Interpreting Interactions In Multinomial Logits

Why Interactions Need Conditional Interpretations

In a multinomial logit model, an interaction means that a predictor’s effect on a category logit depends on another predictor. The coefficient of the interaction is not, by itself, the effect for every category comparison or every covariate value. To interpret it, specify both which categories are being compared and the values of the predictors being held fixed.

This tutorial uses category 0 as the baseline and categories \(j=1,\ldots,J-1\) as the alternatives. Let \(x\) and \(z\) be predictors, and let \(\mathbf{w}\) denote any other covariates, held fixed when making a conditional comparison. Write the category-versus-baseline logit as

The interaction coefficient \(\delta_j\) belongs to category \(j\)'s logit versus category 0. It need not equal the interaction coefficient in another category's logit. For convenience, set the baseline category's coefficients to zero: \(\alpha_0=\beta_0=\gamma_0=\delta_0=0\) and \(\boldsymbol{\lambda}_0=\mathbf{0}\). This lets the same formulas cover comparisons involving the baseline.

Reading the Category-Specific Interaction

When \(x\) is continuous, differentiating the category-\(j\)-versus-baseline logit with respect to \(x\) gives

Thus \(\beta_j\) is the slope when \(z=0\), while \(\delta_j\) is the change in that slope for a one-unit increase in \(z\). If \(z\) is a binary indicator, \(\beta_j\) is the slope in the \(z=0\) group and \(\beta_j+\delta_j\) is the slope in the \(z=1\) group. For a one-unit increase in \(x\), exponentiating the slope gives the corresponding odds ratio against the baseline.

If \(x\) is binary, the same algebra describes the change from \(x=0\) to \(x=1\); the conditional log-odds change is \(\beta_j+\delta_jz\). The interpretation of an interaction therefore depends on predictor coding and reference values. If \(z=0\) is not a meaningful value for a continuous predictor, centering \(z\) at a meaningful value can make the main-effect coefficient easier to interpret, without changing the fitted model.

Deriving Conditional Pairwise Odds Ratios

The model directly gives logits against category 0, but many questions compare two nonbaseline categories. The relevant quantity is obtained by subtracting their logits. This subtraction removes the baseline probability from the odds ratio, leaving a contrast between the two categories’ coefficients.

To verify this, first note that the log of the category-\(j\)-versus-\(k\) odds at \(x\) is

The terms not involving \(x\) cancel when this expression at \(x_0\) is subtracted from the expression at \(x_1\). The remaining difference is \((x_1-x_0)\{(\beta_j-\beta_k)+(\delta_j-\delta_k)z\}\). Exponentiating that log-odds difference gives the displayed odds ratio. This argument also applies when \(j\) or \(k\) is the baseline, using its zero coefficients. \(\square\)

For a one-unit increase in \(x\), the pairwise log-odds slope is \((\beta_j-\beta_k)+(\delta_j-\delta_k)z\). In particular, the interaction relevant to the \(j\)-versus-\(k\) comparison is the difference \(\delta_j-\delta_k\), not just \(\delta_j\). A category-specific interaction coefficient answers a question against the baseline; a nonbaseline pairwise comparison requires a contrast of interaction coefficients.

Binary Interactions as Differences-in-Differences

When both predictors are binary, the interaction coefficient has a particularly direct contrast interpretation. For category \(j\) versus the baseline, compare the treatment log-odds change between the \(z=1\) and \(z=0\) groups. This is a difference between two differences.

The equality follows by substituting the model into each bracket: terms that do not change with \(x\) cancel within each bracket, and subtracting the \(z=0\) treatment contrast from the \(z=1\) treatment contrast leaves \(\delta_j\). On the odds-ratio scale, exponentiating gives

For categories \(j\) and \(k\), the corresponding difference-in-differences is \(\delta_j-\delta_k\), and its exponent is the ratio of their treatment odds ratios across the two groups. This is a contrast on the category-pair log-odds scale.

Choosing the Contrast That Answers the Question

Before calculating an interaction interpretation, translate the research question into a comparison. Identify the outcome categories, the predictor change, and the values of the interacting covariate. Then form the corresponding coefficient contrast.

Multiply a slope by \(x_1-x_0\) when the question compares a change other than one unit, then exponentiate the log-odds contrast to obtain an odds ratio. For a categorical predictor with several levels, use the indicator and reference-level coding in the fitted model: a comparison between two nonreference levels generally requires subtracting their relevant coefficient combinations. The same principle applies to interactions involving multiple indicators.

Common Mistakes and Why the Interpretation Matters

• Reading \(\beta_j\) as the effect at every value of \(z\). With an interaction, it is the effect at \(z=0\); at another value, add \(\delta_jz\).
• Using only \(\delta_j\) for a comparison of two nonbaseline categories. For \(j\) versus \(k\), the interaction contrast is \(\delta_j-\delta_k\).
• Exponentiating before forming the coefficient contrast. First subtract the appropriate logit coefficients and evaluate the covariates; then exponentiate the resulting contrast.
• Leaving the category comparison implicit. “The treatment odds ratio” is incomplete in a multinomial model unless the two outcome categories and the conditioning values are specified.

These distinctions matter because the same fitted model can produce different conditional odds ratios for different category pairs and covariate values. Writing down the target comparison first prevents an interaction coefficient from being assigned a broader interpretation than it supports. It also makes results easier to communicate: report the category pair, predictor change, conditioning values, and odds-ratio scale together.