---
title: "Using `mlmpower` Package to Conduct Multilevel Power Analysis"
output: rmarkdown::html_vignette
author: Brian T. Keller
vignette: >
%\VignetteIndexEntry{Using `mlmpower` Package to Conduct Multilevel Power Analysis}
%\VignetteEncoding{UTF-8}
%\VignetteEngine{knitr::rmarkdown}
editor_options:
markdown:
wrap: 72
---
```{r, include = FALSE}
knitr::opts_chunk$set(
collapse = TRUE,
comment = "#>"
)
# Load Library
library(mlmpower)
# Set seed
set.seed(981723)
# Load Cache
powersim1 <- readRDS('powersim1.rds')
powersim2 <- readRDS('powersim2.rds')
powersim3 <- readRDS('powersim3.rds')
```
## Illustration 1: Cross-Sectional Power Analysis
The first illustration demonstrates a power analysis for a
cross-sectional application of the following multilevel model. To
provide a substantive context, consider a prototypical education example
where students are nested in schools. Intraclass correlations for
achievement-related outcomes often range between .10 and .25 (Hedges &
Hedberg, 2007; Hedges & Hedberg, 2013; Sellstrom & Bremberg, 2006;
Spybrook et al., 2011; Stockford, 2009). To accommodate uncertainty
about this important parameter, the power simulations investigate
intraclass correlation values of .10 and .25.
The multilevel model for the illustration is
$$
\begin{split}
Y_{ij} &= \left( \beta_{0} + b_{0j} \right) + \left( \beta_{1} + b_{1j} \right)\left( X_{1ij} - {\overline{X}}_{1j} \right) + \beta_{2}\left( X_{2ij} - {\overline{X}}_{2} \right) \\
&\phantom{=}+ \ \beta_{3}\left( Z_{1j} - {\overline{Z}}_{1} \right) + \ \beta_{4}\left( Z_{2j} - {\overline{Z}}_{2} \right) + \beta_{5}\left( X_{1ij} - {\overline{X}}_{1j} \right)\left( Z_{1j} - {\overline{Z}}_{1} \right) + \varepsilon_{ij}\\[1em]
\mathbf{b}_{j}\ &\sim\ N\left( 0,\mathbf{\Sigma}_{\mathbf{b}} \right)\ \ \ \ \ \mathbf{\Sigma}_{\mathbf{b}}=\begin{pmatrix}
\sigma_{b_{0}}^{2} & \\
\sigma_{b_{1},b_{0}} & \sigma_{b_{1}}^{2} \\
\end{pmatrix}\mathbf{\ \ \ \ \ }\varepsilon_{ij}\ \sim\ N(0,\sigma_{\varepsilon}^{2})
\end{split}
$$
where $Y_{ij}$ is the outcome score for observation *i* in cluster *j*,
$X_{1ij}$ is a focal within-cluster predictor, ${\overline{X}}_{1j}$ is
the variable's level-2 cluster mean, $X_{2ij}$ is a grand mean centered
level-1 covariate, $Z_{1j}$ is a level-2 moderator score for cluster
*j*, and $Z_{2j}$ is a level-2 covariate. Turning to the residual terms,
$b_{0j}$ and $b_{1j}$ are between-cluster random effects that capture
residual variation in the cluster-specific intercepts and slopes, and
$\varepsilon_{ij}$ is a within-cluster residual. By assumption, the
random effects follow a multivariate normal distribution with a
between-cluster covariance matrix $\mathbf{\Sigma}_{\mathbf{b}}$, and
within-cluster residuals are normal with constant variance
$\sigma_{\varepsilon}^{2}$.
Importantly, group mean centering $X_{1}$ yields a pure within-cluster
predictor, whereas grand mean centering $X_{2}$ gives a predictor with
two sources of variation. Although it need not be true in practice, the
within- and between-cluster parts of any level-1 predictor variables
with non-zero intraclass correlations share a common slope coefficient.
To simplify notation, Equation 1 can be rewritten as
$$Y_{ij} = \left( \beta_{0} + b_{0j} \right) + \left( \beta_{1} + b_{1j} \right)X_{1ij}^{w} + \beta_{2}\left( X_{2ij}^{w} + X_{2j}^{b} \right) + \beta_{3}Z_{1j}^{b} + \beta_{4}Z_{2j}^{b} + \beta_{5}X_{1ij}^{w}Z_{1j}^{b} + \varepsilon_{ij}$$
where the *w* and *b* superscripts reference within- and between-cluster
deviation scores, respectively.
