watched, (c) is watching, (d) was watching, (e) has watched, (f) had watched, (8) has been watching, and (i) had been watching.
As in Hodge and Pollack's data, an improper solution was obtained from the OVLP model. This was due to large proportions of errors in the two data sets. The OVLP model requires pi/i ≥ ∑jpi/j ( J. E. K. Smith, 1980). The difference between the improper solution and the constrained proper solution is much larger, however, in Clark and Stafford's data than in Hodge and Pollack's data. The informed guessing model may be a better choice under this circumstance. However, the informed guessing model suffers from a different kind of problem; model parameters in the informed guessing model are not uniquely determined.
Nakatani's confusion-choice model worked reasonably well. However, the minimum AIC solution was found to be the two-dimensional unsquared Euclidean distance-choice model (AIC = 5.8). The two dimensions in this model roughly corresponded with two of the three defining features of the verb forms: perfective or not perfective and progressive or not progressive. Attempts were made to fit the unique feature models using the defining features of the stimuli and the interactions among them. However, no unique feature models were found to fit better than the best Euclidean distance-choice model.
This chapter compared a number of existing models of stimulus identification data. One important model was omitted, the general recognition model by Ashby and Perrin ( 1988). This model is very general and can explain a variety of phenomena that could not be explained by other models. It was not considered here, despite its promise, primarily because it is still under development and because it is too general. In most cases, only specialized models can be fit, and it is not clear what specializations are necessary in particular situations. This situation can improve rapidly (see chaps. 6-8 and 16), however, and the full comparison between this model and the kinds of models discussed in this chapter would undoubtedly be interesting. Such attempts are already underway ( Ashby & Lee, 1991).
This work has been supported by Grant A6394 from the Natural Sciences and Engineering Research Council of Canada to the first author. Thanks are due to Greg Ashby, Tony Marley, and an anonymous reviewer for their helpful comments on an earlier version of this paper, to Milton Hodge for providing Hodge and Pollack's data, to Jim Corter for providing the ADDTREE and EXTREE structures, to Marion McGlynn for running the analyses in the third section, and to Marina Takane for preparing Figure 13.2.