MerQuR+Statistics Glossary

Statistics Glossary → PLS-SEM (Partial Least Squares Path Modeling)

PLS-SEM (Partial Least Squares Path Modeling)analysis

A variance-based structural equation method that estimates path relations between indicator-measured latent constructs by iteratively building construct scores and running regressions on them, aiming to maximise the explained variance of the endogenous constructs.

Concept

Covariance-based structural equation modelling (CB-SEM) focuses on how well the whole model reproduces the observed covariance matrix, which makes it suited to theory testing. Partial least squares path modelling (PLS-SEM) answers a different question: how much of the variability in the endogenous (dependent) constructs can the antecedent constructs explain and predict? It is often used for prediction-oriented goals, complex models with many constructs, relatively small samples, and measures in which the indicators jointly form the construct (formative measurement).

The algorithm builds a score for each construct as a weighted sum of its indicators. The weights are updated iteratively: in the outer step construct scores are computed from the indicators, in the inner step proxy scores are obtained from neighbouring constructs, and the process stops once the weights stop changing. Path coefficients are then estimated by ordinary least squares regressions on the construct scores. In reflective (Mode A) measurement the indicators are consequences of the construct and are expected to correlate; in formative (Mode B) measurement the indicators jointly make up the construct and are not interchangeable.

Evaluation proceeds in two steps. First the measurement model: for reflective constructs, loadings, reliability (alpha, rho_A, rho_C), convergent validity (AVE ≥ .50) and discriminant validity (HTMT); for formative constructs, the significance of weights and collinearity among indicators. Then the structural model: path coefficients, R², the effect size f² and predictive relevance (Q²). Because the method makes no distributional assumption, significance is assessed by bootstrapping. Being component-based, it can slightly inflate loadings and slightly shrink paths in reflective models, so for confirming theory it should be considered alongside CB-SEM.

When to use it

  • When the aim is to predict particular endogenous constructs (e.g. customer loyalty) and see the explained variance, when the model contains many constructs and paths, or when some constructs are defined by formative indicators.
  • When the aim is to test a well-defined theory with global fit measures and the constructs are reflective, covariance-based SEM (CB-SEM) is more appropriate.
  • If only directional relations between observed variables are of interest, multiple regression or path analysis is enough; to test the structure of a single scale, use confirmatory factor analysis.

An everyday example

A coffee chain asks 250 customers about service quality, price fairness, satisfaction and loyalty (intention to return and to recommend). The model is: service quality and price fairness affect satisfaction, and satisfaction affects loyalty. PLS-SEM might show that satisfaction explains about half the variability in loyalty (R² = 0.50) and that the path from service quality to satisfaction is stronger than the one from price fairness. With this result, the chain might decide to put its limited budget into service quality first.

Parameters in MerQur

Model syntax ([Parameters] text field)

Concept: The model syntax defines both the measurement model (which indicators measure each construct and whether reflectively or formatively) and the structural model (which construct affects which). In a reflective construct the indicators mirror the construct and are interchangeable; in a formative construct each indicator is a distinct component, and dropping one changes what the construct means.

Everyday example: For the coffee chain, 'satisfaction' is reflective: 'I am satisfied', 'it met my expectations' and 'it was a good choice' are different expressions of the same feeling. 'Service quality' can be defined as a formative construct made of speed, friendliness and cleanliness; dropping cleanliness removes part of the concept.

What it does: Defines the measurement and structural model in text syntax: "F =~ x1 + x2" (reflective, Mode A), "F <~ x1 + x2" (formative, Mode B), "Y ~ X1 + X2" (structural path). An undefined column name becomes a single-indicator construct; a cycle, a shared indicator and an unknown name give an explicit error message.

Effect on the result: Determines all outputs. The alpha, rho_C and AVE of Mode B constructs are shown as "not applicable" (Hair et al. 2022).

On the validation data: In the extended model R2: COMP 0.6309, CUSL 0.562, LIKE 0.5576, CUSA 0.2919; paths QUAL->COMP 0.4297, LIKE->CUSA 0.4357. In the simple model COMP alpha 0.776, rho_A 0.8319, rho_C 0.8646, AVE 0.6806; the ρ_C = .869 and AVE = .688 in the interpretation come from the full model.

→ Measurement model and structural model

Inner weighting scheme (Path (default), Centroid, Factor (factorial))

Concept: The inner weighting scheme sets the weights given to neighbouring constructs when the algorithm's inner step builds a proxy score for each construct. The path scheme respects the direction of the arrows in the model; the centroid and factor schemes look only at correlations between constructs. Results usually change only slightly.

