Skip to main content
eScholarship
Open Access Publications from the University of California

UCLA

UCLA Electronic Theses and Dissertations bannerUCLA

Causal Inference Outside of Randomized Trials with the Stability-Controlled Quasi-Experiment: Extensions and Considerations

  • Author(s): Wulf, David
  • Advisor(s): Hazlett, Chad J
  • et al.
Abstract

In many non-randomized, observational settings, treatment assignment mechanisms are opaque and researchers may not trust an assumption of conditional ignorability or the covariate-adjustment-based identification it allows. The Stability-Controlled Quasi-Experiment (SCQE, 2019) avoids this adjustment approach and makes no direct comparison between treated and untreated units. In settings with a change in the usage rate of some treatment of interest between successive cohorts, SCQE transforms an assumption about cohort-wide counterfactual outcome trends into estimates of the treatment's effect on those units that were impacted by the treatment change.

For example, imagine two consecutive cohorts of patients, the latter featuring non-randomized use of a new treatment. SCQE can identify and estimate an Average Treatment Effect among the Treated (ATT). The assumption required is a proposed value of δ, the unobservable difference between the cohorts’ average outcomes that we would have seen had the treatment not been introduced.

SCQE is a partial identification strategy, presenting these effects across values of δ. In doing so, it ties claims about beneficial, harmful, or null effects to the corresponding δ values we would need to defend in order to support them, helping to avoid overconfidence in "suggestive" point estimates. In many applications, δ may also be more intuitive and easy to reason with than sensitivity analyses for conditional-ignorability-based results.

In this thesis, I (1) present new inferential tools for SCQE, including summary-statistic-based inference for use when unit-level data is unavailable or limited by data sharing restrictions; (2) extend the method to accommodate expert-informed prior distributions on δ, applications with only one cohort, and existing treatments experiencing changes in usage rates; (3) apply the method to evaluations of a tuberculosis prophylactic in Tanzania, early COVID-19 treatments, and in-hospital rapid response alerts; and (4) discuss important considerations, conceptual guidance, and best practices for practitioners.

Main Content
For improved accessibility of PDF content, download the file to your device.
Current View