Written and maintained by CASRAI Editorial Board
Last updated
Binary logistic regression gives you one odds ratio per predictor. Multinomial logistic regression gives you a whole set of them — one full set of coefficients for every outcome category except a single reference category — and that structural difference is exactly where most misreadings start. A relative risk ratio (RRR) from a multinomial model looks like an odds ratio, is calculated the same way, and is routinely interpreted as if it answers the same question. It doesn’t. It answers a narrower one: how the relative risk of landing in this specific category versus the reference category changes with the predictor — nothing about the other non-reference categories, and nothing about “this category versus everyone else.”
This guide covers how to set up a multinomial logit model, how to choose a reference category deliberately rather than by software default, how to read RRR coefficients without falling into the single-odds-ratio habit binary logistic regression trains into you, and a worked three-category example that shows where the interpretation actually goes wrong in practice.
When you need multinomial logistic regression
Use multinomial (also called “polytomous” or “generalized”) logistic regression when the outcome variable has three or more categories with no natural order — a person’s chosen employment sector, a manuscript’s disposition (published, revised, rejected without option to resubmit), a patient’s discharge destination. If the categories are ordered (a Likert-style severity scale, a satisfaction rating), an ordinal model with a proportional-odds structure is the better-fitting choice, not multinomial logit run on ranked categories — treating order as unordered throws away information the ordinal model uses.
Setting up the model
Multinomial logistic regression is a direct extension of the binary logit model, fitted as a set of K−1 simultaneous binary comparisons against one held-out category, where K is the number of outcome categories. For outcome category j compared against reference category r, the model estimates:
ln[ P(Y=j) / P(Y=r) ] = β0j + β1jX1 + β2jX2 + … + βpjXp
Exponentiating a coefficient, exp(β1j), gives the relative risk ratio for that predictor on the log-odds of category j versus the reference category. With a three-category outcome and reference category set, the model fits two complete equations — one for category 2 vs. the reference, one for category 3 vs. the reference — each with its own intercept and full set of predictor coefficients. A model with five predictors and a four-category outcome therefore estimates three separate coefficient vectors, not one: the output is three times longer than a binary logistic regression on the same predictors, and every coefficient’s meaning is anchored to whichever category was excluded as the baseline.
Estimation is by maximum likelihood, same as binary logistic regression, and the same practical constraints carry over: watch for separation (a predictor that perfectly or near-perfectly predicts membership in one category, which drives that category’s coefficients toward infinity) and keep an eye on cases-per-category — a common rule of thumb is at least 10–20 observations per predictor within the smallest outcome category, since the K−1 comparisons are only as stable as their thinnest category.
Choosing the reference category — a decision, not a default
Statistical software has to pick a reference category to fit the model at all, and left alone most packages pick it mechanically: the last category by value or alphabetical order (SPSS defaults to the highest-valued category unless told otherwise), the first level in factor order (R’s default for a factor variable is whichever level sorts first), or a base-level option that has to be set explicitly (Stata’s mlogit takes a baseoutcome() option; SAS’s PROC LOGISTIC takes REF=; R’s relevel() or nnet::multinom‘s factor-level order does the same job). None of those defaults are chosen for substantive reasons, and none of them are the right choice automatically.
Pick the reference category on purpose, using one of two defensible logics:
- The most substantively meaningful comparison. If one category is the “status quo” or expected outcome — continuing in an academic track, remaining employed, no adverse event — comparisons against it are what a reader actually wants to see, because every other category’s coefficients then read as “risk of deviating from the expected outcome.”
- The largest category. A bigger reference group gives every non-reference comparison more precision (a smaller standard error on each coefficient), which matters when the outcome is unbalanced across categories — a common, defensible default when there is no clearly “expected” category.
Whichever logic you use, state the reference category explicitly in the write-up, in the same sentence as the first RRR you report, not buried in a methods footnote. A reader who doesn’t know the baseline cannot interpret a single coefficient in the results table, and re-deriving it from context is exactly the kind of extra step that gets skipped, guessed at, or gotten wrong.
