Deviance (statistics)

The unit deviance[2]^[3] ${\displaystyle d(y,\mu )}$ is a bivariate function that satisfies the following conditions:

• ${\displaystyle d(y,y)=0}$
• ${\displaystyle d(y,\mu )>0\quad \forall y\neq \mu }$

The total deviance ${\displaystyle D(\mathbf {y} ,{\hat {\boldsymbol {\mu }}})}$ of a model with predictions ${\displaystyle {\hat {\boldsymbol {\mu }}}}$ of the observation ${\displaystyle \mathbf {y} }$ is the sum of its unit deviances: ${\textstyle D(\mathbf {y} ,{\hat {\boldsymbol {\mu }}})=\sum _{i}d(y_{i},{\hat {\mu }}_{i})}$.

The (total) deviance for a model M₀ with estimates ${\displaystyle {\hat {\mu }}=E[Y|{\hat {\theta }}_{0}]}$, based on a dataset y, may be constructed by its likelihood as:^[4][5]

$${\displaystyle D(y,{\hat {\mu }})=2\left(\log \left[p(y\mid {\hat {\theta }}_{s})\right]-\log \left[p(y\mid {\hat {\theta }}_{0})\right]\right).}$$

Here ${\displaystyle {\hat {\theta }}_{0}}$ denotes the fitted values of the parameters in the model M₀, while ${\displaystyle {\hat {\theta }}_{s}}$ denotes the fitted parameters for the saturated model: both sets of fitted values are implicitly functions of the observations y. Here, the saturated model is a model with a parameter for every observation so that the data are fitted exactly. This expression is simply 2 times the

Suppose in the framework of the GLM, we have two nested models, M₁ and M₂. In particular, suppose that M₁ contains the parameters in M₂, and k additional parameters. Then, under the null hypothesis that M₂ is the true model, the difference between the deviances for the two models follows, based on Wilks' theorem, an approximate chi-squared distribution with k-degrees of freedom.^[5] This can be used for hypothesis testing on the deviance.

Some usage of the term "deviance" can be confusing. According to Collett:^[6]

"the quantity ${\displaystyle -2\log {\big [}p(y\mid {\hat {\theta }}_{0}){\big ]}}$ is sometimes referred to as a deviance. This is [...] inappropriate, since unlike the deviance used in the context of generalized linear modelling, ${\displaystyle -2\log {\big [}p(y\mid {\hat {\theta }}_{0}){\big ]}}$ does not measure deviation from a model that is a perfect fit to the data."

However, since the principal use is in the form of the difference of the deviances of two models, this confusion in definition is unimportant.

Examples

The unit deviance for the Poisson distribution is ${\displaystyle d(y,\mu )=2\left(y\log {\frac {y}{\mu }}-y+\mu \right)}$, the unit deviance for the Normal distribution is given by ${\displaystyle d(y,\mu )=\left(y-\mu \right)^{2}}$.

See also

• Akaike information criterion
• Deviance information criterion
• Hosmer–Lemeshow test, a quality of fit statistic that can be used for binary data
• generalized linear models
  for count data
• Peirce's criterion

Notes