Everyday example: When building the proxy score for satisfaction, the path scheme treats service quality and price fairness jointly (by regression) as its antecedents and loyalty as its successor. The centroid scheme looks only at the sign of each construct's correlation with satisfaction.

What it does: Selects the weighting scheme of the inner proxy variables. The path scheme (default) uses multiple regression for antecedents and correlation for successors; centroid uses the sign of the correlation, the factorial scheme the correlation itself.

Effect on the result: Changes the construct scores, hence loadings, weights and path coefficients, to a small extent.

On the validation data: Each of the four models (corp_simple, corp_ext, mobi, sat) is estimated with all three schemes.

→ Inner weighting scheme (PLS-SEM)

Missing data (Mean replacement (default), Listwise deletion) and Missing value code

Concept: Mean replacement fills a missing value with that indicator's mean; it is simple but, when much is missing, artificially reduces variability. Listwise deletion drops anyone with a missing answer. The missing value code stops the number used to mark missingness (e.g. -99) from being read as a real answer.

Everyday example: If some customers skipped the price question and those cells are coded -99, failing to declare the code lets -99 count as a real score and distorts the price-fairness mean. When the missing rate is low, mean replacement does little harm.

What it does: Mean replacement is the default (reapplied in each bootstrap sample); a warning is given if the missing rate in an indicator exceeds 5%. The missing value code sets the number counted as missing in the indicators (e.g. -99).

Effect on the result: Changes the estimated sample and the values.

On the validation data: -99 code and 11 missing cells in corp_rep_data; mean replacement is applied.

→ Single imputation (mean, median, kNN filling)

Maximum iterations (300) and Stop criterion (10^-k, k = 7)

Concept: The PLS algorithm updates the weights step by step and stops when the weights in two successive steps are almost identical. The stop criterion sets the threshold for 'almost identical', and the maximum iterations sets how many steps the algorithm will try at most.

Everyday example: The algorithm adjusts the weights of the service-quality items a little in each round and stops once the adjustments become too small to notice. In typical models this happens within a few dozen rounds; hitting the limit points to a convergence problem.

What it does: Determines the maximum number of iterations of the PLS path algorithm and the threshold below which the sum of absolute differences of successive outer weights stops.

Effect on the result: Determines convergence accuracy; the result does not change after convergence.

On the validation data: Defaults: at most 300 iterations, stopping criterion 10^-7.

→ Convergence and maximum iterations

Confidence Level (%90, %95, %99, Other (custom))

Concept: Sets how much assurance the bootstrap confidence intervals carry. A higher level gives a wider interval.

Everyday example: If the path from satisfaction to loyalty is 0.55, you can say with 95% confidence that the true path lies between, say, 0.45 and 0.64; asking for 99% widens the interval.

What it does: Determines the level of the bootstrap percentile and BCa confidence intervals.

Effect on the result: Changes only the interval limits.

On the validation data: 90% interval for HTMT: COMP-CUSA MerQur 0.371 to 0.5573; Hair et al. 0.366 to 0.554 (within Monte Carlo tolerance).

→ Confidence level and confidence interval

Bootstrap samples (default 5000) and Seed (default 123)

Concept: Because PLS-SEM makes no distributional assumption, the uncertainty of the coefficients is measured by bootstrapping: many new samples are drawn from the data with replacement, the model is re-estimated in each, and the coefficient's distribution across these samples gives its standard error. More samples give a more stable result; the seed makes the same random samples reproducible.

Everyday example: From the 250 customers' answers, 5000 new samples of 250 are drawn and the service-quality path is computed each time. If the path comes close to zero in very few of these samples, the effect is reliable. With the same seed, another researcher gets exactly the same result.

What it does: Determines the number of resamples for significance (10 000 recommended for the final report; 0 = no bootstrap) and the seed. The p value is computed from t = estimate / SD with the t distribution (df = B - 1) (Hair et al. 2021).

Effect on the result: Determines SD, t, p and confidence limits; does not change the point estimates.

On the validation data: With B = 500 resamples the estimate, SD, percentile and BCa intervals are given; with a different generator (B = 2000) the SD ratio difference is at most 4.6%. QUAL->COMP bootstrap t = 6.5618.

→ Bootstrap (resampling)

BCa confidence interval (check box)

Concept: The percentile bootstrap interval uses the extreme percentiles of the bootstrap distribution directly. The bias-corrected and accelerated (BCa) interval adjusts the limits for the distribution's shift (bias) and skewness, giving more accurate coverage when the distribution is not symmetric.

Everyday example: If the bootstrap distribution of price fairness's indirect effect on loyalty is right-skewed, the percentile and BCa intervals diverge. Because BCa corrects for that skewness, it is a more trustworthy basis for judging whether the indirect effect is significant.

