OBJECTIVES: Randomized controlled trials (RCTs) are the gold standard for clinical evidence and are routinely incorporated into health economic models. However, there are practical and scientific reasons for why observational studies may be preferred to inform model extrapolation in health technology assessment (Zhao, 2016; Ligthelm, 2007; MacLehose, 2000; Ioannidis, 2005; Concato, 2000). This study explored a Bayesian framework for combining information from RCTs and real-world data (RWD) in health economic models, allowing for flexible weighting of the two data sources via prior information.

METHODS: Bayesian exponential survival models were fit to overall and progression-free survival data from a real-world cohort of alectinib- or crizotinib-treated ALK-positive (ALK+) advanced non-small cell lung cancer patients selected from the Flatiron Health electronic health record-derived deidentified nationwide longitudinal database. Posterior mean and 95% credible interval (CrI) estimates were generated for hazard ratios (HRs) and incremental cost-effectiveness ratios (ICERs) comparing alectinib versus crizotinib under different weightings of prior information from RCTs and different cut points of RWD accumulation.

RESULTS: The study included 463 ALK+ patients who were treated with alectinib or crizotinib. Estimated HRs ranged from 0.63 (95% CrI: 0.46-0.85) to 0.71 (0.61-0.83) using uninformative (i.e. not informed by RCT data, only by RWD) to strongly informative priors, while estimated ICERs were ~8% higher when using strongly informative versus uninformative priors. When using strong prior information, model estimates stabilized after 80 alectinib patients were accumulated from RWD. In contrast, when using an uninformative prior, model estimates stabilized after 183 alectinib patients were accumulated from RWD.

CONCLUSIONS: A Bayesian estimation framework allows for controlling the level of borrowing between RCT and RWD according to one’s prior belief. If sample size from RWD is small, adding prior information from RCTs can provide more stable estimates, whereas prior assumptions may be relaxed as more RWD is accumulated.