Relative-Risk Ratios And Multinomial Logit Coefficients

Why Coefficient Interpretation Depends on the Comparison

In a multinomial logit model, a coefficient does not describe a change in the probability of one outcome by itself. It describes how a covariate changes the comparison between one outcome and a chosen base alternative. Exponentiating the coefficient expresses that change as a multiplicative factor, often called a relative-risk ratio (RRR). The interpretation is useful only when the compared alternatives, covariate change, and units are stated clearly.

This tutorial uses the standard multinomial logit probability model and its base-alternative normalization. Alternatives are mutually exclusive categories, and the model represents their probabilities as normalized exponentials of systematic utility. Coefficients describe conditional associations under that model; by themselves, they do not establish causal effects. Estimation by maximum likelihood supplies coefficient estimates, but the interpretation below concerns the model coefficients themselves.

Definition: Let $Y_i$ take one of $K$ alternatives, indexed $0,1,\ldots,K-1$, and choose alternative $0$ as the base. In the normalized multinomial logit model, $V_{i0}=0$ and $V_{ij}=x_i^\top\beta_j$ for each nonbase alternative $j$. The choice probability is $P_{ij}=\exp(V_{ij})/\sum_{m=0}^{K-1}\exp(V_{im})$. Thus, $\log(P_{ij}/P_{i0})=x_i^\top\beta_j$. A coefficient $\beta_{jr}$ is the change in the log probability ratio for alternative $j$ versus the base associated with a one-unit increase in covariate $r$, holding the other included covariates fixed.

The denominator in the probability formula is shared by every alternative. Taking the ratio of two probabilities therefore removes that denominator. This is why the coefficient has a direct interpretation in terms of a log probability ratio, even though the probability of alternative $j$ depends on all alternatives’ utilities.

\frac{P_{ij}}{P_{i0}}=\exp(x_i^\top\beta_j), \qquad \log\left(\frac{P_{ij}}{P_{i0}}\right)=x_i^\top\beta_j.

From a Log-Odds Coefficient to a Relative-Risk Ratio

Suppose $x_{ir}$ is a numerical covariate and $\beta_{jr}$ is its coefficient for alternative $j$ relative to the base. Increase $x_{ir}$ by $\Delta$, leaving the other covariates unchanged. The log probability ratio changes by $\Delta\beta_{jr}$. Exponentiating gives the factor by which the probability ratio $P_{ij}/P_{i0}$ changes.

\frac{(P_{ij}/P_{i0})\text{ after the change}} {(P_{ij}/P_{i0})\text{ before the change}} =\exp(\Delta\beta_{jr}).

For a one-unit increase, the relative-risk ratio is $\exp(\beta_{jr})$. An RRR above one means that the probability of alternative $j$, relative to the probability of the base alternative, is multiplied by that factor. An RRR below one means that this relative probability is multiplied by a factor below one. Neither statement says that $P_{ij}$ itself is multiplied by the RRR: probabilities share a denominator and sum to one.

Worked Example: A Continuous Covariate and Two Alternatives

Suppose a hypothetical multinomial logit model predicts a person’s travel mode, with car as the base alternative and bus and bicycle as the nonbase alternatives. Let $x$ be weekly work-from-home hours. The estimated coefficient of $x$ is $0.30$ for bus versus car and $-0.20$ for bicycle versus car. Interpret a two-hour increase, holding other covariates fixed.

Bus versus car. The log probability ratio changes by $2(0.30)=0.60$. The corresponding RRR is

\exp(2\cdot 0.30)=\exp(0.60)\approx 1.822.

Thus, in this model, two additional weekly work-from-home hours multiply the bus-to-car probability ratio by about 1.822, or increase that ratio by about 82.2%. This does not mean that the probability of taking the bus rises by 82.2%.

Bicycle versus car. The log probability ratio changes by $2(-0.20)=-0.40$, so

\exp(2\cdot -0.20)=\exp(-0.40)\approx 0.6703.

