# Archive for category Probability and Risk

### The Prosecutor’s Fallacy Illustrated

Posted by Bill Storage in Probability and Risk on May 7, 2020

*“The first thing we do, let’s kill all the lawyers.”* – Shakespeare, *Henry VI, Part 2*, Act IV

My last post discussed the failure of most physicians to infer the chance a patient has the disease given a positive test result where both the frequency of the disease in the population and the accuracy of the diagnostic test are known. The probability that the patient has the disease can be hundreds or thousands of times lower than the accuracy of the test. The problem in reasoning that leads us to confuse these very different likelihoods is one of several errors in logic commonly called the prosecutor’s fallacy. The important concept is conditional probability. By that we mean simply that the probability of x has a value and that the probability of x given that y is true has a different value. The shorthand for probability of x is p(x) and the shorthand for probability of x given y is p(x|y).

“Punching, pushing and slapping is a prelude to murder,” said prosecutor Scott Gordon during the trial of OJ Simpson for the murder of Nicole Brown. Alan Dershowitz countered with the argument that the probability of domestic violence leading to murder was very remote. Dershowitz (not prosecutor but defense advisor in this case) was right, technically speaking. But he was either as ignorant as the physicians interpreting the lab results or was giving a dishonest argument, or possibly both. The relevant probability was not the likelihood of murder given domestic violence, it was the likelihood of murder given domestic violence *and* murder. “The courtroom oath – to tell the truth, the whole truth and nothing but the truth – is applicable only to witnesses,” said Dershowitz in *The Best Defense.* In *Innumeracy: Mathematical Illiteracy and Its Consequences.* John Allen Paulos called Dershowitz’s point “astonishingly irrelevant,” noting that utter ignorance about probability and risk “plagues far too many otherwise knowledgeable citizens.” Indeed.

The doctors’ mistake in my previous post was confusing

P(positive test result) vs.

P(disease | positive test result)

Dershowitz’s argument confused

P(husband killed wife | husband battered wife) vs.

P(husband killed wife | husband battered wife | wife was killed)

In *Reckoning With Risk*, Gerd Gigerenzer gave a 90% value for the latter Simpson probability. What Dershowitz cited was the former, which we can estimate at 0.1%, given a wife-battery rate of one in ten, and wife-murder rate of one per hundred thousand. So, contrary to what Dershowitz implied, prior battery is a strong indicator of guilt when a wife has been murdered.

As mentioned in the previous post, the relevant mathematical rule does not involve advanced math. It’s a simple equation due to Pierre-Simon Laplace, known, oddly, as Bayes’ Theorem:

P(A|B) = P(B|A) * P(A) / P(B)

If we label the hypothesis (patient has disease) as D and the test data as T, the useful form of Bayes’ Theorem is

P(D|T) = P(T|D) P(D) / P(T) where P(T) is the sum of probabilities of positive results, e.g.,

P(T) = P(T|D) * P(D) + P(T | not D) * P(not D) [using “not D” to mean “not diseased”]

Cascells’ phrasing of his Harvard quiz was as follows: “If a test to detect a disease whose prevalence is 1 out of 1,000 has a false positive rate of 5 percent, what is the chance that a person found to have a positive result actually has the disease?”

Plugging in the numbers from the Cascells experiment (with the parameters Cascells provided shown below in **bold** and the correct answer in **green**):

- P(D) is the disease frequency =
**0.001**[ 1 per 1000 in population ] therefore: - P(not D) is 1 – P(D) = 0.999
- P(T | not D) =
**5%**= 0.05 [ false positive rate also 5%] therefore: - P(T | D) = 95% = 0.95 [ i.e, the false negative rate is 5% ]

Substituting:

P(T) = .95 * .001 + .999 * .05 = 0.0509 ≈ 5.1% [ total probability of a positive test ]

P(D|T) = .95 * .001 / .0509 = .0019 ≈ **2% ** [ probability that patient has disease, given a positive test result ]

Voila.

I hope this *seeing is believing* illustration of Cascells’ experiment drives the point home for those still uneasy with equations. I used Cascells’ rates and a population of 100,000 to avoid dealing with fractional people:

Extra credit: how exactly does this apply to Covid, news junkies?

Edit 5/21/20. An astute reader called me on an inaccuracy in the diagram. I used an approximation, without identifying it. P = r1/r2 is a cheat for P = 1 – Exp(- r1/r2). The approximation is more intuitive, though technically wrong. It’s a good cheat, for P values less that 10%.

Note 5/22/20. In response to questions about how this sort of thinking bears on coronavirus testing -what test results say about prevalence – consider this. We really have one equation in 3 unknowns here: false positive rate, false negative rate, and prevalence in population. A quick Excel variations study using false positive rates from 1 to 20% and false neg rates from 1 to 3 percent, based on a quick web search for proposed sensitivity/specificity for the Covid tests is revealing. Taking the low side of the raw positive rates from the published data (1 – 3%) results in projected prevalence roughly equal to the raw positive rates. I.e., the false positives and false negatives happen to roughly wash out in this case. That also leaves P(d|t) in the range of a few percent.

### The Trouble with Doomsday

Posted by Bill Storage in Philosophy, Probability and Risk on February 4, 2020

Doomsday just isn’t what is used to be. Once the dominion of ancient apologists and their votary, the final destiny of humankind now consumes probability theorists, physicists. and technology luminaries. I’ll give some thoughts on probabilistic aspects of the *doomsday argument* after a brief comparison of ancient and modern apocalypticism.

**Apocalypse Then**

The Israelites were enamored by eschatology. “The Lord is going to lay waste the earth and devastate it,” wrote Isaiah, giving few clues about when the wasting would come. The early Christians anticipated and imminent end of days. Matthew 16:27: s*ome of those who are standing here will not taste death until they see the Son of Man coming in His kingdom.*

From late antiquity through the middle ages, preoccupation with the Book of Revelation led to conflicting ideas about the finer points of “domesday,” as it was called in Middle English. The first millennium brought a flood of predictions of, well, *flood*, along with earthquakes, zombies, lakes of fire and more. But a central Christian apocalyptic core was always beneath these varied predictions.

Right up to the enlightenment, punishment awaited the unrepentant in a final judgment that, despite Matthew’s undue haste, was still thought to arrive any day now. Disputes raged over whether the rapture would be precede the tribulation or would follow it, the proponents of each view armed with supporting scripture. Polarization! When Christianity began to lose command of its unruly flock in the 1800’s, Nietzsche wondered just what a society of non-believers would find to flog itself about. If only he could see us now.

**Apocalypse Now**

Our modern doomsday riches include options that would turn an ancient doomsayer green. Alas, at this eleventh hour we know nature’s annihilatory whims, including global pandemic, supervolcanoes, asteroids, and killer comets. Still in the *Acts of God* department, more learned handwringers can sweat about earth orbit instability, gamma ray bursts from nearby supernovae, or even a fluctuation in the Higgs field that evaporates the entire universe.

