The Central Limit Theorem
Take samples from almost any distribution (lopsided, lumpy, whatever), average each sample, and those averages pile up into a smooth, predictable bell curve. That is one of the most surprising facts in all of statistics. Let's not take it on faith. Let's watch it happen.
The setup
Imagine a population: say, the reaction times of every student in a huge psychology study. The shape of that population can be anything. Most real data is not a tidy bell curve; reaction times are usually skewed, with a long tail of slow responses.
Now we do something simple and repeat it over and over:
- Reach into the population and grab a random sample of size n.
- Compute the mean of just that sample — one number.
- Write that number down, and do it all again.
The collection of all those sample means has its own distribution, the sampling distribution of the mean. The Central Limit Theorem tells us what it looks like.
🎮 Sampling Distribution Playground
Pick a population shape, choose a sample size, then draw samples and watch the averages collect below.
① The population you're sampling from
② Your latest sample (orange line = its average)
③ The sampling distribution — every orange average lands here. The dashed curve is the bell curve the CLT predicts.
What you should notice
Play with it for a minute, and three things jump out. Together they are the Central Limit Theorem:
1. The averages form a bell curve — even when the population is skewed or lumpy. Switch to the "Skewed" or "Bimodal" population and you'll still get a symmetric bell in panel ③.
2. Bigger samples → tighter bell. Drag n from 1 up to 50 and watch panel ③ get narrower. Averaging more values per sample makes the averages cluster more tightly around the truth.
3. The bell is centered on the population mean. The averages pile up right over the population's true mean instead of drifting away from it. Note that this means the bell is centered on μ, not on the middle of the panel: switch to "Skewed" and the whole bell slides left, because that population's mean is about 16 rather than 50. Watch the dashed μ line move with it.
The one formula worth remembering
The spread of the sampling distribution has a name, the standard error, and a tidy formula (it is the engine behind every confidence interval you will build):
SE = σ / √n
where σ is the population's standard deviation and n is your sample size. Notice the √n on the bottom: to halve the standard error, you need four times the data. That diminishing return is why the playground's bell tightens quickly at first, then more slowly. Compare the "SE" you measured with the "σ/√n" the theorem predicts — they should match closely once you've drawn enough samples.
The fine print on "almost any"
The theorem is generous, not unconditional. It asks that the population have a finite variance, which every population you will meet in a psychology study does. The classic counterexample is the Cauchy distribution, whose tail is so heavy that its variance is infinite, and there the whole mechanism fails: averaging 1,000 Cauchy values leaves the averages spread exactly as wide as single values were. In a simulation of 400,000 repeats the middle half of the averages spanned 2.00 units at n = 1 and 2.02 units at n = 1,000. No amount of data buys precision, because the occasional monstrous draw is always big enough to drag the average with it.
The condition you actually have to think about is the second one, and it is about n. "As n grows" is a limit, so the working question is how far along that road your sample sits. The folk answer is n ≥ 30. On the Skewed population in the playground above, an exponential shape with a population mean near 16 and a skewness of 1.84, it buys the n = 30 row below. Each row is 300,000 simulated studies, each building an ordinary 95% confidence interval for the mean:
| Sample size | Interval really covers μ | Lands entirely below μ | Lands entirely above μ |
|---|---|---|---|
| n = 10 | 90.2% | 9.4% | 0.4% |
| n = 30 | 92.8% | 6.4% | 0.8% |
| n = 50 | 93.6% | 5.4% | 1.0% |
| n = 100 | 94.3% | 4.4% | 1.3% |
| n = 200 | 94.7% | 3.7% | 1.6% |
Thirty gets you 92.8% rather than 95%, and the failures are lopsided: at that size the interval lands entirely below μ eight times for every one time it lands entirely above. A right-skewed population hides most of its mass in a long thin tail, so a typical sample misses the rare huge values and produces a mean a little low (52% of samples of 30 do) together with a standard deviation a little small, which makes the interval both low and narrow. Both errors push the same way.
That asymmetry is the population's skew surviving the averaging in diluted form. The sampling distribution of the mean carries the population's skewness divided by √n: in the simulation it came out at 0.336 for n = 30, against a predicted 1.84/√30 = 0.336. Leftover skew therefore fades on the same schedule the standard error does, and halving it again takes four times the data.
None of this makes 30 a bad number. It makes it a statement about shape, and the shape does arrive quickly, which is what panel ③ shows you. Coverage is a stricter demand than shape, and a badly skewed population meets it later.
Why it matters: confidence intervals, t-tests, p-values: almost everything that comes next leans on the fact that sample means behave predictably like this. The CLT is the bridge from "messy real data" to "clean, usable inference."
Problem 7 of the practice problems takes a clearly right-skewed population of pizza delivery times and asks why a sample of 36 lets you use the normal curve anyway.
Common questions
What sample size is large enough for the Central Limit Theorem?
The folk rule is n ≥ 30, and for mildly skewed data that's usually plenty — sample means settle into a near-normal shape quickly. But it's a rule of thumb, not a law: heavily skewed or outlier-prone populations can need substantially more, while symmetric populations are fine much earlier. When in doubt, look at your data's shape rather than trusting the magic 30.
Does the Central Limit Theorem make my data normally distributed?
This is the classic misreading. The CLT says the distribution of sample means approaches normal as n grows. Your raw data keeps whatever shape it has; skewed data stays skewed no matter how much you collect. The bell curve emerges one level up, in the averages across samples, which is what t-tests and confidence intervals actually rely on.
My data has a few enormous outliers. Will the CLT still rescue me?
Usually yes, but later than you would like. The theorem needs the population to have a finite variance, and genuine heavy tails (income, some reaction-time distributions, network sizes) make the convergence slow rather than impossible: the sampling distribution of the mean inherits the population's skewness divided by √n, so a badly skewed population still shows a lopsided sampling distribution at n = 30 and a confidence interval that covers less often than it claims. Two practical moves: look at whether the extremes are real data or errors, and prefer a bootstrap interval, which does not assume the shape has arrived.