The export–growth relationship: estimating a dose-response function

Abstract

The relationship between individual firms' export behaviour and firm performance has been studied extensively in the economic literature. However, most studies from the field of economics only distinguish between exporting and nonexporting companies, using the firms' export status as a binary treatment variable and comparing the performance of exporting and nonexporting firms. This article introduces the newly developed generalized propensity score (GPS) methodology to the literature of individual firms' export behaviour. Instead of a binary treatment variable, the GPS method allows for continuous treatment, that is, different levels of the firms' export activities. Based on the GPS methodology, a dose-response function is estimated, depicting the relationship between the firms' export-sales ratio and their subsequent sales growth rate as a measure of firm performance.

Acknowledgements

I gratefully acknowledge financial support from the HSBC Innovation and Technology Group and the Anglo-German Foundation for the Study of Industrial Society. Helpful suggestions by Dirk Czarnitzki, Joachim Wagner, and Michael Woywode are also gratefully acknowledged. I thank Marc Rennert for his competent research assistance and Andrew Flower for proofreading. Special thanks go to Marc Cowling and Gordon Murray for carrying out the survey this study is based on in the United Kingdom. Any errors that remain are my own.

Notes

¹ Hirano and Imbens (2004) use a normal distribution for (the logarithm of ) the treatment variable of their model. However, they emphasize that more general models may be considered.

² The bias removal property of the GPS is based on the assumption of weak unconfoundedness which is a generalization of the (strong) unconfoundedness assumption made by Rosenbaum and Rubin (1983) for binary treatments (cf. Imbens, 2000): Let the treatment D take on values in the interval . Assignment to treatment D is weakly unconfounded, given pre-treatment variables X, if Y(d) ⊥ D|X for all , with Y(d) as the outcome associated with treatment level d.

³ Hirano and Imbens state that asymptotic normality for the estimator in Equation 5(5) can be proved.