The Social Cost of Global Energy Consumption Due to Climate Change

The analysis in the paper proceeds in five steps.

1. Historical data on energy consumption and climate are cleaned and merged, along with other covariates needed in our analysis (population and income).
2. Econometric analysis is conducted to establish the energy-temperature empirical relationship.
3. This relationship is used to project future impacts of climate change using an ensemble of climate models
   • Note: this step is exceptionally computationally intensive, and sharable code for this step is a work in progress.
4. These impacts are translated into empirical “damage functions” relating monetized damages to warming
5. Damage functions are used to compute an energy-only partial social cost of carbon.

This master readme outlines the process for each step, and each analysis step has it’s own readme and set of scripts in a subdirectory.

Note, the code currently in this repo performs steps 1,2,4 and 5 outlined above. We will update with more replication code in the future. Code used to project impacts will be held in a separate git repo.

Description of folders

0_make_dataset - Code for constructing the dataset used to estimate all models described and displayed in the paper

1_analysis - Code for estimating and plotting all econometric models present in the paper

data - Repository for storing data related to the 1_analysis part of our paper.

figures - Contains figures produced by codes in this analysis

sters - Contains regression output, saved as .ster files

Codes in step 3 onwards also use data held in an external data repository (currently /{synology}/CIL_energy/code_release_data_pixel_interaction/).

Step 1 - Historical Energy Consumption and Climate Dataset Construction

Data construction is a multi-faceted process. We clean and merge data on energy consumption from the International Energy Agency's (IEA) World Energy Balances dataset, historical climate data from the Global Meterological Forcing Dataset (GMFD), and income and population data from the IEA.

In Part A, we construct data on final consumption of electricity and other fuels, covering 146 countries annually over the period 1971 to 2010.
In Part B, we construct data on historical climate to align with the geographic and temporal definitions used in the energy final consumption dataset. In Part C, we clean data on population and income of each country-year in our data. In Part D, we clean and merge together data produced in each of the previous parts, and output an intermediate merged dataset. In Part E, we prepare merged data for econometric analysis.

Part 1.A - Final Consumption Energy Data

• Data on energy consumption were obtained from the International Energy Agency's (IEA) World Energy Balances dataset.
• This dataset is not public, and not provided in this repository.
• From this raw data, we construct a country-year-product level panel dataset (where product is either electricity or other_energy).
• Due to data quality concerns, we incorporate information on data consistency and quality from the IEA's documentation into this dataset.
  • Details of this can be found in Appendix Section A.1, and in the 0_make_dataset/coded_issues folder of this repo.
  • This allows us to determine which data should be dropped, to contruct a set of fixed effects and FGLS weights to help deal with data quality concerns, and assemble climate data that reflects the geographic and temporal definitions used to report energy consumption data.

Part 1.B - Historical Climate Data

• We take Historical Climata Data on daily average temperature and precip- itation from the Global Meteorological Forcing Dataset v1 (GMFD) dataset.
• The raw GMFD data are at the 0.25 x 0.25 degree gridded resolution. We link climate and energy con- sumption data by aggregating gridded daily temperature data to the country-year level using a procedure detailed in Appendix A.2.4 that preserves nonlinearity in the energy consumption-temperature relationship.
  • This step is highly computationally intensive, and the code for this step is not currently provided in this repo.
• In addition to temperature and precipitation measures, we also calculate other climate measures, such as cooling and heating degree days.
• We then clean these data, so that they match the observations present in our energy consumption data.
  • More documentation of the cleaning process can be found in 0_make_dataset/climate

Part 1.C - Population and income data

• We obtain historical values of country-level annual income per capita from within the International Energy Agency's World Energy Balances dataset, which in turn sources these data from the World Bank.
• Cleaning steps undertaken on these variables can be found in 0_make_dataset/pop_and_income

Part 1.D - Merge energy final consumption, historical climate, population and income data

• As the final part of our dataset construction, we merge all of the data together.
• Codes used in this step can be found in 0_make_dataset/merged

Part 1.E - Clean merged data for econometric analysis

• Motivated by tests that showed that our outcome variable has a unit root, we construct first differenced versions of our variables for use in later econometric analysis.
• To nonlinearly model income heterogeneity in the energy-temperature response, we construct an income spline variable. We decide on knot location based on in sample income deciles.
• Lastly to plot 3 x 3 arrays displaying the interacted model (Methods Equation 1), we save a dataset with information on average income and climate in different terciles of the in sample data.

