Skip to main content

Chapter 6: How to Lie With Statistics

Part Two: Why You Can See It and Others Can't ｜ 📖 Who's Collecting at the Door ｜ 📖 Full Edition

Chapter 6: How to Lie With Statistics

Prologue: A Fight Fought With Numbers​

In September 2026, the online creator Ceylon (錫蘭) and the astrologer Tang Chi-yang (唐綺陽) got into a fight over astrology. On one side were the birth charts of 81 celebrities with romantic scandals, compiled by Tang's show; on the other, Ceylon produced about 40,000 questionnaires on a livestream and asked how big the other side's sample was.

Public opinion quickly boiled it down to "40,000 versus 81," as if whoever had the bigger number won. But both sides' data had the same flaw: each was produced by its own camp for its own audience. Ceylon's questionnaires came mostly from his viewers; Tang's cases were chosen and interpreted by people who believe in astrology. The numbers were not in conversation; they were taking sides.

Watching this fight, I thought of a book I read more than ten years ago: Darrell Huff's How to Lie with Statistics.1 Published in 1954, it is a slim volume entirely about how statistics are used to mislead. By 2026, more than seventy years later, not one of its tricks has gone out of date. In the workplace I have met plenty of so-called senior executives who still don't understand the most basic statistics.

Chapter 5 was about how position decides what a person can see (see Chapter 5). This chapter is about another kind of not seeing: the numbers are right in front of you, plenty of people can read them, and yet no one checks. I'll first pick the five most common tricks in the book, each with a current example; then I'll take up a question the book doesn't address but which I think matters more: why, seventy years on, people still haven't learned.

1. The Biased Sample​

If the sample doesn't represent the population, no amount of precise calculation helps. This is the book's first chapter, and the most common pitfall.

Tang's 81 people were celebrities findable in the news. What was being studied was not only "who is prone to romantic trouble" but also "who is likely to make the news." Fame, newsworthiness and whether the paparazzi were watching all affected who got included.

Ceylon's 40,000 questionnaires look huge, but if the respondents were mainly his viewers, it is a self-selected sample. People willing to spend time on a "star signs and personality" survey are already different from the general public; people who watch his channel even more so in their attitude to astrology.

A bigger sample doesn't make bias go away; it only makes it look weightier. The most famous counterexample in the history of statistics is the Literary Digest poll of the 1936 US presidential election: it collected more than two million questionnaires and predicted a landslide for Landon, and Roosevelt was re-elected by a landslide instead. Its mailing list came mostly from telephone directories and car registrations, which in the Depression meant a relatively well-off group.2

The workplace is the same. The people filling in employee satisfaction surveys are mostly those who haven't left; the truly dissatisfied voted with their feet long ago.

2. The Well-Chosen Average​

"Average" can mean the mean, the median or the mode. Whoever reports the number picks the one that suits them best.

Suppose a company has ten people: the boss earns 10 million a year and nine employees earn 500,000 each. The mean is 1.45 million; the median is 500,000. A job ad saying "average salary over one million" doesn't tell a single lie.

Financial news uses the same trick every day. "Average wages hit a record high," yet most people's pay hasn't moved. Income distributions are skewed to the right; a few high earners can pull the mean up, and the median is closer to most people's situation.

When you see the word "average," ask two things first: which average? And what does the distribution look like?

3. The Undisclosed Data​

The most dangerous number is often the one you don't see: how big the sample is, what it is being compared with, what the baseline was.

"Among 81 celebrities with romantic trouble, 15% have a certain chart configuration" sounds striking. The missing half is: among people without such trouble, how many have that configuration? If 15% of ordinary people have it anyway, nothing has been found at all.

The book uses a toothpaste ad as its example. "Use our product and get 23% fewer cavities"; what it doesn't tell you is that the test group was a dozen or so people. With a sample that small, a manufacturer that runs the experiment over and over will eventually get a good-looking result, and then publish only that one.

The workplace version is even more common. A slide says "revenue up 30% this quarter" without saying that the same quarter last year happened to be a low point. "Conversion rate doubled," without saying it went from 0.5% to 1%. Every number is true; the missing sentence is the point.

