Central limit theorem: definition, formula and a worked example

What is the central limit theorem?

The central limit theorem (CLT) says that if you take many independent random samples of the same size from a population with a finite mean and variance, the averages of those samples follow an approximately normal (bell-shaped) distribution, and the approximation improves as the sample size grows. This holds even when the individual values are skewed, lumpy or far from normal.

In plain terms: single order values, session lengths or page load times are rarely bell-shaped, but the average of a reasonably large sample of them behaves predictably. That predictability is what lets you put a margin of error around an average, compare two averages, or run a z-test on conversion rates.

The formula: mean and standard error

For a population with mean μ and standard deviation σ, the sample mean x̄ of n independent observations is approximately:

x̄ ~ Normal( μ , σ / √n )

The spread of the sample mean, σ / √n, is the standard error. Two consequences matter in practice:

• The center does not move: on average, sample means land on the true mean.
• The spread shrinks with the square root of n. Four times as much data halves the standard error; it does not quarter it.

When σ is unknown (almost always), you plug in the sample standard deviation s. That substitution is why small samples use the t-distribution instead of the normal; see t-test.

Conditions

The classical CLT needs three things:

1. Independent observations. One visitor's order value should not determine another's. Repeated events from the same user are not independent; aggregate per user first.
2. Identically distributed, or at least not dominated by a few values. Mixing a holiday spike with normal days breaks this.
3. Finite variance. Extremely heavy-tailed data (a few customers spending thousands of times the median) can make convergence very slow.

A worked example: skewed order values

We simulated 200,000 order values from an exponential distribution with a mean of $60 (sample data). Individual orders are strongly right-skewed: skewness about 2.0, most orders small, a long tail of big ones. Then we drew 10,000 random samples at each size and looked at the distribution of the sample means.

Three things to read from the table:

• The averages are centered on $60 at every size.
• Their spread matches σ/√n almost exactly, as the theorem predicts.
• Skewness falls from 2.3 to 0.2, but at n = 30 it is still visible, and a textbook "x̄ ± 1.96 × s/√n" interval caught the true mean only 91.9% of the time, not 95%. With skewed revenue data, "30 is enough" is optimistic.

How large does n need to be?

There is no universal number. The rule of thumb n ≥ 30 comes from roughly symmetric data. The more skewed the data, the larger n should be:

• Close to symmetric (time on page after trimming obvious outliers): 30 can be fine.
• Moderately skewed (session duration, items per order): aim for a few hundred.
• Heavily skewed revenue with whales: thousands, or switch to a different statistic such as a trimmed mean, a median with a bootstrap interval, or a log transform.

For proportions such as conversion rate, a common check is that both n × p and n × (1 − p) are at least 10. A 2% conversion rate therefore needs at least 500 visitors before a normal-approximation interval makes sense.

What the central limit theorem does not say

• It does not make your data normal. The raw values stay skewed. Only the distribution of averages becomes approximately normal.
• It is about means and sums. Medians and percentiles such as P95 load time have their own sampling behavior; use bootstrap intervals for those.
• It does not fix biased samples. A large sample of the wrong users still gives the wrong answer; see sampling bias.
• It does not apply to dependent data without care. Daily traffic series are autocorrelated; treat days as a time series, not as independent draws.

Central limit theorem vs law of large numbers

The law of large numbers says the sample mean gets closer to the true mean as n grows. The central limit theorem goes further and describes the shape and size of the remaining error: approximately normal with standard deviation σ/√n. You need the second to build a confidence interval.

Where it shows up in analytics

• Confidence intervals for averages such as average order value or average session duration. See how to calculate a confidence interval and margin of error.
• Comparing two conversion rates with a z-test, which relies on the sampling distribution of a proportion being approximately normal; see statistical significance.
• Deciding whether to report a mean at all. For skewed metrics, the median or a percentile is often the better summary; see mean vs median.

In TapCub, performance cards switch between P95, P75, median and average for exactly this reason: an average load time can look fine while a slow tail hurts real visitors. When you need an interval around an average, export the raw events and compute it per user rather than per event.

Sources

• Kwak, S. G. and Kim, J. H. (2017). Central limit theorem: the cornerstone of modern statistics. Korean Journal of Anesthesiology, 70(2).
• NIST/SEMATECH. e-Handbook of Statistical Methods: sampling distributions.
• OpenStax. Introductory Statistics, chapter 7: The Central Limit Theorem.