Study design
We conducted a retrospective cohort analysis using de-identified data accessed from the NCDB. The NCDB is a joint program established in 1989 by the Commission on Cancer of the American College of Surgeons and the American Cancer Society (19). This comprehensive data set integrates registry records from more than 1500 accredited hospitals, capturing approximately 70% of all incident cancers in the United States (20). According to the agreements executed with each accredited facility, data from Veteran Affairs, Department of Defense, Puerto Rican, and certain other programs are removed from research files. The accreditation requires an annual 90% follow-up rate for all eligible patients diagnosed within five years. Survival outcomes are released only after at least five years of follow-up to avoid censoring bias.
Data are coded using standardized algorithms, and duplicate records are eliminated. Variables include patient demographics, comorbidities, socioeconomic status, and the first course of therapy, defined as all treatment methods recorded in the treatment plan and administered to the patient before disease progression or recurrence. Treatments delivered or withheld because of progression, insufficient response, or other therapy modifications caused by restaging or intercurrent events are not recorded. Specific chemotherapy regimens, doses, or treatment durations are not recorded. Since this study used de-identified data, it was considered exempt from human protection oversight by the Allegheny Health Network institutional review board.
The NCDB provided records of 16,579 patients diagnosed with PCNSL between 2004 and 2015. Cases were identified using primary anatomical site codes C70.0, C70.1, C70.9-C72.1, C72.3-C72.5, C72.8, and C72.9, including the brain, spinal cord, cranial nerves, and meninges. We excluded patients younger than 65 and patients without histologic or cytologic confirmation of the diagnosis. We also excluded patients with HIV positive or unknown status, those with unknown status regarding chemotherapy or radiation administration. We excluded patients treated outside the reporting facility because, otherwise, the NCDB does not require documentation of their treatment and outcomes. We also excluded patients who started treatment with radiation or chemotherapy >120 days after diagnosis or started radiation > 365 days from diagnosis in the combined modality group to account for immortal time bias (Figure 1: Selection process, CONSORT diagram). Treatment was categorized into four groups: chemotherapy alone, radiation alone, combined modality treatment, and no treatment.
Variables and study outcomes
Race was recoded into four categories – non-Hispanic whites, non-Hispanic blacks, Hispanics, and others. Comorbidity was captured using the Charlson/Deyo comorbidity index (21). Socioeconomic data were provided as quintiles of median household income and number of persons with less than high school education in patients’ census tract of residence. The type of facility was assigned according to the Commission on Cancer accreditation category based on annual case volume and available oncology services. Geographic locations corresponded to the U.S. Census Divisions. Insurance status is captured as it appears on the admission face sheet for the patient.
The primary outcome of the study is the overall survival (OS) in the four treatment groups. We estimated median OS as well as 12- and 24-month OS. Overall survival was defined from the time of diagnosis to the time of death. We determined predictors of receiving any treatment modality compared to no treatment and predictors of receiving CMT compared to chemotherapy alone among treated patients. We also calculated the annual percentage change (APC) for CMT to assess its trend over the study period.
Propensity score estimation
In this study, we obtained the average treatment effect (ATE) defined as an estimate of interest. Variables included in the propensity score model included age, sex, race, insurance status, median income, education, treatment facility type, type of area, comorbidity score, and distance from the treating facility.
Inverse probability of treatment weighting was used in estimating weight-adjusted OS in the four treatment groups. We used four different methods to estimate weights: multinomial regression propensity score, generalized boosted propensity score, covariate balancing propensity score, and entropy balancing method. The choice of weighting method was based on achieving small coefficients of variations, large effective sample sizes, and low covariate balance assessed using standardized bias. A standardized bias-cut off less than 0.25 was used (22). We used balance tables and love plots to assess for covariate balance before and after weighting (Figure 2). Generalized boosted modeling was used as a final method to estimate weights based on the above criteria. The absolute standardized mean difference among covariates was used as a balance criterion. We used the propensity score package “WeightIt” coded for R statistical program (23). We used a robust variance estimator to account for within-person homogeneity (24).
Statistical analysis
Descriptive statistics were used to compare baseline characteristics of the four treatment groups. Continuous variables were presented as mean with standard deviation or median with interquartile range (IQR). Categorical variables were presented as absolute numbers and percentages. The means of continuous variables were compared using t-test or ANOVA, and percentages were compared using Pearson chi-square or Fisher’s exact test. Overall survival was compared using log-rank and GehanBreslow-Wilcoxon rank tests. A stratified log-rank test for every three years was used to account for possible variation in available and administered treatments. Univariable logistic regression was used to determine predictors of receiving any treatment versus no treatment and of receiving CMT versus single modality treatment. Those predictors were expressed as odds ratio (OR) and 95% confidence interval (CI). Statistically significant variables on univariable analysis were used to build multivariable logistic regression models.
To account for missing data, we created five multiply-imputed lists using the “mi” package. (25) All fives imputed data sets were analyzed, and the OS estimate was combined using the Rubin procedure (26). Regression diagnostics were used to evaluate model assumptions. All statistical tests were two-sided, and P-values <0.05 were considered statistically significant. We used R-statistical software (version 4.0.3) for statistical analysis (27).