When does a broad capability need several distinct dimensions?
Analytical capability is more than buying technology. A business may also need skilled people and a culture that uses evidence. If those dimensions jointly define the broader capability, treating them as interchangeable manifestations of one trait would make the wrong measurement claim.
Khan, Hinterhuber and Di Benedetto studied 416 managers and used PLS modelling with a higher-order analytics capability and other organisational capabilities to examine market-driving outcomes. Their model included a third-order capability. This lesson deliberately uses an original second-order teaching model to explain dimension assessment, formative higher-order measurement and staged estimation.
Let resources, skills and evidence use form analytics capability
Use 280 fictional business units. Each row has three invented indicators for technical resources, analytical skills, evidence-use culture, learning routines, alliance orientation, market-driving activity and competitive advantage.
Resources, skills and culture form a higher-order construct named ANALYTICS. Their own indicator blocks are reflective. Their relationship to ANALYTICS is formative: changing one dimension changes what the broader capability contains, and the dimensions need not be interchangeable.
Specify ANALYTICS, LEARN and ALLIANCE as predictors of DRIVE, and DRIVE as a predictor of ADV. This model asks how the broader capability relates to market-driving activity alongside two other organisational features. It does not estimate the source paper's complete third-order structure.
Keep all three dimensions because the practice definition includes them. A small higher-order weight is not an automatic reason to redefine analytics capability by dropping a dimension.
On a small screen, swipe the table sideways to see every column.
| Columns | Construct | Measurement role |
|---|---|---|
| TECH1–TECH3 | Technical resources | Reflective lower-order dimension |
| SKILL1–SKILL3 | Analytical skills | Reflective lower-order dimension |
| CULTURE1–CULTURE3 | Evidence-use culture | Reflective lower-order dimension |
| LEARN1–LEARN3 | Learning routines | Reflective predictor |
| ALLIANCE1–ALLIANCE3 | Alliance orientation | Reflective predictor |
| DRIVE1–DRIVE3 | Market-driving activity | Reflective mediator/outcome |
| ADV1–ADV3 | Competitive advantage | Reflective outcome |
| ANALYTICS | Formed from TECH, SKILL and CULTURE | Higher-order construct, no raw column |
Step 1. Build and assess the ordinary dimensions first
The distinction between dimensions comes from the definitions, not from which statistical arrangement fits best. In a real instrument, inspect item wording and content coverage before selecting a formative or reflective higher-order relationship.
Import the indicators
In a writable project choose File > Import Data… and Choose data file…. Inspect Review & import with raw data, Header row 1, comma separator, Decimal convention Dot (1.25) and Continuous indicators. Choose Import dataset and confirm Dataset imported and verified.
Create a named model
Choose File > New Model…, select the imported Dataset version and Cases, then Create model. In Indicators select each matching three-column block and Create construct.
Set names and measurement roles
Use Properties > Model > Name for the seven ordinary constructs. Under Parameter > Measurement model choose Reflective for each. The complete file has no reverse-coded or missing indicators.
Check dimension definitions
Confirm that TECH items describe resources, SKILL items describe people and CULTURE items describe evidence use. Keep case_id outside the model. Save before creating the higher-order structure.
Step 2. Create the broader construct and inspect the stages
A disjoint two-stage analysis first estimates dimension scores and then uses those scores to represent the higher-order construct. It does not mean that the original indicators disappear from the measurement assessment. They remain essential to the validity of the dimensions whose scores enter the next stage.
Select the three dimensions
Select TECH, SKILL and CULTURE together. Open Higher-Order Construct from the Model menu or the selection context menu.
Define the higher-order relationship
Name the new construct ANALYTICS, select the intended dimensions and choose Dimensions form the HOC. Review the derived type and recommended approach. HOC means higher-order construct.
Choose the supported staged approach
Open Advanced. Under Construction approach choose Disjoint two-stage, then Create. If Initial model path appears for a protected project revision, select the intended path before Create. A fresh model need not show that field. Inspect the dimension scores entering the higher-order stage.
Add the substantive paths
Use Connect to specify ANALYTICS → DRIVE, LEARN → DRIVE, ALLIANCE → DRIVE and DRIVE → ADV. Review the stage mapping and distinguish any generated estimation paths from these research relationships.
Calculate the model
Open Calculate… and the supported PLS-SEM Algorithm setup for this higher-order model. Review readiness and the selected approach, then Start calculation. Open the completed result and inspect both stages before reading the final paths.
