Tutorials › Generalized Linear Models › From Research Question To Multinomial Contrast

Home  /  Tutorials  /  Multinomial Logit Models

Multinomial Logit Models · Tutorial 506 of 981

From Research Question To Multinomial Contrast

Connect the research question to a baseline, category-specific predictors, a probability contrast, and an appropriate joint test.

Intermediate 10 min read

What You'll Learn

  • Choose a baseline category that makes the research question easy to express
  • Specify category-specific logits and predictors for a multinomial response
  • Calculate and interpret a probability contrast at two covariate profiles
  • Distinguish a probability comparison from a test of category-specific coefficients
  • Formulate and carry out a joint Wald test for an overall predictor effect
  • Explain why changing the baseline does not change fitted category probabilities

From a Research Question to a Statistical Contrast

A research question about a categorical outcome often asks whether groups differ in the probability of a particular outcome, or whether a predictor is related to the outcome at all. These are related but distinct questions. A probability contrast describes the difference of interest; a hypothesis test evaluates a null claim about model parameters. Moving carefully between them prevents a common error: testing one coefficient when the scientific question concerns several outcome categories.

For a nominal response with \(K\) categories, a baseline-category multinomial logit represents the log odds of each nonbaseline category relative to a chosen baseline. The choice of baseline changes how coefficients are written, but not the probabilities or the scientific contrast. The first task is therefore to name the outcome categories and choose a reference that makes the comparisons interpretable.

Definition: Let \(Y\) take values \(0,1,\ldots,K-1\), and let \(x\) be a covariate vector. With category \(0\) as the baseline, the baseline-category logit model is $$ \log\left\{\frac{P(Y=j\mid x)}{P(Y=0\mid x)}\right\} = \eta_j(x)=x^\mathsf{T}\beta_j,\qquad j=1,\ldots,K-1. $$ Its probabilities are \(P(Y=0\mid x)=1/\{1+\sum_{h=1}^{K-1}\exp(\eta_h(x))\}\) and \(P(Y=j\mid x)=\exp(\eta_j(x))P(Y=0\mid x)\). Each \(\beta_j\) is a category-specific coefficient vector.

A research question should be made operational by specifying both the outcome quantity and the covariate values being compared. For example, “Does treatment increase the probability of full recovery?” is a question about a difference in \(P(Y=\text{full recovery}\mid x)\) between treatment profiles. “Is treatment associated with the outcome distribution?” instead calls for a joint statement about treatment coefficients across all nonbaseline logits.

Worked Example: Translate a Recovery Question into a Probability Contrast

Worked Example: Translate a Recovery Question into a Probability Contrast

Suppose \(Y\) records no improvement (\(0\)), partial improvement (\(1\)), or full improvement (\(2\)). Let \(x=1\) denote treatment and \(x=0\) control, and use no improvement as the baseline. Consider the fitted logits $$ \log\frac{P(Y=1\mid x)}{P(Y=0\mid x)}=-0.9+0.5x,\qquad \log\frac{P(Y=2\mid x)}{P(Y=0\mid x)}=-1.5+0.3x. $$ The research question is whether treatment changes the probability of full improvement.

At \(x=1\), the two nonbaseline logits are \(-0.4\) and \(-1.2\). The denominator for converting logits to probabilities is \(1+\exp(-0.4)+\exp(-1.2)\), which is approximately \(1.971514\). Thus $$ \big(P(Y=0\mid 1),P(Y=1\mid 1),P(Y=2\mid 1)\big) \approx (0.50722,\;0.34000,\;0.15277). $$ At \(x=0\), the logits are \(-0.9\) and \(-1.5\), and the denominator is \(1+\exp(-0.9)+\exp(-1.5)\approx1.629700\). The probabilities are approximately \((0.61361,\;0.24947,\;0.13692)\).

The target contrast is full-improvement probability under treatment minus that under control: \(0.15277-0.13692\approx0.01585\), or about 1.59 percentage points. The other category contrasts are approximately \(0.34000-0.24947=0.09053\) for partial improvement and \(0.50722-0.61361=-0.10639\) for no improvement. Their sum is zero, as it must be because the probabilities in each profile sum to one.

The fitted treatment slopes are \(0.5\) and \(0.3\), but neither is itself the probability contrast. They describe changes in log odds relative to no improvement. The desired probability difference requires transforming both logits jointly into the category probabilities at each profile, then subtracting the probabilities of interest.

