Naked Statistics
Summary
Wheelan explains statistical concepts without mathematics, aimed at readers who need to interpret statistical claims rather than produce them. The coverage runs from descriptive statistics and the problems with averages, through correlation, probability, the central limit theorem, inference and hypothesis testing, polling, regression analysis and program evaluation.
The practical emphasis is on how statistics mislead. Wheelan is specific about selection bias, survivorship bias, publication bias, the difference between correlation and causation, the misuse of averages when a distribution is skewed, the base rate problem in medical and criminal testing, and the ways a regression can produce a confident wrong answer through omitted variables or through fitting a model to data that were used to generate the hypothesis. His chapter on the perils of regression analysis is the most useful, since regression is the technique most commonly cited in support of business and policy claims and most commonly misapplied.
He uses accessible examples throughout, including the Atlanta test score scandal, medical screening statistics, and the mathematics of the Schlitz beer taste test advertisement.
The book teaches interpretation rather than method, and it is deliberately shallow on the technical side. Anyone who needs to run an analysis will need a real text. Its value is in equipping a reader to ask the right questions when presented with a statistical claim, which is the more common requirement for investors and operators. The tone is deliberately light and some readers find the jokes wearing.
Key takeaways 7
- An average conveys almost nothing when the distribution is skewed, and the median plus the shape of the distribution is what a reader actually needs.
- Correlation establishes association and never establishes cause, and the standard error is treating a regression coefficient as a causal effect.
- Selection bias in how a sample was assembled invalidates any inference from it, and it is the most common and least detected flaw in statistical claims.
- A test with high accuracy still produces mostly false positives when the condition being tested for is rare, which is a property of base rates rather than of the test.
- Regression results change substantially when a relevant variable is omitted, and the analyst cannot know which variables were omitted from published work.
- Publication bias means the studies that reach print are systematically those that found an effect, which inflates the apparent evidence for any claim.
- Statistical significance says nothing about whether an effect is large enough to matter, and the two are routinely conflated.
Who should read it
Operators and investors who need to interrogate statistical claims in research, pitches and reporting without doing the analysis themselves.
Where it sits in the reference 3
Who recommended it 1
Similar books 6
Chosen by shared topics, the entries that hold them, and the people who recommended both.
