Prediction-Oriented Segmentation in PLS-SEM (PLS-POS)

[spiking]

I like to seek guidance on how to use the PLS-POS in SmartPLS 3.

Are there any papers published that show the exact detailed steps of using PLS-POS?

Please advise. Thank you.

[janschreier]

I found the process described here quite okay:

Becker, J.-M.; Rai, A.; Ringle, C. M.; et al. (2013): „Discovering Unobserved Heterogeneity in Structural Equation Models to Avert Validity Threats“. In: MIS Quarterly. 37 , S. 665–694.

start reading on 684 at the bottom. I don't think that there is a more detailed process because the process requires some degree of creativity when selecting which segmentation result is actually the best one. I'm currently working on segmenting my data and thus do 10 calculations for each segment number that seems reasonable. With reasonable I mean average R² over the segments (I check the average and the by segment size weighted average) and the share of segments that are above the minimum number of cases I need to detect small effects. The next step is to check which segmentation produces results that seem reasonable and that can be explained. After that I'll look for variables that can resemble the segmentation and make them accessible. The final step is to check the found variables with and MGA.

Hope this gives you and idea, but would also be grateful if someone else would join the discussion.

BR, Jan

[spiking]

I'm currently working on segmenting my data and thus do 10 calculations for each segment number that seems reasonable. With reasonable I mean average R² over the segments (I check the average and the by segment size weighted average) and the share of segments that are above the minimum number of cases I need to detect small effects. The next step is to check which segmentation produces results that seem reasonable and that can be explained. After that I'll look for variables that can resemble the segmentation and make them accessible. The final step is to check the found variables with and MGA.
BR, Jan

Thank you Jan for your feedback.

Yes this paper describes it excellently!

At Stage 2, Steps 1 & 2, I combine the heuristics of well-fitting with substantial segments. I break down into 3 separate segments (see attachment):

Screen Shot 2015-08-27 at 7.33.36 pm.png (34.52 KiB) Viewed 15754 times

Then I look at the distribution percentage, and they seem proportionate.

At Stage 2, Step 3, I run PLS-MGA on the four segments first (see attachment).

Screen Shot 2015-08-27 at 7.49.33 pm.png (49.61 KiB) Viewed 15754 times

The results have 6 outcomes:
1-2, 1-3, 1-4
2-3, 2-4
3-4

Screen Shot 2015-08-27 at 7.51.44 pm.png (38.79 KiB) Viewed 15754 times

Highlighted in red means significant effect.

Please let me know if the above steps are correct?

Thank you.

[janschreier]

Does each group fulfill all model criteria itself?

Apart from that I think what you do sounds reasonable. I'm not one of the experts only another student that tries to learn by helping others :)

[spiking]

Screen Shot 2015-08-27 at 9.01.04 pm.png (92.02 KiB) Viewed 15749 times

I am trying out this now...

[spiking]

The paper used the word segment, which is the number of pre-defined segments for which the segmentation will be performed. After running PLS-POS on SmartPLS 3, it produces 4 different groups since I selected "4 Groups".

But when I ran PLS-MGA, I had to select number of groups in Groups A and Groups B respectively. I ticked all 4 groups for each A and B.

It means that:

3 Groups = 3 Outcomes
1-2, 1-3
2-3

4 Groups = 6 Outcomes
1-2, 1-3, 1-4
2-3, 2-4
3-4

5 Groups = 10 Outcomes
1-2, 1-3, 1-4, 1-5
2-3, 2-4, 2-5
3-4, 3-5
4-5

[janschreier]

could you add the source for your citation?

[spiking]

Becker, J.-M., Rai, A., Ringle, C. M., and Völckner, F. 2013. Discovering Unobserved Heterogeneity in Structural Equation Models to Avert Validity Threats. MIS Quarterly, 37(3): 665-694.

[janschreier]

ooops, lol - that's why i sounded somewhat familiar. Time for a coffee break ;)

[spiking]

Hahaha!

I know that feeling!

My head is spinning now...

[spiking]

Screen Shot 2015-08-28 at 9.43.48 am.png (160.46 KiB) Viewed 15736 times

Request for Advice

In Stage 2, Step 1 => How to apply the selected segmentation method for a consecutive number of segments (e.g., 1 to 10)?

1. When running PLS-POS, we have to select the number of segments. Thus, if we select 4 segments, the results will produce 6 groups of comparisons or 6 pairs of comparisons.

=> Question: when it says above "consecutive number of segments (e.g., 1 to 10)", does it mean segments or groups? There is a big difference here if you refer to the table below. For example, if we select 5 segments, we will have 10 groups. Or if we select 10 segments, we will have 45 groups. Do the authors mean segment or group when applying PLS-POS?

Screen Shot 2015-08-28 at 11.01.05 am.png (60.06 KiB) Viewed 15736 times

=> Question: it does not make practical sense to run the maximum number of iterations based on twice the number of observations due to performance. It simply takes too long to run the program even though I have increased to the maximum capacity of my iMac. What is a good practical number of iterations to run then?

