Do electrical interties stimulate Canadian hydroelectric development? Using causal inference to identify second-order impacts in evolving sociotechnical systems

Image courtsey of American Public Power Association

Abstract

Debates over the scope of environmental impact, life-cycle, and cost-benefit analysis frequently revolve around disagreements on the causal structure of complex sociotechnical systems. Environmental advocates in the United States have claimed that new electrical interties with Canada increase development of Canadian hydroelectric resources, leading to environmental and health impacts associated with new reservoirs. Assertions of such second-order impacts of two recently proposed 9.5 TWh year^-1 transborder transmission projects played a role in their suspension. We demonstrate via Bayesian network modeling that development of Canadian hydroelectric resources is stimulated by price signals and domestic demand rather than increased export capacity per se. However, hydropower exports are increasingly arranged via long-term power purchase agreements that may promote new generation in a way that is not easily modeled with publicly available data. Overall, this work suggests lesser consideration of generation-side impacts in permitting transborder transmission infrastructure while highlighting the need for higher resolution data to model the Quebec-New England-New York energy system at the project scale. More broadly, Bayesian analysis can be used to elucidate causal drivers in evolving sociotechnical systems to develop consensus for the scope of impacts to consider in environmental impact, life cycle, and cost-benefit analysis.

Contents

• Overview
• Repo Contents
• System Requirements
• Installation Guide
• Developed New Functions
• Citation.bib
• Copyrights

Overview

This repository contains data, instructions, and code for the "Do electrical interties stimulate Canadian hydroelectric development? Using causal inference to scope environmental impact assessment in evolving sociotechnical systems" paper. This repository includes R code, a manual for using the code and utilizing the bnlearn^[1] package in this context, and a real dataset for practical application.

Repo Contents

• R files: contains R files
• doc: code reproduction information, manual for using the bnlearn package on our dataset
• data: real dataset to use in the R session
• citation: bib code to use when citing this code or the manuscript
• license: creative commons attribution 4.0 international license
• misc: includes miscellaneous files such as source code to generate Figure 4 in the manuscript

System Requirements

Hardware and OS Requirements

We utilized a MacBook Pro with an Apple M1 Pro chip, featuring an 8-core CPU and 16 GB of memory. The startup disk is the Macintosh HD. The system operates on macOS 13.2.1 (22D68) Ventura.

Software Requirements

This code is tested on macOS operating systems. The Comprehensive R Archive Network (CRAN) package which is the underlying softawre for this code, is compatible with Windows, Mac, and Linux operating systems.

Before setting up the R code, users should have R version 4.3.1 (2023-06-16) Beagle Scouts or higher, and several packages set up from CRAN.

Installation Guide

Install R and R-Studio

You can download and install R via CRAN for free from this link. You can download and install R Studio for free from this link.

Install Packages

Users should install the following packages prior to running the supplied R code, from an R terminal:

install.packages(c("bnlearn", "gRain", "visNetwork", "ggplot2", 
              "zoo", "scales", "gridExtra", "dplyr", "MASS", "svglite", "tidyverse"))

which will install in about a few minutes on a machine with similar specs.

All packages in their latest versions as they appear on CRAN on November 10, 2013. The versions of packages are:

Rgraphviz 2.44.0
bnlearn 4.8.3
gRain 1.3.13
visNetwork 2.1.2
ggplot2 3.4.2
zoo 1.8.12
scales 1.2.1
gridExtra 2.3
dplyr 1.1.2
MASS 7.3.60
svglite 2.1.1
tidyverse 2.0.0

Code Runtime

The runtime on our operating system for this code in R is approximately 15 seconds.

Developed New Functions

We have developed the following functions to simplify the algorithm:

D-Separation Function

We have created the dsep.dag function that can evaluate d-separation for any given node pair. This function uses the optimized DAG results to identify the following for each node pair and then calculates conditional independence: 1. Parents; 2. Neighbors (i.e., parents and children for each node); and 3. Markov-Blanket (i.e., parents, children, and parents of children for each node). Furthermore, this function uses data to evaluate d-separation in addition to the results of DAG discovery. This combines the functionalities of the ci.test and dsep functions available in the bnlearn package. Where ci.test exclusively utilizes data, while dsep.dag solely employs DAGs.

dsep.dag(x, data, z)

Parameters

The dsep.dag function accepts the following parameters:

• x: an object of class bn
• data: a data frame containing the variables in the model
• z: a list, where each element is a character vector representing a pair of node labels

DAG Visualizer

The plot.network function visualizes and returns DAGs returned by the bnlearn package.

plot.network(structure, ht = "400px", title)

Parameters

The plot.network function accepts the following parameters:

• structure: an object of class bn
• ht: a string specifying the height of the plot. If none is specified, the default value will be 400px
• title: a character string, the title of the plot. If none is specified, the title will be blank

Box-Cox Transformer

The transform_and_test function evaluates the data set, checks for normality (using the Shapiro-Wilk test), and transforms the non-Gaussian variables using the Box-Cox transformation. Re-evaluates the transformed variables with the Shapiro-Wilk test and checks for normality. Returns the results.

transform_and_test(data, z)

Parameters

The transform_and_test function accepts the following parameters:

• data: a data frame containing the variables in the model
• z: a list, where each element is a character vector representing a variable

Goodness of Fit Test

The evaluate_fit_continuous and evaluate_fit_discrete functions evaluate each variable's goodness of fit for the DAGs produced by the bnlearn package. The evaluate_fit_continuous evaluate continious variables and the evaluate_fit_discrete evaluates discrete variables.

evaluate_fit_continuous(data, pred)

Parameters

The evaluate_fit_continuous function accepts the following parameters:

• data: a data frame containing continuous variables in the model
• pred: a data frame containing continuous variables' predictions using the model

---

evaluate_fit_discrete(data, pred)

Parameters

The evaluate_fit_discrete function accepts the following parameters:

• data: a data frame containing discrete variables in the model
• pred: a data frame containing discrete variables' predictions using the model

Citations

When using this code or the associated manuscript, please cite using the following citation.bib.

Copyrights

This work is licensed under a Creative Commons Attribution 4.0 International License.

References

[1] M. Scutari. Learning Bayesian Networks with the bnlearn R Package. Journal of Statistical Software, 35(3):1-22, 2010.