Data collection and analysis code for the article From Alternative conceptions of honesty to alternative facts in communications by U.S. politicians.

Cite as
J. Lasser, S.T. Aroyehun, F. Carrella, A.Simchon, D.Garcia, S. Lewandowsky, Alternative conceptions of honesty to alternative facts in communications by U.S. politicians, Nature Human Behaviour, DOI: 10.1038/s41562-023-01691-w (2023).

Contact

• Jana Lasser ([email protected])
• Stephan Lewandowsky ([email protected]).

The code published in this repository needs a number of data sets to execute. These data sets are published in a separate OSF repository, because they are too large to be hosted on GitHub. Therefore, to run the code, after downloading this repository please download the data folder contained in the aforementioned OSF repository and copy it into the main directory of this repository (i.e. on the same level as the code, plots and tables folders).

There are restrictions to publishing all data sets required to fully reproduce our work. These restrictions apply to

1. Tweets: Due to data protection reasons and Twitter's usage agreement for the use of its API, we cannot publish tweet texts. As a workaround, we publish a data set with tweet IDs and statistics that were computed using the tweet text, such as honesty component labels. Tweet IDs can be used to hydrate tweets and retrieve the original text, as long as the tweets are still accessible at the time of hydration. We provide the script hydrate_tweets.ipynb to hydrate tweets from the list of tweet IDs published in the OSF repository. This script produces the input for the script wrangle_data.ipynb and replaces the script get_US_politician_twitter_timelines.ipynb, which collects the same tweets using the Twitter v2 academic API.

2. NewsGuard scores: we cannot publish NewsGuard domain nutrition labels, since the NewsGuard data base is proprietary. For this reason, certain parts of the wrangle_data.ipynb that aggregate NewsGuard scores with the tweet data will not execute. Researchers that want to reproduce our results will have to acquire their own copy of a recent version of the NewsGuard data base and place it in the folder data/utilities named NewsGuard_labels.csv. This file is expected to have the columns Domain and Score as well as the nine columns for the individual criteria such as Does not repeatedly publish false content.

3. New York Times article abstracts: we cannot publish the New York Times article corpus. You can however collect the data yourself using the New York Times' API to collect abstracts from the cateogires "Climate", "Education", "Health", "Science", "U.S.", "Washington", "World" and "Opinion".

Therefore, the following scripts will only execute partly without aquiring the tweet texts, NewsGuard data and New York Times abstracts are

• tweet_collection/wrangle_data.ipynb (requires NewsGuard data base and tweet texts)
• analysis/descriptive_dataset_statistics.ipynb (requires Newsguard data base)
• analysis/label_glove840B_DDR.sh (requires tweet texts and New York Times abstracts)
• analysis/label_lexicon_loop.sh, analysis/label_lexicon_single_word.sh, analysis/label_fasttext-cc_DDR.sh and analysis/label_word2vec-googlenews_DDR.sh (requires tweet texts)
• analysis/bertopic_model.ipynb (requires tweet texts)
• analysis/scattertext.R (requires tweet texts)

Irrespective of these restrictions we publish all data sets necessary to reproduce all figures in our manuscript except for figure 1, since these data sets need not contain original texts nor the NewsGuard data base.

We cannot supply code to reproduce our computation of the LIWC labels for the "authentic", "analytic" and "moral" text components. This is due to the fact that LIWC22 is a proprietary software. Authors interested in reproducing our results should acquire access to LIWC-22 and apply it to the tweet texts. We do hoverwer supply the computed scores for the LIWC text components in the file US_politician_tweets_2010-11-06_to_2022-12-31.csv.gzip

The repository is organized into the two top-level folders code and plots. The folder plots contains only outputs of the analysis scripts. The folder code is subdivided into three subfolders:

