CBSEM using #SmartPLS4 | 6 | Factor Loadings and Model Fit Statistics

Factor loadings and model fit statistics sit at the heart of confirmatory factor analysis (CFA) and structural equation modeling (SEM) because they determine whether observed items meaningfully represent unobservable constructs—and whether the overall measurement and structural setup matches the data closely enough to trust. Factor loadings quantify the effect of an unobservable construct (like job satisfaction) on its indicators. Standardized factor loadings are typically reported because they put indicator weights on a comparable 0–1 scale: a standardized loading of 0.80 implies the construct explains 0.64 (64%) of the indicator’s variance. As a practical rule, standardized loadings above 0.70 (or indicators explaining at least half their variance) are treated as acceptable; weaker indicators may contribute little and can be candidates for deletion, though that decision depends on additional conditions.

Once factor loadings are set, measurement error for each indicator can be derived from the explained variance: measurement error is computed as 1 − r². Lower explained variance means higher measurement error, so factor loadings directly signal how noisy an indicator is relative to the construct it is supposed to measure.

CFA/SEM also requires a metric for each latent variable. In practice, one factor loading per construct is constrained to 1 to set the scale (often called a “reference term”). Without this constraint, SEM estimation can fail with an unidentified error. When comparing multiple groups or samples—such as male versus female—consistency matters: the same indicator should be constrained to 1 in each group so the latent constructs are measured on the same scale across comparisons.

Model fit statistics then assess how closely the model-implied covariance matrix matches the observed covariance matrix. A good fit indicates the specified covariance structure is a close representation of the data; a poor fit suggests the data contradicts the model. Importantly, good overall fit does not guarantee every part of the model is correct. Fit can also be misleading when comparing models with different numbers of indicators per factor: more indicators can make fit harder to achieve, so parsimony is rewarded.

The chi-square (χ²) goodness-of-fit test is a classic starting point, but it is also a “badness of fit” measure: χ² should be non-significant for a good fit. Yet χ² is highly sensitive to sample size, so a relative chi-square (χ² divided by degrees of freedom) is often used. A commonly cited benchmark is a relative χ² between 3 and 5.

Beyond χ², several fit indices are used. Comparative Fit Index (CFI) values above 0.90 indicate good fit and are less affected by sample size. Incremental Fit Index (IFI) and Tucker-Lewis Index (TLI) are also typically considered acceptable when above 0.90. Root Mean Square Error of Approximation (RMSEA) is a badness-of-fit measure where values near zero are best; under 0.05 is good, 0.05–0.08 adequate, and above 0.10 poor. Standardized Root Mean Square Residual (SRMR) is another badness-of-fit metric where values ≤0.05 are good, 0.05–0.09 adequate, and higher values worse.

Finally, there is no single universal “golden rule” for fit thresholds. While Bentler and Bonett (1980) popularized the 0.90 cutoff, later work (including Hu and Bentler, 1999) argued for stricter 0.95 standards, and Marsh and colleagues (2004) pushed for using multiple indices while accounting for sample size and estimation conditions. Even a model that meets fit thresholds can still be misspecified, so fit should be interpreted alongside theory, specification checks, and measurement quality.