A translated scale needs back-translation and conceptual checks, not a literal gloss

What it is

Moving a scale into another language is not a word-for-word mapping. The aim is for items to activate the same construct, intensity, and object in the new language. Back-translation has one translator take the source into the target, and a second translator who has not seen the source take it back; mismatches expose drift. A literal gloss keeps the shell of the words and can still move degree adverbs, the scope of a negation, politeness strength, or a referent that the culture does not have. Without back-translation and a later conceptual check, cross-language scores are not the same ruler.

Why it happens

Lexical matches are not measurement matches. Words such as “quite,” “somewhat,” “convenient,” and “intuitive” drift in intensity and object across languages. Literal transfer also imports source syntax—stacked modifiers, nominalizations—that makes items longer and harder and adds a literacy bias. Back-translation turns invisible drift into a visible source-language difference, forcing a decision to edit the translation, edit the source, or admit the item cannot travel. Even a matching back-translation can be equally wrong on both sides, so target-language cognitive interviews and, with enough sample, measurement-invariance tests remain necessary.

Studying it

Use independent forward and back translators, and compare the back-translation with the source on construct, intensity, object, and negation scope, not on whether the same word reappeared. Cognitive-interview remaining mismatches in the target language: what the item is asking, and how the options differ. With enough sample, test configural, metric, and scalar invariance; items that fail do not enter a cross-language total. Log every edit and every unresolved nonequivalence so later work does not treat the scale as already localized.

Where it stops holding

Back-translation cannot rescue a bad source item, and it does not guarantee equivalence across dialects, registers, or written versus spoken forms. Machine translation plus one back-pass looks fast and shares one bias about degree words; it does not replace independent human translators. Legally or brand-mandated wording may forbid editing the translation; drop the cross-language comparison rather than claim the scores add. Behaviors the culture does not have—certain tipping practices, certain privacy expectations—are construct problems, not translation problems, and need a changed construct or added context.

Applying it

• Before a cross-language fielding, finish independent forward translation, independent back-translation, discrepancy arbitration, and target-language cognitive interviews.
• Do not let a product manager or engineer dictionary-replace sentences and ship.
• Report each language’s distribution separately; merge only items that passed invariance tests.
• Check: any item whose back-translation still disagrees with the source on intensity or object stays out of the production scale until it is aligned or deleted.

Related

• Same group: Q3.15.1 Likert scoring assumes one continuous attitude · Q3.15.2 Option order produces primacy and recency · Q3.15.3 Mixed reverse items detect careless answering at a cognitive cost
• Adjacent: Q3.01 Questionnaire design · Q4.10 Generalizability of research claims
• Search terms: back-translation · measurement invariance · cross-language equivalence