As Stephen Hawking explained bubble nucleation, the Higgs field might be metastable at energies above a certain value, causing a region of false vacuum to undergo catastrophic vacuum decay, causing a bubble of the true vacuum expanding at the speed of light. This might have started eons ago, arriving at your doorstep before you finish this paragraph. Harold Camping, eat your heart out.

Hawking also feared extraterrestrial invasion, a view hard to justify with probabilistic analyses. Glorious as such cataclysms are, they lack any element of contrition. Real apocalypticism needs a guilty party.

Thus anthropogenic climate change reigned for two decades with no creditable competitors. As self-inflicted catastrophes go, it had something for everyone. Almost everyone. Verily, even Pope Francis, in a covenant that astonished adherents, joined – with strong hand and outstretched arm – leftists like Naomi Oreskes, who shares little else with the Vatican, ideologically speaking.

While Global Warming is still revered, some prophets now extend the hand of fellowship to some budding successor fears, still tied to devilries like capitalism and the snare of scientific curiosity. Bioengineered coronaviruses might be invading as we speak. Careless researchers at the Large Hadron Collider could set off a mini black hole that swallows the earth. So some think anyway.

Nanotechnology now gives some prominent intellects the willies too. My favorite in this realm is Gray Goo, a catastrophic chain of events involving molecular nanobots programmed for self-replication. They will devour all life and raw materials at an ever-increasing rate. How they’ll manage this without melting themselves due to the normal exothermic reactions tied to such processes is beyond me. Global Warming activists may become jealous, as the very green Prince Charles himself now diverts a portion of the crown’s royal dread to this upstart alternative apocalypse.

My cataclysm bucks are on full-sized Artificial Intelligence though. I stand with chief worriers Bill Gates, Ray Kurzweil, and Elon Musk. Computer robots will invent and program smarter and more ruthless autonomous computer robots on a rampage against humans seen by the robots as obstacles to their important business of building even smarter robots. Game over.

**The Mathematics of Doomsday**

The Doomsday Argument is a mathematical proposition arising from the Copernican principle – a trivial application of Bayesian reasoning – wherein we assume that, lacking other info, we should find ourselves, roughly speaking, in the middle of the phenomenon of interest. Copernicus didn’t really hold this view, but 20th century thinkers blamed him for it anyway.

Applying the Copernican principle to human life starts with the knowledge that we’ve been around for 200 hundred thousand years, during which 60 billion of us have lived. Copernicans then justify the belief that half the humans that will have ever lived remain to be born. With an expected peak earth population of 12 billion, we might, using this line of calculating, expect the human race to go extinct in a thousand years or less.

Adding a pinch of statistical rigor, some doomsday theorists calculate a 95% probability that the number of humans to have lived so far is less than 20 times the number that will ever live. Positing individual life expectancy of 100 years and 12 billion occupants, the earth will house humans for no more than 10,000 more years.

That’s the gist of the dominant doomsday argument. Notice that it is purely probabilistic. It applies equally to the Second Coming and to Gray Goo. However, its math and logic are both controversial. Further, I’m not sure why its proponents favor population-based estimates over time-based estimates. That is, it took a lot longer than 10,000 years, the proposed P = .95 extinction term, for the race to arrive at our present population. So why not place the current era in the middle of the duration of the human race, thereby giving us another 200,000 thousand years? That’s quite an improvement on the 10,000 year prediction above.

Even granting that improvement, all the above doomsday logic has some curious bugs. If we’re justified in concluding that we’re midway through our reign on earth, then should we also conclude we’re midway through the existence of agriculture and cities? If so, given that cities and agriculture emerged 10,000 years ago, we’re led to predict a future where cities and agriculture disappear in 10,000 years, followed by 190,000 years of post-agriculture hunter-gatherers. Seems unlikely.

Astute Bayesian reasoners might argue that all of the above logic relies – unjustifiably – on an *uninformative prior*. But we have prior knowledge suggesting we don’t happen to be at some random point in the life of mankind. Unfortunately, we can’t agree on which direction that skews the outcome. My reading of the evidence leads me to conclude we’re among the first in a long line of civilized people. I don’t share Elon Musk’s pessimism about killer AI. And I find Hawking’s extraterrestrial worries as facile as the anti-GMO rantings of the Union of Concerned Scientists. You might read the evidence differently. Others discount the evidence altogether, and are simply swayed by the fashionable pessimism of the day.

Finally, the above doomsday arguments all assume that we, as observers, are randomly selected from the set of all existing humans, including past, present and future, ever be born, as opposed to being selected from all possible births. That may seem a trivial distinction, but, on close inspection, becomes profound. The former is analogous to *Theory 2* in my previous post, *The Trouble with Probability*. This particular observer effect, first described by Dennis Dieks in 1992, is called the *self-sampling assumption* by Nick Bostrom*. C*onsidering yourself to be randomly selected from all possible births prior to human extinction is the analog of *Theory 3* in my last post. It arose from an equally valid assumption about sampling. That assumption, called *self-indication* by Bostrom, confounds the above doomsday reasoning as it did the hotel problem in the last post.

Th self-indication assumption holds that we should believe that we’re more likely to discover ourselves to be members of larger sets than of smaller sets. As with the hotel room problem discussed last time, self-indication essentially cancels out the self-sampling assumption. We’re more likely to be in a long-lived human race than a short one. In fact, setting aside some secondary effects, we can say that the likelihood of being selected into any set is proportional to the size of the set; and here we are in the only set we know of. Doomsday hasn’t been called off, but it has been postponed indefinitely.

### The Trouble with Probability

Posted by Bill Storage in Probability and Risk on February 2, 2020

The trouble with probability is that no one agrees what it means.

Most people understand probability to be about predicting the future and statistics to be about the frequency of past events. While everyone agrees that probability and statistics should have something to do with each other, no one agrees on what that something is.

Probability got a rough start in the world of math. There was no concept of probability as a discipline until about 1650 – odd, given that gambling had been around for eons. Some of the first serious work on probability was done by Blaise Pascal, who was assigned by a nobleman to divide up the winnings when a dice game ended unexpectedly. Before that, people just figured chance wasn’t receptive to analysis. Aristotle’s idea of knowledge required that it be universal and certain. Probability didn’t fit.

To see how fast the concept of probability can go haywire, consider your chance of getting lung cancer. Most agree that probability is determined by your membership in a reference class for which a historical frequency is known. Exactly which reference class you belong to is always a matter of dispute. How similar to them do you need to be? The more accurately you set the attributes of the reference population, the more you narrow it down. Eventually, you get down to people of your age, weight, gender, ethnicity, location, habits, and genetically determined preference for ice cream flavor. Your reference class then has a size of one – you. At this point your probability is either zero or one, and nothing in between. The historical frequency of cancer within this population (you) cannot predict your future likelihood of cancer. That doesn’t seem like what we wanted to get from probability.

