Radical Prostatectomy vs. Radiotherapy in Prostate Cancer
Radical Prostatectomy vs. Radiotherapy in Prostate Cancer
This study is based on the PCBaSe Sweden, which has been described previously. Briefly, it is a composite population based dataset of the National Prostate Cancer Registry of Sweden, the Swedish cancer register, the cause of death register, and six other national registers, using the unique 10 digit personal identity number assigned to every resident in Sweden. The dataset covers 98% of all cases of prostate cancer in Sweden diagnosed since 1998 (with coverage from 1996 and 1997 limited to certain regions), and has virtually complete data on year of diagnosis; age; clinical stage (tumour, node, metastases (TNM) classification); tumour grade (either Gleason sum or World Health Organization grade of differentiation); serum level of prostate specific antigen at the time of diagnosis; planned primary treatment within six months of diagnosis; county of residence; marital status; educational level; socioeconomic status; Charlson comorbidity index; and cancer related events during follow-up. The Charlson score was estimated from registrations in the inpatient register, which in a previous study based on the PCBaSe dataset has been shown to have an impact on management and survival.
We identified a total of 109 333 men with a diagnosis of prostate cancer between 1996 and 2010 in PCBaSe Sweden. After exclusion of those whose treatment was unknown (n=4788) or who had died before treatment (n=512), we included all patients provided their primary treatment was listed as radical prostatectomy or radiotherapy (n=34 515). From this analysis we excluded patients who received androgen deprivation or surgical castration as their primary treatment (n=40 502), or watchful waiting (n=29 016). The median follow-up time for the included cohort was 5.37 years (interquartile range 3.00-7.81 years), for the radical prostatectomy group 5.26 (3.03-7.57) years, and for the radiotherapy group 5.60 (2.96-8.18) years. As in previous studies using this dataset, we categorised patients by clinical risk (low, intermediate, or high risk (collectively, non-metastatic prostate cancer), and metastatic disease, Table 1 ), as well as by age (≤64, ≥65) and Charlson comorbidity index (0, ≥1). After stratification by risk group, the study cohort comprised 34 052 cases; 463/34 515 (1.3%) patients had missing data precluding risk categorisation. We merged WHO grade 1 tumours with Gleason scores 2-6, WHO 2 with Gleason score 7, and WHO 3 with Gleason scores 8-10.
The primary outcome of interest was death from prostate cancer. We defined survival time as the interval between date of diagnosis of prostate cancer and the date of death, emigration, or end of follow-up at 31 December 2010.
We used χ and Wilcoxon-Mann-Whitney tests to investigate differences in the distributions of patient characteristics by treatment groups. To visualise cause specific mortality, we plotted cumulative incidence curves for the treatment groups. We investigated differences in each cause of mortality (prostate cancer or other causes) using subdistribution hazard ratios estimated through Fine and Gray proportional hazards regression. To deal with any imbalances in the distribution of covariates among treatment groups, we produced both traditional multivariable model adjusted and propensity score adjusted estimates of subdistribution hazard ratios. We calculated propensity scores using logistic regression, with treatment group as outcome and all adjustment covariates as predictors; we made adjustments by including the resulting logit transformed propensities when modelling subdistribution hazard ratios. We tested heterogeneity of such ratios across risk groups using likelihood ratio tests of the relevant interaction terms. Furthermore, we used the propensity scores for matching, which we carried out within each risk group, using the larger of the two treatment groups and selecting a nearest neighbour 1-to-1 match for those in the smaller treatment group, with a caliper of 0.1 standard deviations for the propensity scores.
We carried out several sensitivity analyses. Assuming that the cancer treatment resulted in no difference in prostate cancer mortality, we assessed the effect required of a hypothetical unmeasured binary confounder to explain the propensity score adjusted subdistribution hazard ratios for prostate cancer mortality of radiotherapy versus radical prostatectomy for varying levels of confounder imbalance between treatment groups. To investigate possibly divergent developments in treatment efficiency we reassessed the comparisons after stratification by year of diagnosis (1996-99, 2000-04, and 2005-09). Furthermore, we used the propensity scores for inverse probability of treatment weight adjustments; we carried out these through weighting data from each individual with weights proportional to the estimated propensity of the treatment not received, using Cox proportional hazards regression.
All tests were performed two sided at the 5% significance level. Statistical analyses were performed with IBM SPSS Statistics (version 20.0, IBM, Armonk, NY), and R software (version 2.15, R Foundation for Statistical Computing, Vienna, Austria) using the cmprsk, survival, rmeta, and Matching packages.
