Models of visual word recognition have been assessed by both factorial and regression approaches. Factorial approaches tend to provide a relatively weak test of models, and regression approaches give little indication of the sources of models' mispredictions, especially when parameters are not optimal. A new alternative method, involving regression on model error, combines these two approaches with parameter optimization. The method is illustrated with respect to the dual route cascaded model of reading aloud. In contrast to previous investigations, this method provides clear evidence that there are parameter-independent problems with the model, and identifies two specific sources of misprediction made by model. (Contains 4 figures and 9 tables.)