Step 3. Check the two stages with this practice file
The Disjoint two-stage model first estimates the reflective scores for TECH, SKILL and CULTURE without ANALYTICS. Those dimensions have no remaining structural arrows in the first stage. Their scores then form ANALYTICS in the second stage, while the other constructs keep their original indicators. The stage plan below explains this calculation; create the higher-order construct through the model workflow above.
ANALYTICS has the largest of the three associations with DRIVE in this model (0.384), followed by LEARN (0.239) and ALLIANCE (0.139). DRIVE → ADV is 0.404. These point estimates describe the specified model; a formal comparison of two coefficients needs its own contrast and uncertainty.
The uncertainty column comes from 1,000 usable full-stage reference refits out of 1,000 requested, with R seed 71010. Both stages are rebuilt in every whole-row resample. The endpoints are the 2.5th and 97.5th percentiles.
About these answers: Separate two-stage reference calculation. Component scores and point estimates were compared with QuickPLS calculation code; this was not a new run through the desktop app.
On a small screen, swipe the table sideways to see every column.
| Second-stage path | Point estimate | 95% reference percentile interval |
|---|---|---|
| ANALYTICS → DRIVE | 0.384 | 0.291 to 0.481 |
| LEARN → DRIVE | 0.239 | 0.141 to 0.331 |
| ALLIANCE → DRIVE | 0.139 | 0.042 to 0.247 |
| DRIVE → ADV | 0.404 | 0.309 to 0.503 |
How the reference answer was calculated
The reference calculated the second stage separately. Its component scores, point paths and R² values were compared with QuickPLS calculation code. This did not run the app's complete higher-order workflow, so it is not an end-to-end check of that workflow. The interval arithmetic was checked against all 1,000 saved full-stage repetitions. Those repetitions were not independently rerun and are not results from the QuickPLS bootstrap command.
Path weighting, +1 starting weights, 3,000 iterations and stopping criterion 0.0000001. For the isolated reflective dimensions, the reference uses the principal-component limit of the own-score iteration, oriented toward the positive item sum.
These reference intervals are not a run of the app’s native bootstrap command. The download retains the sampled row numbers, settings, reference estimates and any failures.
Step 4. 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.
Step 5. Read the dimensions that form the broader concept
A weight describes a dimension’s contribution while the other dimensions remain in the composite. It is not the same quantity as a reflective indicator loading. Keep a theoretically necessary dimension even when its weight needs further investigation.
On a small screen, swipe the table sideways to see every column.
| Dimension | Formative weight | 95% reference percentile interval |
|---|---|---|
| CULTURE | 0.359 | 0.092 to 0.578 |
| SKILL | 0.669 | 0.442 to 0.852 |
| TECH | 0.693 | 0.472 to 0.878 |
Step 6. Do not apply reflective checks to a formative composite
In Results use Search result outputs to find Higher-order component relationships, Higher-order structural paths and Higher-order method and run details. The component table separates Loading from Weight. The details table identifies the higher-order construct, components, method and generated measurement; use it to check the specification behind the estimates.
Read lower-order outer loadings, reliability, average variance extracted and HTMT+ for the reflective blocks. These checks address whether each dimension is measured coherently and can be distinguished from neighbouring constructs.
For ANALYTICS, read its higher-order weights and collinearity among dimension scores. A weight describes a dimension's contribution given the others and the selected estimation method. It is not the same as a loading or the percentage of analytics capability attributable to that dimension.
Do not interpret an alpha or AVE for the formative higher-order construct as though its dimensions were reflective items. Content coverage and the reason for including each dimension remain central. High overlap can make individual weights unstable without proving that a dimension is conceptually dispensable.
In this model TECH, SKILL and CULTURE are reflective dimensions, so read their HTMT+ comparisons. ANALYTICS is formative: assess its dimension weights and collinearity instead of expecting a reflective HTMT+ result for the broader construct.
On a small screen, swipe the table sideways to see every column.
| Level or output | What to assess | What it does not establish |
|---|---|---|
| Reflective dimension blocks | Loadings, reliability, AVE and separation | Validity of the broader formative definition by itself |
| Higher-order weights | Conditional contribution of each dimension | Percentage ownership of the concept |
| Dimension collinearity | Overlap that can destabilise weights | Automatic permission to delete a dimension |
| Stage mapping | Which scores enter each calculation | Interchangeability with scores from another model |
| Structural paths | Relationships among the specified constructs | Extra hypotheses represented by technical estimation paths |
Step 7. Resample the whole staged procedure
For uncertainty, choose the supported full-model higher-order bootstrap. The calculation must refit the required stages from the original business rows. Treating dimension scores estimated once as error-free fixed inputs would omit part of the estimation uncertainty.
