Representativeness

“Representativeness” is a heuristic by which we assume the probability that an example belongs to a particular class is based on how well that example represents the class. On the face of it, this seems like a reasonable heuristic. But it can lead to erroneous results if you’re not careful.

The concept is a bit tricky, but here’s an experiment makes this bias crystal clear.40 Subjects were given the following description of a woman named Linda:
Linda is 31 years old, single, outspoken, and very bright. She majored in philosophy. As a student, she was deeply concerned with issues of discrimination and social justice, and also participated in antinuclear demonstrations.

Then the subjects were given a list of eight statements describing her present employment and activities. Most were decoys (“Linda is an elementary school teacher,” “Linda is a psychiatric social worker,” and so on), but two were critical: number 6 (“Linda is a bank teller,” and number 8 (“Linda is a bank teller and is active in the feminist movement”). Half of the subjects were asked to rank the eight outcomes by the similarity of Linda to the typical person described by the statement, while others were asked to rank the eight outcomes by probability.

Of the first group of subjects, 85% responded that Linda more resembled a stereotypical feminist bank teller more than a bank teller. This makes sense. But of the second group of subjects, 89% of thought Linda was more likely to be a feminist bank teller than a bank teller. Mathematically, of course, this is ridiculous. It is impossible for the second alternative to be more likely than the first; the second is a subset of the first.

As the researchers explain: “As the amount of detail in a scenario increases, its probability can only decrease steadily, but its representativeness and hence its apparent likelihood may increase. The reliance on representativeness, we believe, is a primary reason for the unwarranted appeal of detailed scenarios and the illusory sense of insight that such constructions often provide.”41

Doesn’t this sound like how so many people resonate with movie-plot threats–overly specific threat scenarios–at the expense of broader risks?

In another experiment,42 two groups of subjects were shown short personality descriptions of several people. The descriptions were designed to be stereotypical for either engineers or lawyers. Here’s a sample description of a stereotypical engineer:
Tom W. is of high intelligence, although lacking in true creativity. He has a need for order and clarity, and for neat and tidy systems in which every detail finds its appropriate place. His writing is rather dull and mechanical, occasionally enlivened by somewhat corny puns and flashes of imagination of the sci-fi type. He has a strong drive for competence. He seems to have little feel and little sympathy for other people and does not enjoy interacting with others. Self-centered, he nonetheless has a deep moral sense.

Then, the subjects were asked to give a probability that each description belonged to an engineer rather than a lawyer. One group of subjects was told this about the population:
Condition A: The population consisted of 70 engineers and 30 lawyers.

The second group of subjects was told this about the population:
Condition B: The population consisted of 30 engineers and 70 lawyers.

Statistically, the probability that a particular description belongs to an engineer rather than a lawyer should be much higher under Condition A than Condition B. However, subjects judged the assignments to be the same in either case. They were basing their judgments solely on the stereotypical personality characteristics of engineers and lawyers, and ignoring the relative probabilities of the two categories.

Interestingly, when subjects were not given any personality description at all and simply asked for the probability that a random individual was an engineer, they answered correctly: 70% under Condition A and 30% under Condition B. But when they were given a neutral personality description, one that didn’t trigger either stereotype, they assigned the description to an engineer 50% of the time under both Conditions A and B.

And here’s a third experiment. Subjects (college students) were given a survey which included these two questions: “How happy are you with your life in general?” and “How many dates did you have last month?” When asked in this order, there was no correlation between the answers. But when asked in the reverse order–when the survey reminded the subjects of how good (or bad) their love life was before asking them about their life in general–there was a 66% correlation.43

Representativeness also explains the base rate fallacy, where people forget that if a particular characteristic is extremely rare, even an accurate test for that characteristic will show false alarms far more often than it will correctly identify the characteristic. Security people run into this heuristic whenever someone tries to sell such things as face scanning, profiling, or data mining as effective ways to find terrorists.

And lastly, representativeness explains the “law of small numbers,” where people assume that long-term probabilities also hold in the short run. This is, of course, not true: if the results of three successive coin flips are tails, the odds of heads on the fourth flip are not more than 50%. The coin is not “due” to flip heads. Yet experiments have demonstrated this fallacy in sports betting again and again.44

Probability Heuristics

The second area that can contribute to bad security trade-offs is probability. If we get the probability wrong, we get the trade-off wrong.

