Blog / Can Translated Questionnaires Share a Score Scale? Understanding Measurement Invariance

Can Translated Questionnaires Share a Score Scale? Understanding Measurement Invariance

Matching labels and fluent translations do not establish a common measuring scale. Learn what changes when factor loadings or item intercepts differ.

September 17, 2026 · 5 min read · By Cepingku Editorial Team

Two questionnaires can both label a scale “extraversion” without measuring it in a comparable way. Wording intensity, response conventions, and the situations readers imagine may differ across languages. Measurement invariance asks whether the measurement model works comparably across groups or occasions. The methodological review by Putnick and Bornstein explains why this question matters before interpreting differences in psychological constructs.

Invariance does not mean that groups must have identical average traits. It concerns how the measuring instrument operates. A fluent translation, a successful back-translation, and a statistical comparison of measurement models provide different kinds of evidence; none automatically substitutes for the others.

A fictional intercept difference creates an observed gap

Consider a deliberately simplified continuous-indicator model: E(y) = τ + λ × η. Here η is a latent trait, λ is the item loading, and τ is the item intercept: its expected level when η is zero. The expectation averages over random error with mean zero. This is a teaching model, not the scoring formula of a test on this website.

Suppose both language versions have a loading of 1, and both groups have a mean latent trait of 0.5. Version A has an intercept of 2; version B has an intercept of 2.4. Their expected item means are therefore 2.5 and 2.9. The observed gap is 0.4 even though the latent means are identical. In this example the gap comes entirely from the item intercept.

Fictional versionExpected item mean
A: τ = 2, λ = 1, mean η = 0.52 + 1 × 0.5 = 2.5
B: τ = 2.4, λ = 1, mean η = 0.52.4 + 1 × 0.5 = 2.9

Rescaling both results to 0–100 would not reveal or eliminate this source of difference. Nor could one person validate invariance by answering both translations once. Researchers need responses to multiple items in each group and an identifiable model to separate item parameters from trait distributions. An observed item-mean difference alone is also insufficient to diagnose translation bias: genuine trait differences and other measurement features can produce observed differences.

Three levels answer increasingly specific questions

For common multi-group factor models of continuous indicators, the lavaan tutorial distinguishes the following levels. These constraints concern the measurement model, not a requirement that all respondents give the same answers.

  1. Configural invariance uses the same factor structure across groups, including the proposed mapping of items to factors. It does not yet require equal loadings or item baselines.
  2. Metric, or weak, invariance adds equal corresponding factor loadings. This supports a common unit for the latent variable and helps comparisons of relationships; it does not alone justify latent-mean comparisons.
  3. Scalar, or strong, invariance adds equal corresponding item intercepts to equal loadings. This provides a basis for latent-mean comparisons without requiring the latent means themselves to be equal.

This sequence describes continuous-indicator models. When ordered response categories are analyzed as ordinal data, thresholds and identification conditions also matter. Do not mechanically apply the intercept explanation to every analysis of a Likert questionnaire. A decision also depends on the model, estimator, sample, and intended use; one fit index crossing a universal cutoff cannot certify a translation forever.

Match the evidence to the comparison you want to make

Imagine a project that finds support for equal loadings but not equal intercepts. Its team wants to publish a ranking of average personality levels across language groups. That ranking demands more than the reported metric evidence supports. The team could limit the mean claim, examine wording and response contexts, and evaluate revisions under a stated analysis plan. Quietly ignoring the failed constraints would not repair the comparison.

If a report uses partial invariance, look for which parameters were freed, why they were freed, and whether the remaining constraints are sufficient for the specific comparison. A label alone cannot tell you that. Two high alpha coefficients do not answer these questions either: neither reports whether corresponding loadings and intercepts agree.

What a reader can check

  • Identify the exact language versions, item sets, samples, and proposed factor structure.
  • Look for the level actually examined and the evidence supporting it, rather than a general claim that the questionnaire was translated carefully.
  • Check whether the planned conclusion concerns relationships, means, or individual decisions, and whether the evidence addresses that use.

Even a well-supported invariance model does not make a convenience sample representative or explain why groups differ. Language, age, country, and recruitment may be entangled; a group comparison does not isolate their causal effects. Without validation information, use a questionnaire’s behavioral descriptions cautiously rather than assuming two versions share a numerical ruler. This tutorial does not establish invariance for this site’s Chinese and English assessments. It explains the additional evidence that such a claim would require.

Want to see where you land?

Take the Big Five Personality Test and get your results in minutes.

Take the Big Five Personality Test →

← Back to the blog