The Sampling Distribution of the Mean

"The most useful thing the CLT gives you is the sampling distribution of the mean. It underpins many familiar confidence intervals and mean tests, but not every interval or test."

One Mean, Many Possibilities

Take a sample of size N from a population. Compute the sample mean. Now imagine doing this again — a different random sample of N, a different mean. Repeat thousands of times. The collection of those sample means has its own distribution, called the sampling distribution of the mean.

This is the object the CLT is really speaking about. When statisticians say 'the sampling distribution of the mean is approximately normal for large N,' they mean: if you imagine all the means you'd get from all possible samples of size N, that imagined collection's shape is bell-like, even when the underlying data is not.

The Standard Error

The standard deviation of the sampling distribution has its own name: the standard error (SE). For a sample mean, the SE equals σ / √N — the population standard deviation divided by the square root of the sample size. The SE shrinks as N grows, but only as the square root: to halve the uncertainty in your mean, you need four times the sample size, not twice.

This 'square root law' is one of the most important practical consequences of the CLT. Under simple independent sampling with stable variance, four times the sample size halves the standard error. Survey accuracy also depends on sampling design, nonresponse, weighting, clustering, and measurement error; millions of biased responses need not be accurate.

Why You Should Care

Many familiar intervals use a sampling distribution and an estimated standard error, but methods include exact, design-based, bootstrap, Bayesian, and other constructions. A proportion has its own sampling model. When the CLT's preconditions hold, this interval is meaningful: if you repeated the study many times, about 95% of such intervals would contain the true population value.

If a method's assumptions or standard-error calculation are wrong, nominal coverage can fail. Dependence and heavy tails can sometimes be handled with an appropriate design, model, robust variance, or resampling method. This is one of the major silent failures in published research.

The Operating Rule

The sampling distribution of the mean is the CLT's most practical gift. It is what makes 'inference from a sample to a population' possible at all. Track 05 will turn this object into confidence intervals and hypothesis tests. It is one important route to frequentist inference, alongside exact, likelihood-based, randomization, and other methods. Use it; know its preconditions; don't trust it when they don't hold.

Code

Sampling distribution of the mean from a non-normal population·python

import numpy as np
rng = np.random.default_rng(70)

# Population: highly skewed (lognormal).
population = rng.lognormal(mean=0, sigma=1, size=1_000_000)
pop_mean = population.mean()

# Draw M independent samples of size N each, compute each mean.
N = 100
M = 10_000
sample_means = np.array([
    rng.choice(population, size=N, replace=False).mean()
    for _ in range(M)
])

# The original population is far from normal — heavily right-skewed.
print(f"Population:    mean={pop_mean:.3f}  skew=heavily right (lognormal)")
print(f"Sample means:  mean={sample_means.mean():.3f}  std={sample_means.std():.3f}")
print(f"Theoretical SE = sigma/sqrt(N) = {population.std() / np.sqrt(N):.3f}")

# The sampling distribution of the mean is much narrower (~ sigma/sqrt(N))
# and much closer to normal than the population — even though the population
# itself is far from normal. That's the CLT doing its work on the mean's
# imagined distribution.

Exercise

A poll claims 50% support ±3% with N=1,000 respondents. Estimate the standard error implied. Then ask: under independence (each respondent gives an unbiased random sample), is the ±3% credible? Under correlated samples (respondents from the same geography or social network), would the true uncertainty be larger or smaller? Almost always larger — political polling regularly fails to account for the correlation.

Hint

For a proportion p, SE ≈ sqrt(p(1-p)/N). At p=0.5 and N=1000, SE ≈ 1.6%. A 95% interval is ~1.96 × SE ≈ ±3%. The ±3% is the arithmetic; the trustworthiness is a separate question about preconditions.

One Mean, Many Possibilities

The Standard Error

Why You Should Care

The Operating Rule

Code

External links

Exercise

Progress

Comments 0