Multinomial Logistic Regression Explained

Introduction

Many clinical outcomes have more than two possible categories. A patient might experience a complete response, partial response, stable disease, or progressive disease. A toxicity might be classified as mild, moderate, severe, or life-threatening. A treatment disposition might be continuation, discontinuation for toxicity, discontinuation for lack of efficacy, or another reason.

When the outcome has exactly two categories, ordinary binary logistic regression is often appropriate. But what happens when the outcome has three or more categories?

One important approach is multinomial logistic regression.

What Makes an Outcome Multinomial?

Suppose a clinical endpoint has three possible categories:

Each patient belongs to exactly one category. The outcome is therefore categorical, but it is not necessarily appropriate to treat the categories as numerical values such as 1, 2, and 3.

The crucial question is whether the categories have a meaningful ordering.

A Clinical Example

Suppose a Phase III clinical trial evaluates whether treatment assignment and baseline disease severity are associated with a patient's best overall tumor response.

The outcome is classified as:

• Complete response (CR)
• Partial response (PR)
• Stable disease (SD)
• Progressive disease (PD)

For this example, assume the investigators want to model the probability of these four categories without imposing a proportional-odds assumption.

The outcome has:

categories. One category must be selected as the reference category. Suppose progressive disease is the reference.

The Basic Model

Let:

• \(Y\) = tumor response category
• \(Y=1\) = complete response
• \(Y=2\) = partial response
• \(Y=3\) = stable disease
• \(Y=4\) = progressive disease, the reference category

For each non-reference category, multinomial logistic regression models the log odds relative to the reference outcome.

The three equations are:

This is the central mathematical structure of multinomial logistic regression.

How Many Equations Are Needed?

If the outcome has \(K\) categories, select one reference category. The remaining \(K-1\) categories each receive a separate logit.

This is why multinomial logistic regression is sometimes described as a set of simultaneous logistic regressions. That description is useful intuitively, although the equations are estimated jointly as one multinomial model rather than as unrelated independent binary regressions.

The Reference Category

The reference category is fundamental to interpretation. Suppose progressive disease is the reference. Then every odds ratio answers a question of the form:

For example:

or:

Changing the reference category changes the coefficients and odds ratios, although it does not change the underlying fitted probabilities of the multinomial model.

Adding Treatment as a Predictor

Suppose:

For the complete-response versus progressive-disease logit:

The treatment coefficient is:

Exponentiating it gives:

This is the ratio of the odds of complete response versus progressive disease for treatment compared with control.

Interpreting a Multinomial Odds Ratio

Suppose the fitted model produces:

Then:

The interpretation is: patients receiving experimental treatment have approximately three times the odds of complete response versus progressive disease compared with patients receiving control, assuming other variables in the model are held constant.

This does not mean that treatment triples the probability of complete response.

A Worked Clinical Example

Consider 600 patients randomized to an experimental treatment or control. The primary categorical outcome is best overall response:

Suppose the model adjusts for:

• Treatment group
• Baseline disease severity
• Age

The fitted treatment effects are:

The treatment appears to have the strongest relative association with complete response versus progressive disease.

Interpreting Each Equation

Complete Response vs Progressive Disease

An odds ratio of 3.00 means:

For patients with otherwise equivalent covariate values, the odds of complete response rather than progressive disease are estimated to be three times higher under experimental treatment than under control.

Partial Response vs Progressive Disease

An odds ratio of 1.99 means:

The odds of partial response rather than progressive disease are estimated to be approximately twice as high under experimental treatment.

Stable Disease vs Progressive Disease

An odds ratio of 1.25 means:

The odds of stable disease rather than progressive disease are estimated to be 25% higher under experimental treatment.

The three effects need not be equal.

Why You Should Not Run Three Separate Binary Logistic Models

A tempting approach would be to create three separate binary outcomes:

• CR versus everything else
• PR versus everything else
• SD versus everything else

and fit three logistic regressions.

That is generally not equivalent to multinomial logistic regression.

The multinomial model treats the outcome as a single multicategory random variable whose category probabilities must sum to one.

Separate binary models do not automatically respect this joint probability structure.

From Logits to Probabilities

The model is estimated on the log-odds scale, but clinicians often want predicted probabilities. Suppose there are four categories and PD is the reference. Define:

The probability of progressive disease is:

The other probabilities are:

These probabilities necessarily sum to one.

Example of Predicted Probabilities

Suppose a particular patient has fitted linear predictors:

The denominator is:

Approximately:

Therefore:

Thus, the model predicts approximately:

Odds Ratios Versus Predicted Probabilities

These two quantities answer different questions.

