Get Started

This vignette details how you can set up and execute a basic power analysis for a bivariate random intercept cross-lagged panel model (RI-CLPM) using the powRICLPM package. Throughout, an illustrating example will be used in which we wish to detect a small cross-lagged effect \(\beta_{2}\) (defined here as the effect of \(a_{1}^{*}\) to \(b_{2}^{*}\), where \(a_{1}^{*}\) and \(b_{2}^{*}\) denote the latent within-unit components of \(a_{1}\) and \(b_{2}\), respectively) of 0.2 (standardized). For the design of our power analysis we follow the steps in the strategy as described in Mulder (2023). Various extensions are available for this basic power analysis, and are described in the Vignette Extensions.

Step 1: Determine experimental conditions

Before performing the power analysis, you must first determine the experimental conditions of interest. Experimental conditions (or: simulation conditions) are defined by characteristics of the study design that can impact statistical power. This includes, among others, characteristics like the sample size and the number of repeated measures. Decide on the number of repeated measures that will be used in the simulations, as well as the range of sample sizes over which you want to simulate the power.

For this example, we take a sample size range from 100 to 1000 first, increasing with steps of 100. Let the numbers of repeated measures range from 3 to 5. If these experimental conditions do not lead to the desired amount of power for detecting the small cross-lagged effect, the ranges can be extended later.

Step 2: Choose population parameter values

Next, determine population parameter values for generating data from the RI-CLPM. This requires the specification of:

For our example, the parameter values are set to:

Phi <- matrix(c(.4, .1, .2, .3), ncol = 2, byrow = T)
# The .2 refers to our standardized cross-lagged effect of interest
within_cor <- 0.3
ICC <- 0.5
RI_cor <- 0.3

If you are unsure if you have specified the Phi matrix as intended, you can use the check_Phi() function to give you a summary of how the effects in your Phi are interpreted.

library(powRICLPM)
# Check `Phi` argument
check_Phi(Phi)
## According to this `Phi`, the lagged effects are:
## • Autoregressive effect of A: 0.4
## • Autoregressive effect of B: 0.3
## • Cross-lagged effect of A -> B: 0.2
## • Cross-lagged effect of B -> A: 0.1

Steps 3-5: Perform the power analysis

Steps 3 to 5 are automated by the powRICLPM() function. As input, you must provide:

You can optionally specify:

Options to extend this basic power analysis setup are described in the Vignette Extensions

Now, we can perform the power analysis by running:

# Set number of replications 
n_reps <- 100

output <- powRICLPM(
  target_power = 0.8,
  search_lower = 500,
  search_upper = 1000,
  search_step = 50,
  time_points = c(3, 4),
  ICC = ICC,
  RI_cor = RI_cor, 
  Phi = Phi,
  within_cor = 0.3,
  reps = n_reps
)

Parallel processing using future

Performing a Monte Carlo power analysis with a large number of replications, and across multiple experimental conditions can be time-consuming. To speed up the process, it is recommended to perform the power analysis across simulation conditions in parallel (i.e., on multiple cores). To this end, the powRICLPM() function has implemented future’s parallel processing capabilities.

Load the future package, and use its plan() function to change the power analysis execution from sequential (i.e., single-core, the default), to multisession (i.e., multicore). Use the workers argument to specify how many cores you want to use. Next, run the powRICLPM analysis, and the power analysis will run on the specified number of cores. This can result in a significant reduction of computing time. For more information on other parallel execution strategies using futures, see ?future::plan().

Progress bar using progressr

It can be useful to get an approximation of the progress of the powRICLPM analysis while running the code, especially when running the analysis in parallel. powRICLPM() has implemented progress notifications using the progressr package. Simply put, there are two options through which you can get progress notifications:

Implementing the with_progress({...}) option, as well as parallel execution of the powRICLPM analysis, results in the below code for the example:

# Load `future` and `progressr` packages
library(future)
library(progressr)

# Check how many cores are available
future::availableCores()

# Plan powRICLPM analysis to run on 1 core less than number of available cores
plan(multisession, workers = 7) # For the case of 8 available cores

# Run the powRICLPM analysis 
with_progress({ # Subscribe to progress updates
  output <- powRICLPM(
    target_power = 0.8,
    search_lower = 500,
    search_upper = 1000,
    search_step = 50,
    time_points = c(3, 4),
    ICC = ICC,
    RI_cor = RI_cor, 
    Phi = Phi,
    within_cor = 0.3,
    reps = n_reps
  )
})

# Revert back to sequential execution of code 
plan(sequential)

For more information about progress notification options using progressr for end-users, including auditory and email updates, see https://progressr.futureverse.org.

