Metastatic Prostate Cancer Incidence and Prostate-specific Antigen Testing: New Insights from the European Randomized Study of Screening for Prostate Cancer

The ERSPC is a multicentre, randomised screening trial comparing an intervention arm of men to whom regular PSA screening is offered with those in a control arm to whom such screening is not offered. The trial started in Belgium, The Netherlands, Finland, Sweden, Spain, Italy, Switzerland, and France during the period 1993–1998 [11] . The study methodology has been described in detail previously [12] . Data were obtained by linkage to local cancer registries for PCa incidence and national registries for overall mortality. PCa deaths were ascertained by local, independent, cause-of-death committees [13] .

Data on trial arm, centre, T and M stage [14] , Gleason score, serum PSA at diagnosis, age at randomisation, follow-up time (until 31 December 2010), and vital status were extracted from the trial database.

We selected men aged 55–69 yr (the core age group), excluding those who died, who were diagnosed with PCa, or who emigrated between randomisation and screening invitation (n = 145). The median follow-up for the two French centres, Herault and Tarn, was 6–7 yr and thus too short for their data to be included in this analysis.

The risk classification adopted in the last ERSPC analysis [1] was used. Four risk categories at diagnosis were defined using information available at diagnosis: low risk (category 1) indicated clinical stage T1–2 and Gleason score ≤6; intermediate risk (category 2) indicated T1–2 and Gleason score 7 or T3 and Gleason score ≤7; high risk (category 3) indicated T1–3 and Gleason score 8–10 or T4 and any Gleason score; and metastatic disease (category 4) indicated M1 or PSA >100 ng/ml. Data on disease progression and metastasis status during the entire period of postdiagnosis follow-up have already been published for those centres for which information was available [10] . In the present analysis, only information on risk category at diagnosis, available for all centres, was considered.

The accuracy of risk category definition was evaluated by a sensitivity analysis, restricting the evaluation of metastasis to M1 cases only.

We imputed missing data for the variables used in defining the risk categories at diagnosis, using multiple imputations by the technique of chained equations (MICE) [15] . Briefly, MICE uses the distribution of observed data to identify a set of plausible values for the missing data by incorporating a random component to reflect the uncertainty of the estimate. Several data sets are generated, and the same analysis is conducted on each data set with the aim of obtaining a set of parameter estimates that are combined in a single overall estimate (computed as the average of the estimates in each simulated data set) with relative standard errors (computed including the between- and within-imputation variance) and confidence intervals (CIs) [16] .

For the incomplete variables (T and M stage, Gleason score, and PSA level), we derived imputation models that included the complete variables: arm, centre, vital status, follow-up time, age at randomisation, and interactions between each variable and trial arm. The interactions between each variable and trial arm were included to incorporate potential differences among centres in the mechanism causing the missing data. Ten final data sets were generated after the algorithm converged. In the literature on multiple imputation, three to five imputed data sets are considered adequate [17] . The estimates were combined using Rubin's rules [18] . We conducted a sensitivity analysis by imputing directly the risk categories at diagnosis. We used version 12 of the Stata software (StataCorp, College Station, TX, USA).

PCa incidence rate ratios (IRRs) by centre and risk categories at diagnosis were calculated for the screening versus control arms using Poisson regression analysis. Person-years were calculated as time from randomisation to PCa diagnosis, death, or end of follow-up, whichever came first. We adjusted all analyses for the randomisation ratio 1:1.5 for the screening versus control arms in Finland.

Primary analyses were carried out on an intention-to-treat (intention-to-screen) basis (ie, men were included in the analysis in the arm to which they were randomly assigned, regardless of compliance). Moreover, to provide a sensitivity analysis, we split the cumulative occurrence within the screening arm according to screening status, defining subjects who underwent at least the first screening after first invitation, asattendees.

Complete case analysis (before missing data imputation) and all-case analysis (after data imputation) are presented.

In the core age group, 162 338 men were randomised. The median age at randomisation was 60.2 yr.

With data truncated at 13.0 yr of follow-up, 7408 and 6107 PCa cases were diagnosed in the screening and control arms, respectively. The distribution of missing values was similar in the two arms (8% in the screening and 10% in the control arm); however, for some centres (Belgium, Italy, and Spain), data were missing for between 38% and 77% of cases, and there was an imbalance between the arms of at least 10 percentage points ( Table 1 ).

Imputation did not materially affect the results ( Table 2 ); therefore, the results based on imputed data are shown henceforth.

The IRRs (screening vs control arm) for the low- and intermediate-risk categories at diagnosis were significantly elevated (IRR: 2.14 [95% CI, 2.03–2.25]; and 1.24 [95% CI, 1·16–1·34], respectively); the high-risk category had unity (IRR: 1.00; 95% CI, 0.89–1.13). The IRR was significantly reduced by 40% for the metastatic disease category (IRR: 0.60; 95% CI, 0.52–0.70).

The reduction was even greater when restricting the evaluation to M1 cases only (IRR: 0.50; 95% CI, 0.43–0.57).

The absolute excess in the low-risk category was larger than the deficit of advanced disease: The incidence rate difference (IRD) per 1000 randomised men for the low-risk category was 34.26 (95% CI, 31.99–36.52). The IRD for the intermediate-, high-, and metastatic-risk categories at diagnosis were 4.26 (95% CI, 2.74–5.79), 0.47 (95% CI, −1.09–0.74), and −3.14 (95% CI, −3.93 to −2.35), respectively.

The IRRs (screening vs control arm) for the low- and intermediate-risk categories dropped strongly and consistently with follow-up. The high IRRs in the first 3 yr were associated with the prevalence round and overdiagnosis. There was a sharp decrease of IRR in the high-risk category within the first 6 yr, and a further reduction thereafter as the risk converged with the control arm. A slight decrease of IRR was evident until the ninth year for the metastatic disease category, but afterwards it remained constant ( Table 3 ).

A sharper contrast was seen between the arms for the entire 13-yr period when only the attendees were used to calculate the IRRs: 2.45 (95% CI, 2.32–2.58) for low-risk, 1.32 (95% CI, 1.23–1.43) for intermediate-risk, 1.00 (95% CI, 0.88–1.14) for high-risk, and 0.45 (95% CI, 0.38–0.53) for metastatic disease at diagnosis.

The reduction in cumulative (0–13 yr) IRRs (screening vs control arm) for metastatic disease at diagnosis (IRR: 0.60) was clear and evident within the first 6 yr of follow-up, whereas the reduction in mortality was smaller (IRR: 0.79) and emerged roughly 3 yr later ( Fig. 1 ).

Fig. 1 Cumulative incidence rate ratios for risk category 4 (after data imputation) and cumulative prostate cancer mortality rate ratio by time since randomisation (95% confidence intervals given in parentheses). IRD = incidence rate difference per 1000 randomised men; MRD = mortality rate difference per 1000 randomised men.

In the present study, we examined the incidence rates of PCa by risk category at diagnosis between the two arms of the ERSPC trial.

The net effect in the screening arm was a marked increase of the cumulative incidence of low-risk PCa, along with a significant reduction of cumulative incidence of metastatic disease at diagnosis. A slight increase in the intermediate-risk category and no difference in the high-risk category (with an early excess disappearing by year 10) were observed. When interpreting these results, we must keep in mind that early detection resulting in a stage migration is expected to increase early disease, with corresponding decreases in advanced disease.

In this study, the reduction found for metastatic disease at diagnosis was not reflected as an excess in the high-risk category, but increased risk was found only in the intermediate- and, above all, the low-risk categories. The absolute excess in the low-risk category was substantially larger than the deficit of advanced disease: The difference reflects the excess of incidence in the screening arm, which was reported in the aforementioned update to the ERSPC trial along with the conclusion that one PCa death is averted for 27 excess cases detected in 13 yr of follow-up [1] .

To avoid potential biases due to the imbalance of the proportions of missing data in the two arms and among centres, we used multiple imputation. As the results did not change considerably with imputation, results with imputed data are shown because they are more comprehensive and precise.

The IRRs for the screening versus control arms decreased with follow-up time both overall and for each risk category at diagnosis, though not in a similar fashion. Cumulative incidence for low-risk PCa was three times higher in the screening arm than in the control arm during the first 3 yr of follow-up, and then fell off steadily to an incidence twice as high in the screening arm in the overall period of 13 years. The IRRs for intermediate- and high-risk categories also decreased over time. In contrast, a decrease in the cumulative incidence for the metastatic disease category was not evident in the first 3 yr of follow-up (IRR: 0.94), but thereafter it decreased until the ninth year (IRR: 0.63) and then remained stable.

This pattern was reinforced for those who were attendees (ie, subjects who underwent at least the first screening after first invitation), for example, the IRR (screening vs control arm) for metastatic disease at diagnosis was 0.45. A comparison of attendees with the entire control arm could be biased because of a difference in underlying risk due to selection bias. Nevertheless, such a comparison confirms that PSA-based early diagnosis determined a decrease of metastatic disease at diagnosis.

The current result is consistent with an earlier analysis [10] in showing a reduction in metastatic disease at diagnosis, although the magnitude of effect is slightly smaller in the current analysis, with data from a larger number of centres and longer follow-up and data imputation. In the previous paper, a relative reduction of 50% of metastatic disease at diagnosis was reported (IRR: 0.50; 95% CI, 0.40–0.62).

In the present study, we had no information on disease progression after diagnosis. It is not sufficient for screening to increase detection of early disease; it should affect the course of the disease so that deaths are averted or postponed by early treatment. This requires that there is no compensatory increase in disease progression after diagnosis. In the earlier paper [10] , no differences were reported in metastatic disease emerging during the follow-up period (after diagnosis), although a slight but statistically nonsignificant increase was seen in the post-diagnosis incidence of metastasis in the screening arm (IRR: 1.16; 95% CI, 0.9–1.47). The lack of further reduction in development of metastatic disease after diagnosis is not surprising given the higher incidence of low- and intermediate-risk PCa in the screening arm and the lack of any difference in high-risk PCa between the two arms. Hence, a similar occurrence of metastatic disease after diagnosis seems plausible, likely indicating that early diagnosis and treatment could not stop disease progression in all cases. The lack of reduction in incidence of metastasis after diagnosis is in agreement with the smaller reduction of mortality (IRR: 0.79) than in incidence of metastatic PCa (IRR: 0.60).

An update to the European Randomized Study of Screening for Prostate Cancer (ERSPC) has recently confirmed the reduction in prostate cancer (PCa) mortality within the ERSPC screening arm invited to prostate-specific antigen (PSA) testing [1] . The validity of the result has been questioned[2] and [3]based mainly on the differences in treatment between the arms, ignoring the difference in disease characteristics at presentation [4] . Previous evaluations of treatment and different methods of mortality analysis[5], [6], [7], and [8], however, have shown that differences in treatment are unlikely to explain the impact of screening on mortality. Nevertheless, detailed analysis of key prognostic features of PCa across the study arms is worthwhile to better understand how the mortality difference has emerged.

In principle [9] , assuming that the cancer staging system is consistent for the compared groups, the requirements for an effective screening programme are an ability to produce a stage migration (downstaging) in the screening arm, a stage migration accompanied by a decrease in the incidence of advanced cancers, and a decrease in advanced-stage incidence occurring before a mortality reduction in the screening arm.

Meeting these requirements would not mean that early diagnosis had been the cause (or the only cause) of any observed reduction in mortality; however, if these requirements were not met, it would be unlikely that screening could explain any observed mortality reduction.

A reduction in the risk of diagnosis of metastatic disease in the screening arm of four ERSPC centres was shown in a previous work [10] . That paper addressed the occurrence of metastatic disease both at diagnosis and in the post-treatment follow-up.

In the current study, we present updated results on the stage distribution by arm in seven ERSPC centres. We evaluated whether the aforementioned requirements were met and explored the association of the decrease of advanced-cancer incidence rates (if any) with the reduction in PCa mortality.

The ERSPC is a multicentre, randomised screening trial comparing an intervention arm of men to whom regular PSA screening is offered with those in a control arm to whom such screening is not offered. The trial started in Belgium, The Netherlands, Finland, Sweden, Spain, Italy, Switzerland, and France during the period 1993–1998 [11] . The study methodology has been described in detail previously [12] . Data were obtained by linkage to local cancer registries for PCa incidence and national registries for overall mortality. PCa deaths were ascertained by local, independent, cause-of-death committees [13] .

Data on trial arm, centre, T and M stage [14] , Gleason score, serum PSA at diagnosis, age at randomisation, follow-up time (until 31 December 2010), and vital status were extracted from the trial database.

We selected men aged 55–69 yr (the core age group), excluding those who died, who were diagnosed with PCa, or who emigrated between randomisation and screening invitation (n = 145). The median follow-up for the two French centres, Herault and Tarn, was 6–7 yr and thus too short for their data to be included in this analysis.

The risk classification adopted in the last ERSPC analysis [1] was used. Four risk categories at diagnosis were defined using information available at diagnosis: low risk (category 1) indicated clinical stage T1–2 and Gleason score ≤6; intermediate risk (category 2) indicated T1–2 and Gleason score 7 or T3 and Gleason score ≤7; high risk (category 3) indicated T1–3 and Gleason score 8–10 or T4 and any Gleason score; and metastatic disease (category 4) indicated M1 or PSA >100 ng/ml. Data on disease progression and metastasis status during the entire period of postdiagnosis follow-up have already been published for those centres for which information was available [10] . In the present analysis, only information on risk category at diagnosis, available for all centres, was considered.

The accuracy of risk category definition was evaluated by a sensitivity analysis, restricting the evaluation of metastasis to M1 cases only.

We imputed missing data for the variables used in defining the risk categories at diagnosis, using multiple imputations by the technique of chained equations (MICE) [15] . Briefly, MICE uses the distribution of observed data to identify a set of plausible values for the missing data by incorporating a random component to reflect the uncertainty of the estimate. Several data sets are generated, and the same analysis is conducted on each data set with the aim of obtaining a set of parameter estimates that are combined in a single overall estimate (computed as the average of the estimates in each simulated data set) with relative standard errors (computed including the between- and within-imputation variance) and confidence intervals (CIs) [16] .

For the incomplete variables (T and M stage, Gleason score, and PSA level), we derived imputation models that included the complete variables: arm, centre, vital status, follow-up time, age at randomisation, and interactions between each variable and trial arm. The interactions between each variable and trial arm were included to incorporate potential differences among centres in the mechanism causing the missing data. Ten final data sets were generated after the algorithm converged. In the literature on multiple imputation, three to five imputed data sets are considered adequate [17] . The estimates were combined using Rubin's rules [18] . We conducted a sensitivity analysis by imputing directly the risk categories at diagnosis. We used version 12 of the Stata software (StataCorp, College Station, TX, USA).

PCa incidence rate ratios (IRRs) by centre and risk categories at diagnosis were calculated for the screening versus control arms using Poisson regression analysis. Person-years were calculated as time from randomisation to PCa diagnosis, death, or end of follow-up, whichever came first. We adjusted all analyses for the randomisation ratio 1:1.5 for the screening versus control arms in Finland.

Primary analyses were carried out on an intention-to-treat (intention-to-screen) basis (ie, men were included in the analysis in the arm to which they were randomly assigned, regardless of compliance). Moreover, to provide a sensitivity analysis, we split the cumulative occurrence within the screening arm according to screening status, defining subjects who underwent at least the first screening after first invitation, asattendees.

Complete case analysis (before missing data imputation) and all-case analysis (after data imputation) are presented.

In the core age group, 162 338 men were randomised. The median age at randomisation was 60.2 yr.

With data truncated at 13.0 yr of follow-up, 7408 and 6107 PCa cases were diagnosed in the screening and control arms, respectively. The distribution of missing values was similar in the two arms (8% in the screening and 10% in the control arm); however, for some centres (Belgium, Italy, and Spain), data were missing for between 38% and 77% of cases, and there was an imbalance between the arms of at least 10 percentage points ( Table 1 ).

Imputation did not materially affect the results ( Table 2 ); therefore, the results based on imputed data are shown henceforth.

The IRRs (screening vs control arm) for the low- and intermediate-risk categories at diagnosis were significantly elevated (IRR: 2.14 [95% CI, 2.03–2.25]; and 1.24 [95% CI, 1·16–1·34], respectively); the high-risk category had unity (IRR: 1.00; 95% CI, 0.89–1.13). The IRR was significantly reduced by 40% for the metastatic disease category (IRR: 0.60; 95% CI, 0.52–0.70).

The reduction was even greater when restricting the evaluation to M1 cases only (IRR: 0.50; 95% CI, 0.43–0.57).