• tweet_collection contains all scripts necessary to collect data from the twitter API, clean the data and wrangle it into the various forms required for subsequent analysis.
• article_collection contains all scripts necessary to collect articles from the URLs in the tweets data set and wrangle them into a form suitable for further processing.
• analysis contains scripts for
  • calculating descriptive dataset statistics descriptive_dataset_statistics.ipynb
  • calculating honesty component similarity with various embeddings: label_glove840B_DDR.sh, label_fasttext-cc_DDR.sh, label_word2vec-googlenews_DDR.sh and compute_sbert_avg_lexicon.py
  • calculating keyword similarity: label_lexicon_single_word.sh and compute_sbert_avg_lexicon_reduce_lexicon_single_word.py
  • calculating VADER sentiment scores: compute_VADER_scores.ipynb
  • validating honesty components with human ratings: validation.R create_document_level_validation_sample.ipynb (drawing the sample) and document_level_validation.R (analysis)
  • bootstrapping: bootstrapping.ipynb
  • statistical modelling: lmer_models_tweets.Rmd, OLS_regression_articles.ipynb and mediation.Rmd
  • assessing dictionary robustness label_lexicon_loop.sh and compute_sbert_avg_lexicon_reduce_lexcion_loop.sh
  • topic modelling with BERTtopic: bertopic_model.ipynb
  • reproducing all plots from the article: plots.ipynb and scattertext.R
  • scrubbing all protected information from the data sets for upload to OSF: scrub_data_for_upload.ipynb

Collection of Twitter user accounts associated with members of the U.S. congress is accompished in the script get_US_politician_twitter_accounts.ipynb. We combine lists of accounts from different sources such as socialseer, manually inspect and clean the data and download the account profiles using the Twitter API. Account information is stored in the file users/US_politician_twitter_accounts_clean.csv for later ingestion by wrangle_data.ipynb.

Next, we collect the timelines of all tweets from all accounts in the script get_US_politician_twitter_timelines.ipynb. Since we are interested in tweets from November 6, 2010 to December 31, 2022, we make use of access to historic tweets via the academiv Twitter v2 API access. We use the Python library twarc to collect the tweets and store them after some basic cleaning steps. Cleaned timelines are stored in the file tweets/combined_US_politician_twitter_timelines_2010-11-06_to_2022-12-31_clean.csv.gzip for later ingestion by wrangle_data.ipynb.

A large part of our research relies on the quality of links shared by Congress Members. A large number of URLs are shared in shortened form, using a link shortening service such as bit.ly. We have also observed that links to low-quality websites are much more likely to be shortened than links to high-quality sites. We therefore follow all shortened links to retrieve their true domain. We identify shortened links using a data base of link shorteners. In addition, we manually inspect all domains that were tweeted > 100 times and look for additional link shortening services that were not contained in the initial data base of shorteners. The thus retrieved "unraveled" URLs are stored in the file urls/US_unraveled_urls.csv.xz for later ingestion by wrangle_data.ipynb.

In the script tweet_collection/wrangle_data.ipynb we export a list of urls articles/url_list_for_article_scraping.csv.gzip from which we scrape article texts. This list is ingested by the script article_scraping.R which tries to retrieve the page texts from all links, making use of the Python library newspaper3k.

Note that wrangle_data.ipynb also saves a file articles/url_NG_scores.csv.gzip with the URL, domain, author ID of the account that posted the URL, party, tweet ID in which the URL was posted and NewsGuard score. This file is later ingested by OLR_regression_articles.ipynb. We do not share this file, since it contains the NewsGuard scores, which are proprietary. We do, however, share a file articles/url_independent_scores.csv.gzip, which contains the trustworthiness scores calculated with our independent data base, which can be used to reproduce our results without having access to NewsGuard.

Scraped articles are stored in articles/article_corpus_raw.rds.gz and pre-processed by article_preprocessing.R. Pre-processing also reads the initial list of URLs for article scraping articles/url_list_for_article_scraping.csv.gzip, since this file also contains information about the party of the account that posted the link to the article. This information is needed to filter the articles retrieved from the URLs, since we only retain articles that were linked to by one party, not by both. The filtered articles are then stored in articles/article_corpus_clean.csv.gzip for later ingestion by ``label_glove840B_DDR.sh`.

Validation of the honesty component keywords and documents is done via a survey on qualtrics. The results of the survey are stored in data/validataion.

The dictionary validation shown in Supplementary Figures 1 & 2 is done in the script analysis/keyword_validation.R. The script ingests the files validation/validation_belief.csv and validation/validation_truth.csv. The sample for the document level validation is created in the script analysis/document_level_validation.ipynb. The document level validation analysis, which is shown in supplementary figures S3-S6, is done in the script analysis/document_level_validation.R, which ingests the sample file validation/document_validation_sample.csv and validation/document_validation_data.csv.

