Hello all,
In my model, I have three independent constructs (A, B, and C), one moderator (M), and one dependent construct (Y). The interaction terms for B and C with M are positive (B*M = + Value AND C*M = + Value). However, it is negative for A (A*M = - Value), which is against theory and sense!!
Still, If the model comprised A, M, and Y without the other independent constructs (B and C), the interaction term becomes positive and makes sense (A*M = + Value). So, the interaction terms of B and C affect the interaction term of A.
How to solve/ explain this issue?
An estimate may be negative but still make sense if it is not significantly different from zero.
Is the effect to refer to both negative in sign and significant?
Paul F Burke Thank you for your answer. The effect is negative but insignificant. Still, it does not make sense to be negative! Both the moderator and the independent variable are "positive by nature". To my knowledge, this is an issue of collinearity.
As you think, a multicollnearity issue occurs. you can report the results for each moderator separately; and the results with all moderators that you have. Then, for the results with all moderators, you may discuss that the negative interaction term occurs partially due to a multicollinearity.
Hasan Badran 2 remarks:
1) If the effect is not significant, you cannot even conclude the sign of the effect, per definition, since the coefficient is also compatible with positive values (otherwise it would be significant).
2) As you mentioned, it may be a problem of collinearity or more specifically, of suppression. Without seeing the data and/or the strength of the effects, a sign shift can be indicative of it. On the other hand, you should consider if it is necessary or even useful/indicated to put all predictors in one model. To prevent such problems, you may use directed acyclic graphs (DAGs) to get an impression of the causal relationships and associations between the variables, to know for which variables you should adjust and for which you should NOT adjust (e.g. if you are interested in a total effect of X on Y, you should NOT adjust for a mediator).
- Badges
- Science topic
- Similar topics
- Geoscience
- Atmospheric Sciences
- Climate Science
- Climate Change