Architect: Minoru Yamasaki


I study technology, innovation, investment allocation, and the role of money, trade, and other resources in political competition.

Working Papers:

Political Competition in Dynamic Economies. Presented at the University of Oxford Political Economy of Finance Conference (2018), the Atlanta Conference on Science and Innovation Policy (2017), in the Princeton University Political Economy Colloquium (2017), and the Princeton University International Relations Colloquium (2016).

Does Globalization Bring Peace? A Study of Trade and Interstate War, 1845-1905.  Presented in the Princeton University Political Economy Colloquium (2018), at the Midwest Political Science Association Annual Meeting (2018) and the Princeton University International Relations Colloquium (2017).

Abstract: Testing the link between trade and peace, both between and within countries, has so far been difficult due to the lack of a good instrument for trade. I leverage a novel instrument for trade – the change in effective shipping distances due to the invention and adoption of steam technology (Pascali 2017) – to study the relationship between increases in trade and countries’ propensity for conflict. I find that as countries trade more overall, the probability that they go to war goes down, in a relationship not driven by bilateral effects, and not present for civil war. This has implications both for our understanding of the first wave of globalization, and the one we are experiencing today: I provide evidence in favor of a “commercial peace” that is general, and not within or between particular states.

A sample of commercial routes by sail in 1791-1799 (top) versus shipping routes by modern propulsion (bottom). Both diagrams are based on actual shipping logs, with the former map constructed by the author using data from Climatological Database for the World’s Oceans 1750-1850 (CLIWOC), and the latter due to T. Hengl (2007). These illustrate the large difference between sailing and propeller paths. This paper uses differences in optimal paths based on water currents (steam) and both water currents and wind patterns (sail).

Technology Diffusion and the International System, with Helen V. Milner. Presented at the Princeton International Relations Faculty Colloquium (2018), and the Annual Meeting of the International Political Economy Society (“IPES” 2018). (PDF)

Abstract: Does world politics affect the diffusion of technology? States overwhelmingly rely on technology invented abroad, and their differential intensity of technology use accounts for much of their differences in economic development. Some international relations scholarship suggests states adopt new technology as they seek to avoid vulnerability to attack or coercion by more developed neighbors. We argue the structure of the international system affects the level of competition among states which in turn affects leaders’ willingness to enact policies that speed technology adoption. We examine this systematically by considering states’ adoption of technology over the past 200 years. We find that countries adopted new technologies faster when the international system was less concentrated, that such systemic change Granger-caused technology adoption, and that policies to promote technology adoption are related to concerns about rising international tensions. A competitive international system is an important incentive for technological change, and may underlie global “technology waves.”

Smart Social Networks: How AI-Curation Polarizes and Empowers Extremism in Online Communities

What are the consequences of AI curation for online communities? I build a simulation of a social media platform in which users have dynamic views, generate and share content based on these views, and an AI algorithm connects users with similar content. I find: (1) AI curation leads to view polarization, (2) these effects are magnified the more accurate the AI algorithm is, and (3) that more accurate AI-groupings make users who have views that are extreme and unchanging (trolls/extremists) have a higher impact. I assess the robustness of these findings using a range of AI grouping strategies, number of groups, users, starting views, and content generation procedures. 

Shows starting views (gray) and views after one year of interaction on social media (blue), both on a right-left dimension. The plot builds on a novel computer simulation of a social media platform in which users generate and post content based on their views, this content is shared within groupings constructed by a machine learning algorithm, and users update their views based on the content they see. In the simulation presented here, another layer of complexity is added in that five percent of users are “extremists” – have views at either the far right or far left, and do not update their views. As seen, the more accurate the machine learning is, the more powerful these extremists and group dynamics become: rather than more interaction leading to agreement, it leads to polarization.

The Soil of Politics: Resources, Competition, and the Consequences of Ethnic Inequality, with Rachael McLellan. Presented at the 4th Annual Conference of the History and Politics Network (2017), the Princeton Political Economy Colloquium (2017), and APSA (2016).

