WorksMathematics
Worst-case bias of the independence assumption for functions of two ordinal variables
Abstract
Let ordinal grades and be obtained by cutting a pair of random variables with copula at arbitrary thresholds, and let be a function of the grades. We study the largest error , over all thresholds, that is, over all marginal laws, made when the mean of is computed as if the grades were independent. This is the counterpart of the classical dependence-uncertainty problem: the copula is fixed and the discretisation is free. The error is caused entirely by the interaction of the inputs, measured by an index , the largest absolute -volume of a rectangle of grades, which vanishes if and only if is additive. For every copula, lies between times the largest deviation of from the independence copula and times half the total variation of their difference, and the factor one half cannot be improved. For the Gaussian copula with correlation we prove that equals to first order, with a sharp constant and an explicit remainder. The same first-order law holds for copula families whose first-order departure from independence is a product of unimodal factors, including the Farlie-Gumbel-Morgenstern, Ali-Mikhail-Haq, Frank, Clayton and Plackett families, and we characterise the Archimedean families with this property. An application to the TNM staging of colon cancer, with data from a clinical trial, compares the bounds with the actual error of the independence computation.