How many misses before the third six? The negative binomial knows
Roll a die and count every non-six until the third six appears. That tally follows the negative binomial distribution, a workhorse of statistics that also describes why a few infected people can spread disease far more than others. Its odd name comes from a trick with negative numbers.
The setup is a run of independent trials, each ending in success or failure with the same probability of success. You keep going until a fixed number of successes, called r, has arrived, and the thing you count is how many failures happened along the way. Some versions flip this around and count total trials instead, so the model might describe how many days a machine keeps running before a fixed breakdown point. It is also known as the Pascal distribution, and different textbooks count slightly different quantities, so definitions need checking.
The formula comes from a simple observation. Any particular sequence of r successes and k failures has probability p to the power r times one minus p to the power k. The final trial must be a success, so only the first k plus r minus one positions are free, and the question becomes how many ways the k failures can be placed among them. Another way to see it treats the sequence as r counts of failures between successes, a stars and bars problem. Rewriting that counting coefficient with negative numbers gives a tidier expression, which is where the term negative binomial comes from.
Its real power lies in spread. The Poisson distribution, commonly used for counts, forces the variance to equal the mean. The negative binomial allows the variance to be larger, and it collapses into the Poisson when failures become very rare. That makes it a natural choice for overdispersed data, where events cluster because one occurrence makes another more likely, and it underpins a robust variant of Poisson regression.
Epidemiologists lean on it to model transmission when the number of people each patient infects varies widely between individuals and settings, capturing the fact that a handful of cases can drive most onward spread.
Source: Negative binomial distribution
