Negative weights and loadings. 

Hi Wynne,

In one of my models (20 latent variables and 50 manifest variables, N = 120) for my PHD dissertation (marketing), some constructs have high negative loadings. I don't know why and I would like to explain these results in my dissertation. Have you an explanation for this?

Thanks for your help, 


Here's one possibility:

Currently, the PLS-Graph or the standard PLS algorithm software does not check for the sign of the weights. So among analysis the weights estimated to create the underlying construct scores may be positive or negative. If it is negative, it is equivalent to reverse coding the construct. As a result, the loading between this reverse coded construct and the original indicators will become negative.

So, a key thing is to run all your models first and focus primarily on magnitude. Once you settle on the final model, look at the weights to see if any of the constructs are reverse coded. In turn, the paths among constructs need to be interpreted accordingly. So an expected positive relationship between two constructs will be negative if one of them is reversed coded (but positive if all weights for both constructs are all positive or all negative).