QuickPLS Start with what the items are meant to measure
Cheah and colleagues explain consistent partial least squares, usually shortened to PLSc, using synthetic employee-retention data. Their methods paper compares several estimators and demonstrates corrected measurement and structural results. It is a methods reference, not evidence that the authors used QuickPLS.
In QuickPLS you will create the employee-learning model, save an ordinary PLS result, then run Consistent PLS on the same cases and indicators. You will inspect the correction before comparing paths and use PLSc Consistent Bootstrapping for corrected uncertainty. The new teaching data practise the paper’s estimator comparison without reproducing its data or coefficients.
Your practice study: support, confidence and continued learning
Imagine a professional-development team studying whether employees who feel supported also feel more confident about learning, and whether that confidence is associated with plans to continue. Each row represents one fictional employee. The 320 cases are a convenient teaching dataset, not a sample-size recommendation.
Create three reflective constructs: Support, Confidence and Continuance. Reflective means that the underlying idea is expected to produce similar responses to several questions. Three manifestations of confidence fit this interpretation better than three unrelated training resources.
Draw Support → Confidence, Confidence → Continuance and Support → Continuance. The last arrow allows a remaining direct association alongside the indirect route through Confidence. Decide that model before seeing the estimates. The file contains complete, continuous teaching scores between 1 and 7; it contains no actual employee responses.
On a small screen, swipe the table sideways to see every column.
| Construct | Columns | Role |
|---|---|---|
| Support | support1, support2, support3 | Reflective predictor |
| Confidence | confidence1, confidence2, confidence3 | Reflective mediator |
| Continuance | continue1, continue2, continue3 | Reflective outcome |
| Row label | case_id | Identifier; do not assign as an indicator |
Step 1. Keep the measurement interpretation separate from the calculation
With a real survey, check the wording before reversing a column. On a genuine 1–7 response scale, a reverse-coded value would be 8 minus the original response. That transformation needs an item whose meaning actually runs in the opposite direction. A negative coefficient alone is not a reason to reverse anything.
The practice file is complete. Casewise deletion therefore removes no rows. If you later substitute your own incomplete survey, record which cases were excluded and why. Comparing estimators on different samples makes their differences harder to explain.
Import and verify the employee data
Download the CSV and codebook. Create a writable project and choose File → Import Data… → Choose data file…. In Review & import check raw data, Header row 1, comma separation and Decimal convention Dot (1.25). Keep numeric measurement columns Continuous. Choose Import dataset and confirm Dataset imported and verified with 320 cases and nine measurement columns. Keep case_id outside the model. All items point in the same direction; do not reverse them.
Step 2. Build the same model for both estimators
Create the model
With the imported dataset selected, choose File → New Model…. Enter Employee learning, check Dataset version and the 320-case count, then click Create model.
Create and name the blocks
In Indicators, select support1, support2 and support3 with Ctrl, then choose Create construct. Select the new construct, open Properties → Model, enter Support as Name and press Enter. Repeat for Confidence and Continuance using their own three columns.
Set measurement and arrows
For each construct, open Properties → Parameter → Measurement model and choose Reflective. Select Connect and drag from each source construct to its destination to draw the three stated arrows. Return to Select and save.
Keep an ordinary PLS comparison
Choose Calculate… → PLS-SEM Algorithm. Use Starting weights → Standard (+1) and Missing data → Treatment → Casewise deletion. Review Readiness and choose Start calculation. Save this result as the ordinary PLS comparison; do not attach its uncertainty to PLSc.
Step 3. Run the correction and inspect it before the paths
Return to the same saved model and choose Calculate…. Search for Consistent PLS in Find a method. Set Weighting scheme to Path weighting, Result data to Standardized and Missing data to Casewise deletion. In Advanced settings, record Maximum iterations and Stop criterion as displayed. Review Readiness and click Start calculation.
