Modification Indices in AMOS: Diagnosing Misspecification Without Chasing Model Fit

1. Learning Objectives and Prerequisites for Examining MI

After this guide, you should explain what MI measures, retrieve the correct AMOS table, distinguish residual covariances, cross-loadings, and structural paths, read MI alongside Parameter Change, and decide whether to retain or revise a model using a verifiable decision log.

Inspect MI only after the model is identified, estimation has converged, and no improper solution is present. MI cannot repair root causes such as miscoded data, an excessively small sample, serious distributional violations, or an unclear initial theoretical model (Brown, 2015).

• Finalize a theoretically grounded model before estimation.
• Ensure there are no identification errors, negative variances, or impossible loadings.
• Examine overall fit and standardized residuals.
• Check each item’s content and coding direction.
• Save the original model before making changes.
• Plan to validate the modified model using another sample or sensitivity analysis.

2. What Does a Modification Index Actually Measure?

MI is a score-test statistic for a currently fixed parameter, usually fixed at 0. Under suitable approximation conditions, it predicts the reduction in chi-square if that parameter alone is freed and the model is re-estimated (Sörbom, 1989). A large MI indicates that a constraint may be inconsistent with the data, but does not establish the substantive correctness of the change.

Each MI is a local diagnostic conditional on the entire current model. After a parameter is freed, the remaining MIs may change substantially. Do not add multiple MIs to predict the fit improvement from several changes or simultaneously free every top-ranked parameter.

The threshold 3,84 is sometimes used because of its relationship to a one-df chi-square test at 0,05. It is not a parameter-selection rule. Both MI and chi-square difference tests are sample-size sensitive: with large n, minor misspecification can still produce a significant statistic. Interpret MI with EPC, residuals, and substantive meaning rather than p-values alone (Kline, 2023). MI tables usually contain many tests, so scanning dozens or hundreds of rows increases the risk of modifying the model to fit sample noise.

3. Interpreting MI Alongside Expected Parameter Change

MI primarily reflects statistical evidence against an existing constraint. Expected Parameter Change (EPC), usually labeled Par Change in AMOS, predicts the parameter’s magnitude and direction if freed. Rows with similar MIs can have very different EPCs and therefore different practical priorities.

Unstandardized EPC depends on measurement scales. Standardized Expected Parameter Change (SEPC) makes magnitudes easier to compare across parameter types. Whittaker (2012) shows that standardized EPC can provide more useful parameter-selection information than MI ranking alone.

If MI is large but EPC is tiny, a change may improve chi-square because of sample size without practical importance. If both MI and EPC are substantial but the parameter contradicts theory, retaining the model and discussing the discrepancy may still be appropriate.

4. Enabling and Reading Modification Indices in AMOS

In IBM SPSS Amos 29 Graphics, open Analysis Properties, select Output, enable Modification indices, and enter a display threshold if needed. After estimation, open Text Output and locate Modification Indices. IBM Corp. (2022) describes enabling this table in the AMOS 29 guide.

AMOS divides the table into covariances, variances, and regression weights. In CFA, common suggestions include error correlations and an indicator’s cross-loading onto another factor. In structural models, it may suggest an omitted regression path.

Read variable names carefully. The row e1 <-> e2 represents a residual covariance rather than a correlation between two observed items. The row item3 <--- Factor2 represents a cross-loading. These modifications have different measurement meanings and should not both be described simply as adding arrows.

5. From a Large MI to a Misspecification Hypothesis

A large residual-covariance MI may arise when items share wording, examples, reverse-coding formats, or an unmodeled method source. Correlated errors are reasonable only when a specific shared mechanism beyond the main factor exists.

A large cross-loading MI suggests that an item may also measure another factor. Before adding a cross-loading, examine wording and conceptual definitions. If an item is genuinely ambiguous, revising or removing it in a subsequent study may be clearer than retaining a difficult-to-interpret model.

A large structural-path MI may suggest an omitted variable, a direct relationship beyond mediation, or an inappropriate causal ordering. Cross-sectional data do not independently establish causal direction, so adding paths must begin with theory and research design.

Unusual variance or covariance MIs sometimes signal inappropriate equality constraints, particularly in multigroup models. Examine the invariance objective rather than randomly freeing parameters.

6. A Controlled Model-Modification Workflow

Data-driven specification searches can produce a model that fits the current sample but replicates poorly. MacCallum (1986) cautions that models generated through searching must be treated as post hoc. The objective is therefore to identify explainable, verifiable misspecification rather than obtain the most attractive indices.

