PLS Prediction-oriented Segmentation (PLS-POS)
PLS prediction-oriented segmentation (PLS-POS; Becker et al., 2013) is a distance-based segmentation method for uncovering unobserved heterogeneity in a PLS path model. It follows a clustering approach with deterministic assignment of observations to groups, reassigning observations based on a distance measure rather than on distributional assumptions, which makes it suitable for path models containing both reflective and formative measurement models.
How PLS-POS Works
PLS-POS builds on earlier distance measure-based segmentation work — the PLS typological path modeling (PLS-TPM) approach (Squillacciotti, 2005) and its enhancement, the response-based detection of respondent segments in PLS (REBUS-PLS; Esposito Vinzi et al., 2008).
The PLS-POS algorithm introduces three features that distinguish it from these earlier approaches:
- It uses an explicit PLS-specific objective criterion to form homogeneous groups.
- It includes a distance measure appropriate for PLS path models with both reflective and formative measures, able to uncover unobserved heterogeneity in formative measures.
- It ensures continuous improvement of the objective criterion throughout the algorithm's iterations (a hill-climbing approach).
The segmentation objective is to form homogeneous groups of observations with increased predictive power (R² of the endogenous latent variables) in the group-specific path model estimates, compared to the overall sample model. Because the algorithm can converge on local optima, a repeated application of PLS-POS with different starting partitions is advisable.
Becker et al. (2013) and Hair et al. (2024) describe the PLS-POS method in detail. Sarstedt, Ringle, and Hair (2017) show how to combine PLS-POS with the latent class segmentation method finite mixture PLS-SEM (FIMIX-PLS) in a tandem approach (see also Hair et al., 2024, Chapter 6). Sarstedt et al. (2022) offer a broader review of latent class analysis in PLS-SEM.
PLS-POS Settings in SmartPLS
| Setting | What it controls |
|---|---|
| Number of Segments | The number of pre-defined segments for which the segmentation will be performed. |
| Maximum Iterations | The maximum number of iterations the segmentation algorithm will perform. Should be sufficiently high for a good segmentation solution. |
| Search Depth | The maximum number of observations, from the sorted list of candidates for reassignment, that will be tested for whether they improve the PLS-POS objective criterion. May not exceed the sample size. A reduced search depth can be used for performance in early exploratory stages, but the final segmentation result should use the maximum number of observations to ensure the solution minimizes the objective criterion. |
| Initial Separation | The initial split of data into the pre-specified number of groups, based on either a Random Assignment or a prior FIMIX segmentation solution. If FIMIX Segmentation is chosen, the necessary FIMIX-PLS settings must also be specified in a separate settings tab. |
| Pre-Segmentation | If selected, the algorithm performs a pre-segmentation round that assigns all units to their best-fitting group according to the distance measure, without checking whether this improves the objective criterion. |
Optimization Criterion
The optimization criterion (also objective criterion) will be optimized when estimating the segments in the PLS-POS algorithm. There are two options:
- Sum of All Construct R-Squares: Uses the sum of all R-Squares in the model for all segments as the PLS-POS objective criterion that will be optimized (maximized) when reassigning observations in the course of the segmentation.
- Sum of Target Construct R-Square: Uses the target constructs sum of R-square values over all segments as the PLS-POS objective criterion that will be optimized (maximized) when reassigning observations in the course of the segmentation.
- Sum of All Construct Weighted R-Squares: Uses the sum of all weighted R-Squares in the model for all segments as the PLS-POS objective criterion that will be optimized (maximized) when reassigning observations in the course of the segmentation The weighting of the R-Squares is done by using the relative segment sizes.
- Sum of Target Construct Weighted R-Square: Uses the target constructs sum of weighted R-square values over all segments as the PLS-POS objective criterion that will be optimized (maximized) when reassigning observations in the course of the segmentation. The weighting of the R-Squares is done by using the relative segment sizes.
Target Construct
If Optimization Criterion is Sum of all Construct R-Squares or "Sum of All Construct Weighted R-Squares*, then this option does not have to be specified.
If Optimization Criterion is Sum of Target Construct R-Square or Sum of Target Construct Weighted R-Square, then this option defines the target construct for which the outer residuals or R-square value is calculated.
Frequently Asked Questions
What makes PLS-POS different from other segmentation approaches?
PLS-POS uses an explicit, PLS-specific objective criterion and a distance measure that also works for formative measurement models, and it guarantees continuous improvement of that criterion across iterations (a hill-climbing approach) rather than relying on distributional assumptions.
How many segments should I choose?
The number of segments is a setting you specify before running PLS-POS. Since the algorithm optimizes within the number of segments you request, researchers typically compare solutions across a range of plausible segment counts and evaluate which yields the most interpretable, theoretically meaningful groups.
Should I use Random Assignment or FIMIX Segmentation for the initial separation?
Either can be used as the starting partition. Because PLS-POS can converge on local optima, running the algorithm repeatedly from different starting partitions (including both random and FIMIX-based starts) is advisable to check the stability of the resulting segmentation.
Can PLS-POS be combined with FIMIX-PLS?
Yes. Sarstedt, Ringle, and Hair (2017) show how to use FIMIX-PLS and PLS-POS together in a tandem approach, using each method's strengths to cross-validate the segmentation solution.
What does the Search Depth setting affect?
It caps how many candidate observations are tested for reassignment in each iteration. Lower values speed up exploratory runs, but the final segmentation solution should use the maximum search depth (equal to the sample size) to ensure the result actually minimizes the objective criterion.
Related SmartPLS Methods
- Finite Mixture PLS-SEM (FIMIX-PLS)
- Importance-Performance Map Analysis (IPMA)
- PLSpredict
- Multigroup Analysis (MGA)
References
- Becker, J.-M., Rai, A., Ringle, C. M., & Völckner, F. (2013). Discovering unobserved heterogeneity in structural equation models to avert validity threats. MIS Quarterly, 37(3), 665–694.
- Hair, J. F., Sarstedt, M., Ringle, C. M., & Gudergan, S. P. (2024). Advanced issues in partial least squares structural equation modeling (PLS-SEM) (2nd ed.). Sage.
- Esposito Vinzi, V., Trinchera, L., Squillacciotti, S., & Tenenhaus, M. (2008). REBUS-PLS: A response-based procedure for detecting unit segments in PLS path modelling. Applied Stochastic Models in Business & Industry, 24(5), 439–458.
- Sarstedt, M., Radomir, L., Moisescu, O. I., & Ringle, C. M. (2022). Latent class analysis in PLS-SEM: A review and recommendations for future applications. Journal of Business Research, 138, 398–407.
- Sarstedt, M., Ringle, C. M., & Hair, J. F. (2017). Treating unobserved heterogeneity in PLS-SEM: A multi-method approach. In R. Noonan & H. Latan (Eds.), Partial least squares structural equation modeling: Basic concepts, methodological issues and applications (pp. 197–217). Springer.
- Squillacciotti, S. (2005). Prediction oriented classification in PLS path modeling. In T. Aluja, J. Casanovas, V. Esposito Vinzi, & M. Tenenhaus (Eds.), PLS & marketing: Proceedings of the 4th International Symposium on PLS and Related Methods (pp. 499–506). DECISIA.
- More literature ...
Cite correctly
Please always cite the use of SmartPLS!
Ringle, Christian M., Wende, Sven, & Becker, Jan-Michael. (2024). SmartPLS 4. Bönningstedt: SmartPLS. Retrieved from https://www.smartpls.com
