FIGURE 2.

The (non)collapsibility of common effect measures for binary outcomes as determined by the concavity, convexity, or linearity of the characteristic collapsibility function (CCF) $g_{\nu }(·)=f^{-1}\{f(·)+\nu \}$, where $f$ is the link function and $\nu$ is the conditional effect measure. $f(p)=log\{p/(1-p)\}$ for the logit link, $f(p)=log(-log(1-p))$ for the complementary log‐log link, and $f(p)={\mathrm{\Phi }}^{-1}(p)$, where $\mathrm{\Phi }(·)$ is the CDF of the standard normal distribution, for the probit link