Spurious regressions and near-multicollinearity, with an application to aid, policies and growth

• A classical suppressor has a zero simple correlation with a dependent variable.
• Classical suppressors often lead to spurious inference in multiple regressions.
• Highly correlated classical suppressors are very often statistically significant.
• Parameter inflation factor, pre-tests and outliers signal those spurious regressions.
• Aid has a spurious effect on growth conditional on good policy.

In multiple regressions, explanatory variables with simple correlation coefficients with the dependent variable below 0.1 in absolute value (such as aid/gross domestic product (GDP) with GDP growth) face a problem of parameter identification. They may have very large, statistically significant, estimated parameters which are unfortunately “outliers driven” and spurious. This is obtained by including another regressor which is highly correlated with the initial regressor, such as a lag, a square or interaction terms of this regressor. The analysis is applied on the “Gambia and Botswana outliers driven” Burnside and Dollar (2000) article which found that aid/GDP had an effect on growth only for countries achieving “good” macroeconomic policies.