What Is Average Variance Extracted and What Value Is Acceptable

What Is Average Variance Extracted

Average Variance Extracted, usually abbreviated as AVE, is an index used to assess the convergent validity of a latent variable in the measurement model. In simple terms, AVE shows how much variance a latent variable explains, on average, in the observed variables used to measure it.

For example, the construct “Satisfaction” is measured by SAT1, SAT2, SAT3, and SAT4. If Satisfaction has a high AVE, these observed variables tend to reflect the same underlying construct. If AVE is low, the construct explains only a limited amount of the information in the observed variables.

In PLS-SEM, AVE is usually calculated from the outer loadings of the observed variables. The general formula is:

AVE = sum of squared outer loadings / number of observed variables

Suppose a latent variable has four outer loadings of 0.70, 0.75, 0.80, and 0.65. You square each loading, add the results, and divide the total by four. SmartPLS performs this calculation automatically, so you do not need to calculate it manually when the results table already reports it.

AVE is not the percentage of respondents who selected an option, and it is not the mean score of a Likert scale. It belongs to the group of measurement model assessment indices and is usually examined alongside outer loading, Composite Reliability, Cronbach's Alpha, and HTMT. If you are reviewing the idea of a scale before running your model, see what is a scale to distinguish a scale, a latent variable, and an observed variable.

The Meaning of Average Variance Extracted in Quantitative Research

AVE answers a specific question: do the observed variables converge strongly enough on the latent variable they were designed to measure. This is part of the measurement model assessment, before you interpret path coefficients and hypotheses in the structural model.

When AVE reaches the threshold, you have additional evidence that the latent variable explains a meaningful share of the variance in its indicators. This result is usually reported with Composite Reliability when assessing convergent validity. Under the Fornell and Larcker criterion, CR of 0.7 or above and AVE of 0.5 or above are commonly used benchmarks for assessing reliability and convergent validity (Fornell and Larcker, 1981).

AVE only addresses convergent validity. It does not tell you whether one construct is sufficiently distinct from another construct. To assess discriminant validity, you need to read HTMT or the Fornell-Larcker criterion. HTMT below 0.85, or below 0.90 for theoretically close constructs, is a commonly cited benchmark (Henseler et al., 2015).

AVE does not replace Cronbach's Alpha. Cronbach's Alpha focuses on the internal consistency of a scale, while AVE focuses on the amount of variance in the observed variables explained by the latent variable. Both indices may appear in the same table, but they answer different questions.

In SmartPLS, you usually examine AVE after checking the data, variable names, and measurement model. If the file contains many missing values, an observed variable was entered with the wrong direction, or indicators were assigned to the wrong construct, AVE may be low because of a data problem rather than only a theoretical problem. Issues such as outliers should also be checked separately. See outliers when your data contain unusual observations.

What Average Variance Extracted Value Is Acceptable

The common benchmark for AVE is 0.50 or above, based on the convergent validity criterion proposed by Fornell and Larcker (Fornell and Larcker, 1981). In modern PLS-SEM, AVE should be read together with outer loading, CR, and discriminant validity rather than using one number to judge the entire model.

The table below summarises commonly used thresholds. Place each threshold beside its source when writing Chapter 4, especially if your supervisor asks you to justify the criteria.

You should not write that an AVE of 0.49 automatically makes the model invalid. The conclusion should also consider the number of observed variables, outer loading, CR, and the theoretical basis of the scale. If AVE is only slightly low but the loadings and CR have defensible reasons for acceptance, report the limitation clearly and discuss it with your supervisor. If low AVE occurs together with several low loadings, the issue is more likely to lie in the scale or the data.

Do not use CFI, TLI, or RMSEA thresholds to assess AVE. These indices belong to the fit-index group for CB-SEM, are commonly associated with AMOS, and are not substitutes for AVE in SmartPLS. The CFI, TLI, RMSEA, and SRMR criteria are presented in the source for CB-SEM (Hu and Bentler, 1999).

How to Read Average Variance Extracted in the Output

In SmartPLS 4, run the model by opening the project, checking the Measurement Model, and selecting Calculate > PLS-SEM Algorithm. After the algorithm finishes, open the results under Quality Criteria and select the Construct Reliability and Validity table. The columns are usually named Cronbach's alpha, rho_A, Composite reliability (rho_c), and Average variance extracted (AVE).

The table below is illustrative output, not the result of a real study. Use the Average variance extracted (AVE) column to read each latent variable, then compare it with the 0.50 threshold.

In this example, the AVE of all four constructs is greater than 0.50. You can state that the scales meet the convergent validity criterion based on the benchmark (Fornell and Larcker, 1981). You still need to examine the Outer Loadings table and the discriminant validity results. An acceptable AVE does not mean that every observed variable should automatically be retained.

When reading Outer Loadings, match each variable code to the correct construct. A low loading for SAT3 may pull the AVE of Satisfaction down, but deleting SAT3 immediately may not be the right choice. Check the item content, the number of indicators remaining, and the theoretical model.

If you run the analysis in SmartPLS 3, the menu names and the location of some tables may differ from SmartPLS 4. Focus on the table name Construct Reliability and Validity and the row Average variance extracted (AVE) instead of searching only for an interface screenshot that exactly matches an online guide.

What to Do When Average Variance Extracted Does Not Meet the Threshold

First, check the outer loading of each observed variable. An item with a low loading may be a candidate for review, but the decision to delete it should consider both the scale content and the methodological argument. Save the results table before deleting anything so that you can explain each round of processing.

