The Binomial Distribution

A binomial trial (or Bernoulli trial) is a single event with exactly two possible outcomes, conventionally labeled "success" and "failure," where success occurs with probability p and failure with probability 1−p. A binomial experiment consists of n such trials performed under three conditions: the trials are independent of one another (the outcome of one trial does not change the probability of success in another), p is the same constant for every trial, and n is fixed in advance rather than determined by the outcomes themselves. The binomial random variable X is defined as the total number of successes across the n trials, and it can take any integer value from 0 to n. This is a direct extension of the Multiplication Rule and Independence: if trials are independent, the probability of any particular sequence of successes and failures is the product of the individual trial probabilities. The binomial distribution asks a slightly different, harder question — not the probability of one specific sequence, but the probability of any sequence that contains exactly k successes, regardless of order. That requires counting how many such sequences exist, which is where combinatorics enters. The distribution is a special case of the general framework of Random Variables and Probability Distributions: X is a discrete random variable, and the binomial formula is its probability mass function, assigning a probability to every possible integer outcome from 0 through n. Biologically, "success" is whatever outcome the investigator defines: a wrinkled seed, a female offspring, a surviving fish, a patient who improves. The label is arbitrary — what matters is that each trial is a coin-flip-like event with a fixed, known (or hypothesized) probability, repeated a fixed number of times, with trials that don't influence each other.

Step 1 — Start from one specific ordered outcome. Suppose n = 5 seeds are examined and each seed independently has probability p of being wrinkled. Consider one specific sequence: wrinkled, wrinkled, round, round, round (WWRRR). Because the trials are independent, the probability of this exact sequence is the product of each trial's probability: P(WWRRR) = p · p · (1−p) · (1−p) · (1−p) = p²(1−p)³

Step 2 — Recognize that order doesn't matter for the count. The event "exactly 2 wrinkled seeds out of 5" is not just WWRRR — it also includes WRWRR, RWWRR, RRWWR, and every other arrangement with exactly two W's and three R's. Each of these arrangements has exactly the same probability, p²(1−p)³, because multiplication is commutative — the order of multiplying probabilities doesn't change the product.

Step 3 — Count the arrangements combinatorially. The number of distinct ways to choose which 2 of the 5 trials are the "successes" is given by the binomial coefficient: C(n,k) = n! / [k!(n−k)!] For n = 5, k = 2: C(5,2) = 5!/(2!3!) = 10. There are 10 distinct orderings that all produce exactly 2 wrinkled seeds.

Step 4 — Multiply count by probability of each arrangement. Since the 10 arrangements are mutually exclusive (a sequence can't simultaneously be WWRRR and WRWRR), their probabilities simply add. Because each has identical probability p²(1−p)³, the total probability is: P(X = 2) = C(5,2) · p²(1−p)³ = 10 p²(1−p)³

Step 5 — Generalize. For any n trials and any count k, the same logic gives the binomial probability mass function: P(X = k) = C(n,k) pᵏ(1−p)ⁿ⁻ᵏ, for k = 0, 1, 2, …, n

Step 6 — Derive the mean and variance. Each individual trial is a Bernoulli variable with mean p and variance p(1−p). Because the n trials are independent, their sums add linearly: Mean: E(X) = np Variance: Var(X) = np(1−p) Standard deviation: SD(X) = √[np(1−p)] Intuitively, if p = 0.25 (a recessive phenotype under simple Mendelian dominance) and n = 20 offspring, the expected count is np = 5 wrinkled seeds, with a standard deviation of √(20 × 0.25 × 0.75) ≈ 1.94. Observing anywhere from about 3 to 7 wrinkled seeds is unremarkable; observing 0 or 15 would be a red flag demanding explanation (linkage, lethality, or misclassification).

Step 7 — Shape depends on p and n. When p = 0.5, the distribution is symmetric. When p is far from 0.5, the distribution is skewed, though it becomes more symmetric as n grows — the seed of the later The Normal Distribution approximation. When n is large and p is small (rare-event trials, like mutation per cell division), the binomial collapses toward the shape used in The Poisson Distribution.

The binomial distribution is a mathematical theorem, not an empirical discovery — it follows deductively from the multiplication and addition rules of probability once the assumptions (independence, fixed n, constant p) are granted. But its power in biology was demonstrated empirically, most famously through the reanalysis of Gregor Mendel's pea-plant data (1856–1863) and, in a related discrete-counting logic, the Luria–Delbrück fluctuation test (1943).