The absolute excess in the low-risk category was larger than the deficit of advanced disease: The incidence rate difference (IRD) per 1000 randomised men for the low-risk category was 34.26 (95% CI, 31.99–36.52). The IRD for the intermediate-, high-, and metastatic-risk categories at diagnosis were 4.26 (95% CI, 2.74–5.79), 0.47 (95% CI, −1.09–0.74), and −3.14 (95% CI, −3.93 to −2.35), respectively.

The IRRs (screening vs control arm) for the low- and intermediate-risk categories dropped strongly and consistently with follow-up. The high IRRs in the first 3 yr were associated with the prevalence round and overdiagnosis. There was a sharp decrease of IRR in the high-risk category within the first 6 yr, and a further reduction thereafter as the risk converged with the control arm. A slight decrease of IRR was evident until the ninth year for the metastatic disease category, but afterwards it remained constant ( Table 3 ).

A sharper contrast was seen between the arms for the entire 13-yr period when only the attendees were used to calculate the IRRs: 2.45 (95% CI, 2.32–2.58) for low-risk, 1.32 (95% CI, 1.23–1.43) for intermediate-risk, 1.00 (95% CI, 0.88–1.14) for high-risk, and 0.45 (95% CI, 0.38–0.53) for metastatic disease at diagnosis.

The reduction in cumulative (0–13 yr) IRRs (screening vs control arm) for metastatic disease at diagnosis (IRR: 0.60) was clear and evident within the first 6 yr of follow-up, whereas the reduction in mortality was smaller (IRR: 0.79) and emerged roughly 3 yr later ( Fig. 1 ).

Fig. 1 Cumulative incidence rate ratios for risk category 4 (after data imputation) and cumulative prostate cancer mortality rate ratio by time since randomisation (95% confidence intervals given in parentheses). IRD = incidence rate difference per 1000 randomised men; MRD = mortality rate difference per 1000 randomised men.

In the present study, we examined the incidence rates of PCa by risk category at diagnosis between the two arms of the ERSPC trial.

The net effect in the screening arm was a marked increase of the cumulative incidence of low-risk PCa, along with a significant reduction of cumulative incidence of metastatic disease at diagnosis. A slight increase in the intermediate-risk category and no difference in the high-risk category (with an early excess disappearing by year 10) were observed. When interpreting these results, we must keep in mind that early detection resulting in a stage migration is expected to increase early disease, with corresponding decreases in advanced disease.

In this study, the reduction found for metastatic disease at diagnosis was not reflected as an excess in the high-risk category, but increased risk was found only in the intermediate- and, above all, the low-risk categories. The absolute excess in the low-risk category was substantially larger than the deficit of advanced disease: The difference reflects the excess of incidence in the screening arm, which was reported in the aforementioned update to the ERSPC trial along with the conclusion that one PCa death is averted for 27 excess cases detected in 13 yr of follow-up [1] .

To avoid potential biases due to the imbalance of the proportions of missing data in the two arms and among centres, we used multiple imputation. As the results did not change considerably with imputation, results with imputed data are shown because they are more comprehensive and precise.

The IRRs for the screening versus control arms decreased with follow-up time both overall and for each risk category at diagnosis, though not in a similar fashion. Cumulative incidence for low-risk PCa was three times higher in the screening arm than in the control arm during the first 3 yr of follow-up, and then fell off steadily to an incidence twice as high in the screening arm in the overall period of 13 years. The IRRs for intermediate- and high-risk categories also decreased over time. In contrast, a decrease in the cumulative incidence for the metastatic disease category was not evident in the first 3 yr of follow-up (IRR: 0.94), but thereafter it decreased until the ninth year (IRR: 0.63) and then remained stable.

This pattern was reinforced for those who were attendees (ie, subjects who underwent at least the first screening after first invitation), for example, the IRR (screening vs control arm) for metastatic disease at diagnosis was 0.45. A comparison of attendees with the entire control arm could be biased because of a difference in underlying risk due to selection bias. Nevertheless, such a comparison confirms that PSA-based early diagnosis determined a decrease of metastatic disease at diagnosis.

The current result is consistent with an earlier analysis [10] in showing a reduction in metastatic disease at diagnosis, although the magnitude of effect is slightly smaller in the current analysis, with data from a larger number of centres and longer follow-up and data imputation. In the previous paper, a relative reduction of 50% of metastatic disease at diagnosis was reported (IRR: 0.50; 95% CI, 0.40–0.62).

In the present study, we had no information on disease progression after diagnosis. It is not sufficient for screening to increase detection of early disease; it should affect the course of the disease so that deaths are averted or postponed by early treatment. This requires that there is no compensatory increase in disease progression after diagnosis. In the earlier paper [10] , no differences were reported in metastatic disease emerging during the follow-up period (after diagnosis), although a slight but statistically nonsignificant increase was seen in the post-diagnosis incidence of metastasis in the screening arm (IRR: 1.16; 95% CI, 0.9–1.47). The lack of further reduction in development of metastatic disease after diagnosis is not surprising given the higher incidence of low- and intermediate-risk PCa in the screening arm and the lack of any difference in high-risk PCa between the two arms. Hence, a similar occurrence of metastatic disease after diagnosis seems plausible, likely indicating that early diagnosis and treatment could not stop disease progression in all cases. The lack of reduction in incidence of metastasis after diagnosis is in agreement with the smaller reduction of mortality (IRR: 0.79) than in incidence of metastatic PCa (IRR: 0.60).

An update to the European Randomized Study of Screening for Prostate Cancer (ERSPC) has recently confirmed the reduction in prostate cancer (PCa) mortality within the ERSPC screening arm invited to prostate-specific antigen (PSA) testing [1] . The validity of the result has been questioned[2] and [3]based mainly on the differences in treatment between the arms, ignoring the difference in disease characteristics at presentation [4] . Previous evaluations of treatment and different methods of mortality analysis[5], [6], [7], and [8], however, have shown that differences in treatment are unlikely to explain the impact of screening on mortality. Nevertheless, detailed analysis of key prognostic features of PCa across the study arms is worthwhile to better understand how the mortality difference has emerged.

In principle [9] , assuming that the cancer staging system is consistent for the compared groups, the requirements for an effective screening programme are an ability to produce a stage migration (downstaging) in the screening arm, a stage migration accompanied by a decrease in the incidence of advanced cancers, and a decrease in advanced-stage incidence occurring before a mortality reduction in the screening arm.

Meeting these requirements would not mean that early diagnosis had been the cause (or the only cause) of any observed reduction in mortality; however, if these requirements were not met, it would be unlikely that screening could explain any observed mortality reduction.

A reduction in the risk of diagnosis of metastatic disease in the screening arm of four ERSPC centres was shown in a previous work [10] . That paper addressed the occurrence of metastatic disease both at diagnosis and in the post-treatment follow-up.

In the current study, we present updated results on the stage distribution by arm in seven ERSPC centres. We evaluated whether the aforementioned requirements were met and explored the association of the decrease of advanced-cancer incidence rates (if any) with the reduction in PCa mortality.

The ERSPC is a multicentre, randomised screening trial comparing an intervention arm of men to whom regular PSA screening is offered with those in a control arm to whom such screening is not offered. The trial started in Belgium, The Netherlands, Finland, Sweden, Spain, Italy, Switzerland, and France during the period 1993–1998 [11] . The study methodology has been described in detail previously [12] . Data were obtained by linkage to local cancer registries for PCa incidence and national registries for overall mortality. PCa deaths were ascertained by local, independent, cause-of-death committees [13] .

Data on trial arm, centre, T and M stage [14] , Gleason score, serum PSA at diagnosis, age at randomisation, follow-up time (until 31 December 2010), and vital status were extracted from the trial database.

We selected men aged 55–69 yr (the core age group), excluding those who died, who were diagnosed with PCa, or who emigrated between randomisation and screening invitation (n = 145). The median follow-up for the two French centres, Herault and Tarn, was 6–7 yr and thus too short for their data to be included in this analysis.

The risk classification adopted in the last ERSPC analysis [1] was used. Four risk categories at diagnosis were defined using information available at diagnosis: low risk (category 1) indicated clinical stage T1–2 and Gleason score ≤6; intermediate risk (category 2) indicated T1–2 and Gleason score 7 or T3 and Gleason score ≤7; high risk (category 3) indicated T1–3 and Gleason score 8–10 or T4 and any Gleason score; and metastatic disease (category 4) indicated M1 or PSA >100 ng/ml. Data on disease progression and metastasis status during the entire period of postdiagnosis follow-up have already been published for those centres for which information was available [10] . In the present analysis, only information on risk category at diagnosis, available for all centres, was considered.

The accuracy of risk category definition was evaluated by a sensitivity analysis, restricting the evaluation of metastasis to M1 cases only.

We imputed missing data for the variables used in defining the risk categories at diagnosis, using multiple imputations by the technique of chained equations (MICE) [15] . Briefly, MICE uses the distribution of observed data to identify a set of plausible values for the missing data by incorporating a random component to reflect the uncertainty of the estimate. Several data sets are generated, and the same analysis is conducted on each data set with the aim of obtaining a set of parameter estimates that are combined in a single overall estimate (computed as the average of the estimates in each simulated data set) with relative standard errors (computed including the between- and within-imputation variance) and confidence intervals (CIs) [16] .

For the incomplete variables (T and M stage, Gleason score, and PSA level), we derived imputation models that included the complete variables: arm, centre, vital status, follow-up time, age at randomisation, and interactions between each variable and trial arm. The interactions between each variable and trial arm were included to incorporate potential differences among centres in the mechanism causing the missing data. Ten final data sets were generated after the algorithm converged. In the literature on multiple imputation, three to five imputed data sets are considered adequate [17] . The estimates were combined using Rubin's rules [18] . We conducted a sensitivity analysis by imputing directly the risk categories at diagnosis. We used version 12 of the Stata software (StataCorp, College Station, TX, USA).

PCa incidence rate ratios (IRRs) by centre and risk categories at diagnosis were calculated for the screening versus control arms using Poisson regression analysis. Person-years were calculated as time from randomisation to PCa diagnosis, death, or end of follow-up, whichever came first. We adjusted all analyses for the randomisation ratio 1:1.5 for the screening versus control arms in Finland.

Primary analyses were carried out on an intention-to-treat (intention-to-screen) basis (ie, men were included in the analysis in the arm to which they were randomly assigned, regardless of compliance). Moreover, to provide a sensitivity analysis, we split the cumulative occurrence within the screening arm according to screening status, defining subjects who underwent at least the first screening after first invitation, asattendees.

Complete case analysis (before missing data imputation) and all-case analysis (after data imputation) are presented.

In the core age group, 162 338 men were randomised. The median age at randomisation was 60.2 yr.

With data truncated at 13.0 yr of follow-up, 7408 and 6107 PCa cases were diagnosed in the screening and control arms, respectively. The distribution of missing values was similar in the two arms (8% in the screening and 10% in the control arm); however, for some centres (Belgium, Italy, and Spain), data were missing for between 38% and 77% of cases, and there was an imbalance between the arms of at least 10 percentage points ( Table 1 ).

Imputation did not materially affect the results ( Table 2 ); therefore, the results based on imputed data are shown henceforth.

The IRRs (screening vs control arm) for the low- and intermediate-risk categories at diagnosis were significantly elevated (IRR: 2.14 [95% CI, 2.03–2.25]; and 1.24 [95% CI, 1·16–1·34], respectively); the high-risk category had unity (IRR: 1.00; 95% CI, 0.89–1.13). The IRR was significantly reduced by 40% for the metastatic disease category (IRR: 0.60; 95% CI, 0.52–0.70).

The reduction was even greater when restricting the evaluation to M1 cases only (IRR: 0.50; 95% CI, 0.43–0.57).

The absolute excess in the low-risk category was larger than the deficit of advanced disease: The incidence rate difference (IRD) per 1000 randomised men for the low-risk category was 34.26 (95% CI, 31.99–36.52). The IRD for the intermediate-, high-, and metastatic-risk categories at diagnosis were 4.26 (95% CI, 2.74–5.79), 0.47 (95% CI, −1.09–0.74), and −3.14 (95% CI, −3.93 to −2.35), respectively.

The IRRs (screening vs control arm) for the low- and intermediate-risk categories dropped strongly and consistently with follow-up. The high IRRs in the first 3 yr were associated with the prevalence round and overdiagnosis. There was a sharp decrease of IRR in the high-risk category within the first 6 yr, and a further reduction thereafter as the risk converged with the control arm. A slight decrease of IRR was evident until the ninth year for the metastatic disease category, but afterwards it remained constant ( Table 3 ).

A sharper contrast was seen between the arms for the entire 13-yr period when only the attendees were used to calculate the IRRs: 2.45 (95% CI, 2.32–2.58) for low-risk, 1.32 (95% CI, 1.23–1.43) for intermediate-risk, 1.00 (95% CI, 0.88–1.14) for high-risk, and 0.45 (95% CI, 0.38–0.53) for metastatic disease at diagnosis.

The reduction in cumulative (0–13 yr) IRRs (screening vs control arm) for metastatic disease at diagnosis (IRR: 0.60) was clear and evident within the first 6 yr of follow-up, whereas the reduction in mortality was smaller (IRR: 0.79) and emerged roughly 3 yr later ( Fig. 1 ).

Fig. 1 Cumulative incidence rate ratios for risk category 4 (after data imputation) and cumulative prostate cancer mortality rate ratio by time since randomisation (95% confidence intervals given in parentheses). IRD = incidence rate difference per 1000 randomised men; MRD = mortality rate difference per 1000 randomised men.

In the present study, we examined the incidence rates of PCa by risk category at diagnosis between the two arms of the ERSPC trial.

The net effect in the screening arm was a marked increase of the cumulative incidence of low-risk PCa, along with a significant reduction of cumulative incidence of metastatic disease at diagnosis. A slight increase in the intermediate-risk category and no difference in the high-risk category (with an early excess disappearing by year 10) were observed. When interpreting these results, we must keep in mind that early detection resulting in a stage migration is expected to increase early disease, with corresponding decreases in advanced disease.

In this study, the reduction found for metastatic disease at diagnosis was not reflected as an excess in the high-risk category, but increased risk was found only in the intermediate- and, above all, the low-risk categories. The absolute excess in the low-risk category was substantially larger than the deficit of advanced disease: The difference reflects the excess of incidence in the screening arm, which was reported in the aforementioned update to the ERSPC trial along with the conclusion that one PCa death is averted for 27 excess cases detected in 13 yr of follow-up [1] .

To avoid potential biases due to the imbalance of the proportions of missing data in the two arms and among centres, we used multiple imputation. As the results did not change considerably with imputation, results with imputed data are shown because they are more comprehensive and precise.

The IRRs for the screening versus control arms decreased with follow-up time both overall and for each risk category at diagnosis, though not in a similar fashion. Cumulative incidence for low-risk PCa was three times higher in the screening arm than in the control arm during the first 3 yr of follow-up, and then fell off steadily to an incidence twice as high in the screening arm in the overall period of 13 years. The IRRs for intermediate- and high-risk categories also decreased over time. In contrast, a decrease in the cumulative incidence for the metastatic disease category was not evident in the first 3 yr of follow-up (IRR: 0.94), but thereafter it decreased until the ninth year (IRR: 0.63) and then remained stable.

This pattern was reinforced for those who were attendees (ie, subjects who underwent at least the first screening after first invitation), for example, the IRR (screening vs control arm) for metastatic disease at diagnosis was 0.45. A comparison of attendees with the entire control arm could be biased because of a difference in underlying risk due to selection bias. Nevertheless, such a comparison confirms that PSA-based early diagnosis determined a decrease of metastatic disease at diagnosis.

The current result is consistent with an earlier analysis [10] in showing a reduction in metastatic disease at diagnosis, although the magnitude of effect is slightly smaller in the current analysis, with data from a larger number of centres and longer follow-up and data imputation. In the previous paper, a relative reduction of 50% of metastatic disease at diagnosis was reported (IRR: 0.50; 95% CI, 0.40–0.62).

In the present study, we had no information on disease progression after diagnosis. It is not sufficient for screening to increase detection of early disease; it should affect the course of the disease so that deaths are averted or postponed by early treatment. This requires that there is no compensatory increase in disease progression after diagnosis. In the earlier paper [10] , no differences were reported in metastatic disease emerging during the follow-up period (after diagnosis), although a slight but statistically nonsignificant increase was seen in the post-diagnosis incidence of metastasis in the screening arm (IRR: 1.16; 95% CI, 0.9–1.47). The lack of further reduction in development of metastatic disease after diagnosis is not surprising given the higher incidence of low- and intermediate-risk PCa in the screening arm and the lack of any difference in high-risk PCa between the two arms. Hence, a similar occurrence of metastatic disease after diagnosis seems plausible, likely indicating that early diagnosis and treatment could not stop disease progression in all cases. The lack of reduction in incidence of metastasis after diagnosis is in agreement with the smaller reduction of mortality (IRR: 0.79) than in incidence of metastatic PCa (IRR: 0.60).