What it does: In addition to the percentile interval, gives the bias-corrected and accelerated (BCa) interval; the acceleration is computed by jackknife. When on, the significance flag is based on BCa.

Effect on the result: Affects only the confidence interval.

On the validation data: The same at the 1e-14 level on the same resamples.

→ BCa bootstrap confidence interval

Q2 (blindfolding) (check box) and Omission distance (D, default 7)

Concept: Q² is a measure of the model's predictive relevance: data points are omitted in turn, the model predicts them from the remaining data, and the prediction error is compared with the error of predicting by the mean alone. A Q² above zero means the model predicts the endogenous construct better than the mean; the omission distance sets which data points are hidden in each round.

Everyday example: Some customers' loyalty answers are hidden and the model predicts them from their service-quality, price and satisfaction answers. If these predictions are clearly better than saying 'everyone has average loyalty', Q² comes out positive.

What it does: Gives the Stone-Geisser Q2 by the cross-validated redundancy approach (for reflective endogenous constructs). Every D-th data point is omitted; D should be between 5 and 10 and chosen so that n/D is not an integer.

Effect on the result: Affects only the Q2 output.

On the validation data: The Q2 values are reported in 13 rows.

→ Q² predictive relevance (Stone-Geisser, PRESS)

Extra output (cross-loadings) and Extra output (save construct scores to the dataset)

Concept: Cross-loadings show each indicator's correlation not only with its own construct but with the others too; an indicator should relate most strongly to its own construct (discriminant validity). Saving construct scores makes the estimated construct values available for later analyses.

Everyday example: If the item 'I will come back' relates more strongly to satisfaction than to loyalty, the cross-loading table reveals it and suggests the two constructs are not well separated. If each customer's loyalty score is saved, it can be used to compare branches.

What it does: The first adds the cross-loading table, the second adds the construct scores to the dataset as new columns.

Effect on the result: Does not change the estimation.

On the validation data: In the corp_ext model 248 cross-loadings are given.

→ Cross-loadings

SRMR (automatic output)

Concept: SRMR is a fit measure summarising the average size of the differences between observed correlations and those the model implies. Small values (below .08 by a common guideline) show that the model reproduces the relations in the data well.

Everyday example: If the model closely reproduces most of the correlations among customers' answers, SRMR takes a small value such as .05. A large value such as .12 suggests the model is missing some relations, for example a direct link between price and loyalty.

What it does: Gives the standardized root mean square residual for the estimated and saturated model (model-implied correlation matrix: Mode A common factor, Mode B composite).

Effect on the result: Summarises model fit; does not change the estimates.

On the validation data: SRMR is .057 in the default model.

→ Model fit indices (χ², CFI, TLI, RMSEA, SRMR)

Example on real data

Data: corp_rep_data, Hair et al. 2021, R seminr

Variables: corporate reputation model: QUAL, PERF, CSOR, ATTR (Mode B) -> COMP, LIKE -> CUSA -> CUSL; missing value code -99

  • Question: How do the antecedents of corporate reputation (quality, performance, social responsibility, attractiveness) affect customer satisfaction and loyalty through competence and likeability? (Hair et al. 2021, n = 344)
  • Measurement model: Reliability is adequate for the reflective constructs: COMP α = .776, ρ_C = .869, AVE = .688 (full model); LIKE α = .831, AVE = .747; CUSL α = .831, AVE = .748. The highest HTMT is 0.780 (COMP-LIKE) < 0.85.
  • Structural model: QUAL -> COMP β = .43 (t = 6.56, p < .001), QUAL -> LIKE β = .38 (p < .001), LIKE -> CUSA β = .44 (p < .001), CUSA -> CUSL β = .51 (t = 11.97, p < .001); COMP -> CUSL β = .01 (p = .916) not significant.
  • Explained variance: R²: COMP .631, LIKE .558, CUSA .292, CUSL .562.
  • Indirect effect: LIKE -> CUSA -> CUSL = .220 (p < .001); part of the effect of likeability on loyalty passes through satisfaction.
  • Model fit: SRMR = .057 (< .08).
  • APA: CUSA -> CUSL: β = .51, t = 11.97, p < .001, 95% CI [.42, .59]; R²(CUSL) = .562; SRMR = .057 (5000 bootstrap samples).

Related concepts

Q² predictive relevance (Stone-Geisser, PRESS)Inner weighting scheme (PLS-SEM)BCa bootstrap confidence intervalMeasurement model and structural modelCross-loadingsModel fit indices (χ², CFI, TLI, RMSEA, SRMR)Single imputation (mean, median, kNN filling)Factor (component) loadings

The numbers in the examples are MerQur's output on independent, publicly available validation data; see the MerQur Technical Note for details.