Statistical methods to improve drug and vaccine safety surveillance using big healthcare data
Presenter: Scott Emerson
When: Thursday, July 14, 2016 Time: 9:00 AM - 10:30 AM
Room: Oak Bay 1-2 (Level 1)
Session Synopsis:
Post-marketing surveillance: What are we hoping for when such is a condition of approval?
In order to gain regulatory approval, a sponsor must submit evidence of a drug's safety and efficacy in a target population. Such evidence typically must include the results of "adequate and well-controlled investigations, including clinical investigations, by experts qualified by scientific training" (Kefauver-Harris Amendment, 1962). Even in the best of settings, however, the extent of the data available at the time of approval is not sufficient to ensure that no issues will arise with treatment effectiveness in practice. Hence it is not unusual that the regulatory agencies require that a sponsor conduct additional post-approval studies to investigate specific or more general concerns. In this talk I describe some of the issues that might be addressed in post-marketing surveillance primarily due to convenience, as well as some issues that cannot truly be addressed practically in any other way. I will highlight some specific criteria that should be met to best address any safety concerns that were not resolved in the registrational studies.
Statistical methods to improve drug and vaccine safety surveillance using big healthcare data
Presenter: Robert Platt
When: Thursday, July 14, 2016 Time: 9:00 AM - 10:30 AM
Room: Oak Bay 1-2 (Level 1)
Session Synopsis:
Pooling to estimate treatment effects in distributed networks while respecting privacy concerns: experience from the Canadian Network for Observational Drug Effect Studies (CNODES) network
Distributed networks such as the Canadian Network for Observational Drug Effect Studies (CNODES) and Sentinel conduct studies of drug safety and effectiveness concurrently in separate administrative datasets from different jurisdictions or insurance providers. For privacy reasons, these datasets cannot usually be pooled into a single database for analysis. Typically, a standard protocol is developed, analyses conducted separately in each database following this protocol, and summary results combined using standard meta-analytic methods. Similar issues may arise in individual-patient data meta-analysis. I will review methods for these problems, including standard meta-analytic methods, individual-patient-data meta-analytic methods, and newer distributed-regression methods that allow computation of regression estimates for the whole dataset based only on the distributed pooled sample. Finaly I will describe methods based on principles developed for pooling of biomarker samples, and show how these might be used to conduct analyses in networks like CNODES. I will demonstrate the properties of the various methods via example studies of statins, and via simulations.
Statistical methods to improve drug and vaccine safety surveillance using big healthcare data
Presenter: Andrea Cook
When: Thursday, July 14, 2016 Time: 9:00 AM - 10:30 AM
Room: Oak Bay 1-2 (Level 1)
Session Synopsis:
Group sequential monitoring for drug and vaccine safety using inverse probability of treatment weighting in FDA�s Sentinel project
Conducting observational postmarketing medical product safety surveillance is important for detecting rare adverse events not identified pre-licensure. New systems for the safety surveillance setting have been built using electronic healthcare data that keeps the individual patient data within the health plan and establishes a distributed data network to share deidentified data to answer important safety questions about new medical products. One such network is the FDA Sentinel Initiative. We will present a group sequential inverse probability of treatment method tailored to these networks to estimate a health plan stratified risk difference method that works well for rare event. To assess the performance of such methods we will conduct a simulation study comparing methods in terms of bias, power, and coverage. We will focus our simulation comparison on the rare event setting with strong across health plan/site confounding due to differential uptake of new medical products.
Statistical methods to improve drug and vaccine safety surveillance using big healthcare data
Presenter: Mohammad Karim
When: Thursday, July 14, 2016 Time: 9:00 AM - 10:30 AM
Room: Oak Bay 1-2 (Level 1)
Session Synopsis:
Comparing approaches for Confounding Adjustment in Secondary Database Analyses: High-Dimensional Propensity Score versus two machine learning algorithms: Random Forest and Elastic Net
The uses of retrospective healthcare claims datasets are frequently criticized for the lack of complete information on potential confounders. Utilizing longitudinal information about patient's health status and related information as a set of proxies of unobserved confounders, the high-dimensional propensity score (hd-PS) algorithm enables us to rank empirical covariates with respect to their potential for confounding. The propensity score model that includes only a portion of the top ranked variables has been shown to reduce bias in comparative effectiveness studies. To our knowledge only one study has compared the hd-PS with another confounder selection approach, LASSO, which uses direct adjustment for all potential confounders without prior selection. We compare hd-PS algorithm with two other popular machine learning approaches for confounder selection in high-dimensional covariate spaces: random forest and elastic net.We compare the performance of approaches using data simulated via a plasmode framework that mimicked a previously published cohort of post-myocardial infarction statin use. Performance was evaluated with respect to bias and mean squared error of the estimated treatment effect. Varying various parameters of the simulation (true association measure, confounding effect, degree of unmeasured confounding, exposure and outcome prevalence), we found that in the presence of a large number of potential confounders, estimates from the hd-PS algorithm performed slightly better compared to the estimates from both elastic net and random forest approaches. However, when using these machine learning approaches with 500 covariates selected by the hd-PS algorithm, the estimates were almost identical. In general, more bias was introduced when fewer covariates was taken into account. Across the scenarios under consideration, we found that the use of hd-PS algorithm for confounder selection is preferable to direct adjustment using random forests or elastic net for all potential confounders.
