Statistical Thinking For Everyday Life

There's a reason statistics are taught as formulas rather than frameworks, and why the news gives you relative risk instead of absolute risk. Confusion is profitable. Fear is engaging. A 50% increase in risk sounds terrifying. A risk that goes from 1 in 1,000 to 1.5 in 1,000 sounds manageable. Same fact, wildly different emotional reactions. The media knows this. Pharmaceutical companies know this. Politicians know this.

Statistical thinking is the cognitive immune system you were never trained to develop. This chapter builds it from scratch.

A base rate is the background frequency of something in the relevant population. Before you can assess whether a treatment works, a technology is dangerous, or a trend is real, you need to know what the baseline looks like.

The classic example: you take a test for a rare disease that's 99% accurate. The test comes back positive. How worried should you be?

Most people say very worried. But here's the math. If the disease affects 1 in 10,000 people and you test 10,000 people, one person actually has it. The 99% accurate test will correctly identify that person. But it will also flag 1% of the 9,999 healthy people — about 100 false positives. So when you get a positive result, the odds are roughly 100 to 1 that you don't have the disease.

This is called the base rate fallacy — the failure to account for how rare something is when interpreting test results or probability claims. Doctors make this error. Jurors make this error. You make this error.

The fix is simple but uncomfortable: always ask what the baseline probability is before you assess any claim. What percentage of people in this situation normally experience this outcome? How common is the thing we're worried about in the first place?

Francis Galton discovered regression to the mean in the 1800s while studying the heights of children of tall parents. He found that very tall parents tended to have children who were tall but not as tall — and vice versa. The children "regressed" toward the average.

The underlying principle: extreme observations tend to be partly due to luck, and luck isn't permanent. When you measure extreme performance — good or bad — some portion of it is random fluctuation. Remove the fluctuation (which happens naturally over time), and performance moves toward the mean.

This has practical consequences everywhere:

- Management feedback loops: A manager praises an employee after great performance, then scolds them after poor performance. The praised performance tends to regress downward next period; the scolded performance tends to improve. The manager concludes that praise doesn't work and scolding does. The causal story is wrong — regression to the mean is doing the work.

- Medical interventions: People seek treatment when symptoms are at their worst. Symptoms tend to improve after the peak regardless of treatment. This is why anecdotal evidence for everything from homeopathy to prayer to fad diets looks so compelling — people try things at their worst and improve, and attribute the improvement to the intervention.

- Sports and business: The "sophomore slump" is real. So is the CEO who posts a record quarter then underperforms. Extreme performance tends not to persist at its extreme level.

The diagnostic question: "Is this extreme result likely to be at least partially luck?" If yes, expect regression. Don't build strategy around the outlier.

"Correlation doesn't imply causation" is a cliché. Here's the deeper structure:

There are four possible explanations when two things correlate:

1. A causes B 2. B causes A 3. A third variable C causes both 4. Pure coincidence (especially in small samples)

Most people consider option 1 and stop. Good thinkers ask about all four. The question isn't just "is there a causal arrow?" but "which direction does it point, and is there a hidden driver?"

Examples worth holding in mind:

- Countries with more chocolate consumption win more Nobel Prizes per capita. (Spurious correlation — both are proxies for wealth) - Exercise correlates with better mental health. But does exercise improve mental health, or do mentally healthier people exercise more? (Bidirectional causation possible) - Children with more books at home do better academically. Books or parental education level? (Confound)

The Bradford Hill criteria give you a framework for evaluating whether correlation is likely to be causal. Key criteria: Is the association strong? Is it consistent across different studies and populations? Is there a dose-response relationship (more exposure → more effect)? Is there a plausible mechanism? Does the cause precede the effect? Does removing the cause remove the effect?

No single criterion proves causation. Together, they build a case.

Simpson's Paradox is when a trend appears in several different groups of data but disappears or reverses when these groups are combined. It's not rare — it appears regularly in medical data, economic data, and social statistics.

The canonical example: A university is accused of gender bias in graduate admissions. The aggregate data shows that men are admitted at a higher rate than women. But when you look department by department, women are admitted at equal or higher rates in every single department. How?

Women were applying disproportionately to competitive, low-acceptance departments. Men were applying to easier ones. The aggregate masked the real pattern.

This matters for any disaggregated data claim. Whenever you hear "on average, X" — ask: average of what population? What happens when you break it down by subgroup? The answer is sometimes the opposite of the aggregate.

The prosecutor's fallacy is the error of treating P(Evidence | Innocent) as if it were P(Innocent | Evidence). These are different quantities.

P(Evidence | Innocent): What's the probability of seeing this evidence if the person is innocent? P(Innocent | Evidence): What's the probability the person is innocent, given that we see this evidence?

The second is what matters in a trial. Bayes' theorem connects them through the base rate — the prior probability of guilt before the evidence. If the prior probability of guilt is very low (because the accused was selected from a large population, or because the crime is rare), even strong evidence may leave substantial probability of innocence.

Sally Clark was convicted of murdering her two infant sons in the UK in 1999. The prosecution's statistician testified that the probability of two SIDS deaths in the same family was 1 in 73 million. The jury convicted. The statistician had committed the prosecutor's fallacy — he calculated the probability of two SIDS deaths if the family was a random family, not accounting for the base rate of mothers who actually murder two children. He also assumed statistical independence between SIDS deaths, which may not hold (genetic and environmental factors could make both deaths likelier in the same family). Clark was eventually exonerated, but only after serving years in prison. The statistician's error killed her family.

Understanding Bayesian reasoning — updating your probability estimates based on evidence, weighted by priors — is perhaps the single most important statistical tool for everyday life. It doesn't require math. It requires asking: "What did I believe before? How strong is this new evidence? How should I update?"

Absolute vs. relative risk: Always convert relative risk claims to absolute risk. "50% more likely" means nothing without the baseline.

The 10,000-person visualization: When you hear a probability, imagine 10,000 people in the relevant situation. How many would this affect? Makes small probabilities real and large percentages concrete.

The confound question: When you see a correlation, name three possible third variables before you accept the causal story.

The replication check: Has this finding been replicated? In what sample? How large? One study is not evidence.

The incentive question: Who funded this research? Who benefits from this conclusion? This doesn't disprove the finding — but it should adjust your prior.

A population that can't reason statistically is a population that can be manipulated with statistics. Fear campaigns work. Vaccine scares work. Racialized crime statistics work — not because the numbers are false but because they're presented without base rates, without context, without confounds. A citizenry with statistical literacy would be harder to manipulate with selectively presented data.

More: personal decisions compounded over a lifetime — medical choices, financial risk, career bets — made without statistical reasoning will systematically underperform. Not because of bad luck but because of bad models of how probability and causation actually work.

Statistical thinking is not a specialized skill. It's a minimum requirement for navigating reality.