In PLS-SEM Bootstrapping review the higher-order model and approach, then use Bootstrap samples 5000, Confidence level (%) 95 and Seed 71010 for the teaching run where these controls are shown. Start calculation and inspect the actual interval procedure, usable fits, failed fits and availability of each higher-order target.
Do not assume that every advanced interval method is available for every higher-order quantity. Report the method actually produced. Read intervals for the dimension weights and the substantive paths separately.
A small weight does not mean an irrelevant dimension
Hypothetical example
Illustrative reading example
Suppose illustrative higher-order weights are TECH 0.20, SKILL 0.55 and CULTURE 0.25. Suppose CULTURE's interval is [−0.04, 0.48]. The original estimate is positive, but its separate contribution is imprecisely estimated under the model.
Those weights are not percentages of business value, even though the invented numbers happen to sum to 1. Their interpretation depends on scaling, score construction and the relationships among the dimensions. A different normalisation can change their numerical sum without changing the concept's definition.
If CULTURE overlaps strongly with SKILL, inspect their indicator content and score collinearity. If evidence-use culture is a necessary part of the theoretical definition, its uncertain conditional weight does not justify silently dropping it. Report the uncertainty and consider whether the measures adequately distinguish the dimensions.
Suppose ANALYTICS → DRIVE is 0.40 and DRIVE → ADV is 0.35. Their illustrative indirect association is 0.14. Its interval must come from the complete model bootstrap, not multiplication of separate interval endpoints.
The hypothetical indirect product is 0.40 × 0.35 = 0.14. The weights and intervals in this illustrative example are invented for reading practice. They are not the download's fitted estimates or the source paper's third-order result.
Change an approach only for a stated measurement reason
A comparison between eligible higher-order approaches can be useful if both represent the intended measurement structure. Keep the dataset, dimension definitions and substantive paths fixed, and explain what differs in stage construction. Do not select an approach solely because one path becomes significant.
A model treating ANALYTICS as reflective would make a different claim: the broader concept would explain the dimensions. That is not a routine robustness option for a formative definition. Use it only if theory supports the alternative meaning.
Inspect sensitivity to a genuine item-coding correction at both levels. A change in a lower-order score can propagate into higher-order weights and structural effects. Keep the original and corrected results with the reason for the correction.
Find the level at which the problem occurs
On a small screen, swipe the table sideways to see every column.
| Symptom | Check | Action |
|---|---|---|
| ANALYTICS is missing from raw columns | Whether it is defined through dimensions | Create the higher-order construct; do not invent a raw ANALYTICS item. |
| A formative dimension has a weak weight | Collinearity, content and its interval | Report the contribution and uncertainty; do not delete on weight alone. |
| A reliability table says not applicable | The construct's formative role | Use formative assessment instead of forcing reflective statistics. |
| Stage-two results are interpreted without stage one | Dimension measurement and score mapping | Complete the lower-order checks first. |
| Bootstrap intervals ignore stage uncertainty | Whether the full model is refitted | Use the supported higher-order bootstrap from original cases. |
| Technical paths become extra hypotheses | The selected approach's generated structure | Report only the prespecified substantive relationships as hypotheses. |
Step 8. Describe both levels and the estimation sequence
Results should distinguish lower-order measurement, higher-order weights and structural relationships. Insert the actual weight intervals, collinearity findings and pathway estimates. Explain an uncertain dimension weight without rewriting the construct definition after seeing the result.
| Level | Values and explanation to report |
|---|---|
| Dimensions | Indicator assignments, loadings, reliability and validity |
| Higher-order construct | Formative definition, approach, weights and collinearity |
| Stages | Inputs and score mapping |
| Structural model | Planned paths and indirect effects with intervals |
| Uncertainty | Refitted stages, interval methods and failed fits |
Methods writing outline
We specified ANALYTICS as a formative higher-order construct formed by three reflectively measured dimensions. We used [eligible approach], assessed the original dimension blocks and recorded the stage mapping. The structural model linked ANALYTICS, LEARN and ALLIANCE to DRIVE, then DRIVE to ADV. Uncertainty used [full-model bootstrap settings and counts].
Step 9. 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. Report the dimension weights separately from the paths linking ANALYTICS, LEARN and ALLIANCE to DRIVE and DRIVE to ADV.
Defend the higher-order measurement claim
- Explain why the three dimensions form ANALYTICS.
- Find the indicators used in the first stage.
- Locate the scores used in the later stage.
- Distinguish a weight from a loading.
- Explain why weights are not percentages of conceptual importance.
- State that the exercise is second-order, not the paper's full third-order model.