1Pseudo-contingencies

Running Head: Pseudo-contingencies

Social Judgments Based on Pseudo-Contingencies:

A Forgotten Phenomenon

Klaus Fiedler Peter Freytag

University of Heidelberg, FRG

Author Note: The research underlying the present article was supported by various grants from the German Research Foundation (Deutsche Forschungsgemeinschaft). Correspondence concerning this paper should be addressed to Klaus Fiedler or Peter Freytag, Department of Psychology, University of Heidelberg, Hauptstrasse 47-51, 69117 Heidelberg, FRG. Email: or

Social Judgments Based on Pseudo-Contingencies:

A Forgotten Phenomenon

Is it possible that in the industrious endeavor of contemporary experimental psychology a major phenomenon has simply been missed? To be sure, our starting question refers to a major phenomenon that can be expected to play an important role in everyday adaptive behavior. In this article, we dare to claim to present such a phenomenon in the area of implicit judgment and decision making. It will be illustrated with both anecdotal and experimental evidence, along with a discussion of its theoretical implications. We (Fiedler & Freytag, 2002) have named the phenomenon pseudo-contingencies, which may sound a bit monstrous but precisely conveys its nature. Pseudo-contingencies (PCs) mimic contingencies, but on closer inspection turn out to be essentially different. Although confusing PCs with genuine contingencies involves a serious category mistake, both lay-people and experts treat them alike, precisely because PCs often provide the only way to resolve high environmental uncertainty.

Let us introduce the phenomenon presenting a couple of examples, one referring to lay theories and the other referring to scientific reasoning. To begin, assume that in a minority group with a particular ethnic background the baserate of, say, sexual fantasies is inflated relative to other groups. Assume further that in the same minority group, the baserate of another behavioral tendency, say, violent aggression is also conspicuously high. Such a coincidence of enhanced sexual fantasies and enhanced violence, especially when both tendencies are absent in relevant reference groups, will often give rise to the impression that the two attributes are mutually contingent. Sexual fantasies seem to correlate with violent aggression. One may even feel a sense of causality (e.g., sexual fantasies causing violent aggression). However, closer inspection may reveal the following (see Figure 1). The proportion of individuals showing pronounced sexual fantasies in the minority group is 75%. Further, the proportion of individuals showing violent aggression is 65% among those who show increased sexual fantasies and 85% among those who do not, yielding an overall baserate of (0.75  65%) + (0.25  85%) = 70% violent aggression in the group. Note that the violence rate is lower among group members with pronounced sexual associations. Notwithstanding the fact that enhanced baserates for both attributes hold for the group as a whole, the actual contingency between the two is therefore negative.

Assuming the two attributes to be mutually contingent thus entails an obvious category mistake. As a matter of principle, the coincidence of two elevated baserates must not be confused with a contingency between the attributes involved. Indeed, given high baserates, evidence for a contingency is particularly hard to get. Empirically, a positive contingency not only requires that the violence rate in people with sexual associations is above average, but that it exceeds a baserate that is already very high. Over a wide range of values, contingencies are independent of baserates. Holding baserates constant, the violence rate could increase, remain constant, or decrease with the frequency of sexual fantasies. Psychologically, however, lay observers will often fall prey to a PC illusion under these conditions and erroneously infer to have witnessed a relationship which in fact was not there.

Moving from lay people to experts, consider the following analogy referring to a common scheme in psychological test validation studies. For instance, in an attempt to validate the Implicit Association Test (IAT; cf. Greenwald, McGhee & Schwarz, 1998) in an applied setting, one may report that members of a target group produce strong IAT effects for some sort of sexual associations in a first step. Construct validity may then be "secured" in a second step by showing the baserate of various relevant criterion behaviors to be similarly elevated in the same target group, including violent aggression and other related constructs (e.g., perversion, neuroticism etc.). Whenever test validation studies rely on group comparisons rather than interpersonal variation, PC illusions may be around.

Do PCs constitute a distinct phenomenon?

The question arises whether PCs really represent something new, distinct from other well-known phenomena. One interpretation that suggests itself is in terms of simple associative learning. After all, when two baserates are inflated in a target group, the most frequent event combination will probably refer to the joint occurrence of the most frequent variable levels (e.g., sexual fantasies and violent aggression). Theories of associative learning and conditioning would thus entail the same prediction without postulating a new phenomenon. However, the PC phenomenon is indeed different. It not only predicts an illusory contingency when both baserates are high but also when both baserates are low! For instance, given a low rate of sexual fantasies and a low rate of violent aggression, the two factors should appear to be correlated as well, even though the absolute frequency of the critical observations is particularly low in this case. PCs are therefore distinct from classical conditioning, frequency-based heuristic inference, or density bias (Allan, 1993).

