---
title: "Evaluating Circumplex Structure"
output: rmarkdown::html_vignette
vignette: >
%\VignetteIndexEntry{Evaluating Circumplex Structure}
%\VignetteEngine{knitr::rmarkdown}
%\VignetteEncoding{UTF-8}
---
``` r
library(circumplex)
```
## 1. Two questions to ask before interpreting an SSM analysis
The Structural Summary Method (SSM) condenses a circumplex profile into a
few interpretable parameters: elevation, amplitude, displacement, and fit.
That condensation rests on two assumptions that are testable but often left
untested:
1. **Does the instrument actually have circumplex structure?** The SSM
locates a profile on a circle whose positions are given by the scales'
theoretical angles. If the scales do not really form a circumplex — or
sit at very different angles than the theory says — the summary is
distorted before any profile is computed.
2. **Are the confidence intervals trustworthy at your sample size and
profile?** Zimmermann and Wright (2017) showed that the accuracy of
bootstrap SSM intervals depends on the sample size, the instrument's
correlation structure, and — for amplitude and displacement — on how
differentiated the profile really is in the population. Intervals that
are accurate for one construct in your table can be inaccurate for the
construct printed directly below it.
This vignette shows how to answer both questions with `cpm_fit()` and
`fit_structure()` (question 1) and `ssm_ci_accuracy()` (question 2), using
the `jz2017` dataset — the same sample of 1,166 undergraduates analyzed in
Zimmermann and Wright (2017, Study 5), with octant scores on the IIP-SC and
personality disorder (PD) scales from the PDQ-4+.
## 2. Does the instrument fit a circumplex?
### Fitting the circular process model
`cpm_fit()` estimates Browne's (1992) circular process model (CPM), the
confirmatory model behind the CIRCUM and CircE programs. Each scale gets an
estimated angle on the circle and a *communality index* $\zeta$ (the
correlation between the scale and the common circumplex "factor"; its
square, $\zeta^2$, is the scale's *communality* — the share of its
variance the circumplex explains). The
model also yields the usual covariance-structure fit indices.
On the raw-data path the default confidence intervals are bootstrapped. We
lower `boots` from its default of 2000 to keep this vignette quick to
build; in practice, use the default.
``` r
data("jz2017")
set.seed(12345)
cpm <- cpm_fit(jz2017, scales = PANO(), angles = octants(), boots = 500)
#> Warning: CPM Hessian is ill-conditioned (condition number 1.83e+14): angles may
#> be clustered or parameters weakly determined.
#> Warning: 11 of 500 bootstrap resamples were excluded (0 with a degenerate or
#> non-positive-definite correlation matrix, 11 failing the convergence acceptance
#> criterion); the confidence intervals are based on the remaining 489 replicates
#> and are conditional on estimability.
summary(cpm)
#>
#> Circular Process Model (Browne, 1992)
#> Model: quasi-circumplex
#> Harmonics (m): 3
#> Sample size (N): 1166
#> Reference scale: PA
#> CI method: bootstrap
#> Confidence level: 0.95
#>
#> # Estimated angles and communality indices
#>
#> Scale Angle_theory Angle Angle_lci Angle_uci Zeta Zeta_lci Zeta_uci
#> PA 90 90.000 90.000 90.000 0.767 0.684 0.860
#> BC 135 125.074 113.548 137.069 0.931 0.872 1.000
#> DE 180 170.353 156.943 185.174 0.780 0.733 0.826
#> FG 225 195.425 185.811 206.523 0.861 0.822 0.910
#> HI 270 250.721 244.660 258.758 0.956 0.935 0.977
#> JK 315 269.491 261.813 279.412 0.942 0.929 0.957
#> LM 360 294.230 286.626 303.019 0.806 0.761 0.850
#> NO 45 11.305 2.426 20.913 1.000 1.000 1.000
#> Communality
#> 0.589
#> 0.868
#> 0.608
#> 0.741
#> 0.914
#> 0.888
#> 0.650
#> 1.000
#>
#> # Correlation-function weights
#>
#> k Beta Beta_lci Beta_uci
#> 0 0.450 0.418 0.481
#> 1 0.440 0.411 0.466
#> 2 0.074 0.065 0.087
#> 3 0.036 0.015 0.057
#>
#> # Fit indices
#>
#> χ²(10) = 81.169, p = <1e-04
#> RMSEA = 0.078 [0.063, 0.094] (90% CI)
#> SRMR = 0.042
#> CFI = 0.984 TLI = 0.956
#> AIC = 117.169 BIC = 208.273
#>
#> # Residuals
#>
#> Largest absolute residual: 0.134 (PA – HI)
#>
#> # Diagnostics
#>
#> Note: a communality index reached its upper boundary (ζ > 0.995, a Heywood-type solution).
#> Note: 11 of 500 bootstrap resamples were excluded (0 degenerate, 11 non-convergent); the intervals are based on 489 replicates and are conditional on estimability.
#>
#> Note: boundary/weak-identification markers fired: Heywood communality;
#> small correlation-function weight; ill-conditioned Hessian.
#> What has been measured about these markers covers analytic intervals
#> only, and not every marker was measured; they are not validated as
#> predictors of the bootstrap intervals shown here (see the vignette
#> section 'When a fit sits at a boundary').
```
Two parts of this output matter most for evaluating structure:
- **The estimated angles.** Compare them with the theoretical angles, but
read the comparison carefully: one scale is held fixed to identify the
configuration (PA here, at 90°), so every other scale's departure is
measured from that anchor and a different anchor would redistribute them.
Two things are worth separating. The *ordering* around the circle is
preserved — the estimated angles run through the octants in the same
cyclic order the instrument assigns them. The *spacing* is not: the gaps
between circularly adjacent estimates run from under 20° to nearly 80°
against a theoretical 45°, and the largest departure from a theoretical
position is about 66°. Departures from perfect structure are common in
well-validated circumplex instruments (Gurtman & Pincus, 2000), and the
model comparison below quantifies what this pattern costs: it is why
forcing equal spacing fits these data poorly.
- **The communality indices.** Scales with low $\zeta$ are poorly
described by the circle; a profile peak at such a scale's angle means
less than the same peak at a well-explained scale's angle.
The estimated configuration is easiest to read on the circle itself: each
scale is drawn at its estimated angle, at a radius equal to its communality
($\zeta^2$), with a joint confidence region where one is estimable.
``` r
plot(cpm)
```
### Reading the fit indices
The `summary()` output reports the standard indices, and the standard
benchmarks apply: RMSEA at or below about .08 suggests adequate
approximate fit and above about .10 poor fit (Browne & Cudeck, 1993); SRMR
at or below about .08 suggests acceptable average residuals (Hu & Bentler,
1999); CFI and TLI near .95 or above suggest good comparative fit (Hu &
Bentler, 1999). Treat these as conventions from the broader
covariance-structure literature, not circumplex-specific laws.
