Long-term Trends in Prostate Cancer Incidence by Stage at Diagnosis in Japan Using the Multiple Imputation Approach, 1993–2014

Prostate cancer is the fourth most common cancer globally, and more than 1.2 million prostate cancer cases are estimated to have occurred in 2018 (1). Asia has the second largest share of the prostate cancer incidence in the world (23.3%) next to Europe (35.2%); however, the share of mortality due to prostate cancer is the highest in Asia, which constitutes around 33% of the prostate cancer mortality of the world (1). In Japan, the incidence of prostate cancer has been growing rapidly over the past decades, and the age-standardized incidence rate rose from 7.1 in 1975 to 58.7 per 100,000 in 2014, becoming the fourth common cancer in men (2). Mortality due to prostate cancer has been steady declining since 2000s, yet it ranks the sixth major cause of cancer death in Japanese men (2).

Screening for PSA began in the 1990s globally, including Japan, and has enhanced the detection of asymptomatic prostate cancer (3). To date, however, the question of whether PSA testing has made possible the detection of prostate cancer before it develops to a later stage, or whether the trend is in fact a manifestation of overdiagnosis of asymptomatic tumors that would not otherwise have been clinically detected throughout life has not been clarified. The time–trend analysis of prostate cancer before and after the introduction of PSA testing in Japanese municipalities would help elucidate the impact of screening on stage-specific prostate cancer incidence, but trend analysis of stage-specific incidence is hampered by the fact that stage information was not fully reported in earlier days, a problem common to population-based cancer registries everywhere. Accordingly, the resulting trend estimates would still include selection bias.

In recent years, use of multiple imputation has been shown to be a useful tool to account for missing values in cancer registries (4–7). The multiple imputation (MI) approach yields a set of plausible values using an imputation model, which are then combined to produce a mean estimate and SEs to account for estimation accuracy between the imputed values (8, 9). In this study, we aimed to assess the trend of prostate cancer by stage at diagnosis between 1993 and 2014 using the MI approach with population-based cancer registry data.

Data on cancer incidence in Japan were obtained from three population-based cancer registries (Yamagata, Fukui, and Nagasaki prefectures) as part of the Monitoring of Cancer Incidence in Japan Project (10). These three registries were established to monitor long-term trends in cancer incidence while ensuring a high degree of accuracy and timeliness over a long period of time. Incidence is therefore less likely to be impacted by changes in data quality (11, 12). Furthermore, a previous study confirmed that the incidence rates and trends obtained from these prefectures combined are comparable with those in the overall Japanese population (11).

We extracted 29,418 malignant prostate cancer cases diagnosed between 1993 and 2014 from the three prefectures. The incidence of prostate cancer was classified according to the International Statistical Classification of Diseases and Related Health Problems, 10th Revision (ICD-10) coded as “C61.” The study was approved by the Institutional Review Board of the National Cancer Center in Tokyo (approval number 2004-061).

In the Japanese cancer registries, stage at diagnosis is determined by the degree of spread of cancer and classified as “Localized,” “Regional lymph node metastasis,”, “Adjacent organ invasion,” and “Distant metastasis” (10). Cancer staging in Japanese cancer registries follows the tumor–node–metastasis (TNM) classification, which was developed jointly by the American Joint Commission on Cancer and Union for International Cancer Control, and classification has remained consistent over the study period for prostate cancer (13). In the present study, we grouped clinical stage into three categories according to SEER Summary Staging Manual 2000, namely, “Localized (localized to the tissue of origin)”, “Regional (spread to an adjacent organ, muscle, connective tissue, fat, serosa, or regional lymph node)”, and “Distant (spread to another place in the body)” (14). We treated cases with “unknown” status and those with an unreported extent of disease as missing, and then imputed values by the multiple imputation approach. We assumed that the missing cases were not completely randomly distributed among cases following previous studies (missing completely at random; refs. 15–17), but rather that the missing patterns of stage at diagnosis depended on the observed data (missing at random,; ref. 18), because the missingness of stage depended largely on the year of diagnosis (i.e., earlier diagnoses had considerably more missing cases). Patterns of missing data are described in the Results section.

