Likelihood-Based Goodness of Fit in Two-Way Contingency Tables

Models of contingency tables are based on the counts by category. In a two-way table, models can depend on either, neither, or both of the categories. The likelihood ratio statistic provides a measure of how well a particular model fits the original counts. The null hypothesis is that the chosen model fits the data well. The alternative hypothesis is that the saturated model (the model with predicted counts equal to the actual counts) is needed. A small -value for the statistic indicates the chosen model does not fit the data well. As the counts in the table get large, follows a distribution, and a approximation can be used to obtain a -value provided the predicted counts in the table are not very small.

Use the sliders to adjust the original counts. Select between the four models to get the predicted counts and test statistic for that model of the contingency table.

The underlying model is a log-linear Poisson model. The categories are treated as nominal predictor variables.

The mean count model fits the case where each cell is equally likely. This model predicts an equal number for each cell in the table.

The choice effect model fits the case where the count differs across choice, but is constant across group. The group effect model predicts that counts differ across group, but not across choice.

Choice and group have an additive effect on predicted counts for the additive model. The total predicted counts by choice and the total predicted counts by group for this model match those totals for the original contingency table.