Methods
This study is based on the PCBaSe Sweden, which has been described previously. Briefly, it is a composite population based dataset of the National Prostate Cancer Registry of Sweden, the Swedish cancer register, the cause of death register, and six other national registers, using the unique 10 digit personal identity number assigned to every resident in Sweden. The dataset covers 98% of all cases of prostate cancer in Sweden diagnosed since 1998 (with coverage from 1996 and 1997 limited to certain regions), and has virtually complete data on year of diagnosis; age; clinical stage (tumour, node, metastases (TNM) classification); tumour grade (either Gleason sum or World Health Organization grade of differentiation); serum level of prostate specific antigen at the time of diagnosis; planned primary treatment within six months of diagnosis; county of residence; marital status; educational level; socioeconomic status; Charlson comorbidity index; and cancer related events during follow-up. The Charlson score was estimated from registrations in the inpatient register, which in a previous study based on the PCBaSe dataset has been shown to have an impact on management and survival.
We identified a total of 109 333 men with a diagnosis of prostate cancer between 1996 and 2010 in PCBaSe Sweden. After exclusion of those whose treatment was unknown (n=4788) or who had died before treatment (n=512), we included all patients provided their primary treatment was listed as radical prostatectomy or radiotherapy (n=34 515). From this analysis we excluded patients who received androgen deprivation or surgical castration as their primary treatment (n=40 502), or watchful waiting (n=29 016). The median follow-up time for the included cohort was 5.37 years (interquartile range 3.00-7.81 years), for the radical prostatectomy group 5.26 (3.03-7.57) years, and for the radiotherapy group 5.60 (2.96-8.18) years. As in previous studies using this dataset, we categorised patients by clinical risk (low, intermediate, or high risk (collectively, non-metastatic prostate cancer), and metastatic disease, Table 1 ), as well as by age (≤64, ≥65) and Charlson comorbidity index (0, ≥1). After stratification by risk group, the study cohort comprised 34 052 cases; 463/34 515 (1.3%) patients had missing data precluding risk categorisation. We merged WHO grade 1 tumours with Gleason scores 2-6, WHO 2 with Gleason score 7, and WHO 3 with Gleason scores 8-10.
The primary outcome of interest was death from prostate cancer. We defined survival time as the interval between date of diagnosis of prostate cancer and the date of death, emigration, or end of follow-up at 31 December 2010.
Statistical Analysis
We used χ and Wilcoxon-Mann-Whitney tests to investigate differences in the distributions of patient characteristics by treatment groups. To visualise cause specific mortality, we plotted cumulative incidence curves for the treatment groups. We investigated differences in each cause of mortality (prostate cancer or other causes) using subdistribution hazard ratios estimated through Fine and Gray proportional hazards regression. To deal with any imbalances in the distribution of covariates among treatment groups, we produced both traditional multivariable model adjusted and propensity score adjusted estimates of subdistribution hazard ratios. We calculated propensity scores using logistic regression, with treatment group as outcome and all adjustment covariates as predictors; we made adjustments by including the resulting logit transformed propensities when modelling subdistribution hazard ratios. We tested heterogeneity of such ratios across risk groups using likelihood ratio tests of the relevant interaction terms. Furthermore, we used the propensity scores for matching, which we carried out within each risk group, using the larger of the two treatment groups and selecting a nearest neighbour 1-to-1 match for those in the smaller treatment group, with a caliper of 0.1 standard deviations for the propensity scores.
We carried out several sensitivity analyses. Assuming that the cancer treatment resulted in no difference in prostate cancer mortality, we assessed the effect required of a hypothetical unmeasured binary confounder to explain the propensity score adjusted subdistribution hazard ratios for prostate cancer mortality of radiotherapy versus radical prostatectomy for varying levels of confounder imbalance between treatment groups. To investigate possibly divergent developments in treatment efficiency we reassessed the comparisons after stratification by year of diagnosis (1996-99, 2000-04, and 2005-09). Furthermore, we used the propensity scores for inverse probability of treatment weight adjustments; we carried out these through weighting data from each individual with weights proportional to the estimated propensity of the treatment not received, using Cox proportional hazards regression.
All tests were performed two sided at the 5% significance level. Statistical analyses were performed with IBM SPSS Statistics (version 20.0, IBM, Armonk, NY), and R software (version 2.15, R Foundation for Statistical Computing, Vienna, Austria) using the cmprsk, survival, rmeta, and Matching packages.