Is this scale validated in Spanish? How to judge the evidence

“Is this scale validated in Spanish?” is the hallway question no yes-or-no answer handles well. The official instruction manual for the PHQ-9 and GAD-7 concedes it: most translations are linguistically valid, but few have been psychometrically validated against an independent structured psychiatric interview.

“Validated” is not a switch. It is a ladder, and every instrument sits on a different rung for every population. The useful question is validated for what, against what, and in whom. This guide teaches you to answer it with an abstract in front of you: a scale without local validation is not useless, it just narrows your inferences.

Translation is not validation

The canonical cross-cultural adaptation procedure — Beaton, Bombardier, Guillemin and Ferraz, in the Institute for Work & Health methodological document — has six stages, from initial translation to submission of documentation, including a back translation and an expert committee that must reach four equivalences: semantic, idiomatic, experiential, conceptual. The last two break when an item written in Ohio is administered in Oaxaca.

You need not run that process; you need to recognise when a paper only did it. The IWH document warns that adaptation seeks content equivalence, which suggests internal consistency, validity and reliability might be retained — but not necessarily; and that further psychometric testing is highly recommended, though not required for a translated version to be approved. A version can be official without anyone having measured anything with it.

The surprise is the back translation, which we take as the proof of quality. The international test translation and adaptation guidelines, in their Spanish edition (Muñiz, Elosua & Hambleton, 2013), put it bluntly: in a bad translation, the equivalence between the original and the back-translated version can be very high. If the abstract only boasts a back translation, you have rung one.

The ladder of evidence

The rungs run weakest to strongest, naming each measurement property the way the COSMIN checklist names it.

#	Rung	What the abstract says	Lets you…
1	Translation	“cultural adaptation”, “back translation”	administer it in that language
2	Internal consistency	“Cronbach’s alpha”, “omega”	assume a coherent total in that sample
3	Factor structure	“CFA”, “one-factor model”, “CFI/RMSEA”	read the total as one thing
4	Measurement invariance	“scalar invariance”, “DIF”	compare groups
5	Convergent and discriminant	“construct validity”	argue it measures the construct, not its neighbour
6	Diagnostic accuracy	“sensitivity against a structured interview”	use a cut point
7	Responsiveness	“detects change over time”	read a drop as improvement

One precision reorders the reading: COSMIN holds that the internal-consistency statistic is interpretable only once the items form a unidimensional scale. Rung 3 comes before rung 2, even though nearly every paper reports alpha first.

The rung almost nobody reports

Measurement invariance is what licenses group comparison. The technical statement is blunt: if the factor structures are not comparable across groups, comparing means and correlations is invalid, because there is no interpretation of a mean difference without a common structure behind it. It is tested in levels: configural requires the same items to load on the same latent factor; metric holds the factor loadings equal; scalar additionally holds the intercepts equal, and it is the one you need to compare means. Few cross-cultural studies test that level.

The Spanish-language guidelines turn this into an operating instruction: compare scores across populations only at the level of invariance established for the score scale used in those comparisons. Without that evidence, contrasting your male clients against your female clients, or your sample against a foreign norm, is not risky — it is uninterpretable.

Why cut points travel badly

The most expensive error in the field is treating a cut point as a constant of the instrument. A cutoff is an optimisation performed on one sample, and primary authors tend to choose which to report: in Levis and colleagues’ analysis of selective cutoff reporting (2017), when the instrument was poorly sensitive at the standard cutoff they reported lower cutoffs that yielded optimal results, and when it was highly sensitive, higher ones.

Second: sensitivity and specificity are not fixed attributes of the instrument. They depend on the cutoff and on the sample they were estimated in, which is why the validation sample must match the people you are screening. Third, and most useful: prevalence does not move sensitivity or specificity, but it does move positive predictive value. In the largest individual-participant meta-analysis of the PHQ-9 (58 studies, 17,357 participants), the same cutoff of 10 yields very different predictive values.

Prevalence of major depression	PPV at PHQ-9 ≥ 10, against a semistructured interview
5 %	24 %
14 % (the rate in those studies)	49 %
25 %	66 %

The authors spell out the consequence: 51 % of all positive screens would be false positives. That is not a defect of the PHQ-9: it is base-rate arithmetic. A fourth factor almost nobody reads: what it was measured against. In the same meta-analysis, sensitivity with semistructured interviews ran 5 to 22 percentage points higher than with fully structured interviews. “90 % sensitivity” means nothing until you know against what.

The Latin American case closes it: in 243 primary-care adults in Bucaramanga, Colombia, with the MINI as criterion, the optimal PHQ-9 cutoff was ≥ 7, not ≥ 10 (sensitivity 90.38 %, specificity 81.68 %). The warning we will not soften: that does not authorise you to move your cut point. A cutoff is recalibrated by a study, not by a hunch. When you suspect it does not fit your population, lean on the clinical interview and on within-person change; the PHQ-9 guide develops that point.

