The Corporate Social Responsibility Agent-Based Model (CSR-ABM)

Author

Nick Willems

Preferred Model Citation

Nicholas Willems. (2024). nwillems94/CSR-ABM. Zenodo. https://doi.org/10.5281/zenodo.13363075

---

Welcome to the CSR-ABM: an empirically-grounded, agent-based modeling framework of corporate decision making in the face of economic, regulatory, and social influences. Agent's here consider not only the isolated cost/benefit of specific actions, but also a range of exogenous factors:

• The actions of peer firms. Especially larger, more successful ones.
• Pressure from social activists.
• Shareholders who will value CSR efforts.
• Increased profits from improved reputation. This can manifest through consumers who will pay more for "premium" goods or reduced costs of labor and capital.

In this specific application, the decision to avoid gas flaring is viewed as an act of corporate social responsibility by upstream oil and gas firms. Agents (firms) are assigned heterogenous assets and interact in markets. Assets are empirically grounded oil and gas leases. Agents engage in 2 basic activities: "exploration" (i.e., acquiring new assets) and "development" (i.e., investing in existing assets). An agent's objective is to maximize their composite market value:

$$\begin{align} \text{Market Value} &= (\text{Net Cashflow}) &+ \quad&(\text{Net Social value}) \\ MV &= (\pi + \Delta\pi - C) &+ \quad&(\rho_\theta C - A) \end{align}$$

By spending on limiting flaring (CSR: $C$), firms can avoid pressure from social activists ($A$) and generate additional revenue ($\Delta\pi$) by selling certified gas at a premium price. Simultaneously, shareholders value firms' efforts ($\rho_\theta C$), bolstering stock prices.

The agents, assets, and market conditions are each empirically grounded and validated using freely available public data. To read more about the model and research, see this publication.

Installation

If you encounter issues with either R or package versions, please submit an issue.

Download and install R

This code was originally written for R 4.0.3, but any later versions should also work.

Install required packages

With the goal of minimizing package dependency, there are very few packages required for running the model. Processing the inputs and producing outputs, however, require the use of some more specific packages. I expect the code to continue working as packages evolve over time, but, for the purposes of future-proofing, the correct package versions should be available by calling:

install.packages(..., repos="https://packagemanager.posit.co/cran/2020-07-13")

With the exception of svglite which should be version 2.1.0.

Packages for recreating inputs (not required if using the provided model inputs):

• readxl
• ggplot2
• sf
• maps
• rmarkdown
• formatR
• knitr
• printr

Packages to run the model:

• data.table
• future
• future.apply
• ggplot2 (for validation scripts)
• rmarkdown (for validation scripts)
• Cairo (may be needed to run validation scripts if X11 is not available)

Packages to post-process results & generate graphics:

• DBI
• RSQLite
• svglite
• ggplot2
• ggrepel

The example bash SLURM scripts assume the following commands (standard Unix utilities which are also included as part of e.g., Rtools or msys2) are also available from the command line:

• awk
• cat
• cut
• grep
• sed
• unzip
• gzip
• xz

Download code

You're already here. You can download the code and input data from GitHub.

Running the model

Data

All data needed for running the model are provided in the inputs/processed/ directory. The data were processed and formatted using inputs/firms.Rmd, inputs/leases.Rmd, and inputs/market_history.Rmd. The intermediate results of each script are cataloged in their associated .html files. If you want to experiment with model configurations but use the provided inputs you can skip to the next section. Additionally, users can replicate from a previous initialization by using an existing output database and running a script like scripts/restart.sh (i.e., by creating a copy of the database, running recover_outputs, then optionally reset_db after succesfully writing the new set of runs).

