Dr. Kaufman speaking at UofT’s Medical Sciences Building.
Stats, Stats, Stats
This past Thursday, Dr. Jay S. Kaufman, professor and epidemiology program director at McGill University, Montreal, gave a presentation titled Statistics, Adjusted Statistics and Maladjusted Statistics at the Medical Sciences Building here at UofT. He spoke for a little over an hour to a packed room (spectators stood along the walls when seating ran out) about the 3 main inferential targets of statistical models. These three targets are the real world in the present, the real world in the future, and a hypothetical world in the future (to predict the impact of a change in policy).
The validity of these three models depends on the shielding away of common threats to their validity, Kaufman explained. The three primary threats to models being: confounding bias, selection bias, and information bias. Through adjustments, however, these biases may be diminished, giving a much closer approximation to statistical truth.
A Question of Ethics
A major issue Dr. Kaufman addressed in his presentation is the ethical gap existing in mathematical inequalities between groups, and adjustments rooted in a socially normative perspective that try to correct for these as a matter of aiming for a greater good. To provoke the audience’s thoughts on this matter, he questioned whether women should live longer than men, or whether poor people should die younger than rich people (as is the general trend, currently). The social value inherent in these questions exposes issues for public health statisticians and policymakers, and from an epidemiologist’s perspective, it prompts us to reflect on how we adjust our statistical models. In sum: there is no such thing as a neutral, valueless statistical adjustment. Each adjustment is part of a web of social constructions and tacit value systems, and is thus rife with implications.
Kaufman gave the example of a study on overweightness, which may need to blame or excuse certain causal factors—but these decisions depend on the goals of the study. This statistical contingency is an essential philosophical perspective backing epidemiological inquiry, according to Kaufman, if only because adjustable variables exist on a causal pathway (for example, smoking leads to tar deposition which leads to lung cancer). It is of extreme importance, then, to think about where a variable is located along this causal pathway.
Dr. Kaufman used racial disparities in cervical cancer in relation to hysterectomy procedures to illustrate this point. Removing women who underwent the surgery, and therefore did not have a cervix, returned data showing that cervical cancer among black women had increased at a greater rate than among white women. But by adjusting-out those who had hysterectomies, researchers made an implicit value judgment about why so many of those black women have undergone the procedure.
Adjusting our Adjustments
It is therefore necessary to consider carefully how and why we hold the weight of a variable constant to assess its impact on our outcomes. Dr. Kaufman ended his talk by emphasizing that “adjusted analysis is more sophisticated analysis”, and the way it re-values and re-prioritizes outcomes directs us to large, important questions. In quoting statistician George E.P. Box, who famously said “All models are wrong, but some are useful”, Kaufman prompts us to two larger questions going forward when we use adjusted analysis: “Useful to whom?” and “Useful for what purpose?”–questions that ultimately rely on the thinking of the principal modeler. And with this, Kaufman concluded his riveting hour of discussion.
