Latent Estimation of GDP, GDP per capita, and Population from Historic and Contemporary Sources

Christopher J. Fariss, Charles Crabtree, Therese Anders, Zachary Jones, Fridolin Linder, Jonathan N. MarkowitzarXiv 2017 · DOI · arXiv

Abstract

The concepts of Gross Domestic Product (GDP), GDP per capita, and population are central to the study of political science and economics. However, a growing literature suggests that existing measures of these concepts contain considerable error or are based on overly simplistic modeling choices. We address these problems by creating a dynamic, three-dimensional latent trait model, which uses observed information about GDP, GDP per capita, and population to estimate posterior prediction intervals for each of these important concepts. By combining historical and contemporary sources of information, we are able to extend the temporal and spatial coverage of existing datasets for country-year units back to 1500 A.D through 2015 A.D. and, because the model makes use of multiple indicators of the underlying concepts, we are able to estimate the relative precision of the different country-year estimates. Overall, our latent variable model offers a principled method for incorporating information from different historic and contemporary data sources. It can be expanded or refined as researchers discover new or alternative sources of information about these concepts.

Introduction

Gross Domestic Product (GDP) and population are the fundamental metrics of social science. They serve as the primary independent variables in studies ranging from the causes of civil conflict to the drivers of democratization. Despite their ubiquity, the data underlying these concepts is often fraught with measurement error, strategic manipulation, and significant gaps in historical coverage. Researchers typically rely on a single dataset—such as the World Bank’s World Development Indicators or the Maddison Project—without accounting for the inherent uncertainties or the systematic biases present in these sources.

This paper introduces a dynamic, three-dimensional latent trait model designed to synthesize disparate data sources into a unified, consistent estimation of GDP, population, and GDP per capita. By treating these variables as unobserved latent traits, the model provides a principled way to quantify uncertainty and bridge the gaps in our historical record from 1500 A.D. to 2015 A.D.

The Three Hurdles of Economic Data

The authors identify three primary challenges that plague existing economic and demographic datasets:

  1. Strategic Bias: National governments often have political incentives to manipulate their reported statistics. Some may over-report growth to attract foreign investment or bolster domestic legitimacy, while others might under-report to remain eligible for international aid or concessional loans.
  2. Capacity Gaps: Many developing nations lack the bureaucratic infrastructure required to conduct frequent, accurate censuses or to update the "base years" used for economic calculations. This leads to "gap-filled" data that relies on outdated assumptions.
  3. Dataset Inconsistency: Major datasets frequently disagree on the same country-year units. These discrepancies often stem from different methodologies for handling the informal "shadow" economy or different approaches to historical reconstruction.

Methodology: A Dynamic Latent Trait Model

The core innovation of this research is the move away from simple interpolation toward a dynamic latent trait model. Rather than treating GDP, population, and GDP per capita as independent variables, the model recognizes their deterministic relationship:

$$\text{GDP per capita} = \frac{\text{GDP}}{\text{Population}}$$

By modeling these three dimensions simultaneously, the framework leverages information from one variable to refine the estimates of the others. For example, a high-quality population estimate can help constrain the possible range of GDP per capita even when direct economic data is sparse.

Technical Implementation

The model utilizes Gaussian link functions to map observed data from various sources into a shared latent space. The estimation is performed using Bayesian inference via Markov Chain Monte Carlo (MCMC) simulations. A key feature is the dynamic structure: the model assumes that a country's economic state in a given year is centered on its state from the previous year. This "random walk" approach allows the model to "bridge" years where data is entirely missing by propagating information forward and backward through time.

Data Synthesis and Coverage

The model integrates several major historical and contemporary datasets, including:

  • The Maddison Project: A primary source for long-term historical GDP and population.
  • World Development Indicators (World Bank): The standard for contemporary economic metrics.
  • Gleditsch (2002): Expanded trade and GDP data.
  • Broadberry & Klein and Bairoch: Specialized datasets focusing on European historical accounts.
  • Correlates of War (COW): Specifically for historical population estimates.

The resulting unified dataset covers 35,299 country-year units. Notably, the model generates estimates for 16,145 units that were previously missing in the original source datasets, providing a significantly more complete picture of global development.

Key Insights: Uncertainty as a Feature

A major contribution of this framework is the provision of posterior prediction intervals. Unlike traditional datasets that provide a single point estimate, this model quantifies the precision of every data point. This "uncertainty metric" tells researchers exactly how much they should trust a specific country-year data point.

Identifying Systematic Bias

The model reveals that different organizations exhibit "institutional biases." For instance, World Bank GDP figures are systematically higher than the latent estimates, while certain historical sources tend to fall below the median. By identifying these patterns, the model allows researchers to understand how specific data-gathering methodologies might skew their results.

Information Density and Precision

The precision of the latent estimates is a function of "information density." When a country-year has data from multiple overlapping sources, the model can triangulate a highly precise estimate. Conversely, when data is sparse, the "error bars" naturally widen, signaling to researchers that the data point should be treated with more caution.

Validation and Results

The latent traits demonstrate extremely high convergent validity, with correlations near 0.99 when compared to the original observed variables. However, the latent estimates are more "honest" representations of the data because they incorporate the uncertainty inherent in the observations. Approximately 86% of all observed country-year values fall within $\pm 1$ standard deviation of the model’s predicted range, suggesting a high degree of predictive accuracy.

Limitations and Future Directions

Currently, the model assumes a one-to-one relationship between changes in the latent trait and changes in the observed variables (fixing the discrimination parameter $\beta$ to 1). Future iterations could relax this assumption to determine which datasets are more sensitive to economic shifts.

Additionally, the authors suggest incorporating "covariates of reliability." By including measures of state capacity, bureaucratic quality, or corruption, the model could automatically "down-weight" data from regimes known to manipulate statistics for political gain. This would further enhance the model's ability to provide an objective view of global economic history.

Conclusion

The latent variable model presented here offers a principled, extendable framework for economic and demographic estimation. By synthesizing 500 years of data and accounting for measurement error, it provides a more robust foundation for social science research, transforming "messy" historical records into a rigorous, unified dataset.

Summary updated

Abstract

The concepts of Gross Domestic Product (GDP), GDP per capita, and population are central to the study of political science and economics. However, a growing literature suggests that existing measures of these concepts contain considerable error or are based on overly simplistic modeling choices. We address these problems by creating a dynamic, three-dimensional latent trait model, which uses observed information about GDP, GDP per capita, and population to estimate posterior prediction intervals for each of these important concepts. By combining historical and contemporary sources of information, we are able to extend the temporal and spatial coverage of existing datasets for country-year units back to 1500 A.D through 2015 A.D. and, because the model makes use of multiple indicators of the underlying concepts, we are able to estimate the relative precision of the different country-year estimates. Overall, our latent variable model offers a principled method for incorporating information from different historic and contemporary data sources. It can be expanded or refined as researchers discover new or alternative sources of information about these concepts.
BibTeX
@misc{fariss2017latent,
  author = {Christopher J. Fariss and Charles Crabtree and Therese Anders and Zachary Jones and Fridolin Linder and Jonathan N. Markowitz},
  title = {{Latent Estimation of GDP, GDP per capita, and Population from Historic and Contemporary Sources}},
  year = {2017},
  eprint = {1706.01099},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1706.01099},
}
Figure