Suppose you see a dark cloud and wonder whether it will rain. A likelihood asks how probable that observation would be under each possibility. It does not, by itself, tell you the probability of rain after seeing the cloud. section 2.4 section 2.2
For observed data D and a hypothesis H, the likelihood is Pr(D | H): the probability of the data assuming H. A hypothesis that makes the observation more probable has a higher likelihood. Keep the direction straight: Pr(D | H) is not Pr(H | D). section 2.2 section 11.7
Three tea drinkers out of five
Let p be a proposed proportion of people who prefer tea. If three of five surveyed people prefer tea, the likelihood is 10 × p³ × (1 − p)². The tea preferences contribute p³, the coffee preferences contribute (1 − p)², and 10 counts the possible arrangements of three tea preferences among five people. At p = 0.6, the likelihood is 0.3456; at p = 0.2, it is 0.0512. The observed count fits the first proposal better. section 2.10 section 3.4
binom_coeff = comb(n, k) #A
likelihood = (
binom_coeff
* (ps ** k)
* ((1 - ps) ** (n - k))
)
The NumPy calculation evaluates the same likelihood for every proposed value in ps; comb supplies the arrangement count. section 2.11
To update beliefs, multiply each hypothesis’s likelihood by its prior probability, then divide each product by the sum of the products. The resulting posterior probabilities sum to 1. Thus, a likelihood compares how well hypotheses explain the data; it is not already an updated probability for each hypothesis. A zero likelihood rules a hypothesis out in this update. section 2.2 section 2.6
A positive test is not a diagnosis
For a rare disease affecting about 1 in 150 people, the likelihood of a positive test is 0.98 if the person has the disease and 0.05 if they do not. Accounting for those likelihoods and the disease’s rarity gives about a 12% posterior probability of disease after a positive result—not 98%. section 2.5
The form of a likelihood depends on the model. Tea-preference counts use a binomial likelihood; a house-price regression with normally distributed noise uses a normal likelihood. Likelihoods also support model comparison through a Bayes factor, a ratio of how likely the observed data are under competing models. Very small likelihoods can be more stable to calculate on a logarithmic scale. section 3.4 section 3.13 section 11.5