The key inputs to the model object are the unconditional intraclass
correlation and effect size values. This illustration investigates power
for intraclass correlation values of .10 and .25. The within- and
between-cluster fixed effects are set to $R_{w}^{2}$ = .065 and
$R_{b}^{2}$ = .065, respectively, the sum of which corresponds to
Cohen's (1988) medium effect size benchmark. Note that $R_{b}^{2}$
cannot exceed the outcome's intraclass correlation, as this would imply
that between-cluster predictors explain more than 100% of the available
variance. Following conventional wisdom that interaction effects tend to
produce small effects (Aguinis, Beaty, Boik, & Pierce, 2005; Chaplin,
1991), the cross-level product is assigned $R_{p}^{2}$ = .01. Finally,
the random coefficient effect size is set to $R_{rc}^{2}$ = .03 based on
values from Table 2 in Enders, Keller, and Woller (2023).
By default, `mlmpower` assigns equal weights to all quantities
contributing to a particular source of variance. To assign non-uniform
weights, assume $X_{1}$ and $Z_{1}$ (the interacting variables) are the
focal predictors and $X_{2}$ and $Z_{2}$ are covariates. To mimic a
situation where the covariates explain a small amount of variation, the
fixed effect $R^{2}$ values are predominantly allocated to the focal
predictors using weights of .80 and .20. A small covariate allocation
could be appropriate for a set of background or sociodemographic
characteristics that weakly predict the outcome. Researchers often need
power estimates for individual partial regression slopes. Although the
weights do not exactly carve $R_{w}^{2}$ and $R_{b}^{2}$ into additive
components when predictors are correlated, adopting weak associations
among the regressors allows us to infer that each focal predictor
roughly accounts for 5% of the explained variation at each level (i.e.,
$.80 \times R_{w}^{2}$ or $R_{b}^{2} \approx .05$).
The following code block shows the `mlmpower` model object for the
example.
```{r}
example1 <- (
effect_size(
icc = c(.10, .25),
within = .065,
between = .065,
product = .01,
random_slope = .03
)
+ outcome('y', mean = 50, sd = 10)
+ within_predictor('x1', icc = 0, weight = .80)
+ within_predictor('x2', weight = .20)
+ between_predictor('z1', weight = .80)
+ between_predictor('z2', weight = .20)
+ product('x1','z1', weight = 1)
+ random_slope('x1', weight = 1)
)
```
The `mlmpower` package groups $R^{2}$ values and intraclass correlations
into a single object called `effect_size()`, with the five inputs
separated by commas. The within argument corresponds to $R_{w}^{2}$, the
between argument aligns with $R_{b}^{2}$, the product argument specifies
$R_{p}^{2}$, and the random_slope argument corresponds to $R_{rc}^{2}$.
The icc argument assigns a global intraclass correlation to all level-1
variables, and separate simulations are performed for each requested
level.
Variable attributes are referenced by adding the following objects:
`outcome()`, `within_predictor()`, `between_predictor()`, `product()`,
and `random_slope()`. All five objects have an argument for the variable
name, weight (`weight =`), mean (`mean =`), and standard deviation
(`sd =`). Level-1 variables additionally have an intraclass correlation
argument (`icc =`) that supersedes the global setting in
`effect_size()`.
The previous code block assigns explicit weights to all variables
contributing to a given effect. The unit weights in the `product()` and
`random_slope()` objects result from a single variable determining those
sources of variation. Next, the
`within_predictor('x1', icc = 0, weight = .80)` object overrides the
global intraclass correlation setting, defining $X_{1}$ as a pure
within-cluster deviation variable with no between-cluster variation.
Finally, with the exception of the outcome variable, which has a mean
and standard deviation of 50 and 10, the script accepts the default
settings for the means and variances of the predictors (0 and 1,
respectively).
The multilevel model parameters also require three sets of correlations:
within-cluster correlations among level-1 predictors, between-cluster
correlations among level-2 predictors, and random effect correlations.
