Interpreting Category-Specific Logit Coefficients

What a Category-Specific Coefficient Compares

In a baseline-category logit model, each nonbaseline outcome category has its own logit: its log odds against the chosen baseline category. The coefficient in one such logit describes that category-versus-baseline comparison. It does not, by itself, describe the odds of one nonbaseline category against another.

Let the outcome \(Y\) have categories \(1,\ldots,J\), with category \(J\) as the baseline, and let \(x\) be a vector of predictors. For a particular nonbaseline category \(j\), write its model as

\log\left(\frac{P(Y=j\mid x)}{P(Y=J\mid x)}\right) =\alpha_j+x^\mathsf{T}\beta_j.

The ratio inside the logarithm is the odds of category \(j\) versus the baseline, conditional on the predictor values \(x\). The intercept \(\alpha_j\) and the vector of slopes \(\beta_j\) belong to this particular category-specific logit. A different nonbaseline category has its own coefficients and its own comparison with the baseline.

Definition: For category \(j\) versus baseline \(J\), the category-specific log odds at predictor profile \(x\) are \(\log\{P(Y=j\mid x)/P(Y=J\mid x)\}\). A slope \(\beta_{jr}\) is the coefficient of predictor \(x_r\) in that logit. The conditional odds for this comparison are \(P(Y=j\mid x)/P(Y=J\mid x)\).

An intercept needs a reference profile: it is the log odds when every predictor in the model is at its coded value zero. A slope describes a change in log odds when its predictor changes, with the other predictors held fixed. These interpretations depend on the predictor’s units and coding.

From a Slope to a Conditional Odds Ratio

The key calculation is to compare two predictor profiles in the same category-specific logit. Suppose the profiles differ only in predictor \(x_r\), which increases by \(\Delta\). Subtracting their linear predictors gives a change of \(\Delta\beta_{jr}\) in log odds. Exponentiating that difference gives the ratio of the two conditional odds.

Theorem: In a category-specific logit with no interaction involving \(x_r\), changing \(x_r\) by \(\Delta\) while holding the other predictors fixed multiplies the odds of category \(j\) versus baseline \(J\) by \(\exp(\Delta\beta_{jr})\). In particular, a one-unit increase has conditional odds ratio \(\exp(\beta_{jr})\).

Proof. Let \(x\) be the initial profile and \(x'\) the new profile, with \(x'_r=x_r+\Delta\) and \(x'_s=x_s\) for every \(s\ne r\). The difference in category-specific log odds is

\begin{aligned} \eta_j(x')-\eta_j(x) &=\left(\alpha_j+\sum_{s\ne r}\beta_{js}x_s+ \beta_{jr}(x_r+\Delta)\right) -\left(\alpha_j+\sum_{s\ne r}\beta_{js}x_s+ \beta_{jr}x_r\right)\\ &=\Delta\beta_{jr}. \end{aligned}

Write \(O_j(x)=P(Y=j\mid x)/P(Y=J\mid x)\) for these conditional odds. Since \(\eta_j(x)=\log O_j(x)\), exponentiating the difference yields

\frac{O_j(x')}{O_j(x)} =\exp\{\eta_j(x')-\eta_j(x)\} =\exp(\Delta\beta_{jr}).

This establishes the result. The exponentiated slope is an odds ratio for the stated predictor change and the stated category-versus-baseline comparison. It is conditional on the other predictors being held fixed. \(\square\)

For a binary predictor coded \(0\) and \(1\), the slope compares level \(1\) with level \(0\), so the conditional odds ratio is \(\exp(\beta_{jr})\). Reversing the predictor comparison changes the sign of the log-odds difference and takes the reciprocal of the odds ratio. For a continuous predictor, the one-unit odds ratio may not be the most useful scale: a change of \(\Delta\) units has odds ratio \(\exp(\Delta\beta_{jr})\).

Worked Interpretations

Worked Example: An Intercept and an Age Slope

Suppose the outcome is hospital disposition: admission, discharge, or transfer. Use discharge as the baseline and consider the admission-versus-discharge logit. Let \(x\) be age measured in decades above age 50, and suppose the fitted logit is

\log\left(\frac{P(Y=\text{admission}\mid x)} {P(Y=\text{discharge}\mid x)}\right) =-0.80+0.25x.

At age 50, \(x=0\). The fitted log odds of admission versus discharge are therefore \(-0.80\), and the corresponding odds are

\exp(-0.80)\approx 0.4493.

Thus, at age 50, the conditional odds of admission rather than discharge are about \(0.449\) to \(1\). The intercept is not an overall odds or a probability; it refers to this particular age and to the admission-versus-discharge comparison.

A ten-year increase in age is an increase of one unit in \(x\). The log odds increase by \(0.25\), so the conditional odds ratio is

\exp(0.25)\approx 1.2840.

For example, at age 60 the fitted log odds are \(-0.80+0.25=-0.55\), giving odds \(\exp(-0.55)\approx0.5770\). Their ratio to the age-50 odds is \(\exp(-0.55)/\exp(-0.80)=\exp(0.25)\approx1.2840\), as the slope interpretation requires. This statement concerns odds of admission versus discharge, not odds of admission versus transfer.