The interpretation trap: RRR is not “the” odds ratio
This is where binary logistic regression’s habits actively mislead. In a binary model there is one outcome comparison (event vs. no event) and therefore one odds ratio per predictor, full stop — readers of binary logistic regression output learn to treat “the odds ratio for X” as a single, unambiguous number. Multinomial output has no such single number. Every predictor has K−1 RRRs, one per non-reference category, and each one is a claim about that category specifically relative to the reference category — not about that category versus every other category pooled, and not a claim you can read directly against a different non-reference category.
Three specific misreadings show up repeatedly in practice:
- Treating an RRR as “category j vs. everyone else.” It isn’t. RRRj compares category j against the reference category only. A predictor can have a large RRR for category j vs. the reference and still say nothing about category j vs. category k — that is a different, unfitted comparison (see below).
- Comparing two RRRs from the same model as if they were directly comparable across categories. RRR2 = 3.1 and RRR3 = 1.4 (both vs. the same reference) do not mean “category 2 is roughly 2.2 times more strongly associated with X than category 3 is” in any calibrated sense — they’re two separate log-odds comparisons against a shared baseline, not a ranked effect size. Reading them as a ladder is a common but ungrounded step.
- Wanting the comparison between two non-reference categories directly. If you need category 2 vs. category 3 specifically, that coefficient is not in the output as fitted — you either refit the model with category 3 as the reference (algebraically, RRR2 vs 3 = RRR2 vs r / RRR3 vs r, so it can also be recovered by division without refitting) or use a post-estimation contrast/margins command that computes it directly. Reporting the reference-anchored coefficient as if it already answered a non-reference-vs-non-reference question is the single most common error in applied write-ups of multinomial models.
The correct reading of a single coefficient, stated in full: “Holding other predictors constant, a one-unit increase in X multiplies the relative risk of being in category j rather than the reference category by exp(βj).” Every clause in that sentence is load-bearing — drop “rather than the reference category” and the claim becomes unsupported by the model that produced it.
The independence of irrelevant alternatives (IIA) assumption
Multinomial logit assumes the relative odds between any two categories are unaffected by the presence or attributes of other categories — independence of irrelevant alternatives. The standard illustration is the “red bus / blue bus” problem: if a commuter chooses between car and red bus with some odds, and an identical blue bus is introduced, IIA implies the car-vs.-red-bus odds shouldn’t change — but in practice the two bus options split ridership between themselves rather than drawing proportionally from car users, which violates the assumption. In research contexts the same failure mode shows up when outcome categories are not truly distinct alternatives but partial substitutes for each other (two similar treatment arms, two adjacent job-classification categories).
IIA is testable, most commonly with a Hausman-McFadden test: refit the model after dropping one outcome category and compare coefficients against the full model; a significant difference is evidence against IIA (though the test is known to behave inconsistently in finite samples, so a non-significant result is weak reassurance, not proof). If IIA is genuinely violated, the practical alternatives are a nested logit model (when categories group naturally, e.g. two sub-types within one broader category) or a multinomial probit model, which does not assume IIA but is more computationally demanding and less commonly implemented in standard software.
Worked example: a three-category outcome
The figures below are a constructed illustration for walking through the arithmetic, not results from a real study or dataset.
Suppose a sample of doctoral graduates is followed up one year after graduation, with an outcome recorded in three categories: (1) academic postdoctoral/faculty-track position [set as the reference category, using the “expected outcome” logic above], (2) industry research role, and (3) non-research role. The predictor of interest is number of peer-reviewed first-author publications during the doctorate, entered as a continuous variable.
