This presentation is Mrs. Mishra-Kalyani’s dissertation defense
Advisors: Qi Long, Ph.D. and Brent Johnson, Ph.D.

 
Abstract:
Observational studies, such as those of patient registries may offer valuable patient and disease information that is impossible to study in a randomized trial, but often pose unique challenges that require special care in estimating a causal effect of treatment. Registry data in particular may be incomplete, inconsistent, and subject to selection bias or confounding. This dissertation is motivated by a registry of patients with amyotrophic lateral sclerosis (ALS) maintained by the Emory ALS Clinic, in which the receipt of the treatment, that is, the insertion of a Percutaneous Endoscopic Gastrostomy (PEG) tube, is time-dependent and both the receipt of treatment and clinical outcomes are subject to “censoring” by death. The analysis of the ALS registry is further complicated by non-random treatment assignment, time-dependent covariates, and missing data. In order to identify a causal effect of PEG treatment on outcomes such as short term quality of life or overall survival, we incorporate and build upon various causal inference methods such as principal stratification and propensity score matching

After a review of current literature and a more detailed description of the data in Chapter 1, we aim to delineate a statistical framework to assess the effect of surgical insertion of a percutaneous endogastrostomy (PEG) tube for patients of the Emory ALS registry in Chapter 2. Although all ALS patients are informed about PEG, some agree to PEG while others decline which leads to the potential for selection bias. Assessing the effect of PEG is further complicated due to the fact that ALS can be aggressive and some patients die shortly after diagnosis. As a result, time to death competes directly with both the opportunity to receive PEG as well as any clinical outcome measured after PEG insertion is complete. In this paper, we address the “censoring by death” phenomenon through principal stratification and potential selection bias for PEG treatment through generalized propensity scores. We develop a fully Bayesian modeling approach to estimate the survivor average causal effect (SACE) of PEG on BMI, which is a surrogate outcome measure of nutrition and quality of life. The use of propensity score methods within the principal stratification framework allows for a significant and positive effect of PEG treatment, particularly when time of treatment is included in the treatment definition. In addition, the proposed approach is shown in simulation studies to outperform the method that does not correct for selection bias.

Chapter 3 investigates propensity function matching for estimating treatment effect in observational studies. In observational studies where the treatment can be given at any time during follow-up and there is no structure or intervention to ensure regular clinic visits for data measurement. As a result, complex issues of selection bias and confounding are of paramount concern. Although some propensity score methods, specifically the Generalized Propensity Score of Hirano and Imbens (2004) and Propensity Function of Imai and van Dyk (2004), are able to include time-varying covariates in the treatment assignment model but cannot balance treatment assignment conditional on an entire covariate process. This balancing property is particularly important when time-varying covariates are measured at uneven or inconsistent time intervals that do not coincide with the time of treatment. In this paper we propose the Propensity Process as a method that is able to address these complex features that are common to observational registries with longitudinally measured data. We apply the Propensity Process and standard non-binary propensity score methods to the analysis of the ALS study. Matching by Propensity Process outperforms the naive analysis and other non-binary propensity score methods and achieves covariate balance across treatment groups. Additionally, after matching by Propensity Process, treatment has a significant positive association with change in body mass index from baseline until 24 months. In contrast, the naive analysis and matching by other non-binary propensity score methods result in a negative or non-significant treatment effect.

When observation times are inconsistent within a study, the causal effect of treatment on outcome must take into account the potential for missing outcomes. Chapter 4 extends the methods presented in Chapter 1 to address outcomes that are missing due to a lapse in clinic visits. A single framework is used to address missing outcome data, censoring by death and selection bias. The basis of the framework is principal stratification based on post-treatment survival outcomes, and a model for the mechanism of missing outcomes and generalized propensity score are incorporated to appropriately address the aforementioned issues and to achieve an unbiased estimation of treatment effect. This approach is explored in the data analysis of the Emory ALS Clinic data and is compared to modeling frameworks with more stringent assumptions about the missing mechanism, such as missing completely at random or latent ignorability.

Finally, potential future work is explored in Chapter 5. The data of the ALS registry is rich with complications that could inspire new directions of research, and there is significant interest in the issues of observational studies in the statistical community to fuel this methodological research.
..

Parking available in the Michael Street Visitor parking deck (behind Wayne Rollins Research Building...2nd deck entrance) or at the 1525 Clifton Road Visitor pay parking deck (building directly across the street from Grace Crum Rollins Building). Please visit our webpage at:  http://www.sph.emory.edu/departments_centers/bios/index.html
Questions:  rwaggon@emory.edu