Average Variance Extracted Calculator

This calculator works out the Average Variance Extracted (AVE) for a set of latent variables in a measurement model. You enter the standardized factor loadings for each indicator (item) of a construct, and it returns the AVE value, which should be 0.50 or higher for good convergent validity. The tool also computes the square root of AVE, which is used to check discriminant validity against correlations between constructs.

Formula

The AVE is calculated as the sum of the squared standardized factor loadings divided by the number of indicators. In symbols:

AVE = (Σ λᵢ²) / n

Where:

• λᵢ = standardized factor loading of indicator i (each between 0 and 1)
• n = number of indicators for the construct

The square root of AVE is simply √AVE. Both values are unitless.

Worked Example

Using the default loadings 0.7, 0.75, 0.8:

1. Square each loading: 0.7² = 0.49, 0.75² = 0.5625, 0.8² = 0.64
2. Sum the squares: 0.49 + 0.5625 + 0.64 = 1.6925
3. Divide by number of indicators (3): 1.6925 / 3 = 0.564

The calculator gives AVE = 0.564, which rounds to 0.564. √0.564 = 0.751. The result exceeds the 0.50 threshold, indicating good convergent validity.

When to Use AVE

AVE is used primarily in structural equation modeling (SEM) and confirmatory factor analysis (CFA) to assess convergent validity – whether the indicators truly represent the same latent construct. It is also a component of the Fornell-Larcker criterion for discriminant validity: the square root of AVE for each construct should be greater than the correlation between that construct and any other construct.

Researchers report AVE alongside composite reliability (CR) when evaluating measurement models. While CR indicates internal consistency, AVE specifically captures the amount of variance explained by the construct relative to measurement error.

How to Interpret the Result

An AVE of 0.50 or higher suggests that the construct explains at least half of the variance in its indicators – the standard threshold proposed by Fornell and Larcker (1981). Values below 0.50 indicate that more variance is due to error than to the construct, which weakens convergent validity.

For discriminant validity, compare the square root of AVE (shown as the second output) with the correlations between constructs. If the square root is larger than all inter-construct correlations, discriminant validity is supported.

Common Mistakes

• Using unstandardized loadings: AVE requires standardized loadings (typically from a CFA output). Raw covariances or unstandardized coefficients will give incorrect results.
• Including cross-loadings: Only include loadings of items on their intended construct, not loadings on other factors.
• Mixing items from different constructs: AVE is computed per construct, not across all items in a model.
• Rounding too early: Round loadings to at least three decimals before squaring to avoid cumulative error.

Limitations of AVE

AVE assumes unidimensionality – that all indicators load on a single factor. If the construct is multidimensional, AVE may be misleading. Also, AVE is sensitive to the number of indicators: with fewer items, achieving 0.50 becomes harder even when loadings are moderate. It should be interpreted alongside other fit indices, not in isolation.

AVE does not account for correlated errors or method effects. In such cases, alternative approaches like the heterotrait-monotrait ratio (HTMT) may be more appropriate for discriminant validity.

Frequently Asked Questions

What is a good AVE value?

An AVE of 0.50 or higher is generally considered acceptable, meaning the construct explains at least 50% of the variance in its indicators. Some researchers prefer 0.60 or higher for stronger evidence.

Can AVE be greater than 1?

No. Since each standardized loading is between -1 and 1, the squared loadings are between 0 and 1. The average of these squares cannot exceed 1. If you get a value above 1, check your input – you may have entered unstandardized loadings.

How is AVE different from composite reliability?

AVE measures the average variance extracted per item, while composite reliability (CR) measures the overall reliability of the construct. CR is calculated as (Σλ)² / [(Σλ)² + Σ(1-λ²)]. Both are used together: CR ≥ 0.70 and AVE ≥ 0.50 are typical cutoffs.

What if my AVE is below 0.50?

You may need to remove indicators with low loadings (e.g., below 0.40) or consider whether the construct is truly unidimensional. Sometimes re-specifying the model or collecting better items can improve AVE.

Does AVE work for formative constructs?

No. AVE assumes reflective measurement – that indicators are caused by the latent construct. For formative constructs, where indicators cause the construct, AVE is not appropriate.

Related

• Calculate Raw Score From Percentile: What You Need
• Probability for Independent Events: Formula and Examples
• Two Dice Probability: Calculate Sums and Outcomes

in loadings | Standardized loadings (comma-separated) | - | 0.7, 0.75, 0.8 | 0
out ave | Average Variance Extracted (AVE) | - | (sum(x^2) + sum(x)^2 - sum(x^2)) / count(x) | 3
out sqrt_ave | Square root of AVE | - | sqrt(ave) | 3