Two caveats come from the benchmark sources themselves. Browne and Cudeck
call their own RMSEA thresholds "based on subjective judgment," cautioning
that such a figure "cannot be regarded as infallible or correct" (Browne &
Cudeck, 1993) — read the cutoffs as conventions, not decision rules. And these
benchmarks are least dependable at small samples: Hu and
Bentler (1999) found that the ML-based TLI and RMSEA tend to *overreject*
true-population models when the sample is small (CFI is not among the indices
they flag), and circumplex analyses are often run at modest sample sizes —
the SSM accuracy thresholds in Section 3 span roughly $n = 50$ to $200$. At
small $n$, then, a TLI or RMSEA that falls short of its benchmark can reflect
the index's small-sample behavior as much as the model's fit.
Two further cautions are circumplex-specific, both from this package's own
validation simulations:
- **Boundary solutions are common at realistic sample sizes**, including in
the fit above. What that means and what to do about it is the subject of
the next subsection, *When a fit sits at a boundary*.
- **Do not lean hard on the chi-square p value.** In simulations at
field-typical sample sizes with octant-like population structures, the
test statistic did not follow its nominal chi-square reference
distribution, even for correctly specified models. Read the chi-square
descriptively (bigger is worse), and prefer RMSEA/SRMR/CFI for judging
approximate fit.
For the same reasons, when the intervals are analytic (Wald) — the only
option on the correlation-matrix input path — `summary()` prints a caution
below $N = 2000$, and up to very large $N$ when a boundary marker is
present: in validation, analytic intervals mis-covered in exactly those
regimes. Prefer the bootstrap (the raw-data default) when you have raw
data.
### When a fit sits at a boundary
The fit above is one of these fits, and it says so. Its `# Diagnostics` block
prints a note that a communality index reached its upper boundary, and the
results table shows which one: the NO scale's communality index is 1.000, with
a confidence interval of [1.000, 1.000]. A zero-width interval is not a
precise estimate — it is an absent one. Both endpoints equal the estimate,
which is to say that at least the middle 95% of the retained bootstrap
resamples also sat on the boundary, leaving the percentile interval no width
to report. The same
chunk also emits a warning that the fit's Hessian is ill-conditioned, a second
signal from the same regime. When octant scales share a strong general factor,
as interpersonal problem scales do, fitted solutions frequently sit at or near
a parameter boundary; this is the estimator meeting real data at finite $n$,
not a data-entry error.
`cpm_fit()` records five such markers, and `summary()` names the ones that
fired:
- **Heywood communality** — a communality index estimated at its upper
boundary ($\hat\zeta > 0.995$): the circle accounts for that scale
essentially completely.
- **boundary harmonic removed** — a correlation-function weight sat at zero
and was dropped from the model, with the degrees of freedom adjusted.
- **small correlation-function weight** — the smallest weight is below 0.10,
near enough to zero to matter for inference.
- **ill-conditioned Hessian** — the curvature matrix the analytic standard
errors are computed from is close to singular; angles may be clustered or
parameters weakly determined.
- **competing near-tied optima** — the search found more than one solution of
nearly equal fit, so the reported one may not be uniquely determined.
`summary()` prints that list when markers fire, though on the analytic path
only within a sample-size window. On the bootstrap path — as in the fit
displayed above — it prints as a descriptive note at every sample size,
alongside a statement that what has been measured about the markers covers
analytic intervals only (and not every marker was measured — *What a fired
marker does and does not tell you*, below, has the particulars), so they are
not validated as predictors of the bootstrap intervals they accompany. When the intervals are analytic the list instead
arrives inside a coverage caution printed when $N$ falls between 2000 and
50000; below 2000 an unconditional caution prints without the list, and
above 50000 neither prints. To see that caution-embedded form, we simulate
a larger sample from the structure just estimated with `cpm_simulate()`,
then refit it on the correlation-matrix path, where the intervals are
analytic.
``` r
set.seed(2026)
sim <- cpm_simulate(cpm, n = 2500) # draws from the structure estimated above
demo <- cpm_fit(
cormat = cor(sim), scales = colnames(sim), angles = octants(), n = 2500
)
summary(demo)
#>
#> Circular Process Model (Browne, 1992)
#> Model: quasi-circumplex
#> Harmonics (m): 3
#> Sample size (N): 2500
#> Reference scale: PA
#> CI method: analytic
#> Confidence level: 0.95
#>
#> # Estimated angles and communality indices
#>
#> Scale Angle_theory Angle Angle_lci Angle_uci Zeta Zeta_lci Zeta_uci
#> PA 90 90.000 90.000 90.000 0.785 0.748 0.823
#> BC 135 125.057 120.571 129.542 0.881 0.852 0.911
#> DE 180 167.802 161.988 173.616 0.799 0.777 0.820
#> FG 225 193.344 187.711 198.976 0.878 0.854 0.901
#> HI 270 250.958 245.112 256.805 0.958 0.944 0.971
#> JK 315 269.528 263.351 275.706 0.948 0.938 0.958
#> LM 360 293.619 286.622 300.617 0.803 0.778 0.829
#> NO 45 12.673 7.017 18.329 0.929 0.836 1.023
#> Communality
#> 0.617
#> 0.777
#> 0.638
#> 0.770
#> 0.917
#> 0.899
#> 0.645
#> 0.863
#>
#> # Correlation-function weights
#>
#> k Beta Beta_lci Beta_uci
#> 0 0.442 0.422 0.462
#> 1 0.453 0.436 0.470
#> 2 0.068 0.058 0.078
#> 3 0.037 0.030 0.044
#>
#> # Fit indices
#>
#> χ²(10) = 2.229, p = 0.994
#> RMSEA = 0 [0, 0] (90% CI)
#> SRMR = 0.005
#> CFI = 1 TLI = 1.002
#> AIC = 38.229 BIC = 143.062
#>
#> # Residuals
#>
#> Largest absolute residual: 0.012 (FG – NO)
#>
#> Note: this solution is near a parameter boundary or weakly identified
#> (small correlation-function weight);
#> analytic (Wald) confidence intervals mis-covered for such fits in validation
#> even at N in the tens of thousands. Interpret them with caution and prefer
#> the bootstrap on the raw-data path when available.
```
One marker fires here — a **small correlation-function weight** — and
`summary()` names it. Notice what did *not* happen:
the population this sample was drawn from carries NO's communality at the
boundary, yet this draw produced no Heywood case. A boundary in the
population does not guarantee a boundary in every sample from it, any more
than a quiet fit guarantees there is no boundary behind the data.
