Assessing a Fit
A converged fit is not yet a trustworthy one. Three checks, in increasing strength, tell you whether to believe a result — and they answer different questions. This page is the entry point that ties them together; each step links to where it is documented and driven in full.
1. Is the χ² in its band?
For independent Gaussian observations with fixed, correctly stated errors and a linear model, the minimized \(\chi^2\) has \(\nu=N-N_\mathrm{free}\) degrees of freedom. Nonlinear or bounded fits use this law approximately; the tooltip states “χ² laws are approximate for nonlinear or bounded fits.” Asymmetry reports a two-sided band at confidence \(R\) (default 0.999):
good — the reduced statistic lies inside the band;
poor — it lies in the upper tail;
overdone — it lies in the lower tail, often indicating overestimated errors or excess model flexibility.
A Poisson fit reports Poisson deviance/ν, not a Gaussian χ² statistic. Its likelihood-ratio distribution approaches a χ² law at large expected counts [pdg_statistics]. As a display policy, a χ² quality verdict is shown only when every fitted expected bin count is at least five. This threshold does not guarantee an exact χ² law. For coupled fits, the minimum includes all members of the minimized objective. With a sparse fitted bin, the tooltip says “Poisson deviance has no χ² quality verdict when any fitted expected count is below 5.” If expected counts are unknown, it says “Poisson expected counts were not recorded; no χ² quality verdict is available.” The deviance and convergence status remain visible. Fit ✓ denotes convergence, independently of whether a quality verdict is available.
See Fit statistics for residual checks. Unit weights or scatter-estimated errors do not supply an independent Gaussian quality test; see Parameter trending. Coherent residual structure still signals a missing component even when the scalar statistic looks acceptable.
2. Is the objective recorded and consistent across the series?
Fit summaries record the objective (gaussian or poisson), total and
reduced statistic, fitted parameters and uncertainties. Poisson summaries also
record the minimum fitted expected count needed by the display policy. A
missing quality verdict carries a quality_reason. Run-batch and grouped
series use the same summary format; comparisons require the same likelihood
and data. A mixed-objective series is labelled “Mixed fit objectives.” rather
than assigned one averaged statistic.
An unknown saved objective is explicitly null: it has no χ² verdict and
is labelled “Fit objective was not recorded.” Known Gaussian producers retain
their objective when loaded. A likelihood cannot be inferred from a numeric
field named chi_squared alone. See Project files for saved results.
3. What does a conditional pull experiment show?
The pull diagnostic simulates ideal independent Poisson detector counts at fitted plug-in values, refits each copy, and forms \((\hat\theta-\theta_{\mathrm{gen}})/s_{\hat\theta}\). Its mean standardised residual and sample width describe the common usable subset. They do not establish estimator unbiasedness, interval coverage, model correctness, or uncertainty from omitted acquisition effects. Failures, selection, boundaries, and skewness matter even when a width is near one. The display retains counts, reasons, seed ranges, assumptions, and explicit partial-stop status. Normal reference SEs assume independent standard-normal pulls. See the PDG Statistics review (2025), Secs. 40.2 and 40.4.
ALC fits that subtract an estimated baseline have two uncertainty levels.
Conditional errors hold that baseline fixed. Baseline-propagated errors include
its fitted uncertainty and its correlation with the scan points; Fit peaks
and integral-scan --baseline report these propagated errors. Their validity
still requires a suitable baseline model and non-resonant fitting regions.
See Integral scan mode (avoided-level-crossing field scans) for the covariance construction and its first-order limits.
Before the full pull diagnostic, a quick read on a single parameter’s error is its MINOS asymmetric interval: when the upper and lower excursions differ markedly, the symmetric Hessian error is already suspect — the likelihood is skewed or the parameter is against a bound — and the pull diagnostic is worth running.
Use the three together: the χ² band is a fast triage, the recorded verdict tracks quality across a series, and the pull diagnostic explores error behaviour under its stated simulated conditions.
See also
Fit statistics — reading χ², reduced χ², and residuals.
Parameter trending — the recorded per-fit summary and how it feeds trending, including the χ² quality band’s confidence level.
Conditional simulation: the pull distribution — conditional simulation, failed refits, and pull summaries.
Fit wizard — choosing a starting model when the χ² says the current one is wrong.
Model-selection criteria
Information-criterion comparisons require the same data and likelihood. AICc is unavailable when its sample-size correction is undefined; a smaller AIC value cannot replace it in an AICc ranking. See Fit wizard for formulas, availability and assumptions, and Global fit wizard for series totals and partition BIC.
G. Cowan and L. Lista, “Statistics”, Particle Data Group (2025), sections 40.2 and 40.3.3.1, Statistics review.
Saved-fit information-criterion comparisons require both the same recorded data selection and the same recorded objective. Gaussian and Poisson fits rank in separate groups; a fit without a recorded objective receives no comparison delta or weight.