The Effect of Feed-in Tariffs on Renewable Electricity Generation: An Instrumental Variables Approach

Abstract

While carbon taxes and other market-based instruments are widely regarded as optimal for climate mitigation, political constraints have prevented governments from using them. Instead, narrower instruments, including the feed-in tariff (FIT) for renewable electricity generation, have become popular. However, their causal effect on renewable electricity generation remains subject to uncertainty. We use instrumental variables to estimate the causal effect of FITs on renewable electricity generation in 26 industrialized countries, 1979–2005. We find that increasing the FIT by one U.S. cent (2000 constant prices) per kilowatt hour increases the percentage change in renewable electricity’s share of the total by 0.11 % points. All else constant, if a country implemented for a decade the sample mean FIT of three U.S. cents, the national share of renewable electricity would increase by 3.3 % points, which is more than the sample mean. In addition to demonstrating that the FIT is an effective way to increase renewable electricity generation, our approach lays the foundation for future studies of the causal effects of renewable energy policies.

Appendix

In order to address the potential that the use of a spatial instrument introduces correlation between the errors of adjacent countries, we also estimated a series of models using standard errors that are robust to heteroskedasticity and are clustered by year and region. We categorize the countries in the data in to six regions following the United Nations: North America, Northern Europe, Western Europe, Southern Europe, Eastern Europe, Oceania, and Eastern Asia. See http://millenniumindicators.un.org/unsd/methods/m49/m49regin.htm for a full list of the countries in these regions. In order to estimate a variance-covariance matrix using clustered standard errors, however, the majority of exogenous regressors must be partialled out. By the Frisch–Waugh–Lovell theorem (Frisch and Waugh 1933; Lovell 2008), the coefficients for the regressors that are not partialled out remain the same as they would be if the partialled out variables were included. The conformity of the produced coefficients to the equivalent models without clustered standard errors supports this assertion.

The models are shown in Tables 16 (the first-stage results) and 17 (the second stage results). Models 1 and 2 replicate models 1 and 2 in the main results (Tables 4, 5), and models 3 and 4 replicate models 4 and 5. The list of partialled-out variables is included in the table. Every model includes country and year fixed effects, though these are also partialled out.

Table 16 First stage results using clustered standard errors

Table 17 Second stage results using clustered standard errors

The results found using clustered standard errors are very similar to those found previously. The only notable exception is that the coefficient on Adjacent Country Mean FIT is no longer statistically significant at the $p < 0.05$ level in the first stage models; however, it remains significant at the $p < 0.06$ level in the first two models. More importantly, the coefficient and standard errors for the instrument mean FIT in the second stage results is essentially identical to those reported in the main results. The fit of these models is somewhat questionable, as the adjusted R$^2$ in the second stage results is negative, and the models fail to pass the over-identification test.