Accountability and Cognitive Bias

Implications of Coping Strategies for Judgmental Shortcomings

Accountability means expecting to justify a judgment or choice to others. Skeptics may accept that this changes what people say while questioning whether it changes how they think. For the intuitive-scientist and economist research programs, the central issue is when thought departs from a defined standard of good judgment. The skeptics might even dust off a slightly modified version of the disciplinary division of labor between psychology and the social sciences initially proposed by Miller and Dollard (1946): The mission of the cognitive research program is to shed light on how people think, whereas the mission of the intuitive-politician research program (and other deviant metaphoric traditions) is to shed light on what people think about and when they are willing to say what is on their minds (raising or lowering response thresholds for expressing certain views).

There is some merit in this traditional stance. Certain coping strategies have socially significant but cognitively uninteresting implications for the conditions under which biases and errors manifest themselves. It is easy to imagine the types of heavy-handed social pressures that trigger strategic attitude shifting also suppressing evidence of cognitive bias. People can readily learn that their boss expects them to use certain cues or to disregard others, and it is easy to see why cognitive theorists would find evidence of such bias suppression less than intriguing. There is no great surprise in learning that conformity pressures can alter response thresholds. The acceptability heuristic (as Tetlock, 1992, semi-facetiously labeled it) represents a minor content embellishment on the heuristics-and-biases portrait of the decision maker. The most salient consideration in many decisions is the justifiability of policy options to others. The cognitive research programs tell us that people often use a small number of cues in making up their minds; the intuitive-politician research program tells us that decision makers’ estimates of the probable reactions of those to whom they are accountable will be prominent among the few items considered. The traditional disciplinary division of labor still stands.

In other cases, however, the division begins to crumble. The self-presentational processes triggered by accountability can interact in complex ways with putatively basic cognitive processes. Within the intuitive-politician framework, a key function of private thought is preparation for public performances. Thought frequently takes the form of internalized dialogs in which people gauge the relative justifiability of alternative courses of action by imagining conversations with others in which accounts are exchanged, debated, revised, and evaluated. Indeed, converging lines of evidence demonstrate that certain forms of accountability do change how people think, not just what they are willing to say they think:

1. Predecisional forms of accountability to unknown audiences increase the complexity of argumentation, as revealed by content analysis of confidential thought protocols and the complexity of cue use as revealed by statistical modelling of judgment policies (Hagafors & Brehmer, 1983; Tetlock et al., 1989).

2. Accountability manipulations are much more effective in attenuating certain cognitive biases – primacy, overattribution, overconfidence – when participants learn of being accountable prior to (as opposed to after) exposure to the evidence on which they are basing their judgments (Tetlock, 1983b, 1985; Tetlock & Kim, 1987). There is also evidence that increased complexity of thought at least partly mediates these debiasing demonstrations.

3. The power of certain accountability manipulations to attenuate bias is attenuated by impositions of cognitive load that disrupt more effort and attention-demanding forms of information processing (Kruglanski & Freund, 1983; Tetlock, 1992) – disruption that should not occur if people coped with accountability by relying exclusively on low-effort attitude shifting.

4. When accountability manipulations do improve the quality of judgment and choice, the effects are often too differentiated to reproduce via simple response-threshold adjustment models; for example, enhanced “differential accuracy” from a Cronbach decomposition of person-perception accuracy scores (Tetlock & Kim, 1987) or improved calibration without degradation in resolution from analytical decomposition of Brier scores (Siegel-Jacobs & Yates, 1996; Tetlock & Kim, 1987) or closer correspondence between subjective importance rankings of cues in judgment tasks and “objectively” derived weightings from multiple regression equations (Hagafors & Brehmer, 1983).

5. The effects of accountability manipulations on “thoughts” sometimes persist even after the cancellation of the anticipated interview (Lerner & Tetlock, 1999).

Putting to the side the exact locus of accountability effects, the research of greatest interest to the judgment and decision-making community focused on variants of the hypothesis that accountability, which encourages preemptive self-criticism, also confers protection from cognitive biases rooted in reliance on simple, easy-to-execute heuristics. There is indeed a substantial list of biases that are attenuated, if not eliminated, by these forms of accountability. This list must, however, be followed by a compilation of biases that are either unaffected or even amplified by approximately the same forms of accountability.

On the debiasing side of the ledger, predecisional accountability to audiences with difficult-to-decipher views has been found to

1. Reduce, and occasionally eliminate, the fundamental attribution error (Cornielle, Leyens, Yzerbt, & Walter, 1999; Tetlock, 1985). Preemptively self-critical information processors are more reticent about drawing confident dispositional attributions from conduct that is clearly constrained by the situation (the so-called low-choice conditions of essay-attribution experiments), but no more reticent about drawing such conclusions from conduct that is unconstrained by the situation (the high-choice conditions of such experiments).

2. Reduce primacy and recency effects (Lerner & Tetlock, 1999). Preemptively self-critical thinkers are apparently more cautious about drawing conclusions from incomplete or fragmentary evidence, more willing to revise their preliminary conclusions in response to unexpected feedback, and less likely to be swayed by the last most memorable items of evidence presented.

3. Improve the calibration of the subjective probabilities that people attach to their predictions (Siegel-Jacobs & Yates, 1996; Tetlock & Kim, 1987). Preemptively self-critical thinkers are more likely (according to thought-protocol evidence) to take seriously the possibility that they might be wrong. Accountability that encourages this pattern of thinking has been found to improve the covariation between subjective confidence and predictive accuracy. Moreover, these improvements in “calibration” can be achieved at minimal cost in “resolution” – evidence that subjects are not simply scaling down all confidence estimates indiscriminately, but rather are carefully weighing the balance of evidence for each judgment.

