Zhong, Yongqi
(2021)
Developing and Evaluating Innovative Approaches for Estimating Causal Effects of Low-dose Aspirin on Pregnancy Outcomes.
Doctoral Dissertation, University of Pittsburgh.
(Unpublished)
Abstract
First trimester pregnancy loss occurs in one third of all pregnancies, and recurrent pregnancy loss is also prevalent in up to 30% of women with a prior history. Using intention-to-treat, the Effects of Aspirin in Gestation and Reproduction (EAGeR) trial found that low-dose aspirin (LDA) led to 4.3 (95% CI -1.2 to 9.6) per 100 women at high risk of pregnancy loss. However, the estimated effect, which is based on the assignment to a treatment arm, rather than adherence to a particular treatment protocol, limits the understanding of potential benefits of LDA on pregnancy.
Existing methods for adherence adjustment to estimate per-protocol effects in randomized trials are subject to the limitations of observational studies, including model mis-specification due to incorrect confounder selection, or from strong parametric assumptions.
The objective of this dissertation is to evaluate and develop innovative approaches for estimating the adherence-adjusted effects of LDA on pregnancy. First, to mitigate the impact of incorrect confounder selection, we evaluated the performance of causal discovery methods in a simulation study using the data resampled from EAGeR. We found that, the evaluated causal discovery method yielded low accuracy in selecting sufficient confounder adjustment sets in the M- or Butterfly-structured causal diagrams. Second, to avoid strong parametric assumptions, we developed an R package implementing the augmented inverse probability weighting (AIPW), a doubly robust estimator supporting stacking machine learning. Our simulation study suggests that, our AIPW package has excellent performance compared to existing R packages implementing doubly robust estimators. Finally, we used the AIPW package with stacking machine learning to estimate per-protocol effects of LDA in a time-fixed setting from the EAGeR trial. Our results show that LDA led to 8.0 (95% CI 2.5 to 13.6) more pregnancies per 100 women who adhered to the randomized treatment assignment for at least 5/7 days per week over at least 80% person-week of follow-up, consistent with the previous analysis using parametric g-formula in a time-varying setting. In conclusion, this dissertation does not only provide additional evidence of the benefits of LDA on pregnancy, but also the state-of-the-art approaches for effect estimations in epidemiologic studies.
Share
| Citation/Export: |
|
| Social Networking: |
|
Details
| Item Type: |
University of Pittsburgh ETD
|
| Status: |
Unpublished |
| Creators/Authors: |
|
| ETD Committee: |
|
| Date: |
27 August 2021 |
| Date Type: |
Publication |
| Defense Date: |
1 July 2021 |
| Approval Date: |
27 August 2021 |
| Submission Date: |
16 July 2021 |
| Access Restriction: |
1 year -- Restrict access to University of Pittsburgh for a period of 1 year. |
| Number of Pages: |
137 |
| Institution: |
University of Pittsburgh |
| Schools and Programs: |
School of Public Health > Epidemiology |
| Degree: |
PhD - Doctor of Philosophy |
| Thesis Type: |
Doctoral Dissertation |
| Refereed: |
Yes |
| Uncontrolled Keywords: |
Epidemiologic Methods; Per-protocol Analysis; Randomized Controlled Trials; Directed Acyclic Graph; Causal Discovery; Doubly Robust Estimation; Causal Inference; Machine Learning; Pregnancy; Aspirin |
| Additional Information: |
The Chapter 3 of this dissertation was published by the American Journal of Epidemiology. To cite this chapter, please use "Yongqi Zhong, Edward H. Kennedy, Lisa M. Bodnar, Ashley I. Naimi (2021). AIPW: An R Package for Augmented Inverse Probability Weighted Estimation of Average Causal Effects. American Journal of Epidemiology. doi:10.1093/aje/kwab207" |
| Date Deposited: |
27 Aug 2021 18:16 |
| Last Modified: |
27 Aug 2022 05:15 |
| URI: |
http://d-scholarship.pitt.edu/id/eprint/41429 |
Metrics
Monthly Views for the past 3 years
Plum Analytics
Actions (login required)
 |
View Item |
![[feed]](/images/feed-icon-32x32.png){kind=icon}