0
$\begingroup$

In an ANOVA model, there is a constraint that the coefficients must sum to zero. What does this actually mean? I do understand the reason why you might want to make them sum to zero, i.e. to have two degrees of freedom to estimate two parameters for instance, and not 3 df for estimating 2 parameters.

What parameters actually sum to zero?

Improve this question
$\endgroup$
4
  • 1
    $\begingroup$ Where did you find such information? There is no such constraint. $\endgroup$
    – Tim ♦
    May 15, 2019 at 14:15
  • $\begingroup$ @Tim Consider two-way ANOVA for one example (whether main effects or with interaction); there will have to be some form of regularization or constraint on the parameters or the model is not identifiable; in the usual regression parameterization, a baseline level of each factor will be omitted (constraining its parameter to 0). In another parameterization, sum-to-zero constraints are used. There are other parameterizations still, but they all introduce enough constraints to make the model identifiable. $\endgroup$
    – Glen_b
    May 16, 2019 at 3:18
  • $\begingroup$ @Glen_b I meant that this is not generally true. $\endgroup$
    – Tim ♦
    May 16, 2019 at 4:02
  • $\begingroup$ Related: stats.stackexchange.com/q/257778/119261. $\endgroup$
    – StubbornAtom
    Apr 5, 2020 at 21:09

1 Answer 1

Reset to default
3
$\begingroup$

I think you are confusing the coefficients with the contrasts. The contrast refers to the specific way that the coefficient is estimated. When fitting ANOVAs we describe contrasts that sum to 0 as orthogonal. For instance, in a linear regression model the usual dummy encoding for a factor variable (say Education) is:

$$ \begin{array}{l|ccc} & C_0 & C_1 & C_2 \\ \hline \text{Less than High School} & 1 & -1 & -1 \\ \text{High School} & 0 & 1 & 0 \\ \text{Some College} & 0 & 0 & 1\\ \end{array}$$

So with the exception of the intercept term ($C_0$) the contrasts add up to 0 columnwise. That means that the interpretation of $C_1$ is a mean difference between high school and less than high school and $C_2$ a mean difference from some college to less than high school.

Improve this answer
$\endgroup$

Your Answer

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

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