2. After running PLS-POS, we have to run the PLS-MGA to see if there are significant differences between segments. So for the 4 segments that I ran, it produced 6 groups (pairs) of comparisons. Out of these 6 groups (pairs), only 1 pair has most of the measures achieving p<0.05 or p>0.95. Let's call this pair 2-3. That is, Group 2 versus Group 3.

=> Question: what do we do with Group 1 and Group 4? Do we combine them together so that we are left with Group 2, Group 3, and Group 1&4 Combined? That is, left with just 3 Groups to run PLS-POS again? Again, I am confused with the choice of words here... segment or group?

Any help will be deeply appreciated!

Thank you.





2 Segments: 2 Groups
1-2
3 Segments: 3 Groups
1-2 1-3
2-3
4 Segments: 6 Groups
1-2 1-3 1-4
2-3 2-4
3-4
5 Segments: 10 Groups
1-2 1-3 1-4 1-5
2-3 2-4 2-5
3-4 3-5
4-5
6 Segments: 15 Groups
1-2 1-3 1-4 1-5 1-6
2-3 2-4 2-5 2-6
3-4 3-5 3-6
4-5 4-6
5-6
7 Segments: 21 Groups
1-2 1-3 1-4 1-5 1-6 1-7
2-3 2-4 2-5 2-6 2-7
3-4 3-5 3-6 3-7
4-5 4-6 4-7
5-6 5-7
6-7
8 Segments: 28 Groups
1-2 1-3 1-4 1-5 1-6 1-7 1-8
2-3 2-4 2-5 2-6 2-7 2-8
3-4 3-5 3-6 3-7 3-8
4-5 4-6 4-7 4-8
5-6 5-7 5-8
6-7 6-8
7-8
9 Segments: 36 Groups
1-2 1-3 1-4 1-5 1-6 1-7 1-8 1-9
2-3 2-4 2-5 2-6 2-7 2-8 2-9
3-4 3-5 3-6 3-7 3-8 3-9
4-5 4-6 4-7 4-8 4-9
5-6 5-7 5-8 5-9
6-7 6-8 6-9
7-8 7-9
8-9
10 Segments: 45 Groups
1-2 1-3 1-4 1-5 1-6 1-7 1-8 1-9 1-10
2-3 2-4 2-5 2-6 2-7 2-8 2-9 2-10
3-4 3-5 3-6 3-7 3-8 3-9 3-10
4-5 4-6 4-7 4-8 4-9 4-10
5-6 5-7 5-8 5-9 5-10
6-7 6-8 6-9 6-10
7-8 7-9 7-10
8-9 8-10
9-10

[janschreier]

I think segments and groups are synonyms. What you refer to groups are group comparisons.

Another paper worth reading is this one that describes an omnibus test to check for group differences and comparing multiple groups at once.
Sarstedt, M.; Henseler, J.; Ringle, C. (2011): „Multi-group analysis in partial least squares (PLS) path modeling: alternative methods and empirical results“. In: Sarstedt, Marko; Schwaiger, Manfred; Taylor, Charles R. (Hrsg.) Measurement and research methods in international marketing. Bingley, UK (Advances in International Marketing), S. 195–218.

the described novel confidence set approach is s.th. that I haven't seen in practice yet.

When it comes to the maximum number of iterations I have set it to 2000 and and only in 10 % of the cases I did a PLS-POS segmentation I had more than 1.000 Iterations (highest so far: 1905) but then I have 15k records. When it comes to calculation time it takes between half an hour to two days on my Core I7 (4th gen) laptop, so yes it may take some time to get results.

[spiking]

Hi Jan

Please help me understand your approach.

Is it in the following sequence?

1. Run PLS-POS from 1 Segments to 10 Segments. From each result, it is to get the group breakdowns so that we update the datafile within SmartPLS.

2. Then with the new datafile, we run PLS-MGA to check against the p values to see any significant differences between any two segments.

3. For segments which are not significant, we either combine them with other non-significant segments or we reduce the the 10 Segments in point 1.

Please kindly check the steps above and point out if I made any mistakes?

Thank you.

[janschreier]

Hi Clemen,

I would do it a bit different in 1.: run it for each segment 10 times and take that run which looks best in terms of R²s and model. This is a bit of looking for a needle in the haystack but you should run it more than once to reduce the risk of getting stuck in a local optimum (to reduce calculation time set the hook for "use presegmentation in first round" (see viewtopic.php?f=5&t=3839)).

the rest is fine except that you should not limit the number of segments you test to 10 by default but rather check if the cut off point is may be 12 or 8. Check the average R²and the segment sizes.

[spiking]

Dear Jan

If we apply your approach, how do we determine which respondents are in which segments? I'm very confused by this particular step. How do you determine what go into each segment?

My understanding is that we use PLS-POS to break up the number of segments that we need. And then use the results "group assignment" to feed the respective segments with the corresponding data.

Is my understanding correct?

Thank you.