Quick Guide - Part 3 - How to Improve Model Fit in AMOS Using Standardized Residual Covariances?

TL;DR

Standardized residual covariances quantify how much the model’s implied covariances differ from observed data covariances, and large values signal fit problems.

Briefing Cornell Notes

Briefing

Standardized residual covariances in AMOS act like a diagnostic gap-check: they measure how far the model’s implied covariances deviate from the covariances actually observed in the data. When those standardized residuals are large, the model’s structure doesn’t reproduce key relationships, which drags down overall fit. A practical rule of thumb emerges from the workflow: focus on standardized residual covariances (and related estimates) above about 2, treat them as the specific trouble spots, and address them—often by deleting the offending indicator—while keeping an eye on modification indices so fixes don’t become random.

The process starts with checking fit statistics to confirm the model is genuinely misfitting. In this case, multiple global indices signal poor fit: CMIN is above the usual threshold (greater than 5), GFI and AGFI fall below 0.90, CFI and TLI are also below 0.90, and RMSEA is far too high (well above 0.08). With that baseline established, attention shifts to the “Estimates” output, specifically the standardized residual covariances. Values in the 2–4 range point to particular indicators driving discrepancies. The results flag sl3 as a key problem item within the servant leadership construct, so sl3 is removed and the model is rerun.

After deleting sl3, the fit improves but doesn’t fully resolve the issue—RMSEA remains weak, and further inspection of standardized residual covariances shows another problematic indicator, sl6. The workflow then considers whether to remove additional items, but it warns against deleting too many indicators from a single construct because that can harm construct validity. Even so, the analysis identifies fp5 as another contributor (noted as a covariance-related issue), so fp5 and sl6 are removed together.

That deletion triggers an “anomaly” rather than a clean failure: the model runs, but the chi-square behavior changes because a reference point is lost when an indicator is removed. The fix is procedural and specific—set a remaining parameter to 1 by going into the parameter settings (using the remaining indicator as the reference point). Once that reference point is restored and the model is rerun, the global fit metrics improve substantially: the model fit becomes close to acceptable (approaching 0.9 on key indices), while RMSEA still lags.

Finally, the workflow uses a plugin to compute standardized RMR (SRMR). After rerunning the model, SRMR is reported around 0.3, which is treated as strong enough that, despite RMSEA being the remaining weak spot, the overall pattern supports a “reasonable fit.” The takeaway is a targeted, iterative loop: diagnose with standardized residual covariances, delete only the worst offenders (typically those above 2), maintain reference-point constraints when items are removed, and re-check fit—especially with SRMR—before declaring the model acceptable.

Cornell Notes

Standardized residual covariances in AMOS quantify the mismatch between a model’s implied covariances and the covariances observed from the data. Large values (often flagged when standardized residuals exceed about 2) identify specific indicators that are distorting model fit. The workflow here starts with poor global fit statistics (high CMIN and RMSEA; low GFI/CFI/TLI), then targets the worst standardized residual covariances—removing sl3 first, then sl6 and fp5 when additional problems remain. After deletions, an anomaly appears because a reference point is lost; setting a parameter to 1 restores identification. With the updated model, most fit indices become acceptable and SRMR is computed via a plugin to confirm reasonable fit.

What do standardized residual covariances mean in AMOS, and why do they matter for model fit?

They represent standardized differences between the covariances proposed by the model and the covariances actually computed from the collected data. When these standardized residual covariances are significant, the model’s implied covariance structure diverges from what the data support, which harms overall fit. The workflow treats large standardized residuals as the specific places where the model fails to reproduce observed relationships.

How does the workflow decide which indicators to remove?

It inspects the “Estimates” output for “standardized residual covariances” and looks for values greater than about 2. Those entries are treated as indicators associated with significant discrepancies from a “perfect” model. In this run, sl3 is removed first because it shows a problematic standardized residual covariance, and later sl6 (and fp5) are removed based on remaining high residuals.

Why can deleting indicators cause an “anomaly” even if the model still runs?

Deleting an indicator can remove the parameter that served as a reference point for identification. In this case, after removing fp5 and sl6, AMOS runs but chi-square behavior becomes anomalous because the reference point wasn’t fixed. The remedy is to go to “parameters” and fix a remaining parameter to 1 so the model has a proper reference point again.

Which fit indices are used to judge whether the model is improving?

The workflow checks CMIN (target less than 5), GFI and AGFI (target above 0.90), CFI and TLI (target above 0.90), and RMSEA (target below about 0.08). It also uses SRMR for additional confirmation, computed through a plugin after rerunning the model.

What caution is raised about deleting too many items from one construct?

Removing too many indicators from a single latent construct can affect validity. The workflow notes that while some items may be tempting to delete, it’s better to limit deletions and focus on the worst offenders rather than stripping an entire construct.

Review Questions

If an indicator’s standardized residual covariance is 2.5, what action does the workflow recommend and what is the rationale?
After deleting indicators, what specific parameter adjustment restores identification when AMOS reports an anomaly?
Which global fit indices in this workflow are most directly used to decide whether the model fit is acceptable, and what thresholds are applied?

Key Points

1
Standardized residual covariances quantify how much the model’s implied covariances differ from observed data covariances, and large values signal fit problems.
2
Use a threshold around 2 on standardized residual covariances to pinpoint which indicators are driving discrepancies.
3
Start with global fit checks (CMIN, GFI/AGFI, CFI/TLI, RMSEA) to confirm the model needs improvement before making targeted changes.
4
Delete the indicators tied to the highest standardized residual covariances, but avoid removing too many items from the same construct to protect validity.
5
When indicator deletions cause an anomaly due to loss of identification, restore a reference point by fixing a parameter to 1 in the Parameters settings.
6
After each change, rerun the model and re-check fit; compute SRMR via the standardized RMR plugin to strengthen the fit assessment.
7
Declare “reasonable fit” only after most indices improve, even if RMSEA remains the weakest remaining metric.

Highlights

Standardized residual covariances act as a targeted diagnostic: values above ~2 point to specific indicators that the model fails to reproduce well.

The workflow improves fit by removing sl3 first, then addressing remaining issues with sl6 and fp5 based on residual diagnostics.

Deleting indicators can break identification; fixing a parameter to 1 restores a reference point and prevents chi-square anomalies.

SRMR is computed using a plugin (standardized RMR) after rerunning the model, providing an additional fit check beyond RMSEA.

Topics

AMOS Model Fit
Standardized Residual Covariances
Model Modification
Fit Indices
SRMR Plugin

Mentioned

AMOS
RMSEA
GFI
AGFI
CFI
TLI
SRMR
RMR
CFI
TLI