An efficient Bayesian platform trial design for borrowing adaptively from historical control data in lymphoma.

To reduce a clinical trial's cost and ethical risk to its enrollees, some oncology trial designers have suggested borrowing information from similar but already completed trials to reduce the number of patients needed for the current study. Motivated by competing drug therapies for lymphoma, we propose a Bayesian adaptive "platform" trial design that uses commensurate prior methods at interim analyses to borrow adaptively from the control group of an earlier-starting trial. The design adjusts the trial's randomization ratio in favor of the novel treatment when the interim posterior indicates commensurability of the two control groups. In this setting, our design can supplement a control arm with historical data, and randomize more new patients to the novel treatments. This design is both ethical and economical, since it shortens the process of introducing new treatments into the market, and any additional costs introduced by this design will be compensated by the savings in control arm sizes. Our approach performs well via simulation across settings with varying degrees of commensurability and true treatment effects, and compares favorably to an adaptive "all-or-nothing" approach in which the decision to pool or discard historical controls is based on a simple ad-hoc frequentist test at interim analysis. We also consider a three drug extension where a new imaginary intervention joins the platform, and show again that our procedure performs well via simulation.

Multilevel spatio-temporal dual changepoint models for relating alcohol outlet destruction and changes in neighbourhood rates of assaultive violence.

Many previous studies have suggested a link between alcohol outlets and assaultive violence. In this paper, we evaluate the impact of the "1992 Civil Unrest" in Los Angeles (which followed the "Rodney King incident"), in which many alcohol outlets were damaged leading to a decrease in alcohol outlet density, on crime. We leverage the natural experiment created by the closure of alcohol outlets in certain areas and not others to explore the effects of alcohol availability on assault crimes at the census tract level. We develop a hierarchical model that controls for important covariates (such as race and socio-economic status) and accounts for unexplained spatial and temporal variability. While our model is somewhat complex, its hierarchical Bayesian analysis is accessible via the WinBUGS software. Our results show that, keeping other effects fixed, the reduction in alcohol availability within a census tract was associated with a drop in the assaultive violence rate at the census tract level. Comparing several dual candidate changepoint models using the Deviance Information Criterion, the drop in assaultive violence rate is best seen as having occurred one year after the reduction in alcohol availability, with the effect lasting roughly five years. We also create maps of the fitted assault rates in Los Angeles, as well as spatial residual maps that suggest various spatially-varying covariates are still missing from our model.