Structural Equation Modeling (SEM) Discussion > Can't understand suppression
This is a great question and does indeed sound like a classic case of suppression.
Because each effect is a partial effect in multiple regression (i.e. each predictor's estimated effect is holding all other predictors constant at zero), then each of the predictors is having some degree of suppression effect on every other predictor in the model.
My guess (having not seen your data or model) is that predictor B covaries fairly highly with the other predictors and shares much of the predicted variability of D with A & C also, so what B is negatively associated with isn't strictly D, but rather variability in D that is NOT explained by by A or C.
These types of effects can be very confusing and difficult to interpret, which is why you generally want to avoid multicollinearity in your regression models.
Jeremy Taylor
Jeremy,
Thanks. I'm still trying to sort it all out in my head. My model is the effect of pre-implementation change management (A), post-implementation change management (B), and process management (C) on appropriate use of a healthcare information system (D). The implementation in question is the implementation of the information system.
My predictors do correlate pretty strongly, but there isn't anything I can do about that. When the predictors are allowed to correlate, the coeffients are 0.5, -0.6, and 0.9 respectively. If I remove the correlation lines, AD drops a little, to 0.4, and CD drops a little, to 0.7, but the BD coefficient drops to -0.1. This seems significant but I don't know how (I'm not a stats heavyweight). And as I said, when taken individually, the BD relationship is positive and significant.
The idea that the more post implementation change management that occurs actually has a negative effect on use just doesn't make sense. I have to defend this thesis next month so I want to find a plausible explanation.
I appreciate your help.
Bill Mac
One more thought. Apologies if I'm wearing you out with questions. I'm trying to make sense of what you said: " so what B is negatively associated with isn't strictly D, but rather variability in D that is NOT explained by by A or C."
I did some qualitative interviews with some of my respondents, and several of them said that for many organizations, post implementation change management is not as consistent as pre-implementation change management. Many orgs just drop the ball once the system is implemented. But the nurses (my primary respondents) still use the system extensively. Is it a reasonable explanation that the positive effects of A and C simply overpower the negative effects of B? Is that kind of what you are saying? Sorry to keep pestering.
Thanks
Bill Mac
Bill,
I'm afraid I don't know enough about your project or model to give you any feedback on the interpretation of your model specifically. However, I will say that I would never advise including 2 predictor that are correlated as highly as .9 in the sample predictive model, as the resulting individual coefficients are unlikely to represent what we are intending.
When 2 variables are correlated that strongly, they essentially function as 2 (nearly identical) measurements of the same construct, so having them together in a model will essentially only allow you to determine how predictive one is of the noise in the DV.
As for your explanation, I'm afraid I don't know what you mean by one effect "overpowering" another. The best description I can give is this:
if I have 2 measures of intelligence (extremely highly correlated, because they are 2 measures of the same thing) and I enter them into a regression together to predict annual salary, then the individual effect of neither of the 2 measures will actually be the effect of intelligence on salary (even though both are actually measures of intelligence). Instead, what the effect would be is essentially the predictive effect of the measurement error of each of the two intelligence measures.
Put another way, including two variables as predictors in a regression that correlate as highly as .9 is kind of like saying:
I want to test whether "age" predicts "gray hairs", controlling for "number of years alive on this earth"... the question doesn't make much sense and neither would the results.
Jeremy Taylor
Jeremy,
Thanks for the input. I'll have to chew on it. The 0.9 I mentioned is not a correlation weight but is a regression coefficient, inflated by the suppressor.
Bill
Bill Mac
I have three predictors: A, B, and C. One Dependent, D. A, B, C, and D are all positively correlated with each other. Using single regression, each predictor is positively related to D, but the A-D relationship is non-significant. When analyzed simultaneously, the AD and CD relationships are positive and significant. The BD relationship is negative and significant. (this one is unexpected) So I think there is suppression going on, but which one is suppressing what? I've read the explanations for suppressors but nothing has clicked yet. The BD relationship strengthens the other two, but ends up with a negative coefficient itself.
Any help is appreciated.