The fitted model returns two coefficient sets, both referenced against category 1:
- Industry research role vs. academic (reference): β = −0.18, RRR = exp(−0.18) ≈ 0.84
- Non-research role vs. academic (reference): β = −0.42, RRR = exp(−0.42) ≈ 0.66
Correct reading: each additional first-author publication is associated with the relative risk of ending up in an industry research role, rather than an academic role, being multiplied by about 0.84 (a 16% relative decrease), holding other predictors constant. Each additional publication is associated with the relative risk of a non-research outcome, rather than an academic outcome, being multiplied by about 0.66 (a 34% relative decrease). Both coefficients say something about deviating from the academic-track baseline as publication count rises; neither says anything, as fitted, about industry vs. non-research directly.
Incorrect reading to avoid: “non-research roles are associated with a stronger publication effect than industry roles” as a general two-category ranking claim, or “publications reduce the odds of a non-academic outcome overall” as if there were one pooled non-academic category — the model was never fitted against a pooled non-academic category, only against these two specific comparisons.
If the industry-vs.-non-research contrast is the one actually of interest, it is recoverable without refitting: RRRindustry vs non-research = RRRindustry vs academic / RRRnon-research vs academic = 0.84 / 0.66 ≈ 1.27 — each additional publication multiplies the relative risk of ending up in industry rather than a non-research role by about 1.27, a comparison the two reference-anchored coefficients above do not give you directly.
Reporting checklist
- State the reference category explicitly, in the same sentence as the first reported RRR.
- Report RRR with its confidence interval, not just a point estimate and p-value — the same reporting standard as binary odds ratios.
- Report a fit statistic (McFadden’s pseudo-R² is the most commonly reported for multinomial logit; there is no single R² analogue as clean as OLS’s).
- Report the per-category sample sizes, since the K−1 comparisons are only as stable as the smallest category.
- Note whether IIA was tested and what the result was, especially if outcome categories could plausibly be partial substitutes for each other.
- Never report a single “odds ratio” for a multinomial predictor without naming which non-reference category it belongs to.
Frequently asked questions
Is a relative risk ratio the same thing as an odds ratio?
Mechanically, yes — an RRR from a multinomial logit model is calculated by exponentiating a coefficient exactly the way a binary logistic regression’s odds ratio is. The difference is scope: a binary model’s odds ratio compares the only two outcomes there are, so no baseline needs stating. A multinomial RRR always compares one specific non-reference category against one specific reference category, so it is only interpretable once that reference category is named.
What happens if I don’t set the reference category deliberately?
The software still fits a model — nothing errors out — but it does so against whichever category the default happened to select (often the last category by value, or the first by factor order), which is frequently not the comparison a reader would find meaningful. The coefficients are not wrong, but the results table becomes harder to interpret and easy to misreport, since the reference category has to be reverse-engineered from software defaults rather than stated as a deliberate choice.
Can I compare two non-reference categories directly without refitting the model?
Yes, by dividing their RRRs (each already expressed against the same reference): RRRa vs b = RRRa vs ref / RRRb vs ref. Most statistical packages can also produce this directly via a post-estimation contrast or margins command, which additionally gives a standard error and confidence interval for the derived comparison rather than requiring you to compute one by hand.
How is multinomial logistic regression different from running several separate binary logistic regressions?
Running separate binary models (category 2 vs. reference, category 3 vs. reference, each excluding the other non-reference category’s rows) uses less of the data per model and can give inconsistent coefficient estimates, because each binary model is fit independently rather than jointly. Multinomial logistic regression estimates all K−1 comparisons simultaneously from the full dataset in one maximum-likelihood fit, which is more efficient and is the standard approach when the full set of categories is meaningful together.
For the binary case this model generalizes from, see logistic regression and the logit model. For the broader diagnostic and reporting conventions regression models share, see the assumptions behind regression analysis. Software-specific implementations are covered separately: logistic regression in R, logistic regression in Stata, and binary logistic regression in SPSS. For the odds-ratio interpretation this model extends, see case-control study design and odds ratios and the chi-square test for the categorical-association test multinomial models often follow.
