Central Limit Theorem Calculator

What the central limit theorem says

The central limit theorem describes what happens when you average a batch of numbers and repeat that process. Draw a random sample of size n from any population with mean mu and standard deviation sigma, write down its mean, then draw another sample and do the same. The pile of sample means forms its own distribution, called the sampling distribution of the mean.

The theorem makes three promises. The average of that pile equals mu. Its standard deviation is sigma divided by the square root of n, a quantity called the standard error. And as n grows, the shape of the pile approaches a normal curve, even when the population itself is lopsided. The NIST glossary entry states it the same way: for a random sample of size n, sample means are approximately normal with mean m and variance s squared over n as the sample size increases.

One number is easy to misread. The n in the formula is how many values go into each average, not how many samples you take. Sampling 50 people once gives a standard error based on 50. Sampling 500 people once gives one based on 500. Repeating a 10 person sample a thousand times does not shrink the standard error. It only maps the spread more clearly.

Why a normal curve shows up at all is worth a moment. Each sample mean is an average of independent values, and averaging pulls extreme values toward the middle. A population of incomes with a long right tail produces sample means with a much shorter right tail, and by the time n reaches a few dozen, the pile looks symmetric. Income, die rolls, package weights and wait times behave the same way once averaged.

Standard error and the z-score of a sample mean

Two formulas carry most of the work here. The standard error of the mean is SE = sigma / square root of n. The z-score of an observed sample mean is z = (xbar - mu) / SE. The z-score counts how many standard errors separate your sample mean from the population mean.

Take a population with mean 100 and standard deviation 15, and samples of 25 values. The standard error is 15 / 5 = 3. A sample mean of 105 sits 5 points above the population mean, so z = 5 / 3 = 1.667. On a standard normal curve that puts the sample mean at the 95.2 percentile: about 95.22 percent of samples of this size land at or below 105, and 4.78 percent land above it. A two-sided reading, the chance of being at least that far from 100 in either direction, is 9.56 percent.

Notice what this z-score is not. It is not the z-score of a single person's measurement, which divides by sigma, not by sigma over the square root of n. A person scoring 105 against a mean of 100 with a standard deviation of 15 has z = 5 / 15 = 0.333, an ordinary result. The mean of 25 such people has z = 1.667, which is unusual. Averaging compresses noise, and the compression is what the standard error measures.

The standard error is the number a confidence interval is built on. The NIST handbook chapter on confidence limits for the mean writes the interval as the sample mean plus or minus a critical value times the standard error, which is the same arithmetic used below.

How much a bigger sample helps

Standard error shrinks with the square root of the sample size, so the sample has to grow fast to make a visible difference. Going from 25 to 100 observations, a factor of 4, cuts the standard error in half, from 3.0 to 1.5. Going from 25 to 50, a factor of 2, cuts it by about 29 percent, from 3.0 to 2.12. Each doubling of n multiplies the standard error by 0.707, which is one over the square root of 2.

The right column is the practical number. It says how far a sample mean usually lands from the population mean, with 95 percent confidence. At 25 observations that distance is 5.88 points on a scale where the population standard deviation is 15. At 900 observations it is 0.98 points. To cut the margin from about 6 points to about 3, you need four times the sample, not twice.

A common planning mistake follows from this. Teams often double a sample and expect the error to halve. It does not. Doubling buys a 29 percent reduction. Cutting the margin in half costs four times the data, and cutting it to a tenth costs a hundred times the data. That is why survey firms settle on margins near 3 or 4 points rather than chasing 1 point.

Reading probabilities off the normal curve

Once you have a z-score, a standard normal table or the calculator above turns it into a probability. The function that returns the area to the left of z is usually written as the capital Greek letter phi. For z = 1.667 the area is 0.9522, so the chance that a sample mean lands at or below 105 is 95.22 percent.

Three readings come out of the same number, and mixing them up is the most common error on this kind of page. One-sided above: 100 percent minus 95.22 percent, so 4.78 percent. One-sided below: 95.22 percent. Two-sided, meaning the chance of being at least 1.667 standard errors away in either direction: 2 x 4.78 = 9.56 percent. If a quality team sees a batch mean of 105 and asks whether the process shifted, the two-sided number is the one to quote, because a mean of 95 would be just as surprising.

