Causal inference for complex data: randomization inference for treatment effect heterogeneity, network outcomes, and subgroup specific effects

Fredrickson, Mark M.

Permalink

https://hdl.handle.net/2142/101713 Copy

Description

Title

Causal inference for complex data: randomization inference for treatment effect heterogeneity, network outcomes, and subgroup specific effects

Author(s)

Fredrickson, Mark M.

Issue Date

2018-07-12

Director of Research (if dissertation) or Advisor (if thesis)

• Chen, Yuguo
• Bowers, Jake

Doctoral Committee Chair(s)

• Chen, Yuguo
• Bowers, Jake

Committee Member(s)

• Liang, Feng
• Hansen, Ben B.

Department of Study

Statistics

Discipline

Statistics

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

Randomized controlled trials, networks, treatment effects, subgroups, inference

Abstract

This dissertation presents three new methodologies for analyzing randomized controlled trials using the researcher controlled randomization mechanism as the basis for inference. The first method extends inference for the ``attributable effect'', the total of the difference of outcomes if the treatment group had instead been assigned to the control condition, to count and continuous data using a fast approximation algorithm. Alternative approaches are limited to binary data, require asymptotic approximations, or are computationally expensive. A refinement of the method to allow for including additional information is also included. The second method extends randomization inference to the study of network formation. Previous approaches either required strong parametric assumptions or only allowed for pre-treatment networks to be used. This approach develops several test statistics that can be used to test against common network formation models, based purely on the randomization of treatment. The final method improves inference in cluster randomized trials, where collections of individuals are assigned to treatment conditions simultaneously. Under the appealing assumption that larger clusters will have larger outcomes, on average, the method provides efficient, unbiased estimation of average treatment effects requiring minimal additional assumptions. All three of these methods demonstrate the relevance of randomized controlled trials to key areas of science and statistical development as well as the advantages of carefully crafting study design to fit the problem of interest. Data examples include a large scale field experiment involving health insurance, a gene-wide association study involving high dimensional outcomes, and a policy relevant study of parental social capital and student achievement in schools.

Graduation Semester

2018-08

Type of Resource

text

Permalink

http://hdl.handle.net/2142/101713

Copyright and License Information

Copyright 2018 Mark M. Fredrickson

Owning Collections