Replication of Burnside and Dollar (2000)

Econometrics | Economics

Project
Burnside and Dollar (2000) Replication
Prepared for
Paris 1 Sorbonne Master 2 in Financial Economics

Aid, Policies, and Growth (Burnside and Dollar, 2000) made a huge splash in the economic development literature when it was published. Using a novel policy index built from a weighted average of three indicators of “good governance”: the budget surplus, inflation rate, and a measure of trade openness. This index was interacted with foreign aid in a linear regression to determine the impact of aid on GDP growth. Their controversial result was that in countries with good governance, foreign aid has a positive effect, while in countries with bad governance, it has a negative effect.

The controversy surrounding the paper came in part due to the ideological camps that exist within the aid and development literature. There are those who believe aid never works, and should thus be eliminated all-together, and also those who believe aid is always beneficial, and should be use as aggressively as possible. Neither camp was happy with Burnside and Dollar’s result. More seriously, several methodological issues have been raised as well.  The authors had panel data, and included both fixed and time effects, but their variable of interest is largely invariant. As a result, it is absorbed by the fixed effects and cannot be estimated.

My replication uses the original Burnside and Dollar dataset to re-implement the main regressions of the paper. I then go one step further and use two models designed to allow the inclusion of time-invariant variables into a panel regression. The first is the Mundlak model, that blends a random effects estimator with a between estimator, and the second is a Hausman-Taylor model, which goes one step further by instrumenting endogenous time variant variables with a combination of their between transformations and time invariant variables. This is novel because the main econometrics packages in Python do not include modules for these regressions, and so had to be implemented by hand.

The full reproduction can be found here.