Honesty component similarities for the New York Times article corpus, the tweet corpus and the corpus of articles collected from URLs in the tweet corpus are calculated by the scripts label_glove840B_DDR.sh. In addition, the scripts label_fasttext-cc_DDR.sh and label_word2vec-googlenews_DDR.sh also calcupate honesty component similarities for the tweet corpus based on alternative embeddings (word2vec and fasttext). The scripts internally call compute_sbert_avg_lexicon.py with different parameters to specify settings for each corpus. The script compute_sbert_avg_lexicon.py ingests the file tweets/combined_US_politician_twitter_timelines_2010-11-06_to_2022-12-31_clean.csv.gzip and calculates honesty component similarity using the glove sentence transformer embedding on the New York Times corpus, the tweet corpus and the article corpus. It outputs the honesty component similarity for every corpus in a file:

• The output file tweets/combined_US_politician_twitter_timelines_2010-11-06_to_2022-12-31_honesty_component_scores_glove.csv.gzip is ingested by wrangle_data.ipynb.
• The output file NYT/NYT_abstracts_honesty_component_scores_glove.csv is ingested by scrub_data_for_upload.ipynb.
• The output file articles/article_corpus_clean_honesty_component_scores_glove.csv.gzip is ingested by scrub_data_for_upload.ipynb.

Note that for this to work you will need to download the embeddings and place the respective models in the data/utilities/sentence-transformers directory.

We do not include the results of the dictionary robustness analysis here, due to the size of the generated data. You can however reproduce the results by running the script label_lexicon_loop.sh, which also ingests the file tweets/combined_US_politician_twitter_timelines_2010-11-06_to_2022-12-31_clean.csv.gzip. This calculates honesty component similarity using the glove sentence transformer embedding, but for 100 perturbed versions of the belief-speaking and truth-seeking dictionaries. Note that for this to work you will need to download the embeddings and place the respective models in the data/utilities directory.

The script tweet_collection/wrangle_data.ipynb includes a commented out section ("Add truth seeking & belief speaking scores for dictionary bootstraps") to include the dictionary robustness data tweets/combined_US_politician_twitter_timelines_2010-11-06_to_2022-12-31_honesty_component_scores_glove_bootstrap.csv.gzip in the data processing workflow. If you run wrangle_data.ipynb with these additional lines, it will load the output from label_lexicon_loop.sh, merge it to the rest of the tweet-level data and also include it in the final aggregated tweet data set US_politician_tweets_2010-11-06_to_2022-12-31.csv.gzip.

This file then needs to be run through scrub_data_for_upload.ipynb. The resulting tweets/tweets.csv file can be loaded in dictionary_robustness.ipynb to fit the linear mixed effects model for every one of the 100 perturbed versions of each honesty components. The output of this script is the estimates for the fixed effects of the LME for each of the 100 dictionary versions, which is saved in tweets/LME_results_dictionary_robustness.csv and loaded in analysis/plots.ipynb to generate extended data figure 2.

We processed the full text of each tweet with LIWC-22, the latest version of the Linguistic Inquiry and Word Count Software LIWC-22. We exported the text of each tweet to a csv file with one row per tweet and two columns, one with the tweet id as a string and the second with the string containing the tweet text. We imported this file to LIWC-22 and processed only the column with the tweet text using the LIWC-22 English dictionary for the "authentic", "analytic" and "moral" text components. The result was exported as a csv file including additional columns for each LIWC metric for each tweet. Scores are stored in the file tweets/combined_US_politician_twitter_timelines_2010-11-06_to_2022-12-31_clean_mask_LIWC.csv.gzip for later ingestion by wrangle_data.ipynb.

VADER scores are calculated with the Python library vaderSentiment in the script compute_VADER_scores.ipynb and stored in the file tweets/combined_US_politician_twitter_timelines_2010-11-06_to_2022-12-31_clean_VADER.csv.gzip for later ingestion by wrangle_data.ipynb.

The script wrangle_data.ipynb takes input from all previous data collection and analysis steps to create three output data files that aggregate most information for different downstream analysis tasks:

