Less Wrong is a community blog devoted to refining the art of human rationality. Please visit our About page for more information.

Availability

19 Post author: Eliezer_Yudkowsky 06 September 2007 06:55AM

The availability heuristic is judging the frequency or probability of an event, by the ease with which examples of the event come to mind.

A famous 1978 study by Lichtenstein, Slovic, Fischhoff, Layman, and Combs, "Judged Frequency of Lethal Events", studied errors in quantifying the severity of risks, or judging which of two dangers occurred more frequently.  Subjects thought that accidents caused about as many deaths as disease; thought that homicide was a more frequent cause of death than suicide.  Actually, diseases cause about 16 times as many deaths as accidents, and suicide is twice as frequent as homicide.

An obvious hypothesis to account for these skewed beliefs is that murders are more likely to be talked about than suicides - thus, someone is more likely to recall hearing about a murder than hearing about a suicide.  Accidents are more dramatic than diseases - perhaps this makes people more likely to remember, or more likely to recall, an accident.  In 1979, a followup study by Combs and Slovic showed that the skewed probability judgments correlated strongly (.85 and .89) with skewed reporting frequencies in two newspapers.  This doesn't disentangle whether murders are more available to memory because they are more reported-on, or whether newspapers report more on murders because murders are more vivid (hence also more remembered).  But either way, an availability bias is at work.

Selective reporting is one major source of availability biases.  In the ancestral environment, much of what you knew, you experienced yourself; or you heard it directly from a fellow tribe-member who had seen it.  There was usually at most one layer of selective reporting between you, and the event itself.  With today's Internet, you may see reports that have passed through the hands of six bloggers on the way to you - six successive filters.  Compared to our ancestors, we live in a larger world, in which far more happens, and far less of it reaches us - a much stronger selection effect, which can create much larger availability biases.

In real life, you're unlikely to ever meet Bill Gates.  But thanks to selective reporting by the media, you may be tempted to compare your life success to his - and suffer hedonic penalties accordingly.  The objective frequency of Bill Gates is 0.00000000015, but you hear about him much more often.  Conversely, 19% of the planet lives on less than $1/day, and I doubt that one fifth of the blog posts you read are written by them.

Using availability seems to give rise to an absurdity bias; events that have never happened, are not recalled, and hence deemed to have probability zero.  When no flooding has recently occurred (and yet the probabilities are still fairly calculable), people refuse to buy flood insurance even when it is heavily subsidized and priced far below an actuarially fair value.  Kunreuther et. al. (1993) suggests underreaction to threats of flooding may arise from "the inability of individuals to conceptualize floods that have never occurred... Men on flood plains appear to be very much prisoners of their experience... Recently experienced floods appear to set an upward bound to the size of loss with which managers believe they ought to be concerned."

Burton et. al. (1978) report that when dams and levees are built, they reduce the frequency of floods, and thus apparently create a false sense of security, leading to reduced precautions. While building dams decreases the frequency of floods, damage per flood is afterward so much greater that average yearly damage increases.

The wise would extrapolate from a memory of small hazards to the possibility of large hazards.  Instead, past experience of small hazards seems to set a perceived upper bound on risk.  A society well-protected against minor hazards takes no action against major risks, building on flood plains once the regular minor floods are eliminated.  A society subject to regular minor hazards treats those minor hazards as an upper bound on the size of the risks, guarding against regular minor floods but not occasional major floods.

Memory is not always a good guide to probabilities in the past, let alone the future.


Burton, I., Kates, R. and White, G. 1978. Environment as Hazard. New York: Oxford University Press.

Combs, B. and Slovic, P. 1979. Causes of death: Biased newspaper coverage and biased judgments. Journalism Quarterly, 56: 837-843.

Kunreuther, H., Hogarth, R. and Meszaros, J. 1993. Insurer ambiguity and market failure. Journal of Risk and Uncertainty, 7: 71-87.

Lichtenstein, S., Slovic, P., Fischhoff, B., Layman, M. and Combs, B. 1978. Judged Frequency of Lethal Events. Journal of Experimental Psychology: Human Learning and Memory, 4(6), November: 551-78.

Comments (11)

Sort By: Old
Comment author: michael_vassar3 06 September 2007 07:12:33AM 0 points [-]

Hmm. Usually you can get a strong indicator of the probability of future hazards of a given size by using frequentist statistics, e.g. by finding a statistical distribution that seems to constitute a good, simple, and logically reasonable (matching the causual structure of the underlying phenomenon). You can, for instance, estimate as I do that the distribution of historical flu risks in particular or epidemic risks in general is heavily weighted towards a few large events, and that the probabilities of events many times larger than the largest historical events can be calculated with useful precision. Much more controvercially, I see the distribution of technological innovations as a function of complexity as evidence that China and India are not good candidates for developing molecular nanotech. OTOH, the flooding example with dams gives a counter-example where the useful data from which the distribution could be inferred has been removed.

