Accounting for uncertainty and bias in archaeological and historical evidence on wealth inequality Authors: Mattia Fochesato, Samuel Bowles Published: 2026 ABSTRACT SignificanceIn our era of fundamental changes in technology and institutions along with rising inequality, our assessments of likely future trends and policy options to address inequalities may be enlightened by the historical and prehistoric record of similarly vast changes. But the available data and how it came to be available bears with it the cost of substantial but unknown levels of uncertainty. Going beyond conventional bootstrapping and integrating a suite of statistical tools we provide a comprehensive method and code for estimating levels of bias and uncertainty in the available data on past wealth inequalities and how these uncertainties propagate through every stage of the estimation process. Our estimates of past wealth inequality are probability distributions over likely values. The many insights provided by recent quantitative studies of economic inequality in the distant past have unavoidably been seen as bracketed by substantial error bars (1–11). What Oskar Morgenstern termed “the fact of error” (12) is endemic to research in this area; but convincing estimates of the associated uncertainty are rarely available. The point estimates of inequalities in wealth or other economic metrics on which these and related studies are based are often taken as the raw data on which hypotheses can be tested using conventional measures of statistical significance. But these point estimates are themselves each subject to uncertainty that in many cases is substantial. The uncertainties in these data arise in part from the limited number of mostly small-n datasets that are characteristic of archaeological evidence, but substantial uncertainty is also present in estimates based on historical data, especially for sites and dates prior to the introduction of population-wide taxation by modern national states. In all cases uncertainty is compounded by adjustments to the raw data that are required to make them representative of the underlying populations of interest and comparable across sites and dates. These adjustments are themselves subject to uncertainties beyond the uncertainty of the raw data itself, which then propagate through the estimation process. Ignoring these uncertainties is to claim “specious accuracy” (another Morgenstern expression). Recognition of the extent of uncertainty surrounding point estimates of ancient wealth inequalities could reasonably lead researchers to eschew quantitative studies on these topics. Or it could motivate the creation of probabilistic measures of wealth inequality based on quantitative estimates of the uncertainties, even if these estimates are themselves highly uncertain. We here take the second course, implementing a suggestion made by Arthur Bowley and his coauthors over a century ago. He urged the Royal Statistical Society to collect and present data on the distribution of household incomes so as “to tabulate our knowledge or ignorance … as definitely as possible” thereby allowing “a measurement of precision of the aggregate. … We are bound to make estimates in all cases, whatever our ignorance may be” and so Bowley explained, “we have endeavoured to assign, in light of all the information available…limits within which it seems highly probable that the true measurement must lie.” (13, p. 46–47) Pioneering works in the measurement of national income by Simon Kuznets and others (14, 15) followed Bowley’s advice. But his approach was never widely used by economists (perhaps because of improved measurement of the components of national income). Nor have the methods Bowley advocated been adopted by archaeologists or historians studying past inequalities, notwithstanding the “plea for errors” made by the economic historians Charles Feinstein and Mark Thomas (16).* Here we provide methods (and code) for estimating probability distributions of likely values of the Gini coefficient of material wealth—a standard measure of inequality (20)—at a particular site and date. We illustrate these methods, drawing on wealth data from throughout the world over the past 12,000 y (21). An example of one of these distributions—from Akkadia in ancient Mesopotamia—is in Fig. 1. When we refer to “the Gini coefficient for third millennium BCE Akkadia,” it is this distribution that we mean. Here and throughout, the distribution-free CI that we report are the least span encompassing 95% of the probability mass of the density function, referred to as CI*’s to distinguish them from conventional (symmetric-around-the mean) CI. Further information on our methods underlying Fig. 1 and the other results reported below appear in SI Appendix.† Probability density function of the adjusted Gini coefficient in Akkadian Mesopotamia (2350–2150 BCE). Solid line: mean (0.700); dashed lines: CI* bounds [0.510, 0.878]. Skewness = 0.089. These distributions differ conceptually from conventional bootstrapped estimates based on repeated sampling with replacement of a given dataset in that we have leveraged additional information on what is known or can be inferred about how the data came to be available and other independent data. Not surprisingly, in addition to our distributions not being Gaussian by construction, they also differ quantitatively, in most cases returning much larger estimates of uncertainty than estimates based on bootstrapping methods. Our concerns in this paper are entirely statistical, addressing matters of bias and uncertainty rather than substantive questions concerning the Gini coefficient as a measure of inequality, wealth disparity as an indicator of societal inequality, or causal mechanisms likely to account for the patterns of wealth inequality that we observe, topics that we have addressed in related papers (10, 21). BRIDGE: Bayesian-Resampling and Informed Priors with Data-Driven Gini Estimation By material wealth (hereinafter, “wealth”) we mean alienable assets (e.g., tools, livestock, stored goods, durable household goods, and land) that provide a flow of services (income, reproductive success, or other) contributing to the living standard of a household or other ownership unit, as valued by its members. To measure wealth inequalities, we would ideally have population-level data from randomly selected sites and dates, measuring a suitably aggregated measure of wealth, or at least a form of wealth that is common across our temporal and geographical scope, held by consistently defined households or other wealth-sharing units. What instead we have at our