Confidence level (CL)
The probability that a procedure's interval covers the true value in repeated experiments, such as 90% or 95%.
In frequentist statistics a 90% CL upper limit means the method would exclude the true value in only 10% of identical experiments. Neutrino mass and 0νββ limits are often quoted at 90% CL, oscillation contours at 1σ, 90% or 3σ. Bayesian analyses use credible intervals instead.
In plain words
Suppose a hundred teams each ran the same experiment and each used the same recipe to draw an interval around their result. If the recipe is a 90% confidence-level procedure, about ninety of those intervals would contain the true value and about ten would miss it. The confidence level describes the reliability of the recipe, not the probability that one particular interval is right.
In detail
The frequentist construction, due to Jerzy Neyman in the 1930s, builds for every possible true value the range of outcomes that would occur with the chosen probability, then inverts it: the confidence interval contains all true values whose ranges include the observed result. A procedure that achieves the stated rate is said to have correct coverage. For a Gaussian measurement the familiar correspondences are:
- 1σ: 68.3% CL
- 90% CL: 1.64σ two-sided
- 95% CL: 1.96σ two-sided
- 3σ: 99.73% CL
In likelihood fits the same levels translate into thresholds on −2Δln L, for example 2.71 for one parameter at 90% CL and 4.61 for two parameters, which is why oscillation contours in two-dimensional plots are wider than one-dimensional error bars.
Credible intervals
Bayesian analyses instead report credible intervals: the range that contains a given fraction of the posterior probability, which depends on the prior. Cosmological limits on the sum of neutrino masses are typically quoted this way at 95%. Near a physical boundary, such as a mass that cannot be negative, the two approaches can give noticeably different numbers.
Conventions in neutrino physics
- Mass and half-life limits: usually 90% CL, for example KATRIN and the double beta decay experiments.
- Oscillation parameters: best fit with 1σ errors, plus 90% and 3σ contours.
- When the result is near zero or a boundary, many experiments use the Feldman–Cousins construction to keep coverage correct. If nothing is seen, the interval becomes an upper limit.
Frequently asked questions
Does a 90% CL interval mean there is a 90% chance the true value is inside?
Not in the frequentist sense. The true value is fixed; it is the interval that varies between experiments. The 90% refers to how often the method succeeds. A Bayesian credible interval does have the probability interpretation, but only given its prior.
Why do some papers use 90% and others 95%?
It is a convention of each community. Neutrino and nuclear physics traditionally use 90%, collider physics and cosmology 95%. Comparisons must account for the difference.
Original papers
- TheoryJ. Neyman, Outline of a Theory of Statistical Estimation Based on the Classical Theory of Probability, Phil. Trans. R. Soc. Lond. A 236, 333 (1937)
Go deeper
- Upper limit / exclusionThe largest value of a quantity consistent with the data at a given confidence level when no signal is seen.
- Feldman–Cousins methodA frequentist method for building confidence intervals that moves smoothly from upper limits to two-sided intervals.
Related terms
Mentioned in 20 pages
People 6
Last updated: · Editorial standard
Cite this page
Text
Neutrino Atlas (2026). “Confidence level (CL)”. Retrieved 22 November 2026, from https://neutrino-atlas.com/terms/confidence-level/BibTeX
@misc{neutrinoatlas2026termsconfidencelevel,
title = {Confidence level (CL)},
author = {{Neutrino Atlas}},
year = {2026},
url = {https://neutrino-atlas.com/terms/confidence-level/},
note = {Accessed 2026-11-22},
language = {english}
}