Source
Gary M. Feinman, Gabriela Cervantes Quequezana, Adam Green and six co-authors (2025), “Assessing grand narratives of economic inequality across time”, PNAS 122(16), LIT-087, read in full with its eight-page SI in NOTE-049.
What was actually shown
The data. Site-level Gini coefficients of house floor area from the GINI project’s SiteGiniLevel file: 1,176 sites, 47,019 houses, several world regions, each site-phase dated by the midpoint of its range. Polity predictors (levels of hierarchy, population, governance) are assigned to every site in the polity.
Scale is a weak predictor (Figs. 1–2, S5; values read from the figure panels).
| predictor | all sites | apex | H50 | H100 |
|---|---|---|---|---|
| number of levels | r² 0.158 (n 1,164) | 0.218 (711) | 0.269 (170) | 0.326 (92) |
| log polity population | 0.107 (590) | 0.213 (243) | 0.085 (110) | 0.105 (66) |
All have P < 0.01. The scatters are heteroscedastic fans: the range of Gini widens at three or more levels while the mean rises little.
Who reaches the ceiling. Every apex site in a polity of three or more levels with Gini above 0.75 is in Southwest Asia (Assur, Babylon, Zincirli, Nuzi, Tell el-Amarna), South America (Chan Chan, Galindo) or Europe (Knossos). Collective apex sites have a median Gini of about 0.23 and autocratic ones about 0.41, read from a box plot with no test.
Time. No macroregion shows a consistent subregional trend. The largest upticks postdate the first cities (Southwest Asia after about 3500 BCE, Mesoamerica after about 500 BCE), so inequality did not rise immediately with farming or urbanism (Fig. S2, from the text only).
What could have come out otherwise. Hierarchy and population could have predicted Gini strongly, as stage and Malthusian narratives expect. They explain at most a third of the variance in any subset.
What this does not say
- That governance sets the level. The 0–3 autocracy score does about as well as hierarchy (r² 0.10, 0.18, 0.32), and “Hierarchical Clout”, levels plus inverse governance, reaches 0.30–0.50. But governance was coded by the same regional experts at the same time as the house data, and one of its three dimensions is architecture, so it may partly read the outcome. A rise in r² from summing two predictors is not an interaction, and no two-predictor model was fitted.
- That the P values are right. Sites share polity-level predictor values and are treated as independent in OLS. There is no multilevel model and no control for region, period or house-sample size. The effective n is nearer the number of polities.
- That collapse does not level. Two cases go opposite ways (Roman England down, Teotihuacan up), and the authors say the sample cannot test it.
- That house size is wealth. It is one proxy. A single site can move far on method: Tenochtitlan’s Gini is .13, .30 or .75 by sample adjustment.
- That all grand narratives are “refuted”. The data show scale is not sufficient for high inequality; bivariate correlations cannot carry the paper’s stronger wording.
Connections
- The paper sits against Scheidel’s leveller-only reversals, neo-Malthusian accounts and stage evolutionism. None of these is filed in the record.
- North’s account of institutions (THEORY-429, from LIT-489) is a neighbour in the governance thread: both put the realised outcome in the institutional framework rather than in scale. The record has not read Blanton and Fargher’s collective-action theory, which the governance index simplifies.