Migrating to the efa

EFAtools exposes its functionality through a set of lowercase efa_* functions: efa_screen(), efa_retain(), efa_fit(), and so on. Earlier releases used uppercase names instead (EFA(), N_FACTORS(), OMEGA(), …). The uppercase names have not been removed — they still work — but the efa_* functions are now the recommended interface. This vignette explains the change, gives the full old-to-new mapping, and walks through the migration with the largest change to the argument list: EFA() to efa_fit().

library(EFAtools)

Why the efa_* Interface

The efa_* names are the interface we recommend for new code. They read consistently, share a common prefix that groups them in tab-completion and the documentation index, and they are where the package’s development now happens.

The uppercase names are superseded, not deprecated. In practice that means:

They keep working with their original arguments, and calling them emits no warning or message. Existing scripts continue to run without changes.
Their argument lists are frozen. New capabilities are added only to the efa_* functions, so an uppercase name will never grow a new argument.

So there is no urgency to rewrite working code. Migrate a script when you want the new functions’ additional features (or simply for consistency); until then the old calls remain valid.

The Old-to-New Mapping

Every uppercase function has a lowercase counterpart. For most of them the migration is a pure rename: the arguments are identical and only the name changes.

Old name	Recommended name
`EFA()`	`efa_fit()`
`N_FACTORS()`	`efa_retain()`
`EFA_AVERAGE()`	`efa_average()`
`EFA_POOLED()`	`efa_mi()`
`COMPARE()`	`efa_compare()`
`SL()`	`efa_schmid_leiman()`
`OMEGA()`	`efa_reliability()`
`FACTOR_SCORES()`	`efa_scores()`
`PROCRUSTES()`	`efa_procrustes()`
`BARTLETT()`	`efa_bartlett()`
`KMO()`	`efa_kmo()`
`PARALLEL()`	`efa_parallel()`
`EKC()`	`efa_ekc()`
`KGC()`	`efa_kgc()`
`HULL()`	`efa_hull()`
`SCREE()`	`efa_scree()`
`MAP()`	`efa_map()`
`NEST()`	`efa_nest()`
`SMT()`	`efa_smt()`
`CD()`	`efa_cd()`

efa_reliability() and efa_scores() are broader than the OMEGA() and FACTOR_SCORES() they replace, but they cover the same use cases and are the recommended way to obtain those quantities going forward.

efa_retain() takes estimator and alpha_CD where N_FACTORS() takes method and alpha. Its parallel analysis, Kaiser-Guttman criterion, and scree plot use PCA eigenvalues, and its Hull method uses the CAF and the 95th percentile in its parallel analysis by default. Thus, its results can differ from those of N_FACTORS(), which keeps its own defaults. In the same way, efa_hull() uses ML, the CAF, and PCA eigenvalues with the 95th percentile by default, but HULL() keeps PAF, all three indices, and SMC eigenvalues with the mean rule.

Alongside the renamed functions, the package ships tools that have no uppercase predecessor — they are new, and available only under the efa_* interface (and the two control constructors):

New function	Purpose
`efa_screen()`	Data screening and factorability diagnostics in one report
`efa_group()`	Multigroup EFA with factor congruence
`efa_simulate()`	Simulate data from a factor model
`efa_power()`	Power analysis for EFA
`estimate_control()`	Bundle the estimation tuning knobs (see below)
`rotate_control()`	Bundle the rotation tuning knobs (see below)

For a plain rename, the call is unchanged apart from the name. For example, the Kaiser-Meyer-Olkin criterion:

cor_mat <- test_models$baseline$cormat

# Recommended name -- exactly the same arguments as KMO():
efa_kmo(cor_mat)
#> 
#> ── Kaiser-Meyer-Olkin criterion (KMO) ──────────────────────────────────────────
#> 
#> ✔ The overall KMO value for your data is marvellous.
#> These data are probably suitable for factor analysis.
#> 
#> Overall: 0.916
#> 
#> For each variable:
#>    V1    V2    V3    V4    V5    V6    V7    V8    V9   V10   V11   V12   V13 
#> 0.900 0.914 0.924 0.932 0.923 0.891 0.928 0.919 0.916 0.892 0.928 0.908 0.922 
#>   V14   V15   V16   V17   V18 
#> 0.905 0.924 0.934 0.907 0.923

# The old name still works and returns the same result:
identical(KMO(cor_mat)$KMO, efa_kmo(cor_mat)$KMO)
#> [1] TRUE

Migrating from `EFA()` to `efa_fit()`

EFA() exposed every estimation and rotation setting as a flat argument, which made for a long and somewhat unwieldy signature. efa_fit() keeps the primary choices — the data, factor count, sample size, estimator, rotation, and a few others — as top-level arguments (see ?efa_fit for the full list), and collects the tuning knobs into two small control objects built by estimate_control() and rotate_control().

