Loading...

Causal Inference and CUPED Randomized Experiments

Akrami, Mahdi | 2023

68 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 56470 (02)
  4. University: Sharif University of Technology
  5. Department: Mathematical Sciences
  6. Advisor(s): Haji Mirsadeghi, Mir Omid
  7. Abstract:
  8. In today’s world, large internet companies carry a lot of data. The tests with low triggering rate are designed for causal inference. By regarding the economic scale of Internet companies and high decision -making speeds, it is essential to use variance reduction methods to estimate the causal effect of treatment with sufficient precision. To this goal, the trigger- dilute analysis method is designed and used. When it comes to partial triggering, the stimulation analysis method loses its efficiency, so a new replacement method is designed in which an unbiased estimation along with the reduction of variance is presented to estimate ITT within one-way experiments. This method is based on the idea of increasing efficiency using the CUPED method, which uses identification augmentation in the literature of the principal stratification
    and instrumental variables. The presented unbiased estimator is relying on the zero expectation of augmented term. It is also suggested that if this assumption is not satisfied it has been suggested by using in-experiment observations reduce the biased due to augmented term. this method will sacrifice the amount of variance reduction for biased decreasing. The causal inference is based on the method of potential outputs and using of assumptions such as the principal ignorability, the principal score and the exclusion restriction. the causal effect of the treatment can be estimated by using the principal ignorability and the principal score, they present a coherent method to estimate the causal effect in one sided and two sided experiments context. first the methods to estimate the causal effect are represented and then explained how the uncertainty can be obtained by using the Bootstrap method.
  9. Keywords:
  10. Two-Sided Non-Compliance Experiment ; One-Sided Non-Compliance Experiment ; Randomized Experiments ; Weak Principal Ignorability ; Strong Principal Ignorability ; Principal Score

 Digital Object List

 Bookmark

...see more