The research behind this exercise
Wang and McCarthy studied 345 Chinese consumers in a hotel-review context using PLS models with personality and collectivism interactions. They analysed positive and negative bias separately and included additional robustness analyses. This original exercise teaches careful interaction reading with new observed willingness scores. It does not reproduce the source’s bias construction or claim that all of its interaction hypotheses were supported.
In QuickPLS you will fit the POS regression with LEAD, COLL and their interaction, then inspect the conditional slopes. A separately saved NEG analysis is exploratory in this exercise. These original observed willingness scores adapt the paper’s moderation question; they do not reproduce its latent model or bias measures.
The question you will answer
A traveller who likes to guide others may be willing to post a favourable review, a critical review, or both. Does an individual’s collectivist orientation change the relationship between that guiding tendency and favourable-review willingness?
POS and NEG are separate continuous willingness scores, not two ends of a single rating. Higher values mean greater willingness to write that kind of review. COLL is an individual response score, not a nationality code. These teaching outcomes are not the published paper’s derived bias measures or binary robustness outcomes.
Define bias relative to a justified reference before using that word in research. A favourable review is not automatically biased, and a critical review is not automatically unfair. This exercise avoids that assumption by studying willingness directly. A real bias measure would need its reference, coding and possible floor effects explained.
Step 1. Download the original practice file
The download has 320 complete fictional records: one fictional hotel guest per row. Its continuous values sit within a 1–7 teaching response range; they are not rounded survey categories. The number of rows makes the exercise convenient to practise. It is not a sample-size recommendation for a real study.
Open the codebook before importing. Check the meanings and direction of the columns, then confirm the row count and numeric values in QuickPLS. Keep case_id only as a row identifier. It is not an indicator, predictor or control. These files have no planned missing entries, so an unexpected blank calls for an import check.
For your own study, distinguish a missing response from a genuine zero and document exclusions before analysis. Reverse a question only when its wording and scoring require it. The practice columns are already oriented as described below.
After import, confirm that LEAD, COLL, POS and NEG are separate numeric columns. Keep POS and NEG as different outcomes rather than reverse-coding one to create the other.
Open the import
Open a writable project and choose File → Import Data… or Ctrl+I. In Import dataset click Choose data file… and select this lesson’s CSV.
Review the preview
The dialog moves from Choose file to Review & import. Set Input represents to raw data, Header row to 1, separator to comma and Decimal convention to Dot (1.25). Confirm the teaching columns are Continuous. Keep case_id out of the analysis.
Complete the import
Check that the preview separates the columns correctly, then click Import dataset. Wait for Dataset imported and verified. Open the imported data and confirm the expected case count before building the analysis.
On a small screen, swipe the table sideways to see every column.
| Idea or role | Columns | How this exercise uses them |
|---|---|---|
| Tendency to guide others | LEAD | Observed focal predictor |
| Individual collectivist orientation | COLL | Observed moderator |
| Favourable-review willingness | POS | First observed outcome |
| Critical-review willingness | NEG | Separate follow-up outcome |
Step 2. Use observed scores deliberately
This exercise treats each selected column as an observed score. There are no latent-factor measurement blocks to draw and no automatic correction for measurement error. In actual research, explain how a score was obtained, what higher values mean and whether its reliability is adequate for the question. Giving a column a psychological name does not validate it.
Inspect the range and variation of the predictor, moderator and outcome before fitting. A moderator is the variable whose level may change a relationship. It should not be chosen after trying many columns simply because one interaction looks promising. Keep the main effects: the individual predictor and moderator relationships belong alongside their product.
The file is complete and uses continuous teaching scores. Keep the same complete cases across any planned comparison. If real data contain missing values, record how the shared analysis sample was obtained. Using different respondents for the baseline and interaction model can make an apparent change difficult to interpret.
Step 3. Enter the interaction in the observed-variable workflow
Choose the method and roles
Select the imported dataset, open Calculate and choose Regression under Standalone analysis. Set Regression type to Graph-defined Path Analysis / PROCESS. Under Graph-defined relationships set Outcome to POS and Focal predictor to LEAD.
Enter the direct relationship
Under Directed paths use Add path for LEAD → POS. Keep the intended outcome and source roles visible in the Relationship preview. This is an observed-variable equation, not a latent PLS diagram.
Add the moderator
Use Add moderator and choose COLL. Use Moderate a path to identify LEAD → POS as the relationship being moderated. Keep COLL out of Controls when assigning it the moderator role. Confirm that the preview retains both main effects and the interaction.
Check scale and cases
Review the data profile, sample and intercept. Keep raw units for this teaching model and record Final variable metric if using the extended configuration. Retain the intercept, which defines the baseline of the fitted equation.
Request uncertainty
Under Bootstrap inference choose Case-resampling bootstrap and set Bootstrap samples to 1000, Parallel workers to 1 and Seed to 261107116. Choose Configure PROCESS inference, set Confidence level (%) to 95 and choose a two-sided Test direction. These are practice settings, not claimed settings from the paper.