Open Reliability and validity → PLSc correction reliability. Read rho_A for each block. This estimated reliability is used in the correction; it is not a percentage of truthful answers. Then open Discriminant validity → PLSc construct correlations. Original and Corrected are two columns in the same table.
Check the warnings before interpreting a larger corrected coefficient. The corrected correlations must describe a mathematically possible set of relationships. A correlation above 1 cannot be made acceptable by rounding or clipping it.
Confirm that Consistent PLS is running on the saved Employee learning model. Keep the ordinary PLS result available for comparison, and record the actual corrected result name before opening the reliability and correlation tables.
Check the corrected structural relationships
Use the reflective model in this lesson with consistent PLS, path weighting, standardised indicators, starting weights +1, a 3,000-iteration limit and stopping criterion 0.0000001. Read the corrected path table rather than borrowing ordinary PLS estimates.
The uncertainty column uses 1,000 usable corrected fits out of 1,000 requested, with R seed 202611071 and percentile endpoints. Both programs fitted the same sampled rows. This was a separate repeated-fit exercise, not a run of the app’s Bootstrapping or BCa calculation.
About these answers: Corrected paths, R² and weights agreed between QuickPLS calculation code and cSEM. The check does not establish equal loadings and was not a new desktop-app run.
On a small screen, swipe the table sideways to see every column.
| Corrected path | Point estimate | Common-row 95% percentile interval |
|---|---|---|
| Support → Confidence | 0.610 | 0.533 to 0.674 |
| Confidence → Continuance | 0.499 | 0.382 to 0.608 |
| Support → Continuance | 0.168 | 0.047 to 0.284 |
What agreed and what differed
The 14 comparisons of original-data paths, R² and weights and 3,000 path comparisons from usable repeated fits were all within 0.00000001. The loading columns differ between the two programs. The download keeps both the raw and corrected values so their definitions can be examined; equal loadings have not been demonstrated.
There were 0 unusable QuickPLS repeated fits. Every fit completed using the sampled rows shared with the reference.
Do not clip an invalid reliability estimate or substitute an ordinary PLS fit for a failed corrected fit. Keep requested, usable and failed counts with the result. This answer checks the named structural quantities, not every PLSc output.
Step 4. Why a corrected relationship may be larger
Imperfect measurement can weaken the relationship between scores formed from items. PLSc uses a reliability-based correction under its factor assumptions. That does not mean every corrected path increases by the same amount: when an outcome has several predictors, the whole correlation pattern matters.
Read corrected loadings, paths and R-square under Estimates after checking reliability and corrected correlations. A loading connects an item with its factor. A path describes a modelled association between constructs. R-square is the share of fitted outcome variation explained by the specified predictors. None of these is a general guarantee that the model is right.
Hypothetical example
Illustrative reading example
Suppose a score correlation is 0.48 and the two reliability estimates are 0.80 and 0.72. Dividing 0.48 by the square root of 0.80 × 0.72 gives approximately 0.6325. This explains one correlation correction. It is not an expected result from the practice CSV and is not a formula for multiplying every path.
Step 5. Give corrected estimates corrected confidence intervals
Choose Calculate… → PLSc Consistent Bootstrapping for the same model. Match Path weighting and Standardized results. For this teaching run enter 1000 Bootstrap samples, Confidence level (%) 95, Hypothesis test → Two-sided and Confidence intervals → Percentile and BCa. In Advanced settings enter Seed 202611071 and record the other settings. Leave the optional studentized calculation off for this first exercise.
The bootstrap repeatedly samples employee rows with replacement and refits the corrected model. Some corrected refits can fail even if the original correction was usable. Read attempted, usable and failed counts, then the availability message for the particular interval. A missing BCa interval is not a zero effect; BCa needs additional leave-one-case-out calculations.
For the indirect Support → Confidence → Continuance route, use its own effect-specific interval. Multiplying the lower limits of two path intervals does not give the lower limit of their product. More resamples may reduce simulation noise, but they cannot repair an invalid measurement correction.