Similarly, in the real world, the probabilities of uncommon events and of events with no historical frequency at all are the subject of keen interest. For some predictions of previously unexperienced events, like and airplane crashing due to simultaneous failure of a certain combination of parts, even though that combination may have never occurred in the past, we can assemble a probability from combining historical frequencies of the relevant parts using Boolean logic. My hero Richard Feynman seemed not to grasp this, oddly.

For worries like a large city being wiped out by an asteroid, our reasoning becomes more conjectural. But even for asteroids we can learn quite a bit about asteroid impact rates based on the details of craters on the moon, where the craters don’t weather away so fast as they do on earth. You can see that we’re moving progressively away from historical frequencies and becoming more reliant on inductive reasoning, the sort of thing that gave Aristotle the hives.

Finally, there are some events for which historical frequencies provide no useful information. The probability that nanobots will wipe out the human race, for example. In these cases we take a guess, maybe even a completely wild guess. and then, on incrementally getting tiny bits of supporting or nonsupporting evidence, we modify our beliefs. This is the realm of Bayesianism. In these cases when we talk about probability we are really only talking about the degree to which we believe a proposition, conjecture or assertion.

Breaking it down a bit more formally, a handful of related but distinct interpretations of probability emerge. Those include, for example:

Objective chances: The physics of flipping a fair coin tend to result in heads half the time.

Frequentism: Relative frequency across time: of all the coins ever flipped, one half have been heads, so expect more of the same.

Hypothetical frequentism: If you flipped coins forever, the heads/tails ratio would approach 50%.

Bayesian belief: Prior distributions equal belief: before flipping a coin, my personal expectation that it will be heads is equal to that of it being tails.

Objective Bayes: Prior distributions represent neutral knowledge: given only that a fair coin has been flipped, the plausibility of it’s having fallen heads equals that of it having been tails.

While those all might boil down to the same thing in the trivial case of a coin toss, they can differ mightily for difficult questions.

People’s ideas of probability differ more than one might think, especially when it becomes personal. To illustrate, I’ll use a problem derived from one that originated either with Nick Bostrom, Stuart Armstrong or Tomas Kopf, and was later popularized by Isaac Arthur. Suppose you wake up in a room after suffering amnesia or a particularly bad night of drinking. You find that you’re part of a strange experiment. You’re told that you’re in one of 100 rooms and that the door of your room is either red or blue. You’re instructed to guess which color it is. Finding a coin in your pocket you figure flipping it is as good a predictor of door color as anything else, regardless of the ratio of red to blue doors, which is unknown to you. Heads red, tails blue.

The experimenter then gives you new info. 90 doors are red and 10 doors are blue. Guess your door color, says the experimenter. Most people think, absent any other data, picking red is a 4 1/2 times better choice than letting a coin flip decide.

Now you learn that the evil experimenter had designed two different branches of experimentation. In Experiment A, ten people would be selected and placed, one each, into rooms 1 through 10. For Experiment B, 100 other people would be placed, one each, in all 100 rooms. You don’t know which room you’re in or which experiment, A or B, was conducted. The experimenter tells you he flipped a coin to choose between Experiment A, heads, and Experiment B, tails. He wants you to guess which experiment, A or B, won his coin toss. Again, you flip your coin to decide, as you have nothing to inform a better guess. You’re flipping a coin to guess the result of his coin flip. Your odds are 50-50. Nothing controversial so far.

Now you receive new information. You are in Room 5. What judgment do you now make about the result of his flip? Some will say that the odds of experiment A versus B were set by the experimenter’s coin flip, and are therefore 50-50. Call this Theory 1.

Others figure that your chance of being in Room 5 under Experiment A is 1 in 10 and under Experiment B is 1 in 100. Therefore it’s ten times more likely that Experiment A was the outcome of the experimenter’s flip. Call this Theory 2.

Still others (Theory 3) note that having been selected into a group of 100 was ten times more likely than having been selected into a group of 10, and on that basis it is ten times more likely that Experiment B was the result of the experimenter’s flip than Experiment A.

My experience with inflicting this problem on victims is that most people schooled in science – though certainly not all – prefer Theories 2 or 3 to Theory 1, suggesting they hold different forms of Bayesian reasoning. But between Theories 2 and 3, war breaks out.

Those preferring Theory 2 think the chance of having been selected into Experiment A (once it became the outcome of the experimenter’s coin flip) is 10 in 110 and the chance of being in Room 5 is 1 in 10, given that Experiment A occurred. Those who hold Theory 3 perceive a 100 in 110 chance of having been selected into Experiment B, once it was selected by the experimenter’s flip, and then a 1 in 100 chance of being in Room 5, given Experiment B. The final probabilities of being in room 5 under Theories 2 and 3 are equal (10/110 x 1/10 equals 1 in 110, vs. 100/110 x 1/100 also equals 1 in 110), but the answer to the question about the outcome of the experimenter’s coin flip having been heads (Experiment A) and tails (Experiment B) remains in dispute. To my knowledge, there is no basis for settling that dispute. Unlike Martin Gardner’s boy-girl paradox, this dispute does not result from ambiguous phrasing; it seems a true paradox.

The trouble with probability makes it all the more interesting. Is it math, philosophy, or psychology?

*How dare we speak of the laws of chance. Is not chance the antithesis of all law?* – Joseph Bertrand, *Calcul des probabilités*, 1889

*Though there be no such thing as Chance in the world; our ignorance of the real cause of any event has the same influence on the understanding, and begets a like species of belief or opinion.* – David Hume, *An Enquiry Concerning Human Understanding, 1748*

*It is remarkable that a science which began with the consideration of games of chance should have become the most important object of human knowledge.* – Blaise Pascal, *Théorie Analytique des Probabilitiés*, 1812

### A short introduction to small data

Posted by Bill Storage in Probability and Risk on January 13, 2020

How many children are abducted each year? Did you know anyone who died in Vietnam?

Wikipedia explains that big data is about correlations, and that small data is either about the causes of effects, or is an inference from big data. None of that captures what I mean by small data.

Most people in my circles instead think small data deals with inferences about populations made from the sparse data from within those populations. For Bayesians, this means making best use of an intuitive *informative prior distribution* for a model. For wise non-Bayesians, it can mean bullshit detection.

In the early 90’s I taught a course on probabilistic risk analysis in aviation. In class we were discussing how to deal with estimating equipment failure rates where few previous failures were known when Todd, a friend who was attending the class, asked how many kids were abducted each year. I didn’t know. Nor did anyone else. But we all understood where Todd was going with the question.

Todd produced a newspaper clipping citing an evangelist – Billy Graham as I recall – who claimed that 50,000 children a year were abducted in the US. Todd asked if we thought that yielded a a reasonable prior distribution.