The same standard error builds an interval around the observed mean. Using 1.96 as the 95 percent critical value, the interval is 105 plus or minus 1.96 x 3, which runs from 99.12 to 110.88. That interval is about the population mean, not about individual measurements: a single measurement from this population would usually fall between 70.6 and 129.4, far wider than the interval for a mean of 25.

One caution about the 95 percent label. It describes the procedure, not any single interval. If you repeated the study many times, about 95 percent of the intervals built this way would contain mu. Saying there is a 95 percent chance that mu sits in this particular interval is a different claim, and it is the one statisticians argue about.

Small populations: the finite population correction

When the population is small relative to the sample, the standard error formula overstates the uncertainty, because sampling without replacement from a fixed list gives more information than sampling from an unlimited pool. The fix is one extra factor: SE = (sigma / square root of n) x square root of ((N - n) / (N - 1)), where N is the population size.

Say a company has 2,000 customers and surveys 100 of them, with a satisfaction score whose population standard deviation is 15. The uncorrected standard error is 1.5. The correction factor is the square root of 1,900 over 1,999, which is 0.9749, so the standard error drops to 1.5 x 0.9749 = 1.462. The 95 percent margin falls from 2.94 to 1.96 x 1.462 = 2.87. Modest, but real.

The factor only matters when the sample is a large share of the population. At N = 200 and n = 100 you are measuring half the population, and the factor is 0.709, a 29 percent cut in the standard error. At N = 20,000 it is 0.9975, which changes nothing worth reporting. The usual rule of thumb is to ignore the correction while the sample is under 5 percent of the population and to apply it above 10 percent. When n equals N you have a census, the factor is 0 and there is no sampling error left, only measurement error.

Planning a sample size

The standard error formula runs backwards just as well, which is how sample sizes get chosen before a study starts. To hit a target standard error m, solve for n: n = (sigma / m) squared. With a population standard deviation of 20 and a target standard error of 2, n = 10 squared = 100 samples. To hit a 95 percent margin of 1.5 points with sigma of 12, use the critical value: 1.96 x 12 / 1.5 = 15.68, and 15.68 squared = 245.9, so 246 observations.

Proportions use a variant because the standard deviation depends on the share you are measuring. The standard error of a proportion is the square root of p x (1 - p) / n, and the worst case sits at p = 0.5, where the numerator is 0.25. For a 3 point margin at 95 percent confidence, n = 3.8416 x 0.25 / 0.0009 = 1,067.1, so 1,068 people. That is the number behind the familiar line in polling stories about a margin of error of plus or minus 3 points on a sample of about a thousand adults.

Two adjustments are standard. Round up, always, because a fractional person does not exist. Then inflate for nonresponse: if you expect 40 percent of invited people to answer, divide the target by 0.6. A target of 246 becomes 246 / 0.6 = 410 invitations. Neither step is optional in practice, and leaving them out is why so many student surveys end up with a margin twice as wide as planned.

When the theorem does not hold

The theorem has conditions, and a calculator cannot check them for you. The first is independence: each observation should not be influenced by another. Sampling without replacement breaks independence slightly, which is why the 10 percent rule exists, and sampling clusters such as households or classrooms breaks it badly. A survey of 1,000 people drawn as 50 households of 20 does not carry the same information as 1,000 separate people.

The second condition is the sample size relative to the shape of the population. The familiar rule that 30 is enough comes from textbook practice and holds for mildly skewed data. A population with a heavy tail, such as claim sizes or viral post views, can need hundreds of observations before the pile of means looks normal.

The third condition is that the data are drawn from the population you think you are studying. No sample size fixes a biased frame. A poll that only calls landlines has a standard error of 2 points and a bias of several points, and the bias does not shrink as n grows. For proportions, the working rule is at least 10 successes and 10 failures in the sample; below that the normal curve is a poor stand-in for the real distribution, and exact methods are safer.

Mistakes that change the answer

The t distribution row deserves a number. At n = 25 the degrees of freedom are 24, and the 95 percent critical value is 2.064 rather than 1.96. The margin becomes 2.064 x 3 = 6.19 instead of 5.88, about 5 percent wider. The gap shrinks as n grows: at n = 100 the critical value is 1.984, and at n = 1,000 it is 1.962.

One smaller trap is worth knowing. Mixing units, such as a population mean in dollars with a standard deviation in cents, produces a z-score that is off by 100. Convert both to the same unit before typing them in.