Quick Guide - Part 1 - The Theory of Improving Model Fit in CB-SEM (See Description)

Improving model fit in CB-SEM starts with a disciplined audit of the measurement model: low outer loadings first, then modification indices, and only after that a targeted look at standardized residual covariances. When fit is poor, the fastest path is to identify weak indicators—specifically loadings below 0.50—remove the problematic indicator, and rerun the model to see whether overall fit improves. This step-by-step “delete and reassess” approach prevents chasing fit problems blindly and keeps attention on which items actually measure the intended construct.

If low loadings don’t resolve the fit issue, the next lever is modification indices, a technique that has sparked debate in the literature but is still widely used in practice to improve fit. The key constraint is theoretical and statistical: modification indices suggest adding covariances, but those covariances must be drawn between error terms of indicators that belong to the same construct (or a similar construct context), not between error terms and latent variables, and not between error terms from different constructs. Every suggested covariance should be checked against theory—shared unexplained variance should make substantive sense, not just reduce chi-square.

After considering modification indices, the process shifts to standardized residual covariances (available in the estimates output). Values above 2 are treated as a warning sign that the model’s implied covariances differ meaningfully from the covariances observed in the data. Those discrepancies can undermine overall fit, and the suggested remedy is usually deletion of the item tied to the problematic residual covariance. However, deletion must be handled carefully: removing too many indicators can damage content validity and distort how the construct was operationalized. The guidance is to avoid over-deleting and to maintain at least three items per construct.

The transcript also offers practical decision rules for modification indices. A covariance suggestion is considered meaningful if the chi-square value improves by at least 10 (not merely a small numerical change). In large models, modification indices can easily exceed 20, and sometimes 30–50, so the focus should be on “extreme” suggestions rather than every minor recommendation. At the same time, multiple high modification indices within a construct raise a red flag: shared unexplained error may reflect method bias from how data were collected, not a true measurement relationship.

Finally, the guidance ties measurement-model troubleshooting back to data quality and questionnaire design. Before modeling, the data should be checked for incorrect or inconsistent values, missing values should be handled via imputation, and the questionnaire should include enough indicators per construct—often at least four to six—because items may later be removed due to low loadings, cross-loadings, or residual problems. The recommended workflow is therefore: clean data, build the model in AMOS, remove low-loading indicators iteratively, apply theoretically justified error covariances suggested by modification indices, address standardized residual covariances above 2 through selective deletion, and only then report the finalized measurement model. The overall message is that fit improvement is not just statistical tinkering; it must stay anchored to theory, content validity, and data collection quality.