Mixed-Effects Multinomial Regression

mixed-effectsrandom-effectslongitudinal-dataclustered-dataglmmmaximum-marginal-likelihoodquadraturenominal-datamultilevel-modelmissing-dataempirical-bayes

Definition

The mixed-effects multinomial logistic regression model extends the standard multinomial logit to clustered or longitudinal data by adding subject-specific random effects. For the jjth observation within level-2 unit (subject/cluster) ii and response category cc, the conditional probability is pijc=exp(zijc)/h=1Cexp(zijh)p_{ijc} = \exp(z_{ijc})/\sum_{h=1}^C \exp(z_{ijh}), where zijc=wijαc+xijβicz_{ijc} = \mathbf{w}_{ij}'\boldsymbol{\alpha}_c + \mathbf{x}_{ij}'\boldsymbol{\beta}_{ic}, αc\boldsymbol{\alpha}_c is the fixed-effect vector for contrast cc, and βic=Tcθi\boldsymbol{\beta}_{ic} = \mathbf{T}_c\boldsymbol{\theta}_i with θiN(0,I)\boldsymbol{\theta}_i \sim N(\mathbf{0}, \mathbf{I}) and Σc=TcTc\boldsymbol{\Sigma}_c = \mathbf{T}_c\mathbf{T}_c' (Cholesky). Unlike the Gaussian Mixed-Effects Regression, the marginal likelihood has no closed form and must be approximated numerically.

Key Ideas

How It Works

  1. Conditional likelihood: Given random effects θi\boldsymbol{\theta}_i, observations within subject ii are conditionally independent multinomial draws. The product over time points jj and indicator vector yijc{0,1}y_{ijc} \in \{0,1\} gives (yiθi)=jcpijcyijc\ell(\mathbf{y}_i | \boldsymbol{\theta}_i) = \prod_j \prod_c p_{ijc}^{y_{ijc}}.
  2. Marginal likelihood: h(yi)=(yiθ)g(θ)dθh(\mathbf{y}_i) = \int \ell(\mathbf{y}_i | \boldsymbol{\theta})g(\boldsymbol{\theta})\,d\boldsymbol{\theta}, where g(θ)=N(0,I)g(\boldsymbol{\theta}) = N(\mathbf{0}, \mathbf{I}). This integral has no closed form (contrast with the Gaussian mixed regression model (MRM)); Gauss-Hermite quadrature with QrQ^r points approximates it, where rr is the number of random effects per contrast.
  3. Fisher scoring: Gradient and information matrix with respect to (Γ,Λ)(\boldsymbol{\Gamma}, \boldsymbol{\Lambda}) are computed by summing over subjects and quadrature points; parameters are updated iteratively until convergence.
  4. EAP estimates: Empirical Bayes (EAP) subject-specific random-effect estimates θ~i=E(θiyi)\tilde{\boldsymbol{\theta}}_i = E(\boldsymbol{\theta}_i | \mathbf{y}_i) are recovered via one additional quadrature pass after convergence.
  5. Marginal probabilities: Subject-specific probabilities p^ijcss\hat{p}^{ss}_{ijc} are integrated over g(θ)g(\boldsymbol{\theta}) to obtain population-averaged (marginal) probabilities; for random-intercept models an analytic approximation divides Γ^\hat{\boldsymbol{\Gamma}} by (163/15π)2σ^c2+1\sqrt{(16\sqrt{3}/15\pi)^2\hat{\sigma}_c^2 + 1}.
  6. Quadrature scaling: Standard Gauss-Hermite requires QrQ^r evaluations; for r>5r > 5 or so, adaptive quadrature (Liu-Pierce 1994; Pinheiro-Bates 1995) or Markov Chain Monte Carlo (MCMC) via the Gibbs Sampler is preferred.

Why It Matters

Open Questions

Related