How to Report Measurement Invariance in APA 7

You ran the multigroup CFA, the output handed back four blocks of indices, a significant chi-square difference test at the third step and a ΔCFI of −.029 you cannot tell is large or small. Underneath all of it sits the question that actually matters: whether you are allowed to write "women scored higher on depressive symptoms" in your Results. Most manuscripts settle this with one sentence ("measurement invariance was tested and supported") and no table, which is exactly what a psychometrics reviewer will send back. Here is what each rung constrains, what numbers each line carries, how to decide when Δχ² and ΔCFI disagree, and what to do when the scalar model will not hold. If what you need is the lavaan code to run the sequence, it is in the guide to measurement invariance in R.

What each rung constrains

The four levels are the same confirmatory factor model estimated simultaneously in both groups, with progressively more parameters forced to be equal. The configural model equates nothing: it only requires the same number of factors and the same pattern of loadings, and everything else (loadings, intercepts, residuals) is estimated freely within each group. The metric or weak model equates the factor loadings. The scalar or strong model keeps the loadings equal and adds the item intercepts. The strict model adds the residual variances. The weak/strong/strict terminology comes from Widaman and Reise (1997) and the formal framework from Meredith (1993); both naming schemes coexist in the literature and either is fine as long as you state which parameters you fixed.

The degrees of freedom will tell you whether your output did what you think it did. With nine items, two groups and a mean structure there are 2 × (45 variances and covariances + 9 means) = 108 pieces of information. The configural model estimates 8 loadings (the first is fixed to 1), 9 intercepts, 9 residuals and 1 factor variance per group, so 54 parameters: df = 54. Equating the loadings saves 8 parameters and the metric model rises to df = 62. Here is the trap almost nobody checks: equating the nine intercepts saves 9 parameters, but doing so frees the latent mean of the second group, which had been fixed at 0, so the net saving is 8 and the scalar model rises to df = 70, not 71. The strict model equates the nine residuals and reaches df = 79. If your Δdf at the scalar step is 9 rather than 8, your model is estimating no latent mean difference at all and everything downstream is void.

The configural model is not a level you "pass": it is the baseline the others are measured against. If it does not fit, the ladder stops there, and the problem is not invariance — it is that the factor structure you are assuming does not describe one of the two groups. That gets fixed by going back to the single-group CFA in each group, not by freeing parameters in a multigroup model that was born broken. The configural model is judged by ordinary CFA standards (Hu & Bentler, 1999, as a convention: CFI ≥ .95, RMSEA ≤ .06, SRMR ≤ .08), not by change criteria.

Each rung buys you something different, and it pays to know which one you need before you start. Metric invariance licenses comparisons of correlations, regression coefficients and structural paths across groups: one unit of the latent variable means the same thing in both. Scalar invariance licenses comparisons of means, which is what nearly every paper that gets this far actually wants. Strict invariance licenses treating observed sum scores as measuring the same thing with the same precision, which is what you silently assume when you feed a scale total into a t test or an ANOVA. If your final analysis runs on observed scores rather than latent variables, the level you need is strict, not scalar.

The numbers each line carries

Every model carries its absolute fit and, from the second onwards, its change from the previous one. The full line is χ² with its degrees of freedom and N, CFI, RMSEA with its 90% confidence interval and SRMR; and in the comparison columns, Δχ² with its Δdf and p, ΔCFI, ΔRMSEA and ΔSRMR. In the example running through this article (a nine-item depressive symptom scale, N = 842: n = 431 women and n = 411 men) the configural model gives χ²(54, N = 842) = 118.42, p < .001, CFI = .983, RMSEA = .053, 90% CI [.041, .065], SRMR = .038.

Italics are decided letter by letter, as always. The χ is Greek and stays roman; so does the Δ, so ΔCFI and Δχ² are roman throughout. N, n, p and df are Latin and go in italics. CFI, RMSEA, SRMR and TLI are acronyms, not symbols: roman, never italic. All three indices lose the leading zero, because by convention they are read as proportions: CFI = .983, RMSEA = .053, SRMR = .038, and the same in the changes, ΔCFI = −.001. Chi-square is the APA 7 exception that puts N inside the parentheses, and the difference test does not repeat that N because it is already on the model line: what changes there is Δdf. The full run-through of italics and decimals lives in its own entry.

