Description
Model the cell counts of a three-or-more-way cross-tabulation as main effects plus interactions on a log scale, and prune interactions to find which categorical variables are genuinely associated.
Two-way chi-square tests cannot say whether an association between two survey variables holds across a third or arises from it. Log-linear analysis writes the log of each cell frequency in a multi-way table as a constant plus main effects and all interactions, with no dependent variable, mirroring the general linear model. The saturated model reproduces the table exactly, so the analyst seeks the most parsimonious model whose expected frequencies still match the observed ones. Hierarchical selection by backward elimination starts at the highest-order interaction, using k-way and partial-association chi-square tests, and keeps the lower-order terms of any retained interaction. The chosen model is refitted, checked with likelihood-ratio and Pearson goodness-of-fit tests (a high p-value means it reproduces the table), and interpreted through parameter estimates against reference categories, exponentiated as odds. It shows association structure, not dependence of one variable on others.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Unmigrated
- Stage
- Discovery, Post-launch
- Data
- Quantitative
- Effort
- Days
Steps
- Cross-tabulate three or more categorical variables, recoding sparse categories so few cells hold frequencies of 1 or less.
- Run hierarchical model selection by backward elimination from the saturated model.
- Read the k-way tests (can all effects of order k and above be dropped?) and partial-association tests for each effect.
- At each step remove the term whose deletion changes the likelihood-ratio chi-square least without significant loss of fit, keeping lower-order terms of retained interactions.
- Refit the chosen model in a general log-linear procedure, adding 0.5 to cells to avoid zeros, and check likelihood-ratio and Pearson goodness of fit.
- Try removing further terms and compare fit, then interpret parameter estimates against reference categories as odds and rank effects by their Z ratios.
Sources
- Statistics for Marketing and Consumer Research (Mazzocchi, 2008)
- An Introduction to Categorical Data Analysis (Agresti, 1996)
- Advanced Quantitative Data Analysis (Cramer, 2003)
- Quantitative Social Research Methods (Singh, 2007)
- Bayesian Statistical Modelling (Congdon, 2006)
Also known as
- Hierarchical log-linear analysis of multi-way contingency tables
- Log-linear analysis
- Multi-way frequency analysis
- General log-linear model (GLLM)
- Hierarchical log-linear model selection
Categories
- Analytics & Behaviour
- Statistics & Modelling