Abstract: We demonstrate the importance of the distribution of land between ethnic groups in shaping political development. We argue and show that the concentration of resources, here as measured by the cumulative agricultural potential of ethnic homelands – which we call soil power – makes it easier to form stable ruling coalitions. Where political resources are highly fragmented between groups, coalitions are less secure. Ruling elites therefore have a greater incentive to impose narrow, repressive institutions to retain power and are more likely to face armed challenges to their rule. To test our argument, we leverage the natural experiment of the African borders drawn at the Berlin Conference. We find that higher concentration of power leads to better quality institutions, less civil war and higher levels of economic development. We find similar relationships globally, and that African countries are low-concentration outliers.

Murdock’s (1959) map of of 835 ethnic homelands in Africa. The graphic (including the color shadings) is as made available by Harvard DataVerse (2016). This highly detailed map forms the basis for our analysis of ethnic group resources in Africa. We utilize information on soil, local climate, and area to assess the cumulative agricultural potential of these ethnic homelands, and then assess the within-country concentration of such “soil power”.



The US Patent Data Set (1976-2015). 

Data Set of all patents granted in the United States between 1976 and 2015, with information on patent class, country and city of inventor, and number of times cited by other patents. Includes 5.9 million patents and 31.6 million associated citations.

USTPO Patent
Example U.S. patent with information recorded in the US Patent Data Set underlined in green.

The Military Technology Data Set

  • Country-year adoption rate of 33 military technologies.




Coverage – an R package for seeing what you are missing. Available on my Github. 

The coverage package and associated function provides you with a summary of your time and unit coverage, producing compact tables and a visual representation. This is important for any analysis conducted with row-wise deletion in the presence of missing data, especially if one suspect that patterns of missingness are non-random with respect to variables of interest. Analysis which use row-wise deletion include standard regression analysis and most implementations of maximum likelihood. See my Github for examples and details.

The “out-of-the-box” visual output below shows country and year coverage for a technology-country-year analysis, with darker blue indicating more observations, and grey indicating years in which the unit in question did not exist as an independent country:


QuickReg – Easy OLS with options in R. Available on my Github. 

The QuickReg package and associated function provides an easy interface for linear regression in R. This includes the option to request robust and clustered standard errors, automatic labeling, an easy way to specify multiple regression specifications simultaneously, and a compact html or latex output (relying on the widely used “stargazer” package).

QuickReg also includes several functionalities to speed up OLS computation, including a special implementation of the method of alternating projections. This method reduces calculation time drastically (>60 percent in tests) for analysis with a large number of fixed effects, with a performance gain that is increasing in the number of regression specifications passed to the function simultaneously.

QuickCoefPlot – Easy OLS coefficient plots with options in R. Available on my Github. 

The QuickCoefPlot package turns base R’s linear-model objects into easy-to-read coefficient plots, complete with model statistics, labeling and heteroskedasticity-robust standard errors (by default).  Includes a range of options, including the opportunity to request default or clustered standard errors, the option to calculate and report bootstrapped confidence intervals and/or estimates, and change text-size, plot title, which model statistics or coefficients to report, and so on.

# Use the QuickCoefPlot to produce a regression plot
model <- lm(log_gdppc_mad ~  polity2 + tropical + desert + soil + near_coast + year, data = mydata)



seeAI – Visualizations of Machines Learning in R. Available on my Github. 

The seeAI package aims to visualize common machine learning procedures. Visuals are dynamically generated based on supplied models so as to be useful for both teachers/students and practitioners. It currently supports the popular glmnet implementation of cross-validated lasso regression.

In the graphic, coefficients extend outward (positive) and inward (negative) from the central circle, while the outer blue line tracks decrease in maximum model complexity. The second plot shows mean cross-validation error and standard deviation.