The bicycle-to-car probability ratio is therefore multiplied by about 0.6703, a decrease of about 32.97% in that ratio. These interpretations are conditional on the model’s other included covariates, and concern comparisons with car rather than absolute changes in travel-mode probabilities.

Alternative-Specific Effects and Common Covariates

A covariate can be measured once for each person and still have a different coefficient for each alternative. For example, income may be a person-level covariate, but its association with choosing a particular mode rather than the base mode can differ by mode. In the base-normalized model, $\beta_{jr}$ is alternative-specific: it refers to alternative $j$ compared with alternative $0$. A positive coefficient for one alternative does not imply a positive coefficient for every other alternative.

This differs from an alternative-specific attribute, such as the fare of each travel option, which can have a distinct value for each alternative. Its coefficient describes how a change in that option’s attribute affects its utility relative to the base, according to the model specification. The same interpretation principle applies: identify which alternative’s systematic utility changes and which base is used for comparison.

A “common effect” requires care. If the same covariate enters every alternative’s utility with exactly the same coefficient, it adds the same amount to every utility. Since only differences in utilities determine the probabilities, that shared additive term cancels. A covariate common to people is not automatically such a common effect: it can have alternative-specific coefficients and therefore does not generally cancel.

Worked Example: A Shared Utility Term Cancels

Consider three hypothetical alternatives with utilities for a person given by $V_0=0.4z$, $V_1=0.4z+0.6x$, and $V_2=0.4z-0.2x$. Here $z$ is a person-level covariate with the same coefficient in all three alternatives, while $x$ has different coefficients. Alternative 0 is the base. Find the log probability ratios and interpret a one-unit increase in $x$.

Subtract the base utility from each nonbase utility. For alternative 1, $V_1-V_0=(0.4z+0.6x)-0.4z=0.6x$. For alternative 2, $V_2-V_0=(0.4z-0.2x)-0.4z=-0.2x$. Hence,

\log(P_1/P_0)=0.6x,\qquad \log(P_2/P_0)=-0.2x.

The $z$ term cancels from both comparisons, so this model assigns it no effect on the choice probabilities. For a one-unit increase in $x$, the RRR for alternative 1 versus 0 is $\exp(0.6)\approx1.822$; for alternative 2 versus 0 it is $\exp(-0.2)\approx0.8187$. These are changes in the respective probability ratios, not direct changes in the probabilities. If $z$ had different coefficients across alternatives, it would not cancel in this way.

Comparing Two Nonbase Alternatives

Coefficients reported relative to a base can also be used to compare two nonbase alternatives. For alternatives $j$ and $k$, subtract their log probability ratios to the base. The base terms cancel, leaving the log probability ratio for $j$ versus $k$. For a covariate with coefficients $\beta_{jr}$ and $\beta_{kr}$, its coefficient in that comparison is their difference.

\log(P_{ij}/P_{ik}) =x_i^\top(\beta_j-\beta_k), \qquad \text{one-unit RRR for } j \text{ versus } k =\exp(\beta_{jr}-\beta_{kr}).

This contrast is often more directly relevant than either coefficient alone. Its uncertainty must also be calculated as a contrast, accounting for the covariance between the estimated coefficients; comparing their individual standard errors or significance labels is not a valid substitute for testing the difference.

Worked Example: Comparing Two Nonbase Alternatives

Suppose a hypothetical travel-mode model uses car as its base and includes a continuous cost-sensitivity score $x$. The estimated coefficient is $0.35$ for bus versus car and $-0.20$ for bicycle versus car. Interpret a one-unit increase in $x$ for bus versus bicycle.

The coefficient for bus versus bicycle is the difference $0.35-(-0.20)=0.55$. Therefore the log probability ratio $P_{\text{bus}}/P_{\text{bicycle}}$ increases by $0.55$ per unit of $x$, and its RRR is

\exp(0.35-(-0.20))=\exp(0.55)\approx1.733.

A one-unit increase in the score multiplies the bus-to-bicycle probability ratio by about 1.733, holding other included covariates fixed. It is not correct to interpret the bus coefficient $0.35$ alone as the bus-versus-bicycle effect, because that coefficient compares bus with car. The relevant comparison subtracts the bicycle coefficient.

