nomo_power_simulate() estimates, for each sample size, how often a planned
model recovers its parameters: whether it converges, whether the solution is
proper, how biased the estimates and their standard errors are, how often the
intervals cover the population values, and how often each parameter is
detected. This is Muthén and Muthén's (2002) Monte Carlo approach.
Usage
nomo_power_simulate(
population,
n,
reps = 500,
analysis = NULL,
focus = NULL,
alpha = 0.05,
seed = NULL,
standardized = TRUE
)Arguments
- population
A lavaan model with population values for its parameters.
- n
Sample sizes to simulate.
- reps
Replications per sample size. Default 500; Muthén and Muthén used 10,000, and more replications give steadier estimates.
- analysis
Optional analysis model. By default,
populationwith its numeric values removed.- focus
Optional parameters whose power is of interest, written as in lavaan (
"A~~B","B~A","A=~a1"). By default, every parameter with a nonzero population value.- alpha
Significance level for detection. Default
.05.- seed
Optional integer seed, for a reproducible simulation. The session's random-number state is restored afterwards. A seed reproduces the same samples only under the same version of lavaan:
lavaan::simulateData()changed its default generator in lavaan 0.7-3, so the same seed draws different samples, and gives slightly different estimates of power, bias, and coverage, before and after that version.- standardized
Passed to
lavaan::simulateData(). DefaultTRUE, which gives the observed variables unit variance where the population values allow it (see Details).
Value
A nomo_power object. The fields to read are parameters (for
each sample size and parameter: population value, mean estimate, relative
bias, standard deviation of the estimates, mean standard error, standard
error bias, coverage, and power), summary (for each sample size: the
proportions that converged and that were improper, the smallest power in
focus, the largest absolute biases, the coverage range, and
meets_references), n_required (the smallest simulated sample size that
meets the references, or NA), focus, reps, alpha, and seed (the
seed given, or NA without one). nomo_table() returns the "summary"
table, its default type, or the "parameters" table. print() shows
the table by sample size with a key to its columns and the references;
summary() adds the table by sample size and parameter. plot() draws
the power of each focus parameter by sample size.
Other fields record the call, the population and analysis models, and the
kind of power analysis. They may change between releases and are not part
of the stable interface (see ?nomologR).
Details
Experimental. This function is experimental and remains so after
nomologR 1.0.0. The metric of its estimates and its summaries may be refined
during 1.x, without a deprecation period, as it is extended beyond complete
continuous data. Any change will be described in NEWS; see the package help
page, ?nomologR, for the stability policy.
Data are generated from population, a lavaan model whose parameters carry
their population values (for example A =~ 0.7*a1), with
lavaan::simulateData(). With standardized = TRUE, the default, the
residual variances of the observed variables are set so that each has unit
variance. Values that explain more than a variable's whole variance leave no
residual variance to set, and lavaan then generates the data in another
metric; a warning names the variances population implies when any is not
The
analysismodel, which by default ispopulationwithout its values, is fitted to each data set withlavaan::sem(std.lv = TRUE).
The metric. Each estimate is the analysis model's unstandardized
estimate, and it is compared with the population value as written. Both are
in the metric that std.lv = TRUE sets: an exogenous factor has variance 1,
and an endogenous factor has residual variance 1. With the default
standardized = TRUE, loadings on exogenous factors are therefore
standardized, and covariances between exogenous factors are correlations. A
latent regression is not standardized. In B ~ 0.4*A, B's total variance
is 1.16, its residual variance of 1 plus 0.16 from A, so the completely
standardized coefficient is 0.4 divided by the square root of 1.16, or .37,
and B's indicators have standardized loadings 1.08 times their values. Bias,
coverage, and power are for the values as written.
Muthén and Muthén (2002) suggest choosing the sample size at which three conditions hold, and power for the parameter of interest is close to .80:
parameter and standard error biases are within 10% for every parameter given a population value;
the standard error bias of the parameter whose power is assessed is within 5%;
coverage of the 95% interval lies between .91 and .98.
meets_references records whether they hold at each sample size, for the
parameters in focus. Only parameters given a population value in
population (loadings, regressions, and covariances) are checked, so
residual and factor variances are outside the check.
Wolf, Harrington, Clark, and Miller (2013) showed that the sample size a model needs varies widely with its structure, so rules of thumb such as a fixed N or a ratio of cases to parameters are no substitute. They also counted improper solutions, which are reported here. Summaries use the replications that converged.
References
Muthén, L. K., & Muthén, B. O. (2002). How to use a Monte Carlo study to decide on sample size and determine power. Structural Equation Modeling, 9(4), 599-620. doi:10.1207/S15328007SEM0904_8
Wolf, E. J., Harrington, K. M., Clark, S. L., & Miller, M. W. (2013). Sample size requirements for structural equation models: An evaluation of power, bias, and solution propriety. Educational and Psychological Measurement, 73(6), 913-934. doi:10.1177/0013164413495237
See also
nomo_power_rmsea() for the power of the overall fit tests.
Examples
population <- "
A =~ 0.7*a1 + 0.7*a2 + 0.6*a3 + 0.5*a4
B =~ 0.7*b1 + 0.6*b2 + 0.6*b3 + 0.5*b4
A ~~ 0.3*B
"
# Five replications run quickly and show the result; they are far too few to
# estimate power.
quick <- nomo_power_simulate(population, n = 100, reps = 5,
focus = "A~~B", seed = 2026)
nomo_table(quick, "summary")
#> # A tibble: 1 × 9
#> n converged improper min_power max_abs_bias max_abs_se_bias min_coverage
#> <int> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
#> 1 100 1 0 0.8 0.214 1.32 0.6
#> # ℹ 2 more variables: max_coverage <dbl>, meets_references <lgl>
# \donttest{
pw <- nomo_power_simulate(population, n = c(100, 200), reps = 100,
focus = "A~~B", seed = 2026)
pw
#> <nomo_power> Monte Carlo power and sample size
#> Muthén and Muthén (2002).
#> Replications: 100 per N | alpha = .05 | Focus: A~~B
#>
#> By sample size
#> N Meets Converged Improper Min power Max bias Max SE bias Coverage
#> 100 no 100.0% 0.0% .72 6.9% 10.7% .92 to .97
#> 200 no 100.0% 0.0% .84 7.0% 12.2% .90 to .97
#> No simulated N meets the references; try larger ones.
#>
#> What these columns mean
#> N -- Sample size.
#> Meets -- Whether the references below hold at this N.
#> Converged, Improper -- Share of the replications that converged, and of
#> those, the share with an improper solution.
#> Min power -- Lowest power among the focus parameters.
#> Max bias, Max SE bias -- Largest absolute relative bias of an estimate, and
#> of its standard error (SE), across the parameters given a population
#> value.
#> Coverage -- Lowest to highest proportion of 95% confidence intervals that
#> contain the population value, across those parameters.
#>
#> References: parameter and SE bias within 10% for every parameter given a
#> population value, SE bias within 5% for the focus parameters, coverage
#> .91 to .98, and power .80 for the focus parameters. They are guides for
#> choosing N, not rules.
#>
#> See summary(x) for every parameter at every N and nomo_table(x, "summary") for
#> the table by sample size.
# }