Each flat EFA() tuning argument now lives in one of the two controls:

Control object	Arguments it holds
`estimate_control()`	`type`, `init_comm`, `criterion`, `criterion_type`, `max_iter`, `abs_eigen`, `start_method`
`rotate_control()`	`type`, `normalize`, `precision`, `order_type`, `varimax_type`, `p_type`, `k`, `random_starts`

A few points make the translation mechanical:

The type preset feeds both controls. EFA() had a single type ("EFAtools", "psych", "SPSS", or "none") that governed both estimation and rotation. Pass the same type to each constructor to reproduce it. Because the two controls are independent, you can now give them different presets, but you do not have to.
Three arguments were renamed. The estimator is selected with estimator instead of method — efa_fit() has no method argument, while EFA() keeps it. P_type is now p_type, and randomStarts is now random_starts; EFA() still accepts those old spellings silently.
Rotation-engine extras — the criterion-specific arguments a rotation may take, such as maxit, gam (oblimin), or delta (geomin) — are passed through rotate_control()’s .... It validates the names: an extra that no rotation engine can consume (for example gamma, a misspelling of oblimin’s gam) is rejected with an error, where EFA() silently ignores it.
The default rotation differs. efa_fit(), efa_group(), and efa_mi() use an oblimin rotation unless you set rotation, while EFA() and EFA_POOLED() keep no rotation as their default.

The side-by-side below reproduces an SPSS-style analysis. The old flat call and the new control-based call give identical results:

# Old flat interface
efa_old <- EFA(cor_mat, n_factors = 3, N = 500, rotation = "promax", type = "SPSS")

# New interface: the SPSS preset travels through the two control objects
efa_new <- efa_fit(cor_mat, n_factors = 3, N = 500, rotation = "promax",
                   estimate_control = estimate_control(type = "SPSS"),
                   rotate_control = rotate_control(type = "SPSS"))

# Identical numerical result
all.equal(efa_old$rot_loadings, efa_new$rot_loadings)
#> [1] TRUE

An individual knob is set on the control it belongs to. A flat EFA(..., type = "SPSS", max_iter = 500, k = 3) becomes:

efa_fit(cor_mat, n_factors = 3, N = 500, rotation = "promax",
        estimate_control = estimate_control(type = "SPSS", max_iter = 500),
        rotate_control   = rotate_control(type = "SPSS", k = 3))

efa_fit() has no argument for a tuning knob, so a knob passed directly is an error, not silently ignored:

efa_fit(cor_mat, n_factors = 3, N = 500, max_iter = 500)
#> Error in efa_fit(cor_mat, n_factors = 3, N = 500, max_iter = 500) :
#>   unused argument (max_iter = 500)

The same move — collecting flat estimation knobs into estimate_control() — applies to the other functions that fit a model internally: efa_retain(), efa_schmid_leiman(), and the retention criteria that fit a model or use EFA-based eigenvalues — efa_parallel(), efa_kgc(), efa_scree(), efa_hull(), efa_nest(), and efa_smt(). efa_average() is the exception: it keeps its flat estimation and rotation arguments, and only renames P_type to p_type.

What the `type` Presets Replicate

The type presets are more than shorthand for a bundle of defaults. "SPSS" and "psych" follow how SPSS’s FACTOR procedure and psych::fa() implement principal axis factoring and promax rotation, and "EFAtools", the default, combines the settings that performed best in a systematic comparison of the three (Grieder and Steiner, 2022). There is one place a preset alone is not enough: because every preset keeps EFAtools’ Kaiser normalization, "psych" also needs normalize = FALSE to reproduce psych::fa()’s promax — and even then a small residual difference remains, from ordinary convergence slack between the two implementations.

Returned Objects Keep Their Legacy Classes

Migrating a call does not break code that inspects or dispatches on the result. The renamed functions attach the legacy class alongside the new one, so inherits() checks and S3 methods written against the old classes keep working — including on objects saved from earlier sessions.

class(efa_new)
#> [1] "efa" "EFA"
inherits(efa_new, "EFA")
#> [1] TRUE

The same holds for the other direct renames: efa_retain() objects still inherit "N_FACTORS", efa_bartlett() still inherits "BARTLETT", and likewise for efa_kmo(), efa_schmid_leiman(), efa_compare(), efa_average(), and efa_mi(). The retention criteria (efa_parallel(), efa_ekc(), and the rest) all share the single efa_retention class they have returned since EFAtools 0.8.0. And the uppercase functions themselves keep working unchanged: an EFA() call returns the same result as before, now additionally classed efa so that it picks up the shared methods — inherits() checks against "EFA" are unaffected.

Where to Next

That is the whole migration: rename the function, and — for EFA() only — move the tuning knobs into estimate_control() and rotate_control(). For an overview of the full analysis workflow under the efa_* interface, see the EFAtools vignette; the individual help pages document each function’s arguments in full.

Migrating to the efa_* interface