How should I understand the anova result when comparing two models?


  Res.Df    RSS Df Sum of Sq      F    Pr(>F)    
1      9 54.032                                  
2      7  4.632  2      49.4 37.329 0.0001844 ***

The manpage states: "Compute analysis of variance (or deviance) tables for one or more fitted model objects." However, out professor mentioned that it may be employed for model comparison - that's what I intend to do.

Hence I assume I could use anova(model1, model2) and obtain a p-value which tells me whether I should reject the null hypothesis: "the models are the same".

May I state that if the p-value is less then (let's say) 0.05, the models differ significantly?

  • $\begingroup$ In your example, are model1 and model2 nested? That is, do both models have a shared set of predictor variables and the same outcome variable, but one model has one or more additional predictor variables? $\endgroup$ – EdM May 15 '15 at 15:26
  • $\begingroup$ One is like Y ~ X + X^2 and the second one Y ~ X + X^2 + X^3 $\endgroup$ – petrbel May 15 '15 at 15:30

Assuming your models are nested (i.e. same outcome variable and model 2 contains all the variables of model 1 plus 2 additional variables), then the ANOVA results state that the 2 additional variables jointly account for enough variance that you can reject the null hypothesis that the coefficients for both variables equal 0. This is effectively what you said. If both coefficients equal 0 then the models are the same.

Just as an additional note, in case you weren't aware, ANOVA is always equivalent to doing model comparisons. When you are looking at the ANOVA for a single model it gives you the effects for each predictor variable. That is equivalent to doing a model comparison between your full model and a model removing one of the variables. i.e. $Model 1: y=a+bx_1+cx_2+dx_3; Model 2: y=a+bx_1+cx_2$ will give you the sum of squares (type III) and test statistic for $x_3$. Just note that R gives you type I sum of squares. If you need type III, use car::Anova or use anova and keep changing the order of the variables in the model and only take the sum of squares for the last variable.

  • $\begingroup$ If I understood you properly, the p-value less than 0.05 proves that the models differ, tight? $\endgroup$ – petrbel May 15 '15 at 15:48
  • 3
    $\begingroup$ I wouldn't use those words (i.e. "prove" and "models differ"), but we mean the same thing. I would say that your data does not support the null hypothesis that the coefficients are both 0 or that the data supports the alternative hypothesis that the coefficients are not both 0. $\endgroup$ – le_andrew May 15 '15 at 16:53
  • 1
    $\begingroup$ Before a claim is made that models are proved to be different or that the null hypothesis is not supported, be sure that the data reasonably meet the assumptions of ANOVA that underlie the calculation and interpretation of the p-values. $\endgroup$ – EdM May 15 '15 at 21:37
  • $\begingroup$ Just to be sure, the last part about type I sum of squares only applies to the situation where anova() is used on one model? $\endgroup$ – Jasper Nov 29 '18 at 15:01

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.