Outputs of Step 1

• Step 1 produces datasets ready to run regressions on, and datasets used in later plotting analysis. These can be found in /data. Specifically,
  • Part 1.D produces /data/IEA_Merged_long_GMFD.do -- an intermediate dataset used to construct the final analysis dataset
  • Part 1.E produces:
    • /data/GMFD_*_regsort.data -- the analysis dataset used in Step 2
    • /data/break_data_*.dta -- a dataset used for plotting 3 x 3 arrays
• Within Step 1, we produce two figures that are used in the paper:
  • Figure Appendix A.1: fig_Appendix-A1_ITA_other_fuels_time_series_regimes.pdf
    • This figure is used to visualise the persistent shocks present in energy consumption.
  • Figure Appendix A.2: fig_Appendix-A2_Unit_Root_Tests_p_val_hists_*.pdf
    • These figures are used to visualise the distribution of p-values for unit root tests for within regime energy consumption.

Step 2 - Econometric Analysis to Establish Energy-Temperature Empirical Relationship

This step implements analysis to recover the emprical relationship between temperature and energy consumption. In this step, we take the cleaned data produced in step 1, run regressions and then plot the resulting outputs.

We run three kinds of regressions in this section:

1. Uninteracted global regressions (Appendix Equation C.1)
   • These regressions recover the global population weighted average response of energy consumption to temperature variation.
   • Code for these regressions can be found in 1_analysis/uninteracted_regression
   • This code outputs:
     • ster files for both the first stage and the FGLS regression: FD_global_TINV_clim.ster and FD_FGLS_global_TINV_clim.ster, respectively
     • Appendix Figure C1: fig_Appendix-B1_product_overlay_TINV_clim_global.pdf
2. Decile regressions (Appendix Equation C.3)
   • These regressions are run in order to understand how the sensitivity of energy consumption to climate change modulates with incomoe levels.
   • Code for these regressions can be found in 1_analysis/decile_regression
   • This code outputs:
     • ster files for both the first stage and the FGLS regression: FD_FGLS_income_decile_TINV_clim.ster and FD_FGLS_income_decile_TINV_clim.ster, respectively
     • Figure 1A: fig_1A_product_overlay_income_decile_TINV_clim.pdf
3. Interacted regressions (Appendix Equations C.4, I.1)
   • In these regressions, we model heterogeneity in the energy-climate relationship, by interacting our models with income and climate covariates.
   • Code for these regressions can be found in 1_analysis/interacted_regression
   • This code outputs:
     • Ster files for the first stage and the FGLS regression for the main (TINV_clim), the excluding imputed data (TINV_clim_EX), and temporal trend (TINV_clim_lininter) models:
       • FD_inter_*.ster and FD_FGLS_inter_*.ster
     • The following paper and appendix figures:
       • Figures 1C: fig_1C_*_interacted_TINV_clim*.pdf
       • Appendix Figures I2: fig_Appendix-G2_*_interacted_main_model_TINV_clim_overlay_model_EX*.pdf
       • Appendix Figures I3A: fig_Appendix-G3A_ME_time_TINV_clim_lininter_*.pdf
       • Appendix Figures I3B: fig_Appendix-G3B_*_interacted_main_model_TINV_clim_overlay_model_lininter.pdf

Step 3 - Project Future Impacts of Climate Change

In this stage of our analysis, we take the coefficients identified in Step 2, and use them to project future impacts on energy consumption due to climate change.