disposal are measurements of different forms of wealth, some held by individuals, others by households, often from small samples that in many cases are unlikely to be representative of the population of interest, that are distributed across time and space according to the interests and available resources of researchers, and that are from underlying populations vastly differing in scale (and hence in degree of heterogeneity in wealth relevant characteristics). The corrections and adjustments on which our dataset is based are an attempt to estimate both what the researcher would have found if they had been able to collect the ideal data as well as the degree of uncertainty due to the fact that this was not the case. In many cases, the uncertainty in our estimates arises because the archaeologist, tax collector, or other source of our raw data very often had access to only a subset of evidence, representing a sample of the larger population of interest, whose complete discovery is unlikely or impossible. To estimate the resulting biases and uncertainties we attempt to replicate what might have been the source’s search process, sampling from a more complete set of data to estimate what the researcher would have most likely “found” and “not found.” Our methods necessarily differ depending on the extent to which it is likely that the data that have come to light are a random sample of the population of interest. For example, to estimate the uncertainties associated with small sample size per se (rather than the representativeness of the sample) we proceed as follows. We use a large dataset which we consider to be the relevant hypothetical entire population, and then estimate Gini coefficients from arbitrary subsets of these data of differing sizes that could represent the limited information found by the researcher. We use what we call this not found exercise to estimate the uncertainty and bias of estimates for each possible sample size up to a complete enumeration. Then we use the distributions generated for each of the sample sizes in our dataset to estimate the likely bias and degree of uncertainty of our estimates from other datasets where we are limited to small samples without evidence from a larger (hypothetically complete) dataset. But adjusting for the bias and uncertainties of small random samples is far from sufficient. Suggestive of the nonrandom nature of the search procedures that generated our data is the fact that in a set of ancient Mesopotamian data based on excavations, the frequency of public architecture is twice that of domestic architecture, while in satellite-based data on the same sites, the reverse is true (22). Some of our datasets exhibit glaring evidence that the sample is not representative of the population of interest. A sample entirely of households with substantially above average wealth is probably from a wealthy neighborhood of an unobserved larger settlement, for example. Similarly a sample of excavated house sizes including a palace and the dwellings in a small fraction of the remainder of the occupied area of a settlement cannot alone provide the data for Gini coefficient that is informative about the level of wealth inequality in the settlement as a whole. In these and similar cases we have either excluded the evidence entirely as unrepresentative or used the available evidence to simulate a more complete sample, for example using the available data on commoner housing to simulate a distribution of house sizes in the entire occupied (but not yet excavated) area of the settlement. A related problem arises where the observations are from small settlements that are not representative of the scale of our benchmark population. In these cases we use additional datasets allowing estimates of the degree of bias and uncertainty entailed by sampling from just one or a few of the many subunits making up a settlement of our benchmark scale. Here we provide an overview and the conceptual framework of our adjustment method. We identify five sources of potential bias and imprecision that may drive a wedge between the observed and true value: sample bias from small datasets, difference in wealth ownership unit (individual- vs. household- level wealth), missing data for households without wealth (or more generally, the less wealthy), the appropriate aggregation of different asset types, and taking account of different population scales. For each of these five sources of bias and uncertainty, we design an adjustment procedure that specifies both the parameter capturing the bias or imprecision, and the distribution of that parameter. In each case, the choice of method depends on the nature of the adjustment and the availability of additional data that may contribute to an estimate of the likely degree of bias and uncertainty. For the problem of sample bias in small datasets, we derive the parameters for the adjustment and their uncertainty using nonparametric bootstrapping of the relevant sample sizes, applied to large available datasets to represent the “true, entire” population. This allows us to derive parameters to adjust prehistoric data from actual archaeological datasets, in contrast to parametric bootstrapping, since we resample directly from an alternative archaeological dataset, not from a dataset constructed from a parametric model. When Gini coefficients are estimated from individual-level data (rather than household, our benchmark), we use a Bayesian empirical likelihood (BEL) approach to move from the individual Gini coefficient to the household Gini. We pair individuals into couples using a prior distribution based on historical information from a different, but plausibly similar site. Since a likelihood is not available in closed form, we replace it with its empirical estimate. Combining the prior information with this empirical likelihood we compute the posterior, which can be used to estimate the household Gini. For the problem of the missing households without wealth, differing assets, and population scale adjustments, we specify a prior distribution informed by additional historical and archaeological data. As in the couples’ case, the likelihoods are not available in closed form, so we use their empirical approximation to construct the posterior distribution of the adjusted Gini. We hypothesize that these sources of bias and imprecision define a relationship between the Gini coefficient that one can compute directly on the available wealth data (including those computed by other scholars), called here Gi, for any past society and date i, and our best estimate GiA, adjusted by the above biases and comparable