Paper tutorial / Research in practice

Usefulness, anxiety and intention: two-stage moderation

Estimate whether the usefulness–intention relationship changes with anxiety.

16 min read

We use this paper to show how a similar analysis can be carried out in QuickPLS. Follow the steps with the supplied practice data, then adapt the setup and calculations to your own study. The practice data are a separate teaching example.

An intention model is not a test of clinical benefit

Damar and Koksalmis examined medical learners’ and physicians’ intentions toward metaverse-based education. Their survey included 718 participants in Türkiye and used PLS, including anxiety-related moderation. Most respondents had no prior medical-education metaverse experience, so the study largely concerns expectations and intentions.

In QuickPLS you will add Technology anxiety as a moderator of Expected usefulness → Use intention, fit the revised PLS model and bootstrap the interaction. The original practice data isolate this one relationship from the paper’s wider model. The outcome remains intention, not demonstrated clinical learning, patient care or actual adoption.

Your study: usefulness, anxiety and intention

Imagine 320 fictional medical learners evaluating a described educational tool. Expected usefulness, Technology anxiety and Use intention each have three reflective teaching indicators. The file contains complete continuous 1–7 values, with higher Anxiety meaning more anxiety.

Draw Expected usefulness → Use intention and Technology anxiety → Use intention. Then add a moderating effect in which Technology anxiety changes the Expected usefulness → Use intention relationship. The two ordinary arrows are the main effects and must remain in the model.

Moderation is a changing slope. It asks whether the usefulness–intention relationship differs across anxiety levels. It does not ask whether anxiety is a mediator between usefulness and intention, and it is not tested by dividing the sample into convenient low and high groups and comparing their separate significance labels.

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

ConstructColumnsRole
Expected usefulnessuseful1, useful2, useful3Focal predictor
Technology anxietyanxiety1, anxiety2, anxiety3Moderator and main-effect predictor
Use intentionintend1, intend2, intend3Outcome
Identifiercase_idOne fictional learner

Step 1. Keep expectations separate from experience

For a real pre-adoption survey, record what participants were shown before answering. A description, image or demonstration can influence expectations. Asking an inexperienced learner whether a tool seems useful is different from measuring usefulness after sustained use.

Import the teaching CSV into a writable project using File → Import Data… → Choose data file…. In Review & import use raw data, Header row 1, comma separator and Decimal convention Dot (1.25). Keep the nine measurement columns Continuous. Choose Import dataset and verify 320 cases.

No reversal or missing-value filling is needed. Keep case_id outside the model. The anxiety indicators already point toward greater anxiety. Do not reverse them in response to an unexpected coefficient. Keep a record of the construct definitions so the sign of every effect can be explained plainly.

Step 2. Build and check the ordinary measurement model first

  1. Create the model

    Choose File → New Model…, enter Metaverse intention, check Dataset version and Cases, then choose Create model.

  2. Create three reflective constructs

    Ctrl-select each corresponding three-column block in Indicators and choose Create construct. Select each block and set its Name in Properties → Model. Under Parameter → Measurement model choose Reflective.

  3. Draw both main effects

    Use Connect to draw Expected usefulness → Use intention and Technology anxiety → Use intention. Return to Select and save. These are ordinary structural paths, not control-role paths.

  4. Check the initial measurement

    Run PLS-SEM Algorithm for this baseline. Inspect Outer loadings, Construct reliability and validity, HTMT+ and Inner VIF values. Check that usefulness and anxiety are distinct measures before interpreting their interaction.

Step 3. Add the moderation to the intended relationship

The interaction is a generated model term, not another survey question. Two-stage construction uses the model’s construct-score information to form the moderation analysis. Keep its construction method in the report. A model revision helps you retain the original baseline while checking what the interaction adds.

The moderation model retains both main effects. The interaction is a generated term. Add it through Moderating Effect; keep both measured main-effect predictors.
The interaction is a generated term. Add it through Moderating Effect; keep both measured main-effect predictors.
  1. Open the moderation dialog

    Select the focal Expected usefulness → Use intention structural path. Choose Model → Moderating Effect…. The command may display Moderating Effect (Creates Model Revision)…. The focal path’s context menu also offers this action.

  2. Choose the relationship and moderator

    In the dialog check Relationship is Expected usefulness → Use intention. Set Moderator to Technology anxiety. Keep both measured constructs and both main-effect paths in the model.

  3. Choose the construction

    Set Construction to Two-stage and choose Add moderating effect. Inspect the resulting model revision and save it under a name that distinguishes it from the main-effects baseline.

    What you should see: The revised model includes the generated interaction and both measured main-effect predictors.

  4. Calculate the supported model

    Open Calculate… and choose PLS-SEM Algorithm for the revised model. Review Readiness, the selected interaction and preprocessing settings. Record the settings shown for this model and choose Start calculation. Do not substitute the common-factor CB-SEM product-indicator option for this PLS interaction.

