What is a good HTMT value?

Below 0.85 for conceptually distinct constructs, or below 0.90 for conceptually similar ones. Lower is better: it means the two constructs are clearly separate.

What is the difference between HTMT and the Fornell-Larcker criterion?

Both assess discriminant validity, but the HTMT is far more sensitive. Simulation studies show the Fornell-Larcker criterion often fails to detect a lack of discriminant validity that the HTMT catches, which is why the HTMT is now the recommended standard.

When should I use HTMT2 instead of HTMT?

Use HTMT2 when the indicators of a construct have substantially different loadings (they are not tau-equivalent). It relies on the geometric mean and is less biased in that case.

PLS-SEM METHOD

The HTMT ratio: assessing discriminant validity

The heterotrait-monotrait ratio of correlations (HTMT) is the modern standard for checking whether two constructs in a PLS-SEM model are empirically distinct. This page explains what it measures, how to read it against the 0.85 and 0.90 thresholds, its variants, and the pitfalls that trip up most first submissions.

What the HTMT ratio measures

Discriminant validity asks a simple question: are two constructs that your theory treats as different actually different in the data? The HTMT estimates what the correlation between two constructs would be if they were measured without error. It is the ratio of the average correlation of indicators across the two constructs (heterotrait) to the average correlation of indicators within each construct (monotrait).

The closer that ratio is to 1, the more the two constructs overlap, and the weaker your discriminant validity. Introduced by Henseler, Ringle and Sarstedt (2015), the HTMT detects discriminant-validity problems that the older Fornell-Larcker criterion and the examination of cross-loadings routinely miss.

A worked example

Two constructs, Trust and Quality, are measured with reflective indicators. The average correlation of their indicators across the two constructs (heterotrait), relative to the average within each construct (monotrait), gives HTMT = 0.78. Since 0.78 is below the conservative 0.85 threshold, the two constructs are empirically distinct.

If the value had come out at 0.93, you would conclude the two constructs overlap too much to be treated as separate, and you would revisit the measurement model.

HTMT(Trust, Quality) = 0.78, below the 0.85 threshold, so the two constructs are distinct.

How to read the thresholds

You compare each pairwise HTMT value against a cutoff. Which cutoff applies depends on how conceptually similar the two constructs are.

HTMT < 0.85: The conservative threshold. Use it when the two constructs are conceptually distinct. A value below 0.85 is clear evidence of discriminant validity.
HTMT < 0.90: The liberal threshold. Defensible when the two constructs are conceptually close (for example, cognitive and affective attitude).
HTMT ≥ 0.90: A discriminant-validity problem. The two constructs likely capture the same thing. Consider merging them, dropping overlapping indicators, or rethinking the measurement model.

The choice between 0.85 and 0.90 must be justified on theoretical grounds, before you see the result, not chosen afterwards to secure a pass.

HTMT, HTMT2 and statistical inference

HTMT

The original ratio, based on the arithmetic mean of indicator correlations. The default and most widely reported figure.

HTMT2

A refinement by Roemer, Schuberth and Henseler (2021) using the geometric mean. More appropriate when the indicators of a construct are not tau-equivalent (their loadings differ substantially).

HTMT inference

Instead of comparing against a fixed cutoff, bootstrap a confidence interval for each HTMT and check that its upper bound stays below 1 (or below your threshold). This turns discriminant validity into a formal statistical test.

Assessing discriminant validity step by step

Compute the full HTMT matrix for every pair of constructs in the model.
Pick the threshold per pair: 0.85 for conceptually distinct constructs, 0.90 for conceptually close ones.
Flag any pair at or above its threshold.
For flagged or borderline pairs, bootstrap the HTMT confidence interval and confirm the upper bound stays below 1.
If a pair fails, revise the measurement model before interpreting path coefficients. Nothing downstream is trustworthy until the constructs are distinct.

Common pitfalls

Reporting only the Fornell-Larcker criterion. It is now considered insufficient on its own, and reviewers expect the HTMT.
Using the 0.90 threshold for constructs that are not conceptually similar, just to obtain a pass.
Treating a single point estimate as proof. A value of 0.84 whose confidence interval reaches 0.95 is not clean.
Ignoring HTMT2 when indicator loadings vary a lot within a construct.

HTMT in OpenPLS

OpenPLS reports the full HTMT matrix in the quality-criteria panel the moment you compute a model, alongside HTMT2 and bootstrap-based confidence intervals. Pairs above your chosen threshold are highlighted automatically, so discriminant-validity issues surface without building the matrix by hand. It is free and runs in the browser.

Check discriminant validity on your own model

Build your model in the visual editor and get the HTMT matrix, HTMT2 and bootstrap intervals in a single compute run. Free, no install.

Open OpenPLS What is PLS-SEM?

Frequently asked questions

What is a good HTMT value?: Below 0.85 for conceptually distinct constructs, or below 0.90 for conceptually similar ones. Lower is better: it means the two constructs are clearly separate.
What is the difference between HTMT and the Fornell-Larcker criterion?: Both assess discriminant validity, but the HTMT is far more sensitive. Simulation studies show the Fornell-Larcker criterion often fails to detect a lack of discriminant validity that the HTMT catches, which is why the HTMT is now the recommended standard.
When should I use HTMT2 instead of HTMT?: Use HTMT2 when the indicators of a construct have substantially different loadings (they are not tau-equivalent). It relies on the geometric mean and is less biased in that case.