4. Reduce the incompatibility bias – the tendency for negotiators to assume incorrectly that relationships are zero-sum and that the other party’s interests are opposed to their own (Thompson, 1995).

5. Reduce the influence of an initial and arbitrary reference point on subsequent judgments (the numeric–anchoring effect) by encouraging subjects to consider further relevant evidence and to revise estimates in light of that evidence – a beneficial effect of accountability that holds up as long as subjects are not under time pressure or other cognitive loads (Kruglanski & Freund, 1983).

6. Decrease reliance on simple, noncompensatory decision rules in favor of more complex compensatory ones that explicitly acknowledge trade-offs (McAllister, Mitchell, & Beach, 1979; Tetlock, 1983a).

7. Decrease the influence of incidental affect evoked in one situation on judgments made in a completely unrelated situation (Lerner, Goldberg, & Tetlock, 1998).

8. Decrease the influence of sunk costs on future investment decisions, in part by increasing willingness to adhere to self-imposed earlier limits on the amount planned to invest, and in part by increasing the salience of decision-making procedures (as opposed to outcomes) in evaluation (Brockner et al., 1981; Simonson & Staw, 1992).

9. Reduce illusory correlations and improve accuracy of covariation assessment in part by motivating more complex inferential strategies (Murphy, 1994).

10. Increase correspondence between the judgment strategies that people claim to be using and the strategies that statistical models impute to them (Cvetkovich, 1978; Hagafors & Brehmer, 1983).

On the bias-amplification side of the ledger, preemptively self-critical thinkers seem more, not less, susceptible to the following effects:

1. Ambiguity aversion – the tendency to prefer less ambiguous alternatives when given a choice between options that differ only in uncertainty about the probabilities with which outcomes may occur (Curley, Yates, & Abrahms, 1986). Postexperimental interviews reveal that a preoccupation with how to justify the choice increased preferences for options with well-defined probabilities over those with ambiguous probabilities, holding expected value constant.

2. The compromise effect – the tendency for a product to gain attractiveness simply by virtue of becoming a middle option in a choice set (Simonson & Nye, 1992). Accountable participants are more likely to select the compromise product because they apparently believed that it was more defensible than options that were superior on one dimension but inferior on another.

3. The attraction effect – the power of a relatively inferior alternative (Brand X) when added to a set of closely competing options (Brands A and B) to increase the attractiveness of the preexisting option that happens to be superior to X on all evaluative dimensions (Simonson, 1989). Again, accountable subjects are especially likely to select dominating options, apparently because they thought that the dominating options were less vulnerable to criticism.

4. The dilution effect – the tendency for nondiagnostic evidence to dilute the predictive power of diagnostic evidence (Tetlock & Boettger, 1989). Motivating self-critical thought can sometimes induce people to try too hard to discern relevance amidst irrelevance. Rather than zeroing in on the one relevant cue, accountable subjects often struggle to integrate irrelevant cues into their schematic representations of the target individuals. In environments with unfavorable signal-to-noise ratios, accountability can send people off on inferential wild goose chases.

5. The status quo effect – the tendency for the status quo to occupy a privileged position in decision making that makes it possible to reverse preferences for decision alternatives by holding expected value constant and arbitrarily designating one option as the status quo and the other as the option that requires change (Samuelson & Zeckhauser, 1988). The “change” option is typically held to a higher standard of proof – and accountable subjects are particularly prone to do so when there are identifiable losers as a result of modifying the status quo (Tetlock & Boettger, 1994).

Finally, turning to the null-hypothesis column of the ledger, it is worth contemplating a subset (from the proverbial file drawer) of studies in which accountability has no effect on established judgmental biases. These noneffects have included insensitivity to base-rate information, giving more predictive weight to causal as opposed to merely statistical relationships between variables, preference reversals as a function of choice-versus-matching elicitation procedures, insensitivity to sample size, and the conjunction fallacy.

It is hardly helpful, however, to abandon readers with a list – or, still less helpful, a list of lists. The key theoretical question concerns the nature of the underlying moderators of when one or another class of effect or noneffect emerges. My best guess is that much hinges on the correspondence or lack thereof between the intuitive theories that ordinary people hold about what constitutes good decision making and the more formal theories that academic observers hold about what constitutes rationality (cf. Wegener & Petty, 1995). In cases in which accountability attenuates bias, scholars have made good prima facie cases that the biases result, in part, from lack of mental effort and self-critical attention to one’s judgment processes. Subjects intuit that the social expectation in the accountability relationship is that they are supposed to process task-relevant information in more self-critical and thoughtful ways. Accordingly, they do so and thereby bring their judgments into closer correspondence with those that the expert community, the ultimate arbiters of normative standards, deems defensible. Accountability should be an effective debiasing manipulation to the degree it induces research participants to (1) consider a wider range of relevant cues or arguments; (2) pay more attention to the cues that they do use; (3) anticipate plausible counterarguments, evaluate their strengths relatively impartially, and factor those that pass some threshold of plausibility into their overall assessment.

In cases in which accountability amplifies bias, scholars have made a good prima facie case that the biases result from trying too hard to integrate irrelevant considerations into one’s mental representation of the problem. Here, efforts by subjects to put on the cognitive equivalent of their “Sunday’s best” – to become more thoughtful and to have a “neat justification package” readily available – lead to judgments that depart ever more markedly from those that expert observers deem rational. In effect, there is a negative correlation between lay observers’, and academic observers’, conceptions of good judgment. Finally, in cases in which accountability has virtually no effect, there is a good case either that subjects are completely unaware that they are thinking incorrectly (failing to use relevant cues or using irrelevant ones) or that subjects have conflicting normative intuitions (some of which coincide with the expert community’s conception of procedural rationality and some of which sharply diverge).