Choosing the Baseline and the Predictors

The baseline is a reference for expressing odds, not a claim that its category is unimportant. Choose a category with a clear substantive interpretation and ensure that each predictor in the model corresponds to the comparison asked about. If treatment is binary, it can appear in each nonbaseline logit with a distinct coefficient. If the question concerns treatment’s overall relationship to the outcome, the joint null should set to zero all model terms involving treatment that are relevant to the question, including treatment interactions when present; a question about treatment association at a particular covariate profile instead requires a null formulated for that profile.

For a continuous predictor \(z\), the coefficient in logit \(j\) is the change in log odds of category \(j\) versus baseline per unit increase in \(z\), holding the other predictors fixed. With interactions, the treatment effect or predictor slope depends on the values of the interacting covariates. Therefore, a probability contrast must state the covariate profile at which it is evaluated. A contrast averaged over a population is a different target and requires specifying that population and averaging rule.

Result: Changing the baseline preserves probabilities. Suppose the model is written relative to category \(0\), with logits \(\eta_j(x)=\log\{p_j(x)/p_0(x)\}\), where \(p_j(x)=P(Y=j\mid x)>0\). If category \(r\) is instead the baseline, the new logit for category \(j\ne r\) is \(\eta_j(x)-\eta_r(x)\) when \(j\ne0,r\), and the logit for old category \(0\) is \(-\eta_r(x)\). These are exactly \(\log\{p_j(x)/p_r(x)\}\) and \(\log\{p_0(x)/p_r(x)\}\). Reconstructing probabilities from these ratios gives \(p_j(x)\) for every category: the ratios determine all probabilities up to a common multiplier, and requiring their sum to be one fixes that multiplier. Thus a baseline change alters the coefficient representation, not the fitted probability distribution.

Worked Example: Re-express the Recovery Probabilities with a New Baseline

Use the treatment-profile logits from the preceding example at \(x=1\): \(\eta_1=-0.4\) and \(\eta_2=-1.2\), with category \(0\) as baseline. If full improvement (category \(2\)) becomes the baseline, the log odds of no improvement versus full improvement are \(-\eta_2=1.2\). The log odds of partial versus full improvement are \(\eta_1-\eta_2=-0.4-(-1.2)=0.8\).

The new baseline-category denominator is \(1+\exp(1.2)+\exp(0.8)\), approximately \(6.545658\). The resulting probabilities for full, no, and partial improvement are approximately \(1/6.545658=0.1528\), \(\exp(1.2)/6.545658=0.5072\), and \(\exp(0.8)/6.545658=0.3400\), respectively. These match the earlier probabilities, up to rounding; only the reference category and logit coefficients changed.

Matching a Joint Test to an Overall Research Question

Suppose a predictor \(z\) has a category-specific slope \(\beta_{jz}\) in each nonbaseline logit. If the research question is whether \(z\) is associated with the categorical outcome at all, the null hypothesis is $$ H_0:\ \beta_{1z}=\cdots=\beta_{K-1,z}=0. $$ This is a joint hypothesis with \(K-1\) restrictions. It is not the same as testing one slope, and it is not a test of whether one probability contrast equals zero at one selected covariate profile.

For an estimated coefficient vector \(\hat\beta\), a linear null \(H_0:R\beta=r\) can be assessed by a Wald statistic. If the null has \(q\) independent restrictions and the covariance estimate is \(\widehat V\), the statistic is $$ W=(R\hat\beta-r)^\mathsf{T}\{R\widehat V R^\mathsf{T}\}^{-1}(R\hat\beta-r). $$ Under the null and standard regularity conditions, \(W\) is approximately distributed as \(\chi^2_q\). The number \(q\) is the rank of \(R\), not necessarily the number of rows if some restrictions are redundant.

Theorem: Asymptotic justification for the joint Wald test. Suppose the model is identifiable, the true parameter is an interior point, the information matrix is nonsingular, and the regularity conditions for maximum-likelihood asymptotic normality hold. For a fixed-rank \(q\times p\) matrix \(R\), if \(H_0:R\beta=r\) is true and \(\widehat V\) consistently estimates the asymptotic covariance of \(\hat\beta\), then \(W\xrightarrow{d}\chi^2_q\).

