Open ICPSR - 198941

• Milena Djourelova (corresponding author), Cornell University. Email: [email protected].
• Ruben Durante, National University of Singapore, BSE, IZA, CESifo, and CEPR.
• Elliot Motte, Universitat Pompeu Fabra.
• Eleonora Patacchini, Cornell University, EIEF, IZA, and CEPR.

This folder contains the data, Stata code, and python code required to replicate the analyses in the "Media Slant and Public Policy Views" paper for the 2024 AEA Papers and Proceedings.

• 00_programs/: Stata ado files
• 01_code/: Stata dofiles and Python scripts
• 02_data/: raw and final data
• 03_output/: Tables and Figure in paper
• requirements.txt: Python packages to install before replication
• data_files_summary.xlsx: list of all data files used in replication code

The authors of the manuscript have legitimate access to and permission to use the data used in this manuscript. They have documented permission to redistribute and publish the data contained within this replication package.

The TV segments data are available in this present repository, with the exception of the raw TV segments which can be found under the OpenIPCSR replication archive for this project. The raw data was collected by the authors, and annotations of this data were generated with GPT-4, OpenAI’s large-scale language-generation model. Upon generating the input prompts and output GPT completions the authors reviewed, edited, and revised the language to their own liking and take ultimate responsibility for the content of these annotations. Input prompts are located in 01_code/01_data_prep/01_GPT_questions.

The CES datasets were downloaded from the Harvard Dataverse repository. They are available at the URLs indicated in the data_files_summary Excel file. Users should properly cite Dagonel (2021) if using the cumulative_cces_policy_preferences.dta. Files should be placed under 02_data/02_CCES/.

The NIELSEN Focus Report 2012 data are proprietary and were acquired by the authors. They are not provided as part of this archive. Researchers interested in access to the data may contact Mr. Jonathan Wells, one of Nielsen’s client managers, at [email protected]. The conclusions drawn from the Nielsen data are those of the researcher(s) and do not reflect the views of Nielsen. Nielsen is not responsible for, had no role in, and was not involved in analyzing and preparing the results reported herein.

Data used for county-level merging (fipscode-to-county crosswalk, zipcode-to-county crosswalk) and county-level controls from the US 2010 Census were downloaded from the source URLs listed in the data_files_summary Excel file. They are in the public domain and are provided as part of this archive. County-level vote share information for the 1996 presidential elections are from David Leip’s US Election Atlas. We provide them here and they can also be acquired by visiting the US Election Atlas Web store, looking under the “General Election Results by County” section, “US President General – County Level Vote Data” subsection, and the “1996 President” file.

See the data_files_summary Excel file for a full list of the datasets used by the code files in this archive.

• Software Requirements:

  • Stata: Code was last run with version 15. Required packages include estout, ftools, reghdfe, statastates, and zscore.
  • Python: Code was run on version 3.9.18. See requirements.txt for necessary Python packages.

• Memory and Runtime Requirements:

  • The code was last run on a 6-core Intel-based laptop with 32GB of RAM and 500 GB of HDD, Windows 11.
  • Runtime:
    • 0_master.do: 7 minutes
    • 01_get_tv_transcripts.py and 02_GPT_labelling.py: 38 hours.

• Programs in programs/01_data_prep will extract and reformat all datasets referenced above. Note that an openAI API key is needed for execution of the 02_GPT_labelling.py script. Researchers interested in automatically using GPT models for data annotation should follow the steps indicated in OpenAI’s quickstart guide to using the API.
• Programs in programs/02_analysis generate all tables and figures in the main body of the article. The program programs/02_analysis/0_master.do will run them all.
• Programs in programs/03_appendix will generate all tables and figures in the online appendix. The program programs/03_appendix/master_appendix.do will run them for Tables A2 and A3. Table A1 is created manually based on output of /03_appendix/Table_A1.py
• Ado files have been stored in 00_programs/ado and the 01_code/0_setup.do files set the ADO directories appropriately.
• The program programs/00_setup.do will populate the programs/ado directory with updated ado packages, but for purposes of exact reproduction, this is not needed. The file programs/00_setup.log identifies the versions as they were last updated.

Random seed for reproducibility of openAI GPT completions is set at line 70 of the 02_GPT_labeling.py script. See the openAI API documentation for further details on pseudo-randomness within the GPT class of models. The random seed for analysis in Stata is set at line 32 of the 0_setup.do.

• Adjust the paths in 01_code/0_setup.do.
• Run 01_code/0_setup.do once on a new system to set up the working environment for Stata.
• Acquire the necessary data files. If obtained, place the Nielsen data under 02_data/04_NIELSEN/.
• Run 01_code/02_analysis/0_master.do to replicate the tables and figure in the main body of the manuscript.
• Run 01_code/03_appendix/master_appendix.do to replicate Tables A2 and A3 of the online appendix.

The author-generated data and code in this replication package are licensed under a Creative Commons Attribution 4.0 International License. See LICENSE.txt for details.

• Dagonel, A. (2021): “Cumulative CCES Policy Preferences,” Harvard Dataverse.