Why use a multinomial model?
A nominal variable has categories without a meaningful order: for example, preferred appointment type, transport mode, or medication class. When such a variable is incomplete in a MICE analysis, its imputation model should represent the probabilities of its possible categories conditional on the other variables in the data. Treating category labels as numbers would impose an artificial order and distance; fitting a separate binary model for each category can instead produce probabilities that do not form one coherent distribution.
Baseline-category multinomial logistic regression models each non-reference category relative to one chosen reference category. It is suitable when the outcome categories are mutually exclusive and collectively exhaustive for the cases being modeled, and no genuine ordering is intended. Choosing a baseline gives a convenient parameterization; it does not make the baseline substantively privileged.
Here \(X=(x_1,\ldots,x_p)\) denotes the predictors, and \(P(Y=k\mid X=x)\) is the conditional probability of category \(k\) at predictor values \(x\). For a categorical predictor, indicators are used to represent its levels; its reference level must be stated. A continuous predictor enters linearly in each logit unless a transformation, nonlinear term, or interaction is specified.
What the coefficients mean
A coefficient is an effect on a log relative odds, not a direct change in probability. Holding the other predictors fixed, increasing \(x_j\) by one unit changes the category-\(k\)-versus-baseline logit by \(\beta_{kj}\). Exponentiating gives the corresponding multiplicative change in the relative odds.
This ratio compares category \(k\) with the baseline at two predictor values. It is not a probability and does not, by itself, say how much either category’s probability changes. Those probabilities also depend on the logits for all the other categories. At one fixed predictor profile, however, the relative odds \(P(Y=k\mid x)/P(Y=1\mid x)\) are exactly the ratio of those two category probabilities.
Worked Example: Interpreting a Continuous-Predictor Coefficient
Suppose a hypothetical clinic records a patient’s preferred appointment type as in-person (baseline), video, or telephone. Age is measured in years. A fitted model has an age coefficient of \(0.04\) for video versus in-person. Interpret this coefficient for a five-year age difference.
The model says that, at fixed values of the other predictors, a one-year increase in age changes the video-to-in-person log relative odds by \(0.04\). For five years, the change is \(5(0.04)=0.20\). Exponentiating gives \(\exp(0.20)\approx 1.221\).
Thus, the fitted video-to-in-person relative odds are about 1.221 times as large for a patient five years older, with the other predictors held fixed. This is a relative-odds comparison, not a claim that the probability of choosing video rises by 22.1 percentage points. The interpretation assumes the specified logit is linear in age over this five-year contrast and that the model’s other terms are held fixed.
From logits to predicted probabilities
The fitted logits jointly determine a probability for every category. If \(\eta_k(x)\) denotes the fitted category-\(k\)-versus-baseline linear predictor, set \(\eta_1(x)=0\). The corresponding probabilities can be expressed as \(P(Y=k\mid x)=\exp\{\eta_k(x)\}/\sum_{r=1}^{K}\exp\{\eta_r(x)\}\). This normalization makes each probability nonnegative and makes their sum equal one. In particular, a larger logit means a larger probability relative to the baseline, but it need not make that category the most probable overall: the other category logits also matter.
Worked Example: Comparing Three Predicted Category Probabilities
Consider a hypothetical fitted model with categories A, B, and C, using A as the baseline. At one predictor profile, suppose the fitted logits are \(0\) for A, \(0.1997\) for B versus A, and \(0\) for C versus A. These values are illustrative model outputs, not observations from a real study.
The unnormalized values are \(1\) for A, \(\exp(0.1997)\approx1.221\) for B, and \(1\) for C. Their sum is \(1+1.221+1=3.221\). Therefore,
The probabilities sum to approximately \(1.000\), allowing for rounding, and B is the most probable category at this profile. Its probability is about 1.221 times A’s probability, since \(0.3791/0.3105\approx1.221\). The value 1.221 is not itself a probability; it is the fitted B-to-A relative odds, which at this same profile equals the ratio of the B and A probabilities.
Worked Example: Interpreting Separate Category Effects
In a hypothetical survey, the nominal outcome is preferred insurance plan: HMO (baseline), PPO, or fee-for-service. Suppose age, measured in years, has a fitted coefficient of \(0.04\) in the PPO-versus-HMO logit and \(-0.02\) in the fee-for-service-versus-HMO logit. Consider a five-year increase in age, with other predictors held fixed.
For PPO versus HMO, the log relative odds change by \(5(0.04)=0.20\), so the relative odds are multiplied by \(\exp(0.20)\approx1.221\). For fee-for-service versus HMO, the log relative odds change by \(5(-0.02)=-0.10\), giving a multiplier of \(\exp(-0.10)\approx0.9048\). Thus, the model associates five additional years of age with about 1.221 times the PPO-to-HMO relative odds and 0.9048 times the fee-for-service-to-HMO relative odds, conditional on the other predictors.
