Give examples of actions that would raise the GDP, including undesirable actions.

BASED ON THIS ARTICLE "New Estimates of Over 500 Years of Historic GDP and Population Data"

Fariss, C. J., Anders, T., Markowitz, J. N., & Barnum, M. (2022). New Estimates of Over 500 Years of Historic GDP and Population Data. The Journal of Conflict Resolution, 66(3), 553-591. https://4x10hwww7-mp03-y-https-doi-org.proxy.lirn.net/10.1177/00220027211054432

Journal of Conflict of Resolution (Forthcoming) Abstract Gross Domestic Product (GDP), GDP per capita, and population are central to the study of politics and economics broadly, and conflict processes in particular. Despite the prominence of these variables in empirical research, existing data lack historical coverage and are assumed to be measured without error. We develop a latent variable modeling framework that expands data coverage (1500 A.D-2018 A.D) and, by making use of multiple indicators for each variable, provides a principled framework to estimate uncertainty for values for all country-year variables relative to one another. Expanded temporal coverage of estimates provides new insights about the relationship between development and democracy, conflict, repression, and health. We also demonstrate how to incorporate uncertainty in observational models. Results show that the relationship between repression and development is weaker than models that do not incorporate uncertainty suggest. Future extensions of the latent variable model can address other forms of systematic measurement error with new data, new measurement theory, or both.

