What Is AVE? How to Read and Assess AVE in SmartPLS 4
What is AVE used for in SmartPLS
AVE stands for Average Variance Extracted. In PLS-SEM, AVE shows how much of the variance in its associated indicators a latent variable explains on average. You will find this measure in the measurement model assessment, together with outer loading, Composite Reliability, and HTMT.
The formula is the average of the squared outer loadings of the indicators belonging to the same construct:
AVE = sum of squared outer loadings / number of indicators
For example, suppose a construct has three outer loadings of 0,80, 0,75, and 0,70. AVE is calculated from the squared values of these loadings. SmartPLS performs the calculation automatically, so your task is to check that the construct, indicators, and result table are all correct.
AVE is mainly used to assess convergent validity. When indicators genuinely measure the same concept, they should have adequate outer loadings and the construct should explain a meaningful share of their variance. Under the commonly used criterion, AVE from 0,50 upward indicates convergent validity (Fornell and Larcker, 1981). You can also review the broader [SmartPLS](https://do thesis.vn/blog/vi/chu-de/smartpls) topic collection for PLS-SEM analysis. When publishing, use the correct link path shown here: SmartPLS topic.
Preparing the model and data
Before reading AVE, make sure that the model and data have been prepared correctly. AVE comes from the measurement model, so an error in assigning indicators can produce a misleading result even when SmartPLS performs the calculation correctly.
Distinguishing the measurement model from the structural model
The measurement model describes the relationship between a latent variable and its indicators. For example, the “Service Quality” construct may be measured with CLDV1, CLDV2, and CLDV3. The structural model describes relationships between constructs, such as Service Quality affecting Satisfaction.
You read AVE in the measurement model. Measures such as path coefficient, R², f², and Q² belong to the structural model. Do not use R² to decide whether a scale has convergent validity. If you need to review the basic terminology, read What is SEM and PLS-SEM before running the model.
Use consistent indicator names
Indicator names should be short, contain no diacritics or spaces, and clearly identify the construct. For example, the indicators for “Trust” could be named NT1, NT2, and NT3. In a .csv file, the first row contains column names, and the following rows contain each respondent’s answers.
Check the following points before importing the file:
- Each column is one indicator, not a construct calculated as an average.
- Likert responses are coded numerically in the same direction.
- Reverse-coded indicators have been recoded before running the model.
- Do not leave text characters in a numeric data column unless the software is configured to handle them.
- Column names in the
.csvfile match the indicators placed in the model.
If you import data from SPSS, check missing values and value labels again. SmartPLS usually reads .csv files more reliably when the structure is simple, with one header row and no total row at the bottom of the table.
Steps in SmartPLS 4
The steps below apply to SmartPLS 4. SmartPLS 3 uses a different menu layout, so do not compare every button with screenshots from the older version. AVE is meaningful only after you have run PLS Algorithm and checked the measurement model again.
Step 1: Create a project and import the data file
Open SmartPLS 4, create a new project, and import the .csv file. Check the indicator list after the software reads the data. If variable names have been truncated, numeric variables have been recognised as text, or many missing values appear, stop here and correct the original file.
Step 2: Draw the constructs and connect the indicators
Create the constructs in line with your research model, then drag each indicator to its corresponding construct. For a reflective scale, the arrows usually run from the construct to the indicators. This is the common arrangement when indicators reflect the same latent concept.
You need to determine from the theoretical foundation whether your scale is reflective or formative. AVE is commonly used when assessing convergent validity in a reflective measurement model. Do not apply the same interpretation to every formative construct.
Step 3: Run PLS Algorithm
In SmartPLS 4, select the model and run PLS-SEM Algorithm. Keep the default settings if you do not have a methodological reason to change them. After the run finishes, open the measurement model results and find the Construct Reliability and Validity table.
This table usually contains columns for Cronbach's Alpha, rho_A, Composite Reliability, and Average Variance Extracted (AVE). You should also open Outer Loadings to see which indicators may be pulling AVE down.
Step 4: Run Bootstrapping when testing hypotheses
Select Bootstrapping in SmartPLS 4 to test the statistical significance of path coefficient and related results. The PLS-SEM guidance commonly cited in theses uses 5,000 bootstrap subsamples when reporting results (Hair et al., 2022).
