Numbers only answer the questions they were built to answer

1

Ask the problem first

The PPDAC cycle—Problem, Plan, Data, Analysis, Conclusion—beats any technique; analysis without a real problem answers nobody's question.

Spiegelhalter throws out the image of statistics as calculations you run once data shows up. He offers the PPDAC cycle instead. Order is the point. Jump straight to analysis and you get tidy numbers for a question no one posed.

Harold Shipman, the English GP later convicted of murdering patients, makes the stakes plain. Reviewers of mortality data saw deaths clustered at odd times of day and rates far above comparable doctors. No exotic test found him. The work started by asking whether the pattern fit natural causes.

That habit—what problem am I actually solving—runs through every later topic. Correlation, significance, and algorithms all circle back to whether the question was clean before anyone trusted the output. Literacy here is skepticism about origin, purpose, and what would count against the claim.

Polished analysis aimed at the wrong problem does more harm than leaving the data alone.

In practice: Shipman's excess deaths at odd hours surfaced because statisticians framed natural-causes-versus-not before touching the file.

2

What one number hides

The same mean can come from a bell curve or a long tail; aggregation can reverse every subgroup pattern.

The next trap is the summary itself. A mean or median can spring from wildly different shapes. A symmetric mound and a skewed pile with a long tail can share one average and tell opposite stories.

Income data shows how. Countrywide average income can climb while most people earn less, if a few very high earners drag the mean up. Reports of the average with no spread mislead even when every digit is correct. Histograms and dot plots beat headline single figures.

Simpson's paradox drives the point home: a trend in combined data flips when you split into groups. University admissions once seemed to show women admitted at lower rates overall, yet equal or higher rates inside nearly every department, because women applied more to the tougher ones.

Aggregation is a choice. The level you pick can invent a conclusion the fine grain never held.

In practice: Women looked less likely to win admission in the aggregate while matching or beating men inside nearly every department, because they applied harder to competitive ones.

3

Causation needs a case

A correlation becomes causal evidence only after independent checks—not through statistical strength alone.

Public life abuses the leap from association to cause. Spiegelhalter revives Austin Bradford Hill's checklist, built to weigh whether smoking caused lung cancer: strength of association, consistency across studies, specificity, temporal sequence, a plausible mechanism, and dose-response.

No single item proves cause. Together they can justify action—as they did for smoking long before a definitive randomized trial was possible or ethical. Trials remain the gold standard because random assignment breaks ties to confounders. Most questions in medicine, economics, and policy still cannot wait on trials alone.

Spurious pairings make the opposite point. National chocolate consumption lines up with Nobel prizes per capita; wealth almost certainly sits underneath both. A coefficient is a start for investigation, not a verdict.

Later chapters lean on this stance: a strong statistical link invites work, it does not close the case.

In practice: Chocolate-eating nations racking up Nobels looks like a link until wealth is seen as the common driver of both.

4

Probability's split meanings

Confusion about risk often comes from mixing long-run frequency with degree of belief about one unrepeatable case.

Frequentist probability tracks long-run frequency across repeated trials—how often a fair coin lands heads over many flips. Subjective or Bayesian probability tracks degree of belief about a single event: this defendant's guilt, this patient's response to treatment.

Weather forecasters, insurers, and gamblers live in the frequentist habit fairly well. Courts, doctors, and voters get pushed into the subjective sense, where no long run exists. One version is not crowned correct. Clarity about which is in play is the discipline public talk usually skips.

The 2016 Brexit referendum and US election forecasts looked 'wrong' to many when the less likely outcome won. A well-calibrated forecast that gave an outcome twenty or thirty percent is not falsified when that outcome occurs.

This split sets up courtroom Bayesian use later. Importing frequentist instincts into unique cases produces some of the costliest statistical errors institutions make.

In practice: A forecast giving an outcome a twenty or thirty percent chance is not broken simply because that outcome happens once.

5

The narrow p-value

A p-value scores how surprising data would be if there were no effect—not the probability the hypothesis is true.