Categorical Predictors and Interactions

For a categorical predictor, the coefficient depends on the predictor’s reference category. If an indicator changes from zero to one, the coefficient is the change in the log probability ratio for that discrete contrast, and its exponential is the corresponding RRR. With an interaction, a coefficient may describe an effect only at the reference or zero value of the interacting covariate. The full effect must include all relevant terms.

Worked Example: Interpreting an Interaction

In a hypothetical model, transit is the base and car is a nonbase alternative. Let $G=1$ indicate membership in a specified group and $G=0$ be the reference group. Let income be measured in thousands of dollars. Suppose the car-versus-transit log probability ratio is

\log(P_{\text{car}}/P_{\text{transit}}) =-0.50+0.08\,\text{income}+0.40G+0.03(G\cdot\text{income}).

For $G=0$, the income slope is $0.08$: a $1,000 increase in income multiplies the car-to-transit probability ratio by $\exp(0.08)\approx1.083$, about an 8.329% increase in the ratio. For $G=1$, the slope is $0.08+0.03=0.11$, giving $\exp(0.11)\approx1.116$, about an 11.63% increase in the ratio per $1,000.

The group contrast at income $4$, or $4,000, is $0.40+0.03(4)=0.52$. Thus the car-to-transit probability ratio for $G=1$, compared with $G=0$ at that income, is multiplied by

\exp(0.40+0.03\cdot4)=\exp(0.52)\approx1.682.

The group coefficient $0.40$ alone is the group contrast only when income equals zero in the model’s units. At other income values, the interaction changes the contrast. These are conditional ratio comparisons, not group differences in the probability of choosing car.

Why These Interpretations Matter

The RRR scale is often easier to communicate than a log-odds coefficient, but it does not remove the need to specify the comparison. A useful interpretation names the outcome alternative, the base alternative, the covariate change and its units, and any conditioning variables or interaction values. For example: “A $1,000 increase in income multiplies the car-to-transit probability ratio by about 1.116 for group $G=1$, holding the other modeled covariates fixed.”

Probability changes require evaluating the full probability formula at specified covariate values, since the denominator depends on all alternatives. That is a separate calculation from interpreting coefficients as RRRs. In particular, an RRR of 1.5 does not mean a 50-percentage-point increase in probability, or even a 50% increase in the probability of the numerator alternative.

Key takeaway: A multinomial logit coefficient is a change in a log probability ratio for a specified alternative against a specified comparator. Exponentiating the coefficient gives a multiplicative change in that ratio. For a different comparator, form the appropriate coefficient contrast; for a shared additive utility term, check whether it cancels across alternatives.

Common Mistakes

Calling the RRR a probability ratio. The RRR multiplies $P_j/P_0$ across covariate values; it does not generally multiply $P_j$ itself.
Leaving out the base alternative. A coefficient for bus in a model based on car compares bus with car, not bus with every other alternative.
Using one coefficient to compare two nonbase alternatives. Subtract their coefficients to obtain the relevant log-ratio contrast.
Treating a person-level covariate as having one universal effect. Its coefficients can differ by alternative; only an identical additive effect across every alternative cancels.
Interpreting an interaction coefficient in isolation. Combine it with the main-effect coefficient at the stated value of the interacting covariate.
Inferring causation from the sign of a coefficient. A model coefficient describes a conditional association unless a suitable causal design and assumptions support a causal interpretation.

Check Your Understanding

Use the base-alternative comparison carefully in each answer.

A coefficient for alternative $j$ versus the base is $-0.25$. What is the one-unit RRR, and how should it be described?
A covariate increases by three units and has coefficient $0.10$ for alternative $k$ versus the base. Write the RRR for this change.
Two nonbase alternatives have coefficients $0.40$ and $0.15$ for the same covariate. What is the one-unit RRR for the first versus the second?
Why does a covariate with the same additive coefficient in every alternative’s utility cancel from the choice probabilities?
In a model with an interaction, why may the main-effect coefficient not describe the group contrast at every value of the interacting variable?