Proof. Under the stated conditions, \(\sqrt n(\hat\beta-\beta)\) converges in distribution to a mean-zero normal vector with asymptotic covariance \(\Sigma\), and \(n\widehat V\xrightarrow{p}\Sigma\). Under the null, \(R\beta-r=0\), so \(\sqrt n(R\hat\beta-r)\) converges to a normal vector with mean zero and covariance \(R\Sigma R^\mathsf{T}\). Since \(R\) has rank \(q\) and \(\Sigma\) is positive definite, this covariance is nonsingular. By consistency and Slutsky’s theorem, the Wald statistic converges in distribution to the quadratic form of that normal vector using its covariance inverse. Whiten the vector by multiplying by \((R\Sigma R^\mathsf{T})^{-1/2}\); its components are independent standard normal variables. The quadratic form is therefore the sum of their \(q\) squared values, which has a \(\chi^2_q\) distribution. \(\square\)

Worked Example: Test Whether Treatment Is Related to the Outcome Overall

For the three-category response, let \(\beta_{1T}\) and \(\beta_{2T}\) be the treatment slopes in the partial-improvement and full-improvement logits relative to no improvement. Suppose their estimates are \(0.20\) and \(0.15\), and their estimated covariance matrix is approximately \(0.0107230 I_2\). The scientific question is whether treatment is associated with the outcome distribution at all, so test \(H_0:\beta_{1T}=\beta_{2T}=0\) against the alternative that at least one is nonzero.

Here \(R=I_2\), \(r=(0,0)^\mathsf{T}\), and \(q=2\). The Wald statistic is $$ W=\frac{0.20^2+0.15^2}{0.0107230} =\frac{0.0625}{0.0107230}\approx5.8286. $$ For two degrees of freedom the upper-tail probability is \(\exp(-W/2)\), so \(p\approx\exp(-2.9143)\approx0.0542\). At a 5% significance level, this test does not reject the joint null. That conclusion concerns the overall treatment slopes; it does not prove that every treatment-versus-control probability contrast is zero.

If the prespecified question were instead whether treatment changes the probability of full improvement at a particular covariate profile, the target would be that probability difference. The joint slope test above is a useful overall association test, but it is not automatically the test of that profile-specific probability contrast.

How to Build the Analysis from the Question

1
Name the outcome and target.
State which categories are possible and whether the question concerns a particular category probability, a probability difference, or any association with the outcome distribution.
2
Choose a baseline.
Select a meaningful reference category, then write one logit for every other category against it.
3
Specify predictors and profiles.
Include the predictors relevant to the question, retaining category-specific effects where appropriate. For probability contrasts, state the covariate values being compared.
4
Calculate the estimand or formulate the null.
Convert all logits at each profile into probabilities before taking a probability difference. For an overall predictor test, set all of that predictor’s nonbaseline-logit coefficients to zero jointly.
5
Report the result in context.
Give the probability contrast or test statistic, its degrees of freedom and p-value, and a conclusion framed in the original research question.

Common Mistakes and Why the Distinction Matters

  • Reading a logit coefficient as a probability change. A coefficient is a change in a log odds, not a percentage-point change. Calculate the probabilities at the specified profiles before taking their difference.
  • Testing only one category coefficient for an overall question. A predictor can affect different nonbaseline logits in different ways. An overall association question requires a joint null covering the relevant category-specific coefficients.
  • Changing the baseline and claiming the conclusions must change. The coefficient values and their interpretation relative to the reference do change, but fitted probabilities and correctly defined probability contrasts do not.
  • Leaving the covariate profile unstated. With interactions, the effect of a predictor can depend on other covariates. A conditional probability contrast is incomplete until those values are specified.

The distinction between a scientific estimand and a model test makes results easier to interpret. A probability contrast answers a direct question about predicted outcome frequencies at stated profiles. A joint test addresses whether a set of model coefficients is compatible with a specified null. Both may be useful in the same analysis, but they should not be described as interchangeable.

Check Your Understanding

Try these questions before moving on.

  1. For a four-category response with category \(0\) as baseline, how many logits are specified, and how many restrictions are in a joint test that all slopes for one predictor equal zero?
  2. Why is a treatment coefficient in one baseline-category logit not generally the treatment-associated change in the probability of that category?
  3. A researcher asks whether treatment changes the probability of category \(2\) at a fixed covariate profile. What quantity should be calculated, and how is it defined?
  4. Explain why changing the baseline category cannot change fitted probabilities when the model is re-expressed consistently.
  5. A Wald test for all category-specific slopes of a predictor is not significant. Does this establish that every category probability is identical across predictor values? Explain.

Previous

Next

Back to Tutorials
Clinical Biostats