nomo_power_rmsea() gives the power of MacCallum, Browne, and Sugawara's
(1996) RMSEA-based tests of fit at given sample sizes, or the smallest sample
size that reaches a target power.
Usage
nomo_power_rmsea(
model = NULL,
n = NULL,
df = NULL,
power = 0.8,
test = c("close", "not_close", "exact"),
rmsea_null = NULL,
rmsea_alt = NULL,
alpha = 0.05
)Arguments
- model
The planned model: a lavaan model string, a
nomo_model()object, a fittednomo_cfa()result, or alavaanfit. Its degrees of freedom are used. Givemodelordf, not both. A multi-group fit is refused (see Details).- n
Optional sample sizes at which to compute power.
- df
The model's degrees of freedom, instead of
model.- power
Target power for the smallest sample size. Default
.80.- test
"close"(default),"not_close", or"exact".- rmsea_null, rmsea_alt
Optional null and alternative RMSEA values, replacing the test's defaults. The null is 0 for
"exact"and above 0 for the other tests.- alpha
Significance level. Default
.05.
Value
A nomo_power object. The fields to read are test, df,
rmsea_null, rmsea_alt, alpha, target_power, power (a table of
sample sizes and their power), and n_required (the smallest sample size
reaching target_power, or NA if none up to one million does).
nomo_table() returns the "power" table, its one type. print()
shows the test, its RMSEA values, and the power at each sample size;
summary() shows the same. plot() draws the power curve with the
target power and the smallest sample size that reaches it.
Other fields record the kind of power analysis. They may change between
releases and are not part of the stable interface (see ?nomologR).
Details
MacCallum et al. (1996) framed the power of a covariance structure model's overall test in terms of RMSEA, whose noncentral chi-square distribution has noncentrality \((N - 1) \, df \, \varepsilon^2\). Three tests are available:
"close": the test of close fit, with null RMSEA .05 and alternative .08 by default. Power is the chance of rejecting close fit when the fit is in fact mediocre."not_close": the test of not-close fit, with null .05 and alternative .01. Power is the chance of rejecting not-close fit when the fit is in fact close, which is what supports a claim of good fit."exact": the test of exact fit, with null 0 and alternative .05.
Each test is named by its null hypothesis, so a rmsea_null that
contradicts it is refused: the test of exact fit has a null of 0, and the
tests of close and not-close fit a null above 0.
Power depends on the degrees of freedom: a model with few of them needs a
large sample to test its fit. This is power for the model's overall fit test,
not for any one parameter; nomo_power_simulate() gives that.
The calculation is for a single-group model. A multi-group lavaan fit is
refused, because lavaan's RMSEA for several groups carries a factor of the
square root of their number, and the noncentrality above would overstate
the power for the RMSEA lavaan reports.
References
MacCallum, R. C., Browne, M. W., & Sugawara, H. M. (1996). Power analysis and determination of sample size for covariance structure modeling. Psychological Methods, 1(2), 130-149. doi:10.1037/1082-989X.1.2.130
See also
nomo_power_simulate() for the power to detect particular
parameters.
Examples
model <- nomo_model(list(A = paste0("a", 1:4), B = paste0("b", 1:4)))
nomo_power_rmsea(model)
#> <nomo_power> Power of the test of close fit
#> MacCallum, Browne, and Sugawara (1996).
#> df: 19 | RMSEA: null 0.050, alternative 0.080 | alpha = .05
#> Smallest N for power .80: 453
#>
#> Power by sample size
#> N Power
#> 453 .801
#>
#> RMSEA = root mean square error of approximation; df = degrees of freedom;
#> N = sample size.
#>
#> This is power for the overall fit test, not for any one parameter;
#> nomo_power_simulate() gives that.
#>
#> See nomo_table(x, "power") for the power at each sample size.
nomo_power_rmsea(model, n = c(100, 200, 400), test = "not_close")
#> <nomo_power> Power of the test of not-close fit
#> MacCallum, Browne, and Sugawara (1996).
#> df: 19 | RMSEA: null 0.050, alternative 0.010 | alpha = .05
#> Smallest N for power .80: 490
#>
#> Power by sample size
#> N Power
#> 100 .14
#> 200 .30
#> 400 .67
#>
#> RMSEA = root mean square error of approximation; df = degrees of freedom;
#> N = sample size.
#>
#> This is power for the overall fit test, not for any one parameter;
#> nomo_power_simulate() gives that.
#>
#> See nomo_table(x, "power") for the power at each sample size.