Take a look at the literature on suppressor effects, for example:

Maassen, G. H., & Bakker, A. B. (2001). Suppressor variables in path models: Definitions and interpretations. Sociological methods & research, 30(2), 241-270.

Suppression is not specific to SEM. It also occurs in regular multiple regression with observed variables.

Hi Charlotte,

can you show a path diagram with explanation of the variables involved. Strange negative effects ring the "collider bias" bell. Christian Geiser you pointed into the right direction although I found this paper here much more eye-opening:

Kim, Y. (2019). The causal structure of suppressor variables. Journal of Educational and Behavioral Statistics, 1076998619825679. doi:10.3102/1076998619825679

More on collider bias:

Elwert, F., & Winship, C. (2014). Endogenous selection bias: The problem of conditioning on a collider variable. Annual Review of Sociology, 40, 31-53. doi:10.1146/annurev-soc-071913-043455

Best,

Holger

Thanks to both of you!

That is exactly what I was looking for, I think.

Holger Steinmetz I attached the path diagram. As you can see, we have 6 IVs with one indicator each (those are composite scores from more complex data). When we have only the predictor "G_Social_factor" as IV, the path is positively significant. When other IVs enter the model, this path loading is reversed. When playing around with enterin different combinations of IV's, we see that this reversal/suppression effect wanders from one IV to the other. It is to mention, that the IVs represent expected gains and losses, and the willingness to do sth is the DV. A gain shoud increase the willingness, a loss decrease it. The IVs are somewhat related, as all of them represent different gains and losses of the same thing.

Your help is really appreciated, and thanks for taking the time!

Charlotte

Hi Charlotte,

without knowing the exact meaning of all the variables, it is not possible to say anything.

However, think about wether the following is possible (that would create the problem)

1) Willingness is a cause those IVs that create the phenomenon

2) G_social is also a cause of the IVs

This would render the IVs being colliders, hence, creating a collider bias of the relationship between G_social and willingness.

Two further comments.

a) you have reactance and "CI" uncorrelated with all other exogeneous variables. Why?

b) You have a lot of indicators. I bet the measurement model does not fit cleanly what I would strive to correct. If the measurment model is misspecified, the parameters aren't trustworthy anyway.

Best,

Holger

Dear Holger Steinmetz

thanks for your recommendations!

1) Do you propose that my model is misspecified in the sense that the DV actually causes the IVs? (reversed causality)

And if so, from what indication do you make this assumption?

2) That G Social is the cause of the other IVs is highly unlikely from a theoretical standpoint. This would make no sense at all.

To your comments,

a) the left 6 constructs are IVs, while CI and reactance function at control variables. Should these then be correlated, and if so, why?

In my interpretation, they work as control variables for the DV (willingness), and therefore are only "linked" to Willingness (and not to the IVs).

b) the CFAs (measurement models) all fit well, therefore I don't believe the problem arises here.

However, you previously pointed me toward the suppressor effect. I think that in fact, it is likely that negative suppression is happening between the 6 IVs and willingness.

I come to this conclusion, as the preconditions for suppression are present, i.e.:

- significant zero-order correlations between all IVs and the DV

- positive significant regression weights from IV to DV in SEM when only 1 IV is in the model (and all other 5 IVs are not there)

- a reversal of sign when other DVs enter the model

I would conclude, that our IVs have more in common with each other than with the DV, which is why some of the effects turn negative, although all variables are positively related (ie significant zero-order correlations).

Would you agree?

Thanks so much, and I hope I am not taking up too much of your time here.

Your help is appreciated,

Charlotte Baar