loglikelihood

Loglikelihood, or log-likelihood, is the natural logarithm of the likelihood function used in statistics to quantify how probable a set of observed data is under a statistical model with fixed parameters. For data x = (x1, ..., xn) and a model with parameter θ, the likelihood is L(θ) = p(x1, ..., xn | θ). If observations are independent, L(θ) = ∏i p(xi | θ), and the log-likelihood is l(θ) = log L(θ) = ∑i log p(xi | θ). The log transformation turns products into sums, aiding numerical stability and ease of differentiation.

Maximizing the log-likelihood with respect to θ yields the maximum likelihood estimates (MLEs) of the parameters. Because l(θ) is often easier to differentiate than L(θ), optimization is typically performed on the log-likelihood. Under standard regularity conditions, the MLE is consistent and asymptotically normal, with variance approximated by the inverse of the Fisher information.

A common example is data from a normal distribution with unknown mean μ and known variance σ^2. The log-likelihood is l(μ) = -n/2 log(2πσ^2) - (1/(2σ^2)) ∑ (xi - μ)^2, whose maximizer is μ̂ = x̄. When variance is also unknown, the log-likelihood includes an additional term for σ^2.

Log-likelihood is central to model comparison and hypothesis testing. The likelihood ratio test uses the statistic -2 log(λ), where λ is the ratio of restricted to unrestricted likelihoods, which under regularity conditions follows a chi-squared distribution. Information criteria such as AIC and BIC use -2 times the log-likelihood plus penalties to compare models.

Notes: the log-likelihood depends on the assumed model; misspecification can bias results. For complex or non-iid data, numerical optimization and profile likelihood methods are commonly employed.