In this paper, we test the effect of manipulating discrimination difficulty on subsequent generalization of learning and in particular, on the peak shift effect. Participants learned a discrimination where one stimulus led to an outcome (S+) and another stimulus led to no outcome (S-). Difficulty was manipulated by varying the degree of similarity between the S+ and S- across groups (easy/medium/hard). In contrast to similar studies in animals, we found that increasing the difficulty of the discrimination resulted in less peak shift. Using a hierarchical mixture model, we characterize the effects of discrimination difficulty on relational- and similarity based responding, and show for the first time, a similar mixture of responding on stimulus identification gradients. We conclude that peak shift on generalization and identification measures can be explained by mixtures of participants responding in different ways.

In a causal illusion task, participants rate a cue that has an objectively null contingency with an outcome as causal. Trials are usually organized according to a 2x2 table representing the presence/absence of a binary cue and a binary outcome. Cell A outcomes (cue, outcome) can be attributed to the cue. But how do participants interpret trials from cell C (no cue, outcome), where the cause of the outcome is unspecified? In two experiments we asked participants to provide causal explanations for cell A and C trials in a medicine-recovery causal illusion task. Participants who reported that the cause of cell C outcomes (e.g., strong immunity, spontaneous recovery) did not also apply to cell A outcomes showed the strongest causal illusion. Such a causal reasoning process undermines the logic behind the delta P metric typically used to define a contingency, and thereby provides a potential normative account of causal “illusions”.

Evidence from a variety of learning tasks suggests that cuesthat are more predictive of an outcome attract greater attentionand are learned about more effectively in subsequent tasks. Wetested whether this learned predictiveness effect is due to theobjective strength of the cue-outcome association (cue-outcome correlation), or the degree to which the cue isinformative for making the correct choice on each trial (choicerelevance), by manipulating the possible outcome choicesavailable on each trial. Experiment 1 compared two sets of cuesthat were equally (and imperfectly) correlated with outcomesand showed learning biases in favor of the set of cues that hadinitially been more relevant for choices made on each trial.Experiment 2 used a more conventional learned predictivenessdesign in which the cue-outcome correlation was stronger forone set of cues (perfect predictors) than the other set (imperfectpredictors). However, here we manipulated whether or not theimperfect predictors could be used to make a correct choice,and thus whether the imperfect predictors possessed choicerelevance that was equal to or less than the perfect predictors.In this case, we found no evidence that the relevancemanipulation made any difference; learning biases towards theperfect predictor were evident regardless. The results suggestthat both cue-outcome correlation and choice relevance canlead to changes in associability and learning biases; both wereindividually sufficient but neither were necessary.

The illusory causation effect describes the tendency to judge an unrelated cue and outcome to be causally related. The standard procedure for assessing the illusion is based on the implicit assumptions that participants start as naïve observers with no prior beliefs about the likely relationship between the cue and outcome, and that learning can be adequately captured as a point-estimate causal rating after null contingency training. Here, we use a novel distributional measure to assess participants’ beliefs over a range of causal relationships prior to, as well as after, exposure to non-contingent cues and outcomes. Across two experiments with different causal scenarios and 50% cue and outcome density, we show that participants have an initial bias towards expecting a causal relationship between the cue and outcome, and that this bias is mostly corrected after exposure to the null contingency. We conclude that distributional measures of causal beliefs can offer novel insights in understanding the illusory causation effect.