HTMT Above 0.90: What to Do (and What Not to Do)
Andrzej Szymkowiak · Last reviewed July 2026
TL;DR: An HTMT value at or above .90 (written 0.90 outside APA style) fails both common thresholds — .85 (strict) and .90 (liberal; Henseler et al., 2015) — so it signals a discriminant validity problem you must address, not hide. Diagnose the cause first: an item that shares wording with the other construct, or two constructs that overlap by definition. Then act with a stated reason: run the bootstrap confidence interval test, merge the constructs if theory supports it, or remove one diagnosed item — and report the model before and after. Deleting items one by one until the value passes is p-hacking, and this guide says so plainly.
Your HTMT table came back and one pair reads .93. The HTMT calculator shows a red verdict, the deadline has not moved, and the obvious temptation appears: delete an item, re-run, repeat until the number drops. This guide explains what an HTMT above .90 actually tells you, the honest steps that can fix it, and the one popular "fix" that quietly damages your study. If you are not sure what HTMT is or how it compares with the older criterion, start with HTMT vs Fornell-Larcker and come back.
What an HTMT above .90 actually means
HTMT compares how strongly items correlate across two constructs with how strongly they correlate within their own construct (Henseler, Ringle, & Sarstedt, 2015). A value near 1 means respondents answered the two sets of items almost as if they were one set. In plain words: your data cannot tell the two constructs apart.
That is a real finding about your measurement, not a software error. It has two common causes — one problem item, or two constructs that overlap by definition — and each calls for a different response, which is why diagnosis comes before any fix.
Which HTMT threshold applies: .85 or .90?
The two cutoffs are not interchangeable, and they have different sources:
| Threshold | Reading | Source |
|---|---|---|
| HTMT < .85 | strict; the default for clearly distinct constructs | Henseler, Ringle, & Sarstedt (2015), building on Clark & Watson (1995) and Kline (2016) |
| HTMT < .90 | liberal; defensible only for conceptually close constructs | Henseler, Ringle, & Sarstedt (2015) |
| HTMT ≥ .90 | fails both fixed thresholds; a discriminant validity problem | — |
The important rule: choose your threshold before you see the results, and say why. If your constructs are conceptually close — for example, cognitive and affective satisfaction — the .90 threshold is defensible, and you should write that justification into your analysis plan before data collection — the same plan where you fix your sample size. Switching from .85 to .90 after a value of .87 appears is threshold shopping. Reviewers recognize it, and it weakens every other claim in your paper.
If your value is at or above .90, no fixed threshold saves it. You need the steps below.
Step 1: diagnose before you change anything
Three checks, in order. Each takes minutes and each points to a different remedy.
Check the coding of reverse-worded items. Recoding reverse items before analysis is basic measurement hygiene, so rule it out first — it takes a minute and carries no risk. One honest caveat: most HTMT implementations, including Explonia and the semTools default, compute the ratio from absolute correlations, so HTMT itself is largely unchanged by whether a reverse item was recoded — do not expect recoding alone to move a .93. If your software uses signed correlations, an un-recoded item can distort the value, so recode and re-run to be certain.
Look for one overlapping item. Read the items of both constructs side by side. One item that shares wording or meaning with the other construct — "I would recommend this brand" appearing in both a loyalty scale and a trust scale — can pull the whole pair above the threshold. If one item is the culprit, you have a content reason to act on that item (see Step 3), which is very different from hunting for the statistically weakest one.
Ask whether the constructs overlap by definition. If the two constructs are near-synonyms in your theory — attitude and preference, satisfaction and delight — changing or removing individual items will not separate what respondents do not separate. This diagnosis points to merging, not trimming.
Step 2: run the HTMT bootstrap confidence interval test
A point estimate of .91 is not the same evidence as a point estimate of .98. HTMT is an estimate with sampling error, and there is an inferential version of the test: bootstrap a confidence interval and check whether it includes 1 (Henseler, Ringle, & Sarstedt, 2015). If the interval excludes 1 — say, [.86, .96] — the data still distinguish the constructs in the inferential sense, even though the point estimate fails both fixed cutoffs.
