Expert Political Judgment

by Philip E. Tetlock

  THE POWER OF IMAGINATION

  In the last fifteen years, there has been an intriguing convergence between experimental efforts to correct judgmental biases and the entrepreneurial efforts of scenario consultants to improve contingency planning in business and government. Experimental psychologists have found that many judgmental shortcomings can be traced to a deeply ingrained feature of human nature: our tendency to apply more stringent standards to evidence that challenges our prejudices than to evidence that reinforces those prejudices. These psychologists have also stressed the value of busting up this cozy arrangement. And they have had some success in correcting overconfidence by asking people to look for reasons that cut against the grain of their current expectations,2 in correcting belief perseverance by highlighting double standards for evaluating evidence,3 and in correcting hindsight bias4 by asking people to imagine ways in which alternative outcomes could have come about. Of course, these demonstrations have all been in controlled laboratory conditions. The results tell us little about the mistakes people make in natural settings or about the adverse side effects of treatments.

  Reassuringly, scenario consultants—who cannot be quite so readily dismissed as detached from reality—hold strikingly similar views on the causes of, and cures for, bad judgment.5 They appeal to clients to stretch their conceptions of the possible, to imagine a wider range of futures than they normally would, and then to construct full-fledged stories that spell out the “drivers” that, under the right conditions, could propel our world into each alternative future. Scenario writers know that it is not enough to enumerate pallid possibilities. They must make it easy “to transport” ourselves into possible worlds, to get a feeling for the challenges we would face. They urge us to abandon the illusion of certainty: to adopt the stance “I am prepared for whatever happens.”6

  Scenario consultants should not, of course, be the final judges of their own effectiveness. When pressed for proof, the consultants have thus far offered only anecdotes, invariably self-promoting ones, drawn from a massive file drawer that holds an unknown mix of more or less successful test cases. Their favorite success story is Royal Dutch Shell in the early 1970s. The Shell group was looking for factors that could affect the future price of oil. They suspected that the Arabs would demand higher prices for their oil, but they could not say when. But they needed the ability to read the mind of Anwar Sadat, or a spy in the Egyptian high command, to predict the Yom Kippur war in October 1973. They knew only that storm clouds loomed on the horizon. The United States was exhausting its known reserves. American demand for oil was growing. And OPEC was flexing its muscles. A large fraction of the world reserves was controlled by Middle Eastern regimes that bitterly resented Western support for Israel. One of their scenarios now sounds eerily prescient: massive price shocks that transformed the oil business and contributed to the “stagflation” of the 1970s. Schwartz claims that mentally preparing Shell managers for this medium-term future gave them a critical advantage against their less imaginative competitors.7

  Although scenario writers eschew prediction, they take parental pride when one of their progeny proves prophetic. In addition to the bull’s-eye OPEC scenario, the Shell group advertises its success in anticipating, in 1983 when U.S.-Soviet tensions were running high, radical reform inside the Soviet Union. The Shell futurists persuaded top executives to consider the possibility that the Soviet Union would open its massive untapped resources for development by multinational companies, that the cold war would thaw, and that Europeans would be willing to buy most of their natural gas from the soon-to-be-former Soviet Union. The Shell team also advanced a scenario in which OPEC unity fractured as new supplies came on-line and as demand for oil remained flat. They can thus claim to have foreseen not only the rise of OPEC in the 1970s but also its partial fall in the 1980s.

  These bull’s-eyes are less impressive, however, when we remember that consultants write so many scenarios that they are guaranteed to be right once per gig. For instance, James Ogilvy and colleagues constructed three scenarios for China in 2022 that covered a broad waterfront of possibilities.8 The first envisioned a prosperous, democratic China that becomes, in absolute terms, the world’s largest economy and boasts of a per capita income equivalent to today’s Taiwan. The second depicted a China that is dominated by an oligarchic network of extended families and is so beset by regional factions that it teeters on the edge of civil war. A third anticipated corruption and inequitable distribution of wealth becoming so pervasive that a populist military leader seizes control (after conquering oil-rich territory in Russia’s Far East). The authors advised investors to test the viability of their business plans against each scenario because none could be ruled out.

  There are good reasons to be wary here. Portfolio diversification theory in finance would not command such wide professional acclaim if it had not advanced beyond the folk aphorism “Don’t put all your eggs in one basket.” And scenario consultants cannot expect more than fleeting fame if their advice reduces to “anything is possible” so “be prepared for anything.” There is also the concern that advocates of scenario methods make good livings hawking their wares and have little incentive to explore the negative side effects of leading people down too many overembellished paths. Absent the regulatory equivalent of the Food and Drug Administration to set standards for cognitive self-improvement products, it is hard to say whether consumers are wasting their hard-earned dollars on scenario snake oil.9