Seeing this as a sort of Fermi problem, the class kicked it around a bit. How many kids’ pictures are on milk cartons right now, someone asked (Milk Carton Kids – remember, this was pre-internet). We remembered seeing the same few pictures of missing kids on milk cartons for months. None of us knew of anyone in our social circles who had a child abducted. How does that affect your assessment of Billy Graham’s claim?

What other groups of people have 50,000 members I asked. Americans dead in Vietnam, someone said. True, about 50,000 American service men died in Vietnam (including 9000 accidents and 400 suicides, incidentally). Those deaths spanned 20 years. I asked the class if anyone had known someone, at least indirectly, who died in Vietnam (remember, this was the early 90s and most of us had once owned draft cards). Almost every hand went up. Assuming that dead soldiers and our class were roughly randomly selected implied each of our social spheres had about 4000 members (200 million Americans in 1970, divided by 50,000 deaths). That seemed reasonable, given that news of Vietnam deaths propagated through friends-of-friends channels.

Now given that most of us had been one or two degrees’ separation from someone who died in Vietnam, could Graham’s claim possibly be true? No, we reasoned, especially since news of abductions should travel through social circles as freely as Vietnam deaths. And those Vietnam deaths had spanned decades. Graham was claiming 50,000 abductions *per year*.

Automobile deaths, someone added. Those are certainly randomly distributed across income, class and ethnicity. Yes, and, oddly, they occur at a rate of about 50,000 per year in the US. Anyone know someone who died in a car accident? Every single person in the class did. Yet none of us had been close to an abduction. Abductions would have to be very skewed against aerospace engineers for our car death and abduction experience to be so vastly different given their supposedly equal occurrence rates in the larger population. But the Copernican position that we resided nowhere special in the landscapes of either abductions or automobile deaths had to be mostly valid, given the diversity of age, ethnicity and geography in the class (we spanned 30 years in age, with students from Washington, California and Missouri).

One way to check the veracity of Graham’s claim would have been to do a bunch of research. That would have been library slow and would have likely still required extrapolation and assumptions about distributions and the representativeness of whatever data we could dig up. Instead we drew a sound inference from very small data, our own sampling of world events.

We were able to make good judgments about the rate of abduction, which we were now confident was very, very much lower than one per thousand (50,000 abductions per year divide by 50 million kids). Our good judgments stemmed from our having rich priors (prior distributions) because we had sampled a lot of life and a lot of people. We had rich data about deaths from car wrecks and Vietnam, and about how many kids were not abducted in each of our admittedly small circles. Big data gets the headlines, causing many of us to forget just how good small data can be.

### Use and Abuse of Failure Mode & Effects Analysis in Business

Posted by Bill Storage in Management Science, Probability and Risk on December 5, 2019

On investigating about 80 deaths associated with the drug heparin in 2009, the FDA found that over-sulphated chondroitin with toxic effects had been intentionally substituted for a legitimate ingredient for economic reasons. That is, an unscrupulous supplier sold a counterfeit chemical costing 1% as much as the real thing and it killed people.

This wasn’t unprecedented. Gentamicin, in the late 1980s, was a similar case. Likewise Cefaclor in 1996, and again with diethylene glycol sold as glycerin in 2006.

Adulteration is an obvious failure mode of supply chains and operations for drug makers. Drug firms buying adulterated raw material had presumably conducted failure mode effects analyses at several levels. An early-stage FMEA should have seen the failure mode and assessed its effects, thereby triggering the creation of controls to prevent the process failure. So what went wrong?

The FDA’s reports on the heparin incident didn’t make public any analyses done by the drug makers. But based on the “best practices” specified by standards bodies, consulting firms, and many risk managers, we can make a good guess. Their risk assessments were likely misguided, poorly executed, gutless, and ineffective.

Promoters of FMEAs as a means of risk analysis often cite aerospace as a guiding light in matters of risk. Commercial aviation should be the exemplar of risk management. In no other endeavor has mankind made such an inherently dangerous activity so safe as commercial jet flight.

While those in pharmaceutical risk and compliance extol aviation, they mostly stray far from its methods, mindset, and values. This is certainly the case with the FMEA, a tool poorly understood, misapplied, poorly executed, and then blamed for failing to prevent catastrophe.

In the case of heparin, a properly performed FMEA exercise would certainly have identified the failure mode. *But* FMEA wasn’t even the right tool for identifying that hazard in the first place. A functional hazard anlysis (FHA) or Business Impact Analysis (BIA) would have highlighted chemical contamination leading to death of patients, supply disruption, and reputation damage as a top hazard in minutes. I know this for fact, because I use drug manufacture as an example when teaching classes on FHA. First-day students identify that hazard without being coached.

FHAs can be done very early in the conceptual phase of a project or system design. They need no implementation details. They’re short and sweet, and they yield concerns to address with high priority. Early writers on the topic of FMEA explicitly identified it as being something like the opposite of an FHA, for former being “bottom-up, the latter “top down,” NASA’s response to the USGS on the suitability of FMEAs their needs, for example, stressed this point. FMEAs rely strongly on implementation details. They produce a lot of essential but lower-value content (essential because FMEAs help confirm which failure modes can be de-prioritized) when there is an actual device or process design.

So a failure mode of risk management is using FMEAs for purposes other than those for which they were designed. *Equating* FMEA with risk analysis and risk management is a gross failure mode of management.

If industry somehow stops misusing FMEAs, they then face the hurdle of doing them well. This is a challenge, as the quality of training, guidance, and facilitation of FMEAs has degraded badly over the past twenty years.

FMEAs, as promoted by the Project Management Institute, ISO 31000, and APM PRAM, to name a few, bear little resemblance to those in aviation. I know this, from three decades of risk work in diverse industries, half of it in aerospace. You can see the differences by studying sample FMEAs on the web.

It’s anyone’s guess how FMEAs went so far astray. Some blame the explosion of enterprise risk management suppliers in the 1990s. ERM, partly rooted in the sound discipline of actuarial science, generally lacks rigor. It was up-sold by consultancies to their existing corporate clients, who assumed those consultancies actually had background in risk science, which they did not. Studies a decade later by Protiviti and the EIU failed to show any impact on profit or other benefit of ERM initiatives, except for positive self-assessments by executives of the firms.

But bad FMEAs predated the ERM era. Adopted by US automotive industry in the 1970s, sloppy FMEAs justified optimistic warranty claims estimates for accounting purposes. While Toyota was implementing statistical process control to precisely predict the warranty cost of adverse tolerance accumulation, Detroit was pretending that multiplying ordinal scales of probability, severity, and detectability was mathematically or scientifically valid.