Five models with nine columns will not fit in a paragraph: this is a table, one row per model with the change columns on the right. The note underneath carries what the table cannot say: the estimator, how missing data were handled, the software and its version, and which parameters were constrained in each row. The APA 7 table rules apply without exception: no vertical rules, the configural model is the first row, and its change cells are left empty rather than filled with zeros.

Three things belong in the text rather than the table. The first is each group's sample size, because the power of the Δχ² depends on it. The second is where you put the marker item and in which group you fixed the latent mean at 0, because without that the reader cannot interpret the sign of the mean difference you report at the end. The third is the input matrix: with ordinal items of few categories what you are estimating are thresholds and polychoric correlations, and that changes the whole ladder.

ΔCFI and ΔRMSEA against the chi-square difference test

The Δχ² is the exact test of the constraint: because the models are nested, the difference between their chi-squares is distributed as chi-square on Δdf degrees of freedom, and it tests the hypothesis that the equated parameters really are equal. Its problem is the model chi-square's problem: power grows with sample size, so at large N any trivial departure crosses .05. In the example, the metric step gives Δχ²(8) = 11.34, p = .183, and ΔCFI = −.001 and ΔRMSEA = −.002 say the same thing; all three agree because there is nothing to detect. The disagreement shows up at the last rung, where the strict model gives Δχ²(9) = 17.02, p = .048 while ΔCFI = −.002, ΔRMSEA = .000 and ΔSRMR = .006 barely move.

That is where the change criteria come from, and they should be cited as what they are: simulation conventions, not laws of nature. Cheung and Rensvold (2002) put twenty indices through a simulation study looking for those that did not depend on sample size or model complexity, and proposed that a ΔCFI no larger than .010 in absolute value is consistent with invariance. Chen (2007) refined this by level and by sample size: with N > 300, loadings require ΔCFI ≥ −.010 accompanied by ΔRMSEA ≤ .015 or ΔSRMR ≤ .030, while intercepts and residuals require the same ΔCFI and the same ΔRMSEA ≤ .015 but ΔSRMR ≤ .010, because SRMR is more sensitive to loading non-invariance than to intercept non-invariance. Meade et al. (2008) argued that −.010 is too permissive and proposed −.002. The dominant practice in psychology is Cheung and Rensvold backed by Chen, and what you cannot do is pick the criterion after seeing the result.

When the criteria contradict each other, say so and justify the call. The strict-step sentence gets written out in full: "Constraining the residual variances produced a significant difference test, Δχ²(9) = 17.02, p = .048, but the change indices stayed within conventional limits, ΔCFI = −.002, ΔRMSEA = .000, ΔSRMR = .006; given the sensitivity of Δχ² to sample size (N = 842), strict invariance was retained." That is a defensible decision with its reasoning in front of it. Reporting only the ΔCFI and never mentioning that the Δχ² came out significant is something else.

One estimation detail sinks a lot of tables: if you estimated with MLR or WLSMV you cannot subtract the chi-squares. Robust chi-squares are scaled by a correction factor that differs between the two models, and their difference is not distributed as chi-square. You need the scaled difference test of Satorra and Bentler (2001), with the 2010 correction that avoids negative scaling factors; lavTestLRT() handles it in lavaan and DIFFTEST in Mplus. With ordinal items the ladder itself changes shape: thresholds enter and are not identified the same way, so what is a single "configural → metric" step with continuous data becomes a sequence that includes threshold invariance (Wu & Estabrook, 2016).

Why you cannot compare means without scalar invariance

The reason is one line of algebra. Each item response is modeled as x_i = τ_i + λ_iη + ε_i: the item intercept, plus the loading times the person's standing on the construct, plus error. Taking means within a group, the error drops out and you are left with E(x_i) = τ_i + λ_i × E(η). An item's observed mean, and therefore the mean of the total score, is a sum of two things: where the group sits on the construct and where the item starts. If the τ values differ across groups, an observed mean difference mixes the two and nothing in the observed data can separate them. Equating the intercepts is exactly what the scalar step does, and it is what turns an observed mean difference into an interpretable one.

Concretely, with the example in front of us: the crying item had an intercept of 1.43 in women and 1.22 in men. That means that at the same level of depressive symptoms, a woman scores 0.21 higher on that item than a man. Those 0.21 points are added to every woman's total, whatever her actual level, and the t test hands them back as though they were depression. It is not a flaw in the scale or in the participants: the item is measuring the construct plus something that covaries with group membership.