Two honest caveats. First, this is the most liberal defensible criterion, and a stricter version compares the upper bound of the interval with your chosen threshold instead of with 1 (Franke & Sarstedt, 2019) — a .91 estimate will not pass that version. Second, like the threshold itself, the inferential criterion should be named in your analysis plan, not adopted after the fixed cutoff fails. Used transparently, the interval tells reviewers exactly how bad (or borderline) the problem is; used as a rescue, it looks like what it is.
Step 3: the defensible fixes for discriminant validity
| Option | When it is defensible | What you must report |
|---|---|---|
| Recode a reverse item | always — it is an error correction (though absolute-value HTMT barely moves) | the coding error and the corrected results |
| Remove one item | diagnosis found a content problem (shared wording, double meaning) and the shortened scale still covers the construct | the item, the content reason, and HTMT before and after |
| Merge the constructs | theory supports treating them as one construct, or as facets of a higher-order factor | the respecified model, its fit, and the theoretical argument |
| Keep both, report the overlap | constructs are distinct in theory, the interval excludes 1, and you discuss the overlap openly | the HTMT value, the interval, and the limitation |
Removing an item is legitimate only when Step 1 gave you a reason that exists outside the numbers. State the reason, show the value before and after, and confirm the shortened scale still covers the construct's meaning — losing content validity to gain .02 of HTMT is a bad trade. Keep at least three items per reflective construct — a stand-alone reflective factor is only just-identified with three — so you cannot remove any item from a three-item scale; removal is then not an option, and merging or reporting the overlap remains.
Merging deserves more respect than it gets. If your data say two constructs are one, that can be a finding — some of the literature's "distinct" constructs fail discriminant validity checks routinely (Voorhees, Brady, Calantone, & Ramirez, 2016). Merge them (or model a higher-order factor), refit, and write it up as a result. A model that honestly reflects the data beats a model trimmed only to fit the theory.
What not to do when discriminant validity fails
Do not delete items until the number passes. Removing the "worst" item, re-running, and repeating until HTMT drops below the line is p-hacking applied to measurement. Each deletion is chosen because of the result it produces, so the final "clean" table capitalizes on chance and will not replicate. If you changed the scale, the only honest write-up shows the original model, the change, the reason, and the revised model. A results section that presents the trimmed scale as if it were the planned one misrepresents the study.
Do not switch to Fornell-Larcker because it passes. The Fornell-Larcker criterion is less sensitive and can miss exactly the problem HTMT found (Henseler, Ringle, & Sarstedt, 2015) — when the two tests disagree, you do not get to pick whichever one you prefer. The comparison, and why HTMT is now preferred, is covered in HTMT vs Fornell-Larcker.
Do not bury the original result. Deleting the failed run from the paper — or from your memory — turns a measurement finding into a reporting problem. Transparent before/after reporting is not optional politeness; it is what makes the revised model believable.
A worked example
A study measures trust (5 items) and loyalty (4 items). The Trust–Loyalty HTMT comes back at .93 against a preregistered .85 threshold.
- Diagnosis. All items are coded correctly. Reading the scales side by side shows loyalty item L4, "I would recommend this store to others," sits very close in meaning to trust item T5, "I can rely on this store's recommendations." The constructs themselves are theoretically distinct.
- Test. The bootstrap 95% interval for the pair is [.87, .98] — it excludes 1, so the failure is real but not a complete failure.
- Fix. L4 is removed for the stated content reason — and note the revision is made although the interval excludes 1: the wording overlap is a content problem, not only a statistical one. The three remaining loyalty items still cover the construct (repurchase intention, preference, resistance to switching). The re-estimated HTMT is .81.