An update to the European Randomized Study of Screening for Prostate Cancer (ERSPC) has recently confirmed the reduction in prostate cancer (PCa) mortality within the ERSPC screening arm invited to prostate-specific antigen (PSA) testing [1] . The validity of the result has been questioned[2] and [3]based mainly on the differences in treatment between the arms, ignoring the difference in disease characteristics at presentation [4] . Previous evaluations of treatment and different methods of mortality analysis[5], [6], [7], and [8], however, have shown that differences in treatment are unlikely to explain the impact of screening on mortality. Nevertheless, detailed analysis of key prognostic features of PCa across the study arms is worthwhile to better understand how the mortality difference has emerged.

In principle [9] , assuming that the cancer staging system is consistent for the compared groups, the requirements for an effective screening programme are an ability to produce a stage migration (downstaging) in the screening arm, a stage migration accompanied by a decrease in the incidence of advanced cancers, and a decrease in advanced-stage incidence occurring before a mortality reduction in the screening arm.

Meeting these requirements would not mean that early diagnosis had been the cause (or the only cause) of any observed reduction in mortality; however, if these requirements were not met, it would be unlikely that screening could explain any observed mortality reduction.

A reduction in the risk of diagnosis of metastatic disease in the screening arm of four ERSPC centres was shown in a previous work [10] . That paper addressed the occurrence of metastatic disease both at diagnosis and in the post-treatment follow-up.

In the current study, we present updated results on the stage distribution by arm in seven ERSPC centres. We evaluated whether the aforementioned requirements were met and explored the association of the decrease of advanced-cancer incidence rates (if any) with the reduction in PCa mortality.

The ERSPC is a multicentre, randomised screening trial comparing an intervention arm of men to whom regular PSA screening is offered with those in a control arm to whom such screening is not offered. The trial started in Belgium, The Netherlands, Finland, Sweden, Spain, Italy, Switzerland, and France during the period 1993–1998 [11] . The study methodology has been described in detail previously [12] . Data were obtained by linkage to local cancer registries for PCa incidence and national registries for overall mortality. PCa deaths were ascertained by local, independent, cause-of-death committees [13] .

Data on trial arm, centre, T and M stage [14] , Gleason score, serum PSA at diagnosis, age at randomisation, follow-up time (until 31 December 2010), and vital status were extracted from the trial database.

We selected men aged 55–69 yr (the core age group), excluding those who died, who were diagnosed with PCa, or who emigrated between randomisation and screening invitation (n = 145). The median follow-up for the two French centres, Herault and Tarn, was 6–7 yr and thus too short for their data to be included in this analysis.

The risk classification adopted in the last ERSPC analysis [1] was used. Four risk categories at diagnosis were defined using information available at diagnosis: low risk (category 1) indicated clinical stage T1–2 and Gleason score ≤6; intermediate risk (category 2) indicated T1–2 and Gleason score 7 or T3 and Gleason score ≤7; high risk (category 3) indicated T1–3 and Gleason score 8–10 or T4 and any Gleason score; and metastatic disease (category 4) indicated M1 or PSA >100 ng/ml. Data on disease progression and metastasis status during the entire period of postdiagnosis follow-up have already been published for those centres for which information was available [10] . In the present analysis, only information on risk category at diagnosis, available for all centres, was considered.

The accuracy of risk category definition was evaluated by a sensitivity analysis, restricting the evaluation of metastasis to M1 cases only.

We imputed missing data for the variables used in defining the risk categories at diagnosis, using multiple imputations by the technique of chained equations (MICE) [15] . Briefly, MICE uses the distribution of observed data to identify a set of plausible values for the missing data by incorporating a random component to reflect the uncertainty of the estimate. Several data sets are generated, and the same analysis is conducted on each data set with the aim of obtaining a set of parameter estimates that are combined in a single overall estimate (computed as the average of the estimates in each simulated data set) with relative standard errors (computed including the between- and within-imputation variance) and confidence intervals (CIs) [16] .

For the incomplete variables (T and M stage, Gleason score, and PSA level), we derived imputation models that included the complete variables: arm, centre, vital status, follow-up time, age at randomisation, and interactions between each variable and trial arm. The interactions between each variable and trial arm were included to incorporate potential differences among centres in the mechanism causing the missing data. Ten final data sets were generated after the algorithm converged. In the literature on multiple imputation, three to five imputed data sets are considered adequate [17] . The estimates were combined using Rubin's rules [18] . We conducted a sensitivity analysis by imputing directly the risk categories at diagnosis. We used version 12 of the Stata software (StataCorp, College Station, TX, USA).

PCa incidence rate ratios (IRRs) by centre and risk categories at diagnosis were calculated for the screening versus control arms using Poisson regression analysis. Person-years were calculated as time from randomisation to PCa diagnosis, death, or end of follow-up, whichever came first. We adjusted all analyses for the randomisation ratio 1:1.5 for the screening versus control arms in Finland.

Primary analyses were carried out on an intention-to-treat (intention-to-screen) basis (ie, men were included in the analysis in the arm to which they were randomly assigned, regardless of compliance). Moreover, to provide a sensitivity analysis, we split the cumulative occurrence within the screening arm according to screening status, defining subjects who underwent at least the first screening after first invitation, asattendees.

Complete case analysis (before missing data imputation) and all-case analysis (after data imputation) are presented.

In the core age group, 162 338 men were randomised. The median age at randomisation was 60.2 yr.

With data truncated at 13.0 yr of follow-up, 7408 and 6107 PCa cases were diagnosed in the screening and control arms, respectively. The distribution of missing values was similar in the two arms (8% in the screening and 10% in the control arm); however, for some centres (Belgium, Italy, and Spain), data were missing for between 38% and 77% of cases, and there was an imbalance between the arms of at least 10 percentage points ( Table 1 ).

Imputation did not materially affect the results ( Table 2 ); therefore, the results based on imputed data are shown henceforth.

The IRRs (screening vs control arm) for the low- and intermediate-risk categories at diagnosis were significantly elevated (IRR: 2.14 [95% CI, 2.03–2.25]; and 1.24 [95% CI, 1·16–1·34], respectively); the high-risk category had unity (IRR: 1.00; 95% CI, 0.89–1.13). The IRR was significantly reduced by 40% for the metastatic disease category (IRR: 0.60; 95% CI, 0.52–0.70).

The reduction was even greater when restricting the evaluation to M1 cases only (IRR: 0.50; 95% CI, 0.43–0.57).

The absolute excess in the low-risk category was larger than the deficit of advanced disease: The incidence rate difference (IRD) per 1000 randomised men for the low-risk category was 34.26 (95% CI, 31.99–36.52). The IRD for the intermediate-, high-, and metastatic-risk categories at diagnosis were 4.26 (95% CI, 2.74–5.79), 0.47 (95% CI, −1.09–0.74), and −3.14 (95% CI, −3.93 to −2.35), respectively.

The IRRs (screening vs control arm) for the low- and intermediate-risk categories dropped strongly and consistently with follow-up. The high IRRs in the first 3 yr were associated with the prevalence round and overdiagnosis. There was a sharp decrease of IRR in the high-risk category within the first 6 yr, and a further reduction thereafter as the risk converged with the control arm. A slight decrease of IRR was evident until the ninth year for the metastatic disease category, but afterwards it remained constant ( Table 3 ).

A sharper contrast was seen between the arms for the entire 13-yr period when only the attendees were used to calculate the IRRs: 2.45 (95% CI, 2.32–2.58) for low-risk, 1.32 (95% CI, 1.23–1.43) for intermediate-risk, 1.00 (95% CI, 0.88–1.14) for high-risk, and 0.45 (95% CI, 0.38–0.53) for metastatic disease at diagnosis.

The reduction in cumulative (0–13 yr) IRRs (screening vs control arm) for metastatic disease at diagnosis (IRR: 0.60) was clear and evident within the first 6 yr of follow-up, whereas the reduction in mortality was smaller (IRR: 0.79) and emerged roughly 3 yr later ( Fig. 1 ).

Fig. 1 Cumulative incidence rate ratios for risk category 4 (after data imputation) and cumulative prostate cancer mortality rate ratio by time since randomisation (95% confidence intervals given in parentheses). IRD = incidence rate difference per 1000 randomised men; MRD = mortality rate difference per 1000 randomised men.

In the present study, we examined the incidence rates of PCa by risk category at diagnosis between the two arms of the ERSPC trial.

The net effect in the screening arm was a marked increase of the cumulative incidence of low-risk PCa, along with a significant reduction of cumulative incidence of metastatic disease at diagnosis. A slight increase in the intermediate-risk category and no difference in the high-risk category (with an early excess disappearing by year 10) were observed. When interpreting these results, we must keep in mind that early detection resulting in a stage migration is expected to increase early disease, with corresponding decreases in advanced disease.

In this study, the reduction found for metastatic disease at diagnosis was not reflected as an excess in the high-risk category, but increased risk was found only in the intermediate- and, above all, the low-risk categories. The absolute excess in the low-risk category was substantially larger than the deficit of advanced disease: The difference reflects the excess of incidence in the screening arm, which was reported in the aforementioned update to the ERSPC trial along with the conclusion that one PCa death is averted for 27 excess cases detected in 13 yr of follow-up [1] .

To avoid potential biases due to the imbalance of the proportions of missing data in the two arms and among centres, we used multiple imputation. As the results did not change considerably with imputation, results with imputed data are shown because they are more comprehensive and precise.

The IRRs for the screening versus control arms decreased with follow-up time both overall and for each risk category at diagnosis, though not in a similar fashion. Cumulative incidence for low-risk PCa was three times higher in the screening arm than in the control arm during the first 3 yr of follow-up, and then fell off steadily to an incidence twice as high in the screening arm in the overall period of 13 years. The IRRs for intermediate- and high-risk categories also decreased over time. In contrast, a decrease in the cumulative incidence for the metastatic disease category was not evident in the first 3 yr of follow-up (IRR: 0.94), but thereafter it decreased until the ninth year (IRR: 0.63) and then remained stable.

This pattern was reinforced for those who were attendees (ie, subjects who underwent at least the first screening after first invitation), for example, the IRR (screening vs control arm) for metastatic disease at diagnosis was 0.45. A comparison of attendees with the entire control arm could be biased because of a difference in underlying risk due to selection bias. Nevertheless, such a comparison confirms that PSA-based early diagnosis determined a decrease of metastatic disease at diagnosis.

The current result is consistent with an earlier analysis [10] in showing a reduction in metastatic disease at diagnosis, although the magnitude of effect is slightly smaller in the current analysis, with data from a larger number of centres and longer follow-up and data imputation. In the previous paper, a relative reduction of 50% of metastatic disease at diagnosis was reported (IRR: 0.50; 95% CI, 0.40–0.62).

In the present study, we had no information on disease progression after diagnosis. It is not sufficient for screening to increase detection of early disease; it should affect the course of the disease so that deaths are averted or postponed by early treatment. This requires that there is no compensatory increase in disease progression after diagnosis. In the earlier paper [10] , no differences were reported in metastatic disease emerging during the follow-up period (after diagnosis), although a slight but statistically nonsignificant increase was seen in the post-diagnosis incidence of metastasis in the screening arm (IRR: 1.16; 95% CI, 0.9–1.47). The lack of further reduction in development of metastatic disease after diagnosis is not surprising given the higher incidence of low- and intermediate-risk PCa in the screening arm and the lack of any difference in high-risk PCa between the two arms. Hence, a similar occurrence of metastatic disease after diagnosis seems plausible, likely indicating that early diagnosis and treatment could not stop disease progression in all cases. The lack of reduction in incidence of metastasis after diagnosis is in agreement with the smaller reduction of mortality (IRR: 0.79) than in incidence of metastatic PCa (IRR: 0.60).

An update to the European Randomized Study of Screening for Prostate Cancer (ERSPC) has recently confirmed the reduction in prostate cancer (PCa) mortality within the ERSPC screening arm invited to prostate-specific antigen (PSA) testing [1] . The validity of the result has been questioned[2] and [3]based mainly on the differences in treatment between the arms, ignoring the difference in disease characteristics at presentation [4] . Previous evaluations of treatment and different methods of mortality analysis[5], [6], [7], and [8], however, have shown that differences in treatment are unlikely to explain the impact of screening on mortality. Nevertheless, detailed analysis of key prognostic features of PCa across the study arms is worthwhile to better understand how the mortality difference has emerged.

In principle [9] , assuming that the cancer staging system is consistent for the compared groups, the requirements for an effective screening programme are an ability to produce a stage migration (downstaging) in the screening arm, a stage migration accompanied by a decrease in the incidence of advanced cancers, and a decrease in advanced-stage incidence occurring before a mortality reduction in the screening arm.

Meeting these requirements would not mean that early diagnosis had been the cause (or the only cause) of any observed reduction in mortality; however, if these requirements were not met, it would be unlikely that screening could explain any observed mortality reduction.

A reduction in the risk of diagnosis of metastatic disease in the screening arm of four ERSPC centres was shown in a previous work [10] . That paper addressed the occurrence of metastatic disease both at diagnosis and in the post-treatment follow-up.

In the current study, we present updated results on the stage distribution by arm in seven ERSPC centres. We evaluated whether the aforementioned requirements were met and explored the association of the decrease of advanced-cancer incidence rates (if any) with the reduction in PCa mortality.

The ERSPC is a multicentre, randomised screening trial comparing an intervention arm of men to whom regular PSA screening is offered with those in a control arm to whom such screening is not offered. The trial started in Belgium, The Netherlands, Finland, Sweden, Spain, Italy, Switzerland, and France during the period 1993–1998 [11] . The study methodology has been described in detail previously [12] . Data were obtained by linkage to local cancer registries for PCa incidence and national registries for overall mortality. PCa deaths were ascertained by local, independent, cause-of-death committees [13] .

Data on trial arm, centre, T and M stage [14] , Gleason score, serum PSA at diagnosis, age at randomisation, follow-up time (until 31 December 2010), and vital status were extracted from the trial database.

We selected men aged 55–69 yr (the core age group), excluding those who died, who were diagnosed with PCa, or who emigrated between randomisation and screening invitation (n = 145). The median follow-up for the two French centres, Herault and Tarn, was 6–7 yr and thus too short for their data to be included in this analysis.

The risk classification adopted in the last ERSPC analysis [1] was used. Four risk categories at diagnosis were defined using information available at diagnosis: low risk (category 1) indicated clinical stage T1–2 and Gleason score ≤6; intermediate risk (category 2) indicated T1–2 and Gleason score 7 or T3 and Gleason score ≤7; high risk (category 3) indicated T1–3 and Gleason score 8–10 or T4 and any Gleason score; and metastatic disease (category 4) indicated M1 or PSA >100 ng/ml. Data on disease progression and metastasis status during the entire period of postdiagnosis follow-up have already been published for those centres for which information was available [10] . In the present analysis, only information on risk category at diagnosis, available for all centres, was considered.

The accuracy of risk category definition was evaluated by a sensitivity analysis, restricting the evaluation of metastasis to M1 cases only.

We imputed missing data for the variables used in defining the risk categories at diagnosis, using multiple imputations by the technique of chained equations (MICE) [15] . Briefly, MICE uses the distribution of observed data to identify a set of plausible values for the missing data by incorporating a random component to reflect the uncertainty of the estimate. Several data sets are generated, and the same analysis is conducted on each data set with the aim of obtaining a set of parameter estimates that are combined in a single overall estimate (computed as the average of the estimates in each simulated data set) with relative standard errors (computed including the between- and within-imputation variance) and confidence intervals (CIs) [16] .

For the incomplete variables (T and M stage, Gleason score, and PSA level), we derived imputation models that included the complete variables: arm, centre, vital status, follow-up time, age at randomisation, and interactions between each variable and trial arm. The interactions between each variable and trial arm were included to incorporate potential differences among centres in the mechanism causing the missing data. Ten final data sets were generated after the algorithm converged. In the literature on multiple imputation, three to five imputed data sets are considered adequate [17] . The estimates were combined using Rubin's rules [18] . We conducted a sensitivity analysis by imputing directly the risk categories at diagnosis. We used version 12 of the Stata software (StataCorp, College Station, TX, USA).

