2nd order constructs, 2-step procedure, and VIFs

[schroer]

Dear all,

I am estimating a model that contains two 2nd order latents. To be on the safe side, I used a 2-step procedure to handle to 2nd-order constructs:

1) I ran separate models to do the basic 2nd-order estimation (cf. Wold, 1982, pp. 40). (E.g., how do several aspects of satisfaction relate to a "super-factor" of satisfaction?)

2) I used the individual case values from step 1) as indicators in the final model. (E.g., individual values for satisfaction with colleagues, etc.).

So far, so good. However, I wonder how to interpret multicollinearity/VIF values among the manifest variables (i.e., the level 1 aspects of satisfaction in step 2)). They're far from critical, but substantial.

Wold (1982, p. 42) argues that for formative measurement models "the PLS estimation performs automatically a rotation of the three LVs at the first hierarchical level so as to be intercorrelated…Collinearity problems will arise unless the sample size is sufficiently large".

Now, I wonder, did I artifically inflate intercorrelations/multicollinearity among the level 1 constructs? If so, how to deal with it?

I'd be very grateful for your comments, ideas, or experiences with these kind of things. Thanks!

Joachim


Wold, H. (1982). Soft modeling: The basic design and some extensions. In K. G. Jöreskog & H. Wold (Eds.), Systems under indirect observation: Causality, structure, prediction. Part II (pp. 1-54). Amsterdam: North-Holland.

[viswadatta]

PLS is a non parametric method. It is not affected by multi colinearity or lack of multivariate normality. It will not have a bearing on the beta values of other paths like in ordinary least squares regression and covariance methods using lisrel software. The errors too do not influence the construct relations in PLS path modelling. So carry on. Just look for reliability, convergent validity and discriminant validity.

[yuanho]

Thank you for mentioning the two step procedure.

I have some questions of analyzing measurement model and structural model, should I combine these two models together or seperately?

There are five main construct in my model. Four of them are 2nd order construct.

At first, I use the mean of indicators for the 1st order construct to run the structural model. In this way, there is no 2nd order construct on the structural model. In fact, the indicators of my structural model are the 1st order construct on my measurement model. Is it a right way to use the mean of indicators to represent the value of 1st order construct and run the structural model?

Then I follow the repeated strategy as suggested at viewtopic.php?t=112 and viewtopic.php?t=441 and I can run both measurement and structural model at the same model. Anyway, the path coefficient of the structural model is not the same as the above structural. Some paths are siginficant at the above model and not significant at the both measurement and structural model. Which result should I use? Or Can I use the data from the both measurement and structural model to report the measurement validity , and use the data from the structural model with the composite indicators to report the path coefficient, R square, etc.?

I appreciate anyone who can solve my problem. Thanks in advance.

[viswadatta]

Use the factor scores of first order constructs instead of construct means. That will give you a more accurate result for your model.

[bwilson]

Dear all,

See: Wilson, B. and Henseler, J. (2007), Modeling Reflective Higher-Order Constructs using Three Approaches with PLS Path Modeling: A Monte Carlo Comparison.Thyne, M.; Deans, K. R., and Gnoth, J. in Australian and New Zealand Marketing Academy Conference. Department of Marketing, School of Business. University of Otago. University of Otago., New Zealand.; 791-800.

Happy reading.