In the Weeds of Traffic Fatalities: Replication Dataset
Description
This replication package accompanies the article “In the Weeds of Traffic Fatalities: Revisiting the Effect of Medical Marijuana Laws.” The research re-evaluates the widely cited finding that medical marijuana laws (MMLs) significantly reduce traffic fatalities. The central hypothesis is that previous estimates of MML effects may be biased due to unaccounted-for pre-treatment trends and hard to interpret because of heterogeneity across states. The dataset is a panel of U.S. states from 1990 to 2010, constructed to closely replicate Anderson et al. (2013). It includes annual, state-level traffic fatality rates (log-transformed per 100,000 population), a binary indicator for MML adoption, and a rich set of covariates covering demographics, driving laws, traffic enforcement measures, and substance-related policies. The key finding is that states legalizing medical marijuana were already experiencing declining traffic fatalities before legalization. When accounting for these pre-trends using the Imputation Procedure (Borusyak et al., 2024), the estimated effect of MMLs shifts from negative to either zero or positive—depending on included covariates. The data also reveal large heterogeneity across states, with California disproportionately influencing population-weighted estimates.
Files
Steps to reproduce
Data Sources: Traffic Fatalities: Annual traffic fatality counts by state and year were obtained from the Fatality Analysis Reporting System (FARS) maintained by the National Highway Traffic Safety Administration (NHTSA). Fatalities were normalized by state population and log-transformed. Population Data: State population estimates were sourced from the U.S. Census Bureau. MML Treatment Data: Dates of medical marijuana law (MML) are based on existing literature (including Anderson et al. 2013). A binary treatment variable (MML_cut) equals 1 if an MML was in effect for at least half the year. State-level covariates: mean age — U.S. Census; unemployment rate — BLS; real income — BEA; vehicle-miles per licensed driver — Federal Highway Administration; graduated driver licensing laws — Dee et al. (2005), Insurance Institute for Highway Safety; primary and secondary seat belt laws — Dee et al. (2005), Insurance Institute for Highway Safety; speed limits above 70 mph — FARS; texting and handheld device bans — McCartt (2014), HandsFreeInfo.com; .08 BAC laws — Freeman (2007); administrative license revocation laws — Fell and Scherer (2017); zero-tolerance underage drunk-driving laws — NHTSA; drug per se laws — norml.org; beer tax — Brewers Almanac, Beer Institute; marijuana decriminalization — norml.org Data Construction: Three datasets (available in this package) — fatalities/population, MML timing, and covariates — were merged by state and year. The resulting panel includes all 50 states and Washington, DC. Statistical Methods: TWFE (Two-Way Fixed Effects): I estimate the TWFE regressions using reghdfe, absorbing state and year fixed effects and clustering by state. Specifications vary by inclusion of covariates and state-specific time trends. Imputation Procedure: I apply the Borusyak et al. (2024) imputation method using did_imputation. I implement this both with and without covariates and unit-specific trends. Pre-Trend Checks: To assess differential trends, residuals from each method are regressed on a linear time trend measure in the pre-treatment period. Software & Code: Replication code included in this package allows complete reproduction of results. Additional information can be found in the published manuscript: http://dx.doi.org/10.2139/ssrn.5232936.