Comparing effectiveness of medical interventions, for example drug therapies of surgical procedures, is an ongoing undertaking for medical scientists. As the patient response to any intervention varies from individual to individual, statistical methodology is necessary for any comparison. An ideal tool for this purpose is the controlled clinical trial with patients randomized to treatment. The randomization creates a balance in the many unmeasured factors beyond the treatment that cause variation in patient response. It also allows the calculation, under competing hypotheses of comparative effectiveness, of expectations and probabilities on which decisions can be based. Typically, however, patients enrolled in a clinical trial are drawn from and represent only a small segment of the target population for the intervention. To measure the effectiveness of any intervention in the unselective population at large, it is possible to use other excellent sources of relevant information such cancer registries at national and state levels. However, patients in these databases have chosen their treatment, as opposed to being externally and randomly assigned to it. It is well known that analyzing observational data with techniques designed for clinical trials leads to biased estimation of comparative effectiveness. Two general approaches have been developed for use in observational studies: propensity score matching and employing instrumental variables. This project aims to develop, for the second approach, new model-based methodology with minimal distributional assumptions in two outcome categories that are widely used in comparative effectiveness of medical interventions. These are: a binary outcome such as post-surgical infection or 30-day mortality, and a time-to-event outcome such as survival time after cancer diagnosis or disease recurrence time.
The specific aims of the study are: 1. Develop inference methodology, including computational algorithms, for logistic and probit regression for binary outcomes, using the Bayesian nonparametric approach for instrumental variables. 2. Develop inference methodology, including computational algorithms, for time-to-event outcomes, using the Bayesian nonparametric approach for Cox proportional hazards regression with endogenous regressors predictable by instrumental variables. 3. Using repeated data simulation, compare performance of the methods developed in specific aims 1 and 2 with currently available linear asymptotic methods. As a result of the proposed work, methods for assessing comparative effectiveness of medical interventions from observational data will be improved substantially. The methods will apply to most diseases. Currently, many studies are conducted for various cancers, cardiovascular diseases and geriatric conditions.
To improve the health of the nation, it is important to find out which drugs and medical treatments work better. State and national registries, and other large databases, contain much information that can be used for this purpose. This project will substantially improve currently available scientific methods of making conclusions from this information.
| Sparapani, Rodney A; Logan, Brent R; McCulloch, Robert E et al. (2016) Nonparametric survival analysis using Bayesian Additive Regression Trees (BART). Stat Med 35:2741-53 |
