Machine Learning Approaches to Testing Institutional Hypotheses: the Case of Acemoglu, Johnson, and Robinson (2001)

Boubacar Diallo (Qatar University)

Abstract: In their seminal 2001 work, Acemoglu, Johnson, and Robinson (AJR) argued that institutions influence economic development, using the logarithm of settler mortality as an instrument to establish a causal effect. A number of economists and other social scientists have challenged this work in terms of both data and identification strategy. One of these criticisms concerned the IV estimated coefficients and standard errors, which were nearly twice as large as the OLS coefficients and standard errors. My research uses machine learning to test the robustness of AJR's findings. Using the AJR dataset, which I randomly divide into training data and testing data, I am able to predict the average protection against expropriation risk from settler mortality. These predicted values of property rights protection are then regressed on per capita GDP growth. My results indicate a strong and positive effect of property rights protection on growth, consistent with AJR's earlier results. Moreover, the use of machine learning to obtain institutional values yields estimates close to the OLS estimates, unlike AJR. Removing African countries and neo-European countries such as Canada, Australia, USA, and New Zealand does not alter the sign and significance of the coefficient of interest. These results suggest that machine learning can be helpful to economists facing data issues.