1) In the first situation which I am facing, both indirect (a*b) and direct (c') effects are insignificant, while their sum, i.e. total effect [(a*b)+c'] is significant.
2) In the second scenario, the indirect effect is insignificant, while direct effect is significant, so there is no mediation, but yet, the total effect is significant.
I would appreciate if someone can share interpretation of these scenarios. What role does total effect play in such a scenarios? What is the best way to report such situations in a research paper?
P.S. I am using confidence interval method to assess significance (whether 0 falls in between LL and UL or not). Attached are the screenshots of both the scenarios.
@Harmanjit Indirect effect is the product of regression coefficients of X>>>M (path 'a') and M>>>Y (path 'b') i.e., a*b whereas, total effect is the sum of direct effect of X on Y (path c') and indirect effect (a*b) i.e., c'+ab. It would have been better for providing any suggestions had you attached a screenshot of your mediation model results.
Imran Anwar thanks for your reply. Actually I am facing two situations.
1) Just as you have explained, in the first situation which I am facing, both indirect (a*b) and direct (c') effects are insignificant, while their sum, i.e. total effect [(a*b)+c'] is significant. I am interested in interpretation of such a scenario.
2) Another situation which I am facing is, wherein indirect effect is insignificant, while direct effect is significant, so there is no mediation, but yet, the total effect is significant. I would appreciate if you can share interpretation of this second scenario too.
P.S. I am using confidence interval method to assess significance (whether 0 falls in between LL and UL or not). I have added screenshots for your reference in the question above.
@Harmanjit Are your paths X>>>M and M>>>Y are significant? Because, for establishing mediation, these two should be significant as per Hayes (2009). You should refer to the paper "Beyon Baron and Kenny: Statistical Mediation Analysis in the New Millennium" attached below by A.F. Hayes. I'm sure it'll benefit you in resolving your query. Wishing you best with your research.
Thanks Imran Anwar for your reply.
I am using the confidence interval approach, wherein usually following are the conditions:
But I am particularly interested in the scenarios asked by me in the question because in those cases the total effect is significant (although there is no mediation in both the scenarios). I am not getting answers towards that.
Michael R. Frone, thanks for your detailed reply.
1) Using Lavaan package in R; sample size = 485
Is lack of statistical power a serious concern? How can it be tested and overcome?
2) Using Monte Carlo simulations (produces comparable results to bias-corrected bootstrap) (Preacher and Selig, 2012). So undertaking bias-corrected bootstrap approach will hardly change these results. Further, multivariate normality is violated in my dataset, so going with no inclusion of 0 within confidence intervals as a measure of significance.
Your explanation addresses scenario 2, but scenario 1 still remains unanswered, or am I missing something in your answer?
Finally, should I report only this much that there was no mediation in either of these scenarios or can I tell something more (i.e. despite no mediation, total effect was still significant)? Is there anything interesting here, that despite no mediation, something is happening between X and Y? I am more keen on this. That was the point in asking this question, otherwise, the story ends with no mediation effect. Looking forward to your reply.
Michael R. Frone I was looking forward to your answer last week. I replied to your queries. Please share your thoughts on the same.
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Structural Equation Modeling