Generate Relational Event History Data And Evaluate Variable Selection And Predictive Performance • remreg

Checking Regularized Estimation of Relational Event Models

The remreg package provides tools to generate Relational Event History (REH) data based on predefined effects and to estimate Relational Event Models (REM) using Maximum Likelihood estimation and regularized estimation techniques. The package is designed to facilitate the evaluation of the performance of regularized estimation techniques in the context of REMs. The package is developed by Jonathan Koop

Installation

You can install the development version of remreg from GitHub with:

install.packages("devtools")
devtools::install_github("jonathankoop/remreg")

Then, load the package:

library(remreg)

Workflow

The remreg package is designed to follow a simple workflow lined out in the following steps:

Set Covariates and their Parameters: Define the covariates in the data.frame covariates and their parameters in the list parameters.
Generate Data: Generate relational event history data based on the covariates and parameters using the generate_reh() function.
Estimate Models: Estimate the REM using Maximum Likelihood estimation and regularized estimation techniques using the estimate_rems() function.
Select Variables: Select predictor variables based on Maximum Likelihood estimation and the regularized estimation using the select_variables() function.
Evaluate Performance: Evaluate the performance of the variable selection techniques using the evaluate_selection() function.

Example

The following example demonstrates the workflow of the remreg package. We first define the covariates and their parameters.

Here, it is important to mention that all effects, also the zero effects, need to be defined in the parameters list. This is necessary to ensure that these variables are considered later in the estimation process.

# set seed for reproducibility
set.seed(1)

# Find suitable effects for directed networks
remstats::tie_effects(directed = TRUE)

# Define parameters
parameters <- list(
  baseline = -5,
  inertia = 0.5,
  
  difference_z1 = 0.5,

  
  # Zero effects to be checked
  difference_z2 = 0,
  reciprocity = 0,
  indegreeSender = 0
)

# Define covariates
covar <- data.frame(
  name = 1:10,
  time = 0,
  z1 = rnorm(10),
  z2 = rnorm(10)
)

Next, we generate relational event history data based on the covariates and parameters.

events <- generate_reh(
  parameters = parameters,
  covar = covar,
  M = 500, # Number of events 
  directed = TRUE # directed network
)

Following the data generation, we estimate the REM using Maximum Likelihood estimation and regularized estimation techniques.

results <- estimate_rems(edgelist = events,
                         methods = c("mle", "abr_horseshoe", 
                                     "abr_ridge", "abr_lasso"))

Important Note:

⚠️ In the case of an error message indicating a singular matrix:

Ensure you do not include effects that are linear combinations of others.
Consider increasing the number of events to make the estimation more stable.
If no solution can be found, please consider opening an issue on GitHub for further assistance.

After estimating the models, we can select the variables based on Maximum Likelihood estimation and the regularized estimation. Here, criteria have to be defined to select the variables.

selected <- select_variables(results,
                             criterion_mle = "p<0.05",
                             criterion_abr = "95% hdi")

Finally, we evaluate the performance of the variable selection techniques.

evaluation <- evaluate_selection(selected)

Contact

For questions, suggestions, or issues, please open an issue on GitHub or contact me via email.