Associative models of causal learning predict recency effects: Judgments at the end of a trial series should be strongly biased by recently presented information. Prior research, however, presents a contrasting picture of human performance. López, Shanks, Almaraz, and Fernández (1998) observed recency, whereas Dennis and Ahn (2001) found the opposite, primacy. Here we replicate both of these effects and provide an explanation for this paradox. Four experiments show that the effect of trial order on judgments is a function of judgment frequency, where incremental judgments lead to recency while single final judgments abolish recency and lead instead to integration of information across trials (i.e., primacy). These results challenge almost all existing accounts of causal judgment. We propose a modified associative account in which participants can base their causal judgments either on current associative strength (momentary strategy) or on the cumulative change in associative strength since the previous judgment (integrative strategy).

P. W. Cheng's (1997) power PC theory of causal induction proposes that causal estimates are based on the power (P) of a potential cause, where P is the contingency between the cause and effect normalized by the base rate of the effect. Most previous research using a standard causal probe question has failed to support the predictions of the power PC model but recently Buehner, Cheng, and Clifford (2003) found that participants responded in terms of causal power when probed with a counterfactual test question, which they argued prompted participants to consider the base rate of the effect. However, Buehner et al. framed their counterfactual question in terms of frequency, a factor that has been demonstrated to decrease base rate neglect in judgements under uncertainty. In the experiment reported here, we sought to disentangle the influence of counterfactual and frequency framing of the probe question to determine which factor is responsible for encouraging responses in terms of causal power.

Results from human causal learning tasks that employ multiple cues are often interpreted in terms of the elemental theory of Rescorla and Wagner (1972). However, most results can also be successfully interpreted by the configural model proposed by Pearce (1987, 1994). One method of discriminating between these alternatives is through an investigation of summation and overexpectation. Indeed, demonstrations of these phenomena are fundamental to an elemental approach but are generally incompatible with an account that involves configural processing. Using a procedure in which the magnitude of the outcome varied, evidence for both summation and overexpectation was obtained in two experiments.