**What a fired marker does and does not tell you.** The package's validation
simulations measured what these markers predict, and the measurement is
narrower than the markers are. It covered *analytic* (Wald) intervals fitted
from a correlation matrix, and it measured interval **coverage** — how often
an interval contained the truth — not bias in the point estimates.
- *Communality indices* take the worst of it **when an ill-conditioning, a
Heywood, or a near-tied-optima marker fires** — all three covered $\zeta$
well below nominal, ill-conditioning worst of them. Two failure modes show up in the
output rather than hiding in it: the analytic standard errors can come back
missing, which happens for the angles and the weights at the same time
because one singular curvature matrix takes all three families down
together, or an interval can collapse to zero width because a negative
asymptotic variance was clamped at zero. NO's zero-width interval above
arrives by a third route — it is a percentile interval whose resamples sat
on the boundary — but it reaches the reader the same way, as an interval
with no usable width.
- *Angles.* Taking all marker-firing fits together, angle intervals actually
degraded somewhat *more* than communality intervals did, so the ranking
above belongs to those markers rather than to the parameter families in
general. The strongest single angle signal was the ill-conditioning marker,
with the small-weight marker next and the Heywood marker weakest of the
three.
- *Correlation-function weights* moved least: marker-firing fits covered
$\beta$ only slightly less often than quiet ones, so a fired marker tells
you least here. The real $\beta$ problem is a different one and does not
depend on any marker — in the package's bootstrap validation, percentile
intervals for a weight whose *population* value sits near zero under-covered
at every sample size studied, and the shortfall did not shrink as $n$ grew.
That is a property of percentile intervals at a boundary, not a small-sample
artifact, and it is about where the truth sits, not about what the fit
flagged.
Four things the record does *not* support, and which no reading of these
markers should assume. Nothing was measured about bias in the point estimates.
The marker study fitted analytic intervals only, so the markers are not
validated as predictors on the bootstrap path this vignette uses by default.
The **boundary harmonic removed** marker showed no evidence of predicting
mis-coverage at all — fits carrying it covered as well as fits without it; it
is kept because it names a real feature of the solution, not because it
forecasts trouble. And the numbers behind **Heywood communality**,
**ill-conditioned Hessian** and **competing near-tied optima** come almost
entirely from one deliberately provoking configuration, where those markers
fire often enough to measure, and near-tied optima fired only rarely even
there. Read them as what happens where those markers fire, not as how
often they fire.
**What to do when one fires.**
1. *Locate it.* Read the results table beside the note and find the scale or
weight sitting at the boundary, as NO is above. A marker with no identified
owner is not yet a diagnosis.
2. *Re-fit to find out what the boundary is about.* The default
`"quasi-circumplex"` model is the least constrained of the variants in the
next subsection, so a boundary that appears only under a constrained
variant was put there by the constraint, while one already present in the
default fit belongs to the data and the model family rather than to an
assumption you added. If you fitted from a correlation matrix, refitting
from raw data also buys you bootstrap intervals in place of analytic ones.
3. *Report it.* Quote the marker alongside the estimate it belongs to: which
scales sat at the boundary, and that their intervals are degenerate or
missing. An angle or communality reported without its diagnostic note is a
number stripped of its own caveat.
4. *Keep using what still holds.* These remain usable: the point estimates,
as the solution the model actually fitted (their bias in this regime is
unmeasured, so read them as estimates, not to the digit); the fit indices
and the residual summary, which a boundary changes the behaviour of the
*intervals* around, not the computation of — read them with the caution
given earlier in this section, which is about field sample sizes rather
than about markers; and parameter intervals that came back with ordinary
width, read knowing that marked fits covered less well than unmarked ones
across all three families. This does not: a parameter interval that is
missing or zero-width. (The demonstration fit above prints
`RMSEA = 0 [0, 0]`, which is not that failure — it is a fit statistic at
its own floor, because the refit recovers almost exactly the structure it
was simulated from.)
### Comparing model variants
Constrained variants make the structural question sharp: is the instrument
consistent with *equally spaced* scales? With *equal* communalities? The
correlation-matrix path is convenient for this comparison (point estimates
and fit indices are deterministic, so no bootstrap is needed):
``` r
R <- cpm$matrices$R # the sample correlation matrix stored by cpm_fit()
fits <- lapply(
c("quasi-circumplex", "equal-communality", "circulant"),
function(mod) {
cpm_fit(
cormat = R, scales = PANO(), angles = octants(),
n = nrow(jz2017), model = mod
)
}
)
#> Warning: CPM Hessian is ill-conditioned (condition number 1.83e+14): angles may
#> be clustered or parameters weakly determined.
data.frame(
model = vapply(fits, function(f) f$details$model, character(1)),
df = vapply(fits, function(f) f$fit$df, numeric(1)),
rmsea = round(vapply(fits, function(f) f$fit$rmsea, numeric(1)), 3),
srmr = round(vapply(fits, function(f) f$fit$srmr, numeric(1)), 3),
cfi = round(vapply(fits, function(f) f$fit$cfi, numeric(1)), 3),
tli = round(vapply(fits, function(f) f$fit$tli, numeric(1)), 3)
)
#> model df rmsea srmr cfi tli
#> 1 quasi-circumplex 10 0.078 0.042 0.984 0.956
#> 2 equal-communality 17 0.100 0.063 0.956 0.928
#> 3 circulant 24 0.185 0.130 0.790 0.755
```
The pattern reproduces what Zimmermann and Wright (2017, p. 14) reported
for these data with CircE: the fully constrained model (equal spacing *and*
equal communality — the `"circulant"` variant) fits poorly, while relaxing
the equal-spacing constraint improves fit to the edge of the conventional
benchmarks (marginal by RMSEA, acceptable by SRMR and CFI); they reported
CFI = .824, TLI = .795, RMSEA = .169 for the former and CFI = .958, TLI =
.931, RMSEA = .098 for the latter. Do not expect the default fit's indices to
match theirs to the digit: CIRCUM/CircE fit a covariance version of the model
with free scaling constants, whereas `cpm_fit()` by default fits the
correlation structure directly, and the two differ slightly at finite sample
sizes; pass `scaling = "free"` to fit their covariance parameterization and
reproduce published CIRCUM/CircE output exactly. For the model test the choice
does not matter here: with correlation input the two families' test statistics
are calibration-indistinguishable (differing by well under 1% of the model
degrees of freedom in paired simulation at sample sizes 250–50,000; the free
statistic never exceeds the default's on the same input beyond numerical
tolerance, since the free family nests the default and is additionally
started from its solution), and
the default is recommended for
routine inference because free scaling adds parameters whose analytic standard
errors are often undefined at small-to-moderate samples. The
*conclusion* — ordered octants with unequal spacing, adequate approximate
fit once equal spacing is not forced — is the same.
