Clustered ordinal responses, which are commonplace in behavioural and educational research, are often analysed using mixed-effects ordinal probit models. Likelihood-based inference for these models can be computationally burdensome, and may compromise the consistency of estimators if the model is misspecified. We propose an alternative inferential approach based on generalized estimating equations. We show that systems of estimating equations can be specified for mixed-effects ordinal probit models that avoid the potentially heavy computational demands of maximum likelihood estimation, and can also provide inferences that are robust with respect to some forms of model misspecification--particularly serial effects in longitudinal data.