Geometric vs Negative Binomial Distribution: When to Use Each

The geometric distribution models the number of independent Bernoulli trials needed to obtain the FIRST success. The negative binomial models the number of trials (or failures) needed to obtain the R-TH success. Both share the underlying Bernoulli framework with constant success probability p across independent trials. The geometric is a special case of the negative binomial with r=1. Use geometric when the question is "how long until the first arrival of an event"; use negative binomial when "how long until the third (or fifth, or r-th) arrival." Both distributions are discrete, defined on positive integers (or non-negative integers depending on convention), and both have right-skewed shapes that approach symmetry only when both r and p are large.

There are two conventions and you must check which one any textbook or software uses. Convention A: X = number of TRIALS until the first success, X ∈ {1, 2, 3, ...}. P(X = k) = (1-p)^(k-1) × p. Mean = 1/p. Variance = (1-p)/p². Convention B: X = number of FAILURES before the first success, X ∈ {0, 1, 2, ...}. P(X = k) = (1-p)^k × p. Mean = (1-p)/p. Variance = (1-p)/p². Both produce identical answers to "what is the probability the first success occurs on trial 4" if you map the conventions consistently, but mean and variance differ by 1. R’s pgeom uses convention B, scipy.stats.geom uses convention A. Always check.

Convention A: X = number of trials to get the r-th success, X ∈ {r, r+1, ...}. P(X = k) = C(k-1, r-1) × p^r × (1-p)^(k-r). Mean = r/p. Variance = r(1-p)/p². Convention B: X = number of failures before the r-th success, X ∈ {0, 1, 2, ...}. P(X = k) = C(k+r-1, r-1) × p^r × (1-p)^k. Mean = r(1-p)/p. Variance = r(1-p)/p². With r=1 the negative binomial reduces to the geometric. The variance grows linearly with r, which makes the negative binomial increasingly diffuse as r grows. There is also an alternative "extended" negative binomial parameterization (with shape and dispersion) used in count regression for overdispersed Poisson data — that is the same distribution algebraically but parameterized differently.

A player makes 65% of free throws (p=0.65). What is the probability the first make occurs on the third attempt? Using convention A: P(X=3) = (1-0.65)² × 0.65 = 0.35² × 0.65 = 0.1225 × 0.65 = 0.0796. About 8%. What is the probability of needing MORE than 5 attempts to get the first make? P(X > 5) = (1-p)^5 = 0.35^5 = 0.00525. About 0.5% — the player will almost certainly make at least one in five attempts. Expected number of attempts = 1/p = 1/0.65 ≈ 1.54 attempts. So on average the first make comes on the second attempt.

A manufacturing line produces units with defect probability p=0.02. You want to know the distribution of the number of inspections needed to find 5 defective units. Negative binomial with r=5, p=0.02. Expected number of inspections = r/p = 5/0.02 = 250. Variance = r(1-p)/p² = 5 × 0.98 / 0.0004 = 12,250. Standard deviation = sqrt(12,250) ≈ 110.7. So you expect to inspect about 250 units before finding 5 defects, with one standard deviation of about 111 units of uncertainty. What is the probability of finding the 5th defective unit on exactly the 200th inspection? P(X=200) = C(199, 4) × 0.02^5 × 0.98^195. C(199, 4) = 64,684,950; 0.02^5 = 3.2×10^-9; 0.98^195 ≈ 0.0196. Product ≈ 0.004, about 0.4%. The narrow probability on any single specific value reflects how spread out the distribution is when r is small and p is small.

Binomial asks "in n fixed trials, how many successes did I get?" Negative binomial asks "to get r successes, how many trials did it take?" The roles of count and trial number are swapped. Both share the Bernoulli setup of independent trials with constant p. The binomial has fixed n and random count; the negative binomial has fixed count r and random trial number. This duality means problems can sometimes be solved by either depending on what is fixed and what is random.

Confusion 1: which convention is in use? Always confirm whether your formula or software counts trials or failures. Confusion 2: independence and constant p. If trials are not independent (sampling without replacement) or p changes (learning curve, fatigue), the geometric/negative binomial assumptions fail and you may need the hypergeometric or other alternatives. Confusion 3: confusing "first success" (geometric) with "exactly one success in n trials" (binomial). These are distinct events with different probabilities. Confusion 4: in count regression, "negative binomial" is parameterized by mean μ and dispersion k, and the link to the trials/successes formulation can be obscured. The distribution is the same algebraically.

Snap a photo of any waiting-for-success problem and StatsIQ identifies whether the geometric or negative binomial applies, picks the convention that matches the question (trials vs failures), computes the requested probability or moment, and walks through the C(k-1, r-1) × p^r × (1-p)^(k-r) calculation step by step. The app also flags when independence or constant-p assumptions look implausible.