• Tweets: contains all information pertaining to individual tweets and is stored at tweets/US_politician_tweets_2010-11-06_to_2022-12-31.csv.gzip. It contains the following columns:
  • id: unique tweet ID
  • author_id: unique ID of the author that posted the given tweet. Used to link to the user data file described above.
  • party: party affiliation of the account that posted the tweet.
  • created_at: tweet creation time in UTZ time.
  • retweeted, quoted and reply: whether the tweet is a retweet, quote-tweet or reply. Note that these categories are not exclusive.
  • avg_belief_score and avg_truth_score: float, length-corrected and centered similarity to the belief-speaking and truth-seeking dictionaries. If the robustness analysis is included (see section "Dictionary robustness" above), there are columns with similarity scores calculated using perturbed dictionaries avg_belief_score_i and avg_truth_score_i, where i ranges from 0 to 99. In addition avg_belief_score_word2vec, avg_truth_score_word2vec, avg_belief_score_fasttext and avg_truth_score_fasttext contain the similarity ratings calculated with the word2vec and fasttext embeddings instead of GLoVe.
  • has_url: bool, whether the tweet contained an URL.
  • LIWC_analytic, LIWC_authentic, LIWC_moral, VADER_pos, and VADER_neg: LIWC scores for the text components "analytic", "authentic", "moral" and VADER scores for positive and negative sentiment. Scores are calculated using the LIWC-22 software and the vaderSentiment library.
  • NG_score: float. Only exists for tweets that contained an URL. NewsGuard score determined by matching the domain contained in the URL against the NewsGuard database of information quality. If the tweet contained more than one URL with a NewsGuard score, the scores are averaged.
  • transparency and accuracy: float. Only exists for tweets that contained an URL. Accuracy and transparency score determined by matching the domain contained in the URL against the Independent information quality database. If the tweet contained more than one URL with an accuracy and transparency score, the scores are averaged.
  • NG_unreliable and independent_unreliable: bool. Only exists for tweets that contained an URL. Whether the URL pointed to an "unreliable" website, i.e. a website with a NewsGuard score < 60 or an accuracy score < 1.5 or a transparency score of < 2.5.
• Users: contains all information pertaining to individual twitter accounts and is stored at users/users.csv. It contains the following columns:
  • handle, name and author_id: Twitter handle, screen name and author ID of the user account.
  • followers_count, following_count, tweet_count: number of followers, accounts followed and tweets posted since account creation, retrieved from the Twitter API.
  • created_at: account creation date, retrieved from the Twitter API.
  • N_tweets: number of tweets from the given account contained in the tweets data set collected for this project.
  • party: Party affiliation Democrat, Republican, Independent or Libertarian.
  • congress: Latest congress the Congress Member belonged to, can be 114, 115, 116 or 117.
  • type: type of the account, can be official, campaign or staff.
  • NG_unreliable_share, independent_unreliable_share: share of tweets containing a link to an "unreliable" domain, determined with the NewsGuard and independent data base of domain information quality, respectively.
  • NG_score_mean, accuracy_mean, transparency_mean: Average NewsGuard score as well as accuracy and transparency score over all domains contained in tweets by the given accounts.
  • avg_belief_score, avg_truth_score: average belief-speaking and truth-seeking similarity in tweets by the author.
  • avg_belief_score_2010_to_2013, avg_truth_score_2010_to_2013: average belief-speaking and truth-seeking similarity in tweets by the author in the years 2010 to 2013, respectively.
  • avg_belief_score_2019_to_2022, avg_truth_score_2019_to_2022: verage belief-speaking and truth-seeking similarity in tweets by the author in the years 2019 to 2022, respectively.
  • ideology_count, ideology_mean and ideology_std: number of ideology score entries found in the govtrack database for the given account as well as average ideology score and ideology score standard deviation.
  • pf_score: politifact score.
• URLs: contains all information pertaining to individual URLs posted by twitter accounts of U.S. Congress Members and is stored at urls/US_politician_URLs_2010-11-06_to_2022-03-16.csv.gzip. It contains the following columns:
  • id: ID of the original tweet the URL was contained in
  • author_id: ID of the Twitter user account that posted the tweet the URL was contained in.
  • created_at: creation date of the tweet the URL was contained in.
  • retweeted, quoted and reply: bool. Whether the tweet the URL was contained in was a retweet, quote-tweet or reply.
  • party: Party affiliation of the account that posted the tweet the URL was contained in. Can be Democrat, Republican, Independent or Libertarian.
  • shortened_url: whether the URL was originally shortened using a link shortening service such as bit.ly.
  • avg_belief_score and avg_truth_score: float, similarity to the belief-speaking and truth-seeking dictionaries. If the robustness analysis is included (see section "Dictionary robustness" above), there are columns with similarity scores calculated using perturbed dictionaries avg_belief_score_i and avg_truth_score_i, where i ranges from 0 to 99.
  • NG_score, accuracy and transparency: NewsGuard, accuracy and transparency score of the domain the URL pointed to, determined with the NewsGuard and Independent of information quality.
  • NG_unreliable and independent_unreliable: Whether the URL pointed to an "unreliable" website, i.e. a website with a NewsGuard score < 60 or an accuracy score < 1.5 or a transparency score of < 2.5.