Citing inability to quantify failure rates of basic components and assemblies (an odd claim given the abundance of warranty and repair data), auto firms began to assign scores or ranks to failure modes rather than giving probability values between zero and one. This first appears in automotive conference proceedings around 1971. Lacking hard failure rates – if in fact they did – reliability workers could have estimated numeric probability values based on subjective experience or derived them from reliability handbooks then available. Instead they began to assign ranks or scores on a 1 to 10 scale.

In principle there is no difference between guessing a probability of 0.001 (a numerical probability value) and guessing a value of “1” on a 10 scale (either an ordinal number or a probability value mapped to a limited-range score). But in practice there is a big difference.

One difference is that people estimating probability scores in facilitated FMEA sessions usually use grossly different mental mapping processes to get from labels like “extremely likely” or “moderately unlikely” to numerical probabilities. A physicist sees “likely” for a failure mode to mean more than once per million; a drug trial manager interprets it to mean more than 5%. Neither is wrong; but if those two specialists aren’t alert to the difference, when they each judge a failure *likely*, there will be a dangerous illusion of communication and agreement where none exists.

Further, FMEA participants don’t agree – and often don’t know they don’t agree – on the mapping of their probability estimates into 1-10 scores.

The resultant probability scores or ranks (as opposed to P values between zero and one) are used to generate Risk Priority Numbers (RPN), that first appeared in the American automotive industry. You won’t find RPN or anything like it in aviation FMEAs, or even the modern automotive industry. Detroit abandoned them long ago.

RPNs are defined as the arithmetic product of a probability *score*, a severity *score*, and a detection (more precisely, the inverse of detectability) *score*. The explicit thinking here is that risks can be prioritized on the basis of the product of three numbers, each ranging from 1 to 10.

An implicit – but critical, though never addressed by users of RPN – thinking here is that engineers, businesses, regulators and consumers are risk-neutral. *Risk neutrality*, as conceived in portfolio choice theory, would in this context mean that everyone would be indifferent to two risks of the same RPN, even comprising very different probability and severity values.That is, an RPN formed from the scores {2,8,4} would dictate the same risk response as failure modes with RPN scores {8,4,2} and {4,4,4} since the RPN values (product of the scores) are equal. In the real world this is never true. It is usually *very* far from true. Most of us are not risk-neutral, we’re risk-averse. That changes things. As a trivial example, banks might have valid reasons for caring more about a single $100M loss than one hundred $1M losses.

Beyond the implicit assumption of risk-neutrality, RPN has other problems. As mentioned above, there both cognitive and group-dynamics problems arise when FMEA teams attempt to model probabilities as ranks or scores. Similar difficulties arise with scoring the cost of a loss, i.e., the severity component of RPN. Again there is the question of why, if you know the cost of a failure (in dollars, lives lost, or patients not cured) would you convert a valid measurement into a subjective score (granting, for sake of argument, that risk-neutrality is justified)? Again the answer is to enter that score into the RPN calculation.

Still more problematic is the detectability value used in RPNs. In a non-trivial system or process, detectability and probability are not independent variables. And there is vagueness around the meaning of detectability. Is it the means by which you know the failure mode has happened, after the fact? Or is there an indication that the failure is about to happen, such that something can be observed thereby preventing the failure? If the former, detection is irrelevant to risk of failure, if the latter the detection should be operationalized in the model of the system. That is, if a monitor (e.g, brake fluid level check) is in a system, the monitor is a component with its own failure modes and exposure times, which impact its probability of failure. This is how aviation risk analysis models such things. But not the Project Management Institute

A simple summary of the problems with scoring, ranking and RPN is that adding ambiguity to a calculation necessarily reduces precision.

I’ve identified several major differences between the approach to FMEAs used in aviation and those who claim they’re behaving like aerospace. They are not. Aviation risk analysis has reduced risk by a factor of roughly a thousand, based on fatal accident rates since aviation risk methods were developed. I don’t think the PMI can sees similar results from its adherents.

A partial summary of failure modes of common FMEA processes includes the following, based on the above:

- Equating FMEA with risk assessment
- Confusing FMEA with Hazard Analysis
- Viewing the FMEA as a Quality (QC) function
- Insufficient rigor in establishing probability and severity values
- Unwarranted (and implicit) assumption of risk-neutrality
- Unsound quantification of risk (RPN)
- Confusion about the role of detection

The corrective action for most of these should be obvious, including operationalizing a system’s detection methods, using numeric (non-ordinal) probability and cost values (even if estimated) instead of masking ignorance and uncertainty with ranking and scoring, and steering clear of Risk Priority Numbers and the Project Management Institute.

### Daniel Kahneman’s Bias Bias

Posted by Bill Storage in Probability and Risk on September 5, 2019

*(3rd post on rational behavior of people too hastily judged irrational. See first and second.)*

Daniel Kahneman has made great efforts to move psychology in the direction of science, particularly with his pleas for attention to replicability after research fraud around the *priming effect* came to light. Yet in *Thinking Fast And Slow *Kahneman still seems to draw some broad conclusions from a thin mantle of evidentiary icing upon a thick core of pre-formed theory. He concludes that people are bad intuitive Bayesians through flawed methodology and hypotheticals that set things up so that his psychology experiment subjects can’t win. Like many in the field of behavioral economics, he’s inclined to find bias and irrational behavior in situations better explained by the the subjects’ simply lacking complete information.

Like Richard Thaler and Dan Ariely, Kahneman sees bias as something deeply ingrained and hard-coded, programming that cannot be unlearned. He associates most innate bias with what he calls System 1, our intuitive, fast thinking selves. When called on to judge probability,” Kahneman says, “people actually judge something else and believe they have judged probability.” He agrees with Thaler, who finds “our ability to de-bias people is quite limited.”

But who is the “we” (“our” in that quote), and how is that “they” (Thaler, Ariely and Kahneman) are sufficiently unbiased to make this judgment? Are those born without the bias gene somehow drawn to the field of psychology; or through shear will can a few souls break free? If behavioral economists somehow clawed their way out of the pit of bias, can they not throw down a rope for the rest of us?

Take Kahneman’s example of the theater tickets. He compares two situations:

A. A woman has bought two $80 tickets to the theater. When she arrives at the theater, she opens her wallet and discovers that the tickets are missing. $80 tickets are still available at the box office. Will she buy two more tickets to see the play?

B. A woman goes to the theater, intending to buy two tickets that cost $80 each. She arrives at the theater, opens her wallet, and discovers to her dismay that the $160 with which she was going to make the purchase is missing. $80 tickets are still available at the box office. She has a credit card. Will she buy the tickets and just charge them?

Kahnemen says that the sunk-cost fallacy, a mental-accounting fallacy, and the *framing effect* account for the fact that many people view these two situations differently. Cases A and B are functionally equivalent, Kahneman says.

