It is certainly possible for a variable to have a negative direct effect X --> Y and at the same time a positive effect X --> M on one or more mediator variables that then in turn positively affect the outcome Y. However, in your case, the direct effect (.09) seems small (assuming you're showing standardized path/regression coefficients in the graph). Is the direct effect X --> Y statistically significant? If not, then this may simply indicate that the effect of X on Y is fully mediated (in that case, the direct effect may simply be zero in the population and negative in your sample only due to random sampling error).

Otherwise, you should ask yourself whether a negative direct effect X --> Y is substantively plausible in your study and whether all the other direct paths (X --> M and M --> Y) are also plausible in terms of their sign based on your expectations/theory.

Thank you, Christian.

Yes, coefficients are standardized.

The direct effect after mediating variables are added to the regression (-.09) is not significant. I found your explanation regarding sampling error reasonable.

I am also worried about the small coefficient between X and Y (.18) even though it is statistically significant. Should I worry about the small regression coefficient or is it enough to observe significant mediating effects (indirect paths) and a decrease in direct effect between X and Y after M variables are added?

The negative direct effect plus the positive indirect effects cancel one another out (to a certain extent) when it comes to computing the total effect (which is equal to the sum of the direct effect plus all indirect effects). This is what leads to the total effect of .18. You could set the direct effect to zero in a path analysis model and test whether the model fit (chi-square) declines significantly compared to a model in which this path is estimated. If the fit does not decline significantly (which it should not given that the coefficient is not significantly different from zero), you could drop the direct effect altogether. Then your total effect would simply be the sum of all the indirect effects.

The most interesting question in your model is probably which of the indirect effects is/are statistically significant and how strong they are (i.e., which variables are the relevant mediators vs. which are not). In addition, the R^2 values for the M and Y variables would also be relevant.

Christian Geiser your advices are helpful, but none of them address the question how it can be that the total effect (the correlation) from X on Y is positive, but the partial regression coefficient, when the mediators are included, is negative (in my opinion it doesnt matter if it is significant or not), although all bivariate correlations with the mediator are also positive.

I appears to me that this may be some form of suppression, which can occur in multiple regression. One indicator is that the sign of one predictor changes between the bivariate analysis and the multiple regression. But this is hard to tell without seeing the data. Could you provide them?

I would check all mediation analyses separately for the 4 mediators to see where the change in the sign for X happens and what may be the suppresing variable. I would wager that it occurs with M2 or M3, since they are seemingly not correlated with Y (but this are only the partial regression weights, so... you can never know).

There are some sources about suppression, e.g. in the seminal regression book by Cohen, Cohen, West & Aiken, and for example in MacKinnon, Krull, & Lockwood (2000) - Equivalence of the Mediation, Confounding and Suppression Effect.

Rainer Duesing: The total effect (.18) is not equal to the zero-order correlation between X and Y. Instead, the total effect is defined as the sum of the direct effect (-.09) plus all indirect effects (where the indirect effects are defined as the products of the path coefficients from X to M and then M to Y). Since the direct effect (partial regression coefficient) is negative (-.09) and all indirect effects are positive (the product of two positive numbers is always positive), the total effect is (-.09) + (the [positive] sum of all indirect effects). That's how you get the .18. You get it because the -.09 cancels some of the positive indirect effects. But the direct effect (-.09) is not significantly different from zero. It should be dropped from the model. Then, the total effect will be less misleading (or at least less confusing) and the point estimates and standard errors of the remaining significant path coefficients might be estimated in a more precise way also allowing for better statistical inference.

Christian Geiser it doesnt matter if you take the zero-order correlation, the standardized or the unstandardized regression coefficient. How do you assess/measure the total effect in a simple mediation model?

You are right that the direct and indirect effects sum up to the total effect, fine. But again, how do you measure the total effect for example without knowing the indirect effect? And your explanation does not explain, why the sign changes. And the argument that the effect is not significant does not explain anything in my opinion. Ceteris paribus, if the effect would represent the true population effect and you would increase the sample size, it will become significant at some point and then? Does the situation suddenly change? In my opinion not. It's not about the p-value, but the effects and the sign (or the change of it) in this case. In NHST framework you can only say that you did not find enough evidence for a significant direct effect, but it does not state that there is no change in the sign.

What if X to M is negative and X to Y (total/direct effect) is also negative but M to Y is positive? The indirect effect is negative. The overall result is that the negative effect ot X on Y is reduced partially by the positive M. It seems that the mediator can also reduce and make weaker the relationship between X and Y, not only moderators do?