Home › Stats-Regression › Generalized linear models

GlmOptions

What a GeneralizedLinearModel fit may be told.

public sealed record GlmOptions

Properties — WithIntercept prepends a column of ones; true by default. ConfidenceLevel is the two-sided level the intervals are reported at; 0.95 by default. MaximumIterations is how many IRLS iterations are allowed; 100 by default, which is the reference's own budget. Tolerance is the absolute bound on the change in deviance between iterations — the reference's atol with its rtol left at zero; 1e-8 by default. ThrowOnNonConvergence says whether a fit that did not converge throws instead of returning; true by default. NegativeBinomialAlpha is the dispersion α of GlmFamily.NegativeBinomial, given rather than estimated; null by default, which fits the reference's own default of 1. Link is the GlmLink the mean is fitted through; Default by default, each family's statsmodels default.

Exceptions — ArgumentOutOfRangeException when ConfidenceLevel does not lie strictly inside (0, 1), when MaximumIterations is below one, when Tolerance is not above zero, when NegativeBinomialAlpha is not finite and above zero, or when Link is not a declared GlmLink. Each is thrown where the setting is set, not where the fit reads it: a budget of zero would otherwise skip the IRLS loop entirely and reach the caller as a table of 0/0.

Example — a wider level widens both ends without moving the estimate, the same way it does for OlsOptions.

using Lodestar.Stats.Regression;

double[] design = [0.0, 1.0, 2.0, 3.0, 4.0, 5.0];
double[] response = [0.0, 0.0, 1.0, 0.0, 1.0, 1.0];

GlmSummary ninetyFive = GeneralizedLinearModel.Fit(design, response, 1, GlmFamily.Binomial);
GlmSummary ninetyNine = GeneralizedLinearModel.Fit(
    design, response, 1, GlmFamily.Binomial, new GlmOptions { ConfidenceLevel = 0.99 });

double narrow = ninetyFive.ConfidenceLower[1];  // => -0.57460583007824…
double wide = ninetyNine.ConfidenceLower[1];    // => -1.13663518264790…

Remarks — turning ThrowOnNonConvergence off does not fix a bad fit; it lets you inspect one. A non-converged inference table is plausible and wrong — enormous standard errors and p-values that read like p-values. Set it to false to freeze a non-convergent case in a corpus, or to look at one, and read GlmSummary.Converged before anything else in the table it returns.

NegativeBinomialAlpha belongs to one family. Set for Binomial or Poisson, GeneralizedLinearModel.Fit refuses it rather than ignoring it. Left unset for NegativeBinomial, the fit uses 1 — the value statsmodels falls back to with a warning, stated here because a library has no warning channel a caller reads.

Applies to — net10.0, netstandard2.0.

See also — GeneralizedLinearModel.Fit, GlmSummary.