Statistics Glossary → Box-Tidwell test (linearity of the logit)
Box-Tidwell test (linearity of the logit)concept
A test, in logistic regression, of the assumption that a continuous predictor is linearly related to the logit (log-odds).
Logistic regression assumes a linear relationship between continuous predictors and the logit of the probability, log(p / (1 - p)), not the outcome itself. In the Box-Tidwell approach, an interaction term x·ln(x) is added to the model for each continuous predictor x. If the coefficient of this term is significant, the relationship between x and the logit is not linear and a suitable transformation of x (e.g. log, square root, polynomial or spline) is needed. Hosmer and Lemeshow recommend this test for checking linearity in logistic models.
ln(x) must be defined, so the test applies only to predictors with positive values; zero or negative values are handled by adding a constant to shift the variable. If many predictors are tested, a multiple-comparison correction should be considered. In large samples the test may flag even very small departures, so its result should be weighed together with plots of the relationship between the predictor and the logit (e.g. a smoothed logit plot).
An everyday example
A bank's loan-repayment model assumes that the logit of the probability of default falls steadily as income rises. In reality, a small rise in income may reduce risk a great deal when income is low, while at high incomes extra income hardly changes the risk at all. The Box-Tidwell test catches this kind of curved relationship and suggests entering the logarithm of income into the model.
In MerQur
It is applied in Logistic Regression through the 'Box-Tidwell' option, which adds an x·ln(x) term for each continuous positive predictor.