Running code for in this step is highly computationally intensive. Therefore, we are including the inputs from our analysis that would allow a user to run this step, but we are also providing the outputs of this step that are required for further analysis as stand-alone .csv files, should a user wish to run code in steps 4 and 5 without running step 3.

When complete, we plan to contain in this repo:

• Code for converting the regression coefficient estimates from step 2 into a format that our projection system can use.
• Code for writing configuration files for running the projection system.
• Links to the projection system (external) git repo, and some intructions for how to operate the projection system.
• Code for converting projection system outputs into the data required for steps 4 and 5.

Code for this step is not currently in this repo.

Step 4 - Estimate Empirical Damage Function

In this stage, we take the projected future impacts found in step 3, and use them to construct an empirical damage function.

1. As part of this process, we first plot a selection of the projection system outputs, in order to help us understand the spatial and temporal patterns implied by our projections. Codes for this part of our analysis are contained in 3_post_projection/1_visualise_impacts.

   • These codes output take in outputs extracted from the projection system outputs. See readme in Step 3 for more details.
   • They produce:
     • Figure 2A: fig_2A_*_impacts_map.png. These maps show the spatial distribution of the projected impacts of climate change on energy consumption by fuel types, in the year 2099. We also show the response functions associated with selected impact regions, both historically and in our 2099 projection (fig_2A_city_response_functions_2015_and_2099.pdf).
     • Figure 2B: fig_2B_*_consumption_compared_to_2099_impact_bars.pdf, which shows how projected impacts for certain countries compare to their current consumption.
     • Figure 2C: fig_2C_*_time_series.pdf, which shows aggregated global time series of our projected impacts by fuel type and RCP, with uncertainty.
     • Figure 3 /fig_3/.: Figures in this folder present visualisations of monetized damages, combined across fuel types. We present a map of the damages, to highlight the spatial distribution, with visualisations of uncertainty for selected impact regions (Figure 3A). We also show an aggregated time series showing total projected damages by year as percent of global gdp (Figure 3B). The damage functions in Figure 3C are produced by code in damage function estimation.
     • Appendix figures including
       • Appendix Figure C3: /fig_Appendix-C3_sample_overlap_present_future/.
       • Appendix Figure D1: fig_Extended_Data_fig_5_global_total_energy_timeseries_all-prices-rcp*.pdf
       • Appendix Figure H1: fig_Appendix-H1_SSP3-high_rcp85-total-energy-price014-damages_by_inc_decile.pdf
       • Appendix Figure I1: fig_Appendix-G1_Slow_adapt-global_*_timeseries_impact-pc_CCSM4-SSP3-high.pdf
       • Appendix Figure I3.C: fig_Appendix-G3_lininter-global_*_timeseries_impact-pc_SSP3-high-rcp85.pdf

2. We then use the global damages implied by our projections to construct damage functions. Code for estimating these damage functions is contained in 3_post_projection/2_damage_function_estimation.

   • This code plots visualisations of our damage functions in the year 2099 for electricity, other fuels, and total energy priced using our price014 price scenario, that are shown in the paper as:
     • Figure 3C: fig_3C_damage_function_*_2099_SSP3.pdf
     • The code also outputs damage function coefficients for each price scenario, and quantile regression coefficients.
     • Appendix Figure E1: fig_Appendix-E1_total_energy_damage_function_evolution_SSP3-price014.pdf

Step 5 - Compute Energy-Only Partial Social Cost of Carbon

In the final step of the analysis, we use the empirically derived damage function to calculate an energy-only, partial social cost of carbon.

Code for this step is contained in 3_post_projection/3_SCC. This code takes in the damage function coefficients generated in Step 4, and outputs SCC values.

This code outputs all SCC values used in the tables in the paper.

We are planning to add code that computes the uncertainty of these SCC values, but that is not currently in this repo.