The assessment sequence a PLS study is judged on, reflective versus formative measurement, and the fit indices you should not be reporting.
SmartPLS runs partial least squares structural equation modelling, a variance-based approach oriented towards prediction and explaining variance in target constructs. It is assessed in a fixed sequence: the measurement model first, then the structural model, with significance obtained through bootstrapping rather than from distributional assumptions. The most common reporting error is importing criteria from covariance-based SEM — global fit indices such as CFI and RMSEA are not how a PLS model is judged.
This is the first question asked, and a weak answer sets the tone for everything after it.
Defensible reasons to choose PLS include a focus on prediction and explaining variance in key target constructs rather than confirming an established theory; a model containing formatively measured constructs, which covariance-based approaches handle awkwardly; a complex model with many constructs and indicators relative to the available sample; and exploratory theory development where no well-established model is being tested.
The answer to avoid giving on its own is “my sample was small”. PLS is more tolerant of modest samples than covariance-based estimation, but tolerance is not a rationale, and the claim that PLS solves small-sample problems has attracted substantive criticism in the methodological literature. If your sample is small, say so as a constraint and give a methodological reason for PLS alongside it.
There is genuine, ongoing debate about when PLS is appropriate, with respected work on both sides. That is a reason to cite your justification properly rather than to avoid the method. A candidate who acknowledges the debate and explains their position is in a considerably stronger place than one who appears unaware of it.
This decision determines which criteria your measurement model is judged against, and getting it wrong invalidates the assessment rather than merely weakening it.
In a reflective construct, the latent variable causes the indicators: the items are interchangeable manifestations of an underlying concept, so they should correlate highly, and dropping one does not change what the construct means. Most attitude and perception scales are reflective.
In a formative construct, the indicators cause the construct: they are components that together compose it, they need not correlate, and dropping one removes part of the construct itself. A measure of socioeconomic status built from income, education and occupation is the standard illustration.
The test is directional. Ask whether a change in the construct would change all the items, or whether a change in one item would change the construct. Applying reflective criteria to a formative construct — expecting high inter-item correlation, deleting items with low loadings — damages the measure and is a recognised error.
PLS studies are reported against an established sequence, and following it makes the results chapter straightforward to write and to examine.
Assess in this order: indicator reliability through outer loadings; internal consistency through composite reliability, with Cronbach’s alpha usually reported alongside as a lower bound; convergent validity through average variance extracted; and discriminant validity.
On discriminant validity the field has moved, and this is worth getting right. The long-standing approach comparing the square root of average variance extracted against inter-construct correlations has been shown to perform poorly at detecting problems under common conditions. The heterotrait–monotrait ratio is now widely expected, and reporting only the older criterion increasingly attracts a question. Report HTMT, and include the older criterion as well if your field still expects it.
Threshold values for all of these circulate widely and have themselves shifted as the literature developed. Cite the source of whatever criteria you apply rather than presenting them as settled.
Formative constructs are assessed differently, and applying the reflective checklist to them is a visible error.
What matters is indicator collinearity, since formative indicators that overlap heavily produce unstable weights, and the significance and relevance of the outer weights, obtained through bootstrapping. An indicator with a non-significant weight is not automatically removed — if it is conceptually part of the construct, removing it changes what you are measuring, and content validity takes precedence over the statistic.
Only once measurement holds. Assess collinearity among predictor constructs, then the path coefficients with significance from bootstrapping, then R squared for each endogenous construct, f squared for the contribution each predictor makes, and predictive relevance obtained through blindfolding.
Given that PLS is oriented towards prediction, reporting out-of-sample predictive performance has become increasingly expected, and current versions of the software provide procedures for it. Including it signals familiarity with where the method has gone.
PLS makes no distributional assumptions, so significance cannot come from a theoretical sampling distribution. Bootstrapping supplies it empirically: the software resamples your data many times, estimates the model in each resample, and builds a distribution for every coefficient from the results.
Report the number of resamples, and use a large number — small numbers produce unstable results that will differ if anyone re-runs them. Report confidence intervals alongside significance, since they say more about precision than a p-value alone.
Bootstrapping is also how indirect effects are tested, which is the current standard for mediation across methods, not only in PLS. Report the indirect effect with its bootstrapped interval rather than working through the older causal-steps procedure.
One practical caution: bootstrap results vary slightly between runs because resampling is random. Report the run you used, and do not re-run repeatedly until a borderline path becomes significant — that is a form of selective reporting, and the instability is itself telling you the effect is not robust.
This is the most common error in PLS chapters, and it comes from applying a covariance-based checklist.
PLS does not optimise the reproduction of a covariance matrix, so the global fit indices used in covariance-based SEM — chi-square, CFI, TLI, RMSEA — do not have the same meaning and are not the criteria a PLS model is judged against. Reporting a table of them signals that the two approaches have been conflated.
Some approximate fit measures are available in PLS software and are debated in the literature; where you use them, cite the discussion rather than presenting them as equivalents of covariance-based indices. What a PLS model is actually assessed on is the sequence above: measurement quality, then path coefficients, explained variance, effect sizes and predictive relevance.
If your supervisor expects CFI and RMSEA, that usually means they are thinking in covariance-based terms — which may be a signal that covariance-based SEM is the better fit for your study, and that is worth raising directly.
When your emphasis is prediction and explaining variance in target constructs rather than confirming an established theory, when your model includes formatively measured constructs, when the model is complex relative to the sample available, or in exploratory theory development. A small sample is a constraint rather than a justification on its own — the claim that PLS solves small-sample problems has been criticised, so pair it with a methodological reason and cite your position.
In reflective measurement the construct causes the indicators: items are interchangeable expressions of one underlying concept, should correlate highly, and removing one does not change the construct. In formative measurement the indicators compose the construct: they need not correlate, and removing one removes part of what is being measured. Ask whether a change in the construct would alter all items, or whether changing one item would alter the construct. Assessing a formative construct with reflective criteria is a recognised error.
No. PLS does not work by reproducing a covariance matrix, so those indices do not carry the same meaning and are not the criteria a PLS model is assessed against. Reporting them suggests the two approaches have been conflated. A PLS model is judged on measurement model quality, then path coefficients with bootstrapped significance, explained variance, effect sizes and predictive relevance.
Use the heterotrait–monotrait ratio, which is now widely expected. The older approach comparing the square root of average variance extracted against inter-construct correlations has been shown to detect problems poorly under common conditions, and reporting it alone increasingly attracts a question. Report HTMT, add the older criterion if your field still expects it, and cite the source of whatever thresholds you apply.
A large number, and report the figure you used. Small numbers of resamples give unstable results that will change if anyone re-runs the analysis. Report confidence intervals alongside significance, and do not re-run repeatedly hoping a borderline path becomes significant — the instability is itself evidence that the effect is not robust, and selective re-running is difficult to defend.
Yes, and the current approach is to test the indirect effect directly with a bootstrapped confidence interval, which the software provides. This is the standard across methods now, not only in PLS, and it has superseded the older causal-steps procedure. If your data are cross-sectional, state that the causal ordering is theoretically specified rather than empirically demonstrated — that limitation applies regardless of which software produced the estimate.
Structural results reported before the measurement model holds cannot be repaired afterwards. Send your model and output, and a PhD in your field will work through the sequence with you.
Leave an email or a WhatsApp number — whichever you prefer — and tick how we should reach you. We reach out within 30 minutes.
