12. SEMinR Lecture Series - Evaluating Reflective Measurement Model - Step 1: Indicator Reliability

Evaluating a reflective measurement model in PLS-SEM starts with a practical checklist: confirm indicator reliability, then check internal consistency, followed by convergent validity (via AVE) and discriminant validity (via HTMT). The session frames measurement models as the part of SEM that tests whether constructs are measured reliably and validly, distinct from the structural model that tests relationships among constructs. For reflective models, the quality gates are organized into four steps—indicator reliability, internal consistency reliability, convergent validity, and discriminant validity—each tied to specific statistics and decision rules.

The walkthrough uses a worked example with constructs such as vision development, rewards, collaborative culture, and organizational performance, where all constructs are treated as reflective. It emphasizes that before any reliability or validity statistics can be interpreted, the measurement and structural model must be correctly specified and successfully estimated in the SEMinR workflow. That means loading the SEMinR library, loading the dataset into an object, defining the measurement model by listing constructs and their indicator items, defining the structural paths between constructs, and then running the estimation function. Results are retrieved through a summary object (created via the summary function), and the session stresses that outputs aren’t automatically displayed—you extract them from the stored summary object.

A key operational point is convergence checking. After estimation, the algorithm’s iterations are inspected to ensure the PLS algorithm stopped because the stop criterion was reached—not because it hit the maximum iteration cap. In the example, convergence occurred in five iterations, well below the default maximum of 300, which the session treats as a sign the solution is stable. If convergence fails, two likely causes are highlighted: an overly strict stop criterion (set too small) or data issues such as too small a sample size or indicators with many identical response values that can create singular metrics.

Once the model is estimated and convergence is confirmed, the session moves to indicator reliability. For reflective measurement, indicator reliability is assessed by how much of each indicator’s variance is explained by its construct—computed as the squared indicator loading (because loadings represent the bivariate correlation between an indicator and its construct). The commonly used rule is that squared loadings should exceed 0.50, meaning each indicator explains more than half of its variance through the underlying construct. The session also warns against automatic deletion of indicators that fall below 0.70: researchers should examine whether removing an item actually improves internal consistency reliability and convergent validity enough to justify the change, while also considering content validity. Indicators with extremely low loadings below 0.40 are treated as candidates for elimination, but the decision should still be made carefully.

In the example results, the squared loadings are all above 0.50 (e.g., values around 0.80), so indicator reliability is considered established. The session then signals a transition to construct-level reliability and validity checks as the next step in the reflective measurement model evaluation sequence.