For imputation, we included variables from the cancer registry data that can predict stage at diagnosis and its missingness. Variables included in the imputation model were age at diagnosis; prefecture; year of diagnosis; whether the case was DCN (death certificate notification) or not; whether the case was DCO (death certificate only) or not; whether the tumor was detected by screening, including PSA screening or general health check-up (yes or no); type of primary treatment received, including surgery (yes or no), radiotherapy (yes or no), and chemotherapy (yes or no); observation period (from time of diagnosis to last confirmed year of survival, year of death, or censor year, whichever came first); and vital status during the observation period. Unless death or last confirmed date of survival was reported, the observation period was censored at the last year and month of follow-up, namely, 2016 for Yamagata and Fukui, and 2015 for Nagasaki. We did not include histologic type because the majority of the cases were adenocarcinoma and other histologic types were rarely observed. Similarly, we did not include laparoscopy because the national fee schedule for health insurance only began coverage of this treatment in 2006, and only around 1.9% of total cases received it. Socioeconomic inequalities are significantly associated with cancer incidence and survival in Japan (19, 20) including prostate cancer (21), but these are not collected in Japanese cancer registries. To complement this, we used a prefecture-level Gini coefficient (an index of income inequality), which ranges from 0 (perfect equality) to 1 (perfect inequality; ref. 22). The Gini coefficient is published periodically in the National Survey on Family Income and Expenditure reports conducted by the Ministry of Internal Affairs and Communications (1999, 2004, 2009, and 2015; ref. 23). The Gini coefficient from the most recent year before diagnosis was included as a covariate in the imputation model to account for prefecture-level inequalities.

The MI approach took each case with missing values and imputed for it a specific stage at diagnosis, that is, localized, regional, or distant. For this purpose, sequential multinomial logistic regression, alternatively called multiple imputation with chained equations, was used to impute each variable with missing observations conditional on the remaining variables in the dataset, such that no variable remains with a missing observation (7, 24). In addition, imputed variables included receipt of chemotherapy, surgery, and/or radiotherapy, because other variables had no missing values. Missing data were imputed in order from the most observed to the least observed (7). We set the random-number seed for the Markov chain at 123 to ensure the reproducibility of the results. The same process was repeated 10 times to create 10 sets of completed datasets to account for the variability in imputation. SE was calculated by combining the mean of the squared SEs of the 10 imputed estimates and the variance of 10 incidence rate estimates based on Rubin's rule (25). To assess the validity of the imputation model, we performed a Kaplan–Meier log-rank test of equality with the Bonferroni correction for overall survival by stage at diagnosis and by year of diagnosis, comparing the complete dataset (excluding cases with missing values) and the 10 imputed datasets (5). If the imputation model is valid, the survival probability by each stage in the imputed datasets would yield results that are not significantly different from the complete dataset (5). Details of the test are presented in the Results section.

We estimated the age-standardized incidence rates using a 1985 Japanese standard population in 18 age groups from each of the imputed dataset and then yielded a single incidence rate, which was the mean of 10 imputed data estimates. To detect the transition points at which the trend changes, a joinpoint linear regression model (26) was used to fit log-linear trends with a maximum of four joinpoints using the Joinpoint regression program version 4.6.0.0, developed by the U.S. NCI. We searched for the optimal parameters by grid search with a size of one year and the permutation tests were performed on the basis of 4,500 Monte Carlo iterations. Annual percentage change (APC) was calculated to show the trends between breakpoints with 95% confidence intervals (CI). Because prostate cancer cases identified from DCO had considerably more missing covariates than other cases, we also performed the same imputation and joinpoint regression after excluding all DCO cases (876 cases) as a sensitivity analysis. All analyses except the joinpoint regression used STATA version 14.0 software (StataCorp LP), and P < 0.05 was considered to denote statistical significance.

Table 1 shows the distribution of prostate cancer cases by reported status of stage at diagnosis. Around 76.4% of cases had a confirmed stage. Among the cases that reported an unknown or missing status, around 20.5% are coded as “unknown,” and 3.1% had not reported any status. Table 2 shows details of the proportions of prostate cancer cases by stage at diagnosis, including unknown and missing status for each of the selected covariates. Patients with unknown/missing stage were likely to be older and year of diagnosis tended to be earlier than patients with a reported stage. Nagasaki prefecture was found to have more unknown/missing cases than Yamagata and Fukui prefectures. Screening-detected prostate cancer cases were more likely to report a higher proportion of confirmed stage at diagnosis than those detected otherwise. Patients who received surgery, chemotherapy, or radiotherapy as primary treatment also had a higher proportion of confirmed stage compared with those who did not receive these therapies or those with an unknown therapeutic status.

Figure 1 shows trends in age-adjusted incidence rates by stage at diagnosis, including unknown and missing status, for patients with prostate cancer from 1993 to 2014. Prostate cancer cases with unknown or missing status gradually increased by 2001, then showed a steady decline until 2014.

