Why field TNAs fail at the translation step

A plant manager says “initiative” means taking action without waiting for permission. A line supervisor hears “initiative” as volunteering for extra shifts. The English TNA item asks whether a person “shows initiative.” The Hindi version is grammatically correct. The scores still do not measure the same thing.

That is how a field TNA can fail before a single course is booked. The translation team checks whether sentences sound fluent. Nobody checks whether respondents in different languages are being asked to report the same behaviour. When the scores come back, the gap gets blamed on the workforce, not on the item.

The fix is not to commission a prettier translation. Treat translation as an item-level measurement control: two independent drafts, a documented adjudication, a brief respondent test, and a stop rule that prevents unresolved items entering the analysis.

The failure is meaning drift, not bad grammar

A training needs assessment is often run across plants, branches or contractor groups where the language of the source questionnaire is not the language people use at work. The risk is not just a wrong word. A translated question can change the behaviour being rated, the time period recalled, the group being compared or the meaning of the response scale.

The U.S. Census Bureau’s translation guidance describes preparation, translation, pretesting, revision and documentation as distinct steps. It also warns that a correct translation may still create conceptual problems for respondents. The European Social Survey’s translation framework similarly treats translation, review, adjudication, pretesting and documentation as connected work, rather than equating a back-translation with quality assurance.

Back-translation is useful as a diagnostic. A second translator takes the target-language draft and translates it back into the source language, without seeing the original. If an item comes back with a different concept, that is a signal to investigate. But making the back-translated sentence resemble the source is not proof that workers understand the target version as intended. The source and the back-translation can match while the target phrase carries a different workplace meaning.

For TNA items, that distinction matters because the output drives a decision: who needs practice, on what behaviour, and where. A difference in scores is useful only if the item means roughly the same thing across the groups being compared. If not, the apparent need may be a language artefact.

Start with an item brief, not a dictionary

Before translators touch the questionnaire, write a short brief for every item that carries a decision. Keep the exact source wording, but add the intended behaviour, the setting, the reference period, and what should not count as evidence.

For “shows initiative,” the brief might say: “Starts a practical improvement within their authority when a recurring work problem appears. Do not count taking on extra shifts or bypassing a required safety approval.” This forces the team to translate the construct, not just its English wrapper.

Mark terms that are likely to travel badly: initiative, ownership, accountability, collaboration, customer focus, quality mindset. These words often behave like praise labels. Replace them with observable actions where possible. Instead of “takes ownership,” ask whether the employee follows a defect through to the agreed next step and informs the next shift.

This is also the moment to lock the scale. If “often” means “most shifts” in the source version, specify that in the item notes. Do not allow one language to acquire a different frequency threshold by accident. Keep the subject, timeframe, response options and instructions aligned across versions.

The item brief does not need to become a long psychometric manual. A small table is enough: item ID, intended behaviour, reference period, boundary case, and decision use. The discipline is in writing those fields before reviewing word choices.

Run two independent drafts, then adjudicate disagreements

Ask two fluent translators to produce the target-language version independently. Each should receive the item brief and context, but they should not see the other person’s draft. If your organisation operates across dialects or sites, make the intended respondent group explicit. “Hindi” alone may not tell a translator whether the language should sound formal, conversational or familiar to a particular workforce.

Then put the drafts side by side with the source and the item brief. Review by item, not by page. For each difference, ask four practical questions:

• Does the draft preserve the behaviour, not just the general positive tone?
• Does it preserve who acts, what they do, and in what situation?
• Does it preserve the time window and response scale?
• Could a reasonable worker interpret it as a different action or expectation?

Record the choice and why. Do not settle every disagreement by choosing the more literal phrase. Sometimes the best translation needs a local example or a changed construction to preserve the intended concept. That is adaptation, not error, provided the behaviour and decision meaning remain intact.

A third person adjudicates unresolved differences. This person needs enough knowledge of the TNA purpose to distinguish a stylistic preference from a construct change. The adjudicator should not simply be the senior-most person in the room. A head-office reviewer may know the competency framework but miss how a phrase lands on the shop floor. Bring a local supervisor or field practitioner into review, while keeping the construct brief visible so local familiarity does not silently rewrite the target behaviour.

Back-translation can be added after the draft is stable, especially for high-consequence items. Keep the back-translator blind to the source. Compare the result with the intended behaviour notes as well as the original words. A difference is a prompt for review, not an automatic rejection or pass.

Test understanding with the people who will answer

Before launch, conduct short cognitive interviews in each target language and respondent context. Give the respondent the draft item and ask them to say what it means in their own words, describe a recent example, explain how they chose a response, and identify any phrase they found unnatural or ambiguous.