For clinical communication, predicted probabilities are often easier to understand than odds ratios. However, odds ratios remain useful for inference and formal reporting of covariate associations.

Continuous Predictors

Suppose age is included as a continuous predictor. For the CR versus PD comparison:

If:

then the odds ratio for a one-year increase in age is:

Thus, each one-year increase in age is associated with an estimated 2.9% decrease in the odds of CR versus PD, assuming the linearity assumption on the logit scale is reasonable.

Using a Clinically Meaningful Increment

A one-year change may not always be clinically meaningful. Suppose the age coefficient is \(-0.03\), but investigators want the effect per 10-year increase.

Then:

This means that a 10-year increase in age is associated with approximately a 26% lower odds of CR versus PD.

Categorical Predictors

Suppose baseline disease severity has three categories:

Two indicator variables can be created:

Each indicator receives a coefficient for each non-reference outcome category.

This means the model can become fairly large. For \(K\) outcome categories and a categorical predictor with \(C\) levels, there can be:

coefficients for that predictor, assuming the predictor uses \(C-1\) indicator variables.

The Number of Parameters Can Grow Quickly

Suppose the outcome has five categories and the model contains a four-level categorical predictor. Then that predictor alone requires:

regression coefficients.

Global Tests of a Predictor

Suppose treatment appears in all three equations. There are three treatment coefficients:

A separate p-value for each coefficient answers a specific comparison. But investigators may also ask a broader question: Does treatment have any overall association with the multinomial outcome?

A global test can assess the joint null hypothesis:

against the alternative that at least one is nonzero.

Likelihood Ratio Tests

One common approach is a likelihood ratio test. Compare:

• A full model containing the predictor.
• A reduced model without the predictor.

The likelihood ratio statistic is:

Under standard regularity conditions, the statistic is approximately chi-square distributed.

If the predictor has \(q\) associated coefficients, the degrees of freedom for the global test are generally:

For example, a binary treatment variable in a four-category multinomial outcome produces three treatment coefficients and therefore a three-degree-of- freedom global test.

Confidence Intervals for Odds Ratios

Suppose:

with standard error:

An approximate 95% confidence interval on the coefficient scale is:

giving approximately:

Exponentiating gives the odds-ratio interval:

Thus, the estimated odds ratio is approximately 2.00 with a 95% confidence interval of approximately 1.22 to 3.25.

What Does an Odds Ratio of 1 Mean?

For a particular category-versus-reference comparison:

means the predictor does not change the relative odds between those two outcome categories.

For example:

means the odds of partial response versus progressive disease are the same between the predictor groups, conditional on the other model variables.

What Does an Odds Ratio Below 1 Mean?

Suppose:

This means the predictor is associated with a 40% lower odds of stable disease versus progressive disease:

or 40%.

Again, this is a statement about relative odds, not an absolute probability reduction.

Reference Category Changes Interpretation

Suppose progressive disease is initially the reference. The model reports:

If the reference is changed to complete response, the model can instead report the comparison:

The underlying information is the same.

Baseline-Category Logit Model

The formulation used above is commonly called the baseline-category logit model. For \(K\) outcome categories and reference category \(K\):

Each category \(j\) has its own coefficient vector.

This means:

• Each non-reference category has its own intercept.
• Each predictor can have a different coefficient for each category.
• The model does not assume that the predictor has the same effect across categories.

The Independence of Irrelevant Alternatives Assumption

A traditional multinomial logit model is associated with the independence of irrelevant alternatives, or IIA, assumption. Conceptually, IIA means that the relative odds between two outcome categories do not depend on the presence or characteristics of another alternative.

For example, consider:

• Treatment A
• Treatment B
• Placebo

If B is added or removed, the ratio of the odds of A versus placebo is assumed not to change simply because B exists as another alternative.

Whether IIA is scientifically plausible depends heavily on the application.

Multinomial vs. Ordinal Logistic Regression

This distinction is especially important in clinical research. Suppose toxicity is categorized as:

These categories have a natural order. An ordinal logistic model may therefore be appropriate.

Now consider best response:

• Complete response
• Partial response
• Stable disease
• Progressive disease

Although clinicians may perceive an ordering, the scientific relationship is not necessarily equivalent to a simple ordinal scale. The decision to use an ordinal model should therefore be based on the endpoint's measurement structure and assumptions, not merely on whether one category seems "better" than another.

The Proportional-Odds Assumption Is Not a Multinomial Assumption

A common mistake is to ask whether a multinomial model satisfies the proportional-odds assumption. It does not. The proportional-odds assumption belongs to the standard ordinal logistic model.

A multinomial baseline-category model instead allows different regression coefficients for different category-versus-reference logits.

