Method guide / Groups & conditional effects

PLS-POS

Explore segments whose PLS structural relationships differ using a prediction-oriented partitioning procedure.

16 min read

What the method does

Prediction-oriented segmentation in PLS (PLS-POS) searches for case allocations that improve the specified segment-model criterion. Unlike a prespecified multigroup analysis, the groups are discovered from the same data used to fit the models. Stability, segment size and validation on other data matter when interpreting the discovered structure.

Segment residual criterion: Σₖ Σᵢ∈Gₖ (yᵢ − ŷᵢ,ₖ)².

Gₖ is a candidate segment and ŷᵢ,ₖ the prediction under its fitted model. This expression explains the residual-based idea; use the criterion and outcome aggregation stated in the app output. Different starting allocations may lead to different solutions.

Data and analysis in QuickPLS

Use a suitable PLS model, case-level observations and a defensible segmentation plan. Check the minimum feasible segment size and variation within segments.

QuickPLS provides the partitioning settings and segment-specific results. Compare starts and stability before attaching names or explanations to the segments.

Follow a worked example

The following example shows the data, settings and decisions for one analysis. Use the linked practice files to follow it. For your own study, keep the procedure and choose the variables, model and settings that fit your research question.

Download research-workshop.csv to begin. It contains 360 rows and 31 columns. The example below identifies the variables to use, their roles and any rows to exclude; not every column is an analysis variable.

Would separate groups describe the relationships more usefully?

Suppose one Loyalty model fits some customers poorly. PLS-POS, prediction-oriented segmentation, explores groups whose separate PLS models improve a stated fitting criterion. That criterion is called the objective: the number the search tries to improve.

The groups are discovered from these data. They are different from groups defined beforehand by a recorded characteristic. An improved objective on the fitted sample does not by itself show better prediction for new customers.

QuickPLS distinguishes the published PLS-POS procedure from an extension for interaction models. In that extension, a possible move is evaluated using the scores of the segment the case would join, called the destination segment. Compare objectives only when their score definitions and scales make the comparison meaningful.

Review the search, segment sizes and stability before selecting a solution. Later uncertainty estimates are conditional on that reviewed choice; they do not independently confirm the existence of the discovered groups.

Choose the number of groups and the scoring procedure

K is the proposed number of segments. Decide which K values and which model you will investigate. Keep the outcomes and score definitions comparable across the planned candidates.

Start with a supported model and a clear reason to explore segmentation. For the first workshop exercise, use the published procedure without generated interaction terms. Keep an ordinary whole-sample model as the reference.

Open PLS-POS discovery and inspect its diagnostics

Use the saved no-interaction service model for the first exercise. Discovery produces candidate partitions for review; a lower or higher criterion under another method is not directly comparable with this objective. Return to the same setup for fixed-K inference only after reviewing the chosen partition.

  1. Open the setup

    Click Calculate…, find PLS-POS in Find a method and select it. Click Continue to setup when shown. Under Segmentation stage, select Discover candidates (never auto-select K).

  2. Choose the discovery plan

    For the first exercise, set Interaction profile to Structural model (no interaction terms) and confirm PLS-POS is selected. Choose the Candidate K checkboxes. Expand Calculation settings for Segmentation seed, PLS-POS starts and PLS-POS strict improvement tolerance. Use Validate setup, correct any highlighted issue and click Start calculation.

  3. Review before locking a solution

    In Results, open Diagnostics and warnings → Candidate segmentation diagnostics. Read it with Candidate information criteria, Candidate class or segment shares and Candidate blockers. Review the partition and its stability before returning to setup to choose Lock reviewed algorithm and K for inference.

Screenshot to be addedPLS-POS settings with the intended dataset and completed selections

Read the objective with the groups it produced

In Candidate segmentation diagnostics, check the algorithm, K, State, converged starts and stable starts. Read the objective under its declared definition and inspect the separate outcome contributions where retained. Different score scales or outcome sets can make a numerical comparison misleading.

Use Candidate class or segment shares to look for tiny groups, and Candidate blockers to understand unusable candidates. Keep the case assignments and repeated-start evidence with the saved analysis. A small improvement accompanied by unstable assignments needs a different interpretation from a recurring partition.

After a reviewed fixed-K run, find Class or segment parameters in the result outputs. The estimates and intervals concern that selected segmentation. They are not a fresh confirmatory comparison of groups defined independently before the analysis.

On a small screen, swipe the table sideways to see every column.

OutputHow to read itWhat to check
Objective / comparabilityThe fitting criterion calculated using the selected scoring procedure.Numerically different scales can make objective comparisons misleading.
Assignments and sizesThe partition produced by the selected run.Inspect tiny or unstable segments.
Segment paths and the grouping you reviewedThe relationships within the segmentation you reviewed and selected.Uncertainty estimated after discovery does not provide independent confirmation of the discovered groups.
Segment-specific R-squared values in a separate PLS-POS example. These describe the fitted exploratory segmentation, not established population groups.
Segment-specific R-squared values in a separate PLS-POS example. These describe the fitted exploratory segmentation, not established population groups.

