There appears to be a few possible alternatives to determining whether there is a significant interaction effect.

It is not clear to me whether testing the statistical significance of an interaction term works the same way for different types of scenarios:

(1) 2 WAY ANOVA: Categorical by Binary (where A= 4 level factor, leading to 3 interaction terms being tested)

  • It seems that F-value for specific interaction term is the most widely used.

  • Are the following acceptable alternative indices? incremental tests using chi-square or F-value change with and without the interaction term

(2) Continuous by Continuous

  • Significance of the overall model (Y=A+B+AB, using overall F-value) and the significant p-value for AB

(4) Continuous by Categorical - Anova would still generate an F-value for (A, B, A*B). - Are there alternative indices for this scenario when A is a factor (4 levels), B is binary?

(5) Categorical by Categorical by Continuous (3-way interaction)

Suggestions for any or all of the above (keeping in mind R is the platform) would be much appreciated.


1 Answer 1


I recommend thinking about ANOVA in terms of linear regression. You will find out that all tests of factors and interaction terms are exactly the same test – comparison of model and submodel.

Nominal factor in ANOVA can be represented as a set of dummy variables (example: dummy coding ). To test significance of the main effect, compare your model with a submodel without the dummy variables of the particular factor. The F statistic is used:

$$ F = \frac{\frac{S^2_{submodel} - S^2_{model}}{df_{submodel}- df_{model}}}{\frac{S^2_{model}}{ df_{model}}} \sim F_{df_{submodel}- df_{model}, df_{model}} $$ where $S^2$ is residual sum of squares, $df$ are degrees of freedom.

If your ANOVA contains interaction, add interaction term in regression equation. Interaction term is simply a product of interacting variables. Interaction of continuous or alternative variable $X_1$ and $X_2$ is variable $ X_1 \cdot X_2$. If $X_1$ and/or $X_2$ are nominal factors create interaction term for each pair of their dunny variables (e.g. if you have 4*6 ANOVA design with interaction, you code $X_1$ with 3 dummy variables, $X_2$ with 5 dummy variables, and there are also 15 interaction variables).

The test of significance of interaction is again the comparison of your model with a submodel without the interaction variables. Again, you obtain $F$ statistic. It is the same statistic regardless the fact you test continuous, nominal, interaction…

You mentioned $\chi^2$ test. It does the same but assumes that the amount of error variance is known, which is usually not true. I don’t recommend.

In regression you can also find so called Wald statistic with t distribution. Wald statistic squared is exactly the same as our F statistic. It also produces the same p-value.


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.