Main effect and interaction? Need help interpreting the graph.

#1
We performed a multiple linear regression with valence (negative, positive and neutral) as independent variable, social curiosity as controll variable and recognition performance as our dependent variable.

No we got this graph (see attachments).

Do we have a main effect and interaction going on?
 

hlsmith

Less is more. Stay pure. Stay poor.
#2
You are looking for crossover, otherwise considered unparallel slopes. Given this, what would be your interpretation of the graph? In in social sciences those types of lines are called disordinal.
 
#3
Thanks for your quick reply.

So because the lines are not parallel there is no main effect but an interaction?!

Why would those lines considered to be disordinal since the direction of social curiosity is consistent across all levels of valence?
 

rogojel

TS Contributor
#4
hi,
the interpratation should rather be that you have a main effrect, but the size of it depends on the value of the valence - a smallish effect for negative and positive and a much stronger effrect for neutral.
 

hlsmith

Less is more. Stay pure. Stay poor.
#5
Horizontal lines would represent no effect. Lines that cross represent a different effect size given the variable it is being categorized by. These are things that need to be tested. This is because lines may appear not to be horizontal or may cross, but given the sample size and potential sample variability, we may not be able to make those inferences with confidence.
 

noetsi

No cake for spunky
#6
Say for the sake of argument you have two levels of a predictor (A and B which) impact a Y variable C. If the impact of A and B on C don't change at levels of another predictor there is no interaction between the two predictors. If the impact of A relative to B on C changes at levels of the other predictor you do have interaction.

Assuming you have interaction, If A has greater impact on C than B has on C at some levels of the other predictor , but at other levels of the other predictor B has greater impact on C than A has on C, than the interaction is disordinal (and that is a real pain to interpret). If not the interaction is ordinal.

You always have main effects with interaction (in fact you can't remove main effects even if not significant when you have significant interaction). You just have to interpret the main effect at specific levels of the predictor they are interacting with (which is called simple effects often in the ANOVA literature).
 

hlsmith

Less is more. Stay pure. Stay poor.
#8
Per all of the above posts, the graph appears to show effect modification. This would typically mean you would not want to report the marginal effects of the variables. However, you would want to test for an interaction effect to verify it's presence given your sample (e.g., sample size and directly the standard errors).