Introduction

Gross Domestic Product (GDP), GDP per capita (GDPPC), and population play a vital role in empirical social science. Moreover, they are key variables in the study of international and domes[1]tic conflict processes in particular. Despite the prominence of these variables, existing data used to operationalize them suffers from three problems: (1) measurement coverage (missingness), (2) measurement uncertainty, and (3) measurement bias. First, there is a lack of economic cross-country data coverage prior to 1950 (Gleditsch, 2002), or data are available with very limited temporal coverage even though many other datasets of interest cover variables beginning in the 1800s.1 This leaves researchers unable to consistently estimate the relationship between key variables such as economic development and democratization, conflict, repression, or health prior to this date. Second, measurement error arises because of imprecision or disagreement in available data, which can mask real relationships or exacerbate false ones. Though scholars are aware that data suffers from measurement error, existing estimates of these variables provide only point estimates and offer no method for quantifying uncertainty (measurement error). Third, existing models for estimating GDP, population, and GDP per-capita offer no way to correct for any systematic bias in the data generating process. As previous research has demonstrated, failing to correct for measurement bias leads scholars to draw incorrect inferences about the relationships in their data. We develop a latent variable model of GDP, GDP per-capita, and population that provides remedies for each of these issues.2 First, the latent variable model generates estimates of cross-national data coverage back in time by several centuries to 1500. These estimates are important not only for scholars assessing existing explanations of outcomes in earlier periods of history, but also scholars interested in comparing inferences generated from data across time periods. Given that most of the major interstate wars occurred before 1950, these data will be particularly useful to scholars studying the causes of war, as well as the debate about its relative decline over time (e.g., Fazal, 2014; Lacina, Gleditsch and Russett, 2006). Until now, researchers have been unable to estimate these long-term relationships. 1For example, Bolt et al. (2018) and Maddison (2010) provide extended historical coverage only for specific years, e.g. 1800, 1820, and 1850. 2This measurement model builds on earlier data collection efforts and model versions in which the combined sources were used to estimate these three variables (e.g., Anders, Fariss and Markowitz, 2020; Fariss et al., 2017; Markowitz and Fariss, 2018). These earlier models did not account for the different modes of data generation for each of the datasets (i.e., PPP adjusted vs. exchange rate conversion), which our model accounts for. 2 For example, existing research argues that economic development reduces the risk of conflict both by contributing to the democratization process of states in the international system and by making war more costly (Hegre, 2000). As Souva and Prins state, "democratic regimes are about 37% less likely to initiate fatal militarized disputes given an average level of GDP per capita" (Souva and Prins, 2006, 194). However, other research argues that the relationship with economic development is curvilinear, for both democratization (Treisman, 2020) and conflict (Boehmer and Sobek, 2005; Gartzke, 2012). To date, the relationships between economic development and democratization, repression, or conflict have not been estimated with complete cross-national data prior to 1950. Scholars have instead relied on proxy-measures such as energy consumption per[1]capita (e.g., Markowitz, McMahon and Fariss, 2019), shipping and rail costs (e.g., Lake, 2009; Markowitz and Fariss, 2013), or simple linear interpolation of GDP per-capita (e.g., Treisman, 2020) to account for long-term economic variation. The new estimates we present in this article allow researchers to evaluate whether these empirical relationships are limited to the post-1950 period, or whether they generalize to earlier time periods (using more precise data with better coverage). We show that many correlational patterns between GDP per capita and these other variables vary considerably over time, which means that relationships for the post-1950 period do not necessarily generalize to periods prior to 1950. Critically, this also suggests that these relationships might change in the future. Second, our new latent variable model provides estimates of the relative level of uncertainty for each country-year unit by accounting for variation in the level of coverage within and disagreement between component indicators. This is useful because it allows researchers to evaluate whether measurement error in the data (expressed as the level of uncertainty with which each country-year unit is accurately measured) is large enough to alter the size, or even the direction, of the effect of a given explanatory variable. For illustration, we demonstrate that measurement uncertainty may be large enough to reduce the estimated magnitude of the effect of development (GDP per capita) on repression by as much as a third. If models do not incorporate uncertainty, then the researcher cannot rule out the possibility that the statistical associations of GDP, GDP per capita, or population with some other variables are not a false-negative result (Type 2 error or attenuation bias), or the related possibility that the relationship of these variables are a false-positive result (Type 1 error). In 3 models with multiple indicators (i.e., multiple-variable regression), bias due to measurement error is not always attenuating if un-modeled, higher-order interactions exist. Incorporating measurement uncertainty of variables addresses these difficult-to-model issues and can be further explored with non-parametric regression techniques. In sum, incorporating uncertainty into a regression model provides evidence that effect sizes are probabilistically distinguishable from zero even when we cannot measure a right-hand side variable perfectly (an assumption of standard regression models). Though some existing models that include these variables may have under-estimated an effect size, others may have over-estimated effect size or even incorrectly reversed the direction of the effect. The implications are potentially profound given the wide usage of these variables across the social sciences. Third, the measurement model provides a framework for incorporating theoretical knowledge about the data generating process for information used in the measurement of each variable that can correct for potential bias in the existing data. We describe how the latent variable model we develop in this article can be further developed to address measurement bias and point to other areas of scholarship that have successfully built and extended latent variable models to address measurement bias.3 The new estimates provide ample opportunities for future scholarship to re-visit existing debates and investigate new research questions relating to war, peace, economic development and health and well-being, as well as the information necessary to reduce bias due to measurement uncertainty in GDP, GDP per capita, or population variables. The key output from the model is predicted intervals of the original source variables in the original unit-of-measurement, in addition to the relative level of uncertainty for each country-year estimate in the form of a standard deviation. The level of precision for a particular unit is based on the number of indicators available for that unit, its temporal proximity to other units with available information because the model is dynamic, and the agreement between the indicators. Missing values are inferred from the dynamic structure of the model based on the information contained in the variables associated with a given country year and the estimates from prior and future values. This directly addresses the issue of list-wise deletion that is well known in the conflict literature to potentially bias results (e.g., Boehmer, 3See examples from the measurement of human rights (Fariss, 2014, 2018b,a, 2019; Fariss and Dancy, 2017; Fariss, Kenwick and Reuning, 2020) and civilian control of the military (Kenwick, 2020). 4 Jungblut and Stoll, 2011; Gleditsch, 2002; King et al., 2001). Fourth, scholars can use their preferred variable (e.g. the data series from the Penn World Tables, World Bank, or the Maddison Project) because our model generates a range of estimates for every country-year unit in terms of each and every one of the variables that we include in our model. The new estimates are the posterior predications generated from a dynamic latent variable model that approximates the process economic historians use when creating measures of these variables. Economic historians begin with population estimates, they then use historic data on prices and sector specific production to generate per-capita estimates of economic value. With the estimates of per-capita economic value and population, the calculation of overall GDP becomes possible. Because economic historians are using some data series that are designed to estimate change in productivity within a country over time and others are designed to capture cross-sectional variation, but usually not both, we incorporate this information into our measurement models of GDP and GDP per capita.4 Having demonstrated that solving these problems has important implications for all areas of social science research, we now turn to the task of demonstrating that we have used rigorous and appropriate methods to construct a new dataset to measure each of these concepts of interest. We close with several demonstrations of how our new estimates can be applied by scholars to improve their research.