Generally, we as a species are not very good at dealing with large numbers. An enormous amount has been written about this, by John Paulos27 and others. The saying goes “1, 2, 3, many,” but evolutionarily it makes some amount of sense. Small numbers matter much more than large numbers. Whether there’s one mango or ten mangos is an important distinction, but whether there are 1,000 or 5,000 matters less–it’s a lot of mangos, either way. The same sort of thing happens with probabilities as well. We’re good at 1 in 2 vs. 1 in 4 vs. 1 in 8, but we’re much less good at 1 in 10,000 vs. 1 in 100,000. It’s the same joke: “half the time, one quarter of the time, one eighth of the time, almost never.” And whether whatever you’re measuring occurs one time out of ten thousand or one time out of ten million, it’s really just the same: almost never.

Additionally, there are heuristics associated with probabilities. These aren’t specific to risk, but contribute to bad evaluations of risk. And it turns out that our brains’ ability to quickly assess probability runs into all sorts of problems.
The Availability Heuristic

The “availability heuristic” is very broad, and goes a long way toward explaining how people deal with risk and trade-offs. Basically, the availability heuristic means that people “assess the frequency of a class or the probability of an event by the ease with which instances or occurrences can be brought to mind.” 28 In other words, in any decision-making process, easily remembered (available) data are given greater weight than hard-to-remember data.

In general, the availability heuristic is a good mental shortcut. All things being equal, common events are easier to remember than uncommon ones. So it makes sense to use availability to estimate frequency and probability. But like all heuristics, there are areas where the heuristic breaks down and leads to biases. There are reasons other than occurrence that make some things more available. Events that have taken place recently are more available than others. Events that are more emotional are more available than others. Events that are more vivid are more available than others. And so on.

There’s nothing new about the availability heuristic and its effects on security. I wrote about it in Beyond Fear,29 although not by that name. Sociology professor Barry Glassner devoted most of a book to explaining how it affects our risk perception.30 Every book on the psychology of decision making discusses it.

In one simple experiment,31 subjects were asked this question:
In a typical sample of text in the English language, is it more likely that a word starts with the letter K or that K is its third letter (not counting words with less than three letters)?

Nearly 70% of people said that there were more words that started with K, even though there are nearly twice as many words with K in the third position as there are words that start with K. But since words that start with K are easier to generate in one’s mind, people overestimate their relative frequency.

In another, more real-world, experiment,32 subjects were divided into two groups. One group was asked to spend a period of time imagining its college football team doing well during the upcoming season, and the other group was asked to imagine its college football team doing poorly. Then, both groups were asked questions about the team’s actual prospects. Of the subjects who had imagined the team doing well, 63% predicted an excellent season. Of the subjects who had imagined the team doing poorly, only 40% did so.

The same researcher performed another experiment before the 1976 presidential election. Subjects asked to imagine Carter winning were more likely to predict that he would win, and subjects asked to imagine Ford winning were more likely to believe he would win. This kind of experiment has also been replicated several times, and uniformly demonstrates that considering a particular outcome in one’s imagination makes it appear more likely later.

The vividness of memories is another aspect of the availability heuristic that has been studied. People’s decisions are more affected by vivid information than by pallid, abstract, or statistical information.

Here’s just one of many experiments that demonstrates this.33 In the first part of the experiment, subjects read about a court case involving drunk driving. The defendant had run a stop sign while driving home from a party and collided with a garbage truck. No blood alcohol test had been done, and there was only circumstantial evidence to go on. The defendant was arguing that he was not drunk.

After reading a description of the case and the defendant, subjects were divided into two groups and given eighteen individual pieces of evidence to read: nine written by the prosecution about why the defendant was guilty, and nine written by the defense about why the defendant was innocent. Subjects in the first group were given prosecution evidence written in a pallid style and defense evidence written in a vivid style, while subjects in the second group were given the reverse.

For example, here is a pallid and vivid version of the same piece of prosecution evidence:
On his way out the door, Sanders [the defendant] staggers against a serving table, knocking a bowl to the floor.
On his way out the door, Sanders staggered against a serving table, knocking a bowl of guacamole dip to the floor and splattering guacamole on the white shag carpet.

And here’s a pallid and vivid pair for the defense:
The owner of the garbage truck admitted under cross-examination that his garbage truck is difficult to see at night because it is grey in color.
The owner of the garbage truck admitted under cross-examination that his garbage truck is difficult to see at night because it is grey in color. The owner said his trucks are grey “because it hides the dirt,” and he said, “What do you want, I should paint them pink?”

After all of this, the subjects were asked about the defendant’s drunkenness level, his guilt, and what verdict the jury should reach.

The results were interesting. The vivid vs. pallid arguments had no significant effect on the subject’s judgment immediately after reading them, but when they were asked again about the case 48 hours later–they were asked to make their judgments as though they “were deciding the case now for the first time”–they were more swayed by the vivid arguments. Subjects who read vivid defense arguments and pallid prosecution arguments were much more likely to judge the defendant innocent, and subjects who read the vivid prosecution arguments and pallid defense arguments were much more likely to judge him guilty.

