Viewing Training through
a Fuzzy Lens
Gwendolyn E. Campbell, Wendi L.Van Buskirk, and Amy E. Bolton
The essence of training in many domains is to help people establish effective strategies for interpreting and responding to incomplete and uncertain cues in technologically mediated environments. A number of training researchers have seen the utility of the Brunswik lens model (Brunswik, 1952; Goldstein, this volume) in this endeavor, and have applied the model to the tasks of assessing trainee performance and providing trainees with feedback (e.g., Balzer, Doherty, & O'Conner, 1989). The typical approach has been to generate two regression equations: one from the student data and the other, representing the “correct” model, from either the environment (i.e., ground truth) or a set of data generated by an expert. Training feedback may be based on the correct model alone or on a comparison of the similarities and discrepancies between the correct model and the student model. Finally, the effectiveness of this feedback at promoting improved student performance is assessed. For example, Balzer et al. (1992) asked participants to predict the number of wins a baseball team would have based on performance statistics such as earned run average (ERA), errors, and batting average. They used regression-based modeling and the Brunswik lens model to provide feedback to the participants about the relative weights of these cues and the outcome, and then assessed the extent to which participants improved in their ability to make these predictions. Although linear regression has proven to be effective at predicting both student and expert performance in many domains, feedback based on regression equations has not always led to improved performance in training research (Balzer et al., 1989).
A number of reasons have been hypothesized to explain the mixed results of regression-based feedback. One compelling explanation is that there are a number of characteristics and assumptions inherent in regression that make it unlikely that regression equations are capable of capturing the entire range of reasoning processes that humans employ across all domains (Campbell, Buff, & Bolton, 2000; Dorsey & Coovert, 2003; Rogelberg et al., 1999). Specifically, linear regression equations assume that the relationships between the cues and the final decision are linear and additive. The psychological literature, on the other hand, contains many examples of human reasoning that are noncompensatory and/or based on interactions between cues (Stewart, 1988). For example, when investigating the policies that managers use to distribute merit pay among employees, Dorsey and Coovert (2003) found that some managers reported