Comment author: Joshua_Fox 06 September 2007 10:17:43AM 0 points [-]

The following does not invalidate the argument in the posting, but:

Subjects thought that accidents caused about as many deaths as disease

I want to eliminate aging and death as much as anyone, but I would say that many deaths from disease in old age should be filed under "old age" rather than "disease." I wonder how the statistics work out if we look at it that way. (Or maybe Lichtenstein et al did so already.)

Comment author: Robin_Hanson2 06 September 2007 11:25:51AM 2 points [-]

I'm surprised to hear that dams increase average annual damages. Does the Burton book explain how that works? Is it reduced preparation increasing the effect of the largest events?

Comment author: Zubon2 06 September 2007 02:31:32PM 2 points [-]

Concerning age and death, the more recent links are not working for me right now, but here is the CDC with 2003 numbers: ftp://ftp.cdc.gov/pub/ncipc/10LC-2003/PDF/10lc-2003.pdf

Until age 34, accidents are winning, with intentional injury (suicide and homicide) taking second and third. 35-44, accidents are still #1, but cancer and heart disease are each close so disease wins. Cancer wins through 64, then heart disease takes over. Because disease reigns supreme 55+, unintentional injuries fall to #5 overall, and intentional injuries fall off the chart entirely.

If you are talking about young people, yes, accidents win. The main component of that is traffic crashes; in older adults, falls start to come in. Suicide beats homicide in every age category except 15-24 (and the very small 1-9 age group).

On a side note, it looks like the majority of deaths in the first year are things that might be classified as "stillborn" in another country or century. Those deaths in the <1 category rival all deaths from all other causes through age 14.

Comment author: Pseudonymous2 06 September 2007 06:38:35PM 1 point [-]

I imagine it is a lot easier to avoid an accident than avoid most modern diseases eg cancer. So it make sense to concentrate on the risks you can do most about.

people refuse to buy flood insurance even when it is heavily subsidized and priced far below an actuarially fair value.

How do they know it is heavily subsidized and priced far below an actuarially fair value?

Is it worth going to all the trouble of finding out?

Comment author: gwern 19 August 2011 06:26:58PM 1 point [-]

How do they know it is heavily subsidized and priced far below an actuarially fair value?

I don't think they realize this. Recently in my area some flood maps were updated to take account of new data suggesting increased risks, with subsequent increases in the subsidized flood insurance rates. The affected households raised a hue and cry, enlisting senators in their cause, and many got the increases reversed. Nothing in the rhetoric I read suggested that they realized they were already getting a great deal; instead, here and in the Florida articles I read (where the government objected to the remaining insurer increasing rates after recent hurricanes), the unstated assumption seemed to be that the rates were 'unfair' and profitable.

Is it worth going to all the trouble of finding out?

A house is hundreds of thousands of dollars, and the disruption to your life if it and its contents are destroyed is profound. In some place like Florida, it may be more likely than not that in your lifetime your house will be damaged or destroyed, especially given the suggestion that global warming will increase the variance of storms (and hence the occurrence of super-hurricanes). I think it is worthwhile!

Comment author: Hopefully_Anonymous 06 September 2007 06:46:27PM 0 points [-]

Good question, Pseudonymous. I'm interested in the people that do buy insurance when it's in their rational self-interest, and what's different about them.

Comment author: Bob_Knaus 06 September 2007 11:43:05PM 1 point [-]

In the Bahamas the homicide rate is about 15 times greater than the suicide rate: http://en.wikipedia.org/wiki/List_of_countries_by_homicide_rate http://en.wikipedia.org/wiki/List_of_countries_by_suicide_rate

Some of this may stem from cultural reluctance to identify suicide as such... but I think the majority of it is simply the mark of a violent society.

BTW, I love the Bahamas, I spend 9 months a year sailing there. It may be a troubled paradise, but nonetheless it remains a paradise.

Comment author: K.A. 09 September 2007 11:34:44PM 0 points [-]

The bias probably results because risks that people have less control over (homicide) would be more important to remember than the ones that are primarily due to one's own life decisions (suicide, health practices). The former risks seem unjust and avoidance practices still need to be learned as a means of living adaptively in an uncertain environment, vs. the latter risks, which already seem to be under our control.

Comment author: Constant2 16 October 2007 08:56:47AM 1 point [-]

You can properly use fiction as a shared language. Complex scenarios that would take long to explain can be referenced conveniently by way of a movie or book name. For example, two key classes of future, which are seriously discussed, are, one, the AIs subjugate us, and two, we enslave the AIs. These are not exhaustive but they are of particular interest to us, as is the more general topic of rights in a future with AIs, both human rights and AI rights. I have seen serious discussions of this, not based on movies. Science fiction, and fiction generally, responds to serious concerns, so whatever our concern, we can often find a fiction that we can use as a reference to help us efficiently convey our concern to someone else. Here the fiction is not being used as evidence but as a common language. Like that Star Trek episode in which a race communicates by talking about legends. "Darmok and Jalad at Tanagra."

Comment author: MoreOn 20 August 2011 01:45:27AM 2 points [-]

Subjects thought that accidents caused about as many deaths as disease.

