Path Coeff. of total Group bigger than separated Group PathC

[RobertusMaximus]

Hi,

i´m actually calculating my model with PLS and getting results, that seem on first sight not valid to me.

What happens:

I have a total observation Group of 1200 participants. This Group contains Customers (n=640) and Non-Customers (n=560).
When I Calculate my path coefficients (PC) for the relationship A->B, I get a PC of 0,247 for the total group.
When I subdivide the groups and run separate regressions for Customers and Non-Costumers I get the following PCs for relationship A-->B:

Customers: 0,10
Non-Customes: 0,23
(Complete Group of Customers and Non-Customers: 0,247)

For me, it makes no sense and seems unclear, that the PCs of the subdivided groups are smaller compared to the PC of the complete group (which is actually 0,247).

Is it possible to have PCs like that? Is it due to the calculation behind it?

The dataset is valid and I double checked it already. Customers + Non-Customers= Complete Group.

Thanks for your help and Knowledge in advance!


Cheers
Robert

[RobertusMaximus]

...mhhhh....no one with an idea????

[jmbecker]

It is not surprising what you find.
In brief, PLS estimates outer weights for the indicators of your constructs (which will be represented as weighted composites) and then estimates structural paths from these composites. There will be some iterations of outer weights and structural paths estimation until the outer weights converge and the final inner paths are estimated.

Because the weights can be different between groups, the structural paths must not adhere to the logic that they are basically a weighted average of the path coefficients of the groups. This would only be the case if the weights were equal between groups. That’s why you should test for measurement invariance in a multi group analysis.

In addition, the collinearity of the (exogenous) variables within groups could be different to the collienarity of the combined sample. This could also lead to different forms of "supression effects".

[RobertusMaximus]

Ok. Perfect!!

Thanks for your helpful reply!!