Metric invariance does not cover this, and assuming it does is the most common conceptual error in the whole topic. Equal loadings guarantee that one unit of the construct is worth the same in both groups, so with metric invariance you can already compare correlations, regression slopes and full structural models. Means do not depend only on the unit: they depend on the origin, and the origin is set by the intercepts. Put differently, every between-groups t test, every one-way between-subjects ANOVA, every ANCOVA and every MANOVA run on scale scores is silently assuming scalar invariance at the very least; testing it is the only thing separating a comparison from a coincidence.

The contrast is useful for sizing the problem. In these data, the observed comparison of totals gave women M = 8.41 (SD = 5.61) against men M = 6.59 (SD = 5.38), t(840) = 4.80, p < .001, d = 0.33. The latent mean comparison, once partial invariance was established, gave a difference of 0.34, SE = 0.09, z = 3.78, p < .001, which standardized by the factor's standard deviation in the reference group (1.10) works out at d = 0.31. Here the contamination was small and the substantive conclusion does not change. What you cannot know is that beforehand: an intercept's bias can run with the effect or against it, and there are published cross-cultural comparisons where the observed difference vanishes once you move to latent means.

When scalar fails: partial invariance

In the example, the scalar step failed with all three criteria agreeing: χ²(70, N = 842) = 246.30, ΔCFI = −.029, ΔRMSEA = .026, ΔSRMR = .036 and Δχ²(8) = 116.54, p < .001. When the misfit is that size there is no ambiguity, and the answer is neither to abandon the comparison nor to force it: it is to locate it. Scalar non-invariance is almost never spread evenly across items; typically one or two carry all of it, and the modification indices with their expected parameter change tell you which.

Free them one at a time, largest first, and re-examine the fit after each. Here the crying item's intercept was enough: the partial scalar model gave χ²(69, N = 842) = 141.88, p < .001, CFI = .981, RMSEA = .050, 90% CI [.038, .061], SRMR = .046, and against the metric model, Δχ²(7) = 12.12, p = .097, ΔCFI = −.001, ΔRMSEA = −.001. Notice that Δdf is now 7 rather than 8, because one intercept fewer is constrained and the latent mean is still free; if your output does not drop that degree of freedom, you did not free what you think you freed.

How many can you free before this stops meaning anything? Partial invariance was introduced by Byrne et al. (1989), and the most cited operational rule is Steenkamp and Baumgartner's (1998): besides the marker item, at least one more item per factor must have invariant loading and intercept for the latent mean comparison to be identified and interpretable. That is the mathematical minimum, not a reasonable standard. With nine items, freeing one is defensible and gets reported without apology; freeing five means half the scale behaves differently in the two groups, and the honest move is to write that the instrument does not support the comparison rather than tune the model until the ΔCFI clears.

A modification-index search is exploratory and gets reported as such: what you freed, in what order, on what stopping rule, and that the data made the decision rather than a prior hypothesis. The non-invariant item also deserves a sentence of interpretation, because it is a result in its own right: that crying is endorsed differently at equal levels of depression is well-trodden ground in the gender-differences literature, not a nuisance in the model. Where the sample allows it, what turns this into a genuine finding is replicating the non-invariance in a second sample; Putnick and Bornstein (2016) reviewed how this sequence is reported in practice, and their main complaint was precisely that problem items get freed and never discussed.

The write-up, exactly as it goes into the paper

Two sentences do the job in the Method: "Measurement invariance across men and women was assessed with multigroup confirmatory factor analysis using maximum likelihood estimation, testing the configural, metric, scalar and strict models in sequence. The first item's loading was fixed to 1 in both groups and the latent mean was fixed to 0 in the male group, which served as the reference."

In the Results, the full paragraph: "The configural model fitted adequately, χ²(54, N = 842) = 118.42, p < .001, CFI = .983, RMSEA = .053, 90% CI [.041, .065], SRMR = .038. Constraining the loadings did not degrade fit, χ²(62, N = 842) = 129.76, Δχ²(8) = 11.34, p = .183, ΔCFI = −.001, ΔRMSEA = −.002, so metric invariance was retained. The scalar model did degrade it, χ²(70, N = 842) = 246.30, Δχ²(8) = 116.54, p < .001, ΔCFI = −.029, ΔRMSEA = .026, exceeding the −.010 criterion proposed by Cheung and Rensvold (2002). Modification indices pointed to the intercept of item 4 (crying), estimated at 1.43 in women and 1.22 in men; with that single parameter freed, the partial scalar model fitted equivalently to the metric model, χ²(69, N = 842) = 141.88, Δχ²(7) = 12.12, p = .097, ΔCFI = −.001. Subsequently constraining the residual variances produced a significant difference test, Δχ²(9) = 17.02, p = .048, but the change indices stayed within conventional limits, ΔCFI = −.002, ΔRMSEA = .000, so partial strict invariance was retained (Chen, 2007). On that basis, the latent mean was higher in women, difference = 0.34, SE = 0.09, z = 3.78, p < .001, d = 0.31."