PCa incidence rate ratios (IRRs) by centre and risk categories at diagnosis were calculated for the screening versus control arms using Poisson regression analysis. Person-years were calculated as time from randomisation to PCa diagnosis, death, or end of follow-up, whichever came first. We adjusted all analyses for the randomisation ratio 1:1.5 for the screening versus control arms in Finland.

Primary analyses were carried out on an intention-to-treat (intention-to-screen) basis (ie, men were included in the analysis in the arm to which they were randomly assigned, regardless of compliance). Moreover, to provide a sensitivity analysis, we split the cumulative occurrence within the screening arm according to screening status, defining subjects who underwent at least the first screening after first invitation, asattendees.

Complete case analysis (before missing data imputation) and all-case analysis (after data imputation) are presented.

In the core age group, 162 338 men were randomised. The median age at randomisation was 60.2 yr.

With data truncated at 13.0 yr of follow-up, 7408 and 6107 PCa cases were diagnosed in the screening and control arms, respectively. The distribution of missing values was similar in the two arms (8% in the screening and 10% in the control arm); however, for some centres (Belgium, Italy, and Spain), data were missing for between 38% and 77% of cases, and there was an imbalance between the arms of at least 10 percentage points ( Table 1 ).

Imputation did not materially affect the results ( Table 2 ); therefore, the results based on imputed data are shown henceforth.

The IRRs (screening vs control arm) for the low- and intermediate-risk categories at diagnosis were significantly elevated (IRR: 2.14 [95% CI, 2.03–2.25]; and 1.24 [95% CI, 1·16–1·34], respectively); the high-risk category had unity (IRR: 1.00; 95% CI, 0.89–1.13). The IRR was significantly reduced by 40% for the metastatic disease category (IRR: 0.60; 95% CI, 0.52–0.70).

The reduction was even greater when restricting the evaluation to M1 cases only (IRR: 0.50; 95% CI, 0.43–0.57).

The absolute excess in the low-risk category was larger than the deficit of advanced disease: The incidence rate difference (IRD) per 1000 randomised men for the low-risk category was 34.26 (95% CI, 31.99–36.52). The IRD for the intermediate-, high-, and metastatic-risk categories at diagnosis were 4.26 (95% CI, 2.74–5.79), 0.47 (95% CI, −1.09–0.74), and −3.14 (95% CI, −3.93 to −2.35), respectively.

The IRRs (screening vs control arm) for the low- and intermediate-risk categories dropped strongly and consistently with follow-up. The high IRRs in the first 3 yr were associated with the prevalence round and overdiagnosis. There was a sharp decrease of IRR in the high-risk category within the first 6 yr, and a further reduction thereafter as the risk converged with the control arm. A slight decrease of IRR was evident until the ninth year for the metastatic disease category, but afterwards it remained constant ( Table 3 ).

A sharper contrast was seen between the arms for the entire 13-yr period when only the attendees were used to calculate the IRRs: 2.45 (95% CI, 2.32–2.58) for low-risk, 1.32 (95% CI, 1.23–1.43) for intermediate-risk, 1.00 (95% CI, 0.88–1.14) for high-risk, and 0.45 (95% CI, 0.38–0.53) for metastatic disease at diagnosis.

The reduction in cumulative (0–13 yr) IRRs (screening vs control arm) for metastatic disease at diagnosis (IRR: 0.60) was clear and evident within the first 6 yr of follow-up, whereas the reduction in mortality was smaller (IRR: 0.79) and emerged roughly 3 yr later ( Fig. 1 ).

Fig. 1 Cumulative incidence rate ratios for risk category 4 (after data imputation) and cumulative prostate cancer mortality rate ratio by time since randomisation (95% confidence intervals given in parentheses). IRD = incidence rate difference per 1000 randomised men; MRD = mortality rate difference per 1000 randomised men.

In the present study, we examined the incidence rates of PCa by risk category at diagnosis between the two arms of the ERSPC trial.

The net effect in the screening arm was a marked increase of the cumulative incidence of low-risk PCa, along with a significant reduction of cumulative incidence of metastatic disease at diagnosis. A slight increase in the intermediate-risk category and no difference in the high-risk category (with an early excess disappearing by year 10) were observed. When interpreting these results, we must keep in mind that early detection resulting in a stage migration is expected to increase early disease, with corresponding decreases in advanced disease.

In this study, the reduction found for metastatic disease at diagnosis was not reflected as an excess in the high-risk category, but increased risk was found only in the intermediate- and, above all, the low-risk categories. The absolute excess in the low-risk category was substantially larger than the deficit of advanced disease: The difference reflects the excess of incidence in the screening arm, which was reported in the aforementioned update to the ERSPC trial along with the conclusion that one PCa death is averted for 27 excess cases detected in 13 yr of follow-up [1] .

To avoid potential biases due to the imbalance of the proportions of missing data in the two arms and among centres, we used multiple imputation. As the results did not change considerably with imputation, results with imputed data are shown because they are more comprehensive and precise.

The IRRs for the screening versus control arms decreased with follow-up time both overall and for each risk category at diagnosis, though not in a similar fashion. Cumulative incidence for low-risk PCa was three times higher in the screening arm than in the control arm during the first 3 yr of follow-up, and then fell off steadily to an incidence twice as high in the screening arm in the overall period of 13 years. The IRRs for intermediate- and high-risk categories also decreased over time. In contrast, a decrease in the cumulative incidence for the metastatic disease category was not evident in the first 3 yr of follow-up (IRR: 0.94), but thereafter it decreased until the ninth year (IRR: 0.63) and then remained stable.

This pattern was reinforced for those who were attendees (ie, subjects who underwent at least the first screening after first invitation), for example, the IRR (screening vs control arm) for metastatic disease at diagnosis was 0.45. A comparison of attendees with the entire control arm could be biased because of a difference in underlying risk due to selection bias. Nevertheless, such a comparison confirms that PSA-based early diagnosis determined a decrease of metastatic disease at diagnosis.

The current result is consistent with an earlier analysis [10] in showing a reduction in metastatic disease at diagnosis, although the magnitude of effect is slightly smaller in the current analysis, with data from a larger number of centres and longer follow-up and data imputation. In the previous paper, a relative reduction of 50% of metastatic disease at diagnosis was reported (IRR: 0.50; 95% CI, 0.40–0.62).

In the present study, we had no information on disease progression after diagnosis. It is not sufficient for screening to increase detection of early disease; it should affect the course of the disease so that deaths are averted or postponed by early treatment. This requires that there is no compensatory increase in disease progression after diagnosis. In the earlier paper [10] , no differences were reported in metastatic disease emerging during the follow-up period (after diagnosis), although a slight but statistically nonsignificant increase was seen in the post-diagnosis incidence of metastasis in the screening arm (IRR: 1.16; 95% CI, 0.9–1.47). The lack of further reduction in development of metastatic disease after diagnosis is not surprising given the higher incidence of low- and intermediate-risk PCa in the screening arm and the lack of any difference in high-risk PCa between the two arms. Hence, a similar occurrence of metastatic disease after diagnosis seems plausible, likely indicating that early diagnosis and treatment could not stop disease progression in all cases. The lack of reduction in incidence of metastasis after diagnosis is in agreement with the smaller reduction of mortality (IRR: 0.79) than in incidence of metastatic PCa (IRR: 0.60).

An update to the European Randomized Study of Screening for Prostate Cancer (ERSPC) has recently confirmed the reduction in prostate cancer (PCa) mortality within the ERSPC screening arm invited to prostate-specific antigen (PSA) testing [1] . The validity of the result has been questioned[2] and [3]based mainly on the differences in treatment between the arms, ignoring the difference in disease characteristics at presentation [4] . Previous evaluations of treatment and different methods of mortality analysis[5], [6], [7], and [8], however, have shown that differences in treatment are unlikely to explain the impact of screening on mortality. Nevertheless, detailed analysis of key prognostic features of PCa across the study arms is worthwhile to better understand how the mortality difference has emerged.

In principle [9] , assuming that the cancer staging system is consistent for the compared groups, the requirements for an effective screening programme are an ability to produce a stage migration (downstaging) in the screening arm, a stage migration accompanied by a decrease in the incidence of advanced cancers, and a decrease in advanced-stage incidence occurring before a mortality reduction in the screening arm.

Meeting these requirements would not mean that early diagnosis had been the cause (or the only cause) of any observed reduction in mortality; however, if these requirements were not met, it would be unlikely that screening could explain any observed mortality reduction.

A reduction in the risk of diagnosis of metastatic disease in the screening arm of four ERSPC centres was shown in a previous work [10] . That paper addressed the occurrence of metastatic disease both at diagnosis and in the post-treatment follow-up.

In the current study, we present updated results on the stage distribution by arm in seven ERSPC centres. We evaluated whether the aforementioned requirements were met and explored the association of the decrease of advanced-cancer incidence rates (if any) with the reduction in PCa mortality.

The ERSPC is a multicentre, randomised screening trial comparing an intervention arm of men to whom regular PSA screening is offered with those in a control arm to whom such screening is not offered. The trial started in Belgium, The Netherlands, Finland, Sweden, Spain, Italy, Switzerland, and France during the period 1993–1998 [11] . The study methodology has been described in detail previously [12] . Data were obtained by linkage to local cancer registries for PCa incidence and national registries for overall mortality. PCa deaths were ascertained by local, independent, cause-of-death committees [13] .

Data on trial arm, centre, T and M stage [14] , Gleason score, serum PSA at diagnosis, age at randomisation, follow-up time (until 31 December 2010), and vital status were extracted from the trial database.

We selected men aged 55–69 yr (the core age group), excluding those who died, who were diagnosed with PCa, or who emigrated between randomisation and screening invitation (n = 145). The median follow-up for the two French centres, Herault and Tarn, was 6–7 yr and thus too short for their data to be included in this analysis.

The risk classification adopted in the last ERSPC analysis [1] was used. Four risk categories at diagnosis were defined using information available at diagnosis: low risk (category 1) indicated clinical stage T1–2 and Gleason score ≤6; intermediate risk (category 2) indicated T1–2 and Gleason score 7 or T3 and Gleason score ≤7; high risk (category 3) indicated T1–3 and Gleason score 8–10 or T4 and any Gleason score; and metastatic disease (category 4) indicated M1 or PSA >100 ng/ml. Data on disease progression and metastasis status during the entire period of postdiagnosis follow-up have already been published for those centres for which information was available [10] . In the present analysis, only information on risk category at diagnosis, available for all centres, was considered.

The accuracy of risk category definition was evaluated by a sensitivity analysis, restricting the evaluation of metastasis to M1 cases only.

We imputed missing data for the variables used in defining the risk categories at diagnosis, using multiple imputations by the technique of chained equations (MICE) [15] . Briefly, MICE uses the distribution of observed data to identify a set of plausible values for the missing data by incorporating a random component to reflect the uncertainty of the estimate. Several data sets are generated, and the same analysis is conducted on each data set with the aim of obtaining a set of parameter estimates that are combined in a single overall estimate (computed as the average of the estimates in each simulated data set) with relative standard errors (computed including the between- and within-imputation variance) and confidence intervals (CIs) [16] .

For the incomplete variables (T and M stage, Gleason score, and PSA level), we derived imputation models that included the complete variables: arm, centre, vital status, follow-up time, age at randomisation, and interactions between each variable and trial arm. The interactions between each variable and trial arm were included to incorporate potential differences among centres in the mechanism causing the missing data. Ten final data sets were generated after the algorithm converged. In the literature on multiple imputation, three to five imputed data sets are considered adequate [17] . The estimates were combined using Rubin's rules [18] . We conducted a sensitivity analysis by imputing directly the risk categories at diagnosis. We used version 12 of the Stata software (StataCorp, College Station, TX, USA).

PCa incidence rate ratios (IRRs) by centre and risk categories at diagnosis were calculated for the screening versus control arms using Poisson regression analysis. Person-years were calculated as time from randomisation to PCa diagnosis, death, or end of follow-up, whichever came first. We adjusted all analyses for the randomisation ratio 1:1.5 for the screening versus control arms in Finland.

Primary analyses were carried out on an intention-to-treat (intention-to-screen) basis (ie, men were included in the analysis in the arm to which they were randomly assigned, regardless of compliance). Moreover, to provide a sensitivity analysis, we split the cumulative occurrence within the screening arm according to screening status, defining subjects who underwent at least the first screening after first invitation, asattendees.

Complete case analysis (before missing data imputation) and all-case analysis (after data imputation) are presented.

In the core age group, 162 338 men were randomised. The median age at randomisation was 60.2 yr.

With data truncated at 13.0 yr of follow-up, 7408 and 6107 PCa cases were diagnosed in the screening and control arms, respectively. The distribution of missing values was similar in the two arms (8% in the screening and 10% in the control arm); however, for some centres (Belgium, Italy, and Spain), data were missing for between 38% and 77% of cases, and there was an imbalance between the arms of at least 10 percentage points ( Table 1 ).

Imputation did not materially affect the results ( Table 2 ); therefore, the results based on imputed data are shown henceforth.

The IRRs (screening vs control arm) for the low- and intermediate-risk categories at diagnosis were significantly elevated (IRR: 2.14 [95% CI, 2.03–2.25]; and 1.24 [95% CI, 1·16–1·34], respectively); the high-risk category had unity (IRR: 1.00; 95% CI, 0.89–1.13). The IRR was significantly reduced by 40% for the metastatic disease category (IRR: 0.60; 95% CI, 0.52–0.70).

The reduction was even greater when restricting the evaluation to M1 cases only (IRR: 0.50; 95% CI, 0.43–0.57).

The absolute excess in the low-risk category was larger than the deficit of advanced disease: The incidence rate difference (IRD) per 1000 randomised men for the low-risk category was 34.26 (95% CI, 31.99–36.52). The IRD for the intermediate-, high-, and metastatic-risk categories at diagnosis were 4.26 (95% CI, 2.74–5.79), 0.47 (95% CI, −1.09–0.74), and −3.14 (95% CI, −3.93 to −2.35), respectively.

The IRRs (screening vs control arm) for the low- and intermediate-risk categories dropped strongly and consistently with follow-up. The high IRRs in the first 3 yr were associated with the prevalence round and overdiagnosis. There was a sharp decrease of IRR in the high-risk category within the first 6 yr, and a further reduction thereafter as the risk converged with the control arm. A slight decrease of IRR was evident until the ninth year for the metastatic disease category, but afterwards it remained constant ( Table 3 ).

A sharper contrast was seen between the arms for the entire 13-yr period when only the attendees were used to calculate the IRRs: 2.45 (95% CI, 2.32–2.58) for low-risk, 1.32 (95% CI, 1.23–1.43) for intermediate-risk, 1.00 (95% CI, 0.88–1.14) for high-risk, and 0.45 (95% CI, 0.38–0.53) for metastatic disease at diagnosis.

The reduction in cumulative (0–13 yr) IRRs (screening vs control arm) for metastatic disease at diagnosis (IRR: 0.60) was clear and evident within the first 6 yr of follow-up, whereas the reduction in mortality was smaller (IRR: 0.79) and emerged roughly 3 yr later ( Fig. 1 ).

Fig. 1 Cumulative incidence rate ratios for risk category 4 (after data imputation) and cumulative prostate cancer mortality rate ratio by time since randomisation (95% confidence intervals given in parentheses). IRD = incidence rate difference per 1000 randomised men; MRD = mortality rate difference per 1000 randomised men.

In the present study, we examined the incidence rates of PCa by risk category at diagnosis between the two arms of the ERSPC trial.

The net effect in the screening arm was a marked increase of the cumulative incidence of low-risk PCa, along with a significant reduction of cumulative incidence of metastatic disease at diagnosis. A slight increase in the intermediate-risk category and no difference in the high-risk category (with an early excess disappearing by year 10) were observed. When interpreting these results, we must keep in mind that early detection resulting in a stage migration is expected to increase early disease, with corresponding decreases in advanced disease.

In this study, the reduction found for metastatic disease at diagnosis was not reflected as an excess in the high-risk category, but increased risk was found only in the intermediate- and, above all, the low-risk categories. The absolute excess in the low-risk category was substantially larger than the deficit of advanced disease: The difference reflects the excess of incidence in the screening arm, which was reported in the aforementioned update to the ERSPC trial along with the conclusion that one PCa death is averted for 27 excess cases detected in 13 yr of follow-up [1] .

To avoid potential biases due to the imbalance of the proportions of missing data in the two arms and among centres, we used multiple imputation. As the results did not change considerably with imputation, results with imputed data are shown because they are more comprehensive and precise.