Bootstrapping does not make AVE pass automatically. AVE assesses the measurement model, while bootstrapping helps you assess the stability and significance of estimates for the purpose of your analysis.
Step 5: Save results from each run
Export the result tables or capture the important tables before removing an indicator. If you delete an indicator and run the model again, record the reason, the old outer loading, the old AVE, and the new AVE. The committee may ask why an indicator was removed. A record of your data-processing decisions lets you explain the decision instead of presenting only the final output.
Reading the result table
The key table for reading AVE is Construct Reliability and Validity. Each row represents one construct. The Average Variance Extracted (AVE) column gives the construct’s AVE, while the adjacent columns help you assess reliability.
Illustrative output
The table below is illustrative output, not the result of an actual study. The column names follow the format commonly found in SmartPLS 4 output.
| Construct | Cronbach's Alpha | rho_A | Composite Reliability | Average Variance Extracted (AVE) |
|---|---|---|---|---|
| Service Quality | 0,842 | 0,846 | 0,895 | 0,681 |
| Trust | 0,781 | 0,789 | 0,861 | 0,608 |
| Usage Intention | 0,694 | 0,701 | 0,813 | 0,592 |
| Convenience | 0,733 | 0,741 | 0,815 | 0,471 |
In this example, the first three constructs have AVE from 0,50 upward. Convenience has an AVE of 0,471, below the commonly used threshold, so you need to check its outer loadings and scale content. Do not delete an indicator immediately because of one low value. Indicator removal should consider theory, loading, CR, and the effect on discriminant validity.
To understand why AVE is low, open the Outer Loadings table. If one indicator has a clearly lower loading than the others, you can consider it for removal. However, the decision needs a methodological reason and must be recorded. An increase in AVE after deleting an indicator does not by itself prove that the model is better.
AVE does not replace HTMT. AVE answers how strongly a construct converges with its own indicators. HTMT helps assess discriminant validity between two constructs. You can also read AVE in SmartPLS to compare this measure with the other result tables.
Assessment thresholds
The question “what is a good AVE” is often answered with the 0,50 benchmark. You should still read AVE together with outer loading, Composite Reliability, and HTMT instead of separating one measure from the complete measurement model.
| Measure or criterion | Reference threshold | Interpretation | Source |
|---|---|---|---|
| Average Variance Extracted | From 0,50 upward | The construct explains at least half of the indicator variance on average | (Fornell and Larcker, 1981) |
| Outer loading | From 0,70 upward | The indicator has a strong relationship with the construct | (Chin, 1998) or (Hair et al., 2022) |
| Composite Reliability | From 0,70 upward | The construct’s composite reliability is acceptable | (Fornell and Larcker, 1981) |
| HTMT | Below 0,85 or 0,90 | Supports the assessment of discriminant validity between constructs | (Henseler et al., 2015) |
| VIF | Below 5 | Checks for collinearity in the model | (Hair et al., 2019) |
| R² | 0,75, 0,50, 0,25 | Usually interpreted as substantial, moderate, and weak, respectively, in PLS-SEM | (Hair et al., 2011) |
| Q² | Greater than 0 | Indicates predictive relevance under the Q² assessment logic | (Stone, 1974) or (Hair et al., 2022) |
The 0,50 benchmark assesses convergent validity. It is not the only condition for deciding that a scale is acceptable. For example, AVE may exceed 0,50 while HTMT between two constructs is still high, indicating a problem with discriminant validity. Conversely, an AVE that is slightly low should be considered alongside the number of indicators, outer loadings, and theoretical foundation.
In SmartPLS, do not place CFI, TLI, or RMSEA in the same group of PLS-SEM measures. Those fit indices belong to the CB-SEM context and follow separate criteria (Hu and Bentler, 1999). If your study uses AMOS, the reporting approach differs from a model run in SmartPLS. The article what is partial least squares can help you distinguish the two approaches.
How to report the results in Chapter 4
In Chapter 4, present the results in this order: reliability, convergent validity, discriminant validity, and then the structural model. You can place the Construct Reliability and Validity table in the measurement model section, followed by a short interpretation for each group of constructs.
