Скачать или смотреть The Adjustment Puzzle: All about Collapsibility

Why does adjusting for covariates sometimes change an effect estimate even when there is no confounding?

This question sits at the heart of one of the most misunderstood topics in applied statistics: collapsibility.

In this talk, The Adjustment Puzzle, we unpack why certain effect measures (like odds ratios and hazard ratios) behave differently from risk differences or risk ratios when adjusting for covariates. The answer is not bias, confounding, or model failure: it is a mathematical property of the estimand itself.

This presentation covers:
• What collapsibility actually means (and what it does not)
• Why odds ratios and hazard ratios are non-collapsible
• The difference between marginal and conditional effects
• How nonlinearity and Jensen’s inequality drive the phenomenon
• Why “adjusting more” does not necessarily mean “estimating better”
• Practical implications for regression, epidemiology, and causal inference

The goal is to move beyond rules of thumb and clarify why adjustment changes estimates and when that change is expected, not problematic.

This talk is aimed at statisticians, data scientists, epidemiologists, economists, and anyone working with regression models and causal effects.