Worked Example: A Binary Predictor and Reversed Comparison

Consider the logit for completing a program versus withdrawing, with withdrawal as the baseline. Let \(x=1\) for a public institute and \(x=0\) for a private institute. Suppose

\log\left(\frac{P(Y=\text{complete}\mid x)} {P(Y=\text{withdraw}\mid x)}\right) =0.40-0.75x.

For a private institute, the fitted log odds are \(0.40\), and the odds of completion versus withdrawal are \(\exp(0.40)\approx1.4918\). For a public institute, the fitted log odds are \(0.40-0.75=-0.35\), and the corresponding odds are \(\exp(-0.35)\approx0.7047\).

The public-versus-private comparison changes \(x\) by \(1\), so the conditional odds ratio is

\frac{\text{odds at public}}{\text{odds at private}} =\frac{\exp(-0.35)}{\exp(0.40)} =\exp(-0.75) \approx 0.4724.

Thus, conditional on the other predictors, the odds of completion rather than withdrawal at public institutes are about \(0.472\) times those at private institutes. Reversing the comparison gives \(1/\exp(-0.75)=\exp(0.75)\approx2.117\). Using the rounded value \(0.472\) instead gives \(1/0.472\approx2.119\). The small difference is due to rounding; the exact reciprocal relationship applies to the unrounded odds ratio.

Worked Example: Choosing a More Useful Predictor Change

Suppose a model for partial response versus no response has the category-specific logit

\log\left(\frac{P(Y=\text{partial response}\mid z,w)} {P(Y=\text{no response}\mid z,w)}\right) =0.30-0.12z+0.40w,

where \(z\) is dose measured in five-milligram units and \(w\) is a fixed covariate. Interpret the dose coefficient for a ten-milligram increase, holding \(w\) fixed.

A ten-milligram increase is an increase of \(10/5=2\) units in \(z\). Its change in log odds is \(2(-0.12)=-0.24\). Therefore the conditional odds ratio for partial response versus no response is

\exp(-0.24)\approx0.7866.

For a ten-milligram dose increase, the conditional odds are multiplied by about \(0.787\), or are about \(21.3\%\) lower, because \(1-0.7866=0.2134\). This is a comparison of odds at two doses for the same value of \(w\); it is not a claim that the probability of partial response falls by 21.3 percentage points.

Why the Intercept Depends on Coding

The intercept is the log odds at the predictor profile coded as zero, not necessarily at a typical or scientifically important profile. Centering a predictor can make the intercept refer to a more useful reference value without changing the fitted logit. This is a change in interpretation and parameterization, not a change in the underlying comparison.

Result: For a single predictor \(x\), if \(\eta_j(x)=\alpha_j+\beta_jx\) and the centered predictor is \(x^*=x-c\), then the same logit can be written \(\eta_j(x^*)=(\alpha_j+c\beta_j)+\beta_jx^*\). The new intercept is the log odds at \(x=c\); the slope is unchanged.

Derivation. Since \(x=x^*+c\), substitute this expression into the original logit:

\eta_j =\alpha_j+\beta_j(x^*+c) =(\alpha_j+c\beta_j)+\beta_jx^*.

When \(x^*=0\), the original predictor is \(x=c\), so the new intercept is the log odds at that predictor value. The coefficient of \(x^*\) remains \(\beta_j\), so an equal-sized change in the predictor still has the same log-odds change. \(\square\)

With several predictors, the intercept refers to the profile where each numerical predictor is zero and each categorical predictor is at its reference level. If zero is outside the observed range, the intercept is still defined by the model, but its practical interpretation may be an extrapolation. Centering or choosing a meaningful reference level can clarify what it represents.

Common Misreadings and Why the Distinction Matters

Dropping the baseline from the interpretation. A coefficient is for one named outcome category versus the specified baseline. State both categories when describing it.
Calling the exponentiated slope a probability ratio. \(\exp(\beta_{jr})\) is an odds ratio for the category-versus-baseline comparison. It is not a probability ratio or a fixed change in probability.
Forgetting the predictor’s units. If age is in decades, a one-unit increase means ten years. If dose is measured in five-milligram units, a two-unit increase means ten milligrams.
Interpreting the intercept at the wrong profile. It is the log odds when the predictors take their coded zero or reference values; it is not automatically the log odds for an average subject.
Reversing a comparison without changing the sign. Reversing the predictor levels negates the log-odds contrast and reciprocates the odds ratio.

Clear interpretations preserve three pieces of information: the outcome category, the baseline category, and the predictor contrast. For example, “a ten-year increase multiplies the conditional odds of admission rather than discharge by 1.28, holding the other predictors fixed” identifies the comparison, scale, and conditioning. That precision matters because each category-specific logit answers its own category-versus-baseline question.

Check Your Understanding

For each question, identify the category comparison and predictor contrast before interpreting the coefficient.

In a logit for “employed” versus “not employed,” a predictor has slope \(0.30\) per year. What is the conditional odds ratio for a three-year increase, holding other predictors fixed?
A binary predictor is coded \(0\) for standard care and \(1\) for a new treatment. What comparison does its slope describe in a category-specific logit?
A predictor is centered at 40, so its value is zero for a person aged 40. What profile does the intercept describe, assuming the other predictors are also at their reference values?
If a category-specific slope is negative, how do the log-odds change and the one-unit conditional odds ratio behave?
Why does an odds ratio of \(0.80\) not mean that the category probability is lower by 20 percentage points?