The IRRs for the screening versus control arms decreased with follow-up time both overall and for each risk category at diagnosis, though not in a similar fashion. Cumulative incidence for low-risk PCa was three times higher in the screening arm than in the control arm during the first 3 yr of follow-up, and then fell off steadily to an incidence twice as high in the screening arm in the overall period of 13 years. The IRRs for intermediate- and high-risk categories also decreased over time. In contrast, a decrease in the cumulative incidence for the metastatic disease category was not evident in the first 3 yr of follow-up (IRR: 0.94), but thereafter it decreased until the ninth year (IRR: 0.63) and then remained stable.

This pattern was reinforced for those who were attendees (ie, subjects who underwent at least the first screening after first invitation), for example, the IRR (screening vs control arm) for metastatic disease at diagnosis was 0.45. A comparison of attendees with the entire control arm could be biased because of a difference in underlying risk due to selection bias. Nevertheless, such a comparison confirms that PSA-based early diagnosis determined a decrease of metastatic disease at diagnosis.

The current result is consistent with an earlier analysis [10] in showing a reduction in metastatic disease at diagnosis, although the magnitude of effect is slightly smaller in the current analysis, with data from a larger number of centres and longer follow-up and data imputation. In the previous paper, a relative reduction of 50% of metastatic disease at diagnosis was reported (IRR: 0.50; 95% CI, 0.40–0.62).

In the present study, we had no information on disease progression after diagnosis. It is not sufficient for screening to increase detection of early disease; it should affect the course of the disease so that deaths are averted or postponed by early treatment. This requires that there is no compensatory increase in disease progression after diagnosis. In the earlier paper [10] , no differences were reported in metastatic disease emerging during the follow-up period (after diagnosis), although a slight but statistically nonsignificant increase was seen in the post-diagnosis incidence of metastasis in the screening arm (IRR: 1.16; 95% CI, 0.9–1.47). The lack of further reduction in development of metastatic disease after diagnosis is not surprising given the higher incidence of low- and intermediate-risk PCa in the screening arm and the lack of any difference in high-risk PCa between the two arms. Hence, a similar occurrence of metastatic disease after diagnosis seems plausible, likely indicating that early diagnosis and treatment could not stop disease progression in all cases. The lack of reduction in incidence of metastasis after diagnosis is in agreement with the smaller reduction of mortality (IRR: 0.79) than in incidence of metastatic PCa (IRR: 0.60).

An update to the European Randomized Study of Screening for Prostate Cancer (ERSPC) has recently confirmed the reduction in prostate cancer (PCa) mortality within the ERSPC screening arm invited to prostate-specific antigen (PSA) testing [1] . The validity of the result has been questioned[2] and [3]based mainly on the differences in treatment between the arms, ignoring the difference in disease characteristics at presentation [4] . Previous evaluations of treatment and different methods of mortality analysis[5], [6], [7], and [8], however, have shown that differences in treatment are unlikely to explain the impact of screening on mortality. Nevertheless, detailed analysis of key prognostic features of PCa across the study arms is worthwhile to better understand how the mortality difference has emerged.

In principle [9] , assuming that the cancer staging system is consistent for the compared groups, the requirements for an effective screening programme are an ability to produce a stage migration (downstaging) in the screening arm, a stage migration accompanied by a decrease in the incidence of advanced cancers, and a decrease in advanced-stage incidence occurring before a mortality reduction in the screening arm.

Meeting these requirements would not mean that early diagnosis had been the cause (or the only cause) of any observed reduction in mortality; however, if these requirements were not met, it would be unlikely that screening could explain any observed mortality reduction.

A reduction in the risk of diagnosis of metastatic disease in the screening arm of four ERSPC centres was shown in a previous work [10] . That paper addressed the occurrence of metastatic disease both at diagnosis and in the post-treatment follow-up.

In the current study, we present updated results on the stage distribution by arm in seven ERSPC centres. We evaluated whether the aforementioned requirements were met and explored the association of the decrease of advanced-cancer incidence rates (if any) with the reduction in PCa mortality.

The ERSPC is a multicentre, randomised screening trial comparing an intervention arm of men to whom regular PSA screening is offered with those in a control arm to whom such screening is not offered. The trial started in Belgium, The Netherlands, Finland, Sweden, Spain, Italy, Switzerland, and France during the period 1993–1998 [11] . The study methodology has been described in detail previously [12] . Data were obtained by linkage to local cancer registries for PCa incidence and national registries for overall mortality. PCa deaths were ascertained by local, independent, cause-of-death committees [13] .

Data on trial arm, centre, T and M stage [14] , Gleason score, serum PSA at diagnosis, age at randomisation, follow-up time (until 31 December 2010), and vital status were extracted from the trial database.

We selected men aged 55–69 yr (the core age group), excluding those who died, who were diagnosed with PCa, or who emigrated between randomisation and screening invitation (n = 145). The median follow-up for the two French centres, Herault and Tarn, was 6–7 yr and thus too short for their data to be included in this analysis.

The risk classification adopted in the last ERSPC analysis [1] was used. Four risk categories at diagnosis were defined using information available at diagnosis: low risk (category 1) indicated clinical stage T1–2 and Gleason score ≤6; intermediate risk (category 2) indicated T1–2 and Gleason score 7 or T3 and Gleason score ≤7; high risk (category 3) indicated T1–3 and Gleason score 8–10 or T4 and any Gleason score; and metastatic disease (category 4) indicated M1 or PSA >100 ng/ml. Data on disease progression and metastasis status during the entire period of postdiagnosis follow-up have already been published for those centres for which information was available [10] . In the present analysis, only information on risk category at diagnosis, available for all centres, was considered.

The accuracy of risk category definition was evaluated by a sensitivity analysis, restricting the evaluation of metastasis to M1 cases only.

We imputed missing data for the variables used in defining the risk categories at diagnosis, using multiple imputations by the technique of chained equations (MICE) [15] . Briefly, MICE uses the distribution of observed data to identify a set of plausible values for the missing data by incorporating a random component to reflect the uncertainty of the estimate. Several data sets are generated, and the same analysis is conducted on each data set with the aim of obtaining a set of parameter estimates that are combined in a single overall estimate (computed as the average of the estimates in each simulated data set) with relative standard errors (computed including the between- and within-imputation variance) and confidence intervals (CIs) [16] .

For the incomplete variables (T and M stage, Gleason score, and PSA level), we derived imputation models that included the complete variables: arm, centre, vital status, follow-up time, age at randomisation, and interactions between each variable and trial arm. The interactions between each variable and trial arm were included to incorporate potential differences among centres in the mechanism causing the missing data. Ten final data sets were generated after the algorithm converged. In the literature on multiple imputation, three to five imputed data sets are considered adequate [17] . The estimates were combined using Rubin's rules [18] . We conducted a sensitivity analysis by imputing directly the risk categories at diagnosis. We used version 12 of the Stata software (StataCorp, College Station, TX, USA).

PCa incidence rate ratios (IRRs) by centre and risk categories at diagnosis were calculated for the screening versus control arms using Poisson regression analysis. Person-years were calculated as time from randomisation to PCa diagnosis, death, or end of follow-up, whichever came first. We adjusted all analyses for the randomisation ratio 1:1.5 for the screening versus control arms in Finland.

Primary analyses were carried out on an intention-to-treat (intention-to-screen) basis (ie, men were included in the analysis in the arm to which they were randomly assigned, regardless of compliance). Moreover, to provide a sensitivity analysis, we split the cumulative occurrence within the screening arm according to screening status, defining subjects who underwent at least the first screening after first invitation, asattendees.

Complete case analysis (before missing data imputation) and all-case analysis (after data imputation) are presented.

In the core age group, 162 338 men were randomised. The median age at randomisation was 60.2 yr.

With data truncated at 13.0 yr of follow-up, 7408 and 6107 PCa cases were diagnosed in the screening and control arms, respectively. The distribution of missing values was similar in the two arms (8% in the screening and 10% in the control arm); however, for some centres (Belgium, Italy, and Spain), data were missing for between 38% and 77% of cases, and there was an imbalance between the arms of at least 10 percentage points ( Table 1 ).

Imputation did not materially affect the results ( Table 2 ); therefore, the results based on imputed data are shown henceforth.

The IRRs (screening vs control arm) for the low- and intermediate-risk categories at diagnosis were significantly elevated (IRR: 2.14 [95% CI, 2.03–2.25]; and 1.24 [95% CI, 1·16–1·34], respectively); the high-risk category had unity (IRR: 1.00; 95% CI, 0.89–1.13). The IRR was significantly reduced by 40% for the metastatic disease category (IRR: 0.60; 95% CI, 0.52–0.70).

The reduction was even greater when restricting the evaluation to M1 cases only (IRR: 0.50; 95% CI, 0.43–0.57).

The absolute excess in the low-risk category was larger than the deficit of advanced disease: The incidence rate difference (IRD) per 1000 randomised men for the low-risk category was 34.26 (95% CI, 31.99–36.52). The IRD for the intermediate-, high-, and metastatic-risk categories at diagnosis were 4.26 (95% CI, 2.74–5.79), 0.47 (95% CI, −1.09–0.74), and −3.14 (95% CI, −3.93 to −2.35), respectively.

The IRRs (screening vs control arm) for the low- and intermediate-risk categories dropped strongly and consistently with follow-up. The high IRRs in the first 3 yr were associated with the prevalence round and overdiagnosis. There was a sharp decrease of IRR in the high-risk category within the first 6 yr, and a further reduction thereafter as the risk converged with the control arm. A slight decrease of IRR was evident until the ninth year for the metastatic disease category, but afterwards it remained constant ( Table 3 ).

A sharper contrast was seen between the arms for the entire 13-yr period when only the attendees were used to calculate the IRRs: 2.45 (95% CI, 2.32–2.58) for low-risk, 1.32 (95% CI, 1.23–1.43) for intermediate-risk, 1.00 (95% CI, 0.88–1.14) for high-risk, and 0.45 (95% CI, 0.38–0.53) for metastatic disease at diagnosis.

The reduction in cumulative (0–13 yr) IRRs (screening vs control arm) for metastatic disease at diagnosis (IRR: 0.60) was clear and evident within the first 6 yr of follow-up, whereas the reduction in mortality was smaller (IRR: 0.79) and emerged roughly 3 yr later ( Fig. 1 ).

Fig. 1 Cumulative incidence rate ratios for risk category 4 (after data imputation) and cumulative prostate cancer mortality rate ratio by time since randomisation (95% confidence intervals given in parentheses). IRD = incidence rate difference per 1000 randomised men; MRD = mortality rate difference per 1000 randomised men.

In the present study, we examined the incidence rates of PCa by risk category at diagnosis between the two arms of the ERSPC trial.

The net effect in the screening arm was a marked increase of the cumulative incidence of low-risk PCa, along with a significant reduction of cumulative incidence of metastatic disease at diagnosis. A slight increase in the intermediate-risk category and no difference in the high-risk category (with an early excess disappearing by year 10) were observed. When interpreting these results, we must keep in mind that early detection resulting in a stage migration is expected to increase early disease, with corresponding decreases in advanced disease.

In this study, the reduction found for metastatic disease at diagnosis was not reflected as an excess in the high-risk category, but increased risk was found only in the intermediate- and, above all, the low-risk categories. The absolute excess in the low-risk category was substantially larger than the deficit of advanced disease: The difference reflects the excess of incidence in the screening arm, which was reported in the aforementioned update to the ERSPC trial along with the conclusion that one PCa death is averted for 27 excess cases detected in 13 yr of follow-up [1] .

To avoid potential biases due to the imbalance of the proportions of missing data in the two arms and among centres, we used multiple imputation. As the results did not change considerably with imputation, results with imputed data are shown because they are more comprehensive and precise.

The IRRs for the screening versus control arms decreased with follow-up time both overall and for each risk category at diagnosis, though not in a similar fashion. Cumulative incidence for low-risk PCa was three times higher in the screening arm than in the control arm during the first 3 yr of follow-up, and then fell off steadily to an incidence twice as high in the screening arm in the overall period of 13 years. The IRRs for intermediate- and high-risk categories also decreased over time. In contrast, a decrease in the cumulative incidence for the metastatic disease category was not evident in the first 3 yr of follow-up (IRR: 0.94), but thereafter it decreased until the ninth year (IRR: 0.63) and then remained stable.

This pattern was reinforced for those who were attendees (ie, subjects who underwent at least the first screening after first invitation), for example, the IRR (screening vs control arm) for metastatic disease at diagnosis was 0.45. A comparison of attendees with the entire control arm could be biased because of a difference in underlying risk due to selection bias. Nevertheless, such a comparison confirms that PSA-based early diagnosis determined a decrease of metastatic disease at diagnosis.

The current result is consistent with an earlier analysis [10] in showing a reduction in metastatic disease at diagnosis, although the magnitude of effect is slightly smaller in the current analysis, with data from a larger number of centres and longer follow-up and data imputation. In the previous paper, a relative reduction of 50% of metastatic disease at diagnosis was reported (IRR: 0.50; 95% CI, 0.40–0.62).

In the present study, we had no information on disease progression after diagnosis. It is not sufficient for screening to increase detection of early disease; it should affect the course of the disease so that deaths are averted or postponed by early treatment. This requires that there is no compensatory increase in disease progression after diagnosis. In the earlier paper [10] , no differences were reported in metastatic disease emerging during the follow-up period (after diagnosis), although a slight but statistically nonsignificant increase was seen in the post-diagnosis incidence of metastasis in the screening arm (IRR: 1.16; 95% CI, 0.9–1.47). The lack of further reduction in development of metastatic disease after diagnosis is not surprising given the higher incidence of low- and intermediate-risk PCa in the screening arm and the lack of any difference in high-risk PCa between the two arms. Hence, a similar occurrence of metastatic disease after diagnosis seems plausible, likely indicating that early diagnosis and treatment could not stop disease progression in all cases. The lack of reduction in incidence of metastasis after diagnosis is in agreement with the smaller reduction of mortality (IRR: 0.79) than in incidence of metastatic PCa (IRR: 0.60).

An update to the European Randomized Study of Screening for Prostate Cancer (ERSPC) has recently confirmed the reduction in prostate cancer (PCa) mortality within the ERSPC screening arm invited to prostate-specific antigen (PSA) testing [1] . The validity of the result has been questioned[2] and [3]based mainly on the differences in treatment between the arms, ignoring the difference in disease characteristics at presentation [4] . Previous evaluations of treatment and different methods of mortality analysis[5], [6], [7], and [8], however, have shown that differences in treatment are unlikely to explain the impact of screening on mortality. Nevertheless, detailed analysis of key prognostic features of PCa across the study arms is worthwhile to better understand how the mortality difference has emerged.

In principle [9] , assuming that the cancer staging system is consistent for the compared groups, the requirements for an effective screening programme are an ability to produce a stage migration (downstaging) in the screening arm, a stage migration accompanied by a decrease in the incidence of advanced cancers, and a decrease in advanced-stage incidence occurring before a mortality reduction in the screening arm.

Meeting these requirements would not mean that early diagnosis had been the cause (or the only cause) of any observed reduction in mortality; however, if these requirements were not met, it would be unlikely that screening could explain any observed mortality reduction.

A reduction in the risk of diagnosis of metastatic disease in the screening arm of four ERSPC centres was shown in a previous work [10] . That paper addressed the occurrence of metastatic disease both at diagnosis and in the post-treatment follow-up.

In the current study, we present updated results on the stage distribution by arm in seven ERSPC centres. We evaluated whether the aforementioned requirements were met and explored the association of the decrease of advanced-cancer incidence rates (if any) with the reduction in PCa mortality.

The ERSPC is a multicentre, randomised screening trial comparing an intervention arm of men to whom regular PSA screening is offered with those in a control arm to whom such screening is not offered. The trial started in Belgium, The Netherlands, Finland, Sweden, Spain, Italy, Switzerland, and France during the period 1993–1998 [11] . The study methodology has been described in detail previously [12] . Data were obtained by linkage to local cancer registries for PCa incidence and national registries for overall mortality. PCa deaths were ascertained by local, independent, cause-of-death committees [13] .

Data on trial arm, centre, T and M stage [14] , Gleason score, serum PSA at diagnosis, age at randomisation, follow-up time (until 31 December 2010), and vital status were extracted from the trial database.

We selected men aged 55–69 yr (the core age group), excluding those who died, who were diagnosed with PCa, or who emigrated between randomisation and screening invitation (n = 145). The median follow-up for the two French centres, Herault and Tarn, was 6–7 yr and thus too short for their data to be included in this analysis.

