The Hypergeometric Distribution: When Sampling Without Replacement Really Matters

The binomial distribution requires independent trials with a constant probability of success — a condition that quietly breaks whenever items are drawn from a finite population without replacement, since removing an item changes the composition of what's left. When that violation is severe enough to matter, the correct tool is the hypergeometric distribution — the binomial's close cousin, purpose-built for exactly this situation.

Binomial vs. Hypergeometric: The Core Difference

Both distributions count the number of "successes" in a fixed number of draws from a population containing two types of items. The difference is entirely about replacement: the binomial distribution assumes each draw is independent, with the same success probability every time (true when sampling with replacement, or when drawing from an effectively infinite population). The hypergeometric distribution instead tracks a population of fixed, known size, where each draw removes an item and genuinely changes the probabilities for every draw that follows — sampling without replacement from a finite population.

The Formula

For a population of size N containing K "success" items (and N−K "failure" items), the probability of finding exactly k successes in a sample of n items drawn without replacement is:

The numerator counts the number of ways to choose exactly k successes from the K available successes, multiplied by the number of ways to choose the remaining n−k draws from the N−K available failures. The denominator, C(N,n), is the total number of equally likely ways to choose any n items from the full population of N — the same combinatorics that underpin the binomial formula's C(n,k) term, just applied to the whole finite population instead of trial-by-trial.

Worked Example: Quality Control Sampling

Problem: A shipment contains 50 parts, of which 10 are defective. An inspector randomly selects 5 parts without replacement for testing. What's the probability that exactly 2 of the 5 selected parts are defective?

Here N = 50 (total parts), K = 10 (defective parts), n = 5 (sample size), k = 2 (defectives found).

C(10,2) = 45. C(40,3) = 9,880. C(50,5) = 2,118,760.

P(X=2) = (45 × 9,880) / 2,118,760 = 444,600 / 2,118,760 ≈ 0.2098

There's roughly a 21% chance the inspector's sample contains exactly 2 defective parts. For comparison, the full distribution across all possible outcomes: P(X=0) ≈ 0.3106, P(X=1) ≈ 0.4313, P(X=2) ≈ 0.2098, P(X=3) ≈ 0.0442, with smaller probabilities for X=4 and X=5 — these five values sum to 1, as any complete probability distribution must.

Mean and Variance

For the quality control example: μ = 5 × (10/50) = 1.0 — on average, an inspector examining 5 parts from this shipment expects to find exactly 1 defective one. The variance is σ² = 5 × 0.2 × 0.8 × (45/49) ≈ 0.735, giving σ ≈ 0.857.

Notice the mean formula, n·(K/N), looks identical to the binomial mean np (since K/N plays the role of p here) — but the variance carries an extra factor, (N−n)/(N−1), called the finite population correction. This term is always less than 1 (as long as n > 1), meaning a hypergeometric distribution always has less variance than a binomial distribution with the matching p — sampling without replacement is inherently more consistent than sampling with replacement, because it can't draw the same "lucky" or "unlucky" item twice.

When the Binomial Is a Safe Approximation

As the finite population correction term shows, when the population N is very large relative to the sample size n, (N−n)/(N−1) gets very close to 1, and the hypergeometric distribution converges toward the binomial distribution with p = K/N. In the worked example above, n=5 is exactly 10% of N=50 — right at the common rule-of-thumb boundary (sample should be under about 10% of the population) for when a binomial approximation starts becoming unreliable. Treating the same problem as binomial with p=10/50=0.2 gives P(X=2) = C(5,2)×0.2²×0.8³ ≈ 0.2048 — close to the exact hypergeometric value of 0.2098, but not identical, illustrating exactly how the approximation degrades as the sample size approaches a meaningful fraction of the population.

A Second Worked Example: Card Drawing

Problem: Four cards are drawn from a standard 52-card deck without replacement. What's the probability of drawing exactly 2 hearts (there are 13 hearts in a standard deck)?

N = 52, K = 13, n = 4, k = 2.

P(X=2) = [C(13,2) × C(39,2)] / C(52,4) = (78 × 741) / 270,725 ≈ 0.2135

There's roughly a 21.4% chance of drawing exactly 2 hearts in a random 4-card hand — a calculation that underlies odds computations across many card games, where drawing without replacement is the entire point of the game and the binomial distribution would give a meaningfully wrong answer if mistakenly applied.

Real-World Uses

Quality control and acceptance sampling: exactly the scenario in the worked example — inspecting a sample from a finite batch to decide whether to accept or reject the whole shipment, standard practice in manufacturing.

Lottery and card game probabilities: any game built around drawing without replacement from a fixed, known set of items uses hypergeometric logic under the hood, even when players never see the formula explicitly.

Capture-recapture population estimation: ecologists estimating an animal population's size by tagging some individuals, releasing them, then recapturing a new sample use the hypergeometric distribution's structure to work backward from the recapture data to an estimate of the total population.

Common Mistakes

Applying the binomial formula when the sample is a large fraction of the population. The 10%-of-population rule of thumb exists precisely because this is where the approximation error becomes large enough to matter — always check this ratio before defaulting to the simpler binomial formula.

Forgetting the finite population correction when computing variance. Using the plain binomial variance formula np(1−p) for a without-replacement scenario overstates the true variability, since it ignores the (N−n)/(N−1) correction that without-replacement sampling introduces.

Mixing up which quantity is N and which is n. N is the total population size; n is the sample size drawn from it. Swapping them produces a nonsensical calculation (often literally impossible, since combinations require the "choose" number to be no larger than the total) rather than just a wrong number, which is usually a quick way to catch the error.