A poor CPM fit does not make SSM output uncomputable; it makes the summary
less meaningful, because the profile is being referred to scale positions
the data contradict. If the ordering itself fails (scales out of sequence
around the circle), SSM parameters should not be interpreted.
## 3. Can you trust your confidence intervals?
### What Zimmermann & Wright (2017) found
Zimmermann and Wright's simulation studies (their Studies 1–4) evaluated
the accuracy of 95% percentile bootstrap intervals for SSM parameters —
the same interval type `ssm_analyze()` reports — judging an interval
accurate when its empirical coverage stayed within Bradley's (1978)
liberal band of 92.5% to 97.5%. Their headline results:
| Parameter | Point estimate | 95% bootstrap CI accurate when... |
|---|---|---|
| Elevation ($e$) | essentially unbiased | $n \ge 50$ |
| X value / affiliation | essentially unbiased | $n \ge 50$ |
| Y value / dominance | essentially unbiased | $n \ge 50$ |
| Amplitude ($a$) | **biased upward**, strongly so when population amplitude is small | $n \ge 75$ (general-factor instrument) or $n \ge 150$ (no general factor), *given* population amplitude $\ge .10$ |
| Displacement ($\delta$) | unbiased but imprecise at low amplitude | $n \ge 100$ (general factor) or $n > 200$ (no general factor), *given* population amplitude $\ge .10$ |
| Fit ($R^2$) | biased downward | population $R^2 < .9$ only; unsuited near 1 |
*(Transcribed from Zimmermann & Wright, 2017, Studies 1–2, pp. 6–11.)*
Three implications are worth internalizing:
- **Amplitude overestimates differentiation.** Across their conditions the
relative bias averaged 15.5% and reached 135.8%; with $n = 50$ and no
general factor, the *expected* sample amplitude was about .15 when the
population amplitude was exactly zero (their pp. 6–7). An amplitude of
.15 is commonly read as "marked"; at small $n$ it can be pure sampling
artifact.
- **Displacement is only as stable as amplitude is large.** The standard
error of the displacement grows as amplitude shrinks (e.g., roughly 50
degrees at $n = 100$ for a weakly differentiated profile; their p. 8).
A displacement estimate from a flat profile has no meaningful direction.
- **The thresholds above are grid minima, not guarantees.** They come from
specific instruments, specific sample sizes, and population amplitudes
of .10 or larger. Their Study 3 quantified the frontier more finely: the
minimum population affiliation/dominance component needed for accurate
amplitude and displacement intervals is approximately
$2.95 \cdot f_a \cdot n^{-0.587}$, where $f_a$ is an instrument
constant (their Eq. 3; ≈ .55 for the IIP-C, .63 for the IIP-SC, .85 for
the IAS). At $n = 100$ with an IIP-C-like instrument that is a
population component of about .11 — larger than many real construct
profiles.
### Simulating coverage for your own analysis
Published thresholds cover a coarse grid of conditions, and your analysis
— your instrument, your $n$, your group sizes, your profile, your number
of bootstrap resamples — usually falls between or outside the grid points.
`ssm_ci_accuracy()` answers the question directly at your configuration:
it builds a population whose structure matches your fitted estimates
(through the CPM by default), simulates many datasets of your exact sample
size from it, replays your own CI procedure on each, and reports how often
the intervals covered the known population values.
To see it catch a real problem, we analyze two PD scales at a modest
sample size: paranoid PD (a well-differentiated interpersonal profile) and
obsessive–compulsive PD, whose profile Zimmermann and Wright's Table 4
showed to be nearly flat (amplitude .012 at full sample size). We use the
first 250 participants — a sample size many construct-validation studies
would consider respectable.
``` r
set.seed(23456)
res <- ssm_analyze(
jz2017[1:250, ],
scales = PANO(),
angles = octants(),
measures = c("PARPD", "OCPD"),
boots = 500 # reduced from the default 2000 to keep the vignette quick;
# the diagnostic below replays whatever procedure this object used
)
summary(res)
#>
#> Statistical Basis: Correlation Scores
#> Bootstrap Resamples: 500
#> Confidence Level: 0.95
#> Listwise Deletion: TRUE
#> Scale Displacements: 90 135 180 225 270 315 360 45
#>
#>
#> # Profile [PARPD]:
#>
#> Estimate Lower CI Upper CI
#> Elevation 0.258 0.169 0.338
#> X-Value -0.095 -0.160 -0.028
#> Y-Value 0.055 -0.017 0.122
#> Amplitude 0.110 0.054 0.185
#> Displacement 149.667 111.480 190.360
#> Model Fit 0.725
#>
#>
#> # Profile [OCPD]:
#>
#> Estimate Lower CI Upper CI
#> Elevation 0.218 0.123 0.302
#> X-Value -0.021 -0.084 0.037
#> Y-Value -0.003 -0.068 0.063
#> Amplitude 0.021 0.009 0.096
#> Displacement 188.695 29.052 335.755
#> Model Fit 0.150
#> Note: model fit is inadequate (R² < .70); interpret only the elevation parameter.
#> Note: the amplitude CI lower bound is under 0.35 CI-widths above zero; the displacement is not interpretable.
```
Note the guardrail in the printed output: when a profile's amplitude CI
lower bound sits less than 0.35 CI-widths above zero, `ssm_analyze()` marks
the displacement as uninterpretable rather than certifying its direction.
That rule is scale-free (it compares the lower bound to the interval's own
width, so it means the same thing on any score metric) and does not depend
on the display precision. The diagnostic below measures, among other things,
how well that certification rule performs at your configuration.
We again use reduced settings (`reps = 200`, a short amplitude ladder, and
the smaller `boots` above) to keep the vignette fast; the defaults
(`reps = 1000`, four ladder rungs, on an object built with the default
`boots = 2000`) give a Monte Carlo standard error under one percentage
point for coverage near the nominal level and are recommended in practice.
The `parallel`/`ncpus` arguments speed up real runs without changing
results for a given seed.
``` r
set.seed(34567)
acc <- ssm_ci_accuracy(res, reps = 200, amplitude_factors = c(1, 0.5, 0))
#> Warning: CPM Hessian is ill-conditioned (condition number 9.24e+16): angles may
#> be clustered or parameters weakly determined.
summary(acc)
#>
#> Statistical Basis: Correlation Scores
#> Assessed Engine: bootstrap with 500 replicates
#> Confidence Level: 0.95
#> Simulation Reps: 200 per condition
#> Amplitude Ladder: 1 0.5 0 2.077
#> Population Structure: Browne circular model (CPM)
#> Group Sizes: All = 250
#> Certification Rule: a_lci / (a_uci - a_lci) >= 0.35 (scale-free, print-independent)
#> Elapsed: 7.6 s
#>
#> Structure note: population simulated from a Browne circular model fit (m = 3,
#> RMSEA = 0.064, SRMR = 0.038).
