Rerandomization in Fractional Factorial Experiments
Open Access
- Author:
- Seng, Bibiana
- Graduate Program:
- Statistics
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- May 19, 2021
- Committee Members:
- Kari Lock Morgan, Thesis Advisor/Co-Advisor
Runze Li, Committee Member
Ephraim Mont Hanks, Program Head/Chair - Keywords:
- Design of Experiments
Rerandomization
Applied Statistics
Factorial Experiments - Abstract:
- Covariate balance is an important aspect of causal inference, and it is just as important in experiments. The gold standard for achieving covariate balance in experiments is through randomization. In initial randomizations, there is always the possibility of achieving randomized groups that do not lend themselves to proper balance. This can skew the results of the experiment and introduce bias simply by using a “bad"randomization. To combat this, one can rerandomize, where if a group of covariates is not balanced enough, as judged by an a-priori threshold, the units in the study are simply randomized again before being tested. Recent literature has shown a particular interest in combining randomization with classic Design of Experiment (DoE) procedures.In fact, there have already been successful results in combining rerandomization with full factorial experiments. In cases where a full factorial experiment cannot be used,such as being too computationally intensive or having too many experimental factors of interest, one would use a fractional factorial experiment; but how can you combine that with rerandomization? This study focuses on the application and results of combining rerandomization with experiment allocation for fractional factorial experiment designs,and is applied to a hypothetical experiment on real world data.