Deviance measures the discrepancy between the current model and the full model. The full model is the model that has n
parameters, one parameter per observation. The full model maximizes the log-likelihood function. The full model provides a point of comparison for models with fewer than n
parameters. Comparisons to the full model use the scaled deviance.
The contribution to the scaled deviance from each individual data point depends on the model.
The deviance table is constructed based on the following general result which assumes that ϕ is known. If DI is the deviance associated with an initial model and DS is the deviance associated with a subset of terms in the initial model, then under some regularity conditions, the following relationship exists:
The difference between the deviances is asymptotically distributed as a chi-square distribution with d degrees of freedom. These statistics are calculated for adjusted (type III) analysis and sequential (type I) analysis. The adjusted deviance and the chi-square statistic in the deviance table are equal. The adjusted mean deviance is the adjusted deviance divided by the degrees of freedom.
For the sequential analysis, the output depends on the order that the predictors enter the model. The sequential deviance is the unique portion of the deviance that a predictor explains, given any predictors already in the model. If you have a model with three predictors, X1, X2, and X3, the sequential deviance for X3 shows how much of the remaining deviance that X3 explains given that X1 and X2 are already in the model. To obtain a different sequential deviance, repeat the regression procedure entering the predictors in a different order.
If ϕ is unknown, as for responses that follow a normal distribution, then under some regularity conditions the relationship changes to the following:
Here, the difference between the deviances is asymptotically distributed as an F distribution with d degrees of freedom for the numerator and n − p degrees of freedom for the denominator. To estimate the dispersion parameter, use the initial model.
|yi||the number of events for the ith row|
|the estimated mean response of the ith row|
|mi||the number of trials for the ith row|
|Lf||the log-likelihood of the full model|
|Lc||the log likelihood of the model with a subset of terms from the full model|
|d||the degrees of freedom are the difference between the numbers of parameters in the models to compare|
|ϕ||the dispersion parameter, known to be 1 for the binomial and the Poisson models|
|n||the number of rows in the data|
|p||the regression degrees of freedom for the initial model|