across different time and space. We model this relationship by the following equation[1]GiA=Gi(1λiμiδi)(1−zi)+zi+ki, where as is appropriate for each site and date (using methods to be explained in the sections that follow), λi is the factor to adjust for the sample to population bias just mentioned, and μi, zi, δi and ki are the factors to adjust for, respectively, differences in the wealth ownership unit, inclusion of those without wealth missing from the dataset, differences in the wealth asset type, and differences in population scale. We use this relationship to adjust all the raw Gini coefficients in our dataset and transform them into (ideally) unbiased, representative, and comparable measures, adopting a hypothetical benchmark case with a common asset type (household wealth), unit of observation (household), and population size (1,000 households). A probability distribution of Gini coefficients for a particular site and date (analogous to that shown in Fig. 1) is then estimated using SE-propagation methods, conditional on the errors arising from the sources identified in Eq. 1 being uncorrelated. For each site and date we use Eq. 1 to compute 10,000 values of the Gini coefficient, each based on the raw wealth data and a random draw from the probability distribution of each of the five adjustment parameters of Eq. 1. These probability distributions are shown in Fig. 2. Table 1 lists the adjustments and additional data used implementing them. Uncertainty of the raw data and adjustment factors. (A and B) Frequency distributions of the sample bias adjustment factors from the 1427 Florence and Columbia Plateau datasets. Dashed lines: mean value at each sample size. (C) Probability distribution of the marital adjustment factor and its mean, both computed merging the four archaeological cases. (D) Probability distributions for the fraction of missing zeros used for Ancient Mesopotamia (red) and Roman Empire (blue). Dashed lines: mean values of each distribution. (E) Probability distributions for the asset types factor to convert grave goods inequalities and house area into household wealth inequalities. Table 1.Summary of adjustments and sources Adjustment Parameter in Eq. 1 Section Method Data used Adjustment applied to Sample Size λ 3BootstrappingHohokam 11th c. CE; Florence 1427 CE; Columbia Plateau 10th–18th c. CE; Krummhörn 17th–18th c. CEAll estimates with sample size smaller than 10,000 (n = 380)Individual to household ownership unit μ 4BELColumbia Plateau 10th–18th c. CE; Krummhörn 17th–18th c. CEAll estimates computed on individual-level (rather than household) data (n = 48)Asset type δ 7BELWestern Balkans (Gomolava) 5th millennium BCE; South Mesopotamia (Kafajah and UR) 2nd millennium BCEHouse size and burial goods data (n = 204)Including missing zeros z 5BELSee SI Appendix and textAll datasets with missing zeros (n = 127)Scale k 6BELHohokam 11th CE; Columbia Plateau 10th–18th CE; Finland 1571 CE;All estimates computed on sites with less than 250 households (n = 240)*BEL: Bayesian Empirical Likelihood estimation. Summary of adjustments and sources *BEL: Bayesian Empirical Likelihood estimation. We call our method BRIDGE, an acronym that summarizes our approach and provides a basis for scholars interested in building on our work to account for bias and uncertainty in prehistorical and historical data. Our method exploits different statistical approaches but it is inherently Bayesian in nature, as we use datasets to estimate the distributions of correction parameters and we then sample from these distributions to propagate uncertainty, although, due to the complexity of the likelihood, we replace it with its empirical counterpart. Our approach resembles BEL as we empirically approximate the likelihood from the data, although we do not use a formal BEL framework. We use bootstrapping in only one of the five adjustments (Table 1), our assessment of the uncertainty due to small sample size. Finally, we propagate the parameter’s distribution numerically through an error propagation model. The propagation generates the posterior distribution of the true Gini coefficient and the associated 95% credible interval. A key feature of our method is that when we estimate the parameters for the adjustments, we draw from archaeological datasets that are different from the ones whose inequality measure we adjust. Our approach is, therefore, forward-looking: given an observed Gini and independently estimated adjustment distributions, we propagate uncertainty numerically to obtain the distribution of the true Gini. By using information from other distributions we differentiate ours from other methods that derive the parameters by iterating the observations in the samples to be adjusted. By using error propagation techniques, our work also broadly relates to recent developments in environmental research, where similar models are employed to estimate how error and uncertainties spread through complex and interrelated environmental, climatic, and hydrologic models (23–25). Our context and data structures are different than these cases, but our work similarly suggests that implementing an error propagation model captures a larger degree of uncertainty than the one that would be obtained by looking only at the errors from single stages of the analysis. Bias and Imprecision due to Small Sample Size (λ) We start by exploring the degree of uncertainty and bias in estimates of wealth inequality that would arise from random sampling in what are often very small-n datasets. We are motivated by a challenge faced by archaeologists: the fact that the materials they recover from a site provide evidence on a typically relatively small sample of a larger population that has not been studied or perhaps even discovered. Excepting the past century or so, which in many countries saw the virtually universal taxation of households, this problem is likewise common in historical sources. With few exceptions (one is the survey of the wealth of all households by the tax authorities in 15th century Florence that we use below) records of households’ wealth are very restricted (either deliberately or unavoidably) to relatively small samples of the population of interest. To address the uncertainties and biases arising due to small samples, we make use of large datasets like the wealth data from Florence, and their analogues in the archaeological record. We suppose that the complete set of observations, Ω (of size M), is the whole population, and the Gini coefficient computed on Ω is the true Gini coefficient, G(Ω). We then repeatedly restrict the dataset to a set ωm (of size m, where m