الصدق التمييزي في SmartPLS: جدول فورنل ولاركر ومؤشر HTMT

When researchers build their model in SmartPLS, the measurement model is assessed in three stages: reliability, then convergent validity, then discriminant validity. The last stage is where many master’s and PhD students get stuck, because the table may show values they do not know how to read, or it may show that the criterion is not met, and they do not know what to do next.

In this article, we explain discriminant validity in SmartPLS 4 step by step: the Fornell-Larcker criterion, the HTMT ratio and cross loadings. We show examples from real studies we analyzed at the academy, one where discriminant validity was established and one where it was not. Then we explain what to do in the second case and how to report the result in Chapter Four. If you are still comparing software before you start, read The difference between SPSS, AMOS and SmartPLS.

What Is Discriminant Validity?

Discriminant validity means that each latent variable in the model measures a concept that is truly different from the other variables. If you have two variables such as “job satisfaction” and “organizational commitment”, the data must show that the items of each one measure something distinct, not the same concept under two different names.

Why does this matter? Because a relationship between two variables that are not distinct has no meaning. If you find that variable (A) affects variable (B) with a high coefficient, but they actually measure the same thing, you have not discovered a relationship. You have measured the same concept twice. This is why examiners and supervisors ask you to establish discriminant validity before moving on to the structural model and testing the hypotheses.

The difference between discriminant and convergent validity is simple. Convergent validity asks “Do the items of one variable hang together?” and is measured by the average variance extracted (AVE), which should be 0.50 or higher. Discriminant validity asks “Is this variable separate from the others?”. To check reliability and validity in SPSS before moving to SmartPLS, see Reliability and validity testing in SPSS.

Where to Find Discriminant Validity in SmartPLS 4

  1. After drawing the model, choose PLS-SEM algorithm from the Calculate menu and run it with the default settings.
  2. In the results report, open Quality criteria Then Discriminant validity.
  3. You will find three tables there: Heterotrait-monotrait ratio (HTMT),Fornell-Larcker criterion,Cross loadings.
  4. To get confidence intervals for HTMT, run Bootstrapping (5,000 subsamples are recommended), then open Discriminant validity and then Heterotrait-monotrait ratio (HTMT) in its results to see the 2.5% and 97.5% columns.

Before reading these tables, make sure the outer loadings are acceptable (usually 0.708 or higher) and that composite reliability and AVE are within range, because discriminant validity comes after them.

A Real Example Where Discriminant Validity Was Established

To explain the steps, we use the results of a real study we analyzed at the academy in the healthcare sector. It included 403 nurses, and its model had four variables: transformational leadership, self-efficacy, collaborative teamwork and nursing practice improvement. The numbers are exactly as SmartPLS 4 reported them, and we removed any information that could identify the study’s author.

1. The Fornell-Larcker Criterion

The rule: the square root of each variable’s AVE (the value on the diagonal) must be greater than its correlations with all other variables in the same row and column. In other words, a variable shares more variance with its own items than with any other variable.

VariableTeamworkSelf-efficacyPractice improvementTransformational leadership
Teamwork0.810
Self-efficacy0.2250.804
Practice improvement0.3650.5320.765
Transformational leadership0.2650.5320.4060.824

How to read the table: Take the variable “practice improvement”. Its diagonal value is 0.765, and its highest correlation is 0.532 with self-efficacy. Since 0.765 is greater than 0.532, the criterion is met for this variable. In the same way, every diagonal value in this table is greater than all the values below it and in its row, so discriminant validity is established by this criterion.

Important note: Henseler and colleagues (2015) showed that the Fornell-Larcker criterion performs poorly at detecting discriminant validity problems in PLS-SEM, especially when loadings are similar. That is why it is no longer enough on its own, and it is reported together with HTMT, which has become the main criterion.

2. The HTMT Ratio

HTMT is the ratio of the average correlation between the items of two different variables to the average correlation between the items of the same variable. The closer the ratio gets to 1, the closer the two variables are to being a single concept.