  There is, then, a need for a disinterested assessment. The starting point for our assessment is support theory, the final work of the extraordinary psychologist, Amos Tversky. Support theory posits the likelihoods people attach to outcomes to be monotonic functions of the relative strength of the arguments people can muster for those outcomes. If I feel that the arguments for one set of possibilities are five times more powerful than those for another, that 5:1 ratio will translate into a higher subjective probability (how much higher must be estimated empirically). More controversially, the theory also posits that people are quite oblivious to the complex possibilities implicit in characterizations of events and, as a result, prone to violate a core assumption of formal probability theory: the “extensionality” principle. Odd though it sounds, the expectation is that people will often judge the likelihood of a set of outcomes to be less than the sum of the likelihoods of each member of that set. “Unpacking” stimulates us to imagine sub-possibilities, and arguments for those sub-possibilities, that we would have otherwise overlooked. Thus, unpacking a set of events (e.g., victory in a baseball game) into its disjoint components, A1 ∪ A2 (e.g., victory by one run or victory by more than one run), typically increases both its perceived support and subjective probability.10

  Support theory raises a warning flag: people should quickly become discombobulated, and routinely violate extensionality, when they do what scenario consultants tell them to do: decompose abstract sets of possibilities—say, all possible ways a leader might fall—into increasingly specific and easily imagined sub-possibilities that specify in scenario-like detail the various ways in which that outcome might happen.11 According to Daniel Kahneman and Amos Tversky, such confusion derives from the difficulty that people have in reconciling the tension between “inside” and “outside” approaches to forecasting.12 Unpacking scenarios encourages people to adopt an inside view: to immerse themselves in each case and judge the plausibility of pathways to outcomes by drawing on their detailed case-specific knowledge of the forces at work. People usually find more support for these case-specific predictions than they would have if they had based their judgments on an outside view, if they had stepped back from the details of individual cases and grouped them into summary categories (base rates). This is because the outside view is anchored in external referents that stop people from being swept away by “good stories.” For example, revolutions are rare events even in the “zone of turbulence,” and the outside view reminds us that, no matter how good a story one can tell about i
mpending regime collapse in North Korea or Saudi Arabia, one should adjust one’s inside perspective likelihood estimates by outside perspective base rates for comparable outcomes in sets of comparable cases.

  Of course, in the end we might decide that confusion is a price worth paying if scenario exercises shield us from cognitive biases such as over-confidence and belief perseverance. The best way to combat powerful theory-driven biases could be by activating countervailing biases rooted in our imaginative ability to suspend disbelief and to mobilize support for even far-fetched possibilities.13

  DEBIASING JUDGMENTS OF POSSIBLE FUTURES

  In the 1990s, we conducted a series of small-scale experiments designed to assess whether the hypothesized benefits of scenario exercises outweighed the hypothesized costs. The largest two of these experiments drew participants from the forecasting exercises on Canada and Japan (for details, see Methodological Appendix).

  Figure 7.1. The set of possible futures of Canada unpacked into increasingly differentiated subsets. In the control condition, experts assigned probabilities to the two (exclusive and exhaustive) futures at the top of the figure. In the experimental unpacking condition, experts assigned probabilities to each of the eight exclusive and exhaustive futures at the bottom of the figure.

  Canadian Futures Scenarios

  This experiment compared the likelihood judgments that expert and dilettante, fox and hedgehog, forecasters made before they did any scenario exercises, after they completed scenario exercises, and finally, after they completed “reflective equilibrium” exercises that required reconciling logical contradictions between their pre- and postjudgments by ensuring their probabilities summed to 1.0. Figure 7.1 lays out the scenarios that were judged (a) possible futures involving either a continuation of the status quo (federal and provincial governments agree to continue disagreeing over constitutional prerogatives) or a strengthening of Canadian unity (in which agreements are reached); (b) possible futures in which a secessionist referendum in Quebec succeeds, and controversy shifts to the terms of divorce.

  Figure 7.2 shows that we replicated the well-established finding that “merely imagining” outcomes increases the perceived likelihood of those outcomes: pre-scenario judgments of probability were uniformly smaller than post-scenario judgments.14 We also broke new ground. We discovered that the imagination effect was greatest under three conditions: (a) forecasters were experts rather than dilettantes; (b) forecasters were imagining departures from the status quo (the breakup of Canada) rather than continuation of the status quo; (c) forecasters were foxes rather than hedgehogs. These results took us a bit aback. Our guess had been that expertise would make it easier to winnow out far-fetched scenarios. But the net effect of expertise—especially expertise coupled to an imaginative cognitive style—was to make it easier to get swept away by “change scenarios” that prime rich networks of cause-effect associations.

  Figure 7.2. Effects of scenario-generation exercises on hedgehog and fox, expert and dilettante, forecasters of possible five- and ten-year futures on Canada (1992–1997–2002).