Screenshot to be addedSelect the usefulness relationship and anxiety moderator

Step 4. Read the interaction alongside its main effects

In the revised result, check measurement warnings and the named direct-effect rows before the interaction. The Expected usefulness main coefficient describes its slope at Anxiety equal to zero on the analysis’s score scale. With a standardized anxiety score, zero is the score mean. It is not the usefulness slope for every learner.

Read the interaction coefficient as the change in that slope for a one-unit increase in the moderator on the stated scale. A negative interaction means the usefulness association becomes less positive, or more negative, as anxiety rises. It does not by itself mean anxiety has a harmful overall effect or that the usefulness relationship becomes negative.

Open Simple slope analysis and read Moderator score with Conditional effect. These are point estimates of the usefulness slope at the displayed anxiety values; this table does not provide slope confidence intervals. Keep the score scale beside each estimate rather than assuming that a raw questionnaire score of 4 equals the standardized construct-score mean.

Select the revised interaction result in the Result dropdown. Confirm that the model still contains Expected usefulness → Use intention and Technology anxiety → Use intention before opening Simple slope analysis. A result saved before adding the interaction answers the baseline question.

Hypothetical example

Illustrative reading example

Suppose the standardized usefulness slope is 0.50 - 0.20 × Anxiety. At Anxiety = -1, 0 and 1 the slopes are 0.70, 0.50 and 0.30. All remain positive, while the relationship weakens as anxiety increases. These are invented coefficients and standardized moderator values, not predicted raw responses or expected results from the CSV.

Screenshot to be addedRead the metaverse-model simple slopes

Step 5. Test the change rather than comparing asterisks

Choose PLS-SEM Bootstrapping for the complete revised model, keeping both main effects and the generated interaction. For a first teaching run use 1000 Bootstrap samples, Confidence level (%) 95 and Seed 202611082 where those controls are shown. Record the model’s actual construction and saved settings with the result.

Check attempted, usable and failed refits and the interval method supplied. Read the interaction’s own estimate and interval to assess whether the usefulness relationship changes with anxiety. The point slopes describe the estimated pattern. Their different values alone do not establish that the relationship changes beyond sampling uncertainty.

If the interaction interval includes zero, report the estimated pattern and its uncertainty. The correct conclusion may be that the data do not clearly establish a changing relationship, even though the main usefulness association is positive. Do not claim no moderation exists in every population or seek a threshold that produces two convenient groups.

Hypothetical example

Illustrative reading example

An interaction estimate of -0.10 with a 95% interval of [-0.24, 0.05] would allow both a weakening relationship and a small strengthening one. Report that the direction of moderation remains uncertain. These bounds are hypothetical; they are not the interval for the earlier -0.20 arithmetic example.

Screenshot to be addedRead the interaction’s bootstrap interval

Check the two stages with this practice file

First fit the ordinary PLS model without the interaction. Standardise its saved construct scores using their sample standard deviations. In the second stage, keep those scores fixed and regress each outcome jointly on its incoming predictors and the standardised predictor × moderator product.

Usefulness remains positively associated with Intention at all three anxiety levels. The slopes decrease from 0.461 to 0.360, but the interaction interval includes zero (−0.147 to 0.035). These data do not provide clear evidence that anxiety changes the usefulness slope.

The reference recalculated both stages for 1,000 usable resamples out of 1,000 requested, using R seed 202611082. Each resample selects whole rows, keeping each case’s values together. The uncertainty column gives the 2.5th and 97.5th percentiles of those estimates.

About these answers: First-stage scores checked against QuickPLS calculation code; the second-stage regression was calculated independently in R and NumPy with those scores held fixed.

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

Second-stage pathPoint estimate95% reference percentile interval
Usefulness → Intention0.4110.311 to 0.498
Anxiety → Intention-0.158-0.258 to -0.057
Standardised product → Intention-0.053-0.147 to 0.035
How the interaction answers were checked

The first-stage QuickPLS PLS scores were checked against cSEM. R and NumPy then independently calculated the second-stage regression and slopes with those scores held fixed. This did not run the app’s own second-stage calculation routine. The interval arithmetic was checked using all 1,000 saved reference resamples, each of which refitted both stages. The repeated-fit calculation itself was not independently rerun.

To obtain gamma, divide the coefficient of the standardised product by the sample standard deviation of the raw product of the standardised scores. Then simple slope = b + gamma*w, where b is the predictor’s main-effect coefficient and w is the moderator score. This conversion keeps the product coefficient and moderator probe on compatible scales.

These intervals come from the separate reference calculation, not the app’s Bootstrapping command. The download includes the sampled row numbers, settings, estimates and any failed fits.

Follow the two calculation stages

Use this plan alongside the model diagram. It shows which scores and paths belong to each stage, so an indicator model and a second-stage score model are not confused.

