I know that GLM models are nested when one can obtain the smaller model by removing $\beta$'s from the larger model. However, when doing some exercises I found on the internet, I stumbled about doing a partial deviance test for the following two models:

Here, let covariate A be a continuous covariate and B be a factor covariate with 2 levels

Model 1: Intercept, A

Model 2: No Intercept, A, B coded as [0, 1] and [1, 0] for the two levels.

In the exercise, it has been claimed that Model 1 is nested within Model 2.

Here, obtaining Model 1 by setting parameters of Model 2 to zero is not really possible. However, Model 1 is kind of included when we take $\beta_{intercept} = \beta_{factor1} + \beta_{factor2}$. Does this then imply that the two models are nested? I always thought that the same covariates must be included in the model to have nested models.

Thank you in advance for your help and clarification! :)


Model 1 is nested within Model 2 ( so model 2 is the larger model) means that every fit by Model 1 can be matched exactly by Model 2 (but parameter values can well be different). I explained this in another context here: Why is sum of squared residuals non-increasing when adding explanatory variable?

Such a definition is necessary to take account of different parametrizations, your example case is the same linear model parametrized in two different ways. In the case of linear models $Y_1=X\beta_1 + \epsilon_1, Y_2=Z\beta_2 + \epsilon_2$, this can be paraphrased using the concept (from linear algebra formulation of linear models) of model space, which is the rank space of the design matrix, that is, $$ R(X)=\{ x\beta \colon \beta\in\mathbb{R}^p \} $$ that is, the space of all possible expected values the model can produce. Model 1 is then nested in model 2 if $R(X) \subseteq R(Z)$. With this definition nesting is independent of parametrization.

Your understanding is related to Thomas Lumley's concept of symbolic nesting: What is a "symbolically nested" model?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.