The risk classification adopted in the last ERSPC analysis [1] was used. Four risk categories at diagnosis were defined using information available at diagnosis: low risk (category 1) indicated clinical stage T1–2 and Gleason score ≤6; intermediate risk (category 2) indicated T1–2 and Gleason score 7 or T3 and Gleason score ≤7; high risk (category 3) indicated T1–3 and Gleason score 8–10 or T4 and any Gleason score; and metastatic disease (category 4) indicated M1 or PSA >100 ng/ml. Data on disease progression and metastasis status during the entire period of postdiagnosis follow-up have already been published for those centres for which information was available [10] . In the present analysis, only information on risk category at diagnosis, available for all centres, was considered.

The accuracy of risk category definition was evaluated by a sensitivity analysis, restricting the evaluation of metastasis to M1 cases only.

We imputed missing data for the variables used in defining the risk categories at diagnosis, using multiple imputations by the technique of chained equations (MICE) [15] . Briefly, MICE uses the distribution of observed data to identify a set of plausible values for the missing data by incorporating a random component to reflect the uncertainty of the estimate. Several data sets are generated, and the same analysis is conducted on each data set with the aim of obtaining a set of parameter estimates that are combined in a single overall estimate (computed as the average of the estimates in each simulated data set) with relative standard errors (computed including the between- and within-imputation variance) and confidence intervals (CIs) [16] .

For the incomplete variables (T and M stage, Gleason score, and PSA level), we derived imputation models that included the complete variables: arm, centre, vital status, follow-up time, age at randomisation, and interactions between each variable and trial arm. The interactions between each variable and trial arm were included to incorporate potential differences among centres in the mechanism causing the missing data. Ten final data sets were generated after the algorithm converged. In the literature on multiple imputation, three to five imputed data sets are considered adequate [17] . The estimates were combined using Rubin's rules [18] . We conducted a sensitivity analysis by imputing directly the risk categories at diagnosis. We used version 12 of the Stata software (StataCorp, College Station, TX, USA).

PCa incidence rate ratios (IRRs) by centre and risk categories at diagnosis were calculated for the screening versus control arms using Poisson regression analysis. Person-years were calculated as time from randomisation to PCa diagnosis, death, or end of follow-up, whichever came first. We adjusted all analyses for the randomisation ratio 1:1.5 for the screening versus control arms in Finland.

Primary analyses were carried out on an intention-to-treat (intention-to-screen) basis (ie, men were included in the analysis in the arm to which they were randomly assigned, regardless of compliance). Moreover, to provide a sensitivity analysis, we split the cumulative occurrence within the screening arm according to screening status, defining subjects who underwent at least the first screening after first invitation, asattendees.

Complete case analysis (before missing data imputation) and all-case analysis (after data imputation) are presented.

In the core age group, 162 338 men were randomised. The median age at randomisation was 60.2 yr.

With data truncated at 13.0 yr of follow-up, 7408 and 6107 PCa cases were diagnosed in the screening and control arms, respectively. The distribution of missing values was similar in the two arms (8% in the screening and 10% in the control arm); however, for some centres (Belgium, Italy, and Spain), data were missing for between 38% and 77% of cases, and there was an imbalance between the arms of at least 10 percentage points ( Table 1 ).

Imputation did not materially affect the results ( Table 2 ); therefore, the results based on imputed data are shown henceforth.

The IRRs (screening vs control arm) for the low- and intermediate-risk categories at diagnosis were significantly elevated (IRR: 2.14 [95% CI, 2.03–2.25]; and 1.24 [95% CI, 1·16–1·34], respectively); the high-risk category had unity (IRR: 1.00; 95% CI, 0.89–1.13). The IRR was significantly reduced by 40% for the metastatic disease category (IRR: 0.60; 95% CI, 0.52–0.70).

The reduction was even greater when restricting the evaluation to M1 cases only (IRR: 0.50; 95% CI, 0.43–0.57).

The absolute excess in the low-risk category was larger than the deficit of advanced disease: The incidence rate difference (IRD) per 1000 randomised men for the low-risk category was 34.26 (95% CI, 31.99–36.52). The IRD for the intermediate-, high-, and metastatic-risk categories at diagnosis were 4.26 (95% CI, 2.74–5.79), 0.47 (95% CI, −1.09–0.74), and −3.14 (95% CI, −3.93 to −2.35), respectively.

The IRRs (screening vs control arm) for the low- and intermediate-risk categories dropped strongly and consistently with follow-up. The high IRRs in the first 3 yr were associated with the prevalence round and overdiagnosis. There was a sharp decrease of IRR in the high-risk category within the first 6 yr, and a further reduction thereafter as the risk converged with the control arm. A slight decrease of IRR was evident until the ninth year for the metastatic disease category, but afterwards it remained constant ( Table 3 ).

A sharper contrast was seen between the arms for the entire 13-yr period when only the attendees were used to calculate the IRRs: 2.45 (95% CI, 2.32–2.58) for low-risk, 1.32 (95% CI, 1.23–1.43) for intermediate-risk, 1.00 (95% CI, 0.88–1.14) for high-risk, and 0.45 (95% CI, 0.38–0.53) for metastatic disease at diagnosis.

The reduction in cumulative (0–13 yr) IRRs (screening vs control arm) for metastatic disease at diagnosis (IRR: 0.60) was clear and evident within the first 6 yr of follow-up, whereas the reduction in mortality was smaller (IRR: 0.79) and emerged roughly 3 yr later ( Fig. 1 ).

Fig. 1 Cumulative incidence rate ratios for risk category 4 (after data imputation) and cumulative prostate cancer mortality rate ratio by time since randomisation (95% confidence intervals given in parentheses). IRD = incidence rate difference per 1000 randomised men; MRD = mortality rate difference per 1000 randomised men.

In the present study, we examined the incidence rates of PCa by risk category at diagnosis between the two arms of the ERSPC trial.

The net effect in the screening arm was a marked increase of the cumulative incidence of low-risk PCa, along with a significant reduction of cumulative incidence of metastatic disease at diagnosis. A slight increase in the intermediate-risk category and no difference in the high-risk category (with an early excess disappearing by year 10) were observed. When interpreting these results, we must keep in mind that early detection resulting in a stage migration is expected to increase early disease, with corresponding decreases in advanced disease.

In this study, the reduction found for metastatic disease at diagnosis was not reflected as an excess in the high-risk category, but increased risk was found only in the intermediate- and, above all, the low-risk categories. The absolute excess in the low-risk category was substantially larger than the deficit of advanced disease: The difference reflects the excess of incidence in the screening arm, which was reported in the aforementioned update to the ERSPC trial along with the conclusion that one PCa death is averted for 27 excess cases detected in 13 yr of follow-up [1] .

To avoid potential biases due to the imbalance of the proportions of missing data in the two arms and among centres, we used multiple imputation. As the results did not change considerably with imputation, results with imputed data are shown because they are more comprehensive and precise.

The IRRs for the screening versus control arms decreased with follow-up time both overall and for each risk category at diagnosis, though not in a similar fashion. Cumulative incidence for low-risk PCa was three times higher in the screening arm than in the control arm during the first 3 yr of follow-up, and then fell off steadily to an incidence twice as high in the screening arm in the overall period of 13 years. The IRRs for intermediate- and high-risk categories also decreased over time. In contrast, a decrease in the cumulative incidence for the metastatic disease category was not evident in the first 3 yr of follow-up (IRR: 0.94), but thereafter it decreased until the ninth year (IRR: 0.63) and then remained stable.

This pattern was reinforced for those who were attendees (ie, subjects who underwent at least the first screening after first invitation), for example, the IRR (screening vs control arm) for metastatic disease at diagnosis was 0.45. A comparison of attendees with the entire control arm could be biased because of a difference in underlying risk due to selection bias. Nevertheless, such a comparison confirms that PSA-based early diagnosis determined a decrease of metastatic disease at diagnosis.

The current result is consistent with an earlier analysis [10] in showing a reduction in metastatic disease at diagnosis, although the magnitude of effect is slightly smaller in the current analysis, with data from a larger number of centres and longer follow-up and data imputation. In the previous paper, a relative reduction of 50% of metastatic disease at diagnosis was reported (IRR: 0.50; 95% CI, 0.40–0.62).

In the present study, we had no information on disease progression after diagnosis. It is not sufficient for screening to increase detection of early disease; it should affect the course of the disease so that deaths are averted or postponed by early treatment. This requires that there is no compensatory increase in disease progression after diagnosis. In the earlier paper [10] , no differences were reported in metastatic disease emerging during the follow-up period (after diagnosis), although a slight but statistically nonsignificant increase was seen in the post-diagnosis incidence of metastasis in the screening arm (IRR: 1.16; 95% CI, 0.9–1.47). The lack of further reduction in development of metastatic disease after diagnosis is not surprising given the higher incidence of low- and intermediate-risk PCa in the screening arm and the lack of any difference in high-risk PCa between the two arms. Hence, a similar occurrence of metastatic disease after diagnosis seems plausible, likely indicating that early diagnosis and treatment could not stop disease progression in all cases. The lack of reduction in incidence of metastasis after diagnosis is in agreement with the smaller reduction of mortality (IRR: 0.79) than in incidence of metastatic PCa (IRR: 0.60).

An update to the European Randomized Study of Screening for Prostate Cancer (ERSPC) has recently confirmed the reduction in prostate cancer (PCa) mortality within the ERSPC screening arm invited to prostate-specific antigen (PSA) testing [1] . The validity of the result has been questioned[2] and [3]based mainly on the differences in treatment between the arms, ignoring the difference in disease characteristics at presentation [4] . Previous evaluations of treatment and different methods of mortality analysis[5], [6], [7], and [8], however, have shown that differences in treatment are unlikely to explain the impact of screening on mortality. Nevertheless, detailed analysis of key prognostic features of PCa across the study arms is worthwhile to better understand how the mortality difference has emerged.

In principle [9] , assuming that the cancer staging system is consistent for the compared groups, the requirements for an effective screening programme are an ability to produce a stage migration (downstaging) in the screening arm, a stage migration accompanied by a decrease in the incidence of advanced cancers, and a decrease in advanced-stage incidence occurring before a mortality reduction in the screening arm.

Meeting these requirements would not mean that early diagnosis had been the cause (or the only cause) of any observed reduction in mortality; however, if these requirements were not met, it would be unlikely that screening could explain any observed mortality reduction.

A reduction in the risk of diagnosis of metastatic disease in the screening arm of four ERSPC centres was shown in a previous work [10] . That paper addressed the occurrence of metastatic disease both at diagnosis and in the post-treatment follow-up.

In the current study, we present updated results on the stage distribution by arm in seven ERSPC centres. We evaluated whether the aforementioned requirements were met and explored the association of the decrease of advanced-cancer incidence rates (if any) with the reduction in PCa mortality.

The ERSPC is a multicentre, randomised screening trial comparing an intervention arm of men to whom regular PSA screening is offered with those in a control arm to whom such screening is not offered. The trial started in Belgium, The Netherlands, Finland, Sweden, Spain, Italy, Switzerland, and France during the period 1993–1998 [11] . The study methodology has been described in detail previously [12] . Data were obtained by linkage to local cancer registries for PCa incidence and national registries for overall mortality. PCa deaths were ascertained by local, independent, cause-of-death committees [13] .

Data on trial arm, centre, T and M stage [14] , Gleason score, serum PSA at diagnosis, age at randomisation, follow-up time (until 31 December 2010), and vital status were extracted from the trial database.

We selected men aged 55–69 yr (the core age group), excluding those who died, who were diagnosed with PCa, or who emigrated between randomisation and screening invitation (n = 145). The median follow-up for the two French centres, Herault and Tarn, was 6–7 yr and thus too short for their data to be included in this analysis.

The risk classification adopted in the last ERSPC analysis [1] was used. Four risk categories at diagnosis were defined using information available at diagnosis: low risk (category 1) indicated clinical stage T1–2 and Gleason score ≤6; intermediate risk (category 2) indicated T1–2 and Gleason score 7 or T3 and Gleason score ≤7; high risk (category 3) indicated T1–3 and Gleason score 8–10 or T4 and any Gleason score; and metastatic disease (category 4) indicated M1 or PSA >100 ng/ml. Data on disease progression and metastasis status during the entire period of postdiagnosis follow-up have already been published for those centres for which information was available [10] . In the present analysis, only information on risk category at diagnosis, available for all centres, was considered.

The accuracy of risk category definition was evaluated by a sensitivity analysis, restricting the evaluation of metastasis to M1 cases only.

We imputed missing data for the variables used in defining the risk categories at diagnosis, using multiple imputations by the technique of chained equations (MICE) [15] . Briefly, MICE uses the distribution of observed data to identify a set of plausible values for the missing data by incorporating a random component to reflect the uncertainty of the estimate. Several data sets are generated, and the same analysis is conducted on each data set with the aim of obtaining a set of parameter estimates that are combined in a single overall estimate (computed as the average of the estimates in each simulated data set) with relative standard errors (computed including the between- and within-imputation variance) and confidence intervals (CIs) [16] .

For the incomplete variables (T and M stage, Gleason score, and PSA level), we derived imputation models that included the complete variables: arm, centre, vital status, follow-up time, age at randomisation, and interactions between each variable and trial arm. The interactions between each variable and trial arm were included to incorporate potential differences among centres in the mechanism causing the missing data. Ten final data sets were generated after the algorithm converged. In the literature on multiple imputation, three to five imputed data sets are considered adequate [17] . The estimates were combined using Rubin's rules [18] . We conducted a sensitivity analysis by imputing directly the risk categories at diagnosis. We used version 12 of the Stata software (StataCorp, College Station, TX, USA).

PCa incidence rate ratios (IRRs) by centre and risk categories at diagnosis were calculated for the screening versus control arms using Poisson regression analysis. Person-years were calculated as time from randomisation to PCa diagnosis, death, or end of follow-up, whichever came first. We adjusted all analyses for the randomisation ratio 1:1.5 for the screening versus control arms in Finland.

Primary analyses were carried out on an intention-to-treat (intention-to-screen) basis (ie, men were included in the analysis in the arm to which they were randomly assigned, regardless of compliance). Moreover, to provide a sensitivity analysis, we split the cumulative occurrence within the screening arm according to screening status, defining subjects who underwent at least the first screening after first invitation, asattendees.

Complete case analysis (before missing data imputation) and all-case analysis (after data imputation) are presented.

In the core age group, 162 338 men were randomised. The median age at randomisation was 60.2 yr.

With data truncated at 13.0 yr of follow-up, 7408 and 6107 PCa cases were diagnosed in the screening and control arms, respectively. The distribution of missing values was similar in the two arms (8% in the screening and 10% in the control arm); however, for some centres (Belgium, Italy, and Spain), data were missing for between 38% and 77% of cases, and there was an imbalance between the arms of at least 10 percentage points ( Table 1 ).

Imputation did not materially affect the results ( Table 2 ); therefore, the results based on imputed data are shown henceforth.

The IRRs (screening vs control arm) for the low- and intermediate-risk categories at diagnosis were significantly elevated (IRR: 2.14 [95% CI, 2.03–2.25]; and 1.24 [95% CI, 1·16–1·34], respectively); the high-risk category had unity (IRR: 1.00; 95% CI, 0.89–1.13). The IRR was significantly reduced by 40% for the metastatic disease category (IRR: 0.60; 95% CI, 0.52–0.70).

The reduction was even greater when restricting the evaluation to M1 cases only (IRR: 0.50; 95% CI, 0.43–0.57).

The absolute excess in the low-risk category was larger than the deficit of advanced disease: The incidence rate difference (IRD) per 1000 randomised men for the low-risk category was 34.26 (95% CI, 31.99–36.52). The IRD for the intermediate-, high-, and metastatic-risk categories at diagnosis were 4.26 (95% CI, 2.74–5.79), 0.47 (95% CI, −1.09–0.74), and −3.14 (95% CI, −3.93 to −2.35), respectively.

The IRRs (screening vs control arm) for the low- and intermediate-risk categories dropped strongly and consistently with follow-up. The high IRRs in the first 3 yr were associated with the prevalence round and overdiagnosis. There was a sharp decrease of IRR in the high-risk category within the first 6 yr, and a further reduction thereafter as the risk converged with the control arm. A slight decrease of IRR was evident until the ninth year for the metastatic disease category, but afterwards it remained constant ( Table 3 ).

A sharper contrast was seen between the arms for the entire 13-yr period when only the attendees were used to calculate the IRRs: 2.45 (95% CI, 2.32–2.58) for low-risk, 1.32 (95% CI, 1.23–1.43) for intermediate-risk, 1.00 (95% CI, 0.88–1.14) for high-risk, and 0.45 (95% CI, 0.38–0.53) for metastatic disease at diagnosis.

