What you describe does fit with the concept of suppression, but to make such an analysis meaningful, you need to identify the variable(s) responsible for the suppression effect and provide a reasonable explanation for why this happens.

Depending on how many variables are in your multiple regression, you could examine a series of simpler analyses that contain just your variable and one of the additional variables. Rotating through your control variables one at a time this way should help you identify which one (or more) are responsible for the suppression effect.

@David thank you so much for your help!

I Only have 3 IV so i tried all the combinations. It seems i found the variable responsible for the suppression effect, but in any other combinations that my IV has a positive effect, p value is high and the Confidence intervals include zero. Although when i run the regression with Just this IV and my DV every thing is fine. Whats the Best thing i can do? Should i Just announce the colerration false or Just get the IV fully out of the multiple regression?

When predictors are correlated, suppression effects tend to happen. In your case, you could have a negative suppression which can be explained, with the theoretical nature of predictors in mind, that the contribution of the negative suppressor to the prediction of your DV is primarily by clearing out criterion-irrelevant variance from your other IV. For germane insights, you might refer to the following article.

Tzelgov, J., & Henik, A. (1991). Suppression situations in psychological research: Definitions, implications, and applications. Psychological Bulletin, 109(3), 524–536. https://doi.org/10.1037/0033-2909.109.3.524

Good luck,

Hello Dimitrios,

There are several correlations that can be generated with a data set like yours:

1. Zero-order correlation. The bivariate correlation of scores on a given IV with scores on the given DV, ignoring all other IVs.

2. The partial correlation. The bivariate correlation of scores on a given IV with the unexplained portion (a.k.a., residuals) of scores on the given DV, after taking into account the other IVs.

3. The structure coefficient. The bivariate correlation of scores on a given IV with scores on the linear composite (e.g., estimated/predicted scores on the DV, given the chosen model).

(There are others, but I'll stop here.)

Because of IV overlap/redundancy/collinearity, as well as the suppressor effect, it is quite possible that all three correlations are different for a given IV. That doesn't make them wrong, or incorrect. It simply reflects that each is quantifying a different relationship.

So, as Prof. Morgan suggests, your task is to come up with a reasonable explanation for the patterns you've observed with your data set.

Good luck with your work.

Is the original positive zero-order correlation significant?

Is the negative partial coefficient significant?

Rather than removing the variable from the equation or treating the correlation as "false," you need to explain why controlling for the variable in question has this effect on the direction of the coefficient.

Yes, zero-order correlation and negative partial coefficient are both significant.

I dont know if it adds anything to the case, but ths IV in question is negatively and highly correlated with the other predictors. The correlation is significant. So, could it be because of the negative correlation between them that the suppression effect takes place in the multiple regression, but IV's coefficient is positive during linear regression with just it and the DV?

Again, thank you all for your time. This is new for me and you are helping a lot!