HTMT vs Fornell-Larcker: Discriminant Validity Guide
Andrzej Szymkowiak · Last reviewed June 2026
TL;DR: To show discriminant validity, report the HTMT ratio: below .85 (strict) or below .90 (liberal, for conceptually close constructs) supports it, and a value at or above .90 signals a problem. HTMT is now preferred over the older Fornell-Larcker criterion, which can miss a real lack of discriminant validity. Report HTMT as your main test and keep Fornell-Larcker as a familiar second view.
A reviewer just wrote: "Show that your constructs are distinct." Here is exactly how to do it, and which test to use. Discriminant validity asks a simple question: are two constructs in your model really different from each other? If two factors are almost the same thing, your model is in trouble. For years, researchers checked this with the Fornell-Larcker criterion. Today, most reviewers prefer the HTMT ratio. This guide explains both methods in plain language, shows the correct cutoffs, and helps you decide which to report.
What discriminant validity means
Say you measure "trust" and "satisfaction" with survey items (for help choosing validated scales, see how to find a validated scale). Discriminant validity is supported when the items for trust hang together as trust, the items for satisfaction hang together as satisfaction, and the two factors are clearly distinct. If trust and satisfaction correlate almost perfectly, you may be measuring one construct under two names. You usually check this right after you confirm the measurement model from your CFA.
Method 1: The Fornell-Larcker criterion
The Fornell-Larcker criterion (Fornell & Larcker, 1981) compares each factor's variance in its own items with the variance it shares with other factors.
The rule: for each factor, the square root of its AVE must be larger than its correlation with every other factor.
- AVE (average variance extracted) is the amount of variance a factor explains, on average across its items. AVE should be ≥ .50 (Fornell & Larcker, 1981; Hair et al., 2019).
- You take the square root of each AVE and put it on the diagonal of the factor correlation table.
- Off the diagonal you put the correlations between factors.
- For discriminant validity, each diagonal value must be larger than the other values in its row and column.
Worked example. Trust has AVE = .64, so √.64 = .80. Satisfaction has AVE = .58, so √.58 = .76. The correlation between trust and satisfaction is .55. Because .80 and .76 are both larger than .55, the Fornell-Larcker criterion is met.
Method 2: The HTMT ratio
HTMT stands for the heterotrait-monotrait ratio of correlations (Henseler, Ringle, & Sarstedt, 2015). In plain words, it compares the average correlation of items across two different factors with the geometric mean of the average within-factor item correlations of the two factors (not a single within-factor average). If items correlate almost as strongly across factors as within their own factor, the factors are not distinct.
The rule:
- HTMT < .85 supports discriminant validity (strict threshold).
- HTMT < .90 is a more liberal threshold, used when the two constructs are conceptually close.
- A value at or above .90 signals a discriminant validity problem.
You can also test whether the HTMT confidence interval (from bootstrapping) includes 1. If it does not include 1, discriminant validity is supported. A stricter version tests whether the upper bound stays below the threshold (for example .85 or .90) rather than only below 1 (Franke & Sarstedt, 2019).
Worked example. For trust and satisfaction, the HTMT value is .74. Because .74 is below the strict .85 cutoff, discriminant validity is supported. A bootstrap confidence interval of [.61, .83] does not include 1, which confirms the result.
Why HTMT is now preferred
Simulation studies showed that the Fornell-Larcker criterion often fails to detect a lack of discriminant validity, especially when the indicator loadings within a construct vary only a little. In those cases Fornell-Larcker says "all good" when the constructs are actually too similar. HTMT detects these problems far more reliably (Henseler, Ringle, & Sarstedt, 2015). This is why many journals and reviewers now ask for HTMT.
Side-by-side comparison
| Fornell-Larcker | HTMT | |
|---|---|---|
| What it compares | √AVE vs factor correlations | item correlations across vs within factors |
| Cutoff | √AVE > correlations | < .85 (strict) or < .90 (liberal) |
| Sensitivity | can miss real problems | more sensitive, detects them |
| Confidence interval test | no | yes (bootstrap CI should exclude 1) |
| Current preference | older, still seen | preferred today |
Which should you report?
Report HTMT as your main evidence, because it is the stronger test and reviewers increasingly expect it. You can also include the Fornell-Larcker table, since many readers still know it well. Reporting both is safe and common in current papers. Whatever you choose, state the threshold you used and show the values, ideally in a table.
A short APA sentence:
Discriminant validity was assessed with the HTMT ratio. All HTMT values were below the strict .85 threshold (maximum = .74), and no bootstrap confidence interval included 1, supporting discriminant validity.