  It is hard to make a convincing correspondence or coherence case that these scenario exercises improved judgment. If we adopt the correspondence definition that good judges assign higher probabilities to things that happen than to things that do not, the exercises clearly led forecasters astray. Before the exercise, experts judged continuation of Confederation more likely than disintegration; afterward, they flipped. If we adopt the coherence definition that good judges must be logically consistent, it gets even harder to discern the benefits of scenario exercises. Before the exercise, binary complementarity held: the judged probability of Canada holding together plus that of Canada falling apart summed to nearly exactly 1.0 for both experts and dilettantes. After the exercise, the average probability of these futures summed to 1.58 for all experts and 2.09 for fox experts. Probability judgments became increasingly sub-additive and in violation of the “extensionality” norm of probability theory.

  Why do we find this? Unpacking is mentally disruptive and scenarios are extreme forms of unpacking. One takes a vague abstraction, all possible paths to Canada’s disintegration, and explores increasingly specific contingencies. Quebec secedes and the rest of Canada fragments: the Maritimes—geographically isolated—clings to Ontario, but Alberta flirts with the United States rather than bonding with the other western provinces that have broken with Ontario. One knows in the back of one’s mind that the cumulative likelihood of all these contingent linkages holding true is vanishingly small. But the images are vivid, the plotlines plausible, and it becomes increasingly taxing to keep all the other logical possibilities in focus.

  Mediational analyses underscore this account. “Imaginability” appears to drive the inflation of subjective probabilities. Participants judged easier-to-imagine futures more likely (average r [58] = .46, p < .05). And, after controlling for variation in the imaginability of scenarios, the power of scenario exercises to inflate subjective probabilities disappears, as does the greater susceptibility of experts and foxes to scenario effects.

  In the closing phase of the experiment, it seemed only fair to give defenders of the scenario method a final shot at showing that, although incoherence may be the temporary price of exploring alternative futures in depth, experts encouraged to reflect on their clashing probability estimates would produce “better” judgments than experts not so encouraged. We brought to experts’ attention, as diplomatically as possible, the logical impossibility of the subjective probabilities they had assigned scenarios. As a “consistency check” we asked experts to add the probabilities that they had assigned across “classes of scenarios.” (These sums exceeded 1.0 for 85 percent of respondents.) We then asked them to adjust their estimates so that the sums would equal the mandated ceiling on probabilities of 1.0, and to approach this adjustment process in pursuit of what philosophers call “reflective equilibrium.” “It is well known,” we informed participants, that “different methods of posing questions about complex events frequently elicit different answers. For this reason, we routinely ask participants how they would prefer to reconcile any potential conflicts among their responses.”

  We are now in uncharted territory. One conjecture was that, although scenario exercises puffed up probabilities across the board, they puffed up some more than others, and experts would preserve the new proportional relations among probabilities that emerged from the scenario exercises. So, if an expert judged the continuation-of-Canada and the disintegration-of-Canada scenarios as having the same cumulative likelihood of .70 each (summing to 1.4), he should “normalize” around the values of .5 and .5, regardless of what the expert thought earlier. The second conjecture was that experts would conclude that they had been led down the primrose path, recognize that their estimates had been artificially inflated by the “imaginability” of scenarios, and correct themselves by returning to their pre-scenario estimates. The second conjecture was closest to the mark for hedgehogs who were the most outspokenly skeptical of the scenario exercises at the beginning and remained so at the end; the first conjecture captured the foxes who were the most open to scenarios at the outset and remained convinced the exercises had some lasting value.

  Cynics might say that the scenario exercise was a circuitous journey into incoherence that brought us back to roughly where we started. Although scenarios changed some minds, they did not change many by much after we pressured people to integrate their conflicting intuitions into a coherent stand. And those changes in perspective that did persist through the reflective equilibrium exercise had the net effect of reducing accuracy: assigning higher probabilities to things that did not happen and lower probabilities to things that did.

  Japanese Futures Scenario Experiment

  Defenders of the scenario method could argue that the Canadian results are peculiar to the country, time period, and experts examined. Anticipating this objection, we ran a parallel experiment on forecasters who had already, without sce
nario assistance, attached probability estimates to the possible futures of Japan. Figure 7.3 shows that the “stories about possible futures for the Japanese economy” fell into three categories: possible futures featuring a continuation of the status quo, a substantial improvement on the status quo, and a substantial deterioration. We “unpacked” each economic future into three subclasses of scenarios that specified possible political pathways to each future.

  Figure 7.4 shows that scenario exercises again puffed up probability estimates beyond the bounds of reason. Pre-scenario estimates of each possible future summed roughly to 1.0, whereas the post-scenario estimates (produced by adding the likelihoods of sub-scenarios), substantially surpassed 1.0. Scenario thinking encouraged sub-additive estimates in which the judged likelihood of the whole consistently fell short of the sum of its parts. Overall, the results paralleled those of the Canadian study in several key respects: (1) scenario exercises had more pronounced effects on experts than on dilettantes, especially experts with foxlike cognitive styles; (2) the power of scenarios to inflate subjective probabilities was greater for departures from the status quo, either upward or downward, than for the status quo; (3) when participants performed the “reflective equilibrium” task of reconciling logical contradictions and disciplining probability estimates to sum to 1.0, they mostly split the difference between their pre- and post-scenario estimates.