When all data has been preprocessed, the files US_politician_tweets_2010-11-06_to_2022-12-31.csv.gzip, US_politician_URLs_2010-11-06_to_2022-12-31.csv.gzip, article_corpus_clean_honesty_component_scores_glove.csv.gzip, url_NG_scores.csv.gzip, url_independent_scores.csv.gzip, NYT_abstracts_honesty_component_scores_glove.csv.gzip and NYT_abstracts.csv.gzip are ingested by the script scrub_data_for_upload.ipynb to scrub the data from all columns that contain protected information (tweet texts and New York Time article abstracts) or would allow for a link between domains and NewsGuard scores. The script produces four clean files that are (together with users/users.csv, tweets_for_lme_modelling_NG.csv.gzip and tweets_for_lme_modelling_independent.csv.gzip from wrangle_data.ipynb) provided in the OSF repository:

• tweets/tweets.csv.gzip
• urls/urls.csv.gzip
• NYT/abstracts.csv.gzip
• articles/articles.csv.gzip

The following analysis scripts all only need these files to run.

Timelines shown in our paper (Figure 1, Extended Data Figure 3, Supplementary Figures S3 and S4) include confidence intervals determined via bootstrapping. Bootstrapping is performed in the script bootstrapping.ipynb. Bootstrapping results are saved in the folder bootstrapping for ingestion by plots.ipynb.

Statistical modelling to determine the relation of belief-speaking and truth-seeking to information quality is performed in two scripts: lmer_models_tweets.r for the tweets by politicians, and OLS_regression_articles.ipynb for articles retrieved from the links posted in tweets by politicians.

For tweets we use a linear mixed effects model implemented in lme4 where tweets are grouped within user and users are grouped within party. The script lmer_models_tweets.r ingests the files tweets/tweets_for_lme_modelling_NG.csv.gzip and tweets/tweets_for_lme_modelling_independent.csv.gzip produced by tweet_collection/wrangle_data.ipynb. The files contain belief-speaking and truth-seeking similarity and trustworthiness scores determined by NewsGuard and the independent list, respectively. The LME modelling script outputs the files tweets/LME_predictions_tweets_belief.csv and tweets/LME_predictions_tweets_truth.csv used to plot the prediction lines in Figure 3, panels A and B. It also outputs the file tweets/LME_predictions_tweets_threeway.csv used to plot the three-way interactions visualized in Extended Data Figure 1.

For articles we use a linear regressions model that are fitted using an ordinary least squares method, implemented by the Python library statsmodels. The script OLS_regression_articles.ipynb ingests the files articles/articles.csv.gzip containing the belief-speaking and truth-seeking similarity scores, the NewsGuard scores and party information. It then fits a regression models with one observation for each article. Results of the regression are stored in tables/OLS_table_article_NG_score.txt for input into the manuscript. The script also outputs predictions of average NewsGuard score of depending on the belief-speaking and truth-seeking similarity of a given article. The predictions are stored in the file articles/OLS_predictions_articles_honesty.csv and ingested by plots.ipynb to produce Figure 3.

Mediation analysis is performed in the script analysis/mediation.r. The script ingests the file users/users.csv and outputs the summary statistics of the mediation analysis.

Before further analysis of the texts can be performed, the texts first need to be lemmatized. This is done in the script tweet_collection/wrangle_data.ipynb. The script outputs the file tweets/combined_US_politician_twitter_timelines_2010-11-06_to_2022-12-31_lemma.csv.gzip, which is used by bertopic_model.ipyb. Topic modelling is performed using the library BERTopic. The fitted model is saved at tweets/BERTopic_model for later re-use, since fitting the model takes a while.

The script outputs two results files: tweets/topics_all_docs.csv.gzip and tweets/topics_per_class_ddr.csv, which are used by plots.ipynb to visualize topics in Supplementary Figure 5.

Descriptive statistics of the various data sets used in our analysis and reported in the paper are calculated in the script analysis/descriptive_dataset_statistics.ipynb. The script also contains additional descriptive visualisations not contained in the article.