Chapter 1 noted that US accounting standards require only the total of R&D spending to be disclosed, with no split between development and maintenance (see Chapter 1). Same trick: no number is false; what's missing is the breakdown.

4. The Gee-Whiz Graph​

You don't need to change a single number; just move the axis, and a flat line becomes a surge.

Start the vertical axis at 95 instead of 0, and a change from 96 to 98 fills the whole chart and looks like a doubling. Another trick is the "one-dimensional picture": when a quantity doubles, double both the height and the width of the icon. The area becomes four times larger, and the difference the eye sees far exceeds the number itself.

The bar charts in quarterly company presentations often don't start the vertical axis at zero. It isn't always deliberate deception; a lot of charting software rescales the axis automatically. But for the person looking at the chart, the effect is exactly the same.

Before you read a chart, look at where the axis starts.

5. Mistaking Correlation for Causation​

Two things changing together doesn't mean one causes the other. It may be coincidence, there may be a third factor behind both, or the causation may run in exactly the opposite direction.

The most interesting thing about the astrology fight is precisely the question of direction. By the late 1970s, psychological research had already found that people familiar with astrology describe their own personalities in ways that drift toward their sign's description, while people who know nothing about astrology show no such tendency.3 Even if a "link between star signs and personality" were measured, it might only be that people read the descriptions and gradually lived up to them. The correlation is real; the cause is not in the stars.

That is a problem for both sides. When believers fill in the questionnaire, a link may "show up"; when non-believers fill it in, it is natural that none does. Using self-report questionnaires to test astrology can't answer the question it wants to answer, whether the result comes out positive or negative.

Back to each side's audience: Ceylon's viewers show no link; Tang's viewers very likely would. Both data sets are real, both sides can say "the data supports me," and what both have actually measured is only their own audience.

The workplace version: "employees who attend training perform better." Maybe the training works; maybe only people who were already motivated sign up.

6. Why, Seventy Years On, We Still Haven't Learned​

This book has been reprinted for decades and translated into many languages, with every trick laid out. Yet in meeting rooms the charts still don't start at zero, and nobody asks which "average" a presentation means.

My view is that the problem is not that the knowledge isn't widespread enough, but that statistics inside organizations was never for inference in the first place. It is for endorsement.

Many decisions are made before the numbers appear. The numbers' job is to make the decision look grounded. Actually spelling out base rates, comparison groups and sample bias makes a report less "clean"; the conclusion has to shrink, and reporting upward gets harder.

So organizations reward the people who can produce numbers that support the conclusion, not the people who say "this data only supports us this far." The former are praised for execution; the latter are thought insufficiently positive and unable to see the big picture. It isn't that senior executives can't learn; they don't need to. Learning would get in the way.

Conclusion first, numbers second: that order is applied not only to numbers but to people too. Section 7 comes back to this.

The Veto You Hand Over Yourself​

The most dangerous form of numbers-as-endorsement is not in presentations but in engineering.

Before developing a product, manufacturers carry out a Failure Mode and Effects Analysis (FMEA): list every place something could go wrong, score each for severity, likelihood and difficulty of detection, and any score over the threshold requires a design change. The engineers do the scoring themselves. Lower likelihood by one notch and improve detectability by one, and the risk drops below the threshold: no change needed, no shipping delay.

FMEA is the veto from Chapter 1, written out as a table: its purpose is to let engineers say "this won't do" on solid grounds before the shipping date (see Chapter 1). Chapter 1 described the organization taking the veto away from outside; scoring an FMEA low is engineers handing it over themselves. No one needs to be laid off and no team needs to be replaced; as long as the cost of saying "no" is high enough, the people who would say it will lower the scores on their own.

Outsiders can't spot this kind of manipulation, because seeing it requires expertise in the same field. The people who can see it are fellow experts inside; but their own FMEAs are mostly scored the same way, so exposing others means exposing themselves. Bring in outside experts to review? There is confidentiality, there is cost, and there is the line "we are the experts." So everyone capable of checking the numbers has a reason not to.

The lies in Huff's book can be seen through by a reader asking five questions. This kind can't. It hides behind a professional wall, and the cost of verifying it is so high that only insiders can afford to pay, and insiders won't.

