Laura Puzzello, Monash University
"Does trade matter for growth when geographical instruments are randomly generated?"
In their highly influential paper ‘Does Trade Cause Growth?,’ Frankel and Romer (FR) estimate a trade equation to predict bilateral trade shares, which are in turn used to construct an instrument for trade openness in income regressions. Several papers have followed the FR approach; however, they rarely state whether out-of-sample predictions of bilateral trade flows are included in the instrument set. Using bilateral trade shares predicted from randomly generated geographical characteristics to form instruments for trade openness, this paper shows that the results are highly sensitive to whether out-of-sample predictions are included in the instrument set. We show analytically and empirically that the coefficient of trade openness in income regressions is severely upward biased when out-of-sample predictions are excluded from the instrument set because the instrument captures the number of trading partners and, therefore, violates the exclusion restriction. Thus, out-of-sample predictions should always be included in the instrument set to eliminate mechanical endogeneity.
Contact person: Jakob Roland-Munch