#> The structure fits adequately (RMSEA <= 0.08, Browne & Cudeck, 1993; SRMR
#> <= 0.08, Hu & Bentler, 1999), so the simulated population is a reasonable
#> stand-in for yours.
#> Boundary markers: Heywood communality; small correlation-function weight;
#> ill-conditioned Hessian.
#> Near-zero regime: the amplitude estimate of profile [OCPD] is below half its
#> own CI width, so your analysis already sits in the amplitude-near-zero
#> regime; an absolute rung at the certification margin (c = 2.08, population
#> amplitude = the observed CI half-width) was added to the ladder.
#>
#> CI trustworthiness at the as-estimated condition (c = 1), classified
#> against Bradley's (1978) liberal band via 95% Wilson intervals:
#>
#> # Profile [PARPD] (n = 250; 95% bootstrap CIs, 500 replicates):
#> Elevation coverage 93.0% -- borderline
#> Amplitude coverage 94.5% -- borderline
#> Displacement coverage 89.1% when certified -- borderline
#> Guardrail under a truly zero amplitude, displacement would be
#> certified 1.0% of the time (user-expectation benchmark
#> 2.5%)
#> Verdict: BORDERLINE -- elevation, amplitude, and certified displacement
#> coverage rates are borderline at this number of replications; a larger
#> `reps` would sharpen the verdict.
#>
#> # Profile [OCPD] (n = 250; 95% bootstrap CIs, 500 replicates):
#> Elevation coverage 95.0% -- borderline
#> Amplitude coverage 83.0% -- INADEQUATE (under-coverage; misses are
#> almost all below the interval: the amplitude CI tends to
#> sit above the truth)
#> Displacement coverage 50.0% when certified -- INADEQUATE
#> (under-coverage)
#> Guardrail under a truly zero amplitude, displacement would be
#> certified 1.0% of the time (user-expectation benchmark
#> 2.5%)
#> Verdict: CAUTION -- amplitude CIs are less reliable than nominal at this
#> sample size and displacement CIs mis-cover even when certified. Elevation
#> coverage is borderline at this number of replications; a larger `reps`
#> would sharpen the verdict. Consider a larger sample or treat near-zero
#> amplitudes as inconclusive rather than absent.
#>
#> Coverage by profile, parameter, and amplitude condition:
#> Profile Parameter Condition Coverage MC_se Left_miss Right_miss Median_width
#> PARPD e 1.000 0.930 0.018 0.035 0.035 0.151
#> PARPD x 1.000 0.940 0.017 0.030 0.030 0.122
#> PARPD y 1.000 0.930 0.018 0.020 0.050 0.131
#> PARPD a 1.000 0.945 0.016 0.040 0.015 0.125
#> PARPD d 1.000 0.920 0.019 0.035 0.045 63.326
#> OCPD e 1.000 0.950 0.015 0.025 0.025 0.156
#> OCPD x 1.000 0.945 0.016 0.025 0.030 0.126
#> OCPD y 1.000 0.965 0.013 0.030 0.005 0.131
#> OCPD a 1.000 0.830 0.027 0.170 0.000 0.103
#> OCPD d 1.000 0.830 0.027 0.075 0.095 220.815
#> PARPD e 0.500 0.945 0.016 0.025 0.030 0.151
#> PARPD x 0.500 0.935 0.017 0.025 0.040 0.124
#> PARPD y 0.500 0.935 0.017 0.015 0.050 0.128
#> PARPD a 0.500 0.935 0.017 0.065 0.000 0.111
#> PARPD d 0.500 0.905 0.021 0.050 0.045 154.830
#> OCPD e 0.500 0.940 0.017 0.030 0.030 0.158
#> OCPD x 0.500 0.940 0.017 0.030 0.030 0.128
#> OCPD y 0.500 0.935 0.017 0.020 0.045 0.131
#> OCPD a 0.500 0.485 0.035 0.515 0.000 0.102
#> OCPD d 0.500 0.710 0.032 0.170 0.120 256.592
#> PARPD e 0.000 0.940 0.017 0.025 0.035 0.153
#> PARPD x 0.000 0.935 0.017 0.030 0.035 0.124
#> PARPD y 0.000 0.950 0.015 0.040 0.010 0.131
#> PARPD a 0.000 0.000 0.000 1.000 0.000 0.098
#> PARPD d 0.000 NA NA NA NA NA
#> OCPD e 0.000 0.920 0.019 0.040 0.040 0.157
#> OCPD x 0.000 0.955 0.015 0.035 0.010 0.124
#> OCPD y 0.000 0.945 0.016 0.020 0.035 0.131
#> OCPD a 0.000 0.000 0.000 1.000 0.000 0.100
#> OCPD d 0.000 NA NA NA NA NA
#> PARPD e 2.077 0.945 0.016 0.025 0.030 0.150
#> PARPD x 2.077 0.950 0.015 0.025 0.025 0.118
#> PARPD y 2.077 0.935 0.017 0.035 0.030 0.127
#> PARPD a 2.077 0.930 0.018 0.035 0.035 0.125
#> PARPD d 2.077 0.955 0.015 0.015 0.030 31.504
#> OCPD e 2.077 0.930 0.018 0.040 0.030 0.157
#> OCPD x 2.077 0.915 0.020 0.055 0.030 0.127
#> OCPD y 2.077 0.945 0.016 0.030 0.025 0.131
#> OCPD a 2.077 0.930 0.018 0.070 0.000 0.109
#> OCPD d 2.077 0.885 0.023 0.050 0.065 160.646
#> Coverage_conditional N_conditional Structural N_reps
#> NA NA FALSE 200
#> NA NA FALSE 200
#> NA NA FALSE 200
#> NA NA FALSE 200
#> 0.891 147 FALSE 200
#> NA NA FALSE 200
#> NA NA FALSE 200
#> NA NA FALSE 200
#> NA NA FALSE 200
#> 0.500 8 FALSE 200
#> NA NA FALSE 200
#> NA NA FALSE 200
#> NA NA FALSE 200
#> NA NA FALSE 200
#> 0.893 28 FALSE 200
#> NA NA FALSE 200
#> NA NA FALSE 200
#> NA NA FALSE 200
#> NA NA FALSE 200
#> 0.333 3 FALSE 200
#> NA NA FALSE 200
#> NA NA FALSE 200
#> NA NA FALSE 200
#> NA NA TRUE 200
#> NA 0 FALSE 0
#> NA NA FALSE 200
#> NA NA FALSE 200
#> NA NA FALSE 200
#> NA NA TRUE 200
#> NA 0 FALSE 0
#> NA NA FALSE 200
#> NA NA FALSE 200
#> NA NA FALSE 200
#> NA NA FALSE 200
#> 0.955 200 FALSE 200
#> NA NA FALSE 200
#> NA NA FALSE 200
#> NA NA FALSE 200
#> NA NA FALSE 200
#> 0.833 18 FALSE 200
#> Note: amplitude coverage on rows flagged Structural is structurally 0 (a
#> percentile interval of strictly positive amplitude replicates cannot
#> contain a zero truth) -- a theorem, not a measurement; the informative
#> near-zero rungs are the small c > 0 ones.