Check the Coding and Original Data

Open the .csv or .sav file and check the columns belonging to the same construct. Confirm whether reverse-coded variables have been recoded, whether values outside the Likert scale appear, and whether the variable names in SmartPLS match the questionnaire. A variable entered in the reverse direction can cause its loading to fall sharply.

Review the Outer Loading and Item Content

If an indicator has a low loading, read the wording of the observed variable again. The item may be too broad, difficult to understand, or unsuitable for the research context. Deleting an item only to increase AVE is easy to challenge because you are changing the scale after seeing the results.

Rerun the Model After Establishing a Clear Reason

If you decide to remove an item, run PLS-SEM Algorithm again and check AVE, CR, outer loading, and HTMT. Record the code of the deleted item, the reason for deletion, and the results before and after the change. A model that looks better numerically but lacks a theoretical basis still creates risk at the defence.

Retain an Item When Deleting It Removes Important Construct Content

A scale may need several indicators to cover a construct adequately. If deleting an item makes the construct too narrow, discuss the decision with your supervisor instead of continuing to delete items until AVE exceeds 0.50. In a thesis, transparent reporting of an item that does not meet the threshold is often more defensible than hiding the rounds of trial and error.

Distinguishing Average Variance Extracted from Composite Reliability

AVE and Composite Reliability often appear beside each other in the same SmartPLS table, so they are easy to confuse. Composite Reliability shows whether the observed variables in a construct are consistent when measuring that construct together. AVE shows how much variance the construct explains, on average, in its observed variables.

A construct with acceptable CR but an AVE below the threshold can occur. This indicates that the indicators have relatively good consistency, but the amount of variance explained by the construct is insufficient under the AVE criterion. Check the loadings and the scale content when deciding what to do. Do not use CR as a substitute for AVE.

AVE also differs from R². AVE belongs to the measurement model and concerns the relationship between a latent variable and its observed variables. R² belongs to the structural model and shows how well the predictor variables explain an endogenous variable. In PLS-SEM, R² may be interpreted using the levels 0.75, 0.50, and 0.25 in some guidelines (Hair et al., 2011), but those levels do not apply to AVE.

AVE also differs from extracted variance in EFA. EFA usually reports Total Variance Explained, while AVE appears in the assessment of the PLS-SEM measurement model. Total variance explained in EFA and Average Variance Extracted are related to the idea of explaining variance, but they are calculated in different steps and models.

Finally, AVE is not directly related to a moderator or mediator. A moderator changes the strength or direction of a relationship, while a mediator explains an intermediate mechanism. If your research topic includes these constructs, keep measurement assessment separate from relationship testing instead of including them in your conclusion about AVE. See mediator when you need to distinguish these concepts in the model.

Common Errors

Reporting AVE without reporting CR and loading. A single AVE table is not enough to assess the measurement model. At minimum, report the construct, Cronbach's Alpha, rho_A if you use PLS-SEM, Composite Reliability, and AVE.

Writing that “higher AVE is always better” without checking the scale content. A high AVE can occur alongside indicators that are extremely similar. Check whether the scale still covers the research construct adequately.

Using AVE to conclude the research hypotheses. AVE supports scale assessment only. Acceptance or rejection of H1 and H2 should be based on the path coefficient, t-statistic, p-value, or bootstrapping confidence interval according to the analysis design.

Deleting several variables at once without saving intermediate results. When your supervisor asks why SAT2 and SAT4 were deleted, it will be difficult to answer if you have no before-and-after table. Save each version of the data and model with a clear name.

Mixing SmartPLS criteria with AMOS criteria. SmartPLS focuses on PLS-SEM, while AMOS is commonly used for CB-SEM. Reading the correct family of indices helps you avoid placing CFI, TLI, or RMSEA in the AVE assessment section.

Entering the decimal separator incorrectly. A file that inconsistently uses commas and periods can cause the data to be read as text or as incorrect values. Before running the model, open the first few rows of the file and check the data type of every observed variable.

Frequently asked questions

What is Average Variance Extracted?

Average Variance Extracted is an index showing the average amount of variance in observed variables explained by a latent variable. It is commonly used to assess convergent validity in the PLS-SEM measurement model.

What Average Variance Extracted value is acceptable?

An AVE of 0.50 or above is a commonly used threshold for convergent validity (Fornell and Larcker, 1981). You should still examine it together with outer loading, Composite Reliability, and discriminant validity before concluding that the scale meets the criteria.

Does an AVE below 0.5 mean that I should delete an item immediately?

No. Check the outer loading, the data, the coding of reverse-coded variables, and the theoretical content of the scale. If you delete an item, explain the methodological reason and report the processing steps so that your supervisor can review them.

Where can I read AVE in SmartPLS?

After running Calculate > PLS-SEM Algorithm, open Quality Criteria and find the Construct Reliability and Validity table. The column to read is Average variance extracted (AVE).

Is AVE the same as Composite Reliability?

No. Composite Reliability assesses the composite reliability of the scale, while AVE assesses the average amount of variance explained by the latent variable in the observed variables. The two indices are usually reported together, but they do not replace each other.

Is AVE used for EFA or SmartPLS?

AVE is usually used when assessing the PLS-SEM measurement model and may also appear in related SEM reports. When running EFA in SPSS, you usually read KMO, Bartlett's Test, eigenvalue, factor loading, and Total Variance Explained, rather than using AVE as the main index.

Open the Construct Reliability and Validity table in your SmartPLS file, compare each AVE row with outer loading and CR, and save the results version before making changes. If you need to run the analysis on your own .sav or .csv file, M4 data analysis performs this step using the SPSS and SmartPLS workflow.