Abstract
The literature devoted to the export-led growth (ELG) hypothesis, which is of utmost importance for policymaking in emerging countries, provides mixed evidence for the validity of the hypothesis. Recent contributions focus on the time-dependence of the relationship between export and output growth using rolling causality techniques based on vector autoregressive models. These models focus on a short-term view which captures single policy-induced developments. However, long-term structural changes cannot be covered by examinations related to the short-term. This paper hence examines the time-varying validity of the ELG hypothesis for India for the period 1960–2011 using rolling causality techniques for both the short-run and long-run horizon. For the first time, window-wise optimal lag-selection procedures are applied in connection with these techniques. We find that exports long-run caused output growth from 1997 until 2009 which can be seen as a consequence of political reforms of the 1990s that boosted economic growth by generating foreign direct investment opportunities and higher exports. For the short-run, export significantly caused output in the period 1998–2003 which followed a concentration of liberalization measures in 1997. Causality in the reversed direction, from output to exports, only seems to be relevant in the short-run.
Acknowledgements
We are very grateful to Jonathan B. Hill who provided us with program code for testing Granger-causality in the Toda-Yamamoto approach. We also wish to thank two anonymous referees for their helpful comments and suggestions. All remaining errors, if any, are of course our own.
ORCID
Aviral Kumar Tiwari http://orcid.org/0000-0002-1822-9263
Alexander Ludwig http://orcid.org/0000-0002-4429-8887
Notes
1. Although it can be shown for bivariate models that Granger non-causality at forecast horizon h = 1 is theoretically equivalent to Granger non-causality at each forecast horizon h > 1 [Citation4], such an approach cannot be considered truly long-run if the same test statistic is supposed to be valid for each forecast horizon. This is likely to explain the large fluctuation of test inference from the rolling-window approach based on a VAR model as in Tang [Citation34].
2. We provide gretl code for the VECM-based long-run Granger causality tests upon request.
3. We present the UDmax and SEQ(k + 1|k) test statistics in the notation of Kejriwal [Citation17].
4. As we test restrictions on speed of adjustment coefficients in the long-run case instead of testing restrictions on the short-run coefficients in (see Section 2.1), we expect that the type of lag selection will have less impact on the results of the long-run case.
5. This observation is also likely to be the reason for the borderline-rejection of I(2) for output at the 10% level for the entire sample (see Section 3.1).