Comparing associative, statistical, and inferential reasoning accounts of human contingency learning

Abstract

For more than two decades, researchers have contrasted the relative merits of associative and statistical theories as accounts of human contingency learning. This debate, still far from resolution, has led to further refinement of models within each family of theories. More recently, a third theoretical view has joined the debate: the inferential reasoning account. The explanations of these three accounts differ critically in many aspects, such as level of analysis and their emphasis on different steps within the information-processing sequence. Also, each account has important advantages (as well as critical flaws) and emphasizes experimental evidence that poses problems to the others. Some hybrid models of human contingency learning have attempted to reconcile certain features of these accounts, thereby benefiting from some of the unique advantages of different families of accounts. A comparison of these families of accounts will help us appreciate the challenges that research on human contingency learning will face over the coming years.

Acknowledgments

The development of this paper was possible due to support from Department of Universities, Research, and Technology of the Andalucı´a Government (Junta de Andalucı´a) and NIMH Grant 33881. We would like to thank Jeffrey C. Amundson, Gonzalo Urcelay, and Daniel Wheeler for their comments on an earlier version of this manuscript.

Notes

¹ Although at the implementation level there are models of human learning based on neuroscience, our discussion of HCL models exclusively refers to those within the cognitive tradition.

² The problem for statistical models of not predicting acquisition curves might be solved by assuming that the 2 × 2 contingency table is preexperimentally filled with noise, such that the initial values of the cells are greater than zero—that is, (fa = fb = fc = fd) > 0. However, it is not clear how these cells, for a specific cue and a specific outcome, could be available in memory prior to any kind of experience with these stimuli. If this approach were to be adopted, one would have to assume the existence of virtually an infinite number of cells for any possible combination of cues and outcomes.

³ One might argue that only averaged acquisition curves are gradual because individual curves usually show a step-like function (e.g., Gallistel, Fairhurst, & Balsam, 2004). However, even individual curves do not commonly show an abrupt increase to asymptote after a single cue–outcome pairing, as predicted by statistical and inferential models.

⁴ Even a traditional model like ΔP can be viewed as a hybrid model in that it uses an associative structure to represent the cue–outcome pairings chronicled in Cell a.