What alpha does not tell you

“What Cronbach’s alpha is acceptable” is nearly always answered with an unsourced 0.70. Three corrections: a high alpha does not guarantee internal consistency or unidimensionality — it is well documented that a multidimensional test does not necessarily have a lower alpha than a unidimensional one; it rises with the number of items, so lengthening a test inflates it and a very high value may flag redundant items, with a recommended ceiling around 0.90; and it belongs to the sample, not the instrument, which is why published alphas are a poor thing to lean on. An alpha of 0.92 in Zaragoza is not evidence about your clients in Monterrey.

Where to look for the evidence yourself

Start with PubMed and PMC, but do not stop there. English-language databases index Spanish-language literature poorly, which is why much Mexican and Latin American evidence “does not exist” until you look in the right place. The verified doors:

SciELO México: a virtual library of Mexican journals across all fields, run since 2004 by UNAM’s Dirección General de Bibliotecas y Servicios Digitales de Información, inside the 17-country SciELO network.
Redalyc, run by UAEM: 1,825 journals and 770,337 documents by its own counters, 1,574 of them Spanish-language, with Psychology and Health collections.
LILACS, through the BVS regional portal, coordinated by BIREME/PAHO/WHO: over a million records from more than 2,600 health-science journals across Latin America and the Caribbean, including theses no Anglophone database carries.
Salud Mental (Instituto Nacional de Psiquiatría Ramón de la Fuente Muñiz, bimonthly since 1977) and Salud Pública de México (INSP, bimonthly, bilingual), both mirrored on SciELO México, the second one’s mirror last updated in 2023. One live trap: at the time of writing, the historical domain revistasaludmental.mx no longer belongs to the journal and serves gambling advertising; the current site is revistasaludmental.gob.mx.

With the abstract in front of you, six questions:

Is this a translation or a validation?
In what population, and with what sample size?
Against what reference standard: a structured interview, or another questionnaire?
Does it report invariance, or only alpha?
Matched case-control — artificial prevalence — or a consecutive sample?
Does that sample resemble my clients?

The third is answered in the methods, not the results.

The PHQ-9 in Mexico, rung by rung

Rung 1. An official Spanish-for-Mexico version exists on the official PHQ site, in the public domain; where instruments can legally be obtained is covered in the guide to psychological test licensing.
Rungs 2 and 3. Mexican Teachers’ Cohort (55,555 women): one-factor structure, loadings 0.71 to 0.90, alpha 0.89.
Rung 3, with an asterisk. In 357 Mexican oncology patients the PHQ-9 came out two-factor (alpha 0.837) while the GAD-7 held one: structure is not guaranteed inside one country.
Rung 4. In 768 adolescents in San Luis Potosí, invariance by sex up to the strong (scalar) level was demonstrated: there, a male-female comparison is defensible.
Rung 6. We could not locate a Mexican diagnostic-accuracy study against an independent structured interview. We are not saying none exists — grey literature and theses are not fully indexed — only that we did not find one. A recent study in 272 incarcerated men did plot ROC curves, but against the ICD-11 primary-care scales: another self-report, not an interview. “ROC curve” is not diagnostic accuracy.

In Mexico, then, the PHQ-9 is solid for severity and within-person monitoring, and credible for group comparison in some populations; its cut points are borrowed.

The GAD-7 gives a parallel, thinner picture: the most-cited Mexican evidence is reliability and structure in 163 patients in oncology genetic counselling, with no diagnostic accuracy, and the classic Spanish “validation” used matched case-control sampling with 50 % prevalence by construction — the design that most inflates positive predictive value.

What you can do with a scale that has no local validation

Missing local validation narrows your inferences; it does not forbid measuring. With an instrument that reaches only rungs 1 to 3 in your population, you can:

Monitor the same client against themselves. Same instrument, same version, same conditions: the comparison is within-person, and nothing in the invariance or accuracy literature invalidates it. It is the highest-yield use, and measuring therapy progress covers cadence and thresholds.
Grade severity as a descriptive data point and structure the clinical conversation.

What you cannot do: diagnose from the score, compare groups without invariance evidence, or move the cutoff on your own authority.

One last reminder: if you monitor with the PHQ-9, sooner or later an item 9 will be endorsed. Any risk disclosure calls for an immediate clinical risk assessment under your own protocol and training; this article will not hand you that protocol, and distrust anyone who does in a blog post. Make sure the client leaves knowing the crisis line serving their location and local emergency services.

From data point to chart

All of it is lost if the note keeps only a number. Record the instrument, the version and the language, the date and the previous score; in a SOAP note that lives in the O — and verify what the rules where you practice require of assessment records.

“PHQ-9 (official Spanish-for-Mexico version): 14 (session 3: 19). Screening and monitoring data point, not a diagnosis; cutoffs derived from non-Mexican samples.”

Writing “PHQ-9: 14, moderate depression” and nothing else is one of the clinical documentation mistakes you cannot defend later: it turns a screen into a label.

Sustaining that record across thirty clients is the real work, and there the tool helps: gesell.ai includes built-in validated scales with automatic scoring — PHQ-9 and GAD-7 among them — and the score lands in the client’s structured chart, next to the previous ones and ready for your note. Always as decision support you evaluate: the scale contributes the data point, with the limits described here; the reading stays yours.