Howdy, Stranger!

It looks like you're new here. If you want to get involved, click one of these buttons!

Supported by

JASP BayesFactor interpretation for mixed anova interaction

I am using JASP to analyze the data and bayesfactor of mixed ANOVA, but I got some problem with interpretation.
It's a 3(between)x6(within) ANOVA design. The main effect of within variable(A) is significant, but the main effect of the between variable(B) and interaction( AxB ) are not significant. In the bayesfactor(compare to null model), the BF10 are:
A: BF10 = 1.59e+23
B: BF10 = 0.423
A + B: BF10 = 7.56e+21
A + B + A*B: BF10 = 7.02e+21
Is that I can sure there is a strong evidence for the main effects model and interaction model?
The ANOVA p value is non-significant, I have no idea how to interpret the combine result...
Is the BF10 is as bigger as possible in compare to null model order, and as smaller as possible in compare to best model? I am so confused in these detail...
Thank you very much.


  • Hi Herry,

    1. It may help if you select "compare to best model"
    2. Let's break this down. Model A beats the null model by a tremendous amount. Model B actually does worse than the null model. Model A + B outpredicts the null model by a lot, but not by as much as Model A alone. So if you have main effect A, adding B is a bad idea. The same goes for the model that has the interaction.

    In sum, you need only A. Adding anything else makes the model worse, predictively.

  • edited December 2018

    Hi EJ
    I get it! Thank you so much. Just like the hierarchical regression concept. Thank you so much for the suggestion. I try select compare to best model, the BF10 is more clear!
    A: BF10 = 1
    A + B: BF10 = 0.523
    A + B + A*B: BF10 = 0.466
    Null model(incl. subject) = 7.562e-23
    B: BF10 = 0.423 = 3.465e-23

    Are these result stronger shows that no model is stronger than A?
    So if that means in the result, although the interaction is significant in ANOVA, the BF10 is not strong, is that means the data is also not fit to the theory...? I am confused when p-value and the Bayesian result are opposite, how can I make the choice between p-value and Bayesian...

    Thank you

    1. If you set BF to "BF01" instead of "BF10", then you'll see how many more times the best model predicted the data (numbers > 1 are usually easier to interpret).
    2. The most popular way to test an interaction is to compare A+B to A+B+A*B. The easiest way to do this is to include "A" and "B" in the null model (you can do this under "model"). You will see that there is very little evidence. So the BF says that the data do not allow a strong claim about the presence vs absence of the interaction.
Sign In or Register to comment.