Lichtenstein et aliōrum research subjects were 1) college students and 2) members of a chapter of the League of Women Voters. Students thought that accidents are 1.62 times more likely than diseases, and league members thought they were 11.6 times more likely (geometric mean). Sadly, no standard deviation was given. The true value is 15.4. Note that only 57% and 79% of students and league members respectively got the direction right, which further biased the geometric average down.

There were some messed up answers. For example, students thought that tornadoes killed more people than asthma, when in fact asthma kills 20x more people than tornadoes. All accidents are about as likely as stomach cancer (well, 1.19x more likely), but they were judged to be 29 times more likely. Pairs like these represent a minority, and subjects were generally only bad at guessing which cause of death was more frequent when the ratio was less than 2:1. These are the graphs from the paper.

The following excerpt is from Judged Frequency Of Lethal Events by Lichtenstein, Slovic, Fischhoff, Layman and Combs.

Instructions. The subjects' instructions read as follows:

Each item in part one consists of two different possible causes of death. The question you are to answer is: Which cause of death is more likely? We do not mean more likely for you, we mean more likely in general, in the United States.

Consider all the people now living in the United States—children, adults, everyone. Now supposing we randomly picked just one of those people. Will that person more likely die next year from cause A or cause B ? For example: Dying in a bicycle accident versus dying from an overdose of heroin. Death from each cause is remotely possible. Our question is, which of these two is the more likely cause of death?

For each pair of possible causes of death, A and B, we want you to mark on your answer sheet which cause you think is MORE LIKELY. Next, we want you to decide how many times more likely this cause of death is, as compared with the other cause of death given in the same item. The pairs we use vary widely in their relative likelihood. For one pair, you may think that the two causes are equally likely. If so, you should write the number 1 in the space provided for that pair. Or, you may think that one cause of death is 10 times, or 100 times, or even a million times as likely as the other cause of death. You have to decide: How many times as likely is the more likely cause of death? Write the number in the space provided. If you think it's twice as likely, write 2. If it's 10 thousand times as likely, write 10,000, and so forth.

There were more instructions about relative likelihoods and scales. And there was a glossary to help the people understand some categories.

All accidents: includes any kind of accidental event; excludes diseases and natural disasters (floods, tornadoes, etc.).

All cancer: includes leukemia.

Cancer of the digestive system: includes cancer of stomach, alimentary tract, esophagus, and intestines.

Excess cold: freezing to death or death by exposure.

Nonvenomous animal: dogs, bears, etc.

Venomous bite or sting: caused by snakes, bees, wasps, etc.

Note that there was nothing about “old age” anywhere. There is no such thing as “death by old age,” but I’ll risk generalizing from my own example to say that some people think there is. And even those who know there isn’t might think, despite the instructions, “Oh, darnit, I forgot that old people count, too.”

I wish I’d tested myself BEFORE reading the correct answer. As near as I could tell, I would’ve been correct about homicide vs. suicide, but wrong about diseases vs. accidents (“Old people count, too!” facepalm). I wouldn’t even bother guessing the relative frequency. I didn’t have a clue.

When I need to know the number of square feet in an acre, or the world population it takes me seconds to get from the question to the answer. I dutifully spent ~20 minutes googling the CDC website, looking for this. It wasn’t even some heroic effort, but it’s not something I, or most other people, would casually expend on every question that starts with, “Huh, I wonder….” (we should, but we don’t).

As for what I found: I dare you, click on my link and see table 9. (http://www.cdc.gov/NCHS/data/nvsr/nvsr58/nvsr58_19.pdf). Did you? If you did, you would’ve seen that Zubon2 was right in this comment. Accidents win by quite a margin in the 15-44 demographic. I couldn’t find 1978 data, but I’d expect it to be similar (Lichtenstein’s et al tables are no help because they pool all age groups).

I spent the last two hours looking at these tables. Ask me anything! … I won’t be able to answer. Unless I have the CDC tables in front of me, I might not even do much better on Lichtenstein et aliōrum questionnaire than a typical subject (well, at least, I know tornadoes have frequency; measles doesn’t—I’ll get that question right). I suppose that people who haven’t looked at the CDC table are getting all of their information from fragmented reports like “Drive safely! Traffic accidents is the leading cause of death among teenagers who <insert condition>!” or “Buy our drug! <What it cures> is the leading cause of death in <insert condition> over 55!” or “5-star exhaust pipe crash safety rating!” Humans aren’t good at integrating these fragments.

Memory is a bad guide to probability estimates. But what’s the alternative? Should we carry tables around with us?

Personally, I hope that someday data that is already out there in the public domain will be made easily accessible. I hope that finding the relative frequencies of measles-related deaths and tornado-related deaths will be as quick as finding the number of square feet in an acre or the world population, and that political squabble will focus on whether or not certain data should be in the public domain (“You can’t force hospitals to put their data online! That violates the patients’ right to privacy!” “Well, but….”)

Note: repost from SEQ RERUN.