Mendel's ratios as binomial predictions. Mendel crossed pea plants differing in single traits (seed shape, seed color, flower position) and self-fertilized heterozygous F1 hybrids to produce F2 offspring. His theory of particulate inheritance predicted that F2 offspring would show dominant and recessive phenotypes in a 3:1 ratio — that is, each seed independently has probability p = 0.75 of showing the dominant phenotype and 0.25 of showing the recessive phenotype, with all seeds behaving independently of one another. In one famous cross (seed shape), Mendel examined 7,324 seeds and observed 5,474 round and 1,850 wrinkled — a ratio of 2.96:1, very close to the predicted 3:1. This is precisely a binomial experiment: n = 7,324 independent trials, hypothesized p = 0.75, and the observed count of "wrinkled" successes (1,850) can be compared against the binomial expectation of np = 5,493 round / 1,831 wrinkled. The closeness of the fit — and, decades later, R. A. Fisher's controversial 1936 statistical re-examination showing Mendel's ratios were suspiciously close to expectation across all his experiments (a chi-square analysis suggesting the data fit almost too well to be pure chance, a claim still debated among historians) — established that phenotype counts obey binomial-style sampling, letting biologists use deviations from the expected ratio as evidence of new genetic phenomena, like linkage or lethal alleles, whenever real data deviated significantly more than binomial variance would allow.

The Luria–Delbrück fluctuation test. In 1943, Salvador Luria and Max Delbrück addressed whether bacterial resistance to viral infection (bacteriophage) arose through random, spontaneous mutation before exposure, or through directed adaptation in response to the phage itself. If resistance mutations occur randomly and independently in each bacterium at some low, constant per-generation probability, then in a colony grown to size n, the number of resistant survivors should follow discrete-count statistics related to the binomial (approaching Poisson-like behavior for rare events) with modest variance from culture to culture. If instead resistance were induced by the phage itself (a Lamarckian-style hypothesis), the number of resistant cells should be roughly constant across replicate cultures, reflecting a fixed induction rate applied to a fixed population, not a chance mutational event. Luria and Delbrück grew many independent small cultures of E. coli, then exposed each culture to phage and counted survivors. They found the variance in survivor counts across cultures was far larger than the mean — a huge "jackpot" effect where some cultures had zero resistant cells and others had hundreds — a pattern only possible if resistance mutations arose spontaneously at random times before phage exposure (a mutation early in a culture's growth gets amplified by later cell division, producing a jackpot culture; a mutation occurring late produces few resistant cells). This variance-to-mean signature is the opposite of what constant-probability, independent-trial (binomial/Poisson) induction would produce, and its detection is what proved mutations are spontaneous and pre-existing, not induced by selection — a foundational result for evolutionary biology, for which Luria and Delbrück shared the 1969 Nobel Prize in Physiology or Medicine (with Alfred Hershey). The experiment's logic depended entirely on knowing what variance a simple binomial/Poisson process should produce, so that the actual, much larger variance could be recognized as anomalous and mechanistically informative.

Example 1 — A genetics cross with an unusual ratio. A plant breeder crosses two heterozygous individuals (Aa × Aa) for a completely recessive trait. Standard Mendelian genetics predicts p = 0.25 for the recessive phenotype (aa) in each offspring, independently. The breeder examines n = 16 offspring and observes k = 7 with the recessive phenotype — far more than the expected 4. Is this consistent with simple recessive inheritance, or does it suggest something else (e.g., the trait is not fully recessive, or there's a genotyping error)? Expected count: np = 16 × 0.25 = 4 Variance: np(1−p) = 16 × 0.25 × 0.75 = 3 Standard deviation: √3 ≈ 1.73 Observing 7 is (7−4)/1.73 ≈ 1.73 standard deviations above the mean — a bit unusual, but not extreme. We can compute the exact binomial probability of observing 7 or more: P(X = 7) = C(16,7)(0.25)⁷(0.75)⁹ = 11,440 × 0.0000610 × 0.0751 ≈ 0.0524 Summing P(X ≥ 7) (adding k = 7 through 16) gives a one-tailed probability of roughly 0.079 — an 8% chance of seeing this many or more recessive offspring purely by chance, even if the 3:1 model is exactly correct. By the conventional 5% threshold, this result does not provide strong statistical evidence against simple Mendelian inheritance; the breeder should collect more data before concluding anything unusual (like partial dominance or a modifier gene) is happening. This illustrates why a formal chi-square or exact binomial test is essential — eyeballing "7 instead of 4" can look dramatic, but small-sample binomial variance is large enough to comfortably produce such a difference by chance alone.

