Can a scale be compared across countries or languages?

One and the same "trust in the brand" scale, answered in France and in Germany, does not necessarily measure the same thing in the two countries: a translation can make an item easier to agree with. Comparing mean scores then compares that bias as well. The test of measurement invariance checks, step by step, that the structure, then the weights, then the starting levels of the items are the same.

On synthetic data, where item 3 is known to be shifted in Germany, forcing the starting levels to be equal adds 0.15 standard deviation on average to the estimated trust gap between the two countries — that is the stable result, measured over 100 draws. A model that frees the offending item recovers the true gap, to within 0.006 on average — a remainder not distinguishable from zero at this simulation size. On the draw detailed on this page, the two models give 0.30 and 0.18 respectively, for a true gap of 0.20.

A brand measures its customers' trust with four items, in France and in Germany (600 respondents per country). The raw results show higher trust in Germany. But item 3, translated into German, is worded in a way that is easier to agree with. The data on this page are synthetic: the true trust gap is known to be 0.2 and item 3 is known to be shifted by 0.4 in Germany.

Does the scale measure the same thing in the two countries, to the point where the trust means can be compared? If not, what share of the observed gap comes from trust and what share comes from the instrument?

The mean of the four items is higher by 0.194 scale point in Germany. But the item-by-item gap reveals the problem: 0.17, 0.09, 0.46 and 0.06. Item 3 alone carries a large part of the gap. Adding the items up mixes a real difference in trust with an artefact of translation. Means are compared only once three steps have been cleared — the same structure (called configural), the same weights (metric), the same starting levels (scalar) — set out just below ^[C09].

A model is posited in which each item reflects a latent trust, with a weight (the item's loading) and a starting level (its intercept). The constraints between countries are then tightened, step by step:

1. configural: the same structure in the two countries, which is the first rung ^[C01];
2. metric: the same weights, each item contributing as much to trust in the two countries ^[C02];
3. scalar: the same intercepts, the condition for comparing means ^[C03];
4. partial scalar: if one item fails, its intercept is freed and the others are kept. This has become common practice ^[C08]. Steenkamp and Baumgartner argued that, with at least two loadings and two intercepts left constrained, the latent mean gap remains usable ^[C12]; the source that reports this rule immediately adds that it has been "much debated", notably over the choice of the reference item ^[C16]. The standards of partial invariance vary ^[C17] and rest on no cited empirical study ^[C18].

At each step, if the fit degrades markedly, the imposed equality is rejected.

The individual answers to the items, with each respondent's country; at least three or four items per construct; samples large enough per group (here 600). The answers on this page are continuous for simplicity; real Likert scales on 5 or 7 points call for methods for ordinal data.

This section addresses the practitioner. The reader who decides can skip it and resume at "The declared calculation": the plain-terms sentence that ends it is enough.

For respondent r of country g and item j:

xᵣⱼ = νⱼ⁽ᵍ⁾ + λⱼ⁽ᵍ⁾ ηᵣ + εᵣⱼ

• xᵣⱼ: the answer of respondent r to item j. The superscript (g) denotes the country — fr or de — and marks a parameter allowed to differ from one country to the other; without a superscript, the parameter is common to both countries.
• ηᵣ: latent trust, with mean κ⁽ᵍ⁾ and variance φ⁽ᵍ⁾.
• λⱼ: the item's weight, which practitioners also call its loading; νⱼ: the intercept, the level of answer when trust is 0; εᵣⱼ: the error, with variance θⱼ.
• Metric model: λⱼ⁽ᶠʳ⁾ = λⱼ⁽ᵈᵉ⁾. Scalar model: in addition, νⱼ⁽ᶠʳ⁾ = νⱼ⁽ᵈᵉ⁾. Partial model: the equality is lifted for the intercept of the one item selected by the rule described below.
• Under scalar invariance, Germany's latent mean κ⁽ᵈᵉ⁾ measures the trust gap between countries directly ^[C04].

The scale of the latent variable must be fixed, and it is not fixed the same way at every rung:

This is why the trust gap between countries only appears from the scalar rung onward: before that, it is not identifiable.

In plain terms: if the items weigh as much and start from the same level in the two countries, a gap in answers can only come from a gap in trust.

1. Synthetic survey generated by a declared congruential recursion (seed 20260916; x ← 1103515245 · x + 12345 mod 2³¹), identical in Python and in R.
2. Maximum likelihood estimation with a mean structure, on each country's means and covariances.
3. BFGS optimiser with numerical gradients; Cholesky decomposition; χ² tail by the incomplete gamma function. Everything is written by hand, without any library, so that Python and R print exactly the same lines.
4. Fit indices. The χ² (chi-square) measures the gap between the model and the data. The CFI (comparative fit index) is the share of that gap the model closes relative to a model with no common factor at all: the closer to 1, the better. The RMSEA (root mean square error of approximation) is the mean error per degree of freedom, held to be good below 0.05. In the multigroup case the RMSEA formula is multiplied by √G, where G is the number of groups — here √2.
5. Comparisons of nested models: difference in χ² and change in CFI.
6. Standard error of the latent mean by the numerical Hessian matrix.
7. Partial model: the program computes the intercept gaps between countries under the metric model, then frees the intercept whose absolute gap is the largest ^[C13]. The rule is applied by the code, not decided by hand: on the draw detailed here, the gaps are 0.17, 0.09, 0.46 and 0.06, and item 3 is selected.
8. The same complete procedure — generation, metric model, item selection, scalar model, partial model — is then repeated on 100 independent draws, to tell what is systematic from what depends on the chance of a single sample.

The first three tables describe a single draw: they show the mechanism. The fourth, over 100 draws, shows what is systematic.

