Losing My ReligionCatholic clergy scandals data
Replication

Code & updated results

Everything needed to reproduce the analysis that data-sharing restrictions allow — plus a re-estimation of the headline event studies using modern estimators developed since the paper was published.

On data restrictions. The scandal records are public and shared in full on the Data page. Some outcome data used in the paper (e.g. religious-participation and charitable-giving micro-data) are licensed and cannot be redistributed. We share all code, and provide access instructions for restricted inputs.
Replicated results

Reproducing the paper (Table 1)

Re-running the original Stata specification on the zip-code panel reproduces the published results. A scandal's effect is split into a short-term (0–3 years) and long-term (4+ years) component — measured per scandal event — with a pre-scandal term as a placebo. Controls: zip and year fixed effects, state-specific year trends, and year interacted with 1990 zip characteristics (Catholic share, population, urbanization, race, education, income); standard errors clustered by zip code.

Effect of a scandal (per event)Religious employees
log · Table 1, col 2		Charitable contributions
log · Table 1, col 3
Short-term (0–3 yrs)−0.011 (0.009)−0.003 (0.003)
Long-term (4+ yrs)−0.030 (0.010)***−0.014 (0.004)***
Pre-scandal (placebo)−0.001 (0.005)+0.004 (0.004)
Observations / zip codes247,585 / 19,045175,280 / 25,645

*** p<0.01. Standard errors in parentheses, clustered by zip. Reproduced by replicate.py from regressions.do (Table 1, cols 2–3).

Event-study figures

Replicating the dynamic event studies — and extending them

The paper's event-study graphs trace a scandal's effect year by year (−5 to +5, with the year before the scandal as the reference), using the same controls as the table. We reproduce them exactly, then overlay two modern estimators robust to heterogeneous treatment timing — Callaway & Sant'Anna (2021) and the Borusyak, Jaravel & Spiess (2024) imputation estimator — on a binary first-scandal design, carrying the paper's controls. (We do not use Sun & Abraham: on the sparse IRS giving panel its interaction-weighting is unreliable, so for consistency both figures use the same two estimators.)

Event study of religious participation: replicated original, Callaway-Sant'Anna, and Borusyak imputation, all showing flat pre-trends and a growing decline
Religious participation (log religious employees). The replicated original (gray) shows flat pre-trends and a decline that grows to ≈ −4% five years out, reproducing the paper. Callaway & Sant'Anna (controls partialled out) and the Borusyak, Jaravel & Spiess (2024) imputation estimator track the same pattern, confirming the original two-way fixed-effects result is robust to the staggered-DiD critiques. (Borusyak is plotted post-scandal only; the imputation uses the pre-period as its baseline.)
Event study of charitable giving: the replicated original, Callaway-Sant'Anna, and Borusyak imputation estimator all show a growing post-scandal decline
Charitable giving (log itemized contributions, IRS SOI). Three estimators agree on a growing post-scandal decline: the replicated original (gray), Callaway & Sant'Anna (controls partialled out), and the Borusyak, Jaravel & Spiess (2024) imputation estimator, which fits the fixed effects and controls on the ~25,000 never-treated zip codes and imputes the counterfactual — well-suited to the sparse SOI panel. (Borusyak is plotted post-scandal only; the imputation uses the pre-period as its baseline.)

Replicated and extended from scratch in es_full.py (the original event-study figures) and es_modern.py (both Callaway-Sant'Anna and Borusyak versions, carrying the paper's controls).