Skip to content

Deviance Information Criterion

Compare Bayesian hierarchical models by adding posterior mean deviance to an effective-complexity penalty derived from posterior deviance, using quantities readily estimated from MCMC draws.

Version
v4 · 2026-09-19 · History
Domain-specific #
1658
Domain group
Formal Sciences
Origin domain
Experimental Design & Statistics
Subdomain
Bayesian Model Comparison → Experimental Design & Statistics
Aliases
DIC, Deviance IC

Core Idea

For data (y), parameter \(\theta\), and likelihood \(p(y\mid\theta)\), define deviance \(D(\theta)=-2\log p(y\mid\theta)+C\). The deviance information criterion uses posterior draws to estimate fit and effective complexity:

\[ p_D=\overline{D(\theta)}-D(\bar\theta),\qquad \mathrm{DIC}=\overline{D(\theta)}+p_D. \]

Equivalently, \(\mathrm{DIC}=D(\bar\theta)+2p_D\). Smaller values are preferred among models fit to the same data with the same likelihood and deviance convention.[1] DIC was designed for complex hierarchical models where the nominal parameter count is unclear but posterior deviance is cheap to compute from MCMC output.

How would you explain it like I'm…

The Fit-but-Not-Fancy Score

Say you have a few different guesses about how something works. You want a guess that matches what you saw really well, but isn't too complicated. The Deviance Information Criterion is a score that adds points for matching badly and more points for being too complicated. For guesses tested on the same facts, the one with the lowest score wins.

Fit-Plus-Complexity Score

When scientists have several models to explain the same data, they want one that fits well but is not needlessly complicated. The deviance information criterion, or DIC, gives each model a number. One part measures how badly the model fits the data on average. Another part estimates how complicated the model really is, which is tricky for models with many layers. Add them together, and the model with the smaller DIC is preferred, as long as all the models were compared on the same data in the same way.

Bayesian Fit-and-Complexity Criterion

The Deviance Information Criterion (DIC) is a tool for comparing Bayesian models. It starts from deviance, which is −2 times the log of how likely the data are under the model (plus a constant) — lower deviance means better fit. Using many samples from the posterior (the model's updated beliefs about its parameters), DIC averages the deviance to measure fit, then estimates an "effective number of parameters," p_D, as the average deviance minus the deviance at the average parameter value. DIC equals average deviance plus p_D, so it rewards fit but penalizes complexity. Lower DIC is preferred, but only among models fit to the same data using the same likelihood and deviance convention. It was designed for complex hierarchical models, where it's unclear how many parameters a model "really" has.

 

For data y, parameter θ and likelihood p(y|θ), the deviance is D(θ) = −2 log p(y|θ) + C. The deviance information criterion uses posterior draws, typically from MCMC, to estimate both fit and effective complexity. The posterior mean deviance D̄ measures fit; the effective number of parameters is p_D = D̄ − D(θ̄), the gap between the mean deviance and the deviance at the posterior mean of θ. Then DIC = D̄ + p_D, or equivalently D(θ̄) + 2p_D. Lower values are preferred, but comparisons are valid only among models fit to the same data with the same likelihood and deviance convention, since the constant C and the likelihood definition must match. DIC was designed for complex hierarchical models, where the nominal parameter count is ambiguous but the posterior deviance is cheap to compute from sampler output.

Structural Signature

  • A likelihood-defined Bayesian model.
  • Observed data shared across compared models.
  • A posterior distribution, often represented by MCMC draws.
  • Pointwise or aggregate deviance.
  • Posterior mean deviance as fit term.
  • A posterior summary \(\bar\theta\).
  • Effective complexity (p_D).
  • Fit-plus-penalty score.
  • Lower-is-better comparison.
  • Common likelihood normalization and observation unit.
  • Adequate posterior sampling.
  • Diagnostics for irregular or weakly identified posteriors.

What It Is Not

DIC is not a Bayes factor or posterior model probability. It does not integrate the likelihood against the prior as marginal likelihood does. It is not AIC, although its penalty analogy is deliberate. It is not WAIC or leave-one-out cross-validation, which use pointwise predictive quantities and generally have stronger modern predictive justification.[2]

Scope of Application

DIC is most natural for approximately regular hierarchical models with posterior distributions summarized sensibly by a mean and with a clearly defined likelihood focus. It is easy to compute in BUGS-style workflows.[1]

It can behave poorly under multimodality, skewness, mixtures, weak identification, parameterization changes, negative effective parameter counts, or ambiguous choices about latent variables in the likelihood. The 2014 retrospective emphasizes these limitations.[3]

Clarity

Report the deviance definition, which stochastic nodes count as parameters, the likelihood focus, the (p_D) variant, posterior summary, observation granularity, MCMC diagnostics, and comparison set. A DIC value is meaningless in isolation; only differences under a common construction are interpretable.

