Does Climate Misinformation Affect ESG Performance?

This project contains the replication package for the revised version of the paper, which studies the effect of climate misinformation on firms' ESG performance. The revision replaces the original time-varying-only Composite Misinformation Index with firm-varying misinformation exposure measures (strict and contested variants, with Environmental/Social/Governance pillar components), built from a two-pass LLM re-classification of 530,888 climate-related tweets using the open-weights NVIDIA Nemotron-3-Nano-30B-A3B model (run locally via vLLM; no proprietary classification APIs). Exposure is defined as Misinfo(k)_it = Σ_d m(k)_dt × w_id,t−1: monthly ESG-dimension-specific misinformation intensities interacted with each firm's lagged, within-month-standardized pillar scores. Because the resulting measure varies across firms within a month, the revision's headline specification includes both firm and full calendar-month fixed effects — a design not identified with a time-only index. Headline estimates: standardized coefficients of 0.0154 (strict) and 0.0435 (contested), both p < 0.001, n = 343,537 firm-month observations (5,064 firms, Feb 2012–Nov 2022), driven primarily by the governance channel. The package includes: (1) all derived datasets needed to reproduce every table and figure in the revision, including the firm-month analysis panel and per-tweet model labels keyed by anonymous integer IDs; (2) the complete code pipeline — classification prompts and scripts, exposure-weight and IV construction, and a single analysis script that regenerates all results; and (3) the revision's tables, figures, and run logs. All results were verified by re-running the analysis end-to-end from the packaged data (all 20 tables reproduce exactly). Raw tweet text, timestamps, and platform IDs are not redistributed, consistent with platform terms of service. The per-tweet label file contains no text; the README documents how researchers with independent, legitimate access to the underlying tweet corpora can reconstruct the classification input and re-run the full pipeline. ESG scores and pillar scores are from LSEG (Refinitiv) and are provided only in derived, model-ready form.