Differences in associative change between cues trained in compound are better explained by individual prediction errors than by relative certainty

The present experiments investigated associative changes during compound training involving cues with different learning histories. In six human predictive-learning experiments using the compound-test procedure (Rescorla, 2000), participants learned about the effects of foods (cues) on inflammation (outcome) in a hypothetical patient. Each food could cause decreases in inflammation, increases in inflammation, or had no effect. Cues A and C were trained as excitatory, whereas cues B and D were trained to acquire inhibitory properties. Participants then received compound training in which A and B were presented together and followed by an outcome that should produce associative loss, followed by a final test in which outcome expectations were assessed for the test compounds AD and BC. The initial training designs were expected to produce greater relative certainty about A than B, with A generating a larger individual prediction error than B during the compound training. If cues generating larger individual prediction errors undergo greater associative change, AD should receive lower ratings than BC. Alternatively, if participants preferentially attribute the compound outcome to the less certain cue (theory protection), BC should receive lower ratings than AD. Overall, the compound-test results were more consistent with an individual prediction-error account: AD was rated lower than BC in Experiments 1, 2, 5, and 6. Experiment 3 showed the opposite pattern, consistent with theory protection, but this pattern was not replicated in Experiment 4. We discuss the implications of these findings for the roles of prediction error and certainty in associative learning.