The reduction in cumulative (0–13 yr) IRRs (screening vs control arm) for metastatic disease at diagnosis (IRR: 0.60) was clear and evident within the first 6 yr of follow-up, whereas the reduction in mortality was smaller (IRR: 0.79) and emerged roughly 3 yr later ( Fig. 1 ).

Fig. 1 Cumulative incidence rate ratios for risk category 4 (after data imputation) and cumulative prostate cancer mortality rate ratio by time since randomisation (95% confidence intervals given in parentheses). IRD = incidence rate difference per 1000 randomised men; MRD = mortality rate difference per 1000 randomised men.

In the present study, we examined the incidence rates of PCa by risk category at diagnosis between the two arms of the ERSPC trial.

The net effect in the screening arm was a marked increase of the cumulative incidence of low-risk PCa, along with a significant reduction of cumulative incidence of metastatic disease at diagnosis. A slight increase in the intermediate-risk category and no difference in the high-risk category (with an early excess disappearing by year 10) were observed. When interpreting these results, we must keep in mind that early detection resulting in a stage migration is expected to increase early disease, with corresponding decreases in advanced disease.

In this study, the reduction found for metastatic disease at diagnosis was not reflected as an excess in the high-risk category, but increased risk was found only in the intermediate- and, above all, the low-risk categories. The absolute excess in the low-risk category was substantially larger than the deficit of advanced disease: The difference reflects the excess of incidence in the screening arm, which was reported in the aforementioned update to the ERSPC trial along with the conclusion that one PCa death is averted for 27 excess cases detected in 13 yr of follow-up [1] .

To avoid potential biases due to the imbalance of the proportions of missing data in the two arms and among centres, we used multiple imputation. As the results did not change considerably with imputation, results with imputed data are shown because they are more comprehensive and precise.

The IRRs for the screening versus control arms decreased with follow-up time both overall and for each risk category at diagnosis, though not in a similar fashion. Cumulative incidence for low-risk PCa was three times higher in the screening arm than in the control arm during the first 3 yr of follow-up, and then fell off steadily to an incidence twice as high in the screening arm in the overall period of 13 years. The IRRs for intermediate- and high-risk categories also decreased over time. In contrast, a decrease in the cumulative incidence for the metastatic disease category was not evident in the first 3 yr of follow-up (IRR: 0.94), but thereafter it decreased until the ninth year (IRR: 0.63) and then remained stable.

This pattern was reinforced for those who were attendees (ie, subjects who underwent at least the first screening after first invitation), for example, the IRR (screening vs control arm) for metastatic disease at diagnosis was 0.45. A comparison of attendees with the entire control arm could be biased because of a difference in underlying risk due to selection bias. Nevertheless, such a comparison confirms that PSA-based early diagnosis determined a decrease of metastatic disease at diagnosis.

The current result is consistent with an earlier analysis [10] in showing a reduction in metastatic disease at diagnosis, although the magnitude of effect is slightly smaller in the current analysis, with data from a larger number of centres and longer follow-up and data imputation. In the previous paper, a relative reduction of 50% of metastatic disease at diagnosis was reported (IRR: 0.50; 95% CI, 0.40–0.62).

In the present study, we had no information on disease progression after diagnosis. It is not sufficient for screening to increase detection of early disease; it should affect the course of the disease so that deaths are averted or postponed by early treatment. This requires that there is no compensatory increase in disease progression after diagnosis. In the earlier paper [10] , no differences were reported in metastatic disease emerging during the follow-up period (after diagnosis), although a slight but statistically nonsignificant increase was seen in the post-diagnosis incidence of metastasis in the screening arm (IRR: 1.16; 95% CI, 0.9–1.47). The lack of further reduction in development of metastatic disease after diagnosis is not surprising given the higher incidence of low- and intermediate-risk PCa in the screening arm and the lack of any difference in high-risk PCa between the two arms. Hence, a similar occurrence of metastatic disease after diagnosis seems plausible, likely indicating that early diagnosis and treatment could not stop disease progression in all cases. The lack of reduction in incidence of metastasis after diagnosis is in agreement with the smaller reduction of mortality (IRR: 0.79) than in incidence of metastatic PCa (IRR: 0.60).

An update to the European Randomized Study of Screening for Prostate Cancer (ERSPC) has recently confirmed the reduction in prostate cancer (PCa) mortality within the ERSPC screening arm invited to prostate-specific antigen (PSA) testing [1] . The validity of the result has been questioned[2] and [3]based mainly on the differences in treatment between the arms, ignoring the difference in disease characteristics at presentation [4] . Previous evaluations of treatment and different methods of mortality analysis[5], [6], [7], and [8], however, have shown that differences in treatment are unlikely to explain the impact of screening on mortality. Nevertheless, detailed analysis of key prognostic features of PCa across the study arms is worthwhile to better understand how the mortality difference has emerged.

In principle [9] , assuming that the cancer staging system is consistent for the compared groups, the requirements for an effective screening programme are an ability to produce a stage migration (downstaging) in the screening arm, a stage migration accompanied by a decrease in the incidence of advanced cancers, and a decrease in advanced-stage incidence occurring before a mortality reduction in the screening arm.

Meeting these requirements would not mean that early diagnosis had been the cause (or the only cause) of any observed reduction in mortality; however, if these requirements were not met, it would be unlikely that screening could explain any observed mortality reduction.

A reduction in the risk of diagnosis of metastatic disease in the screening arm of four ERSPC centres was shown in a previous work [10] . That paper addressed the occurrence of metastatic disease both at diagnosis and in the post-treatment follow-up.

In the current study, we present updated results on the stage distribution by arm in seven ERSPC centres. We evaluated whether the aforementioned requirements were met and explored the association of the decrease of advanced-cancer incidence rates (if any) with the reduction in PCa mortality.

The ERSPC is a multicentre, randomised screening trial comparing an intervention arm of men to whom regular PSA screening is offered with those in a control arm to whom such screening is not offered. The trial started in Belgium, The Netherlands, Finland, Sweden, Spain, Italy, Switzerland, and France during the period 1993–1998 [11] . The study methodology has been described in detail previously [12] . Data were obtained by linkage to local cancer registries for PCa incidence and national registries for overall mortality. PCa deaths were ascertained by local, independent, cause-of-death committees [13] .

Data on trial arm, centre, T and M stage [14] , Gleason score, serum PSA at diagnosis, age at randomisation, follow-up time (until 31 December 2010), and vital status were extracted from the trial database.

We selected men aged 55–69 yr (the core age group), excluding those who died, who were diagnosed with PCa, or who emigrated between randomisation and screening invitation (n = 145). The median follow-up for the two French centres, Herault and Tarn, was 6–7 yr and thus too short for their data to be included in this analysis.

The risk classification adopted in the last ERSPC analysis [1] was used. Four risk categories at diagnosis were defined using information available at diagnosis: low risk (category 1) indicated clinical stage T1–2 and Gleason score ≤6; intermediate risk (category 2) indicated T1–2 and Gleason score 7 or T3 and Gleason score ≤7; high risk (category 3) indicated T1–3 and Gleason score 8–10 or T4 and any Gleason score; and metastatic disease (category 4) indicated M1 or PSA >100 ng/ml. Data on disease progression and metastasis status during the entire period of postdiagnosis follow-up have already been published for those centres for which information was available [10] . In the present analysis, only information on risk category at diagnosis, available for all centres, was considered.

The accuracy of risk category definition was evaluated by a sensitivity analysis, restricting the evaluation of metastasis to M1 cases only.

We imputed missing data for the variables used in defining the risk categories at diagnosis, using multiple imputations by the technique of chained equations (MICE) [15] . Briefly, MICE uses the distribution of observed data to identify a set of plausible values for the missing data by incorporating a random component to reflect the uncertainty of the estimate. Several data sets are generated, and the same analysis is conducted on each data set with the aim of obtaining a set of parameter estimates that are combined in a single overall estimate (computed as the average of the estimates in each simulated data set) with relative standard errors (computed including the between- and within-imputation variance) and confidence intervals (CIs) [16] .

For the incomplete variables (T and M stage, Gleason score, and PSA level), we derived imputation models that included the complete variables: arm, centre, vital status, follow-up time, age at randomisation, and interactions between each variable and trial arm. The interactions between each variable and trial arm were included to incorporate potential differences among centres in the mechanism causing the missing data. Ten final data sets were generated after the algorithm converged. In the literature on multiple imputation, three to five imputed data sets are considered adequate [17] . The estimates were combined using Rubin's rules [18] . We conducted a sensitivity analysis by imputing directly the risk categories at diagnosis. We used version 12 of the Stata software (StataCorp, College Station, TX, USA).

PCa incidence rate ratios (IRRs) by centre and risk categories at diagnosis were calculated for the screening versus control arms using Poisson regression analysis. Person-years were calculated as time from randomisation to PCa diagnosis, death, or end of follow-up, whichever came first. We adjusted all analyses for the randomisation ratio 1:1.5 for the screening versus control arms in Finland.

Primary analyses were carried out on an intention-to-treat (intention-to-screen) basis (ie, men were included in the analysis in the arm to which they were randomly assigned, regardless of compliance). Moreover, to provide a sensitivity analysis, we split the cumulative occurrence within the screening arm according to screening status, defining subjects who underwent at least the first screening after first invitation, asattendees.

Complete case analysis (before missing data imputation) and all-case analysis (after data imputation) are presented.

In the core age group, 162 338 men were randomised. The median age at randomisation was 60.2 yr.

With data truncated at 13.0 yr of follow-up, 7408 and 6107 PCa cases were diagnosed in the screening and control arms, respectively. The distribution of missing values was similar in the two arms (8% in the screening and 10% in the control arm); however, for some centres (Belgium, Italy, and Spain), data were missing for between 38% and 77% of cases, and there was an imbalance between the arms of at least 10 percentage points ( Table 1 ).

Imputation did not materially affect the results ( Table 2 ); therefore, the results based on imputed data are shown henceforth.

The IRRs (screening vs control arm) for the low- and intermediate-risk categories at diagnosis were significantly elevated (IRR: 2.14 [95% CI, 2.03–2.25]; and 1.24 [95% CI, 1·16–1·34], respectively); the high-risk category had unity (IRR: 1.00; 95% CI, 0.89–1.13). The IRR was significantly reduced by 40% for the metastatic disease category (IRR: 0.60; 95% CI, 0.52–0.70).

The reduction was even greater when restricting the evaluation to M1 cases only (IRR: 0.50; 95% CI, 0.43–0.57).

The absolute excess in the low-risk category was larger than the deficit of advanced disease: The incidence rate difference (IRD) per 1000 randomised men for the low-risk category was 34.26 (95% CI, 31.99–36.52). The IRD for the intermediate-, high-, and metastatic-risk categories at diagnosis were 4.26 (95% CI, 2.74–5.79), 0.47 (95% CI, −1.09–0.74), and −3.14 (95% CI, −3.93 to −2.35), respectively.

The IRRs (screening vs control arm) for the low- and intermediate-risk categories dropped strongly and consistently with follow-up. The high IRRs in the first 3 yr were associated with the prevalence round and overdiagnosis. There was a sharp decrease of IRR in the high-risk category within the first 6 yr, and a further reduction thereafter as the risk converged with the control arm. A slight decrease of IRR was evident until the ninth year for the metastatic disease category, but afterwards it remained constant ( Table 3 ).

A sharper contrast was seen between the arms for the entire 13-yr period when only the attendees were used to calculate the IRRs: 2.45 (95% CI, 2.32–2.58) for low-risk, 1.32 (95% CI, 1.23–1.43) for intermediate-risk, 1.00 (95% CI, 0.88–1.14) for high-risk, and 0.45 (95% CI, 0.38–0.53) for metastatic disease at diagnosis.

The reduction in cumulative (0–13 yr) IRRs (screening vs control arm) for metastatic disease at diagnosis (IRR: 0.60) was clear and evident within the first 6 yr of follow-up, whereas the reduction in mortality was smaller (IRR: 0.79) and emerged roughly 3 yr later ( Fig. 1 ).

Fig. 1 Cumulative incidence rate ratios for risk category 4 (after data imputation) and cumulative prostate cancer mortality rate ratio by time since randomisation (95% confidence intervals given in parentheses). IRD = incidence rate difference per 1000 randomised men; MRD = mortality rate difference per 1000 randomised men.

In the present study, we examined the incidence rates of PCa by risk category at diagnosis between the two arms of the ERSPC trial.

The net effect in the screening arm was a marked increase of the cumulative incidence of low-risk PCa, along with a significant reduction of cumulative incidence of metastatic disease at diagnosis. A slight increase in the intermediate-risk category and no difference in the high-risk category (with an early excess disappearing by year 10) were observed. When interpreting these results, we must keep in mind that early detection resulting in a stage migration is expected to increase early disease, with corresponding decreases in advanced disease.

In this study, the reduction found for metastatic disease at diagnosis was not reflected as an excess in the high-risk category, but increased risk was found only in the intermediate- and, above all, the low-risk categories. The absolute excess in the low-risk category was substantially larger than the deficit of advanced disease: The difference reflects the excess of incidence in the screening arm, which was reported in the aforementioned update to the ERSPC trial along with the conclusion that one PCa death is averted for 27 excess cases detected in 13 yr of follow-up [1] .

To avoid potential biases due to the imbalance of the proportions of missing data in the two arms and among centres, we used multiple imputation. As the results did not change considerably with imputation, results with imputed data are shown because they are more comprehensive and precise.

The IRRs for the screening versus control arms decreased with follow-up time both overall and for each risk category at diagnosis, though not in a similar fashion. Cumulative incidence for low-risk PCa was three times higher in the screening arm than in the control arm during the first 3 yr of follow-up, and then fell off steadily to an incidence twice as high in the screening arm in the overall period of 13 years. The IRRs for intermediate- and high-risk categories also decreased over time. In contrast, a decrease in the cumulative incidence for the metastatic disease category was not evident in the first 3 yr of follow-up (IRR: 0.94), but thereafter it decreased until the ninth year (IRR: 0.63) and then remained stable.

This pattern was reinforced for those who were attendees (ie, subjects who underwent at least the first screening after first invitation), for example, the IRR (screening vs control arm) for metastatic disease at diagnosis was 0.45. A comparison of attendees with the entire control arm could be biased because of a difference in underlying risk due to selection bias. Nevertheless, such a comparison confirms that PSA-based early diagnosis determined a decrease of metastatic disease at diagnosis.

The current result is consistent with an earlier analysis [10] in showing a reduction in metastatic disease at diagnosis, although the magnitude of effect is slightly smaller in the current analysis, with data from a larger number of centres and longer follow-up and data imputation. In the previous paper, a relative reduction of 50% of metastatic disease at diagnosis was reported (IRR: 0.50; 95% CI, 0.40–0.62).

In the present study, we had no information on disease progression after diagnosis. It is not sufficient for screening to increase detection of early disease; it should affect the course of the disease so that deaths are averted or postponed by early treatment. This requires that there is no compensatory increase in disease progression after diagnosis. In the earlier paper [10] , no differences were reported in metastatic disease emerging during the follow-up period (after diagnosis), although a slight but statistically nonsignificant increase was seen in the post-diagnosis incidence of metastasis in the screening arm (IRR: 1.16; 95% CI, 0.9–1.47). The lack of further reduction in development of metastatic disease after diagnosis is not surprising given the higher incidence of low- and intermediate-risk PCa in the screening arm and the lack of any difference in high-risk PCa between the two arms. Hence, a similar occurrence of metastatic disease after diagnosis seems plausible, likely indicating that early diagnosis and treatment could not stop disease progression in all cases. The lack of reduction in incidence of metastasis after diagnosis is in agreement with the smaller reduction of mortality (IRR: 0.79) than in incidence of metastatic PCa (IRR: 0.60).

An update to the European Randomized Study of Screening for Prostate Cancer (ERSPC) has recently confirmed the reduction in prostate cancer (PCa) mortality within the ERSPC screening arm invited to prostate-specific antigen (PSA) testing [1] . The validity of the result has been questioned[2] and [3]based mainly on the differences in treatment between the arms, ignoring the difference in disease characteristics at presentation [4] . Previous evaluations of treatment and different methods of mortality analysis[5], [6], [7], and [8], however, have shown that differences in treatment are unlikely to explain the impact of screening on mortality. Nevertheless, detailed analysis of key prognostic features of PCa across the study arms is worthwhile to better understand how the mortality difference has emerged.

