External hire premium

Methods

Where the data comes from, how the sample was built, and what was controlled. All public, all reproducible.

Principle

Public data, frozen sample, controls before claims.

Every input is a public filing or a public price. No licensed database is required. The sample was fixed before any financial data was collected, so the rules could not be tuned to the result. The strong claims are the ones that survive controls.

Data sources

Four public feeds

  • Pay. SEC proxy statements (DEF 14A), summary compensation table
  • Operating performance. SEC XBRL company facts, for ROA
  • Stock performance. Public daily prices, for TSR
  • Membership. Point-in-time S&P 500, so no survivor bias

Sample construction

611 transitions

Every classified S&P 500 CEO change from 2010 to 2022: 453 internal and 158 external, a 25.9% external rate. A further 149 transitions were excluded by pre-set rules. The classification was frozen before enrichment.

Measures

Pay, stock, operating, tenure

Premium is the log ratio of new to departing pay. Stock performance is market-adjusted TSR. Operating performance is ROA, net income over total assets. Tenure is time in the job to departure.

Controls and tests

What was held constant

Robust (HC1) standard errors throughout. Controls for firm size, year, and two-digit industry. Returns are read in event time, so a window can be set after the hire. Tenure uses survival models built from the full transition chain.

Reproducibility

No licensed database

The whole pipeline runs from free sources. Collect once, commit the snapshot files, and the analysis runs from one script. The code and the snapshots are in the repository.

Coverage and limits

What the data can and cannot carry

  • Pay. 240 transitions with both new and departing pay. The premium estimate is limited by this coverage, not by the sample size.
  • Operating performance. 461 transitions with pre and post ROA.
  • Stock performance. 443 transitions with computed TSR.
  • Imprecision is structural. Pay is heavy-tailed, so the median is stable and the mean is noisy. Adding recovered cases loosened the estimate rather than tightening it, so the clean estimate was kept and the sample was not padded to chase a p-value.