All visualisations in the main manuscript, extended data and figures and supplement except for Figure 1 and Supplementary Figures 1 & 2 are created in the script plots.ipynb and saved in the folder plots. Figure 1 is created by scattertext.R. Supplementary Figures 1 & 2 are created by keyword_validation.R and Supplemengary Figures 3, 4, 5 and 6 are created by document_level_validation.R. The scripts ingest the following files to create the figures:

• tweets/combined_US_politician_twitter_timelines_2010-11-06_to_2022-12-31_lemma.csv.gzip' (created by tweet_collection/wrangle_data.ipynb`, not provided in repository)

• users/users.csv (created by tweet_collection/wrangle_data.ipynb, provided in repository)
• bootstrapping/belief.csv (created by analysis/bootstrapping.ipynb, provided in repository)
• bootstrapping/truth.csv (created by analysis/bootstrapping.ipynb, provided in repository)

• tweets/tweets.csv.gzip (created by analysis/scrub_data_for_upload.ipynb, provided in repository)
• tweets/LME_predictions_tweets_belief.csv (created by analysis/lmer_models_tweets.Rmd, provided in repository)
• articles/articles.csv.gzip (created by analysis/scrub_data_for_upload.ipynb, provided in repository)
• articles/OLS_predictions_articles.csv (created by analysis/OLS_regression_articles.ipynb, provided in repository)

• LME_predictions_tweets_threeway.csv (created by analysis/lmer_models_tweets.Rmd, provided in repository)

• tweets/LME_results_dictionary_robustness.csv (created by analysis/dictionary_robustness.ipynb, provided in repository)
• tweets/tweets.csv.gzip (created by running tweet_collection/wrangle_data.ipynb with the code for the robustness analysis included and then running the data through analysis/scrub_data_for_upload.ipynb again, not provided in repository.)

• bootstrapping/NG_coverage.csv.gzip (created by analysis/bootstrapping.ipynb, provided in repository)
• bootstrapping/independent_coverage.csv.gzip (created by analysis/bootstrapping.ipynb , provided in repository)

• tweets/tweets.csv.gzip (created by analysis/scrub_data_for_upload.ipynb, provided in repository)
• articles/articles.csv.gzip (created by analysis/scrub_data_for_upload.ipynb, provided in repository)

• validation/validation_belief.csv (output from Qualtrics, provided in repository)

• validation/validation_truth.csv (output from Qualtrics, provided in repository)

• validation/document_validation_sample.csv (created by analysis/document_level_validation.ipynb, provided in repository)
• validation/document_validation_data.csv (output from Qualtrics, provided in repository)

Note that this figure is an analysis of VADER scores, which are also included in the bootstrapping analysis saved in the LIWC files.

• bootstrapping/LIWC.csv.gzip (created by analysis/bootstrapping.ipynb, provided in repository)
• bootstrapping/LIWC_belief.csv.gzip (created by analysis/bootstrapping.ipynb, provided in repository)
• bootstrapping/LIWC_truth.csv.gzip (created by analysis/bootstrapping.ipynb, provided in repository)
• bootstrapping/LIWC_neutral_belief.csv.gzip (created by analysis/bootstrapping.ipynb, provided in repository)
• bootstrapping/LIWC_neutral_truth.csv.gzip (created by analysis/bootstrapping.ipynb, provided in repository)

• bootstrapping/LIWC.csv (created by analysis/bootstrapping.ipynb, provided in repository)
• bootstrapping/LIWC_belief.csv (created by analysis/bootstrapping.ipynb, provided in repository)
• bootstrapping/LIWC_truth.csv.gzip (created by analysis/bootstrapping.ipynb, provided in repository)
• bootstrapping/LIWC_neutral_belief.csv.gzip (created by analysis/bootstrapping.ipynb, provided in repository)
• bootstrapping/LIWC_neutral_truth.csv.gzip (created by analysis/bootstrapping.ipynb, provided in repository)

• tweets/topics_all_docs.csv.gzip (created by analysis/bertopic_model.ipynb, provided in repository)
• tweets/topics_per_class_ddr.csv (created by analysis/bertopic_model.ipynb, provided in repository)

• users/users.csv (created by analysis/scrub_data_for_upload.ipynb, provided in repository)
• utilities/state_names.csv (provided in utilities)
• utilities/popular_vote_2020.csv (provided in utilities)

• tweets/combined_US_politician_twitter_timelines_2010-11-06_to_2022-12-31_honesty_component_scores_glove_singleword.csv.gzip (created by analysis/label_lexicon_single_word.sh, not provided in the repository)