Worked Example: Adding Treatment and Age

Suppose the fitted model for CR versus PD is:

The PR versus PD model is:

The SD versus PD model is:

Notice that age has different coefficients in the three equations.

That is perfectly permissible.

Calculate the Odds Ratios

For treatment:

For age:

Thus age has a negative association with CR versus PD and PR versus PD, but a small positive association with SD versus PD in this hypothetical model.

Predicting a Patient's Outcome

Consider a 60-year-old patient receiving experimental treatment. Then:

For CR versus PD:

For PR versus PD:

For SD versus PD:

Exponentiating:

The denominator is:

Therefore:

The predicted probabilities are therefore approximately:

The most likely predicted category for this patient is stable disease.

Model Fitting in R

A commonly used R implementation is the nnet package.

library(nnet)

fit <- multinom(
  response ~ treatment + age + severity,
  data = clinical_data
)

summary(fit)

The outcome should be a factor, with the desired reference category specified appropriately.

clinical_data$response <-
  relevel(
    factor(clinical_data$response),
    ref = "PD"
  )

fit <- multinom(
  response ~ treatment + age + severity,
  data = clinical_data
)

Obtaining Odds Ratios in R

coef_table <- summary(fit)$coefficients

odds_ratios <- exp(coef_table)

odds_ratios

The exponentiated coefficients give the multinomial odds ratios.

Confidence Intervals in R

fit_summary <- summary(fit)

beta <- fit_summary$coefficients
se   <- fit_summary$standard.errors

lower <- exp(beta - 1.96 * se)
upper <- exp(beta + 1.96 * se)

data.frame(
  OR = exp(beta),
  Lower95 = lower,
  Upper95 = upper
)

Predicted Probabilities in R

The predict() function can return predicted class probabilities.

predicted_probabilities <-
  predict(
    fit,
    newdata = new_patient,
    type = "probs"
  )

predicted_probabilities

The resulting object contains the predicted probability of each outcome category for the specified patient.

Predicting the Most Likely Category

predicted_class <-
  predict(
    fit,
    newdata = new_patient,
    type = "class"
  )

predicted_class

This returns the category with the highest model-predicted probability.

Global Tests in R

A convenient approach is to compare a full and reduced model.

fit_full <- multinom(
  response ~ treatment + age + severity,
  data = clinical_data
)

fit_reduced <- multinom(
  response ~ age + severity,
  data = clinical_data
)

anova(
  fit_reduced,
  fit_full,
  test = "Chisq"
)

This evaluates whether adding treatment significantly improves the model's likelihood.

SAS Implementation

In SAS, multinomial logistic regression can be fit using PROC LOGISTIC with a generalized logit link.

proc logistic data=clinical_data;

    class treatment(ref="Control")
          severity(ref="Mild") / param=ref;

    model response(ref="PD") =
          treatment
          age
          severity
          / link=glogit;

run;

The LINK=GLOGIT option specifies the generalized logit model.

The response(ref="PD") specification identifies progressive disease as the reference category.

Obtaining Odds Ratios in SAS

SAS can produce odds ratios using the ODDSRATIO statement.

proc logistic data=clinical_data;

    class treatment(ref="Control")
          severity(ref="Mild") / param=ref;

    model response(ref="PD") =
          treatment
          age
          severity
          / link=glogit;

    oddsratio treatment;
    oddsratio age;

run;

The resulting output should be interpreted with careful attention to which outcome category is being compared with the reference.

Important SAS Output Consideration

For generalized logit models, SAS output may contain multiple sets of parameter estimates because each non-reference outcome category has its own equation.

For example, with PD as the reference:

The same predictor can therefore appear multiple times, once for each non-reference category.

Testing the Treatment Effect

Suppose the treatment effects are:

The results suggest that experimental treatment is associated with higher relative odds of CR versus PD and PR versus PD, while the evidence for SD versus PD is weaker.

However, these are three related comparisons. If the scientific question is whether treatment has an overall effect on the four-category response outcome, a global test should also be reported.

Multiple Comparisons

A multinomial model naturally generates multiple coefficients. With four outcome categories, there are three category-versus-reference comparisons for each binary predictor.

If many predictors and interactions are also tested, the number of hypothesis tests can become large.

Investigators should therefore distinguish between:

• The prespecified primary hypothesis
• Global predictor tests
• Individual category-specific comparisons
• Exploratory analyses

Multiplicity considerations should be incorporated into the statistical analysis plan when multiple formal inferences are being made.

