The Finite Population Correction: Sampling Without Replacement From a Small Population

The standard error formulas taught in most courses quietly assume the population is effectively endless, or that sampled items are put back before the next draw. Real populations are often finite, and real samples are often drawn without replacement: a survey of 100 employees from a company of 500, an audit of 200 invoices from a batch of 800, a quality check on 30 units from a lot of 60. When the sample is a sizeable slice of the whole, each new observation reveals more of what remains, so the estimate is more accurate than the usual formula suggests. The finite population correction (FPC) is the factor that captures this.

The Idea

Imagine sampling every member of a population. The sample mean then equals the population mean exactly, and there is no sampling error at all. A census has zero standard error. A sample covering nearly the whole population is therefore almost error-free, while a sample of 1% of a huge population behaves as though the population were infinite. The correction interpolates between these extremes. For a sample of size n from a population of size N, the standard error of the sample mean is:

The second factor is the FPC. It equals 1 when n = 1 and falls toward 0 as n approaches N. A common working form replaces N − 1 with N and writes √(1 − n/N), which is nearly identical for all but very small populations.

Worked Example: An Exact Check

Problem: A population contains five values: 2, 4, 6, 8, 10. Its mean is 6 and its population variance is (16 + 4 + 0 + 4 + 16) ÷ 5 = 8. Samples of size n = 3 are drawn without replacement. How variable are the sample means?

There are C(5,3) = 10 equally likely samples, the counting rule from permutations and combinations. Their means are:

Sample mean	4.00	4.67	5.33	6.00	6.67	7.33	8.00
Number of samples	1	1	2	2	2	1	1

The average of these ten means is exactly 6, matching the population mean, and their variance is 1.333. The usual formula, ignoring the finiteness of the population, gives σ² ÷ n = 8 ÷ 3 = 2.667. This is exactly what sampling with replacement produces (all 125 ordered samples give a variance of 2.667). The ratio of the two is 1.333 ÷ 2.667 = 0.5, which equals the FPC applied to variance, (N − n) ÷ (N − 1) = (5 − 3) ÷ (5 − 1) = 0.5. Taking square roots, the correct standard error is √1.333 ≈ 1.155, against 1.633 from the uncorrected formula, a reduction of about 29%.

The correction is not an approximation here. It is an exact result, derived from the fact that draws without replacement are slightly negatively related: a large value already drawn leaves the remaining pool with a lower average, pulling the sample mean back toward the population mean.

A Practical Example: Confidence Intervals

Problem: A company has N = 500 employees. A sample of n = 100 gives a mean commute satisfaction score of 75 with a standard deviation of s = 12. Find a 95% confidence interval.

Without the correction: SE = 12 ÷ √100 = 1.2. With t = 1.984 (99 degrees of freedom), the margin of error is 1.984 × 1.2 ≈ 2.38, giving 72.62 to 77.38.

With the correction: the FPC is √[(500 − 100) ÷ 499] = √0.8016 ≈ 0.895, so SE = 1.2 × 0.895 ≈ 1.074. The margin of error is 1.984 × 1.074 ≈ 2.13, giving 72.87 to 77.13.

The interval is about 10% narrower for the same data, and honestly so: sampling one fifth of the workforce really does pin down the average better than sampling one fifth of an unlimited population would. The interval logic itself is unchanged from confidence intervals; only the standard error shrinks.

Adjusting a Sample Size

The correction also lowers the sample needed for a given precision. Suppose the usual formula for estimating a proportion to within ±5% at 95% confidence gives n₀ = 1.96² × 0.25 ÷ 0.05² ≈ 384, the standard starting point for sample-size formulas (see the sample size calculator). If the population is only N = 1,000, the adjusted requirement is:

About 28% fewer respondents are needed. For a population of 100,000, the same formula gives about 383, essentially the uncorrected figure, which shows how little the correction matters once the population is large.

When the Correction Can Be Ignored

The FPC matters only when the sample is a noticeable share of the population. A widely used rule of thumb says to apply it when n exceeds about 5% of N, with some texts using 10%. For N = 500, the factor is 0.996 at n = 5, 0.976 at n = 25, 0.950 at n = 50, and 0.895 at n = 100. For a sample of 25 (5% of the population) the standard error shrinks by only about 2.4%, which rarely changes conclusions. Most national polls, sampling well under 1% of a huge population, can ignore it completely, which is why standard textbook formulas omit it.

The related hypergeometric distribution is the same idea applied to counts: it is the exact model for successes when sampling without replacement, and its variance contains the same (N − n) ÷ (N − 1) factor. The binomial distribution is its infinite-population approximation.

Real-World Uses

Auditing and quality control: auditors sampling invoices from a fixed ledger, or inspectors testing units from a small production lot, routinely sample large fractions of what exists and apply the correction.

Organizational surveys: a satisfaction survey of a company, school, or membership organization often reaches a large percentage of the total, so intervals based on the uncorrected formula come out needlessly wide.

Official statistics: survey statisticians working with finite population frames, described in sampling methods, build the correction into their variance formulas for simple random and stratified designs.

Common Mistakes

Applying it to sampling with replacement. If items are returned before the next draw, or the population is effectively infinite, the standard formula is already right and the correction would understate the error.

Applying it to a non-random sample. The correction assumes a genuine random draw. A convenience sample of 40% of a population is still biased, and no correction repairs that. See types of bias in research.

Using the wrong population size. N must be the size of the population the sample was actually drawn from, such as the eligible frame, not a larger group you would like to generalize to.

Treating it as a fix for a small sample. The correction helps because the sample is large relative to the population. A small sample from a small population still has a large standard error.

What is the finite population correction factor?

It is √[(N − n) ÷ (N − 1)], the number by which the ordinary standard error is multiplied when n items are drawn without replacement from a population of N. It equals 1 for a single draw and 0 when the whole population is sampled.

Does it apply to proportions too?

Yes. The standard error of a sample proportion, √[p(1 − p) ÷ n], is multiplied by the same factor when sampling without replacement from a finite population.

Does the correction change the sample mean?

No. The sample mean is still an unbiased estimate of the population mean. Only its standard error, and therefore intervals and test statistics built on it, changes. See standard error for the uncorrected version.

Should I use the correction if I am unsure the population is finite?

If the sample is a tiny fraction of the population, it makes no practical difference either way. If the sample is a large fraction of a known finite population, the correction is appropriate and gives a more accurate interval.