Discrete time Markov models of cognitive transitions: Assessing goodness of fit

Authors
Affiliations

Maynooth University

University of São Paulo

Maynooth University

Maynooth University

Latest version

March 26, 2026

Supplementary materials

Multinomial Representation of Discrete-Time Markov Models

In discrete-time settings, transition probabilities between states can be estimated using a multinomial logistic regression framework. Let \(Y_{i,t}\) denote the state occupied by individual \(i\) at time \(t\), where \(Y_{i,t} \in \{1, 2, \dots, K\}\). The probability of transitioning from state \(j\) at time \(t\) to state \(k\) at time \(t + 1\) can be expressed as:

\[ p_{jk}(t, t+1) = P(Y_{i, t+1} = k \; \vert \; Y_{i, t} = j). \]

When covariates \(\mathbf{z}_{i,t}\) are included these probabilities can be modelled using a multinomial logistic regression formulation. For non-reference categories \(k = 2, \dots, K\), the linear predictor is

\[ \eta_{i,t}^{(k)} = log(\frac{p_{jk} (\mathbf{z_{i,t}})}{p_{j1} (\mathbf{z_{i,t}})}) = \alpha_{jk} + \mathbf{z}^\intercal_{i,t} \pmb{\beta}_{jk} \qquad \text{for } K = 2, \dots, K. \tag{1}\]

Within the formulation, \(p_{j1}\) is the reference probability (e.g., remaining in state \(j\)), \(\alpha_{jk}\) is a state-pair specific intercept, and \(\beta_{jk}\) is a vector of coefficients for covariates \(\mathbf{z}_{i,t}\) (of which \(Y_{i,t}\) is inclusive).

The resulting transition probabilities are obtained from the multinomial link function such that the transition probabilities from state \(j\) are given by:

\[ p_{jk}(\mathbf{z}_{i,t}) = \frac{\exp(\eta^{(k)}_{i,t})}{1 + \sum_{k = 2}^{K} \exp(\eta^{(k)}_{i,t})}, \qquad \text{for } k = 2, \dots, K \tag{2}\]

and for the reference category,

\[ p_{j1}(\mathbf{z}_{i,t}) = \frac{1}{1 - \sum^{K}_{k=2}\exp(\eta^{(k)}_{i,t})}. \tag{3}\]

This formulation provides a convenient way to estimate covariate-dependent transition probabilities in discrete-time Markov models using standard multinomial regression software.