We Do Not Know the Population of Every Country in the World for the Past Two Thousand Years

TL;DR

Guinnane argues that the widely used McEvedy and Jones (1978) population estimates for every country over the past two millennia are often guesses, outdated, circular, and affected by non-classical measurement error, so econometric results relying on them are unreliable. A replication of Nunn and Qian (2011) shows their potato–population result loses significance when African hard clones are dropped.

Key findings

• MJ’s estimates often lack firm foundations: MJ say there are “almost no data” for Germany until the late Middle Ages, Poland before the 14th century is based on “general ideas,” and for the Maghreb “there is really nothing” before the 19th century, yet they report totals far back. Their rounding rules make measurement error non-classical: below 1 million to nearest .1 million, 1–10 million to nearest .25 million, 10–20 million to nearest .5 million, 20–100 million to nearest 1 million, etc. Monte Carlo simulations show the correlation between simulated true population and rounding error is significant in 72.2–94.3% of experiments depending on year and distribution (Table 2).
• MJ create “soft clones” by assigning similar population dynamics to countries they think are alike—e.g., Afghanistan assumed to have half of Iran’s population before 1900, Kenya and Uganda identical through 1800, and 35% of countries assigned the same growth rate from years 1 to 1000. Economists then create “hard clones” by disaggregating MJ regions into countries; in Nunn and Qian (2011), hard clones account for 76% of African observations, 36% in Europe, and 41% in Asia.
• Replicating Nunn and Qian (2011), the baseline potato × post-1750 coefficient is 0.0325 (SE 0.0116); dropping African clones gives 0.0296 (0.0172), no longer significantly different from zero; dropping all Africa gives 0.0207 (0.0184). Dropping all clones gives 0.0575 (0.0209).
• Maddison relied heavily on MJ before 1500; he updated 23% of MJ’s observations for year 1000 and 40% for 1500, while in 1850, 84% of 51 observations differ from MJ, with an average absolute difference of 3%. For England and Wales, MJ imply annual growth of .32% in 1600–1650 versus Wrigley and Schofield’s .5%; .26% versus −0.07% in 1650–1700; and .10% versus .26% in 1700–1750.

Methods

• Critical audit of McEvedy and Jones’s *Atlas of World Population History* and its use in economics, examining MJ’s stated methods, rounding rules, and patterns in the MJ database provided by James Fenske.
• Monte Carlo simulations of MJ’s rounding for years 1000, 1500, and 1800 under four distributional assumptions; checks using U.S. state populations 1900–1970 and world country populations 1960–2020; replication exercises for Nunn and Qian (2011) and Ashraf and Galor (2011) with clones dropped; and a survey of every paper citing MJ in the “Top 5” economics journals through 2020.
• No new historical population estimates are produced; the paper uses published replication files, census data, and MJ’s own figures.

Limitations

• The paper states that correct population data are unavailable, so it cannot re-estimate published models with corrected data; the replication exercise is “narrow and specific”; the U.S. state analogy “may not provide a strict analogy” to MJ’s regions; and the simulation addresses only rounding, not MJ’s other flaws.
• Evident from the text: the analysis focuses on economics literature and MJ/Maddison, does not supply alternative population series for most countries, and its hard-clone checks rely on U.S. states and selected replications rather than a full redefinition of all variables.