This website provides software hosted by MD Anderson Cancer Center that enables investigators to incorporate historical data into randomized clinical trial design using Bayesian methods. Three of the web interfaces use hierarchical Bayesian modeling to facilitate “dynamic borrowing,” which enables data-dependent partial pooling of control information acquired from both current and historical studies. For these models, the extent of shrinkage towards the historical information isn’t predetermined, but rather estimated from data. The resulting dynamic Bayesian estimators borrow more strength in the absence of evidence for trial effects, thereby controlling the extent of bias induced from using the historical information. For each analysis interface, the user is given the option of specifying the number of future patients that remain to be randomized. This feature enables the software to be used to conduct adaptive randomization (AR) designs of the type proposed in (Hobbs et al. 2013), wherein the posterior effective historical sample size (EHSS) can be used to adjust the randomization ratio at one or more interim analyses. The software outputs the AR probability that should be used to randomize the next cohort of patients to the novel therapy when the design targets total information balance between treatment and control cohorts. In the absence of evidence for bias arising from inter-trial heterogeneity, the AR design can be used to allocate more newly enrolled patients to novel or lesser studied therapies. The process enables investigators to leverage all of the available information without sacrificing frequentist properties or yielding highly biased estimators of the treatment effect. Effective use of the methodology depends on one’s understanding of the models. Thus, we also provide interfaces designed to elucidate the implications of hyperparameter specification characterized by six induced frequentist properties for posterior inference. From the resultant output, users may ascertain for example the extent to which the resultant estimator of the treatment effect would be biased in the presence of low, intermediate, and high inter-trial effects; to what extent EHSS varies as a function of the true historical bias, and the domain for which the Bayesian estimator dominates the no borrowing and homogeneity estimators. Efforts to develop the software were funded by NIH-NCI grant 1R01CA15745801A1. |
|||

References: |
|||

Methods of inference for normally (Gaussian) distributed data | ||||
---|---|---|---|---|

method | link to user interface | description | last updated | |

Static Borrowing | data analysis | This interface implements posterior inference of user provided sufficient statistics for concurrent control and treatment cohorts, as well as up to 5 historical control cohorts. The software uses data-independent partial pooling of concurrent and historical information, wherein the extent to which the model borrows strength from the historical controls is pre-determined. In addition, the software can be used to balance treatment allocation in the relation to historical borrowing using the dynamic adaptive randomization (AR) design described in (Hobbs et al. 2013). | Apr. 19 2017 | |

frequentist properties | Given sufficient statistics summarizing historical control data from up to 5 studies/cohorts and an effective historical sample size (EHSS), this interface computes several induced frequentist properties for point estimation of the novel treatment effect for varying magnitudes of historical bias arising from arising between-trial effects. | Apr. 19 2017 | ||

Dynamic Borrowing using Empirical Bayes | data analysis | This interface implements posterior inference of user provided sufficient statistics for concurrent control and treatment cohorts, as well as up to 5 historical control cohorts. The software uses data-dependent partial pooling of control information acquired from concurrent and historical cohorts, wherein the extent to which the model borrows strength from the historical controls is estimated using empirical Bayesian methods. In addition, the software can be used to balance treatment allocation in the relation to historical borrowing using the dynamic adaptive randomization (AR) design described in (Hobbs et al. 2013). | Apr. 19 2017 | |

frequentist properties | Given sufficient statistics summarizing historical control data from up to 5 studies/cohorts and an upper bound for the effective historical sample size (EHSS), this interface computes several induced frequentist properties for point estimation of the novel treatment effect and illustrates the extent to which the model borrows strength from the historical information in the presence of varying magnitudes of historical bias arising from between-trial effects. | Apr. 19 2017 | ||

Dynamic Borrowing using Spike and Slab | data analysis | This interface implements posterior inference of user provided sufficient statistics for concurrent control and treatment cohorts, as well as up to 5 historical control cohorts. The software uses data-dependent partial pooling of control information acquired from concurrent and historical cohorts, wherein the extent to which the model borrows strength from the historical controls is estimated using a spike and slab prior. In addition, the software can be used to balance treatment allocation in the relation to the extent of historical borrowing using the dynamic adaptive randomization (AR) design described in (Hobbs et al. 2013). | June 1 2017 | |

frequentist properties | Given sufficient statistics summarizing historical control data from up to 5 studies/cohorts and a set of spike and slab hyperparameters specifying the prior distribution assumed for the effective historical sample size, this interface computes the induced frequentist properties for point estimation of the novel treatment effect when borrowing strength dynamically using the spike and slab method (Hobbs et al. 2012). The interface is intended to assist users in selecting hyperparameters for the spike and slab prior, elucidate implications of various choices for dynamic borrowing, and compare bias/variance tradeoffs for estimation of the novel treatment effect among candidate specifications in the presence of the actual historical data. | June 1 2017 | ||

Methods of inference for right-censored time-to-event data | ||||

method | link to user interface | description | last updated | |

Dynamic Borrowing using piecewise constant hazard | data analysis | This web interface implements posterior inference of user uploaded data for comparing a novel therapy to a previous studied control using the time-to-event piecewise exponential Bayesian model described in Hobbs et al. (2013). The model borrows strength dynamically from the partially informative historical control data using a hierarchical model with spike and slab hyperprior. The software can be used at interim analyses of an ongoing trial to balance treatment allocation in relation to the extent of historical borrowing using the dynamic adaptive randomization (AR) design described in (Hobbs et al. 2013) using the provided AR probability. | Apr. 19 2017 | |

© 2014 Department of Biostatistics (PID 706) The University of Texas MD Anderson
Cancer Center