What to Do When AVE Is Below 0.5 in SmartPLS

The problem you are seeing

You run the PLS Algorithm in SmartPLS, open the Construct Reliability and Validity table, and see that the AVE column for a latent variable is below 0.5. For example, the PU construct has AVE = 0.442 while Composite Reliability is still 0.781. This often happens when some indicators have low outer loadings or when the indicators in the same scale do not reflect the construct consistently.

AVE is used to assess the convergent validity of a latent variable. In practical terms, it shows how much of the average variance in its indicators is explained by the latent variable. An AVE of 0.5 or above is commonly used when assessing a PLS-SEM measurement model, based on (Fornell and Larcker, 1981). When AVE is below 0.5, do not delete an indicator immediately. Review the outer loading, Composite Reliability, item content, and theoretical model together.

If you are looking at the overall results screen, you can also review the posts in the SmartPLS topic. This post stays with the specific problem of low AVE and the order of decisions that you can explain to your supervisor.

Common causes

Low AVE often appears because one indicator with a low outer loading pulls down the average for the whole scale. Another cause is an incorrectly entered measurement direction, especially for a reverse-coded item, or data in the .csv file being read as text. A scale may also have been assembled from different contexts without sufficient adjustment for your research topic.

Check the table below before changing the model. This helps you separate a data problem from a genuine problem with the scale.

In PLS-SEM, an outer loading of 0.7 or above is commonly used when assessing an indicator (Chin, 1998). However, a loading between 0.4 and 0.7 does not automatically mean that the indicator must be deleted. The decision also depends on Composite Reliability, the AVE after deletion, and the substantive meaning of the scale.

AVE is also different from Cronbach's Alpha. Alpha reflects a form of internal reliability, while AVE focuses on the proportion of indicator variance explained by the latent variable. A scale can therefore have an acceptable Alpha and still have a low AVE.

How to handle it in priority order

Work from the cause that changes the model least to the cause that could remove important construct content. After every change, record the deleted indicator, its outer loading before and after the change, the new AVE, Composite Reliability, and the reason for the decision. Do not delete several indicators and only then inspect the output, because you will no longer know which indicator actually changed the result.

Check the data file and measurement direction

Open the .csv file that you imported into SmartPLS. Each row should represent one respondent, each column should represent one indicator, column names should be unique, and Likert values should remain within the range you designed. If one column contains text such as Disagree while the other columns contain numbers from 1 to 5, SmartPLS may not process the data in the way you expect.

Next, compare the reverse-coded items with the original questionnaire. If an item has a negative meaning, recode it so that a high score points in the same direction as the other items. For a Likert scale from 1 to 5, the usual reverse-coding rule is new score = 6 minus old score. Perform this operation on a copy of the data, keep the original file, and record which column was recoded.

Review Outer Loadings before deleting an indicator

In SmartPLS 4, select Calculate > PLS-SEM Algorithm, keep the default settings that fit your model, and run it. Open Results > Outer Loadings to review the loading for each indicator. Mark the item with the lowest loading, any item with a negative loading, and any item whose content is furthest from the definition of the latent variable.

If a loading is below 0.4, it is the first candidate to review. If the loading is between 0.4 and 0.7, test what happens to AVE and Composite Reliability after deletion. For a loading of 0.7 or above, delete the indicator only when there is a clear problem with its content, data entry, or measurement structure.

Test one indicator at a time, with a reason

Create a copy of the model before editing it. Delete the indicator with the lowest loading, run PLS-SEM Algorithm again, and compare AVE, Composite Reliability, and the number of remaining items. If AVE increases but Composite Reliability falls sharply or the scale is left with only two items, the new result may not be theoretically better.

You can use the post what to do when the model does not converge if changing AVE produces a warning about the measurement model. A defensible change needs both quantitative evidence and a content-based reason, such as an item being inconsistent with the construct or having been mistranslated.

Reassess the complete scale

After each round, open Construct Reliability and Validity. This table usually contains the columns Cronbach's Alpha, rho_A, Composite Reliability, and AVE. CR of 0.7 or above and AVE of 0.5 or above are commonly used benchmarks for composite reliability and convergent validity (Fornell and Larcker, 1981).

Also check what to do when discriminant validity is not achieved if two latent variables become too similar after an item is removed. If HTMT exceeds the threshold, review what to do when HTMT is greater than 0.9. An acceptable AVE does not resolve a discriminant validity problem.

Compare the scale with its source and original model

Before finalising the model, reread the construct definition, hypotheses, and scale source. An item with a low loading may measure a dimension that the construct must include, so retaining it and explaining the result may be appropriate. Conversely, an item with a high loading should not be retained solely because it increases AVE if its code was entered incorrectly or it does not fit your research context.

If you use a reflective scale, the indicators generally represent manifestations of the same construct. A formative measurement model follows a different logic, so do not mechanically delete indicators just to increase AVE. Identify the measurement model type before making changes.

Recheck the structural model after editing

Once you have finalised the measurement model, run the PLS Algorithm and bootstrapping again. Deleting an indicator can change the path coefficient, R², and the significance of a hypothesis. If bootstrapping produces an error, review what to do when bootstrapping fails, but do not rerun the model repeatedly in search of a visually attractive p-value.

The table below is illustrative output, not the result of a specific study. The column names follow the output groups commonly used in SmartPLS, so you know what to record after each round of decisions.

