UBC Theses and Dissertations
Dynamic information model of identification performance Mori, Shuj
This dissertation examined analysis methods and models of sequential dependencies in absolute identification responses. It has been reported that observers' absolute identification responses are strongly affected by previous stimuli and responses, although there is no agreed-upon method of analysis of these sequential dependencies. In this thesis, I used for this purpose multivariate information analysis (Garner, 1962; Garner & McGill, 1956; McGill, 1954), which is an extension of one-input one-output contingent uncertainty (information transmission) to the multivariate case. Multivariate information analysis is preferred to other methods because statistically it is a nonmetric analysis of categorical data, such as those from an absolute identification experiment (Krippendorf, 1986). However, there are some difficulties in its application to empirical data. For example, it is known that information measures are likely to be inflated (or overestimated) when there are a small number of observations per stimulus relative to a large number of variables involved in the calculation (e.g., Houtsma, 1983). Since no previous research had dealt with the inflation problem of multivariate information measures, I ran extensive computer simulations of absolute identification and calculated the multivariate information measures as a function of the number of observations and the number of variables used in the calculation. As expected, the multivariate information measures were inflated for a small number of observations, and they reached their theoretical and/or asymptotic values as larger numbers of observations were used to calculate them. To solve the inflation problem, I used the results of the computer simulations and a method of pooling individual data to estimate the amount of inflation of the information measures and correct them accordingly. Previous studies had suggested that there are three important factors affecting sequential dependencies in absolute identification responses: the amount of stimulus information available to the observers (measured by the amount of information transmission), the number of stimulus/response categories, and giving observers trial-by-trial feedback. To investigate these three factors systematically, I conducted seven absolute identification experiments and analyzed the resulting data by multivariate information analysis with the correction method mentioned above. The results confirmed previous results as follows: (1) The amount of sequential dependencies was inversely related to the amount of information transmission (McGill, 1957; Mori, 1989; Ward & Lockhead, 1971). (2) The amount of sequential dependencies increased with an increasing number of stimulus/response categories (Garner, 1953). (3) The dependency on the previous stimulus was larger when feedback was given than when it was not, and the dependency on the previous response was smaller when feedback was given (Ward & Lockhead, 1971). The results (1) and (2) can be interpreted as an increase of the amount of sequential dependencies with the increasing complexity of making judgments in the task (Garner, 1953). Since the present results were obtained across stimulus modalities (e.g., sound frequency, brightness, visual position), they support the idea that sequential dependencies in absolute identification responses arise mostly from the observer's response processes in the absolute identification task (e.g., Garner, 1953; Ward & Lockhead, 1971). Finally, two general models of absolute identification (Braida & Durlach, 1988; Treisman, 1985) were examined to interpret the pattern of sequential dependencies and other results obtained in this thesis. While some aspects of Braida & Durlach's (1988) model were disconfirmed by the present results (although the model does not make explicit predictions about the type of sequential dependencies obtained in the present study), the present results fit quite well with Treisman's (1985) model, with a few exceptions.
Item Citations and Data