Multinomial Distribution

What is the multinomial distribution?

The multinomial distribution models the counts of three or more mutually exclusive outcomes in a fixed number of independent trials, where the outcomes are discrete and finite. The probability of each of the k outcomes is constant from trial to trial.

What are some examples of the multinomial distribution?

Examples of data that could be modeled with the multinomial distribution include:

When should I consider using the multinomial distribution?

Use the multinomial distribution when you repeat the same type of trial n times and each trial can end in one of three or more outcomes. The trials must be independent, with the probability of each outcome remaining constant from trial to trial. You can use observed counts to estimate those probabilities. For more advanced analyses, you can also model counts or proportions using methods such as logistic regression, penalized regression, or generalized linear models.

What are the characteristics of the multinomial distribution?

The multinomial distribution extends the binomial distribution from two possible outcomes to three or more. For outcome $i$, $p_i$ is its probability and $n_i$ is its observed count. All probabilities add to 1, and all counts add to n.

Model parameters $p_i$ for each outcome $i$
Mass function

$p(X=\{n_1,n_2,\ldots,n_k\})=\frac{n!}{n_1\times n_2\times\cdots\times n_k}p_1^{n_1}p_2^{n_2}\cdots p_k^{n_k}$,

where the $n_i$ are nonnegative integers such that $\sum_{i=1}^{k} n_i = n$.

Mean $np_i$ for each outcome $i$
Variance $np_i(1$ - $p_i)$ for each outcome $i$

Example of a multinomial random variable

Suppose a list of blood donors has the following percentages of different blood types:

O+ O- A+ A- B+ B- AB+ AB-
38.4% 6.5% 32.0% 5.6% 11.8% 1.5% 3.7% 0.5%

You randomly call 20 donors. What is the probability of reaching at least five O+ donors and at least one O- donor? Assume each call is independent and the percentages stay the same.

Only the O+ and O- counts matter for this question, so you can combine all the remaining blood types into an “Other” group. The three probabilities are 0.384, 0.065, and 0.551:

O+ O- Other
38.4% 6.5% 55.1%

Let X represent the three group counts. We want all outcomes in which the O+ count is at least 5, the O- count is at least 1, and the three counts total 20.

For each allowed combination of counts, calculate its multinomial probability, then add those probabilities:

$\begin{aligned}P(O+ \geq 5 \,\&\, O- \geq 1)&=\sum_{x_1=5}^{20}\sum_{x_2=1}^{20-x_1}\frac{20!}{x_1!\,x_2!\,(20-x_1-x_2)!}(0.384)^{x_1}(0.065)^{x_2}(0.551)^{20-x_1-x_2}\end{aligned}$

$=0.682$

The sum is approximately 0.682. Therefore, the probability of reaching at least five O+ donors and at least one O- donor is about 68.2%.

This example shows a useful feature of the multinomial distribution: outcomes that are not central to the question can often be combined into a single group by adding their probabilities together.