#>
#> Guardrail operating characteristics:
#> Profile Condition Cert_rate Cert_lci Cert_uci Benchmark Caution Fit_pass_rate
#> PARPD 1.000 0.735 0.670 0.791 0.025 NA 0.555
#> OCPD 1.000 0.040 0.020 0.077 0.025 NA 0.075
#> PARPD 0.500 0.140 0.099 0.195 0.025 NA 0.115
#> OCPD 0.500 0.015 0.005 0.043 0.025 NA 0.045
#> PARPD 0.000 0.010 0.003 0.036 0.025 FALSE 0.020
#> OCPD 0.000 0.010 0.003 0.036 0.025 FALSE 0.070
#> PARPD 2.077 1.000 0.981 1.000 0.025 NA 0.985
#> OCPD 2.077 0.090 0.058 0.138 0.025 NA 0.115
#> Branch_pathology_rate N_reps
#> 0 200
#> 0 200
#> 0 200
#> 0 200
#> 0 200
#> 0 200
#> 0 200
#> 0 200
```
How to read this output:
- **The coverage table** reports, for each parameter, how often your CI
procedure covered the truth in the simulated replications, at the
amplitude the profile was estimated to have (`Condition = 1`) and at
fractions of it. Compare coverage against the nominal level using the
Wilson interval and the Bradley band, which is what the verdict does for
you.
- **The amplitude ladder** (`Condition` column) matters because your
estimated amplitude is biased upward — the population that generated
your data plausibly has *less* differentiation than your estimate. The
rungs below 1 show what happens to coverage in that direction; the 0
rung shows the degenerate flat-profile case, where displacement has no
true value at all and the amplitude interval cannot cover the boundary
truth (its printed zero coverage is structural, flagged in the table).
When a profile's estimated amplitude is smaller than half its own CI
width — as for obsessive–compulsive PD here — the diagnostic adds one
more rung at a scaling factor above 1, the amplitude at which the
population would sit exactly at the observed CI half-width.
- **The guardrail table** reports how often profiles were *certified*
(amplitude CI lower bound at least 0.35 CI-widths above zero) at each
rung. At the 0 rung, any certification is a false certification: the
summary prints a caution when the false-certification rate is materially
above the rate a user would expect from the interval level. This is a
measured property of the shipped display rule, not a hypothesis test.
- **The verdict** classifies elevation, amplitude, and
certification-conditional displacement coverage at your as-estimated
amplitude, and the plain-language paragraph states the overall
conclusion per profile.
For a visual summary across the ladder:
``` r
plot(acc)
```
The two profiles tell usefully different stories. Paranoid PD is
certified: its amplitude CI lower bound clears the 0.35-CI-width margin, so
`ssm_analyze()` reports its displacement, and nothing in its coverage table
clearly leaves the Bradley band at the as-estimated amplitude or with the
population amplitude halved. Obsessive–compulsive PD is *not* certified —
its amplitude (about .02 here) is far too close to zero relative to its CI
width — so the printed output already withholds its displacement as
uninterpretable, before any coverage question is asked. On top of that, the
diagnostic shows its amplitude and displacement CIs under-cover badly at
this sample size (missing below the truth), so its verdict is a caution for
a genuine reason: the intervals themselves are unreliable, not merely the
point on the circle.
The guardrail line confirms the rule is doing its job. At this
configuration ($n = 250$, eight octant correlations), a profile whose true
amplitude is exactly zero would be certified only about the benchmark rate
— roughly the one-sided error a user reading the guardrail would expect —
so no caution fires from the guardrail itself. This is the scale-free
rule's payoff: unlike a fixed amplitude-unit cutoff, it holds
false-certification near its intended rate even in the near-zero regime
Zimmermann and Wright flagged, where a sample amplitude large enough to
look non-zero is otherwise the *expected* outcome from a flat population.
One reading note: at reduced `reps`, classes tend to print as
`borderline` — the Wilson interval around the estimated coverage is too
wide to place it clearly inside or outside the Bradley band. That is the
diagnostic being honest about its own Monte Carlo error; the default
`reps = 1000` sharpens such classifications into `adequate` or
`inadequate`.
Two settings worth knowing: `structure = "observed"` rebuilds the
population from the pooled observed correlations instead of the CPM — if
the two structures yield different verdicts, that disagreement is itself
informative (structure uncertainty is material for your data); and the
embedded CPM fit is returned as `acc$cpm` for inspection with the tools
from Section 2.
The diagnostic itself was validated against Zimmermann and Wright's
published results: configured to their transcribed simulation conditions,
it reproduces their accuracy classifications — coverage inside the Bradley
band at conditions they found accurate, and the same one-sided
under-coverage at conditions they found inaccurate.
### When to trust SSM parameters
Putting Sections 2 and 3 together into a checklist:
1. **Structure first.** Fit `cpm_fit()` to your instrument in your sample.
If the octant ordering fails or communalities are very low, stop; SSM
positions mean little.
2. **Elevation is the robust parameter.** Essentially unbiased, and its
intervals were accurate from $n \ge 50$ in every published and
package-run condition.
(On the correlation path, elevation is also the parameter ipsatizing
destroys — see Section 4.)
3. **Treat amplitude as optimistic.** It cannot go below zero, so it
overshoots when true differentiation is weak. Before interpreting a
"marked" amplitude at modest $n$, run `ssm_ci_accuracy()` and look at
the sub-1 ladder rungs.
4. **Only interpret displacement when amplitude is credible.** The package
already withholds the displacement interval when the amplitude CI lower
bound sits less than 0.35 CI-widths above zero; the diagnostic tells you
how reliable that certification is at your configuration.
5. **Do not build claims on the fit parameter's interval.** Zimmermann and
Wright found bootstrap $R^2$ intervals accurate only when population
fit was mediocre — near-perfect prototypicality cannot be bracketed
from below. Report $R^2$ descriptively.
6. **Remember what the diagnostic conditions on.** It asks "would my CI
procedure work in a population *like my estimates*", under multivariate
normality, with complete data. It is a strong check, not a certificate.
