Guides · Published · Research Corner editorial team

AVE below .50: what to do

Average variance extracted (AVE) should be at least .50, meaning the construct explains at least half of the variance of its items (Fornell & Larcker, 1981; Hair et al., 2019). If AVE is below .50, the usual cause is one or two items with weak outer loadings, below about .70. The standard fix is to inspect those items and remove an item with a loading between .40 and .70 only if doing so raises AVE above .50 without harming content validity (Hair et al., 2022). Items loading below .40 are normally removed (Hair et al., 2022).

What AVE is

AVE is the mean of the squared standardised loadings of a construct's items. A loading of .708 squared is .50, which is why Hair et al. (2019) recommend loadings of .708 or higher: an item at that level shares half its variance with the construct. When AVE drops below .50, more of the item variance is error than construct, and convergent validity is in question. Fornell and Larcker (1981) introduced the measure, together with the rule that each construct's AVE should exceed its squared correlation with any other construct, the familiar Fornell-Larcker criterion.

Worked example

A four-item construct, Service Quality, has standardised loadings of .82, .78, .61 and .55. The squared loadings are .672, .608, .372 and .303, which sum to 1.955.

AVE = 1.955 / 4 = .49, just below .50.

Composite reliability is still fine: (sum of loadings)² / [(sum of loadings)² + sum of error variances] = 2.76² / (7.618 + 2.045) = .79, above .70 (Hair et al., 2019). So the scale is reliable but its convergent validity is borderline. The weakest item is SQ4 at .55. Removing it gives AVE ≈ (.672 + .608 + .372) / 3 = .55. In practice the loadings change slightly when you re-estimate the model, so always re-run it rather than recalculating by hand.

Before deleting SQ4, the student reads its wording. If SQ4 is the only item covering an aspect of the construct (say, "responsiveness"), removing it narrows what Service Quality means, and that is a cost the examiner will ask about. If it duplicates another item or was confusingly worded, removal is easy to defend.

A step-by-step checklist

  1. Check data problems first. Unrecoded reverse items, straight-lining respondents or coding errors depress loadings. Clean the data and re-run.
  2. Sort the outer loadings. Items below .40: remove. Items between .40 and .70: candidates (Hair et al., 2022).
  3. Remove one item at a time, starting with the lowest, and re-estimate. Stop as soon as AVE reaches .50.
  4. Protect content validity. Keep at least three items per reflective construct where possible, and do not remove items that carry a distinct facet of the definition.
  5. Re-check everything. Composite reliability, HTMT and the structural paths can all change after removal.
  6. Report every deletion with the original loading and the reason.

The "AVE above .40 is acceptable if CR is above .60" claim

You will find many theses stating that an AVE below .50 is acceptable when composite reliability exceeds .60, attributed to Fornell and Larcker (1981). The argument is that convergent validity can be judged on composite reliability alone. Be careful: this is a widely repeated interpretation rather than a clearly stated rule, and some examiners do not accept it. If you rely on it, quote the exact page you are citing and present it as a secondary justification after attempting the item-level fix.

What not to do

Reporting example

The initial AVE for Service Quality was .49, below the .50 threshold (Fornell & Larcker, 1981). Item SQ4 had an outer loading of .55, below the .708 guideline (Hair et al., 2019), and duplicated SQ2 in content. After its removal, AVE was .55 and composite reliability was .81, establishing convergent validity.

Run it free

Run your model in the free PLS-SEM tool. It shows outer loadings, AVE, composite reliability and HTMT with sourced cut-offs and lets you drop an item and re-estimate in seconds. For a CFA-based check, open the CB-SEM tool (Vietnamese interface).

FAQ

Is an AVE of .48 really a problem?

Strictly, it is below the threshold. Examine the weakest item; if removal is not defensible, report the value honestly and discuss it.

Does low AVE affect discriminant validity?

Yes, under the Fornell-Larcker criterion a low AVE makes it harder for the square root of AVE to exceed inter-construct correlations.

Does AVE apply to formative constructs?

No. AVE, composite reliability and alpha are for reflective measurement. Formative indicators are judged by weights, VIF and convergent validity with a global item (Hair et al., 2022).

References