It is possible to test for interaction in Analysis of Covariance (ANCOVA) between the independent variable and the co-variate, but isn't “the homogeneity of regression slopes” an assumption that should not be violated in ANCOVA? If this is the case, then if we have an interaction between independent variable and co-variate, does this not indicate non-homogeneity of slopes (and hence violation of the former assumption)?

So if we find a significant interaction in ANCOVA, is it OK to accept the results? Or should we run ANCOVA at a specified level of co-variate?

Furthermore, if ANCOVA is part of the family of regression analyses, why is it inappropriate to have different slopes (different according to the independent variable) when having different slopes would not be a problem in a regression analysis? And so, when interactions exist, should we run a multiple regression analysis instead of an ANCOVA model?


2 Answers 2


+1 to @FrankHarrell. To be honest, I find a lot of terminology in statistics to be used inconsistently, confusing, or generally unhelpful. It's best to concentrate on the underlying logical structure of your situation. For example, an ANOVA isn't fundamentally different from a multiple regression model. An ANOVA is just a MR where all the explanatory / predictor variables are categorical. An ANCOVA is just a MR where there are categorical explanatory variables (that are of primary interest), and also some continuous covariates (that are assumed to contribute to the DV, but are regarded as nuisance variables not of substantive interest), but no interactions between the factors and the covariates (hence, the assumption of parallel lines, as you state). Note that not everyone seems to use the term ANCOVA in this (traditional) way. Of course, you can also have a MR model with both categorical and continuous variables and interactions between them. The world does not end when this occurs, you just no longer have an 'ANCOVA'.


This is just a nomenclature problem. ANCOVA in its original incarnation often implied an additive model for which non-parallelism was feared and tested. If we used the more general name "linear model" we would avoid this connotation (or perhaps the even more general phrase "multivariable regression model").

Besides rightly worrying about the additivity assumption you should spend a lot of effort examining linearity assumptions (plus constant variance, normality, etc.). The linearity assumption, in my experience, is the most frequently violated assumption that has high impact, causes such problems as apparently significant interactions that are just stand-ins for omitted main effects or nonlinear terms.

  • $\begingroup$ Thank you for your answer, if i got it right you mean for example running a GLM in spss with a categorical IV and a couple of covariates wouldnt be called ANCOVA if interaction exist and the results are still valid. But about the other assumptions( normality, homoscedasticty) as you mentioned I thought the same way, but the reason for my confusion was that i read in many resources that violation of these assumptions are negligable especialy if group sizes are equal. But in wikipedia(assumptions of ancova) it says u shouldnt do ANCOVA if interaction exist. $\endgroup$ Commented Dec 3, 2012 at 21:15
  • 1
    $\begingroup$ I wouldn't get too attached to the word ANCOVA and its traditional definition. I don't like to change the names of techniques when a term is added to the model. So I prefer linear or multivariable model. $\endgroup$ Commented Dec 3, 2012 at 22:04
  • $\begingroup$ Again thanks, can you please recommend me a refrence where i can find such inferences about different models, or is it a matter of experience? Sorry if the question sounds dumb but im a rookie in this field. Also i have asked another question about assumptions of ancova here id be greatful if u could give a hand on that too. $\endgroup$ Commented Dec 3, 2012 at 22:33
  • 1
    $\begingroup$ You need to have had a certain number of stat courses before starting, and to read some good standard regression books. There are many to choose from. $\endgroup$ Commented Dec 4, 2012 at 16:17

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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