Researchers lean on significance testing as a referee. A p-value answers one technical question: how odd is this data if nothing real is going on. It does not give the chance your hypothesis is correct. It does not give effect size or importance.

Treating a cross of the conventional 0.05 line as proof, and anything above as nothing, helped produce what much of science now calls a replication crisis. Many psychology studies failed when independently repeated. P-hacking—running analyses until one clears the line by chance, then reporting only that one—became a named habit.

The tool is not thrown out. It needs pairing with effect size, design quality, and prior plausibility. Confidence intervals and full spreads of possible effects beat a pass/fail stamp.

Even a study strong enough for a causal claim can still mislead if journals and journalists overread what a p-value allows them to say.

In practice: Psychology's wave of failed replications followed years of treating a 0.05 threshold as a truth switch.

6

Base rates in court

Juries convict on rare-event figures while skipping how rare the guilt hypothesis is too—the prosecutor's fallacy.

Criminal courts raise the stakes. Sally Clark, a British solicitor, was wrongly convicted in 1999 of murdering her two infant sons. An expert said the chance of two natural cot deaths in one family was about 1 in 73 million, from squaring the single-death probability.

That math was wrong on its own terms: it treated the deaths as independent when shared genes or environment could make a second cot death more likely. The deeper error was the prosecutor's fallacy. Even a correct tiny chance is not impossibility. The jury had to weigh that rarity against the also-rare alternative that a mother killed both children.

Clark's conviction was overturned in 2003 once the flawed statistic was exposed. Bayesian updating—revising a prior in light of evidence—asks which story, murder or natural death, fits the combined facts better, rather than how shocking the coincidence looks alone.

Literacy for judges, jurors, and experts is a safeguard against wrongful conviction, not a classroom ornament.

In practice: The 1-in-73-million cot-death figure helped convict Sally Clark until the fallacy and the independence error were exposed and the verdict quashed.

7

Accurate without understanding

Models trained only to predict can score well with no causal mechanism, then lock in biased history when used on people.

Machine learning now drives decisions in medicine, lending, and criminal justice. Prediction is not explanation. A model can forecast who defaults, reoffends, or responds to treatment from historical patterns without any piece standing for a cause.

Deploy that model on individuals and the cracks show. Biased training data gets reproduced and often amplified. A black-box score may flag someone as high risk with no usable account of why, so the person cannot contest the reasoning. Race-correlated variables can seep in even when race itself is barred.

Algorithmic prediction is not scrapped. Real gains show up in places like medical imaging. Predictive power still needs scrutiny of training data, ongoing error audits, and a clear line between a tool that aids a clinician and one that replaces judgment.

An algorithm's link to an outcome is not evidence it understands that outcome. Institutions that blur the line hand decisions to a process no one can interrogate.

In practice: Risk tools in UK and US justice can absorb race-correlated signals indirectly and leave defendants no way to challenge the model's path.

8

Absolute risk or drama

Relative risk can double a tiny baseline and sound like a crisis; absolute figures and natural frequencies keep the scale honest.

Risk talk is where statistics meets the public hardest. In 1995 UK regulators said a newer contraceptive pill doubled blood-clot risk. True as a relative claim. Absolute risk moved from roughly 1 in 7,000 to 2 in 7,000 women per year.

Reported without the baseline, the doubling set off panic. Women quit the pill in waves. Unwanted pregnancies and abortions rose afterward. Relative figures get chosen, often without bad intent, because they sound larger. Absolute risks belong beside every relative comparison.

Natural frequencies such as '2 out of 7,000' beat percentages for intuitive grasp. The same demand reaches breast cancer screening: benefits in deaths prevented get overstated when overdiagnosis and unneeded treatment stay offstage.

Good risk communication is honesty about uncertainty—what is known, what is estimated, and how firm the estimate is. A clean question and careful analysis still fail if the closing translation misleads the people it was for.

In practice: The pill scare's relative 'double' hid a shift from about 1 in 7,000 to 2 in 7,000 and helped drive a rise in unwanted pregnancies.