Really? Finding that $160 is missing from a wallet would cause most people to say, “darn, where did I misplace that money?”. Surely, no pickpocket removed the cash and stealthily returned the wallet to her purse. So the cash is unarguably a sunk cost in case A, but reasonable doubt exists in case B. She probably left the cash at home. As with philosophy, many problems in psychology boil down to semantics. And like the trolley problem variants, the artificiality of the problem statement is a key factor in the perceived irrationality of subjects’ responses.

By *framing effect*, Kahneman means that people’s choices are influenced by whether two options are presented with positive or negative connotations. Why is this bias? The subject has assumed that some level of information is embedded in the framer’s problem statement. If the psychologist judges that the subject has given this information too much weight, we might consider demystifying the framing effect by rebranding it the gullibility effect. But at that point it makes sense to question whether framing, in a broader sense, is at work in the thought problems. In presenting such problems and hypothetical situations to subjects, the framers imply a degree of credibility that is then used against those subjects by judging them irrational for accepting the conditions stipulated in the problem statement.

Bayesian philosophy is based on the idea of using a specific rule set for updating a “prior” (meaning *prior belief* – the degree of credence assigned to a claim or proposition) on the basis of new evidence. A Bayesian would interpret the framing effect, and related biases Kahneman calls anchoring and priming, as either a logic error in processing the new evidence or as a judgment error in the formation of an initial prior. The latter – how we establish initial priors – is probably the most enduring criticism of Bayesian reasoning. More on that issue later, but a Bayesian would say that Kayneman’s subjects need training in the use of uninformative priors and initial priors. Humans are shown to be very trainable in this matter, against the behavioral economists’ conclusion that we are hopelessly bound to innate bias.

One example Kahneman uses to show the framing effect presents different anchors to two separate test groups:

Group 1: Is the height of the tallest redwood more or less than 1200 feet? What is your best guess for the height of the tallest redwood?

Group 2: Is the height of the tallest redwood more or less than 120 feet? What is your best guess for the height of the tallest redwood?

Group 1’s average estimate was 844 feet, Group 2 gave 282 feet. The difference between the two anchors is 1080 feet. (1200 – 120). The difference in estimates by the two groups was 562 feet. Kahneman defines anchoring index as the ratio of the difference between mean estimates and difference in anchors. He uses this anchoring index to measure the robustness of the effect. He rules out the possibility that anchors are taken by subjects to be informative, saying that obviously random anchors *can be* just as effective, citing a 50% anchoring index when German judges rolled loaded dice (allowing only values of 3 or 9 to come up) before sentencing a shoplifter (hypothetical, of course). Kahneman reports that judges rolling a 3 gave 5-month sentences while those rolling a 9 assigned the shoplifter an 8-month sentence (index = 50%).

But the actual study (Englich, et. al.) cited by Kahneman has some curious aspects, besides the fact that it was very hypothetical. The judges found the fictional case briefs to be realistic, but they were not judging from the bench. They were working a thought problem. Englich’s Study 3 (the one Kahneman cites) shows the standard deviation in sentences was relatively large compared to the difference between sentences assigned by the two groups. More curious is a comparison of Englich’s Study 2 and the Study 3 Kahneman describes in *Fast and Slow. *Study 2 did not involve throwing dice to create an anchor. Its participants were only told that the prosecutor was demanding either a 3 or 9 month sentence, those terms not having originated in any judicial expertise. In Study 3, the difference between mean sentences from judges who received the two anchors was only two months (anchoring index = 33%).

Studies 2 and 3 therefore showed a 51% higher anchoring index for an explicitly random (clearly known to be random by participants) anchor than for an anchor understood by participants to be minimally informative. This suggests either that subjects regard pure chance as being more useful than potentially relevant information, or that something is wrong with the experiment, or that something is wrong with Kahnemnan’s inferences from evidence. I’ll suggest that the last two are at work, and that Kahneman fails to see that he is preferentially selecting confirming evidence over disconfirming evidence because he assumed his model of innate human bias was true before he examined the evidence. That implies a much older, more basic fallacy might be at work: *begging the question*, where an argument’s premise assumes the truth of the conclusion.

That fallacy is not an innate bias, however. It’s a rhetorical sin that goes way back. It is eminently curable. Aristotle wrote of it often and committed it slightly less often. The sciences quickly began to learn the antidote – sometimes called the scientific method – during the Enlightenment. Well, some quicker than others.

### A Bayesian folly of J Richard Gott

Posted by Bill Storage in Philosophy of Science, Probability and Risk on July 30, 2019

Don’t get me wrong. J Richard Gott is one of the coolest people alive. Gott does astrophysics at Princeton and makes a good argument that time travel is indeed possible via cosmic strings. He’s likely way smarter than I, and he’s from down home. But I find big holes in his Copernicus Method, for which he first achieved fame.

Gott conceived his Copernuicus Method for estimating the lifetime of any phenomenon when he visited the Berlin wall in 1969. Wondering how long it would stand, Gott figured that, assuming there was nothing special about his visit, a best guess was that he happened upon the wall 50% of the way through its lifetime. Gott saw this as an application of the Copernican principle: nothing is special about our particular place (or time) in the universe. As Gott saw it, the wall would likely come down eight years later (1977), since it had been standing for eight years in 1969. That’s not exactly how Gott did the math, but it’s the gist of it.

I have my doubts about the Copernican principle – in applications from cosmology to social theory – but that’s not my beef with Gott’s judgment of the wall. Had Gott thrown a blindfolded dart at a world map to select his travel destination I’d buy it. But anyone who woke up at the Berlin Wall in 1969 did not arrive there by a random process. The wall was certainly in the top 1000 interesting spots on earth in 1969. Chance alone didn’t lead him there. The wall was still news. Gott should have concluded that he saw the wall near in the first half of its life, not at its midpoint.

Finding yourself at the grand opening of Brooklyn pizza shop, it’s downright cruel to predict that it will last one more day. That’s a misapplication of the Copernican principle, unless you ended up there by rolling dice to pick the time you’d parachute in from the space station. More likely you saw Vini’s post on Facebook last night.

Gott’s calculation boils down to Bayes Theorem applied to a power-law distribution with an uninformative prior expectation. I.e., you have zero relevant knowledge. But from a Bayesian perspective, few situations warrant an uninformative prior. Surely he knew something of the wall and its peer group. Walls erected by totalitarian world powers tend to endure (Great Wall of China, Hadrian’s Wall, the Aurelian Wall), but mean wall age isn’t the key piece of information. The *distribution* of wall ages is. And though I don’t think he stated it explicitly, Gott clearly judged wall longevity to be scale-invariant. So the math is good, provided he had no knowledge of this particular wall in Berlin.