These correlations are optional, but they can specified using the
`correlations()` object. The earlier code block omits this object,
thereby accepting the default specification. When the `correlations()`
object is omitted, the `mlmpower` package iteratively samples all
correlations from a uniform distribution between .10 and .30, such that
the resulting power estimates average over a distribution of possible
associations. This default range spans Cohen's (1988) small to medium
effect size benchmarks, and it brackets common correlation values from
published research (Bosco, Aguinis, Singh, Field, & Pierce, 2015; Funder
& Ozer, 2019; Gignac & Szodorai, 2016).
The default specifications could be invoked by explicitly appending the
`correlations()` object to the earlier script, as follows.
```{r}
example1 <- (
effect_size(
icc = c(.10, .25),
within = .065,
between = .065,
product = .01,
random_slope = .03
)
+ outcome('y', mean = 50, sd = 10)
+ within_predictor('x1', icc = 0, weight = .80)
+ within_predictor('x2', weight = .20)
+ between_predictor('z1', weight = .80)
+ between_predictor('z2', weight = .20)
+ product('x1','z1', weight = 1)
+ random_slope('x1', weight = 1)
+ correlations(
within = random(0.1, 0.3),
between = random(0.1, 0.3),
randeff = random(0.1, 0.3)
)
)
```
Researchers can modify the upper and lower limits of each correlation
range, or they can specify a constant correlation by inputting a scalar
value. For example, specifying `randeff = 0` would define
$\mathbf{\Sigma}_{\mathbf{b}}$ as a diagonal matrix. The illustrative
simulations presented in Enders et al. (2023) suggest that predictor and
random effect correlations tend not to matter very much.
It may be instructive to inspect the population parameters prior to
running the simulation. Executing `summary(example1)` returns tabular
summaries of the multilevel model parameters.
```{r}
summary(example1)
```
Having specified the target model, you next use the `power_analysis()`
function to conduct simulations. The function requires four inputs: the
model argument specifies the parameter value object (e.g., `example1`),
the replications input specifies the number of artificial data sets,
`n_between` is a vector of level-2 sample sizes, and `n_within` is a
vector of level-1 sample sizes (i.e., the number of observations per
cluster). The code block below pairs four level-2 sample size values
($J$ = 30, 60, 90, and 120) with three level-1 sample sizes ($n_{j}$ =
10, 20, or 30 observations per cluster), and it requests 2,000
artificial data sets for each combination.
```{r, eval = FALSE}
# Set seed for replicable results
set.seed(2318971)
# Run Power Analysis
powersim1 <-
power_analysis(
model = example1,
replications = 2000,
n_between = c(30, 60, 90, 120),
n_within = c(10, 20, 30)
)
```
The package uses `lme4` (Bates et al., 2021) for model fitting, and it
defaults to an alpha level of .05 for all significance tests.
Significance tests of random slope variation use a likelihood ratio test
with a mixture chi-square reference distribution (i.e., a chi-bar
distribution; Snijders & Bosker, 2012, p. 99), as implemented in the
`varTestnlme` package (Baey & Kuhn, 2022). Executing `summary(powersim1)`
returns the tabular summaries of the simulation results shown
```{r}
summary(powersim1)
```
## Illustration 2: Growth Curve Power Analysis
The second illustration demonstrates a power analysis for a longitudinal
growth curve model with a pair of cross-level interactions involving a
binary level-2 moderator. Intraclass correlations for longitudinal and
intensive repeated measures data often reach values of .40 or higher
(Arend & Schäfer, 2019; Bolger & Laurenceau, 2013; Singer & Willett,
2003). To accommodate uncertainty about this important parameter, the
simulation investigates intraclass correlation values of .40 and .60.