Do not turn this into a quiz about whether they can guess the “right” answer. You are testing the wording and the response process. Ask neutral probes: “What came to mind when you read this?” “What does ‘taking initiative’ look like on your shift?” “What period were you thinking about?” “What made you choose ‘sometimes’ rather than ‘often’?”

A compact interview log can record item ID, language and work context, respondent’s paraphrase, example recalled, response choice and reason, interviewer observation, and proposed action. Avoid collecting names unless there is a clear need. The purpose is to find recurring interpretation problems, not to create a performance record.

Use a stop rule. If the respondent’s paraphrase changes the behaviour, timeframe or decision boundary, mark the item red. If the phrase is understood but people cannot map their experience to the scale, mark it amber. Green means the meaning and response logic appear aligned in the tested context. These colours are a workflow device, not statistical validation.

Any red item goes back to the translators and adjudicator. Revise it, then test the revised wording with a fresh respondent where practical. Do not repair a bad item by changing the English after only one language has been tested; if the construct changes, update the source and all versions together. An amber item needs a specific fix, such as an example, clearer timeframe or better scale label, followed by a check that the fix did not create a new issue.

Worked example: “takes ownership” on a shift handover

Imagine an illustrative assessment used across three maintenance shifts. One item reads: “Takes ownership of unresolved issues.” The target-language draft uses a phrase that workers commonly interpret as “accepts blame.” During a cognitive interview, a technician says they chose the lowest rating because they do not want to claim responsibility for a machine fault caused by another shift.

That answer reveals more than a vocabulary problem. The item has merged follow-through with fault attribution. A practical revision could ask: “When a fault remains open at handover, does the person record the next action and confirm who will take it?” The response scale can then refer to how consistently that action happens during the agreed review period.

The translator checks natural phrasing. The reviewer checks that the action and timeframe remain intact. The adjudicator approves the target version. A second interview tests whether respondents describe the same handover behaviour and use the scale for frequency rather than blame. If the example still produces different interpretations, stop the item. Use a different observable behaviour or do not compare that item across language groups.

The numbers in this example are deliberately absent. A small wording test can identify a defect, but it cannot establish that the scale is reliable or that scores are comparable across language groups. Do not turn a few interviews into a claim of validation.

Keep the evidence attached to the item

Create a version-controlled translation record. One row per item is usually enough. Keep the source text and intended behaviour, each independent draft, the reviewer’s comparison, adjudication rationale, any back-translation note, pretest observations, final approved wording and version date.

This record solves a recurring operational problem: months later, someone asks why one plant’s wording differs from the head-office version. Without a decision trail, the answer becomes “the vendor translated it.” With a trail, you can see whether the change preserved the construct, addressed a respondent misunderstanding or was an unreviewed edit.

It also helps when the instrument changes. If the English source changes “during the last month” to “during the last quarter,” every translation and field version needs review. A translation is not a reusable asset independent of its source version. Give versions identifiers, and make the deployment checklist verify that the language packs match the approved questionnaire release.

Keep the field TNA decision separate from the translation decision. First decide whether the item works as intended in each language. Only then compare results, interpret gaps and decide whether a training intervention is justified. This sequencing fits an evidence-first training needs analysis process, where a requested course is not itself proof of a capability gap.

For a broader assessment workflow, the assessment features should support the actual method: consistent item versions, clear response instructions and an audit trail. Software can help control distribution. It cannot adjudicate whether “ownership” means follow-through or blame in a particular plant.

The limit: translation review does not validate the instrument

Two translators, adjudication and respondent testing reduce avoidable ambiguity. They do not prove that an item predicts job performance, that a scale has the same measurement properties across languages, or that two groups can be ranked fairly. Those are separate questions that require suitable data and a design matched to the use of the scores.

The literature on cross-cultural adaptation distinguishes translation and pretesting from field testing and psychometric evaluation. A small pretest is a fault-finding step. It is not a substitute for validation, nor does an identical back-translation certify equivalence. If the TNA will determine high-stakes promotion, selection or pay decisions, get qualified measurement advice before interpreting cross-language score differences.

There is also a real constraint: field teams may not have time to test every item in every language. Prioritise items that drive costly decisions, use abstract competency labels, or have already produced disagreement between translators or sites. The lower-risk items can follow the same review protocol proportionately, but do not pretend a rushed translation is evidence of equivalence.

Before the next multilingual TNA goes live, select one high-consequence item, write its behaviour brief, obtain two blind independent drafts and run a short cognitive check with the intended respondents. If the paraphrase changes the behaviour, hold the item out of the comparison until it is repaired. That one stop rule is worth more than another round of polishing the English.