But he did. He knew its provenance; it was Soviet. Believing the wall would last eight more years was the same as believing the Soviet Union would last eight more years. So without any prior expectation about the Soviet Union, Gott should have judged the wall would come down when the USSR came down. Running that question through the Copernican Method would have yielded the wall falling in the year 2016, not 1977 (i.e., 1969 + 47, the age of the USSR in 1969). But unless Gott was less informed than most, his prior expectation about the Soviet Union wasn’t uninformative either. The regime showed no signs of weakening in 1969 and no one, including George Kennan, Richard Pipes, and Gorbachev’s pals, saw it coming. Given the power-law distribution, some time *well* after 2016 would have been a proper Bayesian credence.

With any prior knowledge at all, the Copernican principle does not apply. Gott’s prediction was off by only 14 years. He got lucky.

### Representative Omar’s arithmetic

Posted by Bill Storage in Probability and Risk on July 28, 2019

Women can’t do math. Hypatia of Alexandria and Émilie du Châtelet notwithstanding, this was asserted for thousands of years by men who controlled access to education. With men in charge it was a self-fulfilling prophecy. Women now represent the majority of college students and about 40% of math degrees. That’s progress.

Last week Marcio Rubio caught hell for taking Ilhan Omar’s statement about double standards and unfair terrorism risk assessment out of context. The quoted fragment was: “I would say our country should be more fearful of white men across our country because they are actually causing most of the deaths within this country…”

Most news coverage of the Rubio story (e.g. Vox) note that Omar did not mean that everyone should be afraid of white men as a group, but that, e.g., “violence by right-wing extremists, who are overwhelmingly white and male, really is a bigger problem in the United States today than jihadism.”

Let’s look at the numbers. Wikipedia, following the curious date-range choice of the US GAO, notes: “of the 85 violent extremist incidents that resulted in death since September 12, 2001, far-right politics violent extremist groups were responsible for 62 (73 percent) while radical Islamist violent extremists were responsible for 23 (27 percent).” Note that those are incident counts, not death counts. The fatality counts were 106 (47%) for white extremists and 119 (53%) for jihadists. Counting fatalities instead of incidents reverses the sense of the numbers.

Pushing the *terminus post quem* back one day adds the 2,977 9-11 fatalities to the category of deaths from jihadists. That makes 3% of fatalities from right wing extremists and 97% from radical Islamist extremists. Pushing the start date further back to 1/1/1990, again using Wikipedia numbers, would include the Oklahoma City bombing (white extremists, 168 dead), nine deaths from jihadists, and 14 other deaths from white wackos, including two radical Christian antisemites and professor Ted Kaczynski. So the numbers since 1990 show 92% of US terrorism deaths from jihadists and 8% from white extremists.

Barring any ridiculous adverse selection of date range (in the 3rd week of April, 1995, 100% of US terrorism deaths involved white extremists), Omar is very, very wrong in her data. The jihadist death toll dwarfs that from white extremists.

But that’s not the most egregious error in her logic – and that of most politicians armed with numbers and a cause. The flagrant abuse of data is what Kahneman and Tversky termed *base-rate neglect*. Omar, in discussing profiling (sampling a population subset) is arguing about *frequencies* while citing raw *incident counts*. The base rate (an *informative prior*, to Bayesians) is crucial. Even if white extremists caused most – as she claimed – terrorism deaths, there would have to be about one hundred times more deaths from white men (terrorists of all flavors are overwhelmingly male) than from Muslims for her profiling argument to hold. That is, the base rate of being Muslim in the US is about one percent.

The press overwhelmingly worked Rubio over for his vicious smear. 38 of the first 40 Google search results on “Omar Rubio” favored Omar. One favored Rubio and one was an IMDb link to an actor named *Omar Rubio*. None of the news pieces, including the one friendly to Rubio, mentioned Omar’s bad facts (bad data) or her bad analysis thereof (bad math). Even if she were right about the data – and she is terribly wrong – she’d still be wrong about the statistics.

I disagree with Trump about Omar. She should not go back to Somalia. She should go back to school.

### My Trouble with Bayes

Posted by Bill Storage in Philosophy of Science, Probability and Risk, Uncategorized on January 21, 2016

In past consulting work I’ve wrestled with subjective probability values derived from expert opinion. Subjective probability is an interpretation of probability based on a degree of belief (i.e., hypothetical willingness to bet on a position) as opposed a value derived from measured frequencies of occurrences (related posts: Belief in Probability, More Philosophy for Engineers). Subjective probability is of interest when failure data is sparse or nonexistent, as was the data on catastrophic loss of a space shuttle due to seal failure. Bayesianism is one form of inductive logic aimed at refining subjective beliefs based on Bayes Theorem and the idea of rational coherence of beliefs. A NASA handbook explains Bayesian inference as the process of obtaining a conclusion based on evidence, “Information about a hypothesis beyond the observable empirical data about that hypothesis is included in the inference.” Easier said than done, for reasons listed below.

Bayes Theorem itself is uncontroversial. It is a mathematical expression relating the probability of A given that B is true to the probability of B given that A is true and the individual probabilities of A and B:

P(A|B) = P(B|A) x P(A) / P(B)

If we’re trying to confirm a hypothesis (H) based on evidence (E), we can substitute H and E for A and B:

P(H|E) = P(E|H) x P(H) / P(E)

To be rationally coherent, you’re not allowed to believe the probability of heads to be .6 while believing the probability of tails to be .5; the sum of chances of all possible outcomes must sum to exactly one. Further, for Bayesians, the logical coherence just mentioned (i.e., avoidance of *Dutch book* arguments) must hold across time (synchronic coherence) such that once new evidence E on a hypothesis H is found, your believed probability for H given E should equal your prior conditional probability for H given E.

Plenty of good sources explain Bayesian epistemology and practice far better than I could do here. Bayesianism is controversial in science and engineering circles, for some good reasons. Bayesianism’s critics refer to it as a religion. This is unfair. Bayesianism is, however, like most religions, a belief system. My concern for this post is the problems with Bayesianism that I personally encounter in risk analyses. Adherents might rightly claim that problems I encounter with Bayes stem from poor implementation rather than from flaws in the underlying program. Good horse, bad jockey? Perhaps.

Problem 1. Subjectively objective

Bayesianism is an interesting mix of subjectivity and objectivity. It imposes no constraints on the subject of belief and very few constraints on the *prior* probability values. Hypothesis confirmation, for a Bayesian, is inherently quantitative, but initial hypotheses probabilities and the evaluation of evidence is purely subjective. For Bayesians, evidence E confirms or disconfirms hypothesis H only after we establish how probable H was in the first place. That is, we start with a *prior *probability for H. After the evidence, confirmation has occurred if the probability of H given E is higher than the prior probability of H, i.e., P(H|E) > P(H). Conversely, E disconfirms H when P(H|E) < P(H). These equations and their math leave business executives impressed with the rigor of objective calculation while directing their attention away from the subjectivity of both the hypothesis and its initial prior.

