Deviance (statistics)

A likelihood-based goodness-of-fit quantity comparing a fitted statistical model with a saturated model, conventionally twice their maximized log-likelihood difference.

Version

v2 · 2026-09-19 · History

Domain-specific #

4136

Domain group

Formal Sciences

Origin domain

Experimental Design & Statistics

Subdomain

Statistical Modeling and Generalized Linear Models → Experimental Design & Statistics

Core Idea

Deviance generalizes residual sums of squares to likelihood models and supports residual diagnostics, nested-model comparisons, dispersion estimation, and asymptotic tests only under declared regularity and scaling conventions. The fitted model and saturated benchmark are optimized on the same observations; their log likelihoods are differenced and scaled, and nested changes or unit contributions attribute lack of fit under model-specific distribution assumptions. The abstraction is therefore identified by a declared carrier, a transformation or constraint over that carrier, and an invariant that tells an analyst whether the named structure is genuinely present.

How would you explain it like I'm…

How Far From Perfect

Imagine drawing a line through a bunch of dots. A 'cheater' drawing could zig-zag through every single dot exactly. Deviance is a number that tells how much worse your drawing fits the dots than the cheater drawing does. A small number means your drawing fits almost as well as the perfect cheat.

Fit Compared to a Perfect Copy

When statisticians build a model to explain data, they want to know how well it fits. Deviance compares their model to a 'saturated' model, a model so flexible it matches every data point exactly. It uses likelihood, a measure of how well a model explains the data, and looks at how much worse the real model does than the perfect-fit one. Smaller deviance means a better fit. It's a lot like adding up how far the points are from a line, but it works for many more kinds of data, like counts or yes/no answers.

Likelihood-Based Lack of Fit

Deviance is a measure of lack of fit for statistical models based on likelihood, used especially with generalized linear models. It compares the fitted model to a saturated model, which has enough parameters to fit every observation exactly, on the same data: deviance is twice the difference in their log-likelihoods, possibly scaled by a dispersion parameter. For ordinary linear regression with normal errors, it essentially reduces to the familiar residual sum of squares, which is why it's described as a generalization of that idea. Deviance can be broken into contributions from each observation (deviance residuals), used to compare nested models by the change in deviance, and used to estimate dispersion. The approximate statistical tests built on it only hold under stated assumptions about the distribution, scaling and sample size.

 

Deviance generalizes the residual sum of squares to likelihood-based models. For a given response distribution, data and weights, the fitted model and the saturated model (one parameter per observation) are both maximized on the same observations; the deviance is twice the difference of their maximized log-likelihoods, and the scaled deviance divides by the dispersion or scale parameter where one exists. Because it is a sum of unit contributions, it yields deviance residuals for diagnostics; differences in deviance between nested models attribute lack of fit to the added terms and, under regularity conditions, have approximate chi-squared distributions with degrees of freedom equal to the difference in parameters. It is also used to estimate dispersion. These uses depend on declared conventions — the distribution and likelihood, constants, scale treatment, nesting, degrees of freedom, asymptotic conditions and residual definition — and without them the numbers are not comparable.

Scope of Application

Deviance (statistics) belongs to statistical modeling and generalized linear models and is useful where the analyst can specify the typed statistical modeling and generalized linear models carrier, defining objects and relations, parameters, conventions, evidence, boundary cases, and comparison targets, then evaluate the response distribution and likelihood, data and weights, fitted and saturated models, parameter estimates, scale or dispersion, unit and total formula, constants, nesting, degrees of freedom, asymptotic conditions, and residual convention are explicit. The scope is broad within that domain but bounded by the need for the response distribution and likelihood, data and weights, fitted and saturated models, parameter estimates, scale or dispersion, unit and total formula, constants, nesting, degrees of freedom, asymptotic conditions, and residual convention are explicit.

Clarity

The abstraction clarifies a crowded vocabulary by making the response distribution and likelihood, data and weights, fitted and saturated models, parameter estimates, scale or dispersion, unit and total formula, constants, nesting, degrees of freedom, asymptotic conditions, and residual convention are explicit the center of the account. A claim should name the carrier, the governing operation or relation, the applicable assumptions, and the recognition test.

Manages Complexity

Without the abstraction, an analyst must reason directly over many local details: the carrier roles, admissibility assumptions, competing conventions, derived invariants, boundary cases, and proof or validation obligations specific to Deviance (statistics). Deviance (statistics) compresses them into the roles in the structural signature. That compression permits comparison across instances without erasing the variables that determine validity. It also exposes which details may be varied safely and which are constitutive.

Abstract Reasoning

1. Identify the carrier. State what the elements, states, objects, or observations are: the typed statistical modeling and generalized linear models carrier, defining objects and relations, parameters, conventions, evidence, boundary cases, and comparison targets. Reject examples whose alleged carrier belongs to a different problem. 2.

Knowledge Transfer

Knowledge transfers strongly among subfields of statistical modeling and generalized linear models because they reuse the typed statistical modeling and generalized linear models carrier, defining objects and relations, parameters, conventions, evidence, boundary cases, and comparison targets, The fitted model and saturated benchmark are optimized on the same observations; their log likelihoods are differenced and scaled, and nested changes or unit contributions attribute lack of fit under model-specific distribution assumptions., and type the carrier, state every parameter and convention in the definition, test that the response distribution and likelihood, data and weights, fitted and saturated models, parameter estimates, scale or dispersion, unit and total formula, constants, nesting, degrees of freedom, asymptotic conditions, and residual convention are explicit, compare the nearest accepted identity, and report counterexamples, uncertainty, and limiting cases.

Relationships to Other Abstractions

Current abstraction Deviance (statistics) Domain-specific

Parents (1) — outgoing typed relations

• Deviance (statistics) is a kind of Measurement Prime

  It remains its own entry because its identity is fixed by the response distribution and likelihood, data and weights, fitted and saturated models, parameter estimates, scale or dispersion, unit and total formula, constants, nesting, degrees of freedom, asymptotic conditions, and residual convention.

Hierarchy path (1) — routes to 1 parentless root

• Deviance (statistics) → is a kind of Measurement

Neighborhood in Abstraction Space

Deviance (statistics) sits in a crowded region of the domain-specific corpus (14^th percentile for distinctiveness): several abstractions share nearly its structure, so a description that fits it tends to fit its neighbors too.

Family — Statistical Inference & Hypothesis Testing (34 abstractions)

Nearest neighbors

• Interaction (statistics) — 0.92
• Hierarchical generalized linear model — 0.92
• Standard error — 0.91
• Principle of marginality — 0.91
• Nuisance parameter — 0.90

Computed from structural-signature embeddings · 2026-10-25