Pedestrian choice between co-located stairs and escalators in Toronto transit stations was modelled using a set of standard binary and mixed-logit models, incorporating dynamic variables like crowding. While all models had good fit, the ascending direction and restricted-mobility individual choice were more readily predicted. Performance was measured predictive ability of ten-second aggregate flows after implementation in the pedestrian simulator MassMotion. The mixed-logit models performed consistently better than the standard models, with all showing good predictive ability (nearing 90%). There were also significant spreads of accuracy of up to 10% when the way the models were applied by simulation agents was varied.