The multilevel model for the illustration is
$$
\begin{split}
Y_{ij} &= \left( \beta_{0} + b_{0j} \right) + \left( \beta_{1} + b_{1j} \right)X_{1ij}^{w} + \beta_{2}\left( X_{2ij}^{w} + X_{2j}^{b} \right) + \beta_{3}\left( X_{3ij}^{w} + X_{3j}^{b} \right)\\
&\phantom{=}+ \ \beta_{4}Z_{1j}^{b} + \ \beta_{5}Z_{2j}^{b} + \beta_{6}Z_{3j}^{b} + \beta_{7}X_{1ij}^{w}Z_{1j}^{b} + \varepsilon_{ij} \\
\mathbf{b}_{j}\ &\sim\ N\left( 0,\mathbf{\Sigma}_{\mathbf{b}} \right)\ \ \ \ \ \mathbf{\Sigma}_{\mathbf{b}}=\begin{pmatrix}
\sigma_{b_{0}}^{2} & \\
\sigma_{b_{1},b_{0}} & \sigma_{b_{1}}^{2} \\
\end{pmatrix}\mathbf{\ \ \ \ \ }\varepsilon_{ij}\ \sim\ N(0,\sigma_{\varepsilon}^{2})
\end{split}
$$
Following established notation (see Illustration 1), the *w* and *b*
superscripts reference within- and between-cluster deviation scores,
respectively. The explanatory variables include a time score predictor
with a random coefficient, a pair of time-varying covariates, a binary
level-2 moderator, a pair of level-2 covariates, and a cross-level
(group-by-time) interaction. For the purposes of weighting, we
designated $X_{1}$ and $Z_{1}$ (the interacting variables) as focal
predictors, $X_{2}$ and $X_{3}$ as level-1 (time-varying) covariates,
and $Z_{2}$ and $Z_{3}$ as level-2 covariates. To mimic the scaling of a
typical temporal index, we assume the major time increments are coded
$X_{1}^{w}$ = (0, 1, 2, 3, 4).
A brief discussion of $Z_{1}$ (the binary moderator) is warranted before
continuing. First, like other level-2 variables, this variable's
population mean must be 0 (i.e., $Z_{1}$ is centered in the population)
in order to maintain the orthogonality of the cross-level interaction
terms. Grand mean centering a level-2 binary predictor creates an
ANOVA-like contrast code, such that $\beta_{0}$ is the grand mean and
$\beta_{4}$ is the predicted group mean difference when $X_{1}^{w}$ (the
time score predictor) equals 0. In this case, a code of 0 corresponds to
the baseline assessment. Because $\beta_{4}$ is a conditional effect
that represents the group mean difference at the first occasion,
$Z_{1}$'s contribution to the between-cluster effect size depends on
whether we view this regressor as a naturally occurring classification
or an intervention assignment indicator. In the former case, we might
expect a group mean difference at baseline, and the presence of such a
difference would require a non-zero weight. In contrast, random
assignment to conditions would eliminate a group mean difference at
baseline, and $Z_{1}$'s weight would equal 0. Note that this conclusion
changes if the time scores are coded differently. For illustration
purposes, we assume $Z_{1}$ is an intervention assignment indicator.
The key inputs to the model object are the unconditional intraclass
correlation and effect size values. This illustration investigates power
for intraclass correlation values of .40 and .60. The within- and
between-cluster fixed effects are set to $R_{w}^{2}$ = .13 and
$R_{b}^{2}$ = .065, respectively; the former corresponds to Cohen's
(1988) medium effect size benchmark. Note that $R_{b}^{2}$ cannot exceed
the outcome's intraclass correlation, as this would imply that
between-cluster predictors explain more than 100% of the available
variance. Following conventional wisdom that interaction effects tend to
produce small effects (Aguinis et al., 2005; Chaplin, 1991), $R_{p}^{2}$
= .05 is assigned to the pair of cross-level product terms. Finally, the
random coefficient effect size is set to $R_{rc}^{2}$ = .03 based on
values from Table 2 in Enders et al. (2023).
To refresh, we designated $X_{1}$ and $Z_{1}$ (the interacting
variables) as focal predictors, $X_{2}$ and $X_{3}$ as level-1
(time-varying) covariates, and $Z_{2}$ and $Z_{3}$ as level-2
covariates. To mimic a situation where the linear change predominates
the level-1 model, we used weights of .50, .25, .25 to allocate the
within-cluster $R^{2}$ to $X_{1}$, $X_{2}$, and $X_{3}$. At level-2, we
used weights equal to 0, .50, .50 to allocate the between-cluster
$R^{2}$ to $Z_{1}$, $Z_{2}$, and $Z_{3}$. As noted previously, $Z_{1}$'s
slope represents the group mean difference at baseline, and we are
assuming that random assignment nullifies this effect. Finally,
$R_{p}^{2}$ and $R_{rc}^{2}$ do not require weights because a single
variable determines each source of variation.
