Evaluate general-factor strength in bifactor and higher-order models
Source:R/nomo_hierarchical.R
nomo_hierarchical.Rdnomo_hierarchical() answers the question behind many published scales that
report both a total score and subscale scores: how much of each score
reflects a general factor, and how much reliable variance does each subscale
carry beyond it? It evaluates a fitted bifactor or higher-order CFA and
reports omega total, omega hierarchical, explained common variance (ECV),
the percentage of uncontaminated correlations (PUC), and subscale omega and
omega hierarchical subscale, each with its estimand stated.
Usage
nomo_hierarchical(
fit,
general = NULL,
obs.var = TRUE,
guidance = nomo_defaults()
)Arguments
- fit
A
nomo_cfaobject or fittedlavaanmodel containing a bifactor or higher-order measurement model. Single-group, single-level models only.- general
Optional name of the general (bifactor) or second-order (higher-order) factor. If
NULL, it is identified from the model and the call is refused if that is ambiguous.- obs.var
Logical.
TRUE(default, matchingnomo_reliability()) uses observed covariances in each omega's denominator;FALSEuses model-implied covariances, which reproduces the formulas in Rodriguez, Reise, and Haviland (2016).- guidance
Guidance settings returned by
nomo_defaults().
Value
A nomo_hierarchical object. The fields to read are:
structure:"bifactor"or"higher_order", as detected.generalandgroups: the general (or second-order) factor, and each group factor with its items.estimand:"observed_composite"with continuous indicators and"latent_response"with ordered ones (see Estimands).indices: omega total, omega hierarchical, the share of omega total that is general, ECV, and PUC, each with its estimand and interpretation.subscales: omega subscale and omega hierarchical subscale for each group factor.factors: for the general factor and each group factor, factor determinacy, its square, the minimum correlation between competing sets of factor scores, and construct replicability H (see Factor scores).loadings: each item's standardized general and group loadings, communality, and share of common variance that is general.notes: each note'stopic,severity, and full text (note), with a one-sentence version (brief); anddecision_log.
Other fields record the call, the settings used, and the fitted model.
They may change between releases and are not part of the stable interface
(see ?nomologR).
print() shows the indices for the total score and the item set, the
subscale and factor-score tables, and one sentence for each flagged note;
summary() adds each index's interpretation and every note in full.
Details
Structure. The structure is read from the fitted model, so syntax written
by hand works as well as syntax from nomo_model():
Higher-order: one second-order factor measured by first-order factors, each measured by its own items. The first-order factors' disturbances are the group sources. Their loadings are the Schmid-Leiman decomposition of the higher-order solution, which is exact for a confirmatory model.
Bifactor: one general factor measured by every item, plus group factors measured by disjoint subsets of items. Items may load on the general factor alone.
The general and group sources must be orthogonal, because variance can be attributed to a source only when sources do not covary. A model that allows them to correlate is refused with an explanation.
Estimands. Every omega is the proportion of the variance of a unit-weighted composite that a set of sources explains. With continuous indicators the composite is the observed sum score. With ordered indicators the indices describe the latent-response composite, an upper bound on the reliability of the observed ordinal sum score; the observed ordinal-scale version is not provided. They are computed on the latent responses' correlation scale, so they are the same whether the model was fitted with lavaan's delta or theta parameterization. ECV and PUC describe the item set and do not depend on the denominator.
No verdicts. The indices are reported with plain-language interpretation, but no value is treated as a pass/fail threshold. Reise (2012) notes that no benchmark value of ECV establishes when a general factor is strong enough, and that when PUC is very high, even modest ECV can yield relatively unbiased estimates from a unidimensional model.
