Efficiency evaluation of electricity distribution utilities in India: A two-stage DEA with bootstrap estimation

Abstract

The efficiency of the energy sector in India, the world’s third-largest energy consumer, is becoming an important issue amid growing concerns about global warming. This paper examines the efficiency of electricity distribution utilities in India using panel data from 2005 to 2012. We perform a two-stage data envelopment analysis with a bootstrap estimation, in which bias-corrected efficiency estimates are calculated in the first stage and then regressed on external environmental variables in the second stage. First, we find positive effects of customer structure and population density on the efficiency of utilities. Second, we find efficiency advantages of public utilities in the Indian power distribution sector. However, the interaction between ownership and population density is negative, implying that public utilities are less efficient than private enterprises in high population density areas. Finally, we find that government subsidies are negatively related to the efficiency of utilities.

Keywords:

• Data envelopment analysis (DEA)
• bootstrap
• environmental variable
• electric power distribution
• India

Notes

1. Global Energy Statistical Yearbook 2014 (https://yearbook.enerdata.net).

2. $E_{\widehat{\mathcal{P}}}({\widehat{\theta }}_i^{\ast })-{\widehat{\theta }}_i$ is the estimator of $E_{\mathcal{P}}({\widehat{\theta }}_i)-{\theta }_i$.

3. Specifically, ${\tilde{\theta }}_i={\widehat{\theta }}_i-\{E_{\widehat{\mathcal{P}}}({\widehat{\theta }}_i^{\ast })-{\widehat{\theta }}_i\}=2{\widehat{\theta }}_i-E_{\widehat{\mathcal{P}}}({\widehat{\theta }}_i^{\ast })$. See Daraio and Simar (2007) and Bogetoft and Otto (2011) for details.

4. A random effects approach assumes that individual effects are uncorrelated with the regressors. By contrast, a fixed effects approach allows for correlation between the individual effects and regressors, but it cannot estimate the coefficients of the time-invariant variables.

5. The parameters of the random effects model can be estimated consistently but not efficiently using OLS. The FGLS estimator is efficient (Greene, 2012).

6. Banker and Natarajan (2008) and McDonald (2009) discussed in detail the application of OLS to a two-stage DEA model. Simar and Wilson (2007, 2011) examined a truncated regression for such a model.

7. Executive Summary of Power Sector, the Ministry of Power, Government of India (February 2014).

8. Specifically, the R packages Benchmarking (Bogetoft & Otto, 2011) and FEAR (Wilson, 2008) are used for the bootstrap DEA estimates (2,000 repetitions for bootstrapping).

9. We reject the null hypothesis of the CRS assumption for Model 1 from a statistical viewpoint. However, the test statistics are still practically close to 1 in Model 1. Thus, the CRS and VRS assumptions may not make a considerable difference to the analysis in this case.

10. In a regression analysis, logarithms are usually used for the variables that are strictly positive and possibly large values. Regression results are less sensitive to outlying or extreme observations, as taking logarithms usually narrows the range of variables. Moreover, strictly positive variables often have a skewed distribution. Taking logarithms usually mitigates this problem as well (e.g., Wooldridge, 2009, Chapter 6).