Example 2 — A clinical drug-response scenario. A new antifungal drug is reported, from prior large trials, to clear infection in 70% of patients (p = 0.70). A small rural clinic treats n = 12 patients and observes only k = 5 successes — a 42% response rate, seemingly much worse than advertised. Should the clinic suspect the drug is not working as claimed, a batch problem, or patient non-compliance? Expected successes: np = 12 × 0.70 = 8.4 Variance: np(1−p) = 12 × 0.70 × 0.30 = 2.52 SD: √2.52 ≈ 1.587 We want P(X ≤ 5) under the hypothesis that p is truly 0.70: P(X=5) = C(12,5)(0.70)⁵(0.30)⁷ = 792 × 0.16807 × 0.0002187 ≈ 0.02913 P(X=4) = C(12,4)(0.70)⁴(0.30)⁸ = 495 × 0.2401 × 0.0000656 ≈ 0.00780 P(X=3) = C(12,3)(0.70)³(0.30)⁹ ≈ 220 × 0.343 × 0.0000197 ≈ 0.00149 P(X≤2) is smaller still, contributing under 0.0004 combined Summing: P(X ≤ 5) ≈ 0.0291 + 0.0078 + 0.0015 + 0.0004 ≈ 0.039 Only about a 3.9% chance of seeing 5 or fewer successes out of 12 if the drug truly works 70% of the time. This is below the conventional 5% threshold, so the clinic's disappointing result is statistically unusual enough to warrant investigation — perhaps a shipping/storage problem degraded the drug, the local patient population differs (resistant strain, poor compliance), or (with much lower likelihood, given a single result) the originally reported 70% efficacy doesn't generalize. Crucially, this conclusion required computing an exact binomial tail probability from a small sample — a normal approximation would be shaky here since n = 12 is small and p is not close to 0.5.

The binomial distribution's mathematical foundations were laid by Swiss mathematician Jacob Bernoulli, whose posthumously published Ars Conjectandi (1713) formalized the probability of a fixed number of successes in repeated independent trials — work now named the Bernoulli trial in his honor. Bernoulli was not a biologist; he was working out abstract problems in games of chance and demography. But his framework became indispensable to biology only once biologists began treating heredity as a probabilistic, particulate process rather than a blending phenomenon. That translation was made concrete by Gregor Mendel, an Augustinian friar working in Brno (in the modern Czech Republic) in the 1850s–60s, who bred pea plants (Pisum sativum) through multiple generations, meticulously counted phenotypes by the thousands, and inferred that inheritance follows discrete, independently segregating "factors" (later called genes) — implicitly treating each offspring's phenotype as an independent binomial trial with a fixed probability set by parental genotype. Mendel's work went largely unnoticed until its rediscovery around 1900 by Hugo de Vries, Carl Correns, and Erich von Tschermak, after which R. A. Fisher, a towering figure in 20th-century statistics and genetics, formalized the statistical machinery (chi-square goodness-of-fit tests, analysis of variance) that lets biologists rigorously compare observed binomial-type counts against Mendelian or other theoretical expectations — the very apparatus used in the worked examples above.

Clinical trials and drug approval. Regulatory agencies evaluate whether a new drug's response rate exceeds a placebo or standard-of-care rate by treating patient outcomes as binomial trials, computing exact or approximate binomial confidence intervals and p-values — exactly the logic in Worked Example 2, scaled up to hundreds or thousands of patients.

Genetic counseling and prenatal risk. For couples carrying known recessive alleles, genetic counselors use binomial probabilities to communicate risk — e.g., the chance that at least one of three future children will be affected, given a 25% per-pregnancy risk, computed as 1 − (0.75)³ ≈ 58%.

Conservation and wildlife survival studies. Mark-recapture and survival analyses often model the number of tagged animals surviving to a census date as binomial, with p representing an estimated survival probability — informing harvest quotas, endangered-species population viability, and reintroduction program success rates.

Genome-wide association studies (GWAS) and population genetics. Allele counts sampled from a population under Hardy-Weinberg equilibrium follow binomial (and, for two alleles across many loci, multinomial) sampling distributions, underlying statistical tests for whether an allele's frequency differs between case and control cohorts.

Quality control in agriculture and biotechnology. Seed germination rates, contamination rates in cell culture batches, and defect rates in manufactured biological products (like vaccine doses) are monitored using binomial-based control charts, flagging batches whose defect counts deviate implausibly from the expected rate.

Public health surveillance. Disease prevalence surveys sample a fixed number of individuals and count how many test positive, using the binomial to build confidence intervals around the true population prevalence and to detect outbreak clusters that exceed background binomial variance.

Ask the tutor: How do I compute a binomial confidence interval for an estimated proportion, and when should I use the normal approximation instead of the exact method?