Deviance information criterion explained

The deviance information criterion (DIC) is a hierarchical modeling generalization of the Akaike information criterion (AIC). It is particularly useful in Bayesian model selection problems where the posterior distributions of the models have been obtained by Markov chain Monte Carlo (MCMC) simulation. DIC is an asymptotic approximation as the sample size becomes large, like AIC. It is only valid when the posterior distribution is approximately multivariate normal.

Definition

Define the deviance as

D(\theta) =-2 log (p(y|\theta)) + C , where y are the data, \theta are the unknown parameters of the model and p(y|\theta) is the likelihood function. C is a constant that cancels out in all calculations that compare different models, and which therefore does not need to be known.

There are two calculations in common usage for the effective number of parameters of the model. The first, as described in, is

p _{D=\overline{D(\theta)}-D(\bar{\theta})} , where \bar{\theta} is the expectation of \theta . The second, as described in, is p _{D} = p _{V} = (1)/(2) \overline{\operatorname{var}\left(D(\theta)\right)} . The larger the effective number of parameters is, the easier it is for the model to fit the data, and so the deviance needs to be penalized.

The deviance information criterion is calculated as

DIC = p _{D+\overline{D(\theta)},}

or equivalently as

DIC = D(\bar{\theta}) +2 p _{D.}

From this latter form, the connection with AIC is more evident.

Motivation

The idea is that models with smaller DIC should be preferred to models with larger DIC. Models are penalized both by the value of

\bar{D} , which favors a good fit, but also (similar to AIC) by the effective number of parameters p _{D} . Since \bar D will decrease as the number of parameters in a model increases, the p _{D} term compensates for this effect by favoring models with a smaller number of parameters.

An advantage of DIC over other criteria in the case of Bayesian model selection is that the DIC is easily calculated from the samples generated by a Markov chain Monte Carlo simulation. AIC requires calculating the likelihood at its maximum over

\theta , which is not readily available from the MCMC simulation. But to calculate DIC, simply compute \bar{D} as the average of D(\theta) over the samples of \theta , and D(\bar{\theta}) as the value of D evaluated at the average of the samples of \theta . Then the DIC follows directly from these approximations. Claeskens and Hjort (2008, Ch. 3.5) show that the DIC is large-sample equivalent to the natural model-robust version of the AIC.

Assumptions

In the derivation of DIC, it is assumed that the specified parametric family of probability distributions that generate future observations encompasses the true model. This assumption does not always hold, and it is desirable to consider model assessment procedures in that scenario.

Also, the observed data are used both to construct the posterior distribution and to evaluate the estimated models.Therefore, DIC tends to select over-fitted models.

Extensions

A resolution to the issues above was suggested by, with the proposal of the Bayesian predictive information criterion (BPIC). Ando (2010, Ch. 8) provided a discussion of various Bayesian model selection criteria. To avoid the over-fitting problems of DIC, developed Bayesian model selection criteria from a predictive view point. The criterion is calculated as

IC = \bar{D} +2 p ^{\theta[} _{D =-2 E} log (p(y|\theta))] +2 p _{D.}

The first term is a measure of how well the model fits the data, while the second term is a penalty on the model complexity. Note that the in this expression is the predictive distribution rather than the likelihood above.

See also

• Akaike information criterion (AIC)
• Bayesian information criterion (BIC)
• Focused information criterion (FIC)
• Hannan-Quinn information criterion
• Kullback–Leibler divergence
• Jensen–Shannon divergence
• Watanabe–Akaike information criterion (WAIC)

References

• Tomohiro . Ando . 2007 . Bayesian predictive information criterion for the evaluation of hierarchical Bayesian and empirical Bayes models . . 94 . 443–458 . 10.1093/biomet/asm017 . 2 .
• Ando, T. (2010). Bayesian Model Selection and Statistical Modeling, CRC Press. Chapter 7.
• Tomohiro . Ando . 2011 . Predictive Bayesian Model Selection . . 31 . 13–38 . 10.1080/01966324.2011.10737798 . 1–2 . 123680697 .
• Claeskens, G, and Hjort, N.L. (2008). Model Selection and Model Averaging, Cambridge. Section 3.5.
• Book: Gelman . Andrew . Andrew Gelman. Carlin . John B.. Stern . Hal S.. Rubin . Donald B. . Donald Rubin. 2004 . Bayesian Data Analysis: Second Edition. Texts in Statistical Science . CRC Press. 978-1-58488-388-3 . 2003051474 . 2027492.
• van der Linde, A. (2005). "DIC in variable selection", Statistica Neerlandica, 59: 45-56.
• David J. . Spiegelhalter . David Spiegelhalter . Nicola G. . Best . Nicky Best . Bradley P. . Carlin . Angelika . van der Linde. Angelika van der Linde . 2002 . Bayesian measures of model complexity and fit (with discussion) . . 64 . 4 . 583–639 . 10.1111/1467-9868.00353 . 1979380 . 3088806 . free.
• David J. . Spiegelhalter . David Spiegelhalter . Nicola G. . Best . Nicky Best . Bradley P. . Carlin . Angelika . van der Linde . Angelika van der Linde . 2014 . The deviance information criterion: 12 years on (with discussion) . . 76 . 3 . 485–493 . 10.1111/rssb.12062 . 119742633 .

External links

• Web site: Richard . McElreath . January 29, 2015 . Statistical Rethinking Lecture 8 (on DIC and other information criteria) . . https://ghostarchive.org/varchive/youtube/20211221/vSjL2Zc-gEQ . 2021-12-21 . live.