A constructive replication of Patel, Worsham, Liu & Jena (2026), "Smartphones, Online Music Streaming, and Traffic Fatalities," NBER Working Paper 34866. [Local PDF]

The detailed response is available as [pdf].

Patel et al. (2026) analyze 10 major album releases from 2017-2022 and report:

• 139.1 deaths on album release days vs 120.9 on control days (+18.2 deaths, +15.1%)
• 123.3M streams on release days vs 86.1M control (+43%)
• Proposed mechanism: smartphone distraction from streaming while driving

"We find an additional 18.2 traffic fatalities (139.1 versus 120.9; p < 0.01) on album release days compared to control days..." — Patel et al. (2026), Figure 2B

We successfully replicate the paper's main result:

• Difference: 0.6 deaths (< 1 SE)
• Same methodology: week-of-year fixed effects, day-of-week, year, holiday indicators
• Same sample: Tier 1 albums, 2017-2022
• Randomization inference confirms significance (p < 0.001)
• Note: Our SE uses cluster-robust estimation (clustering by album) to account for only 10 independent units

The statistical effect replicates. The question is how to interpret it. See t12_paper_replication.md.

Several findings are consistent with the paper's claims.

We test 35 specifications varying window size, sample period, and album set:

Across the multiverse of reasonable analytical choices, the effect is remarkably consistent. See t29_multiverse.md.

The effect is concentrated on day 0:

The sharp spike on release day (with some spillover to day +1) is consistent with an event-specific effect. See t13_dynamic_effects.md.

If distraction (not alcohol) drives the effect, sober crashes should show a larger effect:

The effect is 4x larger for sober crashes, consistent with the distraction mechanism rather than alcohol-related confounding. See t22_drunk_mechanism.md.

The effect is robust to weather controls:

See t21_fars_controls.md.

Friday album releases are not cherry-picked by the researchers. Since 2015, Friday has been the global standard release day for new music ("New Music Friday"), established by the IFPI. All 9 Friday releases in the study follow this industry norm.

The estimated +16-18 deaths represents approximately a 15% increase over baseline. For comparison:

• Major holidays show similar or larger effects (July 4th, Labor Day weekend)
• The magnitude is plausible given that ~100 daily fatalities × millions of concurrent streamers × some fraction driving

Several findings raise questions about interpretation.

9 of 10 albums in the study were released on Friday. This creates an identification challenge:

We test this directly: randomly selecting ANY 10 Fridays from the 939 available in 2017-2022:

The observed effect (16.1) is below average for random Friday selection. However, this does not rule out a streaming effect—the Friday coincidence is suspicious but album releases genuinely do occur on Fridays. See t18_friday_fpr.md and t24_balance_check.md.

The effect does not persist in the post-2022 period. Possible explanations include: changes in streaming behavior, smaller sample size (n=7), or that the original finding was a chance result. See t20_extended_series.md.

If streaming causes distracted driving deaths, more streams should produce more deaths:

Pearson r = -0.17 — the correlation is in the wrong direction. This could reflect a ceiling effect (all albums have enough streams to saturate the driving population) or measurement error in first-day streaming data. See t03_dose_response.md.

We apply the same methodology to outcomes that should not respond to album releases:

Additionally, the joint F-test for pre-trends is significant (p = 0.03), and day -6 shows a large spike (+16.5 deaths, t = 3.54) before any album release. These suggest the methodology may be sensitive to noise. See t28b_structural_fars_placebos.md and t32_parallel_trends.md.

This analysis cannot definitively resolve the causal question:

• The Friday coincidence is suspicious but does not prove the effect is spurious—albums genuinely release on Fridays
• The effect could be real but smaller than estimated due to selection on high-Friday dates
• The post-2022 reversal could reflect changes in streaming behavior rather than refuting the original finding
• We cannot distinguish between "streaming causes crashes" and "Friday baseline is high"
• With only 10 albums in the core sample, all inference is inherently uncertain

The most honest conclusion is that the paper's finding is fragile but not definitively refuted.

Bottom line: We successfully replicate the paper's statistical finding (+17.6 vs +18.2 deaths). The effect is robust across 86% of reasonable specifications, concentrated on release day, and larger for sober crashes. However, Friday selection bias remains a concern (100% FPR), the effect does not persist post-2022, and some placebo tests show unexpected results. The finding is fragile but not refuted.

• FARS: NHTSA Fatality Analysis Reporting System
• Streaming data: Spotify Newsroom, Billboard, Chart Data (see albums_sources.md)

# Install dependencies
pip install pandas numpy matplotlib scipy requests scikit-learn

# Run analysis
make extract        # Extract FARS CSVs from zips
make run            # Run main analysis

# Extended analysis (includes 2023-2024 albums)
python3 -m src.pipeline --extended

1. Download FARS zip files from NHTSA → data/raw/
2. Run make extract to extract accident CSVs
3. Album data in data/albums.csv with sources in data/albums_sources.md

farce/
├── Makefile
├── README.md
├── w34866.pdf              # Paper
│
├── data/
│   ├── albums.csv          # Album release dates & streams
│   ├── albums_sources.md   # Data provenance
│   ├── fars/               # Extracted accident CSVs (not tracked)
│   └── raw/                # FARS zip files (not tracked)
│
├── src/
│   ├── constants.py        # Load albums from CSV
│   ├── pipeline.py         # Main entry point
│   └── ...                 # Analysis modules
│
├── tabs/                   # Output tables (Markdown)
└── figs/                   # Output figures (PNG)

• Patel, Worsham, Liu & Jena (2026). "Smartphones, Online Music Streaming, and Traffic Fatalities." NBER Working Paper 34866. [PDF]
• Harvard Gazette coverage
• Freakonomics podcast
• New York Times