How to write this in your thesis: “The measurement model assessment shows that AVE values for the constructs range from [lowest value] to [highest value]. The values for [list constructs] are all greater than 0,50, meeting the convergent validity criterion proposed by (Fornell and Larcker, 1981). The [construct name] construct has an AVE of [value]. Therefore, the research team also checked the outer loading and theoretical foundation before deciding to [retain or remove] indicator [indicator code].”
Do not write “AVE reached 68%” when you are reading 0,681 without explaining that the value represents extracted variance. You can state that the construct explains approximately 68,1% of the variance in its indicators on average, but use only the actual values from your own output. Do not transfer the numbers in the illustrative table into your thesis.
After the AVE table, add a Discriminant Validity or HTMT table if required by your research design. A construct can have good convergent validity and still fail to differ adequately from other constructs in the model. This is why the committee often asks about HTMT after you report AVE.
Common mistakes
Using AVE to assess the structural model
AVE belongs to the measurement model. When you want to explain how much an independent variable explains of a dependent variable, read R². When you want to assess whether an effect is significant, check the path coefficient, t-value, or p-value in the bootstrapping results.
Seeing AVE below 0,50 and deleting an indicator immediately
A low-loading indicator may reduce AVE, but it may still be important to the scale’s content. Check the loading, CR, HTMT, and theory before deleting it. Removing several indicators only to increase AVE can strip the scale of its original content.
Confusing AVE with outer loading
Outer loading is the relationship between one indicator and its construct. AVE is a combined measure for the entire construct. One table reports the loading of each indicator, while the Construct Reliability and Validity table reports the construct’s AVE.
Mixing CB-SEM criteria into a SmartPLS report
CFI, TLI, and RMSEA commonly appear in CB-SEM references. If you run PLS-SEM with SmartPLS, keep the relevant group of measures: outer loading, CR, AVE, HTMT, VIF, R², f², Q², and path coefficient.
Reporting AVE without a source
When you state the 0,50 threshold, include the source (Fornell and Larcker, 1981). If your supervisor requires Vietnamese-language references, you can compare the terminology and presentation in (Nguyễn Đình Thọ, 2011), but do not omit the original source of the criterion.
Frequently asked questions
What is AVE and what does it assess?
AVE stands for Average Variance Extracted. It assesses the convergent validity of a reflective construct by showing how much indicator variance the construct explains on average.
What is a good AVE?
AVE from 0,50 upward is the commonly used threshold for assessing convergent validity (Fornell and Larcker, 1981). You still need to read it together with outer loading, Composite Reliability, and HTMT because a high AVE does not automatically prove that the entire measurement model is acceptable.
What should I do if AVE is below 0,5?
First, open the Outer Loadings table to find indicators with low loadings, check whether reverse-coded data were handled correctly, and confirm that every indicator was assigned to the correct construct. Then consider removing an indicator only when you have a clear methodological and theoretical reason. Do not keep deleting indicators simply to raise AVE.
How is AVE different from Composite Reliability?
AVE focuses on convergent validity and the share of indicator variance explained by the construct. Composite Reliability focuses on the construct’s overall reliability. The two measures are usually read together in the Construct Reliability and Validity table, but they answer different questions.
Can AVE be used to assess discriminant validity?
AVE mainly assesses convergent validity. To assess discriminant validity, you usually examine HTMT and may also examine the Fornell-Larcker criterion. Under the HTMT criterion, the value is generally expected to be below 0,85 or 0,90, depending on how conceptually close the constructs are (Henseler et al., 2015).
Do I need to run Bootstrapping to calculate AVE?
SmartPLS calculates AVE after you run PLS Algorithm. Bootstrapping is used for inferential tests and to assess the stability of estimates, not to make AVE reach the threshold. If you are testing hypotheses, report bootstrapping using the selected settings and state the number of subsamples.
Open your SmartPLS 4 file, check Construct Reliability and Validity, Outer Loadings, and HTMT in that order, then record the values before changing the model. If you need to run the analysis on your own .sav or .csv file and turn the results into Chapter 4 tables, see M4 data analysis.