How are you determining the significance of the indirect effect? Also do you have a sense of how correlated path's a and b are? The selection for testing the indirect effect and the degree to which the two indirect paths are correlated could both be factors that lead to a non-significant indirect effect, though they are not the only factors to consider. 

Thanks for your respond, Matthew.

The indirect effect ist bca bootstrapped with 5000 bootstraps. I am not sure what is meant with the correlation of a and b. Is there a way to calculate this? Else, a is the influence of a leadership style on how much you are willing to follow your leader's expectation. B is the influence of this will on your engagement. There might be a correlation to some extend, but I wouldn't expect it to be large.

Hi Paul,

seems to me like a simple power issue. The indirect effect is .11 (.30*.38) which is likely to become insignificant due to the moderate / small sample size.

In addition, although you are satisfied by the model fit, there may by some potential for a fruitful respecification. The chi-square is significant - although the N is small - hence, there may be some serious misspecifications which could be solved. Perhaps this changes the problem with the indirect effect.

For instance, you say that your IV is leadership style. Did you specify this as a singular latent variable? This could not only be a misspecification (as most leadership scales are multidimensional). Also this could out valid variance of the which could serve to predict M...

Just my 2c,

Holger

I agree with Holger that power could certainly be an issue, and your model fit indices are not "great", particularly with such a small sample ( I missed the N in your description initially). 

I would definitely be sure that your the measurement model for your latent variables is correctly specified and that your constructs are not multi-dimensional.  However, if you believe that your measurement model is correctly specified, then you may want to consider examining modification indices to determine whether the measurement model would benefit from adding error covariances. You should not add error covariances or other paths solely because a modification index suggests doing so will improve model fit, but examining modification indices can provide a useful guide for adding theoretically defensible paths to a model. Improving model fit would be one way of reducing your standard errors and increasing your overall power. 

Even if you are able to improve fit, the model may still be underpowered to detect this indirect effect. You may want to consider a power analysis, even as a post hoc means of better understanding your null effect (assuming it remains a null effect following any improvements you make to the model). 

I am not sure what software you are using. I tend to use R for my SEM models these days, and I have found that Beaujean's (2014) description of conducting power analyses to be useful in the past.

Best of luck. 

https://www.amazon.com/Latent-Variable-Modeling-Using-Step/dp/1848726996

Hi Holger, hi Matthew,

first of all, many thanks for both of your answers! I checked the phantom models and spotted a typo restraining the paths :P

It turned out that there now is an indirect effect with a CI not including 0. However, the indirect effect seems to be quite large (it's >1). I  think that this is, because it's an unstandardized effect (isn't it?) and the outcome variable is "hours per week" while the IV is measured using a likert-scale. Thus, the interpretation would make sense, at least. Please correct me if I am wrong in that point.

I also respecificated the model in that the IV now is a 2nd order construct consisting of the different dimensions of the leadership style. However, model fit didn't improve significantly.

Thanks again for your help, Paul

It depends on the software being used as each program has different defaults, but my guess is that you are indeed describing an unstandardized effect.