- Write-up:
Discriminant validity was assessed with the HTMT ratio against the preregistered .85 threshold (Henseler et al., 2015). The initial Trust–Loyalty value was .93, and its bootstrap 95% confidence interval [.87, .98] excluded 1. Item inspection showed that loyalty item L4 overlapped in content with trust item T5. L4 was removed on content grounds; the revised loyalty scale retained three items covering repurchase intention, preference, and switching resistance. All HTMT values in the revised model were below .85 (maximum = .81). Both measurement models are reported in Appendix A.
Four sentences of honesty take little effort — and they answer the reviewer question before it is asked.
Common mistakes to avoid
- Choosing the .85 or .90 threshold after seeing which one your values pass.
- Deleting items repeatedly until HTMT passes, then reporting only the final scale.
- Replacing HTMT with Fornell-Larcker because Fornell-Larcker "passes" — it is the less sensitive test.
- Skipping basic data checks (item coding, obvious wording overlap) and jumping straight to model changes.
- Removing an item from a three-item scale — keep at least three items per reflective construct (a stand-alone factor with fewer is not even just-identified).
- Treating a merge as a failure; two constructs collapsing into one can be a publishable finding.
Frequently asked questions
Is HTMT above 0.90 always fatal for my study?
No, but it always requires action. If the bootstrap confidence interval excludes 1 and the constructs are theoretically distinct, you can report the overlap as a limitation; otherwise you revise the measurement — by correcting coding, removing one diagnosed item, or merging the constructs — and report both versions.
Can I just use the .90 threshold instead of .85?
Only if the two constructs are conceptually close and you chose .90 with that justification before the analysis (Henseler et al., 2015). Switching thresholds after seeing a .87 is threshold shopping, and a value at or above .90 fails both cutoffs anyway.
How many items can I remove to fix discriminant validity?
Think in reasons, not counts. Every removal needs a content justification found in diagnosis, the shortened scale must still cover the construct with at least three items, and both the original and revised models must be reported. Repeated removals chosen only because they lower HTMT are p-hacking.
What if the bootstrap confidence interval includes 1?
Then the data genuinely cannot distinguish the constructs, and changing or removing individual items is unlikely to help. Merge the constructs, model them as facets of a higher-order factor, or — if the distinction is central to your theory — collect new data with sharper measures.
Can Explonia recompute HTMT after I change the model?
Yes. Explonia computes HTMT from your CFA or PLS-SEM model — with bootstrap confidence intervals in PLS-SEM — and flags pairs above your chosen threshold, so after you recode, remove, or merge, you re-run and compare the before/after values directly. The "Article description" feature drafts the APA sentences with your actual numbers, which you edit and sign as your own.
References
Clark, L. A., & Watson, D. (1995). Constructing validity: Basic issues in objective scale development. Psychological Assessment, 7(3), 309–319. https://doi.org/10.1037/1040-3590.7.3.309
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
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. (2016). Principles and practice of structural equation modeling (4th ed.). Guilford Press.
Voorhees, C. M., Brady, M. K., Calantone, R., & Ramirez, E. (2016). Discriminant validity testing in marketing: An analysis, causes for concern, and proposed remedies. Journal of the Academy of Marketing Science, 44(1), 119–134. https://doi.org/10.1007/s11747-015-0455-4
How Explonia does this for you
Explonia computes HTMT from your CFA or PLS-SEM model — with bootstrap confidence intervals in PLS-SEM — and flags every pair above the threshold you set (.85 or .90). When you recode an item, drop one, or merge constructs, you re-run the model and see the before/after values side by side — the transparent reporting this guide asks for is simply what the platform shows. The "Article description" feature drafts APA-style sentences from your actual results; you stay the author and edit every word. Tables export to Word, R, SPSS/.sav, and LaTeX. For a quick check of a few values by hand, the free HTMT calculator needs no account, and our HTMT methods page covers the formula and reporting details.
Ready to fix your discriminant validity the honest way? Run HTMT in Explonia and compare your model before and after the change. Start free →
How to cite this article
If this article helped your research, you can cite it (APA 7):
Szymkowiak, A. (2026). HTMT Above 0.90: What to Do (and What Not to Do). Explonia. https://explonia.com/blog/htmt-above-090/
Last updated: July 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