Assess reliability and convergent validity

Before you can trust the paths between constructs, you must show that the constructs are measured well. For reflective constructs this means four checks, in this order: indicator reliability, internal consistency reliability, convergent validity and discriminant validity (Lesson 5). Formative constructs are assessed differently, as explained at the end of this lesson.

Step 1: indicator loadings

A loading is the correlation between an indicator and its construct. Squared, it is the share of the indicator’s variance explained by the construct. The guideline is a loading of at least 0.708, because 0.708² ≈ 0.50: the construct explains at least half of the indicator’s variance. In the sample study, all 20 loadings are between 0.814 (PE3) and 0.932 (BI1), and all are significant (p < .001, bootstrap CIs far from zero).

What if a loading is lower? Indicators between 0.40 and 0.708 should be removed only if deleting them raises composite reliability or AVE above the threshold, and if content validity does not suffer. Indicators below 0.40 should usually be removed. Report every deletion and the reason.

Indicator collinearity

The VIF column shows how much each indicator overlaps with the other indicators of its construct. For reflective items some overlap is expected; values above 5 suggest redundant wording. The highest here is BI1 at 3.32, slightly above the conservative guideline of 3 but well below 5, which is acceptable for a three-item intention scale whose items are deliberately similar.

Step 2: internal consistency reliability

Three reliability coefficients are reported in PLS-SEM, and the true reliability usually lies between them:

• Cronbach’s alpha assumes equal loadings and is a conservative lower bound.
• Composite reliability ρC weights indicators by their loadings and tends to be a liberal upper bound.
• ρA (Dijkstra-Henseler’s rho) lies between them and is the recommended single measure for PLS constructs.

Values between 0.70 and 0.90 are satisfactory to good (0.60 to 0.70 acceptable in exploratory research). Values above 0.95 are problematic: they suggest redundant items, straight-lining or inflated correlations. All constructs here are between 0.84 and 0.94, so reliability is established. The Studio also provides bootstrap CIs for each coefficient; for a strict test, check that the upper bound of ρC stays below 0.95 (BI: 0.952, at the limit).

Step 3: convergent validity

Convergent validity asks whether the construct explains most of the variance of its indicators. The average variance extracted (AVE) is the mean of the squared loadings and should be at least 0.50. All AVEs here are between 0.723 and 0.843, so convergent validity is established.

Formative constructs

For a formative construct, loadings and reliability make no sense, because the indicators need not correlate. Instead, assess (1) convergent validity with a redundancy analysis against a global single item measuring the same construct (path ≥ 0.70), (2) indicator collinearity (VIF below 3, problems above 5), and (3) the significance and size of the outer weights. A non-significant weight is kept if its loading is at least 0.50 and significant; otherwise, theory decides.

Reporting

All indicator loadings exceeded 0.708 (range 0.814 to 0.932, p < .001), indicating satisfactory indicator reliability. Cronbach’s alpha (0.841 to 0.907), ρA (0.852 to 0.909) and composite reliability ρC (0.903 to 0.941) were above 0.70 and below 0.95, establishing internal consistency reliability. AVE values ranged from 0.723 to 0.843, above the 0.50 threshold, supporting convergent validity (Table 2).

For a cross-check outside the Studio, the reliability and validity tool computes α, CR, ω, AVE, Fornell–Larcker and HTMT from a CFA, the CB-SEM perspective on the same questions.