2. Rational formulation of the prior

Problem 2 follows from the above. Paranoid, crackpot hypotheses can still maintain perfect probabilistic coherence. Excluding crackpots, rational thinkers – more accurately, those with whom we agree – still may have an extremely difficult time distilling their beliefs, observations and observed facts of the world into a prior.

3. Conditionalization and old evidence

This is on everyone’s short list of problems with Bayes. In the simplest interpretation of Bayes, old evidence has zero confirming power. If evidence E was on the books long ago and it suddenly comes to light that H entails E, no change in the value of H follows. This seems odd – to most outsiders anyway. This problem gives rise to the game where we are expected to pretend we never knew about E and then judge how surprising (confirming) E would have been to H had we not know about it. As with the general matter of maintaining logical coherence required for the Bayesian program, it is extremely difficult to detach your knowledge of E from the rest of your knowing about the world. In engineering problem solving, discovering that H implies E is very common.

4. Equating increased probability with hypothesis confirmation.

My having once met Hillary Clinton arguably increases the probability that I may someday be her running mate; but few would agree that it is confirming evidence that I will do so. See Hempel’s raven paradox.

5. Stubborn stains in the priors

Bayesians, often citing success in the business of establishing and adjusting insurance premiums, report that the initial subjectivity (discussed in 1, above) fades away as evidence accumulates. They call this *washing-out* of priors. The frequentist might respond that with sufficient evidence your belief becomes irrelevant. With historical data (i.e., abundant evidence) they can calculate P of an unwanted event in a frequentist way: P = 1-e to the power -RT, roughly, P=RT for small products of exposure time T and failure rate R (exponential distribution). When our ability to find new evidence is limited, i.e., for modeling unprecedented failures, the prior does not get washed out.

6. The catch-all hypothesis

The denominator of Bayes Theorem, P(E), in practice, must be calculated as the sum of the probability of the evidence given the hypothesis plus the probability of the evidence given *not* the hypothesis:

P(E) = [P(E|H) x p(H)] + [P(E|~H) x P(~H)]

But ~H (“not H”) is not itself a valid hypothesis. It is a family of hypotheses likely containing what Donald Rumsfeld famously called unknown unknowns. Thus calculating the denominator P(E) forces you to pretend you’ve considered all contributors to ~H. So Bayesians can be lured into a state of false choice. The famous example of such a false choice in the history of science is Newton’s particle theory of light vs. Huygens’ wave theory of light. Hint: they are both wrong.

7. Deference to the loudmouth

This problem is related to no. 1 above, but has a much more corporate, organizational component. It can’t be blamed on Bayesianism but nevertheless plagues Bayesian implementations within teams. In the group formulation of any subjective probability, normal corporate dynamics govern the outcome. The most senior or deepest-voiced actor in the room drives all assignments of subjective probability. Social influence rules and the wisdom of the crowd succumbs to a consensus building exercise, precisely where consensus is unwanted. Seidenfeld, Kadane and Schervish begin “*On the Shared Preferences of Two Bayesian Decision Makers*” with the scholarly observation that an outstanding challenge for Bayesian decision theory is to extend its norms of rationality from individuals to groups. Their paper might have been illustrated with the famous photo of the exploding Challenger space shuttle. Bayesianism’s tolerance of subjective probabilities combined with organizational dynamics and the shyness of engineers can be a recipe for disaster of the Challenger sort.

All opinions welcome.

### More Philosophy for Engineers

Posted by Bill Storage in Engineering, Philosophy of Science, Probability and Risk on January 9, 2015

In a post on Richard Feynman and philosophy of science, I suggested that engineers would benefit from a class in philosophy of science. A student recently asked if I meant to say that a course in philosophy would make engineers better at engineering – or better philosophers. Better engineers, I said.

Here’s an example from my recent work as an engineer that drives the point home.

I was reviewing an FMEA (Failure Mode Effects Analysis) prepared by a high-priced consultancy and encountered many cases where a critical failure mode had been deemed highly improbable on the basis that the FMEA was for a mature system with no known failures.

How many hours of operation has this system actually seen, I asked. The response indicated about 10 thousand hours total.

I said on that basis we could assume a failure rate of about one per 10,001 hours. The direct cost of the failure was about $1.5 million. Thus the “expected value” (or “mathematical expectation” – the probabilistic cost of the loss) of this failure mode in a 160 hour mission is $24,000 or about $300,000 per year (excluding any secondary effects such as damaged reputation). With that number in mind, I asked the client if they wanted to consider further mitigation by adding monitoring circuitry.

I was challenged on the failure rate I used. It was, after all, a mature, ten year old system with no recorded failures of this type.

Here’s where the analytic philosophy course those consultants never took would have been useful.

You simply cannot justify calling a failure mode *extremely rare* based on evidence that it is at least *somewhat rare*. All unique events – like the massive rotor failure that took out all three hydraulic systems of a DC-10 in Sioux City – were very rare before they happened.

The authors of the FMEA I was reviewing were using unjustifiable inductive reasoning. Philosopher David Hume debugged this thoroughly in his 1738 *A Treatise of Human Nature*.

Hume concluded that there simply is no rational or deductive basis for induction, the belief that the future will be like the past.

Hume understood that, despite the lack of justification for induction, betting against the sun rising tomorrow was not a good strategy either. But this is a matter of pragmatism, not of rationality. A bet against the sunrise would mean getting behind counter-induction; and there’s no rational justification for that either.

In the case of the failure mode not yet observed, however, there is ample justification for counter-induction. All mechanical parts and all human operations necessarily have nonzero failure or error rates. In the world of failure modeling, the knowledge “known pretty good” does not support the proposition “probably extremely good”, no matter how natural the step between them feels.

Hume’s problem of induction, despite the efforts of Immanuel Kant and the McKinsey consulting firm, has not been solved.

A fabulously entertaining – in my view – expression of the problem of induction was given by philosopher Carl Hempel in 1965.

Hempel observed that we tend to take each new observation of a black crow as incrementally supporting the inductive conclusion that all crows are black. Deductive logic tells us that if a conditional statement is true, its contrapositive is also true, since the statement and its contrapositive are logically equivalent. Thus if all crows are black then all non-black things are non-crow.

It then follows that if each observation of black crows is evidence that all crows are black (compare: each observation of no failure is evidence that no failure will occur), then each observation of a non-black non-crow is also evidence that all crows are black.

Following this line, my red shirt is confirming evidence for the proposition that all crows are black. It’s a hard argument to oppose, but it simply does not “feel” right to most people.

Many try to salvage the situation by suggesting that observing that my shirt is red is in fact evidence that all crows are black, but provides only unimaginably small support to that proposition.

But pushing the thing just a bit further destroys even this attempt at rescuing induction from the clutches of analysis.

If my red shirt gives a tiny bit of evidence that all crows are black, it then also gives equal support to the proposition that all crows are white. After all, my red shirt is a non-white non-crow.

## Recent Comments