Statistics & inference
The statistical framework used to quantify measurements, discoveries and limits.
Neutrino analyses use maximum-likelihood fits with nuisance parameters, frequentist confidence intervals (often Feldman–Cousins) and Bayesian credible intervals, together with conventions such as 5σ for discovery.
In plain words
A neutrino detector does not hand out answers; it hands out a list of events, many of them background, each measured imperfectly. Statistics is the toolkit that turns such a list into a statement like 'θ₁₃ is not zero' or 'the neutrino mass is below 0.45 eV'. Because neutrino data are often sparse and the effects subtle, the field has shaped several of the methods now used across particle physics.
The building blocks
- Fitting: a maximum-likelihood fit finds the parameter values under which the observed data are most probable. Many experiments combined form a global fit.
- Intervals: a confidence level states how often a procedure covers the true value. When nothing is seen, the result is an upper limit, and the Feldman–Cousins method handles the awkward case of small signals near a physical boundary.
- Discovery: statistical significance, expressed in σ, translates a p-value into a familiar scale; the look-elsewhere effect corrects for searching in many places.
- Uncertainties: statistical uncertainty comes from limited event counts, systematic uncertainty from imperfect knowledge of the detector, flux and physics.
- Detector performance: energy resolution limits how sharply spectral features can be seen.
- Planning and integrity: sensitivity predicts what an experiment should reach, blind analysis protects against wishful thinking, and Monte Carlo simulation supplies the predictions every step relies on.
Frequentist and Bayesian views
Most accelerator, reactor and double beta decay results are frequentist: they describe how a procedure would behave in many imagined repetitions of the experiment. Cosmological bounds on the sum of neutrino masses are usually Bayesian credible intervals, which depend on a chosen prior. Both are legitimate, but numbers from the two schools are not directly interchangeable, and careful papers often show both.
Conventions worth knowing
- 3σ is called 'evidence', 5σ 'observation' or discovery.
- Limits are commonly quoted at 90% CL in neutrino physics and at 95% CL in collider and cosmology papers.
- Oscillation contours are drawn at 1σ, 90% and 3σ for one or two parameters, with different Δχ² thresholds for each.
Why it matters
Several famous episodes in neutrino physics turned on statistics: long-standing 2–4σ anomalies such as LSND and MiniBooNE, the superluminal OPERA claim that turned out to be a hardware fault, and the slow climb of the mass ordering and CP violation signals in global data. Reading neutrino results critically means reading their statistical statements. The standard reference is the statistics review of the Particle Data Group.
Frequently asked questions
Why do neutrino physicists insist on 5σ?
Because many 3σ effects have disappeared with more data or turned out to be systematic. A high threshold, combined with independent confirmation, keeps the rate of false discoveries low.
Which statistical approach is correct, frequentist or Bayesian?
Both answer different questions. Frequentist intervals describe the long-run behaviour of a method; Bayesian intervals describe degrees of belief given a prior. Results are most convincing when both agree.
Go deeper
- Confidence level (CL)The probability that a procedure's interval covers the true value in repeated experiments, such as 90% or 95%.
- Statistical significance (σ)How unlikely an observed excess would be from background alone, expressed in Gaussian standard deviations.
- Statistical uncertaintyThe uncertainty from the random fluctuation of event counts, roughly √N for N events.
- Systematic uncertaintyUncertainty from imperfect knowledge of the detector, the source or the theory, which does not shrink with more data.
- Energy resolutionHow precisely a detector measures energy, usually quoted as a fractional width such as 3%/√E(MeV).
- SensitivityThe median result an experiment expects in the absence (or presence) of a signal, used for design and comparison.
- Blind analysisHiding the signal region or key results until the analysis procedure is finalised, to prevent unconscious bias.
- Maximum-likelihood fitFinding the parameter values that make the observed data most probable under a model.
- Monte Carlo simulation & event generatorsComputer simulation of neutrino fluxes, interactions and detector response, used to predict and interpret data.
Mentioned in 2 pages
People 1
Terms 1
Last updated: · Editorial standard
Cite this page
Text
Neutrino Atlas (2026). “Statistics & inference”. Retrieved 22 November 2026, from https://neutrino-atlas.com/terms/statistics-methods/BibTeX
@misc{neutrinoatlas2026termsstatisticsmethods,
title = {Statistics & inference},
author = {{Neutrino Atlas}},
year = {2026},
url = {https://neutrino-atlas.com/terms/statistics-methods/},
note = {Accessed 2026-11-22},
language = {english}
}