Use an HTMT value below .85 as the conservative threshold for discriminant validity, and below .90 when the two constructs are conceptually very similar; both cut-offs come from Henseler, Ringle and Sarstedt (2015). Hair et al. (2019) repeat the same pair of values in their PLS-SEM reporting guidelines. A third, statistical test is to check that the bootstrap confidence interval of HTMT does not include 1 (Henseler et al., 2015). Choose the cut-off before you look at the results and explain why it fits your constructs.
What HTMT measures
The Heterotrait-Monotrait ratio of correlations (HTMT) compares how strongly the items of two different constructs correlate with each other (heterotrait correlations) with how strongly items within each construct correlate (monotrait correlations). If items of construct A correlate with items of construct B almost as much as they correlate among themselves, the two constructs are not empirically distinct. Henseler et al. (2015) showed in simulations that the older Fornell-Larcker criterion (Fornell & Larcker, 1981) often fails to detect this problem in variance-based SEM, while HTMT detects it reliably. That study is why HTMT became the default discriminant validity check in SmartPLS output and in thesis examinations.
Where .85 and .90 come from
- HTMT.85: Henseler et al. (2015) based this stricter value on Kline's (2011) guideline that correlations above .85 between constructs suggest a lack of discriminant validity.
- HTMT.90: the more lenient value follows Gold, Malhotra and Segars (2001), and Henseler et al. (2015) recommend it when constructs are conceptually close, for example cognitive and affective satisfaction, or two dimensions of the same higher-order construct.
- HTMT inference: if the upper bound of the bootstrap confidence interval is below 1, the constructs are statistically distinct (Henseler et al., 2015). Many reviewers now prefer testing against .85 or .90 rather than 1, because 1 is very lenient.
How to choose
Ask whether your theory says the two constructs should be clearly different. Perceived usefulness and price, for instance, are different concepts, so .85 is the natural choice. Attitude towards a brand and brand love overlap by definition, so .90 is defensible. Do not switch thresholds after seeing that one pair fails at .85: examiners notice. State in your method section which threshold you apply and why, and apply it to every pair.
Worked example
Suppose construct A (Trust, three items) and construct B (Satisfaction, three items). The nine correlations between A-items and B-items average .52. The three correlations among A-items average .64 and the three among B-items average .58.
HTMT = mean heterotrait correlation / √(mean monotrait A × mean monotrait B) = .52 / √(.64 × .58) = .52 / .609 = .85.
The result sits exactly on the strict cut-off. If Trust and Satisfaction are theorised as distinct but related, .85 is fairly tight here; the student should look at the bootstrap interval. If the 95% interval is [.78, .91], it excludes 1 but includes .90, so a careful write-up would say: "HTMT for Trust and Satisfaction was .85, at the conservative threshold (Henseler et al., 2015). Given the conceptual proximity of the constructs, the .90 criterion was applied, and the value met it." It would also examine cross-loadings to find the items that drive the overlap.
What to do when HTMT is too high
- Look for items that load highly on both constructs and check their wording; an item about "trusting the service to satisfy me" mixes both ideas.
- Consider whether the two constructs should be merged or modelled as dimensions of one higher-order construct, if theory supports that.
- Do not delete items only to lower HTMT; removal must keep each construct's content complete (Hair et al., 2019).
- Report the problem openly if it remains, and discuss it as a limitation.
Reporting example
Discriminant validity was assessed with the HTMT criterion. All HTMT values were below .85 (range .41–.79), and no bootstrap confidence interval included 1, indicating discriminant validity (Henseler et al., 2015).
Run it free
The free PLS-SEM tool computes the full HTMT matrix with bootstrap intervals, flags pairs above .85 and .90 and writes the reporting sentence. Results are compared with the R package seminr.
FAQ
Is HTMT only for PLS-SEM?
No. It is computed from item correlations, so it works for CB-SEM and CFA as well. It became popular through PLS-SEM.
Should I still report the Fornell-Larcker criterion?
Many examiners still expect it. Report it as a secondary check, but rely on HTMT, which performs better in simulations (Henseler et al., 2015).
My HTMT is .88. Is that a fail?
It fails .85 and passes .90. Whether that is acceptable depends on whether you justified .90 before analysis on conceptual grounds.
Can HTMT be above 1?
Yes, it is a ratio, not a correlation. A value above 1 is a clear sign the two constructs are not distinct.
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/10.1177/002224378101800104
- Gold, A. H., Malhotra, A., & Segars, A. H. (2001). Knowledge management: An organizational capabilities perspective. Journal of Management Information Systems, 18(1), 185–214. https://doi.org/10.1080/07421222.2001.11045669
- 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. https://doi.org/10.1108/EBR-11-2018-0203
- 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
- Kline, R. B. (2011). Principles and practice of structural equation modeling (3rd ed.). Guilford Press.