The Normalization of Deviance​

What does this add up to? The risk doesn't disappear because the score is low; it only moves off the paperwork, into the product, onto the users, into the future. Every time a low score passes without incident, the next low score feels more natural. Studying the Challenger space shuttle disaster, the sociologist Diane Vaughan called this process the "normalization of deviance": when departures from the standard repeatedly have no consequences, the departure becomes the standard.4 What new engineers learn is no longer how to analyze risk, but "this is how FMEAs are scored here."

Here, too, is the asymmetry of measurement from Chapter 1: when prevention succeeds there is no evidence, while every delay and every design change is counted precisely (see Chapter 1). The engineer who scores high leaves behind a delay; the engineer who scores low leaves behind a record of on-time shipping. At the year-end review, only the latter is visible.

Until one day something really does go wrong. Pull out the old FMEA, and there it is in black and white: "Analyzed; low risk." That document was never an analysis; it was an endorsement proving that everyone had done their duty.

This has the same structure as the astrology fight. Online, the audience is fans; inside a company, the audience is the boss and the board. In both places, data plays the same role: a badge of identity, not a tool of verification.

7. Numbers and Labels: A Combination Punch​

Looking back at the astrology fight, one thing is easy to miss: what the two sides were really fighting over was not numbers but personality. Astrology is a system for labeling people's personalities; the 40,000 questionnaires and the 81 birth charts were arguing over whether those labels are accurate. The numbers were only called in to testify for the labels.

This is not unique to astrology. In the workplace, numbers and personality labels often appear together, with a clear division of labor: the numbers supply authority, the labels assign blame.

  • A performance ranking puts you in the bottom 20%. The number itself gives no reason; then comes a line like "can't handle pressure" or "not positive enough," and now there is a reason, and it lies in you.
  • An employee satisfaction score drops, and the report reads it as "the new generation is fragile," not as pay, hours or management.
  • A department has high turnover, and the conclusion is "young people have no patience." As Section 1 noted, the people who stay to fill in surveys were never the ones who left; those who left voted with their feet, but their reasons were never counted.

Each has its weakness. Numbers can be checked: where the sample came from, what the baseline is, which average was used, all the questions from the first five sections can be asked. Personality labels have no authority: "you're just that kind of person" can be said by anyone, and disbelieved. Put together, each covers the other's weakness. Question the numbers, and you're told you're defensive and won't face facts; deny the label, and you're told the numbers speak for themselves. The numbers make the label look grounded; the label gives the numbers someone to hold responsible.

Labels Wrapped in Numbers​

Even the labels themselves often arrive wrapped in numbers. The personality tests popular in corporate training and online sort people into a few animals, a few numbers or four letters, with percentages and radar charts in the report, so that it looks like statistics. In the comment section of the podcast in Chapter 7, the "Type 8s," the "INFJs," the owls and the sharks that viewers pinned on one another are among the most common of these systems (see Chapter 7).

In 1949, the psychologist Bertram Forer ran an experiment. He gave his students a personality test and then handed each a "personalized" analysis, asking them to rate its accuracy from 0 to 5; the average was 4.26. In fact, everyone had received the same report, most of it copied from an astrology book.5 A description true of anyone, once placed in a report that seems to have been measured, is taken as insight into oneself.

Three Steps, and the Structure Disappears​

So the order of this combination punch is exactly the same as the endorsement described above: conclusion first, numbers second, label last. First decide who the problem is, then find a number that supports it, then use personality to nail the blame to that person.

No step involves a lie, and every step can be said aloud in a meeting room; yet after the three steps, the structure has vanished from the story. Why the ranking is designed that way, why the survey only asks the people who stayed, why the people who left left: none of these questions needs asking any more, because the answer is already in: it was that person.

Chapter 7 takes this blade apart from another direction: not through numbers, but through an answer given three times on one show and translated, each time, into personality.

8. The Man Who Wrote the Liar's Handbook​

The most ironic lesson comes from Huff himself.

Huff was not a statistician but a freelance journalist. In 1964, the US Surgeon General issued the report that formally linked smoking to lung cancer. That same year, the tobacco industry hired Huff to write a pamphlet questioning the report's evidence, for use by congressional committees and industry lawyers. In March 1965 he testified before the Senate Commerce Committee, citing "eight red flags" in the report.6

His weapon was exactly the trick his own book had taught: correlation is not causation. He compared the link between smoking and lung cancer to classic fallacies like "storks and babies." When a senator asked whether he truly believed the two were equally arbitrary, he answered that as far as he could see, they were the same.