Thresholds:

  • Below 0.85: the strict threshold, used when the variables are clearly different concepts.
  • Below 0.90: acceptable when the variables are theoretically close (such as dimensions of job satisfaction).
  • Bootstrapping confidence interval: the more rigorous test is that the upper bound of the confidence interval (the 97.5% column) is below 0.85 or 0.90, depending on the threshold you chose, as Franke and Sarstedt (2019) recommend. The older rule that “the interval must not include 1” is too lenient.

These are the HTMT values from the same study:

VariableTeamworkSelf-efficacyPractice improvement
Self-efficacy0.308
Practice improvement0.5190.776
Transformational leadership0.3440.7100.566

The highest value is 0.776 between self-efficacy and practice improvement, which is below the strict 0.85 threshold, so discriminant validity is established by both criteria. Notice that the HTMT values are higher than the correlations in the Fornell-Larcker table (0.776 vs. 0.532). This is normal, because HTMT corrects for measurement error, which is exactly why it is more sensitive to overlap.

3. Cross Loadings

This table shows the loading of each item on every variable. The rule is that each item should load higher on its own variable than on any other variable. It is the weakest of the three criteria for judging discriminant validity, but the most useful for diagnosis: when HTMT fails between two variables, this table shows you which items load strongly on the other variable and cause the overlap.

A Real Example Where Discriminant Validity Was Not Established

In another study in organizational behavior, the researcher measured several variables at the employee level, including mindfulness, proactive personality, turnover intention and employee performance. These are the HTMT values as they appeared:

VariableMindfulnessProactive personalityTurnover intention
Proactive personality0.930
Turnover intention0.7420.410
Employee performance0.8630.9220.450

Three values exceeded 0.85, and two of them also exceeded 0.90. This means the data do not clearly separate proactive personality, mindfulness and employee performance, even though they are different concepts in theory. This often happens when employees rate positive statements about themselves and about their own performance in the same questionnaire.

Many researchers stop here with a sentence like “the overlap between the variables is theoretically expected” and move on to the hypotheses. That justification alone is not enough, and examiners or journal reviewers often pick up on it.

What to Do If Discriminant Validity Is Not Established

Follow these steps in order:

  1. Review the theory first: are the two variables truly distinct in the literature? If they are theoretically close, 0.90 may be the right threshold instead of 0.85, as long as you state this explicitly with a reference. In the example above even this is not enough, because two values exceeded 0.90.
  2. Check the cross loadings: look for items whose loading on the other variable is close to their loading on their own variable. You will often find one or two items worded too close to the other concept.
  3. Remove overlapping items carefully: one item at a time, then rerun the model and recheck HTMT, AVE and reliability. Do not remove an item if dropping it weakens the content of the concept, and report any removal in Chapter Four with the reason.
  4. Run bootstrapping and check the confidence intervals: if a value is close to the threshold, the upper bound of the confidence interval settles the question.
  5. If the overlap remains: it may be more appropriate to merge the two variables into one, or to build a second-order construct if the theory allows it. This is a theoretical decision to discuss with your supervisor, because it changes the hypotheses.
  6. In future studies: collect performance data from another source (such as the direct manager) instead of self-ratings, to reduce common method bias.

When HTMT Exceeds 1: The Problem Is the Sample

In a third study, some variables were measured at the organization level with a sample of only 15 managers. HTMT values such as 1.01 and 1.15 appeared, and the Fornell-Larcker table showed diagonal values above 1. The square root of AVE cannot logically exceed 1, so these values are not “weak validity”. They are a sign that the model estimation itself is unstable.

The lesson: do not try to interpret or justify these numbers. Instead, review the sample size against the number of variables and items, the missing values, and how the data were aggregated if the study is multilevel. It is best to calculate the right sample size before collecting the data, not after.