Step 6: Summarize results

The powRICLPM() function creates a powRICLPM object: A list with results, upon which we can call print(), summary(), give(), and plot() functions to print, summarize, extract results, and visualize the results, respectively.

print() outputs a textual summary of the power analysis design contained within the object it was called upon. It does not output any performance metrics computed by the power analysis.

summary() can be used in one of four ways. First, summary can be used simply like print() to get information about the design of the power analysis (the different experimental conditions), as well as the number of problems the occurred per condition (e.g., non-convergence, fatal estimation errors, or inadmissible results). Second, by specifying the parameter = "..." argument in summary(), the function will print the results specifically for that parameter across all experimental conditions. Third, if you specify a specific experimental condition using summary()’s sample_size, time_points, ICC and reliability arguments, performance measures are outputted for all parameters in that experimental condition.

The interpretation of the various performance measures available is explained in the function documentation ?summary.powRICLPM().

# Summary of study design
summary(output)

# Summary of results for a specific parameter, across simulation conditions
summary(output, parameter = "wB2~wA1")

# Summary of all parameter for a specific simulation condition
summary(output, sample_size = 500, time_points = 4, ICC = 0.5, reliability = 1)

give() extracts various bits of information from an powRICLPM object. The exact information to be extracted is given by the what = "..." argument:

  1. what = "conditions" gives the different experimental conditions per row, where each condition is defined by a unique combination of sample size, number of time points and ICC.
  2. what = "estimation_problems" gives the proportion of fatal errors, inadmissible values, or non-converged estimations (columns) per experimental conditions (row).
  3. what = "results" gives the average estimate average, minimum estimate minimum, standard deviation of parameter estimates SD, the average standard error SEavg, the mean square error MSE, the average width of the confidence interval accuracy, the coverage rate coverage, and the proportion of times the p-value was lower than the significance criterion power. It requires setting the parameter = "..." argument.
  4. what = "names" gives the parameter names contained within the powRICLPM object.
# Extract experimental conditions
give(output, what = "conditions")

# Extract estimation problems
give(output, what = "estimation_problems")

# Extract results for cross-lagged effect "wB2~wA1" 
give(output, what = "results", parameter = "wB2~wA1")

# Extract parameter names
give(output, what = "names")

Finally, plot() creates a ggplot2-plot for a specific parameter (specified using the parameter = "..." argument) with sample size on the x-axis, the simulated power on the y-axis, lines grouped by number of time-points, and plots wrapped by proportion of between-unit variance. plot() returns a ggplot2 object that can be fully customized using ggplot2 functionality. For example, you can change the scales, add titles, change geoms, etc. More information about options in the ggplot2 framework can be found at https://ggplot2-book.org/index.html. In the below example, I add a title and change the labels on the x-axis:

# Create basic plot of powRICLPM object
p <- plot(output, parameter = "wB2~wA1")
p

# Adjust plot aesthetics
p2 <- p + 
  ggplot2::labs(
    title = "Power analysis for RI-CLPM",
    caption = paste0("Based on ", n_reps, " replications.")
  ) +
  ggplot2::scale_color_discrete("Time points") + 
  ggplot2::guides(
    color = ggplot2::guide_legend(title = "Time points", nrow = 1),
    shape = ggplot2::guide_legend(title = "Reliability", nrow = 1), 
    fill = "none"
  ) + 
  ggplot2::scale_x_continuous(
    name = "Sample size",
    breaks = seq(500, 1000, 50),
    guide = ggplot2::guide_axis(n.dodge = 2)
  )
p2

References

Mulder, Jeroen D. 2023. “Power Analysis for the Random Intercept Cross-Lagged Panel Model Using the powRICLPM r-Package.” Structural Equation Modeling: A Multidisciplinary Journal 30 (4): 645–58. https://doi.org/10.1080/10705511.2022.2122467.

mirror server hosted at Truenetwork, Russian Federation.