Why the destination score definition matters

A case’s score can depend on the segment model used to construct it. The illustration below explains why moving the case is more than changing a group label, and why objectives from incompatible score definitions should not be ranked together.

Illustrative example, not a software benchmark

Moving a case can change how the destination segment's scores are constructed. Evaluating that move requires the destination-scored definition. Comparing its objective with a value calculated from fixed scores on a different scale would not establish a predictive advantage.

Explore the published procedure on the service model

Use research-workshop.csv and build reflective Quality, Trust, Satisfaction and Loyalty constructs. Assign quality_1 through quality_3 to Quality, trust_1 through trust_3 to Trust, satisfaction_1 through satisfaction_3 to Satisfaction, and loyalty_1 through loyalty_3 to Loyalty. Draw Quality → Satisfaction, Trust → Satisfaction and Satisfaction → Loyalty. Save the model and leave group and weight unused.

Begin without generated interactions. The published PLS-POS procedure improves a sum of in-sample R² values across the segments and their outcome constructs. R² describes the variation explained in a fitted outcome equation. Here, prediction-oriented names the segmentation objective; it does not mean that new cases have already been predicted successfully.

The teaching file supplies no true segmentation. You will inspect whether a useful partition recurs across starts. A stable partition is still an exploratory model result, and an unstable result is worth reporting.

Discover, review and then lock the same procedure

  1. Inspect the whole-sample model

    Check indicator coding, score directions and the structural equations. Keep the model and outcomes unchanged across the candidate segmentations. Changing the outcome set changes the objective, so the old and new values do not provide a simple improvement comparison.

    What you should see: The whole-sample model is available.

    How to read it: Keep it as the comparison for discovered segments.

  2. Choose the published discovery route

    Open Calculate… → PLS-POS and Continue to setup. Select Discover candidates (never auto-select K). Set Interaction profile to Structural model (no interaction terms) and confirm PLS-POS is selected. Choose Candidate K values 2 and 3. K = 1 remains the pooled reference; it is not a selectable segmentation solution.

    What you should see: The intended segmentation procedure is selected.

    How to read it: This is an exploratory partition, not a prespecified group test.

  3. Enter the search settings and run

    Expand Calculation settings. Set Segmentation seed to 20261105, PLS-POS starts to 10 and PLS-POS strict improvement tolerance to 1e-10. The starts field accepts 10 to 100. Use Validate setup, resolve any issue and click Start calculation. A permissible case move is evaluated through full segment-specific PLS refits, including scores and structural equations, rather than simply clustering fixed scores.

    What you should see: The search plan and seed are recorded in the completed run.

    How to read it: Compare multiple starts where the workflow permits.

  4. Read the candidate diagnostics

    Open Candidate segmentation diagnostics, Candidate information criteria, Candidate class or segment shares and Candidate blockers under Diagnostics and warnings. Review the objective histories, sizes and failed starts. At least two starts must reproduce the aligned partition and parameter solution. Each accepted move must improve the objective; a failed refit is not a completed search.

    What you should see: Candidate diagnostics identify segment sizes, fit and stability.

    How to read it: Reject an interpretation based only on a tiny high-fit segment.

  5. Lock a defensible reviewed choice

    Return to setup and select Lock reviewed algorithm and K for inference only after reviewing discovery. Choose Locked algorithm and Locked K, then tick I reviewed the discovery results and confirm this method and segment count for inference. The segmentation seed and optimisation settings remain tied to discovery. Run a new discovery to change them.

    What you should see: The selected candidate has a stated substantive rationale.

    How to read it: Preserve competing plausible solutions.

  6. Run fixed-K uncertainty if justified

    For an initial exercise, set Bootstrap replicates to 500, Confidence level (%) to 95 and Bootstrap seed to 20261106. Validate and start the calculation. Every sample repeats the relevant segmentation with K fixed and matches the segments to the reviewed solution. Inspect failures and the 90% usable requirement before interpreting the percentile intervals.

    What you should see: The fixed-K uncertainty result names the selected partition plan.

    How to read it: Report it separately from the discovery process.

Screenshot to be addedPLS-POS discovery settings

Add the contributions without calling the sum an R²

In a hypothetical two-segment solution, let Satisfaction and Loyalty R² be 0.40 and 0.50 in the first segment, and 0.60 and 0.30 in the second. The published objective is the sum: 0.40 + 0.50 + 0.60 + 0.30 = 1.80. It is unweighted, so it is neither a whole-sample R² nor 180% explained variance.

Suppose a permissible case move raises the objective to 1.81 after the required refits. The improvement is 0.01 under this definition. It does not show that the person’s true identity changed, that the grouping is causal or that predictions for new cases improved.

At K = 3, more segment contributions enter the sum. A larger raw objective therefore does not, by itself, justify preferring three groups to two. Compare the research purpose, sizes, stability and usefulness while retaining the pooled reference and every candidate result.

Suppose 500 fixed-K bootstrap samples produce 470 usable aligned results and 30 failures. The usable proportion is 94%, which meets the stated minimum. These are hypothetical numbers, not workshop output. If many otherwise converged samples cannot be uniquely matched, the uncertainty concerns segment identity; confident segment names do not resolve it.