In a similar vein, PCs are distinct from illusory correlations (Fiedler, 2000; Garcia-Marques & Hamilton, 1996), spurious correlations, Simpson's paradox, and other multivariate paradigms (cf. Fiedler, Walther, Freytag & Stryczek, 2002) that all presuppose that genuine contingency information is encoded in the first place. A premise for all these models and paradigms would be that observers actually receive information about the joint occurrence of both attributes in the same individuals. Unless information about the joint frequencies or conditional frequencies (of one attribute given different values of the other attribute) can be assessed directly, a contingency task cannot be performed. In contrast, PCs are not restricted to this (seemingly) necessary condition. One might observe sexual fantasies on one occasion, and violent aggression on another, may-be even years later, without having the slightest chance to coordinate these detached data with the same reference persons, and still infer a strong relationship between these separately encoded attributes.

That PCs are not confined inductive learning is evident in the scientific community's readiness to accept PC-like theoretical explanations at the level of deductive reasoning. For example, the following explanation of the confirmation bias in social hypothesis testing was published by Zuckerman, Knee, Hodgins & Myake (1995) in JPSP and accepted by expert reviewers and readers as logically sound. When the task in different experimental conditions is either to test the hypothesis that an interview partner is extraverted or that an interview partner is introverted, either hypothesis tends to be confirmed. Those focusing on extraversion tend to find their partners rather extraverted, whereas those focusing on introversion find them rather introverted. The explanation offered by the authors involves the joint application of two well-established trends. First, it is well known that hypothesis testers engage in positive testing. When focusing on extraversion, they ask more extraversion questions (e.g., referring to time spent with friends). When focusing on introversion, they raise more introversion questions (e.g., referring to time spent on one’s own). This tendency is complemented by an acquiescence response tendency. Respondents tend to provide more affirming than disconfirming responses, regardless of the question focus. Both tendencies together, the questioner's question focus and the respondent's tendency to provide affirming answers, appear to provide a plausible account for the confirmation bias.

This account is based on a PC illusion though, because acquiescence is by definition non-contingent; that is, regardless of whether the majority of questions refers to extraverted or introverted behavior, acquiescence means, say, 75% yes responses in general. Logically, then, the account offered by Zuckerman et al. (1995) cannot explain that conversation results in hypothesis confirmation, simply because the confirmation rate remains constant. Rather, an appropriate account must explain why the same confirmation rate is worth more when more data had been gathered on a specific hypothesis (cf. Fiedler, Walther & Nickel, 1999).

Origins of PC illusions

Granting that PCs may intrude inductive as well as deductive reasoning and given that the underlying mistake can be readily understood, one might speculate about the illusion's evolutionary origin (for related topics, see Haselton & Buss, this volume). Indeed, a closer inspection of our learning environment renders PCs quite plausible – and even indicative of adaptive intelligence. No doubt, assessing (multiple) contingencies between aggression and antecedent factors is of adaptive value. Figure 2 identifies three potentially relevant factors: frustrations, sexual fantasies, and eliciting stimuli. To assess the comprehensive relation between all variables, one would have to encode aggressive events contingent on all three factors at the same time. Each memory trace referring to aggressive behavior should not be stored in isolation, but instead be conditionalized on a certain eliciting condition, a certain frustration level in the aggressor, and a certain level of sexual fantasies. Such a full three-factorial design is hardly ever available. Frustration is a hidden internal state, sexual fantasies may be kept secret, and eliciting conditions may be less apparent to observers than to actors. Even when the full design was available, acquiring systematic data would be a tedious and time-consuming task that would often exceed an observer's capacity or motivation. And even when we engaged in complex multi-factorial assessment, we may still regret the effort because at some later time we have to decide on completely different questions. For instance, having just encoded aggression as a function of frustration, sexual fantasies, and eliciting conditions, we may find ourselves in a situation that requires a decision on whether aggression increases with, say, heat or skin color.

As a matter of principle, it is impossible to simultaneously gather empirical data on all contingency problems that happen to become relevant at some later time. What could we do in face of this dilemma? A useful heuristic – which is by no means irrational but which is in fact the heuristic used in empirical sciences – is to rely on approximations from less complex, sub-ordinate contingencies. To predict the risk of violent aggression in a particular target group, it would be important, first of all, to recognize the baserate of aggression. If the baserate is extreme enough, assessing contingencies between aggression and eliciting conditions may be dysfunctional; it may be more adaptive to avoid such a situation altogether, rather than trying to acquire contingency information in too dangerous an environment. As noted by Kareev (2000), the risks and information costs of contingency learning are only justified when the baserate is not too extreme.

In the latter case, however, more control might be achieved by a model that explains when aggression occurs. Lacking a full design of all relevant factors, a useful proxy might be to assess other relevant factors having enhanced (or reduced) baserates in the target group. When the sexual fantasy baserate is high, and the violent aggression baserate too, this coincidence could provide the basis for the inference that sexual fantasies and aggression are somehow related to each other. If another baserate is elevated, say the frustration baserate, this triple coincidence supports an explanation of aggression in terms of both factors; for example, both sexual fantasies and heightened frustration levels could reflect the same sort of deprivation that explains the target group's aggressiveness. Such an inference would not only be in line with common attribution theories (Jones & McGillis, 1976; Kelley, 1973) but also with reasoning schemes in politics and economics. In any case, taking PCs as a proxy is presumably the best strategy an organism can apply in a complex environment for it helps to identify highly probable event combinations in the absence of covariation information.

