Hummers or Hybrids — Replication & Extension

A replication and extension of Kahn (2007), and an experiment in using Claude as a full research assistant — from data acquisition through analysis, writing, and peer review response.

Source Code

Hummers or Hybrids is two things at once: a replication and extension of Kahn (2007), “Do Greens Drive Hummers or Hybrids?”, and an experiment in using Claude as a primary research assistant on an empirical economics project.

The research question is substantive — does the “greens drive hybrids” finding hold with modern data, and did Elon Musk’s political transformation change the Tesla–ideology link? But the meta-question is equally interesting: how far can an AI assistant take a research project, and where does it fall short?

Claude (via Claude Code) handled the full pipeline: writing and running all data acquisition scripts, building geographic crosswalks, constructing the ideology index via PCA, running panel regressions and event studies, producing figures, drafting the paper, and even responding to a simulated peer review. The collaborator (John Morehouse) contributed the first-difference event study extension and the LaTeX paper draft. The PI directed the research design, verified outputs, and made all substantive judgment calls.

On using Claude as a research assistant: The workflow surfaced real tradeoffs. Claude is fast and capable at scaffolding an empirical pipeline — it can write clean acquisition scripts, handle geographic crosswalk logic, and produce readable regression tables without much prompting. It requires careful oversight on data integrity (the project CLAUDE.md enforces a strict no-fabricated-values rule) and makes opinionated choices about specification that need PI review. It is not a substitute for domain expertise, but it meaningfully compresses the time from research design to working code.

Research questions:

  1. Replication — Do greener California communities still exhibit lower-carbon transportation behavior in 2024? (Yes — the core Kahn finding holds.)
  2. EV extension — Does climate ideology predict EV ownership, and does it predict Tesla vs. non-Tesla EVs differently?
  3. The Elon Effect — Did the ideology–Tesla correlation shift after Elon Musk’s Twitter acquisition (Oct 2022) or his role in the Trump administration (2025)?
  4. Status signal migration — Has the “green status signal” moved from Tesla toward other EV brands?

Key findings:

  • The Kahn replication holds robustly: tracts with stronger climate beliefs have significantly more EVs, higher transit use, and lower truck registrations, consistent with Kahn’s original results
  • No Elon Effect detected through 2024: the ideology–Tesla coefficient is stable across the full 2018–2024 panel; non-Tesla EVs, by contrast, show a rising ideology gradient — “democratization” of EV adoption across the ideology spectrum
  • Results are robust to three alternative ideology specifications (county/YCOM, tract/no-YCOM, Prop 30 share only) and survive spatial correction (Moran’s I = 0.58; SAR ρ = 0.78)
  • Strong spatial autocorrelation in residuals confirms the geographic clustering of EV adoption — neighboring tracts’ ideology spills over into local EV markets

Ideology construct: Rather than Green Party registration (Kahn’s proxy), we build a composite Climate Ideology Index via PCA across Yale Climate Opinion Maps (YCOM), Democratic–Republican voter registration share, and California environmental ballot measure vote shares (Prop 30 2022 and others). PC1 explains 84.7% of variance.

Data: California Energy Commission ZEV registration panel (2018–2024), ACS 2023 5-year estimates, Yale Climate Opinion Maps (county), California Secretary of State voter registration and Statement of Vote (precinct-level), TIGER shapefiles for geographic crosswalks.

Tech: Python, pandas, geopandas, scikit-learn (PCA), statsmodels, linearmodels (two-way FE), pysal (spatial regression), matplotlib