The coefficients have different signs because each compares a different category with HMO. Neither one alone gives a probability change, and the two comparisons do not imply that one plan is chosen because of age. The interpretations rely on the model’s linear age effect in both logits and on the chosen predictor specification.
Assumptions and use in MICE
For the regression model to represent the conditional distribution adequately, the outcome must be recorded in well-defined, mutually exclusive categories, and the specified predictors and functional forms must be appropriate. The model assumes linearity in the logit for each continuous predictor as entered; it does not require a linear relationship between a predictor and the category probabilities. Observations are ordinarily treated as conditionally independent given the predictors, unless clustering or repeated measurements are handled in the model. There must also be enough information to estimate the category-specific coefficients without non-identifiability or severe instability.
In MICE, the multinomial regression is a conditional imputation model for a nominal variable. It is fitted using cases where that variable is observed, with the other analysis variables serving as predictors according to the chosen specification. During an update, missing values are drawn from the fitted conditional category distribution rather than being replaced deterministically by the most likely category. This preserves categorical uncertainty. The regression form alone does not guarantee valid multiple imputation: the broader missing-data assumptions and suitable representation of uncertainty in the imputation procedure also matter.
A model suitable for imputation should include predictors that help explain the incomplete variable and, as appropriate, variables associated with its missingness or the substantive analysis. Predictions should be plausible across the predictor patterns where imputations are needed. The fitted model should not be treated as a causal model: its coefficients describe conditional associations under the model, not effects of interventions.
Diagnostics, alternatives, and limitations
Before relying on the model, examine category counts and cross-tabulations for predictors, confirm that every modeled outcome category is present, and check fitting warnings, convergence, and coefficient uncertainty. Very large coefficients or standard errors, unstable estimates across fitting attempts, or predicted probabilities extremely close to zero or one may indicate sparse patterns or separation. These signs matter because an unstable conditional distribution can yield implausible or overly deterministic imputations. Also compare observed category proportions with the model’s predicted proportions across useful predictor groups, and investigate whether continuous predictors need nonlinear terms or interactions.
The key structural limitation is that each category’s log relative odds is modeled as linear in the included terms. A standard multinomial-logit specification also implies that the log odds of any two categories can be written as a difference between their baseline logits. If categories are close substitutes or have other structure this relationship cannot represent, the model may be too restrictive. Predictions should therefore be checked for plausibility, not accepted solely because the fitting routine converged.
| Approach | When it is appropriate | Important distinction |
|---|---|---|
| Baseline-category multinomial logistic regression | Outcome has three or more unordered categories | Models each non-baseline category relative to the chosen baseline |
| Binary logistic regression | Outcome has exactly two categories | There is one logit, rather than a set of category-specific logits |
| Ordinal logistic regression | Categories have a meaningful order and its assumptions are suitable | Uses the ordering; it is not interchangeable with a nominal model |
| Separate one-versus-rest binary models | May be used for distinct classification objectives | Fitted probabilities need not sum to one, so they do not automatically define a coherent imputation distribution |
If the outcome has a genuine order, an ordinal model may use that information more efficiently when its structure is appropriate. If it has only two categories, binary logistic regression is the simpler model. Conversely, assigning integer scores to unordered categories and fitting a standard regression would make arbitrary numerical distances part of the model, which is generally unjustified.
Common mistake
A useful check is to keep the three quantities distinct: \(\beta_{kj}\) is a log-relative-odds coefficient, \(\exp(\beta_{kj})\) is a relative-odds multiplier for a one-unit predictor change, and \(P(Y=k\mid x)\) is a conditional probability at a particular predictor profile. Interchanging these quantities is a frequent source of incorrect interpretations.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- In a three-category model with category A as baseline, what does \(\beta_{B,1}=0.30\) mean for a one-unit increase in predictor \(x_1\), with other predictors fixed?
- At a given profile, suppose the logits for categories A, B, and C are \(0\), \(\log(2)\), and \(0\), respectively. What are the three predicted probabilities?
- Why is a multinomial model usually preferable to coding unordered categories as 1, 2, and 3 and fitting a linear regression?
- Name two diagnostics that could signal an unreliable nominal imputation model.
Solutions. (1) The B-to-A log relative odds increase by \(0.30\); the relative odds are multiplied by \(\exp(0.30)\approx1.350\). This is not a 30-percentage-point probability increase. (2) The unnormalized values are \(1,2,1\), with sum \(4\), so the probabilities are \(0.25,0.50,0.25\). (3) The numbers would impose an artificial order and distances between labels, while multinomial regression models the unordered categories directly. (4) Examples include convergence warnings, large unstable coefficients or standard errors, sparse outcome-predictor combinations, and implausibly extreme predicted probabilities.