After testifying, he received tobacco industry funding to expand the pamphlet into a book titled How to Lie with Smoking Statistics. The manuscript was nearly finished when, at the end of 1968, the publication plan was suddenly dropped. Only after the 1998 tobacco settlement were the related documents and manuscript made public, and between 2012 and 2014 statisticians dug them up again.6 7

The man who wrote about how statistics lie ended up using the same techniques to lie for tobacco. This doesn't rob the book of its value; every trick it teaches still holds. But it shows one thing: understanding statistics and using statistics honestly are two different things. Understanding more only makes you clearer about which trick to pick.

9. Five Questions, Then Read It Yourself​

In the book's last chapter, Huff leaves readers five questions. Faced with any number, ask first:

  1. Who says so? What reason do they have to want you to believe it?
  2. How do they know? Where did the sample come from?
  3. What's missing? The sample size, the baseline for comparison, the sentence left unsaid.
  4. Did somebody change the subject? The data is about A, but the conclusion jumps to B.
  5. Does it make sense?

Take these five questions back to "40,000 versus 81," and in under three minutes you'll see that neither side won.

The five tricks I picked are only half the book; the rest I leave you to read yourself. It is short; an afternoon will do. As of October 2026, the easiest edition to find in Taiwan is Lohas Culture's 2024 translation by Wen Yi-chun; CommonWealth Publishing's 2005 translation by Cheng Wei-hou is the same book.1

Only don't assume that once you've read it, you can't be fooled. The people fooling you have often read it more closely than you.

Summary​

The three chapters of Part Two answer the same question: why can you see what others can't? Chapter 4 was about position: standing inside a separate restaurant, you can't see the shared queue. Chapter 5 was about structure: the same person sees completely different human natures in companies with money to share and companies without. This chapter is about the cost of checking: the numbers are right there, and everyone who can read them has a reason not to look.

None of these three kinds of blindness comes from stupidity. But they leave a consequence: when a structure can't be seen, the costs it causes are charged to some individual. Part Three is about how that bill gets charged to you, and why you don't have to accept it. It begins with a show in which an answer was given three times and no one heard it.

← Chapter 5: You Didn't Meet a Villain, You Met a Structure With No Money to Share📖 Back to ContentsChapter 7: Psychological Manhunt: The Answer Was Given Three Times, and No One Heard It →

Footnotes​

  1. Darrell Huff, How to Lie with Statistics, W. W. Norton, 1954. Taiwanese translations: 《統計操控的真相與謊言》, trans. Wen Yi-chun (聞翊均), Lohas Culture (樂金文化), 2024; 《別讓統計數字騙了你》, trans. Cheng Wei-hou (鄭惟厚), CommonWealth Publishing (天下文化), 2005. Mainland Chinese translation: 《統計數字會撒謊》, trans. Liao Yinglin (廖穎林), China City Press, 2009. ↩ ↩2

  2. Peverill Squire, "Why the 1936 Literary Digest Poll Failed," Public Opinion Quarterly, 52(1), 1988, pp. 125–133. ↩

  3. J. Mayo, O. White & H. J. Eysenck, "An Empirical Study of the Relation between Astrological Factors and Personality," Journal of Social Psychology, 105(2), 1978, pp. 229–236; K. Pawlik & L. Buse, "Selbst-Attribuierung als differentiell-psychologische Moderatorvariable," Zeitschrift für Sozialpsychologie, 10, 1979, pp. 54–69. [To be checked] ↩

  4. Diane Vaughan, The Challenger Launch Decision: Risky Technology, Culture, and Deviance at NASA, University of Chicago Press, 1996. ↩

  5. Bertram R. Forer, "The Fallacy of Personal Validation: A Classroom Demonstration of Gullibility," Journal of Abnormal and Social Psychology, 44(1), 1949, pp. 118–123. ↩

  6. Alex Reinhart, "Huff and Puff," Significance, 11(4), 2014. ↩ ↩2

  7. Tim Harford, "Cigarettes, Damn Cigarettes and Statistics," April 2015. ↩