Absolute vs. relative assessments in the detection of covariation

Previous research has shown that causal attributions can be made from patterns of covariation (Cheng, 1997). While the study of how humans learn contingencies goes back decades (e.g., Ward & Jenkins, 1965), cue interaction effects, involving covariations with two or more cues, have taken on particular importance (e.g., Shanks, 1985), due to their rich potential for theoretical insights. One such effect is causal discounting (Goedert & Spellman, 2005): People believe a cue is less contingent if they learned about it in the presence of a more contingent cue.
Using a new method for investigating covariation detection, the steamed-trial technique (Allen et al., 2008), Art Markman, Kelly Goedert and I (Laux et al., 2010) have established that differences in bias underlie causal discounting. We argued that this implies discounting is an effect of a process employed to make causal judgments after learning has occurred. Analyses of how different theories account for discounting, especially simulations of associative models, establishes that this is not necessarily correct; several learning models can reproduce our data. However, model and data explorations show that the key feature of those data is that they track relative, not absolute, magnitudes.
My dissertation extends this work establishing the plausibility of a comparative judgment process as the locus of causal discounting. I replicate the finding that responding tracks relative magnitudes. By conducting experiments that parametrically manipulate the contingency of the alternative cue (and thereby the relative contingency of the cues), I show that causal discounting is due to responding to contingencies as a linear function of their relative magnitude. I further verify that discounting manifests identically in response to contingencies presented via summary tables. Because summary tables do not afford the series of experiences necessary to build an association, this enhances the credibility of the theory that discounting is due to a shared process employed subsequent to learning—namely, a judgment process. These investigations reveal that discounting is not a cue interaction effect at all, but rather is a manifestation of a fundamental aspect of the systems that subserve covariation detection.