Figure 2 shows the age-adjusted incidence rates for prostate cancer by stage at diagnosis before and after imputing the missing values (also shown in Supplementary Table S3). The incidence rate for localized prostate cancer increased sharply from 2002 to 2010, and then started to decline after 2011 before imputing missing cases. After imputing the missing and unknown cases, a pronounced increase in localized cancer incidence was seen, from 12.4 per 100,000 in 2001 to 34.3 per 100,000 in 2003. Regional prostate cancer also showed an increasing trend until 2006. Supplementary Fig. S1 shows the same age-adjusted incidence rates separately for the Nagasaki registry and the combined Yamagata and Fukui registries. Despite the high proportion of cases with unknown and missing status in the Nagasaki registry, its overall trends did not differ from those of the Yamagata and Fukui registries after imputation.

Table 3 shows the detailed results of the joinpoint trend analysis. For each trend segment, the numbers listed as “original” and “imputed” represent the APC for that specific segment, with an asterisk indicating a statistically significant trend at the P < 0.05 level. For localized prostate cancer, a significant increase was seen in both the original (APC, 7.0%; P < 0.001) and the imputed dataset (APC, 4.1%; P < 0.001) between 2003 and 2014, and an insignificantly increasing trend was seen between 2000 and 2003 in the original (APC, 50.9%; P = 0.200) and the imputed dataset (APC, 41.8%; P = 0.104). For regional cancer, the original data showed a significantly increasing trend until 2009 (APC, 12.6%; P < 0.001), while the imputed data showed that the significantly increasing trend stopped earlier in 2006 (APC, 12.5%; P < 0.001). For distant cancer, the original data showed an increasing trend until 2014 with no change in trend throughout (APC, 2.4%; P < 0.001), while the imputed data showed that the increasing trend stopped in 2004 (APC, 2.4%; P < 0.001), and showed a marginal decrease thereafter (APC, −2.2%; P < 0.001). Our sensitivity analyses (Supplementary Table S1) after excluding all DCO cases show consistent trends, except for distant cancer between 2004 and 2014, which indicated an insignificant decline in incidence.

Supplementary Table S2 shows the results of the Kaplan–Meier log-rank test for equality of survival probability by year of diagnosis and by stage at diagnosis comparing the complete dataset and imputed datasets. Estimated survival probability for the imputed data differed significantly from the complete data for localized cancer only between 2005 and 2009 (P < 0.001). However, this significant difference resulted from the imputed data reporting better survival probability than the complete data, meaning that the imputation model correctly assigned cancer stage to “Localized.”

This is one of only a few studies that have investigated trends in stage-specific prostate cancer over 20 years which can compare data before and after the introduction of PSA screening. The strength of our study lies in the use of high-quality cancer registry data over time that is comparable with the national estimates of cancer incidence. Furthermore, the three cancer registries in our study are spread out from the Tohoku (northeastern Japan) to Kyushu (southern Japan) regions, minimizing bias induced by geographic distribution.

Several findings of this study warrant emphasis. First, we found a pronounced hike in the incidence of localized prostate cancer between 2000 and 2003 after imputing missing information on stage at diagnosis, albeit that the trend was not significant, possibly due to rapid fluctuation of the imputed incidence rates before and after 2003 (joinpoint analysis found a significant increase after 2003). Second, trends in the incidence of distant prostate cancer also changed after imputing the missing information: the original data before imputation showed a constantly increasing trend until 2014, while the imputed data showed that the increasing trend stopped in 2004 then turned to a declining trend. The observed trends in the incidence of distant prostate cancer around 2004 are consistent with those of a previous study that reported trends in prostate cancer mortality, in which a significant decline in prostate cancer mortality started in 2004 (12). However, trends seen in distant prostate cancer were not as distinct as those in localized prostate cancer.

Our findings suggest a stage shifting, an increase in localized cancer coupled with a marginal decrease in the incidence of distant cancer, which coincides with the time PSA screening spread among Japanese municipalities in the early 2000s. A significant increase in regional cancer between 1993 and 2006 also occurred when PSA screening spread rapidly in early 2000s. A study reported that the percentage of municipalities implementing PSA screening has increased from 14.7% in 2000 to 49.9% in 2003, 71.2% in 2006, and eventually to 83.0% in 2015 (27). In our study area, the percentage of municipalities implementing PSA screening increased rapidly between 2000 (Yamagata, 4.5%; Nagasaki, 3.8%), 2003 (Yamagata, 64.7%; Fukui, 55.6%; Nagasaki, 30.2%), and 2006 (Yamagata, 74.1%; Fukui, 78.6%; Nagasaki, 76.2%; ref. 27). Consequently, screening uptake has also increased; another report covering 252 municipalities in Japan showed that the number of people undergoing PSA testing increased from 31,917 in 2002 to 109,902 in 2003, and to 156,695 in 2004 (28). The year 2003 also coincides with the time the former Emperor Akihito was diagnosed with prostate cancer and underwent surgery in January 2003 (29), which fueled public concern for prostate cancer and the importance of screening (30).