How to Report Discriminant Validity in Chapter Four

Present two tables (Fornell-Larcker, then HTMT), state the threshold you used and its reference, and report the highest value. Here is a sample paragraph using the numbers from the first example:

Discriminant validity was assessed using the Fornell-Larcker criterion and the HTMT ratio. Table (9) shows that the square root of the average variance extracted for each variable (from 0.765 to 0.824) is greater than its correlations with the other variables, which meets the Fornell and Larcker (1981) criterion. Table (10) shows that all HTMT values were below the strict 0.85 threshold proposed by Henseler et al. (2015), with the highest being 0.776 between self-efficacy and nursing practice improvement. Discriminant validity is therefore established for all study variables.

Once the measurement model is validated, you move on to the structural model and hypothesis testing. To read path coefficients and p-values correctly, see Interpreting statistical analysis results step by step.

Try the free SmartPLS & AMOS model diagnostic lab

Enter your model results exactly as your software reports them and see which indicators are acceptable and which need attention: loadings, AVE, CR, HTMT, Fornell–Larcker and fit indices, with a short measurement-model paragraph in English and Arabic.

Open the Free Lab

Frequently Asked Questions

What is the difference between discriminant validity and convergent validity?

Convergent validity checks that the items of one variable measure the same thing (AVE above 0.50), while discriminant validity checks that each variable is separate from the other variables.

Which should I use: Fornell-Larcker or HTMT?

Use HTMT as the main criterion because it detects overlap more accurately, and report Fornell-Larcker alongside it because many supervisors and journals still ask for it.

Is the threshold 0.85 or 0.90?

0.85 is the safer choice. 0.90 is acceptable when the variables are theoretically close, as long as you state this and its reference in your thesis.

Why don’t I see the HTMT table in SmartPLS?

Make sure you ran the PLS-SEM algorithm and did not just draw the model, and that each variable has at least two items. HTMT is not calculated for single-item variables.

Does discriminant validity work the same way for formative variables?

No. These criteria are designed for reflective variables. Formative variables are assessed with other criteria, such as VIF and the significance of the weights.

Conclusion

  • Discriminant validity shows that each variable in your model measures a distinct concept.
  • In SmartPLS 4, you find it under Quality criteria, then Discriminant validity.
  • HTMT below 0.85 (or 0.90 for closely related variables) is the main criterion, and Fornell-Larcker is reported alongside it.
  • If it is not established, do not rely on a theoretical justification alone. Review the cross loadings, the items and the confidence intervals.
  • An HTMT value above 1 signals a problem with the model or the sample, and it should not be interpreted.

If your model does not establish discriminant validity, or you want your SmartPLS results reviewed before submitting Chapter Four, the Science Club Academy team can review your model and items with you, and the first consultation is free.

Do you need help analyzing your research data?

The Science Club Academy team helps you prepare your data, analyze it, and interpret the results for master’s theses, doctoral dissertations, and research papers. Learn more about our Integrated statistical analysis service, and the consultation is free.

Send your data details via WhatsApp

References

  • Fornell, C., & Larcker, D. F. (1981). Evaluating structural equation models with unobservable variables and measurement error. Journal of Marketing Research, 18(1), 39–50.
  • Henseler, J., Ringle, C. M., & Sarstedt, M. (2015). A new criterion for assessing discriminant validity in variance-based structural equation modeling. Journal of the Academy of Marketing Science, 43(1), 115–135.
  • Franke, G., & Sarstedt, M. (2019). Heuristics versus statistics in discriminant validity testing: A comparison of four procedures. Internet Research, 29(3), 430–447.
  • Hair, J. F., Risher, J. J., Sarstedt, M., & Ringle, C. M. (2019). When to use and how to report the results of PLS-SEM. European Business Review, 31(1), 2–24.

Mr. Mohamed AbdelfattahAuthor posts

Avatar for Mr. Mohamed Abdelfattah

Mr. Mohamed Abdelfattah holds a master's degree in statistics from the Institute of Statistics. At Science Club Academy, he writes about statistical tests, SPSS, and how they are applied in master's and PhD theses.

No comment

Leave a Reply

Your email address will not be published. Required fields are marked *