Can likelihood exceed one?

A likelihood based on continuous densities can, depending on units and model scale.

Does maximum likelihood select the true parameter?

It selects a parameter maximizing the observed-data criterion; accuracy depends on data and assumptions.

Is likelihood the same as probability in everyday language?

Everyday usage often overlaps, but statistical likelihood has the specific fixed-data, varying-parameter meaning described here.

AI Glossary / Statistics and probability / Intermediate

Likelihood

Definition

Likelihood is the support a statistical model assigns to the observed data as its parameters or specified hypotheses vary. For fixed observations, it is evaluated using the model's probability mass or density, viewed as a function of the unknown parameter. Likelihood is central to estimation, model comparison and Bayesian updating. It is not, by itself, the probability that a parameter value or hypothesis is true. The distinction matters because probability of data under a hypothesis and probability of a hypothesis after seeing data are different quantities.

What it means

In a probability calculation, a model is fixed and possible data outcomes vary. In a likelihood calculation, the data have been observed and candidate model parameters vary. The same mathematical expression can be used in these two directions, but its interpretation changes. A likelihood ratio compares how strongly two specified candidates support the same observed data. Its value depends on the assumed sampling model, including independence, measurement and selection. A high relative likelihood cannot rescue a model that ignores important features of how the data were generated.

How it works

Specify the observed data and a probability model for their collection. Evaluate the joint probability mass or density at candidate parameter values, retaining any parameter-dependent factors. For independent observations, the joint expression often factors into a product; with dependence, the model must represent that dependence. Ratios or log-likelihood differences compare candidates. Maximization gives a maximum likelihood estimate, while Bayesian analysis combines likelihood with a prior and normalizes the result. Absolute likelihood magnitudes are usually less interpretable than comparisons made for the same data representation and model framework.

Three practical examples

Parameter estimation: compare candidate success probabilities against observed binary outcomes.
Reliability modeling: evaluate lifetime-model parameters while including appropriate contributions from censored units.
Machine learning: train a probabilistic prediction model by maximizing likelihood or minimizing negative log-likelihood. Good training likelihood alone does not establish calibration, generalization or suitability for a consequential decision.

Think of it this way

Think of several candidate explanations each assigning a score to the observations you actually obtained. Likelihood compares those scores while keeping the evidence fixed. It does not automatically assign odds to the explanations themselves. To obtain posterior odds, a Bayesian calculation also needs prior odds and a clearly defined set of hypotheses.

Worked example

Observe six successes and two failures in eight independent trials with a common success probability p. Ignoring a factor that is identical for all p, likelihood is p to the sixth power times (1-p) squared. Comparing p = 0.75 with p = 0.50 gives a likelihood ratio of 729/256, approximately 2.8477, in favor of 0.75. This means the observed pattern has about 2.85 times the probability under the first candidate, not that the first candidate has probability 2.85 or is 2.85 times as probable without further assumptions. In a deliberately restricted Bayesian comparison containing only these two hypotheses, prior odds of 1 to 3 become posterior odds 729 to 768. The posterior probability of the 0.75 hypothesis is then 243/499, approximately 0.487. The likelihood favors it, yet the stated prior leaves its posterior probability below one-half.

Illustrative case study

In a fictional case, Maple Experiments compares two prespecified response-rate models for a pilot. A dashboard labels the likelihood ratio as the probability that the new process works. Analysts correct the label and report what was actually compared, including the trial assumptions. They also show uncertainty over a wider range of possible rates instead of treating the two chosen candidates as exhaustive truth. The corrected report separates evidence from prior assumptions and from the practical effect size needed for a rollout decision.

Common misconceptions and mistakes

Likelihood is not a posterior probability, a p-value or the probability of future success. For continuous observations, likelihood uses densities, so it need not be bounded by one and does not equal the probability of an exact point. Multiplying by a positive factor independent of parameters leaves relative likelihoods unchanged. Normalizing likelihood over a parameter range is not automatically a uniquely justified posterior. Raw likelihoods from different datasets or incompatible measurement representations should not be compared as though they were on a universal evidence scale.

Business use

Businesses should ask what data, model and alternatives lie behind any likelihood-based claim. Use relative evidence with effect sizes, uncertainty and validation appropriate to the decision. For predictive systems, evaluate held-out log scores and calibration where relevant, and check whether dependence or selective sampling has been ignored. When communicating a Bayesian conclusion, state the prior assumptions separately. A numerical evidence ratio can clarify a comparison, but it does not determine business utility, causality or the cost of being wrong.

Check your understanding

Why can the example favor p = 0.75 by likelihood while its posterior probability remains below 0.5?

Answer: The stipulated prior odds favor p = 0.50 by three to one. Multiplying those odds by the likelihood ratio does not quite reverse them. Posterior probability combines the evidence with the prior within the stated two-hypothesis model.

Related terms

Likelihood function
Log likelihood
Posterior probability
likelihood ratio
maximum likelihood estimation
prior odds
probability density
negative log-likelihood

Learn next

Learn how likelihood ratios update prior odds, then compare likelihood-based intervals with confidence and credible intervals. Study how dependence, censoring and selection affect the observed-data likelihood before applying it to complex models.

Sources

NIST/SEMATECH e-Handbook, Maximum likelihood estimation: https://www.itl.nist.gov/div898/handbook/apr/section4/apr412.htm . Cousins, What is the likelihood function, and how is it used in particle physics?: https://arxiv.org/abs/2010.00356 . The binary-trial comparison and fictional case above are original.

Previous: Lightweight ontology Next: Likelihood function