V. Núñez Antón, M. A. García Pérez, R. Alcalá Quintana
The analysis of contingency tables uses chi-square statistics. If the null is rejected, residual analyses are conducted to identify influential cells. Such analyses are conditional on the results of the omnibus test and may lose control of Type I error rates. Simulations show that residual analyses conditional on significant chi-square tests yield Type I error rates up to five times larger than the nominal rate. We describe an unconditional approach consisting of testing the omnibus hypothesis through familywise residual analyses with a correction that takes into account the interdependence of residuals. The correction is based on bootstrap results tailored to the cell dependencies in the table, and this strategy is shown to maintain the nominal Type I error rate of both the omnibus test and the residual analyses. Application of this method is illustrated with empirical examples involving tests of homogeneity, tests of independence, and goodness-of-fit tests for parametric models.
Palabras clave: Contingency tables, Residual analysis, Chi-square tests, Multiple testing, Bootstrap
Programado
L06.3 Sesión de la Sociedad Española de Biometría
5 de septiembre de 2016 12:55
Aula 21.08
