No label defined (Q3972862)

[For the entire collection see Zbl 0717.00021.] The collapsibility means that a measure of association remains unchanged when a contingency table is collapsed into a low-dimensional table. The well-known Simpson's paradox shows that a measure of association may be even reversed after collapsing. The authors consider an \(I\times J\times K\) contingency table of a background variable \(V_ 1\), a treatment variable \(V_ 2\) and an outcome variable \(V_ 3\). The strong collapsibility of the odds ratio, the collapsibility of the local--global odds ratio, the collapsibility of half local-global odds ratios and collapsibility of global odds ratio are defined and investigated. It is shown that the results are closely related with collapsibility of logit models. Two main conditions for this collapsibility are \(V_ 1\perp V_ 2\mid V_ 3\) and \(V_ 1\perp V_ 3\mid V_ 2\).