Conclusion

In this article, we identified and discussed the limitations of existing GDP, GDP per capita, and population variables. We then introduced and described a new latent variable model that estimates posterior predictions for these important variables. These new estimates provide several primary advantages over existing variables: (1) they extend the temporal and spatial coverage of existing data, (2) they include multiple manifest indicators of each concept and quantify the uncertainty for each of the resulting country-year estimates, and (3) the modeling framework itself can be expanded to account for new data or knowledge about systemic error in existing data. With the expanded temporal coverage of our estimates, we showcase several new empirical insights about the relationship between economic development and democracy, repression, conflict, power projection capabilities, literacy rates, and life expectancy, respectively. We also demonstrate how to incorporate uncertainty in commonly used observational models. These new results show that the relationship between repression and development is weaker than models that do not incorporate uncertainty suggest. Future extensions of the latent variable model can address other forms of systematic measurement error with new data, new theory, or both. To encourage the extension of our model, we make publicly available the data and code used to construct it. The code in our replication archive has been carefully annotated to encourage easy replication, modification, and we hope theoretically informed extensions. The primary feature of our measurement model is that it provides a principled means to incorporate information for comparing measures of the economy from different datasets (across time, across country, or both). To inform our estimates, the measurement model only includes the dimension of information that is comparable. For example, using dataset values that are designed for comparison within a country over time, we include a country fixed effect (c). When we produce the estimates of the country-year distributions, these additional parameters are also included. To be clear, we only include the relevant dimension of comparison from each of the datasets. The distributions of plausible estimates for each country-year unit are converted back into the relevant unit of measure for each of these datasets. Future users who are interested in within-country change should use one of the PPP estimates that are designed for this dimension of comparison (e.g., MDP 2018 RGDPpcNA in 2011 $US). Users interested in making between-country comparisons should 41 use the estimates generated for the MDP 2018 CGDPpc in 2011 $US variable. This is because the model includes parameters that preserve the built in methodologies used to generate the original dataset.26 All this to say, we take seriously the data production process of each variable included in our latent variable model, which should help guide scholars to the appropriate estimates generated from our model and alleviate concerns about combining different variables together in a unified latent variable framework. To close, we wish to reiterate that the uncertainty estimates associated with our data should always be incorporated into any statistical analysis and that, though these uncertainty estimate are useful, they do not eliminate concerns about systematic measurement error that may arise because of potential bias in the information sources used to estimates these data (such as lying or low state capacity). Moving forward, we plan to update our new data each year and incorporate new information as it becomes available. We will maintain both a full historic data set (1500- present) and a smaller dataset from 1800-present. The goal in maintaining two datasets is to keep users informed about the relative level of data coverage in the early historic period relative to the historic data coverage available over the last 200 years. The smaller dataset will work well with other data sets like the Varieties of Democracy and Correlates, which also begins in 1790. Our new estimates are already helping scholars expand the scope of their inquiries (e.g. Anders, Fariss and Markowitz, 2020; Gerring and Veenendaal, 2020; Markowitz et al., 2020). We are excited to introduce these new estimates to the broader community of scholars and to see what new studies discover.