Propensity score matched augmented controls in randomized clinical trials: A case study

Abstract

Existing statutes in the United States and Europe require manufacturers to demonstrate evidence of effectiveness through the conduct of adequate and well-controlled studies to obtain marketing approval of a therapeutic product. What constitutes adequate and well-controlled studies is usually interpreted as randomized controlled trials (RCTs). However, these trials are sometimes unfeasible because of their size, duration, cost, patient preference, or in some cases, ethical concerns. For example, RCTs may not be fully powered in rare diseases or in infections caused by multidrug resistant pathogens because of the low number of enrollable patients. In this case, data available from external controls (including historical controls and observational studies or data registries) can complement information provided by RCT. Propensity score matching methods can be used to select or "borrow" additional patients from the external controls, for maintaining a one-to-one randomization between the treatment arm and active control, by matching the new treatment and control units based on a set of measured covariates, ie, model-based pairing of treatment and control units that are similar in terms of their observable pretreatment characteristics. To this end, 2 matching schemes based on propensity scores are explored and applied to a real clinical data example with the objective of using historical or external observations to augment data in a trial where the randomization is disproportionate or asymmetric.

Historical Controls in Randomized Clinical Trials: Opportunities and Challenges

“…The use of concurrent controls may provide some assurance by providing a point of reference, but the extent of allocation of matching to historic or concurrent controls is important. If the number of patients borrowed is very large, then the prior information may be too informative 81 . Various propensity matching schemes exist and involve either pair‐matching or caliper‐matching combinations of historic controls, concurrent controls, and the experimental arm 3,81,83 …”

“…An inherent drawback of such matching is that, whereas measured confounders may be balanced, unmeasured factors, including those related to time frame, definitions, or other design characteristics, may not be well‐controlled. The historical controls need to have the same eligibility, treatment, and evaluation as the randomized controls, with no unexplained indications, and all important prognostic factors must be measured in the same fashion, as first suggested by Pocock 2,81 . Many of the differences in patient‐level variables may be controlled through use of covariates.…”

“…If comparable, pool the data Propensity score match Match HC to RC and/or AT across known characteristics with propensity score. 3,81,83 Bayesian approaches…”

Incorporating Innovative Techniques Toward Extrapolation and Efficient Pediatric Drug Development

“…Alternatively, the informative prior can be defined proactively to come from specific adult populations that more closely resemble the pediatric population, for example, younger adults and adults with lower body weight or similar disease severity that are known to modify responses to treatments. 38 Another important consideration with this approach is whether there was sufficient quality in how the data were collected and how patients were assigned to treatments. The higher the quality, the more likely extrapolation will be appropriate for regulatory submissions.…”

Propensity‐score‐based priors for Bayesian augmented control design

“…In this paper, we use the propensity scoring technique to construct informative prior information for augmenting active control data in limited population antibacterial drug development against MDR pathogens that uses disproportionate randomization (see also Lin et al). We will explore strategies on how historical data, eg, registries or patient level data, collected from previous randomized clinical trials can be used effectively in the context of clinical trials to form a basis for drug approval, ie, how can the active control arm data be augmented with matched observations from existing historical data from the same control.…”