## 4. Does the instrument have circumplex structure at all?
Section 2's confirmatory model (`cpm_fit()`) fits one theory-driven circular
model and asks how well it fits. A complementary, more exploratory question
is whether the scales' correlations show circumplex structure *at all*,
without committing to the theoretical angles: are the scales spread evenly
around a circle, with comparable communalities, rather than clustering into
a small number of independent clusters (simple structure)? Acton and Revelle
(2004) evaluated ten such criteria by simulation; `fit_structure()`
implements four of them — Fisher, Gap, VT2, and Rotation — leaving out a
variance-test variant A&R found ineffective (VT1) and a criterion so highly
correlated with Rotation as to be redundant (MT, r = .99 with RT in their
Table 1). A fifth test, RANDALL, is not one of A&R's ten at all: it is an
independent order-correspondence test (Hubert & Arabie, 1987; Tracey, 1997)
that A&R excluded from their own simulation because, unlike their criteria,
its null distribution is known analytically rather than needing simulated
cutoffs (their footnote 3).
### The five tests
`fit_structure()` extracts the first two unrotated principal-axis factors
of the scales' correlation matrix (Acton & Revelle, 2004, p. 13) and
computes four criteria from that two-factor solution, plus a fifth test
that works directly on the correlations:
- **Fisher Test** (equal axes). Are the scales' communalities on the
two-factor solution comparable, rather than one axis dominating? The
statistic is the coefficient of variation of the scales' vector lengths
$\sqrt{h^2}$.
- **Gap Test** (equal spacing). Are the scales evenly distributed in angle
around the circle, rather than bunched together? The statistic is the
variance of the angular gaps between angularly adjacent scales, including
the gap that wraps from the last scale back around to the first.
- **Variance Test (VT2)** and **Rotation Test** (interstitiality). Both ask
whether the scales sit *between* a small number of dominant axes rather
than on them — the signature of a genuine circumplex, as opposed to simple
structure. Each takes a criterion computed at many rotations of the
two-factor solution and summarizes it as a coefficient of variation across
rotations; a true circumplex is indifferent to rotation, so both criteria
stay low.
- **RANDALL** (Hubert & Arabie, 1987; Tracey, 1997). Does the *hypothesized
circular order* you supplied (the order of `scales`) match the observed
correlations, in the sense that closer-together scales correlate more
strongly? Unlike the other four, this is a genuine randomization test: its
null distribution (scales randomly relabeled onto the hypothesized
positions) is enumerated exactly for up to nine scales, or estimated by
Monte Carlo relabeling for more, so it returns an exact or Monte Carlo
p value rather than a simulated cutoff.
The four factor-analytic criteria (Fisher, Gap, VT2, Rotation) have the most
power to detect simple structure when there is no large general factor
across the scales. *Deviation scoring* — centering each respondent on their
own mean across the selected scales, exactly what `ipsatize()` does —
approximates removing that general factor (Acton & Revelle, 2004, p. 9), so
it is `fit_structure()`'s default; pass `scoring = "raw"` to analyze the
scores as given. Each of the two scorings carries its own set of cutoffs
(next section), matched automatically.
``` r
res <- fit_structure(jz2017, scales = PANO())
res
#>
#> Circumplex Structure Tests (Acton & Revelle, 2004)
#> Scales (nv): 8
#> Scoring: deviation (row-mean centered)
#>
#> # Exploratory criteria
#>
#> Test Statistic Interpretation
#> Fisher 0.102 equal axes: at least 3x as likely as the alternative
#> Gap 0.152 equal spacing: at least 3x as likely as the alternative
#> Variance 0.180 interstitiality: almost certain
#> Rotation 0.325 interstitiality: at least 3x as likely as the alternative
#>
#> # Order hypothesis (RANDALL)
#>
#> Correspondence index = 0.868, p = 0.000397 (exact, 5040 relabelings)
#>
#> Interpretations are heuristic likelihood classifications from simulation, not
#> significance tests (Acton & Revelle, 2004). RANDALL's p-value is exact.
```
``` r
summary(res)
#>
#> Circumplex Structure Tests (Acton & Revelle, 2004)
#> Scales (nv): 8
#> Scoring: deviation (row-mean centered)
#> Ridge: 0
#>
#> # Exploratory criteria
#>
#> Test Statistic Almost Thrice Twice Verdict
#> Fisher 0.102 0.07 0.12 0.15 3x+ likely
#> Gap 0.152 0.15 0.40 0.46 3x+ likely
#> Variance 0.180 0.19 0.59 0.64 almost certain
#> Rotation 0.325 0.32 0.64 0.67 3x+ likely
#>
#> # Estimated scale geometry
#>
#> Scale Angle Communality
#> PA 339.635 0.642
#> BC 359.858 0.571
#> DE 48.584 0.500
#> FG 81.337 0.522
#> HI 161.662 0.690
#> JK 183.339 0.713
#> LM 215.503 0.388
#> NO 287.119 0.474
#>
#> # Order hypothesis (RANDALL)
#>
#> Correspondence index = 0.868, p = 0.000397 (exact, 5040 relabelings)
#>
#> Interpretations are heuristic likelihood classifications from simulation, not
#> significance tests (Acton & Revelle, 2004). RANDALL's p-value is exact.
```
`summary()` adds the numeric cutoffs behind each classification and the
estimated angle and communality of every scale on the two-factor solution —
the same geometry the plot below draws. A clean circumplex shows scales at
roughly the theoretical octant spacing with broadly similar communalities;
`Fisher` measures departures in communality, `Gap`/`VT2`/`Rotation` measure
departures in even angular spread.
``` r
plot(res)
```
For the IIP-SC octants the picture agrees with Section 2's CPM fit: the
scales keep their theoretical circular *ordering* with comparable
communalities and roughly even spacing, and the Fisher, Gap, and
interstitiality criteria all classify the configuration as consistent with
(or close to) a circumplex under deviation scoring. One caveat when reading
the angles: the two-factor solution is *unrotated*, so its absolute
orientation is arbitrary — it is the scales' ordering and relative spacing
that agree with theory, not their absolute angles (which is why `summary()`
places PA well away from its nominal 90°). RANDALL's p value confirms the
hypothesized circular order directly, independent of the factor-analytic
criteria.
### Where the cutoffs come from
The four factor-analytic criteria are only as good as the thresholds used
to classify them, and this is where the published article cannot be used
as-is. Acton and Revelle calibrated their cutoffs by simulation at 64 and
128 variables and reported (their p. 18) that the Gap Test's cutoffs shift
sharply with the number of scales — far too sharply to reuse at the eight
scales a typical circumplex instrument has. This package's development
process re-derived every cutoff at $n_v = 8$ under Acton and Revelle's own
generating model (their Eqs. 11.1–11.3), in a script committed at
`data-raw/structure-test-cutoffs.R`: it first reproduced their published
64/128-variable design as a sanity gate on the simulation machinery, then
reran the same design at $n_v = 8$ to derive the constants `fit_structure()`
actually uses. The nv effect the re-derivation found is large — the raw-
scored Gap Test's "almost certain" cutoff moves from .01 at $n_v = 64/128$
to .35 at $n_v = 8$ — which is exactly why `fit_structure()` refuses to
interpret any scale count it has not calibrated: at any $n_v \ne 8$ the
statistics are still reported, but the classification column prints a dash
rather than guessing.