Users wishing to replicate or update bundled datasets will need to compile the Rmarkdown scripts after downloading the raw input data:

Usage(s)	Data Type	File / Directory Name	Source
Lease Data
Lease numbers Production Data	csv	PDQ_DSV.zip	Texas Railroad Commission
Play Boundaries	shp	PermianBasin_Boundary_Structural_Tectonic	EIA
		ShalePlays_AboYeso_GlorietaYeso_Spraberry_EIA
		EagleFord_Play_Boundary_Elevation_Isopach_EIA
Gas processing plants	shp	NaturalGas_ProcessingPlants_US_EIA	EIA
Crude oil pipelines	shp	CrudeOil_Pipelines_US_EIA	EIA
Well latitude, longitude, and depth	csv	dbf900.txt.gz	Texas Railroad Commission
Well depth and type	csv	daf804.txt.gz	Texas Railroad Commission
Decline Curve reference parameters	csv	decline_curves.xlsx	EIA
Market Data
Drilling productivity	csv	dpr-data.xlsx	EIA
Drilling cost index	csv	NG_ENR_WELLCOST_S1_A.xls	EIA
Drilling cost index	pdf	MIT_composite_index.csv	US DOE, MIT
Drilling cost index	pdf		IEA
Oil prices	csv	Cushing_OK_WTI_Spot_Price_FOB.csv	EIA
Natural gas prices	csv	Henry_Hub_Natural_Gas_Spot_Price_weekly.csv	EIA
		Henry_Hub_Natural_Gas_Spot_Price.csv
Natural gas marketed production	csv	U.S._Natural_Gas_Marketed_Production.csv	EIA
		Texas_Natural_Gas_Marketed_Production.csv	EIA
Electricity prices	csv	f8612017_table6.xlsx	EIA
Green electricity premiums	csv	utility-green-pricing-program-list.xlsx	NREL

These data are not ingested by the model directly, they are either processed and combined into a new, formatted input file or used for the purposes of validation or parameter selection. Note file names may change as datasets are updated by their respective publishers.

Configuration

The model configuration can be altered either by editing flaringABM_exe.R directly or by using command line arguments. The default parameters are specified below:

Parameter	Value	Source
Number of runs	32	NA
Number of agents	2,000	Number of operators producing oil and gas in Texas
Total activist funds	$500,000 per month	Greenpeace annual spending on energy activism
Flaring threshold	0.05	Half of baseline
Proportion of green market	Linear growth from 0 to 0.075	New York City residential demand
Premium for green gas	1.07-1.3	Green electricity prices
Shareholder valuation	0.5	Calibrated
Percent of time doing exploration	11/12	NA
Probability of discovering a new asset	1.23	Validated based on EIA Drilling productivity report
Probability of imitation	0.5	NA

Execution

Once the inputs are in place, running the model is as simple as calling Rscript flaringABM_exe.R from the command line. The model configuration can be altered either by editing flaringABM_exe.R directly or by using command line arguments. There are several example bash scripts included in the scripts/ directory which demonstrate those options. The scripts are designed to be queued sequentially (using SLURM) by calling bash run_jobs.sh. start.sh or restart.sh should run first, with graphics.sh running last. The sample scripts are parallelized and designed 32 stochaistic runs on a machine with 48 threads.

Script	Run Description
start.sh	A baseline model (everything on)
reporting.sh	Simulate different kinds of false reporting
social.sh	Configurations without: imitiation, product differentiation, activism, shareholder valuation
market.sh	Vary certified market premium and threshold
actvism.sh	Vary activist strategies
sensitivity/actvism.sh	Sensitivity to activist budget
sensitivity/size.sh	Sensitivity to certified market size
sensitivity/threshold.sh	Sensitivity to certified market threshold

Processing

A significant volume of outputs are generated by a standard model run. To accommodate this, a post processing script is provided (flaringABM_postproc.R) which consolidates the output files and stores them in an sqlite database. The script assumes that the data have been concatenated. To avoid memory overhead in R, the data should be concatenated directly from the command-line. Each of the SLURM scripts contains an (the same) example on how to do so.

The focus is on completeness and portability: the data stored in the database replicate 100% of the data output by the model while occupying significantly less disk space. Some repetitive string columns are stored as with manual dictionary compression: they are converted to integers with integer-key pairs saved in the string_lookup table. To recover string values, see examples in flaringABM_graphics.R.

Creating outputs

Users are encouraged to create their own analyses and graphics. A set of svg outputs are included and can be reproduced by calling flaringABM_graphics.R. Importantly, this script requires that the outputs have been formatted by flaringABM_postproc.R.