To illustrate how to modify default correlation settings, we sampled
within-cluster predictor correlations between the range of .20 and .40
under the assumption that the time scores could have stronger
associations with other time-varying predictors. We similarly sampled
the random effect correlations between the range of .30 and .50 to mimic
a scenario where higher baseline scores (i.e., random intercepts) are
associated with higher (more positive) growth rates. Finally, we adopted
the default correlation range for the between-cluster predictors. The
simulations from the previous cross-sectional example suggest that the
correlations are somewhat arbitrary and would not have a material impact
on power estimates.
The following code block shows the `mlmpower` model object for this
example.
```{r}
example2 <- (
effect_size(
icc = c(.40, .60),
within = .13,
between = .065,
product = .03,
random_slope = .10
)
+ outcome('y', mean = 50, sd = 10)
+ within_time_predictor('x1', weight = .50, values = 0:4)
+ within_predictor('x2', weight = .25)
+ within_predictor('x3', weight = .25)
+ between_binary_predictor('z1', proportion = .50, weight = 0)
+ between_predictor('z2', weight = .50)
+ between_predictor('z3', weight = .50)
+ product('x1','z1', weight = 1)
+ random_slope('x1', weight = 1)
+ correlations(
within = random(.20, .40),
between = random(.10, .30),
randeff = random(.30, .50)
)
)
```
The `mlmpower` package groups $R^{2}$ values and intraclass correlations
into a single object called `effect_size()`, with the five inputs
separated by commas. The within argument corresponds to $R_{w}^{2}$, the
between argument aligns with $R_{b}^{2}$, the product argument specifies
$R_{p}^{2}$, and the random_slope argument corresponds to $R_{rc}^{2}$.
The icc argument assigns a global intraclass correlation to all level-1
variables, and separate simulations are performed for each requested
level.
Variable attributes are referenced by adding the following base objects:
`outcome()`, `within_predictor()`, `between_predictor()`, `product()`,
and `random_slope()`. All five objects have an argument for the variable
name, weight (`weight =`), mean (`mean =`), and standard deviation
(`sd =`). Level-1 variables additionally have an intraclass correlation
argument (`icc =`) that supersedes the global setting in
`effect_size()`.
This illustration additionally uses the `within_time_predictor()` object
to specify a set of fixed time scores for $X_{1}$, and it uses
`between_binary_predictor()` object to define the level-2 moderator
$Z_{1}$ as a binary predictor. In addition to a name and weight, the
`within_time_predictor()` object requires a vector of time scores as an
argument (`values =`). The `between_binary_predictor()` object requires
a name, weight, and the highest category proportion (`proportion =`).
First, the `within_time_predictor()` object specifies $X_{1}$ as a
temporal predictor with the fixed set of time scores described earlier.
The `values = 0:4` argument specifies an integer sequence, but unequally
spaced increments can also be specified using a vector as input (e.g.,
`values = c(0,1,3,6)`). This object does not require a mean or standard
deviation argument, as these quantities are determined from the time
scores. Additionally, the variable's intraclass correlation is
automatically fixed to 0 because the time scores are constant across
level-2 units. Next the
`between_binary_predictor('z1', proportion = .50, weight = 0)` object
specifies $Z_{1}$ (the moderator variable) as a binary predictor with a
50/50 split. This object does not require a mean or standard deviation
argument, as the category proportions determine these quantities.
Finally, within the exception of the outcome, the code block uses
default values for all means and standard deviations (0 and 1,
respectively). The means and standard deviations of the time scores and
binary predictor are automatically determined by the user inputs.
The multilevel model parameters also require three sets of correlations:
within-cluster correlations among level-1 predictors, between-cluster
correlations among level-2 predictors, and random effect correlations.
These correlations are optional, but they can specified using the
`correlations()` object. This example samples within-cluster predictor
correlation values between .20 and .40 (e.g., to mimic a situation where
the time scores have salient correlations with other repeated measures
predictors). Random effect correlation values are similarly sampled
between the range of .30 and .50 to mimic a scenario where higher
baseline scores (i.e., random intercepts) are associated with higher
(more positive) growth rates. Finally, the script specifies the default
setting for between-cluster correlations, which is to iteratively sample
correlation values between .10 and .30.
It may be instructive to inspect the population parameters prior to
running the simulation. Executing `summary(example2)` returns tabular
summaries of the multilevel model parameters.
