LoginSign Up
← Back to all sparks
P

profileCI

INFRA · APIS
Velocity0.0

Profiling a Log-Likelihood to Calculate Confidence Intervals

github.com/paulnorthrop/profileci ↗

Profile-likelihood confidence intervals for any fitted model, in a feed that publishes out of order.

profile-likelihoodconfidence-intervalsstatisticsnumerical-robustnessr-package
◆Current state
profileCI computes confidence intervals from the profile log-likelihood for user-supplied fitted models, generalising what confint.glm does for GLMs to any model object exposing a log-likelihood. The releases handle the awkward cases that make profiling fail in practice: infinite limits when the profile never drops below the interval threshold, bounded profiling ranges, and interpolation that breaks down near the limits. Only convex log-likelihoods are supported, so disjoint intervals are out of scope by design.
◆Where it's heading
Work is concentrated on numerical reliability rather than scope: 1.1.1 replaced quadratic with monotonic cubic spline interpolation because the quadratic form could fail, and corrected parameter values stored near the confidence limits. The feed publishes these out of order, with the v1.0.0 entry stamped six months after v1.1.0 and carrying the package's full description rather than a changelog, so release order should be read from the version numbers rather than the dates. The same maintainer's revdbayes has been in pure maintenance across this period, which places profileCI as the more active project.
◆Prediction
Expect further robustness work at the profiling limits and more logLikFn methods for common model classes, following the nls method added in 1.1.0.

◆Recent moves

  1. 6mo ago

    Cubic spline interpolation replaces a quadratic that could fail

    Switches to monotonic cubic spline interpolation for negative epsilon, where the quadratic form could fail outright, and fixes slightly incorrect parameter values stored near the confidence limits. Reliability work at exactly the point where profiling is hardest.

    View source ↗
  2. 7mo ago

    CRAN 1.0.0 release of profile-likelihood interval computation

    The 1.0.0 entry, whose body is the package description rather than a changelog. Note the feed stamps it six months after v1.1.0, so the dates here do not reflect release order; read the version numbers instead.

    View source ↗
  3. 1y ago

    Infinite limits and bounded profiling ranges handled

    Adds the flat argument for profiles that never drop below the interval threshold, lb and ub to bound the profiling range, and a logLikFn method for nls objects. The release that made the package usable on models where profiling does not terminate cleanly.

    View source ↗