Manages Complexity

DIC compresses a posterior sample into one fit–complexity coordinate without requiring a nominal dimension for a hierarchical model. It also decomposes the score so analysts can see whether a model wins through fit or pays for flexibility. The convenience is computational, not a guarantee that the approximation matches predictive performance.

Abstract Reasoning

  1. Fix comparable data and likelihood definitions.
  2. Obtain converged posterior draws.
  3. Compute deviance for every retained draw.
  4. Average the deviance.
  5. Compute \(D(\bar\theta)\) under a declared parameterization.
  6. derive (p_D) and inspect its plausibility.
  7. Add fit and penalty.
  8. Compare differences, not absolute scores.
  9. Cross-check with posterior predictive diagnostics, WAIC, or LOO when irregularity matters.

Knowledge Transfer

The portable pattern is estimate flexibility from the gap between average realized fit and fit at a representative state, then penalize average fit by that effective flexibility. The proposed immediate parent is Optimality Criterion.

Examples

A multilevel model may contain hundreds of latent effects yet have (p_D) far below the raw count because partial pooling constrains them. Two models can have similar mean deviance but different DIC when one consumes much more effective flexibility. Conversely, a mixture model can yield nonsensical (p_D) because the posterior mean is not a representative fitted state.[4]

Structural Tensions

  • Computational convenience versus predictive validity.
  • Nominal dimension versus effective complexity.
  • Posterior averaging versus representative point evaluation.
  • Hierarchical shrinkage versus latent-variable focus choice.
  • Parameterization dependence versus desired model invariance.
  • One-number ranking versus diagnostic uncertainty.

Structural–Framed Character

Fit–complexity tradeoff, posterior averaging, and penalty construction are structural. Deviance, likelihood focus, MCMC, hierarchical parameters, and Bayesian posterior conventions are constitutive. The identity is domain-specific.

Structural Core vs. Domain Accent

The portable core is average fit + effective-flexibility penalty. The domain accent is Bayesian deviance computed over posterior draws.

This entry is a kind of Optimality criterion.

Optimality Criterion is the broader abstraction this entry instantiates. Complexity Penalty, Model Selection, Approximation, and Bias–Variance Tradeoff are related. Focused Information Criterion is a sibling, not coverage.

Relationships to Other Abstractions

Local relationship map for Deviance Information CriterionParents appear above the current abstraction, mutual partners to the right, and children below. Node labels state whether each abstraction is prime or domain-specific; colors identify relation types.Deviance InformationCriterionDOMAINDomain-specific abstraction: Optimality criterion — is a kind ofOptimalitycriterionDOMAIN

Current abstraction Deviance Information Criterion Domain-specific

Parents (1) — outgoing typed relations

  • Deviance Information Criterion is a kind of Optimality criterion Domain-specific

    Optimality Criterion is the broader abstraction this entry instantiates.

Hierarchy paths (2) — routes to 2 parentless roots

Neighborhood in Abstraction Space

Deviance Information Criterion sits in a sparse region of the domain-specific corpus (98th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (2769 abstractions)

Nearest neighbors

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

Not to Be Confused With

  • Akaike information criterion.
  • Bayesian information criterion.
  • Bayes factor.
  • WAIC.
  • Leave-one-out cross-validation.
  • Residual deviance alone.
  • Focused information criterion.

References

[1] David J. Spiegelhalter, Nicola G. Best, Bradley P. Carlin, and Angelika van der Linde, “Bayesian Measures of Model Complexity and Fit,” JRSS B 64, no. 4 (2002): 583–639, doi:10.1111/1467-9868.00353. registry ↩a ↩b

[2] Aki Vehtari, Andrew Gelman, and Jonah Gabry, “Practical Bayesian Model Evaluation Using Leave-One-Out Cross-Validation and WAIC,” Statistics and Computing 27 (2017): 1413–1432, doi:10.1007/s11222-016-9696-4. registry ↩

[3] David J. Spiegelhalter et al., “The Deviance Information Criterion: 12 Years On,” JRSS B 76, no. 3 (2014): 485–493, doi:10.1111/rssb.12062. registry ↩

[4] Andrew Gelman, Jessica Hwang, and Aki Vehtari, “Understanding Predictive Information Criteria for Bayesian Models,” Statistics and Computing 24 (2014): 997–1016, doi:10.1007/s11222-013-9416-2. registry ↩