What is statistics actually used for?

Statistics is learning from data under uncertainty

Statistics is the discipline of drawing conclusions from data when the data is incomplete and noisy. You almost never get to measure everyone. You measure some, then say something responsible about everyone, with an honest account of how wrong you might be.

That is the whole subject in one move: go from a sample to a population, and carry an uncertainty estimate with you. A confidence interval is that move written down; a hypothesis test is that move turned into a decision. Everything in the rest of this page is a variation on it.

The vocabulary follows the same split. A number computed from your sample is a statistic, such as the sample proportion $\hat{p}$ (read "p-hat"). The matching number for the whole population is a parameter, such as $p$, and you almost never get to see it. See parameter vs statistic and descriptive vs inferential statistics.

Medicine and clinical trials

A clinical trial is a designed experiment. Subjects are randomly assigned to a treatment or a control, often with a placebo, and often with blinding so that neither the patients nor the people assessing outcomes know who got what. The random assignment is not a formality; it is the only reason the trial can end with a claim about cause rather than a claim about association.

The course idea doing the work is experimental design, followed by two-sample inference. When the trial ends you compare a recovery rate in one group against the other, which is a two-proportion procedure, or you compare mean outcomes, which is a two-sample $t$ procedure. See experimental design, Unit 3, and Unit 4.

The p-value in a trial report answers one narrow question: if the treatment did nothing, how surprising would a result like this be? That is why a trial can be statistically significant and still clinically unimportant, and why sample size and effect size get reported alongside it. See what a p-value means.

Public health

Public health usually cannot randomize. You cannot assign people to smoke or to live beside a busy highway, so most of the evidence is observational. That constraint is why epidemiologists are so careful about confounding: a third variable moving with both the exposure and the outcome can manufacture an association out of nothing.

The ideas here are observational study design, confounding, and conditional probability. A rate such as cases per 100,000 people is a conditional statement about a subgroup, and swapping which group you condition on changes the meaning completely. See correlation vs causation, confounding vs lurking variable, and conditional probability.

Screening tests are the sharpest case of that reversal. The probability a test is positive given that someone has the disease is not the probability someone has the disease given a positive test, and when the disease is rare those two numbers can differ enormously. See conditional probability and why order matters.

Polling and public opinion

A poll asks a thousand people and reports a claim about millions. That is only legitimate because of how the thousand were chosen. Random selection is what makes a sample a fair miniature of the population, and it is what makes a margin of error mean anything at all.

The ideas are sampling methods, bias, and the confidence interval for a proportion. A reported margin of error of 3 points is the half-width of a 95% confidence interval, and it covers random sampling variation only. It says nothing about people who refused to answer, or about a question worded to nudge them. See random sampling, how to identify the type of bias, and the margin of error calculator.

This is also where extra sample size stops helping. A larger sample shrinks random error but does nothing to a biased sampling method, which is why a huge voluntary-response web poll is worth less than a small random one. See does a bigger sample fix bias.

Sports

Sports analytics is mostly regression plus a hard-won respect for variability. You fit a model relating an input to an outcome, read the slope as a rate of change, and use $r^2$ (r-squared, the coefficient of determination) to say how much of the variation in the outcome the model explains. Then people argue about what the residuals, the gaps between actual and predicted values, are really telling them.

The other half of the job is knowing how much of a hot streak is noise. Short runs of data are wildly variable, so a player who goes 8 for 12 across three games has told you very little about their true ability. That is the sampling variability idea from Unit 2, and it is why analysts trust season-long numbers over weekly ones. See Unit 5, the least-squares regression line, and r vs r-squared.

Prediction also has a range limit. A model fitted on players aged 22 to 34 says nothing trustworthy about a 40-year-old, because you have left the data behind. See interpolation vs extrapolation.

Quality control and manufacturing

A factory cannot test every unit it makes, so it samples on a schedule and watches for the process drifting off target. The tools are the normal model, the standard deviation, and the sampling distribution of the sample mean. A control chart is a sampling distribution turned sideways, with limits drawn a fixed number of standard deviations from the target value.

The course ideas are the normal distribution, z-scores, and the central limit theorem. A z-score, $z = \frac{x - \mu}{\sigma}$, reports how many standard deviations a value sits from the mean, with $\mu$ (mu) the population mean and $\sigma$ (sigma) the population standard deviation. Because sample means vary less than individual measurements do, a chart built on means catches a drift that single readings would hide.

That last point is the practical payoff of the central limit theorem, and the worked example below makes it concrete. See how to find a z-score, the empirical rule, and the central limit theorem.

Machine learning

Machine learning is statistics with far more parameters and far less patience for closed-form answers, but the underlying problem has not changed: you hold a sample and you want to generalize. Splitting data into training and test sets is a sampling argument. Overfitting is reading noise as signal, which is the same mistake as taking a three-game hot streak seriously.

Linear regression is the base case that nearly every supervised model generalizes, and $\hat{y} = a + bx$, where $\hat{y}$ (read "y-hat") is the predicted value, is where the family starts. Model evaluation borrows vocabulary straight from Unit 3: a false positive is a Type I error and a false negative is a Type II error, with the same trade-off between them. See Type I vs Type II errors and the least-squares regression line.

The limit is the same one too. A model trained on observational data predicts; it does not establish a cause, however accurate its predictions are. See correlation vs causation.

Randomization is what licenses the leap

Every field above turns on one question: what does the design let you say? The Course and Exam Description states the rule in two halves, and they do different jobs.

• Random selection from a population lets you generalize a result to that population. Without it, bias may result and a sample statistic cannot be assumed to describe the population at all.
• Random assignment to treatments lets you conclude that a statistically significant difference was caused by the treatment, because assignment tends to balance the effects of uncontrolled variables across the groups.

Mixing the two up is one of the most common errors students make, and the Course and Exam Description singles it out. A random sample with no assignment gives you an association that applies to the population. A randomized experiment on volunteers gives you a cause, but only for people like those volunteers. See experiments vs observational studies and experiment vs survey.

If you want to build these ideas from the start, work through learn statistics, try dropping a dataset into the data analyzer, or see how the same ideas are assessed in the AP Statistics course.