```{r}
summary(example2)
```
Having specified the target model, you next use the `power_analysis()`
function to conduct simulations. The function requires four inputs: the
model argument specifies the parameter value object (e.g., `example2`),
the `replications` input specifies the number of artificial data sets,
`n_between` is a vector of level-2 sample sizes, and `n_within` is a
vector of level-1 sample sizes (i.e., the number of observations per
cluster). The code block below specifies six level-2 sample size
conditions ($J =$ 50, 60, 70, 80, 90, and 100), each with five repeated
measurements and 2,000 replications.
```{r, eval = FALSE}
# Set seed for replicable results
set.seed(12379)
# Run Power Analysis
powersim2 <-
power_analysis(
model = example2,
replications = 2000,
n_between = c(50, 60, 70, 80, 90, 100),
n_within = 5
)
```
The package uses `lme4` (Bates et al., 2021) for model fitting, and it
defaults to an alpha level of .05 for all significance tests.
Significance tests of random slope variation use a likelihood ratio test
with a mixture chi-square reference distribution (i.e., a chi-bar
distribution; Snijders & Bosker, 2012, p. 99), as implemented in the
`varTestnlme` package (Baey & Kuhn, 2022). Executing
`summary(powersim2)` returns the tabular summaries of the simulation
results shown below.
```{r}
summary(powersim2)
```
## Illustration 3 Vignette: Cluster-Randomized Design
The third vignette demonstrates a power analysis for a
cluster-randomized design (Raudenbush, 1997) where level-2 units are
randomly assigned to one of two experimental conditions. To provide a
substantive context, consider a prototypical education example where
students are nested in schools, and schools are the unit of
randomization. Intraclass correlations for achievement-related outcomes
often range between .10 and .25 (Hedges & Hedberg, 2007; Hedges &
Hedberg, 2013; Sellström & Bremberg, 2006; Spybrook et al., 2011;
Stockford, 2009). To accommodate uncertainty about this important
parameter, the simulation investigates intraclass correlation values of
.10 and .25.
The multilevel model for the illustration is
$$
\begin{split}
Y_{ij} &= \beta_{0} + \beta_{1}\left( X_{1ij}^{w} + X_{1j}^{b} \right) + \beta_{2}\left( X_{2ij}^{w} + X_{2j}^{b} \right)\\
&\phantom{=}+ \ \beta_{3}\left( X_{3ij}^{w} + X_{3j}^{b} \right) + \beta_{4}\left( X_{4ij}^{w} + X_{4j}^{b} \right) + \ \beta_{5}Z_{1j}^{b} + b_{0j} + \varepsilon_{ij}
\end{split}
$$
where $X_{1}$ to $X_{4}$ are grand mean centered level-1 covariates, and
$Z_{1}$ is a binary intervention assignment indicator. Following
established notation, the *w* and *b* superscripts reference within- and
between-cluster deviation scores, respectively. The notation conveys
that all level-1 predictors contain within- and between-cluster
variation. By default, all predictors are centered, including the binary
dummy code. Grand mean centering a level-2 binary predictor creates an
ANOVA-like contrast code, such that $\beta_{0}$ is the grand mean and
$\beta_{5}$ is the predicted mean difference, adjusted for the
covariates. Turning to the residual terms, $b_{0j}$ is a between-cluster
random effect that captures residual variation in the cluster-specific
intercepts, and $\varepsilon_{ij}$ is a within-cluster residual. By
assumption, both residuals are normal with constant variance.
The key inputs to the model object are the unconditional intraclass
correlation and effect size values. This illustration investigates power
for intraclass correlation values of .10 and .25. To mimic a scenario
with a strong covariate set (e.g., one that includes a pretest measure
of the outcome), the within-cluster effect size is set to $R_{w}^{2}$ =
.18 (a value roughly midway between Cohen's small and medium
benchmarks). Most of this variation was allocated to $X_{1}$ (the
pretest) by assigning it a weight of .70, and the remaining three
predictors had their within-cluster weights set to .10. We previously
argued that the allocation of the weights among covariates is arbitrary
because these predictors are not the focus. To demonstrate that point,
we conducted a second simulation that weighted the four level-1
covariates equally.
Turning to the level-2 predictor, researchers often prefer the Cohen's
(1988) *d* effect size when working with binary explanatory variables.
