BIG DATA, DEEP ROOTS — SURNAME AGGREGATES
Van Pelt & Kirkegaard (2026), Comparative Sociology.
https://doi.org/10.1163/15691330-bja10162

These 10,989 surname records are the processed data used by the site's game and explorer, at the precision stored in the game. They are not the full raw record-level research datasets. No photos or individual records are included. Surname-level estimates describe groups in the source samples, not any individual bearing a name. Coverage and selection into each source differ.

CSV: one row per surname, UTF-8, header row. Empty cells are missing, not zero.
JSON: metadata and compact surname rows. Missing numeric values are null.
Tuple positions (zero-based):
0 surname; 1 ancestry proportions; 2 observed estimates; 3 observed percentiles;
4 ancestry-predicted estimates; 5 predicted percentiles; 6 Census [count, shares] or null;
7 standard errors; 8 lower and 9 upper 95% confidence limits;
10 lower and 11 upper percentile confidence limits;
12 grouped ancestry proportions; 13 ancestry PCA [PC1, PC2].
Array order is defined by ancestries, metrics, censusLabels, and clusters in JSON metadata.

CSV METRIC COLUMNS
*_estimate, *_percentile, *_predicted, *_predicted_percentile, *_se, *_ci_low, *_ci_high, *_ci_low_percentile, *_ci_high_percentile
The prefix identifies a metric. Estimates, standard errors and raw confidence limits use the same metric scale; percentiles use 0–100. Percentiles compare surnames with available observations for that metric, not people. Higher criminal-record percentiles mean MORE representation, not better outcomes. Percentiles and confidence limits are rounded as in the game.

METRICS
s_factor: composite socioeconomic score combining government salary, occupational prestige, physician licensure and reversed criminal-record representation. Wikipedia is excluded.
salary: adjusted standardized government-pay measure, NOT dollars.
opr: adjusted government occupational-prestige measure, on the supplied model scale.
physician and crime: log10 representation residuals adjusted for surname frequency. 10^estimate gives representation relative to expected counts. These are not individual probabilities.
Predicted S factor uses XGBoost; other predictions use linear models based on ancestry.
Confidence intervals describe model estimates and do not account for all source-selection biases.

ANCESTRY AND CENSUS
ancestry_*: 45 source ancestry proportions (0–1); may not sum exactly to 1 due to rounding and source coverage.
cluster_*: 23 grouped ancestry proportions used in the game. JSON clusterLabels supplies display names.
census_count: surname population count from the source Census surname table, not the number of observations used for each outcome.
census_share_1: White (percent, 0–100)
census_share_2: Black (percent, 0–100)
census_share_3: Hispanic (percent, 0–100)
census_share_4: Asian/Pacific Islander (percent, 0–100)
census_share_5: American Indian/Alaska Native (percent, 0–100)
census_share_6: Two or More Races (percent, 0–100)
ancestry_pc1/pc2: PCA coordinates combining standardized ancestry proportions and Census shares, used by the game. These are distinct from the socioeconomic S factor.

All original aggregate values are preserved. See the paper and site supplement for methods and limitations.