Calculate and read the complete equation
Start the calculation and select the POS result. Use Search result outputs to open Equation coefficients and find the LEAD × COLL term. Open Simple slopes and conditional plots for the LEAD slope at each reported COLL value. Read resampling intervals in Bootstrap inference and counts in Bootstrap summary, matching the effect label and scale.
What you should see: The selected output refers to POS and the intended COLL probes, with the matching interaction and uncertainty.
Check this 320-case observed-score analysis
Use the exact equations below, all 320 complete cases, raw units and intercepts. The reference uses 1,000 whole-row resamples, R seed 261107116, and 95% percentile intervals with quantile type 7. It preserves the joint variation of every equation in the system.
The POS interaction estimate is 0.135 with a reference interval of 0.040 to 0.239. The optional NEG interaction is 0.027 with an interval of −0.073 to 0.123. The NEG run is exploratory here; its pointwise intervals are not adjusted for a family of tests. Different significance labels alone are not a test that the two interactions differ.
About these answers: Reference answer calculated in R and independently repeated with NumPy.
| Check | Answer |
|---|---|
| Cases | 320 |
| Requested resamples | 1000 |
| Reference seed | 261107116 |
| Independent numerical comparisons | 14048 |
What was checked
NumPy independently recalculated every regression coefficient, indirect effect, conditional effect and saved resampled fit using least squares. All 14,048 comparisons were within 0.00000001. This checks the answers calculated in R; it does not test the desktop app’s PROCESS or Bootstrapping command. To reproduce the reference exactly, use its saved sampled rows: the same seed can select different rows in different programs.
Read the positive model
Fit this equation with an intercept and the original observed units: POS = LEAD * COLL. Here the R formula shorthand LEAD * COLL includes LEAD, COLL and their product, each with its own coefficient; it does not mean regressing POS on the product alone.
The fixed original-sample moderator probes are 2.915, 4.004, 5.093 (mean minus one sample standard deviation, mean, mean plus one sample standard deviation). They are raw scores, not standardised −1, 0 and 1.
About these answers: Reference answer calculated in R and independently repeated with NumPy.
On a small screen, swipe the table sideways to see every column.
| Equation / quantity | Estimate in raw units | 95% reference percentile interval |
|---|---|---|
| POS/intercept | 4.121 | 2.539 to 5.799 |
| POS/LEAD | -0.295 | -0.705 to 0.089 |
| POS/COLL | -0.276 | -0.692 to 0.104 |
| POS/LEAD:COLL | 0.135 | 0.040 to 0.239 |
| Slope at low | 0.099 | -0.051 to 0.250 |
| Slope at mean | 0.247 | 0.138 to 0.348 |
| Slope at high | 0.394 | 0.238 to 0.548 |
Equation fit and refit counts
POS R² = 0.151
1,000 resamples requested; 1,000 usable. Every equation is refitted on the same resampled rows. No failed refit has been silently removed.
Read the negative model
Fit this equation with an intercept and the original observed units: NEG = LEAD * COLL. Here the R formula shorthand LEAD * COLL includes LEAD, COLL and their product, each with its own coefficient; it does not mean regressing NEG on the product alone.
The fixed original-sample moderator probes are 2.915, 4.004, 5.093 (mean minus one sample standard deviation, mean, mean plus one sample standard deviation). They are raw scores, not standardised −1, 0 and 1.
This is the separately saved exploratory negative-review outcome. It is not part of the primary POS equation and has no multiplicity-adjusted inference in this answer pack.
About these answers: Reference answer calculated in R and independently repeated with NumPy.
On a small screen, swipe the table sideways to see every column.
| Equation / quantity | Estimate in raw units | 95% reference percentile interval |
|---|---|---|
| NEG/intercept | 4.133 | 2.338 to 6.058 |
| NEG/LEAD | 0.151 | -0.272 to 0.576 |
| NEG/COLL | -0.291 | -0.718 to 0.146 |
| NEG/LEAD:COLL | 0.027 | -0.073 to 0.123 |
| Slope at low | 0.228 | 0.065 to 0.390 |
| Slope at mean | 0.257 | 0.161 to 0.363 |
| Slope at high | 0.286 | 0.140 to 0.425 |
Equation fit and refit counts
NEG R² = 0.106
1,000 resamples requested; 1,000 usable. Every equation is refitted on the same resampled rows. No failed refit has been silently removed.
Step 4. Read the relationship at stated moderator values
The interaction coefficient describes how the predictor slope changes as the moderator changes. The predictor main effect describes the slope when the moderator equals zero on the fitted scale. If zero is outside the observed range, that main-effect reference may be far from an ordinary case. Do not mistake it for the relationship for an average participant.
