On the many uses and abuses of regression models (A reprise of the ISCB43 President’s Invited Talk)
johncarlin_iscb43_final_vicbiostat.pdf
As a PhD student, I was puzzled when my supervisor told me that techniques are not the important things in statistics. What he meant is that a statistician who is collaborating with an empirical scientist should focus on gaining a deep understanding of the substantive problem that the collaborator is seeking to solve, rather than turning immediately to the bank of technical knowledge about models and methods that s/he might have acquired. How exactly one should do this was less clear, but the practice and teaching of statistics have generally failed to recognise the essential truth of this point. Recently this has begun to change, and the key issues are articulated particularly well in Hernan’s elucidation of the “three tasks of data science” (description, prediction, causal inference). However, we have a long way to go, evidenced by the fact that the applied literature is still full of papers that fit multivariable regression models for ill-defined purposes such as “identifying risk factors” or “exploring the effects” of a list of variables on an outcome. This talk focusses on the familiar (generalised) regression models that dominate statistical practice, illustrating their (mis)use in applications first by considering a range of examples and then by examining more closely how and why they might be used to address each of the three types of empirical question. Is our aim to identify the correct model, or is it to develop a model that is useful for a particular purpose? How should that influence our approach to model specification and interpretation? Does our teaching of the fundamental concepts of regression models help or hinder better use in applications? I outline a proposed reform agenda that I believe would encourage better practice and thus better science.
This seminar will be held at the Doherty Institute and streamed on Zoom. Please be aware the presentation will be recorded.
To join via Zoom, click this link or go to https://monash.zoom.us/join and enter meeting ID: 858 1099 6525 and passcode: 606479