Source
How to Lie with Statistics
Two-part review from book_cube with a walkthrough of 10 chapters and a practical fact-check checklist.
Source
How to Lie with Statistics
Two-part review from book_cube with a walkthrough of 10 chapters and a practical fact-check checklist.
How to Lie with Statistics
Authors: Darrell Huff
Publisher: W. W. Norton & Company
Length: 1954
A compact classic on how statistical framing can be used to persuade. It is short, but highly practical for reviewing dashboards, reports, and argument quality.
Original
TranslationMath / Statistics / Management
How to Lie with Statistics was published more than 70 years ago, but its core warning still holds: numbers do not speak for themselves. They depend on sampling, calculation method, visualization choices, and interpretation.
1. Book map: 9 manipulation patterns
Chapter 1
Biased sample
If the sample is not representative, the output describes a convenient subgroup, not the real population.
Chapter 2
Convenient average
Mean, median, and mode can diverge heavily. Manipulation starts when only the favorable one is shown.
Chapter 3
Hidden methodological caveats
Without sample size, failed runs, and formula transparency, a metric turns into a narrative device.
Chapter 4
Noise sold as signal
Point values without confidence bounds or significance framing create false precision.
Chapter 5
Chart manipulation
Axis truncation, scale switching, and selective time windows can exaggerate tiny differences.
Chapter 6
Misleading pictorial charts
A 2x numeric change drawn as a 3D object can look like an 8x jump to the eye.
Chapter 7
Pseudo-grounded number
A real number from an authoritative source is paired with a biased interpretation.
Chapter 8
After does not mean because of
Correlation does not prove causation: hidden variables, feedback loops, and coincidence still apply.
Chapter 9
Stati-culations
Percent tricks, double counting, and mixing incomparable objects into one average produce fake insight.
2. Chapter 10: putting the statistician in place
Who says this?
Check incentives and conflicts of interest: who funded the work and who benefits from the claim.
How do they know?
Demand transparency in data collection, experiment design, and metric computation.
Was the object of measurement swapped?
A proxy can silently replace the real goal. Validate the Goal -> Signal -> Metric chain.
Does this make sense?
Numbers should explain reality, not distract from it. Semantics first, statistics second.
For engineering leadership, the proxy-swap question is usually the most valuable. A practical anchor is the Goal -> Signal -> Metric chain discussed inEngineering Productivity Measurement.
3. How to apply this as a tech lead
Product hypothesis reviews
Do not discuss conversion in isolation. Inspect sample composition, experiment duration, and effect stability beyond one reporting window.
Engineering metrics
For DORA/SPACE/DevEx, make definitions explicit. Otherwise cross-team comparisons are methodologically weak.
Stakeholder communication
Present target values together with confidence bounds, bias risks, and alternative explanations.
Anti-manipulation culture
Normalize hard questions in metric reviews. It reduces KPI theater and improves decision quality.
4. Common anti-patterns
Presenting one metric without baseline, sample size, and uncertainty range.
Making causal claims from correlation without checking alternatives.
Changing metric definitions mid-discussion and calling it trend evidence.
Using a polished chart as a replacement for methodological scrutiny.
5. Recommendations
For every report, document sample size, inclusion criteria, and observation window.
Add a mandatory section in metric reviews: limitations, interpretation risks, and unknowns.
Separate descriptive, diagnostic, and causal claims instead of merging them into one story.
Use the 4 questions from Chapter 10 before committing to a high-impact decision.