Interactions

Multinomial logistic regression can also include interactions. Suppose treatment effect may differ by baseline severity. The model can include:

For example:

fit <- multinom(
  response ~ treatment * severity + age,
  data = clinical_data
)

This allows the treatment association to differ across severity groups.

Model Assumption: Linearity for Continuous Predictors

For a continuous predictor such as age, ordinary multinomial logistic regression assumes a linear relationship between age and each logit. For example:

This assumes a constant change in the log relative odds for each one-unit increase in age.

That assumption may be unrealistic.

Possible alternatives include:

• Restricted cubic splines
• Polynomial terms
• Clinically motivated categories
• Other flexible functional forms

Why Categorizing Continuous Variables Can Be Dangerous

It may be tempting to turn age into:

• <50
• 50–65
• >65

Categorization can simplify presentation, but it also discards information and may introduce artificial thresholds.

If age has a nonlinear relationship with the multinomial outcome, flexible continuous modeling is often preferable.

Sparse Data and Separation

Multinomial models can become unstable when some combinations of predictors and outcome categories are rare. For example, suppose complete response occurs in only a few patients and all of those patients happen to receive the experimental treatment. The model may have difficulty estimating the CR equation reliably.

Potential warning signs include:

• Very large coefficients
• Very large standard errors
• Extremely large odds ratios
• Convergence warnings
• Very wide confidence intervals
• Predicted probabilities extremely close to 0 or 1

Sample Size Considerations

There is no single universal sample-size formula for multinomial logistic regression that works for every clinical study. The required sample size depends on:

• Number of outcome categories
• Prevalence of each category
• Number of predictors
• Effect sizes
• Covariate distributions
• Interactions
• Desired precision or power

A simple rule such as "10 events per variable" should not be treated as a universal requirement for multinomial models.

Simulation-based planning can be particularly useful when the model is complex.

Missing Outcome Data

Multinomial regression generally requires the outcome category to be known for the patient included in the analysis. If outcome assessments are missing, investigators need a prespecified strategy.

Possible approaches include:

• Complete-case analysis
• Multiple imputation
• Prespecified missing-as-category approaches when scientifically justified
• Sensitivity analyses

The appropriate strategy depends on the endpoint, missingness mechanism, estimand, and clinical trial design.

Goodness of Fit

Model assessment should go beyond simply examining p-values. Useful considerations include:

• Likelihood-based measures
• AIC or other information criteria
• Calibration of predicted probabilities
• Observed versus predicted category frequencies
• Influential observations
• Residual diagnostics where available
• Clinical plausibility of fitted probabilities

Observed vs Predicted Frequencies

A useful descriptive check is to compare observed outcome frequencies with model-predicted frequencies. Suppose the observed proportions are:

This suggests reasonably close aggregate calibration, although aggregate agreement alone does not establish good individual-level calibration.

What Multinomial Logistic Regression Does Not Tell You

An odds ratio does not directly provide:

• Absolute risk difference
• Risk ratio
• Probability ratio
• Number needed to treat
• Individual patient outcome certainty

If the clinical question is about absolute probabilities, predicted probabilities or marginal standardization may be more useful.

Marginal Predicted Probabilities

Suppose investigators want to estimate the probability of each response category under treatment versus control while averaging over the observed distribution of other covariates. This can be accomplished using model-based marginal predictions.

Conceptually:

where \(\hat{P}_{ij}\) is the predicted probability for patient \(i\) and outcome category \(j\).

This can provide a more clinically interpretable summary than reporting only category-specific odds ratios.

Example of a Clinical Probability Summary

This presentation may be much easier for a clinical audience to interpret than a table containing only multinomial odds ratios.

Why Odds Ratios and Probabilities Can Tell Different Stories

The categories compete for probability mass. If treatment increases the probability of CR and PR, those increases necessarily come at the expense of probability assigned to other categories.

Therefore, changing one category can alter the probabilities of all other categories even when the interpretation of an individual coefficient is framed as a category-versus-reference comparison.

A Practical Reporting Table

A useful clinical report might present the category-specific odds ratios like this:

The table should explicitly state the reference outcome.

For example:

Reference outcome: Progressive disease.

