Key takeaways
- Four conditions must hold: a fixed number of trials, two outcomes, a constant success probability, and independence.
- The coefficient counts arrangements. C(n, k) is the number of orders in which k successes can appear among n trials.
- Mean is n × p, variance is n × p × (1 − p). The spread peaks at p = 0.5 and shrinks near 0 or 1.
- “At least” means add the tail, not evaluate a single k.
- Normal approximation needs a large expected count in both successes and failures, plus a continuity correction.
- Independence is the fragile assumption. Clustered or without-replacement sampling breaks the model.
What the binomial model describes
A binomial experiment is a sequence of n identical trials, each with exactly two outcomes: success or failure. The probability of success p stays the same on every trial, the trials do not influence one another, and the count X is the total number of successes. Tossing a coin 20 times, testing 50 parts for a defect, and asking 400 voters whether they approve are all binomial in form.
The model answers questions such as “what is the chance of exactly 8 successes,” “what is the chance of at least 8,” and “what range of counts is plausible.” It does not answer “how many trials until the first success” — that is the geometric distribution — or “how many trials until the r-th success,” which is the negative binomial.
The formula and the binomial coefficient
For X ~ Binomial(n, p), the probability mass function is:
P(X = k) = C(n, k) × p^k × (1 − p)^(n − k)
C(n, k) = n! ÷ (k! × (n − k)!)
C(n, k), read “n choose k,” counts the orderings. With 10 trials and 7 successes there are C(10, 7) = 120 arrangements, because the seven successes can occupy any seven of the ten positions. Each arrangement has probability p^7 × (1 − p)^3, and the 120 arrangements are mutually exclusive, so their probabilities add.
Factorials grow quickly, so compute the coefficient with cancellation or a calculator rather than multiplying large factorials directly. C(10, 7) = C(10, 3) = (10 × 9 × 8) ÷ (3 × 2 × 1) = 120.
Worked example: 10 free throws at 70%
Suppose a player makes 70% of free throws and takes 10 independent shots. For exactly 7 baskets: C(10, 7) = 120, 0.7^7 = 0.0823543, and 0.3^3 = 0.027. Multiplying, 120 × 0.0823543 × 0.027 = 0.266828, or 26.68%.
The neighbouring probabilities are 23.35% for exactly 8, 12.11% for exactly 9, and 2.82% for all 10. Exactly 7 is the single most likely count, but it is still less than a 27% chance — a useful reminder that the mode of a distribution is not a prediction.
For exactly 0 baskets the probability is 0.3^10 = 0.0000059, about 1 in 169,350. Small probabilities at the extremes are normal even when the model fits well.
Mean, variance, and standard deviation
The expected count is μ = n × p, the variance is σ² = n × p × (1 − p), and the standard deviation is the square root of that variance. For n = 10 and p = 0.7: μ = 7, σ² = 10 × 0.7 × 0.3 = 2.1, and σ = 1.449.
The mean need not be a possible count. With n = 5 and p = 0.5 the mean is 2.5, which cannot occur. The variance is largest at p = 0.5 and approaches zero as p approaches 0 or 1, because the outcome becomes nearly certain. Doubling n doubles both the mean and the variance, so the standard deviation grows only with the square root of n.
Cumulative probabilities and “at least”
Questions usually ask for a range, not one count. “At least 8” means P(X ≥ 8) = P(8) + P(9) + P(10) = 0.233474 + 0.121061 + 0.028248 = 0.382783, or 38.28%. “At most 2” means add P(0) through P(2). “Between 5 and 7 inclusive” means P(5) + P(6) + P(7).
For a strict “more than 7,” start at 8. For “fewer than 8,” stop at 7. The complement rule is often faster: P(X ≥ 1) = 1 − P(X = 0). With a 2% defect rate and 50 parts, the chance of at least one defect is 1 − 0.98^50 = 1 − 0.3642 = 0.6358, or 63.6%.
Normal approximation and continuity correction
When n is large, the binomial histogram becomes bell-shaped and can be approximated by a normal distribution with the same mean and standard deviation. A common rule is to require n × p ≥ 10 and n × (1 − p) ≥ 10; some texts use 5. For n = 10 and p = 0.7 the expected failures are only 3, so the approximation is marginal.
A continuity correction improves accuracy because a discrete count is being replaced by a continuous curve. To approximate P(X ≥ 8), use the boundary 7.5: z = (7.5 − 7) ÷ 1.449 = 0.345, giving a tail of about 0.365. The exact answer is 0.383, so the approximation is close but not exact. For small n, compute the binomial sum directly.
Binomial, Poisson, and hypergeometric
The Poisson distribution is the limiting case when n is large, p is small, and λ = n × p is moderate; it counts rare events rather than a fixed number of trials. The hypergeometric distribution applies when sampling is without replacement from a finite population, so p changes after each draw. The negative binomial counts trials needed to reach a fixed number of successes.
Choosing among them is a modelling decision, not a formula preference. Drawing 5 cards from a deck is hypergeometric, not binomial, because the composition of the deck changes. Counting arrivals at a help desk in an hour is Poisson, because there is no fixed n. State the sampling scheme before quoting a probability.
Common mistakes
- Using a percentage as p. Enter 0.7, not 70, or every probability will be wildly wrong.
- Forgetting the coefficient. p^k × (1 − p)^(n − k) is the probability of one specific ordering, not of any ordering.
- Answering “at least” with a single term. Sum the tail or use the complement.
- Assuming independence when sampling without replacement. Use the hypergeometric distribution, or a finite-population correction.
- Applying the normal approximation with tiny expected counts. Check n × p and n × (1 − p) first.
- Rounding p before computing. Carry full precision until the final percentage.
Practical caveats
Real trials are rarely perfectly independent. Free throws can be affected by fatigue or confidence, defects can cluster by machine or shift, and survey responses cluster by household. If the observed variance is much larger than n × p × (1 − p), the data are overdispersed and a binomial model understates uncertainty; beta-binomial or quasi-binomial models are the usual alternatives.
The probability p is itself an estimate. A rate measured from 30 trials carries substantial uncertainty, and plugging it in as if it were exact overstates confidence. Report the sample size behind p, and treat binomial answers as model-based projections rather than guarantees. When n is very large, even a tiny violation of independence can make exact binomial intervals too narrow.
Frequently asked questions
What is the difference between binomial and Bernoulli?
A Bernoulli trial is a single yes-or-no experiment. The binomial distribution is the sum of n independent Bernoulli trials with the same success probability.
Can a binomial distribution have a mean that is not a whole number?
Yes. The mean n × p is an expected value, not a count that must occur. With 5 trials at p = 0.5 the mean is 2.5.
When should I use the normal approximation?
When both n × p and n × (1 − p) are at least about 10, and preferably when n is several dozen. Use a continuity correction and compare with the exact sum when in doubt.
What does “independent trials” mean in practice?
The outcome of one trial does not change the probability of success on another. Sampling without replacement from a small population violates this; sampling with replacement or from a very large population approximately satisfies it.
Why is exactly 7 successes not the most likely outcome overall?
It is the most likely single count, but many other counts are possible. The probabilities of all counts must sum to 1, so the mode can still be well below 50%.
How do I handle “at least one” questions?
Use the complement: P(X ≥ 1) = 1 − P(X = 0) = 1 − (1 − p)^n. This is faster and less error-prone than summing every term from 1 to n.
Further reading