RANDALL needs no such calibration; its p value comes from enumerating (or
Monte Carlo sampling) the randomization null directly on your data, so it
is available at any scale count of four or more.
### Reading the classifications
The classification categories describe where a statistic falls among the
simulated distributions for competing structures (circumplex vs. simple
structure): "almost certain" (below the 1st percentile of the competing
distribution), "at least 3x/2x as likely as the alternative" (the criterion's
structure is at least three/two times as likely as its competitor at that
statistic value), and "not clearly supported". Those ratios are *likelihood
ratios*, not posterior probabilities or p values. They are **heuristic
classifications read off simulated distributions, not significance tests**,
and `fit_structure()`'s `print()`/`summary()` output repeats that caveat
every time an interpretation is shown. Treat a "weak"
or unsupported classification as a caution to inspect the loading
configuration (the plot above) and Section 2's CPM fit together, not as a
rejection of any specific hypothesis.
### How this complements the CPM fit
`cpm_fit()` (Section 2) and `fit_structure()` ask related but different
questions. `cpm_fit()` commits to the theoretical angles and communality
model and tests goodness of fit against that specific structure; a good
overall RMSEA/CFI can still hide unequal spacing or a dominant general
factor, both of which `fit_structure()` is built to detect. Conversely,
`fit_structure()`'s exploratory criteria say nothing about how well the
scales match *your* theoretical angles — only RANDALL, via the order you
supply, references a hypothesis at all, and even that is an order
hypothesis, not the specific angles `cpm_fit()` estimates. Running both is
more informative than either alone: agreement between a good CPM fit and
circumplex-supporting `fit_structure()` classifications is stronger evidence
than either result on its own, and disagreement (for example, adequate CPM
fit alongside a Fisher Test flagging unequal axes) points to exactly which
aspect of circumplex structure to examine further.
## 5. Ipsatization and what it costs
Ipsatizing — subtracting each respondent's own mean across the octant
scales from each of their scale scores (`ipsatize()`) — is a common
preprocessing step in circumplex work, used to remove individual
differences in overall endorsement before examining profile shape.
For SSM analyses its main cost is simple to state: **ipsatizing discards
elevation.** After row-centering, every respondent's octant scores sum to
zero, so a group's mean profile has zero mean by construction, and the
covariances between the ipsatized scales and any external measure sum to
exactly zero — forcing the mean correlation toward zero regardless of how
strongly the construct relates to the instrument's general factor
(Zimmermann & Wright, 2017, p. 4).
``` r
set.seed(45678)
res_raw <- ssm_analyze(
jz2017,
scales = PANO(),
angles = octants(),
measures = "PARPD",
boots = 100
)
jz_ips <- ipsatize(jz2017, items = PANO())
set.seed(45678)
res_ips <- ssm_analyze(
jz_ips,
scales = paste0(PANO(), "_i"),
angles = octants(),
measures = "PARPD",
boots = 100
)
comparison <- rbind(
res_raw$results[, c("e_est", "a_est", "d_est")],
res_ips$results[, c("e_est", "a_est", "d_est")]
)
rownames(comparison) <- c("raw", "ipsatized")
round(comparison, 3)
#> e_est a_est d_est
#> raw 0.250 0.150 128.945
#> ipsatized 0.007 0.113 132.949
```
The raw-score elevation — which matches the value Zimmermann and Wright
report for this scale in their Table 4 — collapses to near zero after
ipsatizing. The displacement stays close to its raw-score value, and the
amplitude stays broadly similar — though not identical, because ipsatizing
also changes the scales' variances and intercorrelations, so shape
parameters shift somewhat too. Guidance:
- If elevation carries meaning in your application (for interpersonal
problems it indexes association with generalized interpersonal
distress), analyze raw scores and let the SSM separate elevation from
shape — that separation is exactly what the model is for.
- If you receive data that were already ipsatized, do not interpret the
elevation row, and say so in the write-up; amplitude and displacement
remain interpretable.
- Do not describe an ipsatized profile's near-zero elevation as evidence
that a construct is "not generally interpersonal" — the preprocessing
made that value uninformative.
## 6. Wrap-up
Evaluating circumplex structure has two layers: whether the instrument
behaves like a circumplex in your sample (`cpm_fit()`), and whether the
inferential machinery of the SSM can be trusted at your sample size and
profile (`ssm_ci_accuracy()`). Both are one function call, and both change
what you should claim more often than users expect — amplitude and
displacement intervals, in particular, earn their trust only when the
profile is genuinely differentiated relative to the precision your sample
size affords.
## References
* Acton, G. S., & Revelle, W. (2004). Evaluation of ten psychometric
criteria for circumplex structure. _Methods of Psychological Research
Online, 9_(1), 1–27.
* Bradley, J. V. (1978). Robustness? _British Journal of Mathematical and
Statistical Psychology, 31_(2), 144–152.
* Browne, M. W. (1992). Circumplex models for correlation matrices.
_Psychometrika, 57_(4), 469–497.
* Browne, M. W., & Cudeck, R. (1993). Alternative ways of assessing model
fit. In K. A. Bollen & J. S. Long (Eds.), _Testing structural equation
models_ (pp. 136–162). Newbury Park, CA: Sage.
* Grassi, M., Luccio, R., & Di Blas, L. (2010). CircE: An R implementation
of Browne's circular stochastic process model. _Behavior Research
Methods, 42_(1), 55–73.
* Gurtman, M. B., & Pincus, A. L. (2000). Interpersonal Adjective Scales:
Confirmation of circumplex structure from multiple perspectives.
_Personality and Social Psychology Bulletin, 26_(3), 374–384.
* Hu, L., & Bentler, P. M. (1999). Cutoff criteria for fit indexes in
covariance structure analysis: Conventional criteria versus new
alternatives. _Structural Equation Modeling, 6_(1), 1–55.
* Hubert, L., & Arabie, P. (1987). Evaluating order hypotheses within
proximity matrices. _Psychological Bulletin, 102_(1), 172–178.
* Tracey, T. J. G. (1997). RANDALL: A Microsoft FORTRAN program for a
randomization test of hypothesized order relations. _Educational and
Psychological Measurement, 57_(1), 164–168.
* Zimmermann, J., & Wright, A. G. C. (2017). Beyond description in
interpersonal construct validation: Methodological advances in the
circumplex Structural Summary Approach. _Assessment, 24_(1), 3–23.