#SmartPLS4 Webinar Day 1: Measurement Model Assessment

Measurement model assessment in Smart PLS starts long before factor loadings and validity tables—data cleaning is treated as the gatekeeper for trustworthy results. Skipping it can cascade into misleading measurement outcomes, including failed convergent validity (low AVE) and weak discriminant validity (constructs that don’t look distinct). The workflow begins with checking minimum/maximum values to catch impossible entries (e.g., an age of 79 when the expected range tops out at 65), then correcting them by tracing back to the questionnaire or fixing data-entry errors. Next comes missing data handling: interpolation is only appropriate when missingness is limited; if a large share of responses is empty, filling values can distort the measurement and should be avoided.

After range checks and missing-data decisions, the process turns to outliers and respondent misconduct. Outliers can be identified via box plots or standardized Z-scores, with extreme standardized values (beyond ±3.3) flagged for removal. Respondent misconduct is handled differently: the method computes standard deviation across items within a construct (e.g., Vision 1–4). If responses show near-zero variation—like repeated patterns (all 5s or all 1s)—the data may reflect inattentive responding. But deletion isn’t automatic; the guidance is to use subject-matter judgment about whether identical responses are theoretically plausible for that population and construct.

Once the dataset is “filtered,” Smart PLS setup begins with defining a workspace, importing the dataset (including scale ranges and missing-value markers if needed), and sanity-checking distribution diagnostics such as skewness and kurtosis (noting that Smart PLS is less dependent on normality than covariance-based SEM, but extreme violations still matter). The first modeling milestone is the measurement model, which evaluates how well survey items represent latent constructs. Four quality criteria drive the assessment: factor loadings (item-to-construct representation), reliability (Cronbach’s alpha and composite reliability/CR, with thresholds commonly around 0.70), convergent validity (AVE, typically requiring >0.50), and discriminant validity (construct distinctiveness).

Factor loadings are checked first, with a common benchmark of ~0.70; items below that level aren’t necessarily deleted immediately, because AVE and CR can still be acceptable and content validity may be harmed by aggressive item removal. Reliability is then verified using alpha (more conservative) and composite reliability (more aligned with PLS’s loading-based logic). Convergent validity is assessed through AVE, which is computed from squared loadings; if AVE is low, the response is not just “delete more items,” but to consider whether the questionnaire design, item wording, or sample characteristics are the real problem.

Discriminant validity is assessed using multiple lenses: HTMT (with typical acceptance below 0.85), Fornell–Larcker (the square root of AVE should exceed inter-construct correlations), and cross-loadings (each item should load highest on its own construct, with differences ideally above 0.10). When discriminant validity fails—such as a high HTMT between Vision and Rewards—the fix can involve diagnosing respondent misconduct (items producing identical patterns across constructs) or removing problematic items with cross-loading (e.g., deleting Reward 4 when it loads too similarly on another construct). The result is a measurement model that can then support the later structural model stage, where path coefficients and hypothesis testing come into play.