The overall interaction effect appears to be insignificant (p .081) but the moderator is significant at average and higher levels (p .01 and p .000) respectively in conditional effects. What inferences are plausible to draw from this result?
Key: Predictor = Education Stress (ESSA), Moderator = Perceived Social Support (PSS), Dependent Variable = Psychological Wellbeing (PWB).
Hello Aleem,
I think, the first main message would be: PSS significantly predicts Wellbeing. Then, in general (before coming to your specific question), regarding the other predictors and the moderation (or interaction), it depends on whether you did a Power analysis beforehand. Assuming there was none, the remaining thing you can say is, the result is inconclusive for all other predictions and the moderation (or: We don't know, whether its there or not). Assuming you did a Power analysis, then the usual conclusion would be, ESSA has no effect on wellbeing; and there is no interaction between ESSA and PSS (instead of just being inconclusive).
So without knowing more, I simply would be inconclusive regarding the interaction. The marginal effects you got, pretty much show the same picture (of being inconclusive), as they show, sometimes the data is less probably than 5% under the null hypothesis (i.e. no effect of ESSA at the mean and +1SD of PSS), and then the data also is more probable than 5% under the Null (at -1SD of PSS). In general, the data is more probable than 5% under the Null (i.e. assuming that there is no interaction between ESSA and PSS).
I see no conflict. This is not an unusual result. Also, I think the way of thinking/asking about it, shows some lack of confidence (in statistical sense) in how to think about the non-significant effect, which is also contributing to the answer. Because, if we would know how much we can trust a p-value of 8% in our experiment, we would be able to draw a conclusion. So, being simply inconclusive seems appropriate :)
Best wishes, René
René Schlegelmilch thanks for your reply. It clears some doubts. However, I am still unsure what these negative beta values (-.48 and -.71) mean in perceived social support row. I expected perceived social support to interact with educational stress to enhance (+) psychological well-being. But these negative beta values are confusing me. Can you or anyone else chime in to solve this query?
I see, that's a different question, but good that you mention it :) ... I did not recognize. It reminds me of two things.
1 such tests (usually) depend on whether the continuous predictors have been mean-centered or not, because the interaction is constructed as multiplicative term ESSA*PSS. This means, if these two IV'S are not centered, they will correlate with ESSA*PSS (e.g. 1*1=1; 10*10=100, and so on), which is always bad in a regression (multi-collinearity of predictors). Another way of viewing at centering is: the constant (intercept; as well as other factor means), are always estimated at the 0's of the continuous predictors (i.e. where they have no influence), which 1) might not even be present or plausible in the data (e.g. nobody has a PSS=0), and 2) is usually better addressed by saying that the continuous predictor 'has no influence' at the mean of itself (i.e. mean-centering).
2 If it is true what one sees there (i.e. that the overall effect of ESSA is positive), but that the marginal effects of ESSA are negative at PSS=53 and even more negative at PSS=73, this indeed seems to be an issue of mean-centering, or at least is very suspicious. For example, if most people would have a PSS of about 30 (i.e. with some outliers making a mean of 64), and the ESSA effect becomes more positive with decreasing PSS, then this output would make sense...
In other words, you might want to check the distribution of PSS and ESSA, clear outliers, mean-center them, run again :)
Hope this helps,
René
- Badges
- Science topic
- Similar topics
- Philosophy
- Esthetics
- Interpretation