However, whether PSA screening alone can explain the sharp increase in the incidence of localized prostate cancer requiring any type of treatment is questionable. Although no study to date has reported the nationwide coverage of PSA screening, estimated coverage appears to be low; for instance, 15.7% of men ages above 40 years in Osaka prefecture were estimated to have received PSA screening (31), versus only around 5.8% men above 50 years in Gunma prefecture (32). This level of coverage is considerably lower than that in the United States, where 43.6% of U.S. men ages 60 to 74 years were screened even after the 2012 U.S. Preventive Services Task Force Recommendation discouraged PSA screening (33).

Another possibility to explain the shift in stage is overdiagnosis. In the European Randomized Study of Screening for Prostate Cancer (ERSPC), Schroder and colleagues reported that around 20% of the reduction in prostate cancer mortality is due to PSA screening, but also estimated that the rate of overdiagnosis of prostate cancer, in which the participants would not have shown clinical symptoms during the lifetime had it not been for screening, is 50% in the screened group (34). Results obtained from simulation modeling for PSA screening also estimate the overdetection rate to be around 50% (35). Although no study to date has yet reported evidence in Japan, a similar portion of incident prostate cancer after introduction of PSA screening might be explained by the overdiagnosis of asymptomatic tumors.

Furthermore, if we assume that early detection of prostate cancer and subsequent stage shift leads to a decline in mortality, the reduction in late-stage mortality around 2004 can be partly attributed to PSA screening (12). Nevertheless, the stage shift induced by screening cannot account for all of the observed changes. First, exposure to PSA screening still remains too low in Japan to have had an effect. Second, the introduction of PSA screening is unlikely to have reduced mortality so rapidly because only a small fraction of patients with early-stage cancer be predicted to die from the disease even if treated conservatively in the absence of PSA screening (36). Other factors such as earlier use of prostatectomy and antiandrogen deprivation drugs may also have played a role, as evidenced in the case of the United States (37). Third, our sensitivity analysis excluding DCO cases did not show a significant decline in distant prostate cancer after 2004, which suggests the need for cautious interpretation of the trends. Hence, the overall impact of prostate cancer control should be assessed in light of changing treatment practices and subsequent improvement in cancer survival, which warrant further investigation.

Our study has several limitations. Although the completeness and timeliness of the three high-quality cancer registries have remained relatively stable over time, the incidence of prostate cancer might have diverged from other population-based registries in Japan. However, as Katanoda and colleagues validated in their study, the incidence rates and trends estimated from these registries are representative of the overall Japanese population (11). Second, the completeness of cancer registration in these three registries has changed over time, with DCO cases fluctuating from 9.0% in 1993 to 1.5% in 2014, and DCN cases shifting from 19.0% in 1993 to 2.1% in 2014, which could yield biased estimates. However, because we have imputed the cancer stage for DCN and DCO cases using the multiple imputation approach, we consider that the influence of changes in the completeness of the registry is negligible. Furthermore, the Nagasaki registry had a large proportion of missing and unknown status compared with the Yamagata and Fukui registries. This is because the Nagasaki registry collects incidence data from pathology reports, in addition to abstraction from medical records and screening programs (38), and cases identified only from pathology reports do not include clinical stage at diagnosis. Such regional differences might yield biased estimates. Nevertheless, trends for Nagasaki and for Yamagata and Fukui combined were concordant in our sensitivity analyses.

In conclusion, our study confirmed a significantly pronounced increase in the incidence of localized prostate cancer after the spread of PSA screening in Japan. Furthermore, this increase in distant prostate cancer leveled off after 2004 and turned to a marginally declining trend. Assessing the impact of PSA screening on mortality reduction would require a comprehensive analysis, including changes in mortality, survival and treatment practices over time, and warrants further investigation.

No potential conflicts of interest were disclosed.

Conception and design: E. Saito, K. Katanoda

Development of methodology: E. Saito, D. Yoneoka

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): E. Saito, M. Hori, T. Matsuda, K. Katanoda

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): E. Saito, M. Hori, D. Yoneoka, Y. Ito, K. Katanoda

Writing, review, and/or revision of the manuscript: E. Saito, M. Hori, T. Matsuda, D. Yoneoka, Y. Ito, K. Katanoda

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): M. Hori, K. Katanoda

Study supervision: K. Katanoda

This work was supported by the Health Labour Sciences Research Grant in Japan (Cancer Control—General—016, 2018 FY).

The authors sincerely thank the members of the cancer registries of Fukui, Yamagata, and Nagasaki prefectures. We also thank Dr. Hiromi Sugiyama (Radiation Effects Research Foundation) for her valuable input.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.