Quantitative Issues in Cancer Research Working Group Seminar

Christian Covington, PhD Student, Department of Biostatistics, Harvard T.H. Chan School of Public Health

Multiverse Analysis for Causal Inference (and vice versa)

Abstract: Multiverse analysis is a framework developed in the quantitative social and behavioral sciences, designed to represent “non-statistical” uncertainty that arises in data analysis when making choices about conceptual operationalization, data preparation, etc. Multiverse analysis has become quite popular in the past few years, but there is disagreement about how it ought to be used. In particular, there are unresolved questions about how to construct a reasonable multiverse and how to interpret the results.

We’ll start with a brief history of how multiverse analysis was developed and has been built upon since its inception. I’ll argue that debates around constructing a “reasonable” multiverse lack a coherent philosophical underpinning and, as such, have contributed to general confusion among users. As one particular solution, I’ll propose a simple framework for thinking about multiverse analysis which is rooted in the basic tenets of causal inference and provide a case study in which the framework is applied. Finally, I’ll give some very early-stage ideas for future methodological work combining ideas from multiverse analysis, Bayesian model averaging, and causal inference.