Select the completed PLSc bootstrap result for Employee learning. Check the interval method and the usable corrected-fit count before copying a path or indirect-effect interval into your comparison table.
Hypothetical example
Illustrative reading example
A corrected path of 0.36 with a 95% interval of [0.11, 0.58] would support a positive association under this model. If 950 of 1000 attempted corrected fits were usable, report the 50 failures and their reasons as well. These invented values describe how to read a result, not what the download will produce.
When the correction makes a problem visible
On a small screen, swipe the table sideways to see every column.
| What you notice | What to inspect | What to do |
|---|---|---|
| A corrected correlation exceeds 1 | Reliability, repeated item assignments and overlap between concepts | Investigate the measurement model; retain the failed result rather than clipping it. |
| rho_A looks implausible | Item direction, variation and block meaning | Correct documented coding errors; reconsider whether the factor interpretation is defensible. |
| PLSc and PLS paths differ | Same cases, indicators, paths and result scale | Explain the different estimation targets before judging agreement. |
| Many bootstrap corrections fail | Original correction warnings and failed-refit reasons | Report failure counts and investigate instability before relying on the interval. |
| A path interval includes zero | Its width and the size of effects still compatible with the data | Describe uncertainty; do not switch estimators solely to obtain significance. |
Step 6. Make one comparison that answers a real question
Hold the sample, indicators and arrows fixed when comparing ordinary PLS with PLSc. First compare the original and corrected construct correlations, then the named paths. Keep separate columns for each estimator and use each estimator’s own intervals. This isolates the effect of changing the estimator.
If an item seems to measure something different, assess its content before changing the block. Save any justified revision as a separate model. Explain which part of Support or Confidence the revised measure loses. A higher reliability statistic can accompany a narrower and less useful construct.
Do not treat the synthetic factor design as evidence that your own questionnaire has the same structure. In a real study, item development, content review and the sampling design remain part of the argument. The calculation cannot supply them after data collection.
Write an explanation another researcher can check
Methods outline: We treated Support, Confidence and Continuance as reflective common factors, each measured by [items]. We estimated the stated paths with Consistent PLS using [weighting, scale and missing-data handling]. We inspected correction reliability, corrected construct correlations and warnings before interpreting structural results. We obtained [interval method] intervals from [attempted] corrected bootstrap fits, of which [usable] were usable.
Results outline: The corrected Support → Confidence association was [estimate], with a [level] interval of [lower, upper]. The specific indirect association through Confidence was [estimate and interval]. The corresponding ordinary PLS analysis is reported separately. [Describe correction warnings and failed fits.] Replace every placeholder from the saved run, not from the illustrative numbers above.
| Keep in the report | Why it matters |
|---|---|
| Factor rationale and indicators | Shows why a consistency correction was chosen. |
| Original and corrected correlations | Makes the operation and any validity concern inspectable. |
| Corrected paths with matching intervals | Avoids combining different estimators. |
| Bootstrap counts and interval names | Shows what uncertainty was actually calculated. |
| Ordinary PLS sensitivity result | Explains differences without selecting the most convenient answer. |
Report corrected relationships with their own uncertainty
Name consistent PLS and justify the common-factor interpretation of the reflective blocks. Report the corrected estimate and an interval obtained from the same corrected method. Ordinary PLS uncertainty cannot be attached to a PLSc point estimate.
This teaching check requested 1,000 shared-row refits and retained 1,000 usable corrected fits. The displayed percentile intervals describe that reference exercise. Use the actual settings and counts from your own analysis when writing a research paper.
Finish when you can explain these decisions
- Show why each block is reflective rather than a list of ingredients.
- Find Original and Corrected in the correlation table and explain one changed value.
- Explain why a correlation above 1 must be investigated.
- Identify the estimator that produced every reported interval.
- Save the original model, both point calculations, the corrected bootstrap and your interpretation notes.