Empirical evidence

Positive PCs. To demonstrate PC illusions experimentally, one ought to rule out prior knowledge and expectancies. Accordingly, the first series of PC experiments (cf. Fiedler & Freytag, 2002) referred to two unspecified personality tests, X and Y. The experimental cover story mentioned two groups of people, those in which psychotherapy is successful and those in which psychotherapy fails. For convenience, let us refer to groups A and B, respectively. In the first part of the stimulus presentation series, participants were informed about the outcomes on test X of various members of both groups. Rather than merely providing a numerical test score, each stimulus displayed the profile of a respondent's outcome on 12 items of the test. Aggregating over all 12 items, participants had to assess within a few seconds whether the overall (average) value of the respondent on a test was high (mean value about 1.0) or low (mean value about 3.0). To control how accurately participants had identified the test value in question, they were asked to provide online estimates of the mean test value they had just seen using a vertical graphical bar that could be adjusted on a scale ranging from 0 to 4. In this manner, participants learned over the course of 36 trials that high test X values prevailed in members of one group whereas low test X values prevailed in members of the other group. Specifically, the stimulus distribution was 24 high and 12 low test scores in group A, as compared to 12 high and 24 low scores in group B. In a second run, participants learned about test Y scores of members of the same two groups. The group in which high test X scores had predominated (i.e., A) also had more high (24) than low (12) scores on test Y, and the group in which low X scores had prevailed (i.e., B) also happened to have more low (24) than high (12) test Y scores. Again, graphical online ratings secured that participants correctly encoded the baserate distributions.

In fact, the X and Y test scores assessed in the first and second part of the stimulus series belonged to the same individuals from the two groups. However, as X and Y scores were presented in separate runs, participants had no chance to encode the actual contingency, which accorded to the distribution in Figure 3. However, the alignment of two skewed baserate distributions should create a PC illusion. Participants should come to believe that X and Y are correlated within both groups, due to the co-occurrence of either high X and high Y values (group A) or low X and low Y values (group B).

For a suitable test, a simple prediction task was administered. In each of the 16 prediction task trials a new member of either group A or group B was presented whose score on either test X or test Y was given in the same graphical format as in the online rating during the learning phase, and participants were asked to predict the new member’s score on the other test using the same graphical adjustment device as before. A PC effect would be evident in a positive correlations between the given and the predicted values on this task. That is, given a high value on the predictor test, the values adjusted for the other test should be higher than when a low value was given on the predictor test and – regardless of whether X was given and Y left open to be predicted or Y was given and X to be predicted. Moreover, the PC illusion should hold for (the high-scoring) group A as well as (for the low-scoring) group B.

In fact, the correlations between X and Y produced on the prediction test were fairly above zero for most of the participants (see the first pair of bars in Figure 4). There was a regular tendency to predict higher values on one test when the given values on the other test were high rather than low, and to predict lower values when given values were low rather than high. Virtually the same trends held for predictions with test X (Y) as the predictor and test Y (X) as the criterion. The positive relation was not confined to those prediction task trials that involved the frequently paired test scores. When the most frequently paired test scores in a group were high X and high Y values, predictions from low Y and low X values yielded the same positive relation, emphasizing the independence of PCs from simple associative learning effects.

Thus, even though no contingency information was provided – observations of X and Y were presented in separate runs, making contingent encoding impossible – the alignment of two skewed baserate distributions created the illusion that X and Y were positively related, as evident in participants' predictions. To be sure, a manipulation check showed that judges did recognize the skewed baserates. That is, they ought to have known that a coincidence of high X and high X scores in the one group (and of low X and low Y scores in the other) could be expected from the baserates alone.

Given the tendency to infer PCs from joint baserate manipulations, the idea suggests itself that the same tendency might be at work in tasks that look like a contingency task. Thus, when information about a persons' X and Y scores is given simultaneously rather than in successive runs, making the contingent encoding of both variables possible, there is no guarantee that participants encode the contingency proper. Instead, they might still be sensitive to the alignment of X and Y baserates, which might actually override the genuine contingency information. To test this consideration empirically, we repeated the experiment described above with simultaneous rather than successive presentation of X and Y test scores. Note that while the baserates of high (low) X and high (low) Y were aligned in the two target groups, the proportion of high versus low scores on one test were the same for high and low levels of the other test – yielding a zero correlation (Figure 3). Nevertheless, the participants' predictions yielded a similar PC pattern as with successive presentation (see the second pair of bars in Figure 4).