To illustrate, consider power for *d* = .40, the midway point between
Cohen's small and medium benchmarks. Substituting this value into the
conversion equation below returns $R_{b}^{2}$ = .038.
$$R^{2} = \frac{d^{2}}{d^{2} + 4}$$
The following code block shows the `mlmpower` model object for the
example.
```{r}
example3 <- (
effect_size(
icc = c(.10, .25),
within = .18,
between = .038,
)
+ outcome('y')
+ within_predictor('x1', weight = .70)
+ within_predictor('x2', weight = .10)
+ within_predictor('x3', weight = .10)
+ within_predictor('x4', weight = .10)
+ between_binary_predictor('z1', proportion = .50, weight = 1)
)
```
The `mlmpower` package groups $R^{2}$ values and intraclass correlations
into a single object called `effect_size()`, with three inputs separated
by commas. The within argument corresponds to $R_{w}^{2}$, the between
argument aligns with $R_{b}^{2}$, and the icc argument assigns a global
intraclass correlation to all level-1 variables. Separate simulations
are performed for each requested level.
Variable attributes for this example require three objects: `outcome()`,
`within_predictor()`, and `between_binary_predictor()`. The first two
objects have an argument for the variable name, weight (weight =), mean
(`mean =`), standard deviation (`sd =`), and an intraclass correlation
argument (`icc =`) that supersedes the global setting in
`effect_size()`. The `between_binary_predictor()` object requires a
name, weight, and the highest category proportion (`proportion =`).
The previous code block assigns explicit weights to all variables
contributing to a given effect, as described previously. The unit weight
in the `between_binary_predictor()` object reflects the fact that this
$Z_{1}$ solely determines $R_{b}^{2}$. Finally, the proportion argument
assigns a 50/50 split to the level-2 groups.
The multilevel model parameters also require two sets of correlations:
within-cluster correlations among level-1 predictors, and
between-cluster correlations among level-2 predictors (in this case, the
cluster means and the intervention assignment indicator). These
correlations are optional, but they can specified using the
`correlations()` object. The earlier code block omits this object,
thereby accepting the default specification. When the `correlations()`
object is omitted, the `mlmpower` package iteratively samples all
correlations from a uniform distribution between .10 and .30, such that
the resulting power estimates average over a distribution of possible
associations. This default range spans Cohen's (1988) small to medium
effect size benchmarks, and it also brackets common correlations from
published research (Bosco et al., 2015; Funder & Ozer, 2019; Gignac &
Szodorai, 2016). The default specifications could be invoked by
explicitly appending the `correlations()` object to the earlier script,
as follows.
```{r}
example3 <- (
example3
+ correlations(
within = random(0.1, 0.3),
between = random(0.1, 0.3)
)
)
```
Researchers can modify the upper and lower limits of each correlation
range, or they can specify a constant correlation by inputting a scalar
value. The illustrative simulations presented in Enders et al. (2023)
suggest that predictor correlations tend not to matter very much.
It may be instructive to inspect the population parameters prior to
running the simulation. Executing `summary(example3)` returns tabular
summaries of the multilevel model parameters.
```{r}
summary(example3)
```
Having specified the target model, you next use the `power_analysis()`
function to conduct the simulations. The function requires four inputs:
the model argument specifies the parameter value object (e.g.,
`example3`), the `replications` input specifies the number of artificial
data sets, `n_between` is a vector of level-2 sample sizes, and
`n_within` is a vector of level-1 sample sizes (i.e., the number of
observations per cluster). The code block below pairs four level-2
sample size values ($J =$ 30, 60, 90, and 120) with two level-1 sample
sizes ($n_{j} =$ 15 or 30 observations per cluster), and it requests
2,000 artificial data sets for each combination.
The package uses `lme4` (Bates et al., 2021) for model fitting, and it
defaults to an alpha level of .05 for all significance tests. Executing
`summary(powersim)` returns the tabular summaries of the simulation
results shown below.
```{r, eval = FALSE}
# Set seed for replicable results
set.seed(981723)
# Run Power Analysis
powersim3 <-
power_analysis(
model = example3,
replications = 2000,
n_between = c(30, 60, 90, 120),
n_within = c(15, 30)
)
```
The package uses `lme4` (Bates et al., 2021) for model fitting, and it
defaults to an alpha level of .05 for all significance tests. Executing
`summary(powersim3)` returns the tabular summaries of the simulation
results shown below.
```{r}
summary(powersim3)
```