Simple Accurate Approximation of Likelihood Profiles

Likelihood profiles for arbitrary functions of the model parameters are useful for constructing likelihood ratio confidence intervals, diagnosing linear approximation intervals, suggesting linearizing transforms, and many other purposes. This article investigates a simple integration method for producing accurate approximations to likelihood profiles that avoid problems associated with producing exact profiles. A basic theorem guarantees that the approximation can achieve any desired precision. In many cases the method requires no more than adding a few lines of code to that required to produce the maximum likelihood estimate of the parameter vector. Standard methods for computing likelihood profiles are based on solving a sequence of constrained maximum likelihood problems. Each problem generates a point on the profile. An earlier article proposed an integration method that generates the entire profile directly by solving a differential equation. That method, however, requires the Hessian of the log-likelihood which can be difficult to produce. The method considered here requires only the gradient. It may be particularly useful in conjunction with the EM algorithm which typically does not produce any type of confidence interval.