The two-stage calculation for medical metaverse moderation. Exact two-stage interaction plan. The second-stage outcome includes both main effects and the generated product; the first-stage indicator weights are not re-estimated at that step.
Exact two-stage interaction plan. The second-stage outcome includes both main effects and the generated product; the first-stage indicator weights are not re-estimated at that step.

Read the slope at an actual moderator level

The probes −1, 0 and +1 mean one sample standard deviation below the mean, the mean, and one above. The product coefficient in the first table uses a standardised product. Divide it by the sample standard deviation of the unstandardised score product before calculating these slopes.

About these answers: Reference second-stage coefficients and full-stage resampling; all moderator probes use standard-deviation units.

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

Quantity / moderator scorePoint estimate95% reference percentile interval
Slope at -10.4610.341 to 0.589
Slope at 00.4110.311 to 0.498
Slope at 10.3600.205 to 0.486

Keep the moderator values within a credible range

Before explaining an extreme simple slope, check how much data support that anxiety level. A coefficient formula can produce a value outside the observed range, but the arithmetic alone does not make it a credible description of learners. State the range and scale used in the interpretation.

Do not dichotomise a continuous anxiety score at its median just to obtain a more striking result. That discards information and changes the question. If a genuine pre-existing group comparison is needed, it requires its own measurement-comparability and group-analysis plan.

The proposed outcome remains intention toward an educational tool. Even a clearly supported interaction would not demonstrate that an anxiety-reduction intervention improves clinical performance. That conclusion would need a design with an intervention and appropriate performance outcomes.

Check the focal arrow, score scale and uncertainty

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

ProblemCheckNext action
Moderation cannot be addedMeasured focal constructs, additional moderator and structural path roleUse the intended structural relationship; the workflow blocks control-role paths.
The wrong arrow is moderatedRelationship and Moderator in the dialogCorrect the model revision before calculating.
The usefulness main coefficient seems surprisingModerator zero point and interaction termInterpret it conditionally, not as the slope for all learners.
A negative interaction is called a negative usefulness effectSimple slopes across supported anxiety valuesDistinguish a weakening positive slope from a negative slope.
Point slopes look differentInteraction estimate and its own intervalUse the interaction uncertainty to assess a changing relationship; the simple-slope table supplies point estimates.
Bootstrap failures or unavailable intervals occurGenerated-term specification and failure messagesReport availability and investigate rather than attaching baseline intervals.

Compare the baseline and interaction as different models

Keep the main-effects baseline and the revised interaction model. Explain why the moderation question was added and compare the relevant results without selecting a model only because it has a larger R-square. Adding a term usually changes fitted explanation; it needs a scientific interpretation and uncertainty assessment.

A useful sensitivity discussion considers whether anxiety was measured before or after exposure to the tool description, whether experience changes the construct meaning, and whether the chosen probes are supported by the data. Those design questions may matter more than increasing the bootstrap count.

Describe exactly which relationship anxiety changes

Methods outline: We fitted reflective measures of Expected usefulness, Technology anxiety and Use intention. We retained both main effects and added a Two-stage PLS moderating effect of Technology anxiety on Expected usefulness → Use intention. We inspected measurement quality and evaluated the interaction with [complete-model bootstrap settings and counts]. We described the point slopes from Simple slope analysis at [actual moderator values and score scale].

Results outline: The interaction estimate was [value], with a [level and method] interval of [bounds]. At [moderator value], the usefulness point slope was [estimate]; at [other value], it was [estimate]. [State whether the interaction itself is supported or uncertain.] The simple-slope table supplies point estimates, so we report uncertainty for the interaction rather than attach that interval to each slope. These results concern pre-adoption intentions, not clinical performance or proven intervention benefits.

ReportInclude
Focal relationshipExpected usefulness → Use intention.
ModeratorTechnology anxiety, coding direction and score scale.
ConstructionTwo-stage interaction with both main effects retained.
Conditional readingActual moderator values, score scale and point slopes from Simple slope analysis.
InferenceInteraction interval, usable fits and any failures.
ScopeTool expectations and intentions rather than experienced clinical benefits.

Explain what this staged model actually tested

Describe the two stages, measurement roles, score scaling and resampling procedure. Use the same model and the same coefficient scale for the estimate and its interval. Keep a main-effect conclusion separate from a claim that one variable changes another variable’s slope.

Explain a negative but uncertain interaction

  • Find the exact structural arrow selected in the moderation dialog.
  • Show that both main effects remain after adding the interaction.
  • Recalculate the hypothetical slopes 0.70, 0.50 and 0.30, then interpret the checked practice slopes and the interaction interval. Explain why decreasing point slopes do not establish moderation when that interval includes zero.
  • Explain why a negative interaction need not produce a negative usefulness slope.
  • Report an interval crossing zero without declaring universal absence or searching for favourable groups.

Further reading

Related guides