Free only one theoretically approved parameter per iteration. Then re-estimate the entire model and check the solution, fit, residuals, loadings, correlations, and new MIs. If the change produces abnormal coefficients or weakens conceptual meaning, return to the previous model.

7. Hypothetical Numerical Example: Two Similarly Worded Items

The following example is illustrative rather than actual research data. A three-factor CFA model has chi-square = 246,80 with df = 87. The standardized residual between SAT2 and SAT3 is 0,19. AMOS reports MI = 31,60 for eSAT2 <-> eSAT3 and Par Change = 0,14.

Both items begin with the same phrase, use the same service example, and are adjacent in the questionnaire. The team anticipated a wording effect in the design log, providing a substantive rationale for residual covariance. After freeing exactly one covariance, chi-square decreases to 216,10 with df = 86. Because M1 is nested in M0 and only one additional parameter is freed, Δχ² = 246,80 - 216,10 = 30,70 with Δdf = 1 is the likelihood-ratio chi-square difference test for the two nested models.

MI is a score test, whereas Δχ² is a likelihood-ratio test after re-estimation. Both have asymptotic one-df chi-square distributions in this example, but MI uses information only at the constrained solution to approximate improvement. Δχ² uses two fully estimated solutions. Thus, MI = 31,60 is close to but need not equal the actual decrease of 30,70 (Sörbom, 1989; Kline, 2023).

Without evidence of shared wording, the same MI is insufficient justification for correlating the errors. Alternative explanations include cross-loadings or an omitted subdimension.

8. Three Cases Requiring Different Responses

For residual covariances, report the common method source and keep the correlation local. Do not connect a chain of errors merely because items belong to the same factor, since the common factor already explains their principal covariation.

For cross-loadings, consider alternative measurement models, wording, ESEM, or item revisions in subsequent research. Adding a cross-loading may be more appropriate than removing an item, but acknowledge that the original simple structure no longer holds.

For structural paths, alternative-model comparisons are valid only when each model has a clear causal rationale. MI does not establish that X causes Y. With cross-sectional data, path direction still depends on design and theory.

Saris et al. (2009) emphasize detecting local misspecification rather than relying solely on overall fit. A model with acceptable global fit may still contain important incorrect parameters, while one large MI does not automatically invalidate the entire model.

9. From Local Fit to Out-of-Sample Replicability

Global fit indicates overall model–data discrepancy, while standardized residuals and MI locate concentrations of discrepancy. A model with acceptable overall fit may still have a large residual for a crucial variable pair. Conversely, reducing one large MI may not repair the entire misspecified structure.

Kline (2023) emphasizes that modifying CFA models using the same data progressively shifts the analysis from confirmation to exploration. MacCallum, Roznowski, and Necowitz (1992) examine specification searches across repeated samples and show that selected changes can be inconsistent and cross-validation results highly variable. Every path added after inspecting MI should therefore be treated as a new hypothesis rather than a confirmed structure. The strongest evidence is testing the modified model in an independent sample; otherwise, use a holdout sample or at least explicitly state the lack of validation.

Do not use MI to construct a causal argument. A suggested structural path indicates only that a constraint of 0 is locally inconsistent with the current covariance structure. Causal direction still requires research design, temporal ordering, no-confounding assumptions, and theory.

10. A Decision-Log Template for Each Modification Iteration

A decision log makes specification searching verifiable. Each row records the model before modification, signal source, proposed parameter, substantive rationale, anticipated effect, decision, and result after re-estimation.

MacCallum (1986) cautions that data-driven specification searches can capitalize on sample noise. The log must therefore also record considered but rejected changes. This prevents a post hoc narrative describing only decisions that produced attractive fit.

11. Errors That Produce Attractive but Hard-to-Defend Models

12. APA 7 Reporting Template and Final Checklist

Every MI-based change is post hoc and must be named accordingly. Reports should let readers reconstruct the decision sequence rather than see only the final model.

• Save and report the initial hypothesized model.
• Check residuals before reading MI.
• Record MI with EPC or Par Change.
• Correctly classify residual covariances, cross-loadings, and paths.
• Write a substantive rationale before modification.
• Free only one parameter per iteration.
• Re-estimate the entire model and check the solution.
• Assess the effect on structural meaning.
• Explicitly label modifications post hoc.
• Validate using other data or state the absence of validation.