is less than for compliers, but that the incidence rate of metastases is higher. This is due to the fact that the risk is calculated using the binary data, while a substantial proportion of the men became noncompliant after the first screening round (30%). In the risk calculations, those men are considered noncompliant, although they are screened in the first round. After taking the person-years into account, this effect has disappeared, as those screened person-years are now added to the compliers group. Similarly, the contaminating controls had a much higher incidence rate, but a lower proportion had metastases. This difference might be due to unrecorded screening tests, as we used only data from the general practitioners laboratory. It is of interest that adjustment for contamination only gave a relative risk of 0.73 compared to 0.75 for the intention-to-screen analysis.
We conclude that after 3 rounds of 4-yearly screening by PSA testing, the risk of metastatic prostate cancer is reduced by 32% based on the contamination and compliance-adjusted analysis. For men who are considering screening for prostate cancer, the effect estimate for those who are actually screened is more clinically relevant.