Common Mistakes

1. Treating a nominal outcome as continuous. Coding CR, PR, SD, and PD as 1, 2, 3, and 4 and fitting ordinary linear regression imposes assumptions that are generally inappropriate.
2. Using binary logistic regression automatically. A multicategory endpoint requires a model that accounts for all categories unless a scientifically justified binary contrast is the actual endpoint.
3. Fitting multiple one-versus-rest logistic regressions without considering the joint outcome structure. This is not generally equivalent to a multinomial model.
4. Forgetting the reference category. Every multinomial odds ratio is relative to a specific outcome category.
5. Interpreting odds ratios as probability ratios. An OR of 2 does not mean the probability is twice as high.
6. Assuming all category-specific coefficients are equal. Multinomial models generally permit different predictor effects across outcome-category comparisons.
7. Ignoring the possibility of an ordinal model. If categories are genuinely ordered, ordinal regression may be more efficient and scientifically appropriate.
8. Ignoring sparse outcome categories. Very rare categories can produce unstable estimates and enormous standard errors.
9. Reporting only individual p-values. A global test may be needed to answer whether a predictor has an overall association with the multinomial outcome.
10. Ignoring multiplicity. Many categories and predictors can generate many hypothesis tests.
11. Ignoring nonlinear continuous effects. The default model assumes linearity on each logit scale.
12. Reporting predicted classes without probabilities. The most likely category does not communicate how certain the prediction is.

Multinomial Logistic Regression Workflow

How to Explain a Multinomial Model to Clinicians

A technically correct statistical explanation can still be difficult for a clinical audience. A useful strategy is to begin with the predicted probabilities.

For example:

After adjusting for baseline characteristics, the model estimated that the probability of complete response was 13% under experimental treatment compared with 6% under control.

Then provide the relative odds comparison:

The corresponding adjusted odds of complete response versus progressive disease were approximately three times higher under experimental treatment.

This gives the clinical audience both an absolute probability interpretation and a model-based relative effect.

Multinomial Regression in a Clinical Trial SAP

If multinomial logistic regression is prespecified as a primary or secondary analysis, the statistical analysis plan should clearly define:

• Outcome categories
• Reference outcome category
• Analysis population
• Predictors
• Categorical-variable reference levels
• Continuous-variable functional forms
• Interaction terms
• Missing-data handling
• Model-fitting method
• Hypothesis-testing strategy
• Multiplicity strategy
• Effect measures
• Confidence interval level
• Predicted-probability summaries, if applicable

Example SAP Language

A statistical analysis plan might specify that:

The actual SAP should of course reflect the specific endpoint, estimand, analysis population, and statistical objectives of the study.

Multinomial Regression vs. Binary Response Analysis

Sometimes a clinical trial has a multicategory endpoint but the primary analysis deliberately collapses it. For example:

Then binary logistic regression could be used for that specific binary endpoint.

That is not inherently wrong. The question is whether the collapsed endpoint is the scientifically intended estimand.

When Multinomial Logistic Regression Is Particularly Useful

• Best overall response has multiple nominal categories.
• A clinical outcome has several mutually exclusive states.
• The categories should not be assumed to follow a proportional-odds structure.
• Separate predictor effects for different outcome categories are scientifically important.
• Researchers want predicted probabilities for all categories simultaneously.
• A binary collapse of the outcome would discard clinically relevant information.

When Another Model May Be Better

• The outcome has only two categories → binary logistic regression.
• The categories are genuinely ordinal → consider ordinal logistic regression.
• The outcome represents repeated longitudinal states → consider a longitudinal or transition model.
• The outcome represents competing event times → consider competing-risks survival methods.
• The outcome categories are generated by a hierarchical process → consider a model reflecting that hierarchy.
• Events are extremely sparse → consider alternative modeling strategies or category definitions.

Key Mathematical Summary

For \(K\) nominal outcome categories with category \(K\) as the reference, the baseline-category multinomial model is:

The category-specific odds ratio for a one-unit increase in \(X_r\) is:

The predicted probability of category \(j\) is:

for \(j=1,\ldots,K-1\), while the reference-category probability is:

Complete Worked Example Summary

The Most Important Concept

The most important idea in multinomial logistic regression is simple: the model compares each non-reference outcome category with a selected reference category while allowing predictors to have different effects for different comparisons.

For a four-category endpoint:

are three separate logits. Each can have its own intercept and predictor coefficients.

The coefficients are interpreted through relative odds, while the complete model can be converted into predicted probabilities that sum to one.

References

Agresti, A. (2013). Categorical Data Analysis, 3rd ed. Wiley.
Hosmer, D.W., Lemeshow, S. & Sturdivant, R.X. (2013). Applied Logistic Regression, 3rd ed. Wiley.
Long, J.S. & Freese, J. (2014). Regression Models for Categorical Dependent Variables Using Stata, 3rd ed. Stata Press.
McCullagh, P. (1980). Regression models for ordinal data. Journal of the Royal Statistical Society: Series B, 42, 109–142.
Venables, W.N. & Ripley, B.D. (2002). Modern Applied Statistics with S, 4th ed. Springer.
Harrell, F.E. (2015). Regression Modeling Strategies, 2nd ed. Springer.