BAYESIAN NETWORK META-ANALYSIS OF DCIS RECURRENCE AFTER BREAST CONSERVING SURGERY: A TIME DEPENDANT HAZARD APPROACH
Candidate for the Lee B. Lusted Student Prize Competition
Purpose: To synthesize evidence on the potentially time-dependent hazards of ductal carcinoma in situ (DCIS) recurrence after breast conserving surgery plus radiation therapy (BCS+RT) and BCS alone for inclusion in a decision-analysis.
Methods: We abstracted time to recurrence directly from Kaplan-Meier curves in five published RCTs involving BCS+RT and BCS alone. We used Bayesian random-effects meta-analysis to estimate posterior distributions of survival model parameters under both an exponential model (assuming constant hazards) and a Weibull model (allowing for time-dependent hazards), following the approach of Ouwens, Phillips and Jansen (Res Synthesis Meth 2010; 1:258-271). Base case time-to-event curves were constructed from posterior median parameters and compared between the models over a 30 year horizon.
Results: Estimated posterior medians of the Weibull shape parameters suggested decreasing hazards of DCIS recurrence under BCS+RT (Treatment) and BCS alone (Control) (see Figure). Comparing cumulative incidence over 5 years, the Weibull model predicted more early recurrences but fewer recurrences overall by year 30 when compared to the exponential model. The magnitude of the estimated differences comparing the Weibull model to the exponential model was substantial: ranging from 5.3% more recurrences at year 3 to 8.3% fewer recurrences at year 30 with BCS alone, and from 1.8% more recurrences at year 3 to 5.1% fewer recurrences at year 30 with BCS+RT. The differences between models differed between BCS+RT (Treatment) and BCS alone (Control), with the Weibull model implying more prevented recurrences from the addition of RT to BCS in the first 12 years, but fewer prevented recurrences by year 30.
Conclusion: Comparing the exponential and Weibull model approaches to meta-analysis of DCIS recurrence, we found that model choice led to substantial differences in predicted recurrence: increasing the comparative effectiveness of BCS+RT relative to BCS alone over the short run, but decreasing it over the long run. Differing base-case parameter estimates (and posterior distributions to be used in probabilistic sensitivity analysis) are likely to produce substantially different results when combined with other factors in a broader decision analytic model. The Weibull approach makes use of the entire survival curve, but requires that the survival curve be available. The simpler exponential approach requires only counts of events, but is unable to identify potentially consequential time varying hazards.