In principle [9] , assuming that the cancer staging system is consistent for the compared groups, the requirements for an effective screening programme are an ability to produce a stage migration (downstaging) in the screening arm, a stage migration accompanied by a decrease in the incidence of advanced cancers, and a decrease in advanced-stage incidence occurring before a mortality reduction in the screening arm.

Meeting these requirements would not mean that early diagnosis had been the cause (or the only cause) of any observed reduction in mortality; however, if these requirements were not met, it would be unlikely that screening could explain any observed mortality reduction.

A reduction in the risk of diagnosis of metastatic disease in the screening arm of four ERSPC centres was shown in a previous work [10] . That paper addressed the occurrence of metastatic disease both at diagnosis and in the post-treatment follow-up.

In the current study, we present updated results on the stage distribution by arm in seven ERSPC centres. We evaluated whether the aforementioned requirements were met and explored the association of the decrease of advanced-cancer incidence rates (if any) with the reduction in PCa mortality.

The ERSPC is a multicentre, randomised screening trial comparing an intervention arm of men to whom regular PSA screening is offered with those in a control arm to whom such screening is not offered. The trial started in Belgium, The Netherlands, Finland, Sweden, Spain, Italy, Switzerland, and France during the period 1993–1998 [11] . The study methodology has been described in detail previously [12] . Data were obtained by linkage to local cancer registries for PCa incidence and national registries for overall mortality. PCa deaths were ascertained by local, independent, cause-of-death committees [13] .

Data on trial arm, centre, T and M stage [14] , Gleason score, serum PSA at diagnosis, age at randomisation, follow-up time (until 31 December 2010), and vital status were extracted from the trial database.

We selected men aged 55–69 yr (the core age group), excluding those who died, who were diagnosed with PCa, or who emigrated between randomisation and screening invitation (n = 145). The median follow-up for the two French centres, Herault and Tarn, was 6–7 yr and thus too short for their data to be included in this analysis.

The risk classification adopted in the last ERSPC analysis [1] was used. Four risk categories at diagnosis were defined using information available at diagnosis: low risk (category 1) indicated clinical stage T1–2 and Gleason score ≤6; intermediate risk (category 2) indicated T1–2 and Gleason score 7 or T3 and Gleason score ≤7; high risk (category 3) indicated T1–3 and Gleason score 8–10 or T4 and any Gleason score; and metastatic disease (category 4) indicated M1 or PSA >100 ng/ml. Data on disease progression and metastasis status during the entire period of postdiagnosis follow-up have already been published for those centres for which information was available [10] . In the present analysis, only information on risk category at diagnosis, available for all centres, was considered.

The accuracy of risk category definition was evaluated by a sensitivity analysis, restricting the evaluation of metastasis to M1 cases only.

We imputed missing data for the variables used in defining the risk categories at diagnosis, using multiple imputations by the technique of chained equations (MICE) [15] . Briefly, MICE uses the distribution of observed data to identify a set of plausible values for the missing data by incorporating a random component to reflect the uncertainty of the estimate. Several data sets are generated, and the same analysis is conducted on each data set with the aim of obtaining a set of parameter estimates that are combined in a single overall estimate (computed as the average of the estimates in each simulated data set) with relative standard errors (computed including the between- and within-imputation variance) and confidence intervals (CIs) [16] .

For the incomplete variables (T and M stage, Gleason score, and PSA level), we derived imputation models that included the complete variables: arm, centre, vital status, follow-up time, age at randomisation, and interactions between each variable and trial arm. The interactions between each variable and trial arm were included to incorporate potential differences among centres in the mechanism causing the missing data. Ten final data sets were generated after the algorithm converged. In the literature on multiple imputation, three to five imputed data sets are considered adequate [17] . The estimates were combined using Rubin's rules [18] . We conducted a sensitivity analysis by imputing directly the risk categories at diagnosis. We used version 12 of the Stata software (StataCorp, College Station, TX, USA).

PCa incidence rate ratios (IRRs) by centre and risk categories at diagnosis were calculated for the screening versus control arms using Poisson regression analysis. Person-years were calculated as time from randomisation to PCa diagnosis, death, or end of follow-up, whichever came first. We adjusted all analyses for the randomisation ratio 1:1.5 for the screening versus control arms in Finland.

Primary analyses were carried out on an intention-to-treat (intention-to-screen) basis (ie, men were included in the analysis in the arm to which they were randomly assigned, regardless of compliance). Moreover, to provide a sensitivity analysis, we split the cumulative occurrence within the screening arm according to screening status, defining subjects who underwent at least the first screening after first invitation, asattendees.

Complete case analysis (before missing data imputation) and all-case analysis (after data imputation) are presented.

In the core age group, 162 338 men were randomised. The median age at randomisation was 60.2 yr.

With data truncated at 13.0 yr of follow-up, 7408 and 6107 PCa cases were diagnosed in the screening and control arms, respectively. The distribution of missing values was similar in the two arms (8% in the screening and 10% in the control arm); however, for some centres (Belgium, Italy, and Spain), data were missing for between 38% and 77% of cases, and there was an imbalance between the arms of at least 10 percentage points ( Table 1 ).

Imputation did not materially affect the results ( Table 2 ); therefore, the results based on imputed data are shown henceforth.

The IRRs (screening vs control arm) for the low- and intermediate-risk categories at diagnosis were significantly elevated (IRR: 2.14 [95% CI, 2.03–2.25]; and 1.24 [95% CI, 1·16–1·34], respectively); the high-risk category had unity (IRR: 1.00; 95% CI, 0.89–1.13). The IRR was significantly reduced by 40% for the metastatic disease category (IRR: 0.60; 95% CI, 0.52–0.70).

The reduction was even greater when restricting the evaluation to M1 cases only (IRR: 0.50; 95% CI, 0.43–0.57).

The absolute excess in the low-risk category was larger than the deficit of advanced disease: The incidence rate difference (IRD) per 1000 randomised men for the low-risk category was 34.26 (95% CI, 31.99–36.52). The IRD for the intermediate-, high-, and metastatic-risk categories at diagnosis were 4.26 (95% CI, 2.74–5.79), 0.47 (95% CI, −1.09–0.74), and −3.14 (95% CI, −3.93 to −2.35), respectively.

The IRRs (screening vs control arm) for the low- and intermediate-risk categories dropped strongly and consistently with follow-up. The high IRRs in the first 3 yr were associated with the prevalence round and overdiagnosis. There was a sharp decrease of IRR in the high-risk category within the first 6 yr, and a further reduction thereafter as the risk converged with the control arm. A slight decrease of IRR was evident until the ninth year for the metastatic disease category, but afterwards it remained constant ( Table 3 ).

A sharper contrast was seen between the arms for the entire 13-yr period when only the attendees were used to calculate the IRRs: 2.45 (95% CI, 2.32–2.58) for low-risk, 1.32 (95% CI, 1.23–1.43) for intermediate-risk, 1.00 (95% CI, 0.88–1.14) for high-risk, and 0.45 (95% CI, 0.38–0.53) for metastatic disease at diagnosis.

The reduction in cumulative (0–13 yr) IRRs (screening vs control arm) for metastatic disease at diagnosis (IRR: 0.60) was clear and evident within the first 6 yr of follow-up, whereas the reduction in mortality was smaller (IRR: 0.79) and emerged roughly 3 yr later ( Fig. 1 ).

Fig. 1 Cumulative incidence rate ratios for risk category 4 (after data imputation) and cumulative prostate cancer mortality rate ratio by time since randomisation (95% confidence intervals given in parentheses). IRD = incidence rate difference per 1000 randomised men; MRD = mortality rate difference per 1000 randomised men.

In the present study, we examined the incidence rates of PCa by risk category at diagnosis between the two arms of the ERSPC trial.

The net effect in the screening arm was a marked increase of the cumulative incidence of low-risk PCa, along with a significant reduction of cumulative incidence of metastatic disease at diagnosis. A slight increase in the intermediate-risk category and no difference in the high-risk category (with an early excess disappearing by year 10) were observed. When interpreting these results, we must keep in mind that early detection resulting in a stage migration is expected to increase early disease, with corresponding decreases in advanced disease.

In this study, the reduction found for metastatic disease at diagnosis was not reflected as an excess in the high-risk category, but increased risk was found only in the intermediate- and, above all, the low-risk categories. The absolute excess in the low-risk category was substantially larger than the deficit of advanced disease: The difference reflects the excess of incidence in the screening arm, which was reported in the aforementioned update to the ERSPC trial along with the conclusion that one PCa death is averted for 27 excess cases detected in 13 yr of follow-up [1] .

To avoid potential biases due to the imbalance of the proportions of missing data in the two arms and among centres, we used multiple imputation. As the results did not change considerably with imputation, results with imputed data are shown because they are more comprehensive and precise.

The IRRs for the screening versus control arms decreased with follow-up time both overall and for each risk category at diagnosis, though not in a similar fashion. Cumulative incidence for low-risk PCa was three times higher in the screening arm than in the control arm during the first 3 yr of follow-up, and then fell off steadily to an incidence twice as high in the screening arm in the overall period of 13 years. The IRRs for intermediate- and high-risk categories also decreased over time. In contrast, a decrease in the cumulative incidence for the metastatic disease category was not evident in the first 3 yr of follow-up (IRR: 0.94), but thereafter it decreased until the ninth year (IRR: 0.63) and then remained stable.

This pattern was reinforced for those who were attendees (ie, subjects who underwent at least the first screening after first invitation), for example, the IRR (screening vs control arm) for metastatic disease at diagnosis was 0.45. A comparison of attendees with the entire control arm could be biased because of a difference in underlying risk due to selection bias. Nevertheless, such a comparison confirms that PSA-based early diagnosis determined a decrease of metastatic disease at diagnosis.

The current result is consistent with an earlier analysis [10] in showing a reduction in metastatic disease at diagnosis, although the magnitude of effect is slightly smaller in the current analysis, with data from a larger number of centres and longer follow-up and data imputation. In the previous paper, a relative reduction of 50% of metastatic disease at diagnosis was reported (IRR: 0.50; 95% CI, 0.40–0.62).

In the present study, we had no information on disease progression after diagnosis. It is not sufficient for screening to increase detection of early disease; it should affect the course of the disease so that deaths are averted or postponed by early treatment. This requires that there is no compensatory increase in disease progression after diagnosis. In the earlier paper [10] , no differences were reported in metastatic disease emerging during the follow-up period (after diagnosis), although a slight but statistically nonsignificant increase was seen in the post-diagnosis incidence of metastasis in the screening arm (IRR: 1.16; 95% CI, 0.9–1.47). The lack of further reduction in development of metastatic disease after diagnosis is not surprising given the higher incidence of low- and intermediate-risk PCa in the screening arm and the lack of any difference in high-risk PCa between the two arms. Hence, a similar occurrence of metastatic disease after diagnosis seems plausible, likely indicating that early diagnosis and treatment could not stop disease progression in all cases. The lack of reduction in incidence of metastasis after diagnosis is in agreement with the smaller reduction of mortality (IRR: 0.79) than in incidence of metastatic PCa (IRR: 0.60).

An update to the European Randomized Study of Screening for Prostate Cancer (ERSPC) has recently confirmed the reduction in prostate cancer (PCa) mortality within the ERSPC screening arm invited to prostate-specific antigen (PSA) testing [1] . The validity of the result has been questioned[2] and [3]based mainly on the differences in treatment between the arms, ignoring the difference in disease characteristics at presentation [4] . Previous evaluations of treatment and different methods of mortality analysis[5], [6], [7], and [8], however, have shown that differences in treatment are unlikely to explain the impact of screening on mortality. Nevertheless, detailed analysis of key prognostic features of PCa across the study arms is worthwhile to better understand how the mortality difference has emerged.

In principle [9] , assuming that the cancer staging system is consistent for the compared groups, the requirements for an effective screening programme are an ability to produce a stage migration (downstaging) in the screening arm, a stage migration accompanied by a decrease in the incidence of advanced cancers, and a decrease in advanced-stage incidence occurring before a mortality reduction in the screening arm.

Meeting these requirements would not mean that early diagnosis had been the cause (or the only cause) of any observed reduction in mortality; however, if these requirements were not met, it would be unlikely that screening could explain any observed mortality reduction.

A reduction in the risk of diagnosis of metastatic disease in the screening arm of four ERSPC centres was shown in a previous work [10] . That paper addressed the occurrence of metastatic disease both at diagnosis and in the post-treatment follow-up.

In the current study, we present updated results on the stage distribution by arm in seven ERSPC centres. We evaluated whether the aforementioned requirements were met and explored the association of the decrease of advanced-cancer incidence rates (if any) with the reduction in PCa mortality.

The ERSPC is a multicentre, randomised screening trial comparing an intervention arm of men to whom regular PSA screening is offered with those in a control arm to whom such screening is not offered. The trial started in Belgium, The Netherlands, Finland, Sweden, Spain, Italy, Switzerland, and France during the period 1993–1998 [11] . The study methodology has been described in detail previously [12] . Data were obtained by linkage to local cancer registries for PCa incidence and national registries for overall mortality. PCa deaths were ascertained by local, independent, cause-of-death committees [13] .

Data on trial arm, centre, T and M stage [14] , Gleason score, serum PSA at diagnosis, age at randomisation, follow-up time (until 31 December 2010), and vital status were extracted from the trial database.

We selected men aged 55–69 yr (the core age group), excluding those who died, who were diagnosed with PCa, or who emigrated between randomisation and screening invitation (n = 145). The median follow-up for the two French centres, Herault and Tarn, was 6–7 yr and thus too short for their data to be included in this analysis.

The risk classification adopted in the last ERSPC analysis [1] was used. Four risk categories at diagnosis were defined using information available at diagnosis: low risk (category 1) indicated clinical stage T1–2 and Gleason score ≤6; intermediate risk (category 2) indicated T1–2 and Gleason score 7 or T3 and Gleason score ≤7; high risk (category 3) indicated T1–3 and Gleason score 8–10 or T4 and any Gleason score; and metastatic disease (category 4) indicated M1 or PSA >100 ng/ml. Data on disease progression and metastasis status during the entire period of postdiagnosis follow-up have already been published for those centres for which information was available [10] . In the present analysis, only information on risk category at diagnosis, available for all centres, was considered.

The accuracy of risk category definition was evaluated by a sensitivity analysis, restricting the evaluation of metastasis to M1 cases only.

We imputed missing data for the variables used in defining the risk categories at diagnosis, using multiple imputations by the technique of chained equations (MICE) [15] . Briefly, MICE uses the distribution of observed data to identify a set of plausible values for the missing data by incorporating a random component to reflect the uncertainty of the estimate. Several data sets are generated, and the same analysis is conducted on each data set with the aim of obtaining a set of parameter estimates that are combined in a single overall estimate (computed as the average of the estimates in each simulated data set) with relative standard errors (computed including the between- and within-imputation variance) and confidence intervals (CIs) [16] .

For the incomplete variables (T and M stage, Gleason score, and PSA level), we derived imputation models that included the complete variables: arm, centre, vital status, follow-up time, age at randomisation, and interactions between each variable and trial arm. The interactions between each variable and trial arm were included to incorporate potential differences among centres in the mechanism causing the missing data. Ten final data sets were generated after the algorithm converged. In the literature on multiple imputation, three to five imputed data sets are considered adequate [17] . The estimates were combined using Rubin's rules [18] . We conducted a sensitivity analysis by imputing directly the risk categories at diagnosis. We used version 12 of the Stata software (StataCorp, College Station, TX, USA).

PCa incidence rate ratios (IRRs) by centre and risk categories at diagnosis were calculated for the screening versus control arms using Poisson regression analysis. Person-years were calculated as time from randomisation to PCa diagnosis, death, or end of follow-up, whichever came first. We adjusted all analyses for the randomisation ratio 1:1.5 for the screening versus control arms in Finland.

Primary analyses were carried out on an intention-to-treat (intention-to-screen) basis (ie, men were included in the analysis in the arm to which they were randomly assigned, regardless of compliance). Moreover, to provide a sensitivity analysis, we split the cumulative occurrence within the screening arm according to screening status, defining subjects who underwent at least the first screening after first invitation, asattendees.

Complete case analysis (before missing data imputation) and all-case analysis (after data imputation) are presented.

In the core age group, 162 338 men were randomised. The median age at randomisation was 60.2 yr.

With data truncated at 13.0 yr of follow-up, 7408 and 6107 PCa cases were diagnosed in the screening and control arms, respectively. The distribution of missing values was similar in the two arms (8% in the screening and 10% in the control arm); however, for some centres (Belgium, Italy, and Spain), data were missing for between 38% and 77% of cases, and there was an imbalance between the arms of at least 10 percentage points ( Table 1 ).