In this example, AVE increases from 0.442 to 0.548 after PU2 is deleted, but this is only an illustration of how to read the output. Do not copy these numbers into Chapter 4. Your figures must come from the .sav file, .csv file, and output produced by the model you are running.

When you should collect more data

Collecting more data is not the first solution for every low AVE problem. If the cause is an unreversed item, an incorrect data type, or a mismatched variable code, adding respondents will not correct the error. Fix the data and run the model again first.

Consider collecting more data when the current sample is too small for the number of indicators, many responses are incomplete, respondents appear to have selected the same response for most items, or the sample does not match the target population. Even then, a larger sample does not guarantee that AVE will exceed 0.5. If the scale content is unsuitable, additional responses may only make the problem more visible.

When an item is conceptually important but has a low loading, deleting it can remove part of the construct content. A more defensible option may be to report the current result, state the limitation, and discuss possible adjustments with your supervisor. Do not collect more data solely to reach a benchmark without a methodological reason.

How to explain the decision to your supervisor

You can use the paragraph below and replace the placeholders with your actual figures:

How to write this in your thesis: “The initial assessment of the measurement model showed that the AVE of the [CONSTRUCT NAME] scale was [INITIAL AVE], below the 0.5 threshold. Inspection of the Outer Loadings table showed that [INDICATOR CODE] had an outer loading of [LOADING], the lowest loading in the scale. After jointly considering the theoretical basis, indicator content, and Composite Reliability, the study removed [INDICATOR CODE] from the model. The rerun produced an AVE of [POST-DELETION AVE] and Composite Reliability of [POST-DELETION CR]. Therefore, the scale was used for the structural model assessment.”

If you retain an item while AVE remains low, state the reason clearly and do not describe convergent validity as acceptable if it has not reached the benchmark you are using. You can write: “The [CONSTRUCT NAME] scale had an AVE of [AVE], below the reference benchmark. However, the indicators were retained because of [CONTENT-BASED REASON], and this limitation is reported in [SECTION NAME].” This is more transparent than deleting indicators repeatedly without recording the history.

Preventing the problem during scale design

When designing the questionnaire, write a short definition for each latent variable and note which aspect each indicator measures. Indicators within the same scale should use a consistent direction unless you deliberately include reverse-coded items and prepare a clear recoding procedure.

Before official data collection, check the questionnaire with a small group to identify unclear wording, double negatives, or terms that are unfamiliar to the target respondents. This does not replace PLS-SEM assessment, but it can reduce wording and data-entry errors.

When collecting data, assign stable variable codes from the beginning. Do not change PU3 to PU_03 between the questionnaire and the data file. After exporting the .csv, open it in spreadsheet software and check the column names, number of rows, blank cells, and values outside the scale before importing it into SmartPLS.

Finally, save three separate versions: the original data, the cleaned data, and the model after each round of indicator removal. A folder structure such as v1_original, v2_recode, and v3_remove_PU2 will help you explain why the current output differs from the initial output.

Common mistakes when fixing low AVE

The most common mistake is deleting an indicator because AVE has not reached the benchmark without checking the outer loading. AVE is an aggregate result, while the decision usually begins with each indicator. You can also accidentally delete an item with a low loading that is necessary for the construct content.

Another mistake is deleting several items in the same run. This prevents you from tracking the effect of each change and may leave the scale with too few indicators. Do not mix CB-SEM criteria such as CFI, TLI, or RMSEA into a SmartPLS assessment, because those indices belong to a different evaluation framework.

Some students rerun the model many times but do not save the output from each round. Save screenshots or export a report containing Outer Loadings, Construct Reliability and Validity, Discriminant Validity, and bootstrapping results after each major decision. If discriminant validity becomes problematic after editing, address that issue separately instead of continuing to delete indicators.

Frequently asked questions

Is an AVE below 0.5 an automatic reason to delete an indicator

No. Review the outer loading for each indicator, Composite Reliability, the number of remaining items, and the theoretical meaning of the scale. Deleting an indicator only to increase AVE can remove important scale content.

Is AVE = 0.48 acceptable

AVE = 0.48 is still below the 0.5 benchmark commonly cited from (Fornell and Larcker, 1981). Whether to retain the scale should be based on the complete evidence, including the loadings, CR, scale source, and explanation to your supervisor. Do not change the benchmark to acceptable simply because the difference is small.

How should I interpret acceptable Composite Reliability with low AVE

Acceptable CR indicates that composite reliability may be adequate, but it does not automatically establish convergent validity. Report both statistics, state that AVE is low if it is, and explain why you retained or removed the indicators.

Should I delete an item with an outer loading of 0.6

Do not delete it immediately. Run the model with and without the item, then compare AVE, CR, the number of indicators, and the construct content. If the item is theoretically necessary, retaining it may be more appropriate, provided that you report the limitation clearly.

What if AVE increases after deleting an indicator but HTMT becomes greater than 0.9

Reassess discriminant validity between the latent variables. A higher AVE does not mean that the overall model is sound. Check the content of the scales, the correlations between constructs, and HTMT using the relevant guidance before finalising the model.

Open your data file, record AVE, CR, and all Outer Loadings from the current run, then review each indicator in the order above before deleting anything. If you need the analysis run on your own .sav or .csv file, see DoThesis M4 data analysis.