Choosing a structure. A bifactor model will usually fit at least as well
as a correlated-factors or higher-order model of the same items, including
when it is not the model that generated the data (Reise, 2012), and a
higher-order model is a constrained version of the bifactor model (Yung,
Thissen, & McLeod, 1999). Bonifay, Lane, and Reise (2017) treat that
tendency to show superior goodness of fit as a particular concern, and note
that the superior performance "may be a symptom of overfitting", capturing
unwanted noise as well as real trends. Murray and Johnson (2013) compared
these two structures directly and found the comparison itself biased: unless
there was essentially no unmodeled complexity, their simulation favored the
bifactor model even when a higher-order model generated the data, and they
concluded that the choice "should not rely on which is better fitting".
Compare the alternatives with
nomo_compare() and choose on substantive grounds, not on fit alone.
nomo_hierarchical() does not choose.
Factor scores: two indices that answer different questions. The
factors table reports, for the general factor and each group factor,
factor determinacy and construct replicability.
Factor determinacy (Beauducel, 2011; Rodriguez, Reise, & Haviland, 2016)
is the correlation between a factor and its estimated factor score. It is
computed from the whole model-reproduced correlation matrix, so a group
factor's score estimate can use the other items to partial out the general
factor. determinacy_r2 is its square, the proportion of variance in factor
scores explained by the factor, and min_competing_r is the minimum possible
correlation between two equally valid sets of factor scores, twice the
squared determinacy minus one. A negative value there means two researchers
scoring the same data could rank people in opposite orders and both be
consistent with the model.
Construct replicability H (Hancock & Mueller, 2001) is the proportion of variance in a factor explainable by its own indicators when optimally weighted. It uses only that factor's loadings and treats the remainder of each item as uncorrelated residual.
When the data are unidimensional, H equals determinacy_r2, the squared
determinacy. The two can differ under a bifactor model, where an item's
residual with respect to one factor contains the other factors and is
correlated across items. Rodriguez et al. (2016) state this and decline to
prefer either, so both are reported.
Determinacy is always computed from the model-reproduced matrix, whatever
obs.var is set to, because that is what the formula is defined on.
Gorsuch (1983) recommended using factor score estimates only when determinacy exceeds .90, with competing score sets correlating above .70, and Hancock and Mueller (2001) proposed .70 as a standard for H. These are reported as their authors' recommendations where a value falls below them, as with fixed fit-index cutoffs elsewhere in the package, and are never applied as rules.
Improper solutions. A negative variance estimate, a residual variance or
the variance of a general or group source (in a higher-order model, a
first-order factor's disturbance), makes the solution improper. The
indices that depend on a negative source variance are NA, and a concern
note says which variance is negative.
References
Beauducel, A. (2011). Indeterminacy of factor score estimates in slightly misspecified confirmatory factor models. Journal of Modern Applied Statistical Methods, 10(2), 583-598. doi:10.22237/jmasm/1320120900
Bonifay, W., Lane, S. P., & Reise, S. P. (2017). Three concerns with applying a bifactor model as a structure of psychopathology. Clinical Psychological Science, 5(1), 184-186. doi:10.1177/2167702616657069
Gorsuch, R. L. (1983). Factor analysis (2nd ed.). Lawrence Erlbaum.
Hancock, G. R., & Mueller, R. O. (2001). Rethinking construct reliability within latent variable systems. In R. Cudeck, S. du Toit, & D. Sörbom (Eds.), Structural equation modeling: Present and future (pp. 195-216). Scientific Software International.