The moral here is that people will be persuaded more by a vivid, personal story than they will by bland statistics and facts, possibly solely due to the fact that they remember vivid arguments better.

Another experiment34 divided subjects into two groups, who then read about a fictional disease called “Hyposcenia-B.” Subjects in the first group read about a disease with concrete and easy-to-imagine symptoms: muscle aches, low energy level, and frequent headaches. Subjects in the second group read about a disease with abstract and difficult-to-imagine symptoms: a vague sense of disorientation, a malfunctioning nervous system, and an inflamed liver.

Then each group was divided in half again. Half of each half was the control group: they simply read one of the two descriptions and were asked how likely they were to contract the disease in the future. The other half of each half was the experimental group: they read one of the two descriptions “with an eye toward imagining a three-week period during which they contracted and experienced the symptoms of the disease,” and then wrote a detailed description of how they thought they would feel during those three weeks. And then they were asked whether they thought they would contract the disease.

The idea here was to test whether the ease or difficulty of imagining something affected the availability heuristic. The results showed that those in the control group–who read either the easy-to-imagine or difficult-to-imagine symptoms, showed no difference. But those who were asked to imagine the easy-to-imagine symptoms thought they were more likely to contract the disease than the control group, and those who were asked to imagine the difficult-to-imagine symptoms thought they were less likely to contract the disease than the control group. The researchers concluded that imagining an outcome alone is not enough to make it appear more likely; it has to be something easy to imagine. And, in fact, an outcome that is difficult to imagine may actually appear to be less likely.

Additionally, a memory might be particularly vivid precisely because it’s extreme, and therefore unlikely to occur. In one experiment,35 researchers asked some commuters on a train platform to remember and describe “the worst time you missed your train” and other commuters to remember and describe “any time you missed your train.” The incidents described by both groups were equally awful, demonstrating that the most extreme example of a class of things tends to come to mind when thinking about the class.

More generally, this kind of thing is related to something called “probability neglect”: the tendency of people to ignore probabilities in instances where there is a high emotional content.36 Security risks certainly fall into this category, and our current obsession with terrorism risks at the expense of more common risks is an example.

The availability heuristic also explains hindsight bias. Events that have actually occurred are, almost by definition, easier to imagine than events that have not, so people retroactively overestimate the probability of those events. Think of “Monday morning quarterbacking,” exemplified both in sports and in national policy. “He should have seen that coming” becomes easy for someone to believe.

The best way I’ve seen this all described is by Scott Plous:
In very general terms: (1) the more available an event is, the more frequent or probable it will seem; (2) the more vivid a piece of information is, the more easily recalled and convincing it will be; and (3) the more salient something is, the more likely it will be to appear causal.37

Here’s one experiment that demonstrates this bias with respect to salience.38 Groups of six observers watched a two-man conversation from different vantage points: either seated behind one of the men talking or sitting on the sidelines between the two men talking. Subjects facing one or the other conversants tended to rate that person as more influential in the conversation: setting the tone, determining what kind of information was exchanged, and causing the other person to respond as he did. Subjects on the sidelines tended to rate both conversants as equally influential.

As I said at the beginning of this section, most of the time the availability heuristic is a good mental shortcut. But in modern society, we get a lot of sensory input from the media. That screws up availability, vividness, and salience, and means that heuristics that are based on our senses start to fail. When people were living in primitive tribes, if the idea of getting eaten by a saber-toothed tiger was more available than the idea of getting trampled by a mammoth, it was reasonable to believe that–for the people in the particular place they happened to be living–it was more likely they’d get eaten by a saber-toothed tiger than get trampled by a mammoth. But now that we get our information from television, newspapers, and the Internet, that’s not necessarily the case. What we read about, what becomes vivid to us, might be something rare and spectacular. It might be something fictional: a movie or a television show. It might be a marketing message, either commercial or political. And remember, visual media are more vivid than print media. The availability heuristic is less reliable, because the vivid memories we’re drawing upon aren’t relevant to our real situation. And even worse, people tend not to remember where they heard something—they just remember the content. So even if, at the time they’re exposed to a message, they don’t find the source credible, eventually their memory of the source of the information degrades and they’re just left with the message itself.

We in the security industry are used to the effects of the availability heuristic. It contributes to the “risk du jour” mentality we so often see in people. It explains why people tend to overestimate rare risks and underestimate common ones.39 It explains why we spend so much effort defending against what the bad guys did last time, and ignore what new things they could do next time. It explains why we’re worried about risks that are in the news at the expense of risks that are not, or rare risks that come with personal and emotional stories at the expense of risks that are so common they are only presented in the form of statistics.