What to do when HTMT is too high
A high HTMT (at or above your threshold) does not always mean you must drop a construct. First, check whether the two constructs are conceptually close — if they are, the more liberal .90 threshold may be defensible, and you should say why. Next, inspect the items: a single item that cross-loads or shares wording with another factor can inflate HTMT, and rewording or removing it (with a stated reason) can help. If two factors still overlap heavily, consider whether they are really one construct, or whether a higher-order factor explains them both. Run the bootstrap inference test (Franke & Sarstedt, 2019) to see whether the upper confidence bound clears the threshold; a point estimate just over the line with a confidence interval that stays below it tells a different story than a clear failure. Whatever you decide, report the change and the reasoning, not just the final clean table. For a full step-by-step guide — diagnosis, the bootstrap test, and the fixes that survive review — see HTMT above 0.90: what to do.
Common mistakes to avoid
- Reporting only Fornell-Larcker when the journal expects HTMT.
- Treating .85 and .90 as interchangeable without saying which you used and why.
- Skipping the AVE check — both methods assume AVE ≥ .50.
- Reading the HTMT confidence interval against 1 only, when a stricter test compares the upper bound with the threshold.
- Adding a leading zero to HTMT, AVE, or correlations (write .74, not 0.74).
A note on AVE
Both methods depend on AVE, so make sure each factor's AVE is ≥ .50 first (Fornell & Larcker, 1981; Hair et al., 2019). If AVE is below .50, the factor does not explain, on average across its items, even half of their variance, and discriminant validity is hard to claim no matter which method you use.
Frequently asked questions
Is HTMT better than Fornell-Larcker?
For most studies, yes. Simulation work shows the Fornell-Larcker criterion can miss a real lack of discriminant validity, while HTMT detects it more reliably (Henseler, Ringle, & Sarstedt, 2015). Reporting HTMT as your main test, and Fornell-Larcker as a familiar second view, is a safe choice.
What is a good HTMT value?
Below .85 is the strict threshold and clearly supports discriminant validity. Below .90 is a more liberal threshold for conceptually close constructs. A value at or above .90 signals a problem you should explain or fix.
Do I still need to report Fornell-Larcker?
Not always, but many reviewers and readers still expect to see it. Including both tables costs little space and avoids a follow-up request. If you report only one, make it HTMT.
What does the HTMT confidence interval tell me?
A bootstrap confidence interval that does not include 1 supports discriminant validity. A stricter inference test checks whether the upper bound stays below your threshold, such as .85 or .90 (Franke & Sarstedt, 2019).
Can Explonia compute HTMT for me?
Yes. Explonia returns HTMT, the Fornell-Larcker table, and AVE from your CFA or PLS-SEM model — with bootstrap confidence intervals for HTMT in PLS-SEM — and flags any value above your chosen threshold.
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. https://doi.org/10.2307/3151312
Franke, G., & Sarstedt, M. (2019). Heuristics versus statistics in discriminant validity testing: A comparison of four procedures. Internet Research, 29(3), 430–447. https://doi.org/10.1108/IntR-12-2017-0515
Hair, J. F., Black, W. C., Babin, B. J., & Anderson, R. E. (2019). Multivariate data analysis (8th ed.). Cengage.
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. https://doi.org/10.1007/s11747-014-0403-8
How Explonia does this for you
Explonia computes both HTMT and the Fornell-Larcker table from your CFA or PLS-SEM model, along with AVE — and, in PLS-SEM, the bootstrap confidence intervals for HTMT. It flags any value above your chosen threshold (.85 or .90) so you can spot problems at a glance. If you just want to check a few values by hand, the free HTMT calculator works out the ratio for you, no account needed. The "Article description" feature drafts an APA sentence with your actual numbers, which you then edit and keep as your own. You can export the tables to Word, R, SPSS/.sav, or LaTeX. If you currently use SmartPLS for PLS-SEM, see how Explonia compares with SmartPLS. Learn more about our HTMT methods.
Want to check discriminant validity in minutes? Run HTMT and Fornell-Larcker in Explonia and get a ready-to-edit results sentence. Start free →
How to cite this article
If this article helped your research, you can cite it (APA 7):
Szymkowiak, A. (2026). HTMT vs Fornell-Larcker: Discriminant Validity Guide. Explonia. https://explonia.com/blog/htmt-vs-fornell-larcker/
Last updated: June 2026
Run your whole study in one place
Design it, collect it, and run real CFA and SEM on your own data — with a draft APA paragraph you can edit. Free to start.
Start free