Consistent PLS algorithm and outer loadings (form. cons.)

[DHoppe]

Dear all,

I have a research model with one reflectively and two formatively measured constructs. The formatively measured constructs had been measured reflectively on the lower dimensions (but these dimensions were eliminated by applying the two-stage approach to obtain latent factor scores).

See image for details.

As the new paper from Dijkstra and Henseler (Consistent Partial Least Square Path Modeling) suggests, I'm applying the PLS (c) algorithm to consider reflective measurement model correlations for more consistent parameter estimation. However, application of the PLS (c) algorithm renders outer loadings of formatively measured constructs 0.0000. As I understand the routine, the PLS (c) algorithm should only perform a correction of reflective construct's correlations. However, it seems that it does infer with formative measures.

Applying the "standard" PLS algorithm yields plausible outer loadings for formatively measured constructs.

Can you offer any advice? Thanks in advance.

[jmbecker]

If you install the latest version (3.2.1) you will get the loadings. This has been a small bug in prior versions which however, does not interfere with the other results. It is simply just not showing the correct values for the loadings of formative constructs.
And you are correct, that PLSc does not perform any correction on the loading of formative constructs (single-order correlations between the indicators and the composite).

[DHoppe]

Thank you for your quick and helpful reply.

I would like to ask another question related to the application of the PLSc algorithm.

As I want to conduct multi-group-analysis for moderator analysis, I was wondering whether the "standard" PLS or the PLSc alrogithm will be used within the MGA procedure. Looking at different results when perfoming data-group specific PLSc computations and data-group comparisons with MGA, I suppose that the MGA uses the "standard" PLS algorithm for calculations.

Is there a possiblity to use the PLSc algorithm when applying the MGA? This is important because I would like to assess statistical differences using the PLS-MGA rather than the parametric approach (which could be easily done manually with the consistent bootstrapping routine for the data-groups). Any insight is highly appreciated.

[jmbecker]

We have not yet implemented the MGA for PLSc. This is definitely on our to-do list as PLSc gains popularity.
Until then, you have to do it yourself using PLSc bootstrapping results, which is all your need to do the necessary calculations.

[krismurphy01]

Hello,

I have been testing the consistent PLS algorithm and have made a few observations. Before I begin note that the program is running on a new Mac – very exciting. He is where I am having challenges:

Bootstrapping with 5000 samples
1. When I run the bootstrapping the first time I get ? marks on many paths – I start to think maybe my data is suspect.
2. When I run the bootstraps the second time I get all data and related significance levels

Bootstrapping with 6000 samples
1. First time I get the same ? marks
2. The second time I get the same ? marks but the numbers change

The 6000 samples takes 30 minutes to run versus the 5000 samples which take about 2 minutes.

Bootstrapping with 10,000 samples – Java Heap error

Is there a setting or configuration that can improve consistency? This problem can be replicated and each run results in a different p or t score.[/i]


When I run the SEM not using consistent PLS all relevant values are significant. Based on the recent paper I believe we should use the consistent PLS with the option connecting all latent variables.

[jmbecker]

This is an inherent problem of the consistent PLS algorithm. It sometimes produces inadmissable solutions.

Please read these related discussions:
viewtopic.php?f=5&t=3531
viewtopic.php?f=3&t=3766