Talisa, Victor
(2021)
Novel statistical approaches for practical problems in individualized medicine.
Doctoral Dissertation, University of Pittsburgh.
(Unpublished)
This is the latest version of this item.
Abstract
Conditions such as sepsis are clinically and biologically heterogeneous, often deadly, and yet lack demonstrably effective treatments. In these cases, learning patient-level conditional average treatment effects (CATEs) using methodological and computational advancements could improve trial execution and delivery of care. In this dissertation, I develop methodologies addressing two practical challenges in individualized medicine. The first stems from the existing regulatory procedure for drug development programs. Regulatory agencies usually require two successful demonstrations of efficacy trials before granting approval. One way to improve the likelihood of follow-up success is to identify a subpopulation associated with a treatment benefit, and enroll future studies from this subpopulation. In part 1 of this dissertation, I define Confirmable Responder Class (CRC) learning, where the objective is to find a subpopulation in which the predicted conditional average treatment effect (CATE) suggests benefit, with high power of being confirmed empirically in an independent sample, and strong type 1 error control. I define a set of CRC learners and study their performance across simulated trial datasets. Simulations show that performance depends on population characteristics, and that confirmation rates can be improved by increasing the use of all data in relatively simple ways. A method based on cross-validation is a possible avenue for further development of model selection procedures. In part 2 I address the challenge of estimating CATEs when some covariate information is missing. I design and study 6 new algorithms for estimation and inference of the CATEs in this setting. I describe novel approaches that respond adaptively to missingness in the covariate vector, which could contain elements with different marginal missingness mechanisms. I use simulated datasets to show that the best-performing algorithm is a Bayesian Causal Forest (BCF) model expanded to include a prior on missing data elements, in combination with the Missingness Incorporated in Attributes (MIA) decision tree splitting approach.
Public Health Significance: I develop tools that could be used for effective subgroup identification in the context of missing covariate data. Identifying treatment responders could in turn advance development of tailored treatment approaches for a variety of public health and medical conditions without effective treatments.
Share
| Citation/Export: |
|
| Social Networking: |
|
Details
| Item Type: |
University of Pittsburgh ETD
|
| Status: |
Unpublished |
| Creators/Authors: |
|
| ETD Committee: |
|
| Date: |
19 January 2021 |
| Date Type: |
Publication |
| Defense Date: |
20 November 2020 |
| Approval Date: |
19 January 2021 |
| Submission Date: |
10 December 2020 |
| Access Restriction: |
1 year -- Restrict access to University of Pittsburgh for a period of 1 year. |
| Number of Pages: |
102 |
| Institution: |
University of Pittsburgh |
| Schools and Programs: |
School of Public Health > Biostatistics |
| Degree: |
PhD - Doctor of Philosophy |
| Thesis Type: |
Doctoral Dissertation |
| Refereed: |
Yes |
| Uncontrolled Keywords: |
Individualized medicine, subgroup identification, conditional average treatment effect, missing data, biostatistics |
| Date Deposited: |
19 Jan 2021 20:01 |
| Last Modified: |
19 Jan 2022 06:15 |
| URI: |
http://d-scholarship.pitt.edu/id/eprint/40063 |
Available Versions of this Item
-
Novel statistical approaches for practical problems in individualized medicine. (deposited 19 Jan 2021 20:01)
[Currently Displayed]
Metrics
Monthly Views for the past 3 years
Plum Analytics
Actions (login required)
 |
View Item |
![[feed]](/images/feed-icon-32x32.png){kind=icon}