lls: Local Least Squares for Information Provision Experiments

This R package implements the Local Least Squares (LLS) estimator for identifying causal effects in information provision experiments. The LLS estimator consistently estimates Average Partial Effects (APE) even when there is strong dependence between belief updating and belief effects.

Installation

You can install lls from GitHub using:

# install.packages("devtools")
devtools::install_github("dballaelliott/lls")

Alternatively, you can install from source by downloading this repository:

# From the source folder
install.packages("path/to/lls", repos = NULL, type = 'source')

# Or from a built tarball
install.packages("path/to/lls_1.0.0.tar.gz", repos = NULL)

Quick Start

devtools::install_github("dballaelliott/lls")
# build_vignettes = TRUE gives you some local examples with 
devtools::install_github("dballaelliott/lls", build_vignettes = TRUE)
# now you have vignette('lls-intro')
library(lls)

Quick Start

# Basic usage - Active/Passive Control
result <- iv.lls(
  dat = your_data,
  y = "outcome_variable", 
  x = "posterior",
  r = "learning_rate_variable",
  control.fml = "~ prior + other_controls"
)

# Panel Design
result <- panel.lls(
  dat = your_data,
  dy = "outcome_change",
  dx = "belief_change"
)

# View results and plot
summary(result)
plot(result)

A Simple Example with Simulated Data

The lls package comes with a simulated dataset info.sim to highlight how the syntax works and to make some example plots.

# Load the package
library(lls)
library(data.table)

# Load packaged simulated data
data(info.sim)
setDT(info.sim)  # ensure data.table

info.sim[, alpha_est := (posterior - prior) / (signal - prior)]

# Estimate using IV mode
est <- iv.lls(info.sim, y = "Y", x = "posterior", r = "alpha_est",
    bandwidth = 0.05, control.fml = "prior",
    bootstrap = TRUE, bootstrap.n = 100)
#> Bootstrapping with 100 iterations

# Print summary
print(est)
#> Local Least Squares (LLS) Estimation
#> ====================================
#> 
#> Average Partial Effect (APE):
#>   Estimate:   0.9454
#>   Std. Err:   0.0912
#>   t-value:   10.3613
#>   p-value:   <0.001
#> 
#> Normal CI (95%): [ 0.7666,  1.1243]
#> Percentile CI (95%): [ 0.7853,  1.1509]
#> 
#> Estimation Details:
#>   Bandwidth:   0.0500
#>   Bootstrap reps: 100
#>   Observations: 500
#>   Support points: 500

There is a plot function for lls objects. It returns a ggplot object, so you can customize it further with standard ggplot2 functions.

library(ggplot2)
plot(est) + 
labs(title = "Heterogeneous Effects of Beliefs by Update Size",
     x = "Change in Beliefs",
     y = "",
     subtitle = "Conditional Average Partial Effect")

Citation

If you use this package, please cite the working paper:

Balla-Elliott, Dylan (2025). “Identifying Causal Effects in Information Provision Experiments.” arXiv:2309.11387

License

MIT License - see LICENSE file for details.