Holzinger, K. J., & Swineford, F. (1937). The bi-factor method. Psychometrika, 2(1), 41-54. doi:10.1007/BF02287965
Murray, A. L., & Johnson, W. (2013). The limitations of model fit in comparing the bi-factor versus higher-order models of human cognitive ability structure. Intelligence, 41(5), 407-422. doi:10.1016/j.intell.2013.06.004
Reise, S. P. (2012). The rediscovery of bifactor measurement models. Multivariate Behavioral Research, 47(5), 667-696. doi:10.1080/00273171.2012.715555
Reise, S. P., Bonifay, W. E., & Haviland, M. G. (2013). Scoring and modeling psychological measures in the presence of multidimensionality. Journal of Personality Assessment, 95(2), 129-140. doi:10.1080/00223891.2012.725437
Rodriguez, A., Reise, S. P., & Haviland, M. G. (2016). Evaluating bifactor models: Calculating and interpreting statistical indices. Psychological Methods, 21(2), 137-150. doi:10.1037/met0000045
Schmid, J., & Leiman, J. M. (1957). The development of hierarchical factor solutions. Psychometrika, 22(1), 53-61. doi:10.1007/BF02289209
Yung, Y.-F., Thissen, D., & McLeod, L. D. (1999). On the relationship between the higher-order factor model and the hierarchical factor model. Psychometrika, 64(2), 113-128. doi:10.1007/BF02294531
Examples
set.seed(2026)
n <- 500
g <- rnorm(n)
s <- matrix(rnorm(n * 3), n, 3)
dat <- as.data.frame(sapply(1:9, function(i) {
.6 * g + .45 * s[, ceiling(i / 3)] + rnorm(n, sd = .65)
}))
names(dat) <- paste0("x", 1:9)
factors <- list(
A = c("x1", "x2", "x3"),
B = c("x4", "x5", "x6"),
C = c("x7", "x8", "x9")
)
bf <- nomo_cfa(nomo_model(factors, structure = "bifactor"), data = dat)
h <- nomo_hierarchical(bf)
h
#> <nomo_hierarchical> Hierarchical model evaluation
#> Rodriguez, Reise, and Haviland (2016).
#> Bifactor model | General factor: G | Group factors: A, B, C
#> Estimand: unit-weighted observed composite
#>
#> Total score
#> Index Estimate
#> Omega total .90
#> Omega hierarchical .74
#> Omega hierarchical, relative .83
#>
#> Item set
#> Index Estimate
#> ECV .61
#> PUC .75
#>
#> Subscales
#> Subscale Items Omega subscale Omega hierarchical subscale
#> A 3 .80 .33
#> B 3 .79 .35
#> C 3 .80 .24
#>
#> Factor scores
#> Factor Role Determinacy Min competing r Replicability H
#> G general .87 .50 .83
#> A group .70 -.01 .49
#> B group .72 .02 .50
#> C group .64 -.18 .41
#>
#> Flagged
#> - Review: Factor determinacy is at or below .90 for G, A, B, C, the value
#> above which Gorsuch (1983) recommended using factor score estimates.
#> - Review: Two equally valid sets of factor scores could correlate as low as
#> G (.50), A (-.01), B (.02), C (-.18), below the .70 Gorsuch (1983)
#> suggested.
#> - Review: Construct replicability H is below .70, the standard Hancock and
#> Mueller (2001) proposed, for A, B, C.
#>
#> What these abbreviations mean
#> ECV -- Explained common variance: the share of the items' common variance
#> that the general factor explains.
#> PUC -- Percentage of uncontaminated correlations, shown as a proportion: the
#> share of item correlations that reflect the general factor alone.
#> Min competing r -- The lowest correlation two equally valid sets of factor
#> scores could have, twice the squared determinacy minus one.
#> Replicability H -- Construct replicability (Hancock & Mueller, 2001): how
#> well a factor's own indicators, optimally weighted, define it.
#>
#> These indices do not choose between a bifactor and a higher-order structure,
#> and no value is treated as a pass/fail threshold.
#>
#> See summary(x) for each index's meaning with the notes in full and
#> nomo_table(x, "factors") for every factor-score value.
nomo_table(h, "subscales")
#> # A tibble: 3 × 6
#> subscale n_items omega_subscale omega_hierarchical_s…¹ omega_hs_relative items
#> <chr> <int> <dbl> <dbl> <dbl> <chr>
#> 1 A 3 0.798 0.330 0.414 x1, …
#> 2 B 3 0.791 0.354 0.448 x4, …
#> 3 C 3 0.799 0.239 0.298 x7, …
#> # ℹ abbreviated name: ¹omega_hierarchical_subscale