jackdaw
03-16-2009, 08:28 PM
I have a question regarding an ANOVA test I ran.
A 2x2 within-subjects ANOVA provided evidence of an interaction. However, the post hoc tests I've run looking for simple effects (within-subject t-tests comparing B1 to B2 at A1, B1 to B2 at A2, B1 at A1 to B1 at A2, and, finally, B2 at A1 to B2 at A2) are all non-significant! How can I get a significant interaction without any underlying significant simple effects?
I've been reading in some books as well as in an online article (ftp://ftp.spss.com/pub/spss/statistics/nichols/articles/testsnoa.txt) and found some help. Although I believe I am starting to understand the take-home message of how this the situation can arise (logically contradictory results due to sampling error), I am not clear on how I should proceed with my interpretation of this result. I expected this interaction, and it makes sense given previous results. But because I cannot find any simple effects that could have given rise to the interaction, must I conclude that the interaction is nothing more than sample error?
By extension, any time I cannot find significant results at every level down (descending by dropping factors from the ANOVAs), do I need to conclude that they are baseless results caused by sample error? Or is it possible that sampling error caused the simple-effect tests to be non-significant and the interaction may in fact still be valid? Or is it impossible to know?
Many thanks for any help anyone can give! :)
A 2x2 within-subjects ANOVA provided evidence of an interaction. However, the post hoc tests I've run looking for simple effects (within-subject t-tests comparing B1 to B2 at A1, B1 to B2 at A2, B1 at A1 to B1 at A2, and, finally, B2 at A1 to B2 at A2) are all non-significant! How can I get a significant interaction without any underlying significant simple effects?
I've been reading in some books as well as in an online article (ftp://ftp.spss.com/pub/spss/statistics/nichols/articles/testsnoa.txt) and found some help. Although I believe I am starting to understand the take-home message of how this the situation can arise (logically contradictory results due to sampling error), I am not clear on how I should proceed with my interpretation of this result. I expected this interaction, and it makes sense given previous results. But because I cannot find any simple effects that could have given rise to the interaction, must I conclude that the interaction is nothing more than sample error?
By extension, any time I cannot find significant results at every level down (descending by dropping factors from the ANOVAs), do I need to conclude that they are baseless results caused by sample error? Or is it possible that sampling error caused the simple-effect tests to be non-significant and the interaction may in fact still be valid? Or is it impossible to know?
Many thanks for any help anyone can give! :)