Mind the units: the gap of 0.194 seen above is in scale points, those in this table in standard deviations of latent trust. They are not to be read against one another. Brought back to scale points, the corrected gap is 0.12: the raw score overstates the trust gap by about 60%. The estimated intercept gap of item 3 is 0.33, against a true value of 0.40.

The selection rule picked item 3 in 100 draws out of 100.

The simulation error is the standard deviation over the 100 draws divided by √100. The bias of the scalar model, +0.156, far exceeds it; that of the partial scalar model, +0.006, is of the order of its own: it is not distinguishable from zero.

• A single construct measured by the four items. Do the items not also measure something else, such as satisfaction?
• Approximate normality of the answers, for maximum likelihood. Are your answers on 5 points, with floor or ceiling effects? If so, ordinal methods are needed.
• At least some of the items invariant: partial invariance assumes one knows which items are trustworthy. An assumption that cannot be fully tested: the choice of the item to free is guided by the data, which always carries a risk of ad hoc reasoning ^[C14], and must be replicated. Here the partial model leaves the four weights and three of the four intercepts constrained, so the rule quoted above ^[C12] is satisfied — which does not make it established: its own source reports it as much debated ^[C16].
• Partial model chosen after the fact: the standard error, the interval and the p of the partial model assume a model fixed in advance. They do not account for the freed item having been chosen after inspecting the data, and are therefore too optimistic.
• Independent samples between countries. Do your two samples come from separate panels, or could one respondent have answered for both countries or on two waves?
• Two groups only here. Beyond that, the same steps apply, but the usual thresholds and the √G factor change, and the choice of items to free becomes more delicate.

• The χ² test is very sensitive to small gaps of no importance in large samples ^[C06]; the change in CFI is therefore also read. A drop of more than 0.01 between nested models is the commonly used threshold ^[C05]. Here the scalar model loses 0.036 of CFI, a clear signal of non-invariance; the partial model loses nothing.
• The RMSEA tells the same story: 0.040 and 0.031 for the configural and metric models, 0.101 for the scalar one — double the usual threshold of 0.05 — and 0.028 for the partial one.
• The intercept gap of item 3 locates the source of the problem.
• Standard error of the latent gap: 0.07. The approximate 95% interval, computed on the unrounded values at about two standard errors, runs from 0.04 to 0.31. The direction of the gap is established; its size is not, since that interval is compatible both with a very weak effect and with an effect seven times larger — and it does not account for the item selection.

The structure and the weights are the same in the two countries; the intercepts are not. The analysis locates the gap on item 3: in Germany that item collects higher answers at an equal level of trust. It does not say why. Here we know the translation is the cause because we built the data that way; on real data, a non-invariant intercept names the item, never its cause — the translation still has to be reread, the item order checked, or a local use of the term examined. Forcing the intercepts to be equal attributes that gap to trust: over 100 draws, it adds 0.15 standard deviation to the estimate. Unequal intercepts severely distort the comparison of means ^[C07].

• Allowed: "Once item 3 is freed, latent trust is higher in Germany by about 0.18 standard deviation of the French distribution, with notable uncertainty (standard error 0.07) and a model chosen after inspecting the data."
• Allowed: "Forcing the intercepts to be equal, despite a non-invariant item, systematically inflates the estimated gap: by 0.15 standard deviation on average in our simulation."
• Forbidden: comparing the two countries' raw mean scores without having established scalar invariance, at least partial. It is only under scalar invariance that latent mean gaps account for the gaps in item means ^{[C09, C10, C11]}.
• Forbidden: concluding from the analysis that the translation is at fault; it names an item, not a cause.
• Forbidden: presenting these figures as a market result; the data are synthetic.

Establish the cause of the gap on item 3 first — translation, wording, local use — then correct or replace the item in the German version before the next wave. In the meantime, publish cross-country comparisons from the partial model, never from the raw scores. The decision to revise the translation stays human.

• Use it: as soon as scale scores are compared across countries, languages, segments or survey waves. The demonstration on this page covers two groups, four items and continuous answers.
• Do not use it alone: with two items per construct, where the model is identifiable only at the price of extra constraints — the very ones the test is meant to put to the proof — or with very small samples.
• Alternatives: methods for ordinal data. Partial invariance, which has become common practice ^[C08] but whose standards vary ^[C17]. Approximate invariance, of Bayesian inspiration, which constrains the parameters more tightly than partial invariance instead of freeing an item completely ^[C15].

• Python 3.14.5, standard library only: python msc-p042-reference.py
• R 4.6.1, base only: Rscript msc-p042-reference.R
• The two programs print exactly the same lines. The sha256 digest of the output, after line-ending normalisation: a0b6fce517d9c53684494d9aa6ccaffc9f0614d1299c464054f04dda98393972. Declared seed 20260916. No third-party library: the software does not validate the method, it reproduces it.
• External check with the R package lavaan: planned, not yet carried out to date.

The table of the four models and of the comparisons, the failing item and its intercept gap, the latent gap between countries with its standard error under the model retained, and the mean bias the scalar model introduces over repeated draws.

• Putnick, D. L. & Bornstein, M. H. (2016). Measurement Invariance Conventions and Reporting. Developmental Review 41, 71–90. ↗ — open-access NIH author manuscript, full text verified ^{[C01–C10, C17, C18]}.
• Van de Schoot, R., Schmidt, P., De Beuckelaer, A., Lek, K. & Zondervan-Zwijnenburg, M. (2013). Facing off with Scylla and Charybdis: a comparison of scalar, partial, and the novel possibility of approximate measurement invariance. Frontiers in Psychology 4, 770. ↗ — open access under a CC BY licence, the statement read in the article itself, full text verified ^[C11–C16].
• Evidence level: demonstration on synthetic data, one detailed draw and 100 repeated draws, reproducible in Python and in R. This is not a market result.