Use instability to decide how far the conclusion can go

Start with Candidate segmentation diagnostics and Candidate blockers. A large objective, stable numerical convergence and a useful scientific grouping are different things. The checks below help distinguish them and identify when an interval or contrast should remain unavailable.

On a small screen, swipe the table sideways to see every column.

SymptomCheckReasonable action
Higher K always has a larger objectiveThe unweighted sum and additional segment contributionsAvoid treating raw objective size as an automatic best-K rule.
Starts disagree about the partitionWhether cases and coefficients agree after matching segments across starts, alongside the recorded objective valuesReview a justified larger start plan or report unstable discovery; do not keep one convenient start.
A move or start failsWhether full segment refits can be estimated uniquely, converge and meet the minimum sizeInspect the reason and retain the failure. Do not substitute an approximate score-only fit.
Interaction contrasts are unavailableCommon metric and compositional comparabilityReport local segment descriptions without the unsupported contrast.
Bootstrap segments cannot be matchedWhether segments have a unique match supported by majority overlapTreat the interval as unavailable when its usable requirement fails; matching segment numbers alone does not establish matching groups.
The discovered groups seem to predict wellWhether any cases were held out of segmentation and fittingPlan separate predictive validation before making a generalization claim.

Check whether the partition survives a meaningful change

Compare planned K values and start budgets with the data and substantive model held fixed. If a small model change produces very different assignments, explain that dependence. Even a reproducible partition may not offer a useful explanation; look for a pattern that can be tested independently.

If comparing FIMIX with PLS-POS, retain their different objectives and procedures. A cross-tabulation or adjusted Rand index describes agreement between assignments. It does not tell you which partition is true.

For later confirmation, define external group variables or an assignment rule and evaluate it on independent data. Fixed-K inference after discovery holds the chosen method and K fixed; it does not include the uncertainty from searching and selecting the segmentation.

Name the objective and keep discovery separate from validation

After the exercise, a methods paragraph could say: “We explored the published PLS-POS procedure without generated interactions for K = 2 and K = 3. We used 10 starts, segmentation seed 20261105 and strict-improvement tolerance 1e-10. Candidate moves required full segment-specific PLS refits. We reviewed objective histories, sizes and reproduction across starts before locking a candidate for fixed-K inference.” Adapt it to the actual saved settings.

For the hypothetical example, write: “The two-segment objective was 1.80, the sum of four segment-by-outcome R² contributions. We reported those contributions and segment sizes separately.” Explain that subsequent intervals are conditional on the reviewed method and K, and that discovery has not established out-of-sample predictive performance.

Report itemDetail to include
ProcedurePublished no-interaction method or destination-scored extension
DiscoveryCandidate K, starts, seed, tolerance and failed starts
ObjectiveDefinition, outcome contributions and comparability
SegmentsSizes, aligned partition stability and reviewed lock
InferenceBootstrap plan, alignment failures and common-metric requirements for contrasts

Check that you can explain the segmentation

  • Reconstruct the 1.80 objective and explain why it is not 180% explained variance.
  • Distinguish an in-sample objective improvement from a held-out prediction gain.
  • Describe the full refit needed when a case moves.
  • Identify when to use PLS-POS for interaction models and its destination-scored definition.
  • Respect a descriptive-only result when between-segment comparisons are unsupported.
  • Explain the uncertainty that remains after K is fixed for bootstrap inference.

Record the procedure as well as the selected groups

  • Name the segmentation procedure, candidate and selected K, seed and objective definition.
  • Report objective comparability, segment sizes, assignments and the result you reviewed before locking it.
  • Describe stability, the exploratory status of the groups and what later fixed-K intervals do and do not include.

Separate an improved fit from a discovered population

Compare objectives only when they use compatible score metrics, meaning comparable numerical scales and definitions. A larger number from a different metric is not evidence of a better segmentation.

Giving discovered groups appealing names does not make them independently observed populations. Keep the exploratory interpretation and plan a separate assessment before claiming generalisation to new cases.

Use the separate procedure for interaction models

An interaction allows a relationship to depend on another variable. With eligible two-way or bounded three-way interactions, use the separate destination-scored extension, labelled PLS-POS for interaction models in setup. Select the appropriate Interaction profile and confirm that algorithm. A possible move rebuilds the destination segment’s scores and products and refits the full model.

Before comparing interaction coefficients or conditional effects between segments, the scores need a common metric. The measurement review checks configural comparability, meaning compatible measurement arrangements, and compositional comparability, meaning that the scores can represent the constructs comparably.

The setup offers Request segment-to-segment interaction contrasts on the pooled common metric. Select it only with a justified measurement review. If the required checks fail, local segment estimates remain descriptive and unsupported differences are suppressed. Step 3 mean or variance differences are reported, but alone do not block standardised coefficient comparison after the required earlier stages pass.

For example, a coefficient of 0.50 on one locally constructed scale and 0.30 on another does not automatically support a comparable difference of 0.20. First establish the common scale and measurement meaning. Do not manually calculate a suppressed contrast and present it as a supported app result.

Screenshot to be addedSeparate setup for interaction models

Further reading

Related guides