For a continuous moderator, this workflow reports probes generated from the analysed sample: its mean and one sample standard deviation below and above that mean. Standard deviation describes spread. For a binary moderator the probes are 0 and 1. Read the actual reported values and their support in the data rather than assuming the screen accepts arbitrary custom probes.
A significant interaction does not mean that every conditional slope is positive or distinguishable from zero. Likewise, one slope being significant and another not does not itself establish their difference. Use the interaction or a specifically supplied contrast with its own uncertainty. Keep the original outcome units and the moderator values in the explanation.
Step 5. Inspect the equation before telling the story
An interaction can be difficult to estimate when the predictor and moderator cover only a narrow range, overlap strongly, or have a few unusual combinations. Inspect those cases for genuine coding problems and consider whether the proposed linear relationship is appropriate. Do not remove an unusual but valid person merely because their response weakens the interaction.
Read usable and failed bootstrap refits before quoting a narrow interval. A requested count is not a guarantee that every equation fitted successfully. Keep the same variables, coding and analysis sample when comparing the baseline with the interaction model. A revised specification answers a revised question and should be saved separately.
Step 6. Keep praise and criticism as separate outcomes
Fit the planned POS model first. NEG is available for a separately named follow-up analysis; do not include it as a predictor simply because it appears in the file. A person can be willing to praise a good stay and criticise a poor one, so the two responses need not sum to a constant.
If you planned to study NEG as well, save a separate specification, change Outcome to NEG, retain LEAD as Focal predictor and moderate LEAD → NEG with COLL. Check the new Relationship preview and case count before calculating. Report both planned analyses and the approach used to handle the resulting family of tests. Do not silently select whichever outcome gives the smaller probability.
Read the table more carefully than the headline
An interaction needs its own estimate and uncertainty. A broad story about collective values cannot replace the row for the actual outcome. If favourable and critical willingness show different patterns, describe both without claiming that one difference is established merely because one test passes and the other does not.
Hypothetical example
Illustrative reading example
Suppose the LEAD-by-COLL interaction for POS is 0.10 with a 95% interval from -0.01 to 0.21. That interval includes zero, so the evidence is uncertain under that procedure. A second invented NEG interaction of -0.02 with interval -0.13 to 0.09 is also uncertain. Neither result supports a sweeping claim that collectivism suppresses criticism. These are invented examples.
Interpret the moderator at the level measured
COLL describes differences among the individuals in this exercise. It does not compare nations and does not establish that all members of a culture behave alike. Within-country variation can be important; nationality cannot be substituted for the measured score without changing the question.
For a real hotel study, record the stay context and whether a review was actually posted. Willingness may differ from behaviour when guests face account requirements, time costs or fear of identification. A stronger follow-up could observe the defined action with suitable consent, while keeping genuine criticism distinct from measurement error.
What to check when the result is unexpected
If positive and negative outcomes appear exact opposites, check whether the instrument forced a single choice or whether a coding rule created that dependence. Separate constructs and a single bipolar scale require different interpretations.
If an interaction appears only under one unplanned threshold, report its estimate and interval instead of changing the significance rule after seeing the result. Preserve the planned confidence level and identify exploratory analysis.
If the result changes after excluding low-willingness guests, inspect the exclusion reason. Those cases may be a genuine part of the population, and removing them can redefine the question rather than improve the evidence.
Save the original model and the reason for each genuine correction. Keep unsupported relationships in the results you report. A more convincing study explains a weak or uncertain finding and plans better evidence; it does not erase cases, items or paths until every test is favourable.
Write a clear Methods and Results account
Methods outline: We defined favourable and critical review willingness separately. The primary model predicted POS from the guiding tendency, individual collectivist orientation and their interaction. We used [case and scale rules], [uncertainty settings] and [planned scope of outcome comparisons].
Results outline: For POS, the interaction was [estimate and interval] and the conditional slopes at [probes] were [values]. The separately specified NEG analysis showed [results, if performed]. The findings concern individual response tendencies and do not establish national stereotypes or the fairness of actual reviews.
Include the analysed case count, exact file or data version, model diagram, measurement decisions, calculation settings and the uncertainty procedure. Attach the relevant full tables rather than only the strongest row. Explain which decisions were planned and which followed inspection.
For this exercise, call the records simulated teaching data. In an actual paper, replace every bracketed placeholder with the result from the named saved run. Remove any illustrative numbers from this lesson before drafting your findings. Cite the original study for the research context and the relevant methodological references for the analysis.
Explain the finding in this file
Write the outcome and the precise path being discussed. Include the estimate, its matching interval, the metric and the actual probe value when relevant. In a real paper, report the procedure you ran and its usable refits; do not copy reference intervals as if they came from a different calculation.
Check your conclusion before you finish
- Explain why favourable is not synonymous with biased.
- Keep the second outcome out of the first model’s predictor list.
- Report the actual interaction row and its uncertainty.
- Do not infer a between-outcome difference from two significance labels.
- Interpret collectivism at the individual level measured.