That is a single paragraph, and it contains everything a reviewer will ask for: the five models, their degrees of freedom, both decision criteria at every step, the freed parameter with its value in each group, and the mean comparison placed after you earned the right to make it. If your Results section says "measurement invariance was supported (ΔCFI < .01)" and nothing else, you are not reporting an analysis, you are announcing that you ran one.

Running the ladder by hand takes a while, and most of the errors are not in the code — they are in the degrees-of-freedom bookkeeping and in the criterion used to decide. The measurement invariance checker estimates the four models, returns the Δχ², ΔCFI and ΔRMSEA for each step, and lays out the table with the columns in APA 7 order. The judgment about whether to free an intercept or to concede that the scale does not support the comparison stays yours.

Frequently asked questions

How do you report measurement invariance in APA 7?

With a table carrying one row per model (configural, metric, scalar, strict) showing χ² with its degrees of freedom and N, CFI, RMSEA with its 90% interval and SRMR, plus the change columns Δχ² with Δdf and p, ΔCFI and ΔRMSEA. The text carries each step's decision with the criterion that made it: "constraining the loadings did not degrade fit, Δχ²(8) = 11.34, p = .183, ΔCFI = −.001, so metric invariance was retained". χ and Δ are Greek and stay roman; N and p are Latin and go in italics; CFI, RMSEA and SRMR are acronyms and drop the leading zero.

What ΔCFI value counts as acceptable?

The most widely used convention is Cheung and Rensvold's (2002): a ΔCFI of −.010 or smaller in absolute value is consistent with invariance. Chen (2007) refined it by level and sample size, and with N > 300 asks for that same ΔCFI accompanied by ΔRMSEA ≤ .015 or ΔSRMR ≤ .030 for loadings, and ΔRMSEA ≤ .015 or ΔSRMR ≤ .010 for intercepts and residuals. Meade et al. (2008) proposed −.002 on the grounds that the earlier value is too permissive. These are simulation conventions, not laws: choose your criterion before you look at the output and state it in the Method.

Can I compare group means without scalar invariance?

Not with observed scores. An item's mean is τ + λ × latent mean, so if intercepts differ across groups the observed mean difference confounds the construct with where the item starts, and nothing in the data separates them. With partial scalar invariance you can: it is enough for the marker item and at least one other item per factor to have invariant loadings and intercepts (Steenkamp & Baumgartner, 1998), and the comparison is then made between latent means, not between sum scores.

What is the difference between metric and scalar invariance?

Metric invariance equates the factor loadings and guarantees that one unit of the construct is worth the same in both groups: with it you can compare correlations, regression coefficients and structural models. Scalar invariance keeps the loadings equal and adds the intercepts, which set the origin of the scale: with it, and only with it, you can compare means. In degrees of freedom, the metric step gains as many degrees of freedom as you have free loadings and the scalar step one fewer than you have intercepts, because equating them frees the second group's latent mean.

Why is the chi-square difference test almost always significant?

Because its power grows with sample size and it detects departures that change no conclusion. At N = 842, a step where ΔCFI = −.002 and ΔRMSEA = .000 can still give Δχ²(9) = 17.02, p = .048. That is why the change criteria displaced the exact test. Report both, and when they disagree explain which one decided and why. If you estimated with MLR or WLSMV, you also cannot subtract the chi-squares directly: you need the scaled difference test of Satorra and Bentler (2001, 2010).

Do I need to reach strict invariance?

It depends on what you analyze next. To compare latent means inside a structural equation model, scalar invariance is enough. If your final analysis runs on the observed sum score (a t test, an ANOVA, a correlation with the total), you are assuming the scale measures with the same precision in both groups, and that is exactly strict invariance. It is the rung most often skipped and the one most papers actually need.

With the four models in one table, degrees of freedom that add up, both decision criteria at every step and the mean comparison placed only after you earned it, your measurement invariance is reported the way APA 7 asks. When the scalar model fails on more than one item, when there are ten groups rather than two, or when the items are ordinal and you have to decide what to do about thresholds, we can work through it together in my statistical consulting.