Benchmarking Water Retail Cost Efficiency in England and Wales

Abstract

Privatised water companies in England and Wales are subject to economic regulation by the industry regulator (Ofwat). Ofwat sets 5-year caps on the prices’ companies can charge their customers. These caps are in part based on the benchmarking of companies’ costs. Ofwat has not previously used econometrics to benchmark domestic retail costs, but it undertook such modelling in its 2019 price review analysis. This is the first journal article to present an efficiency analysis of domestic water retail costs in England and Wales. Our approach is different from Ofwat’s, as we propose two new ways of accounting for differences in the effect on cost of the number of single and dual service customers (water/sewerage-only and water and sewerage). Some companies’ cost efficiencies vary greatly between the two ways. Depending on the approach to price capping, this could possibly have non-negligible implications for companies’ caps.

Keywords:

• Periodic price caps
• frontier models
• single service (water/sewerage-only) customers
• dual service (water and sewerage) customers
• bad debt costs

JEL CLASSIFICATION:

• D24
• L95
• L51

Acknowledgements

The authors acknowledge comments on an earlier version of this paper from colleagues at Economic Insight, and participants in a special session on ‘Regulatory Cost Assessment in the Water Industry’ at the 16th European Workshop on Efficiency and Productivity Analysis, June 2019, London. The authors also acknowledge the constructive comments on an earlier version of the paper from an editor and two anonymous reviewers.

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

1 Cambridge Economic Policy Associates (CEPA) advised Ofwat on the econometric modelling for PR14.

2 Ultimately econometric benchmarking by Ofwat of companies’ domestic retail cost efficiencies did not form the basis of the final determinations of the companies’ PR19 domestic retail price controls. This was because Ofwat set these controls for the companies for 1 April 2020–31 March 2025 at the levels’ companies proposed in their business plans (Ofwat 2019). In doing so, Ofwat acknowledged the performance improvements inherent in the levels the companies proposed. On the other hand, it is evident that econometric benchmarking of domestic retail cost efficiencies was conducted by companies as part of their analysis for PR19, as they reported this benchmarking in their PR19 business plans (e.g., Welsh Water 2019; United Utilities 2018; Wessex Water 2018).

3 This academic paper emerged from earlier work by Economic Insight (2018).

4 It should be noted that Ofwat’s modelling of retail costs was undertaken completely independently of the modelling in this paper.

5 Doubtful debt relates to debts that are unlikely to be repaid and represents the movement in bad debt provision (and the bad debt charge). Expenditure on debt management relates to monitoring, issuing reminders, follow-up calls and field recovery which includes customer visits and costs relating to third party debt collectors and legal fees.

6 Customer services expenditure relates to billing, payment handling, vulnerable customer schemes and customer enquiries and complaints.

7 Specifically, Welsh Water (2019, 7) note that ‘Ofwat’s approach to residential retail modelling uses Corrected Ordinary Least Squares (COLS) to determine the “efficient” level of expenditure’.

8 SFA, DFA and DEA all have their advantages. For discussions of the relative merits of different frontier methods, see Hjalmarsson, Kumbhakar, and Heshmati (1996), Ruggiero (2007) and Greene (2008).

9 Technical change is Hicks-neutral if the change does not influence the productivity of the inputs. Relating this to the specifications of our cost models, where the dependent variable reflects the input cost to produce the output, technical change is Hicks-neutral as t is not interacted with output, and so the change in the input cost when the output level changes is independent of technical change. This contrasts with the translog cost function, where technical change is non-neutral because t is interacted with output, and so how the input cost changes when the output level changes depends on, among other things, the magnitude of technical change.

10 There has recently been a resurgence of methodological interest in the SS efficiency estimator, as it is the basis for a method to calculate efficiency performance spillovers (e.g., Glass, Kenjegalieva, and Sickles 2014, 2016). This draws parallels with the well-established literature on productivity performance spillovers.

11 The common way of using the random effects from an estimate of a model such as Equation (2)(2) $$c_{it}=\alpha +c\left(y_{it},w_{it},t\right)+{\gamma }^{\mathrm{\prime }}{z}_{it}+{\theta }_i+{\epsilon }_{it}.$$(2) to calculate time-varying efficiency estimates using the CSS approach is a three-step process. First, using a single model regress the disturbances from the fitted random effects model on, for each firm, a time trend and its square. For example, if there are 17 firms in the sample, there would be 34 time and time squared regressors in total. We cannot apply the CSS approach in our empirical analysis because in the first step this would involve estimating, for each firm, the coefficients on the time trend and its square using only five observations. To guard against this type of problem, the CSS efficiency estimator should only be applied to samples where T is big. In practice, researchers do apply the CSS estimator when T is not strictly big, but even in these cases T is much bigger than five. Second, for each firm add its random effect to the sum of the time trend and its square pre-multiplied by their firm specific coefficients. Third, using the values for the firms from the second step, calculate the time-varying efficiency estimates in the same fashion as COLS (see Equation (3)(3) $${\widehat{\mathrm{Eff}}}_{it}=\exp \left[\underset{i}{\min }\left({\widehat{\epsilon }}_{it}\right)-{\widehat{\epsilon }}_{it}\right],$$(3) ).

12 Note that the sign of inefficiency in a cost frontier model specification is positive. This is because firms aim to minimise costs so inefficient firms will lie above the cost frontier (see Figure 2).

13 Another approach that has widely been applied to COLS by regulators, in spite of it being conceptually problematic, is to split the residuals between inefficiency and other components (Ofwat 2008; ORR 2013). This typically involves applying a fixed percentage reduction to all firms’ residuals (Cubbin 2004; Chung 2011) and effectively pulls all the observations closer to the regression line, so that all firms’ distances to the lowest residual are reduced. While this approach directly addresses the problem at hand, the choice of percentage reduction is arbitrary and assumes that the split between inefficiency and other residual components is the same across all firms. We do not therefore use this approach here.

14 We note that, at earlier price controls, regulators including Ofwat tended to address this issue by selecting the ‘frontier’ firm as the benchmark (i.e., not the upper quartile) and then applying downwards adjustments to the residuals, to reflect the fact that they do not entirely relate to inefficiency. Note, this is separate from other downwards adjustments to residuals made by regulators for reasons other than the efficiency gap – which we discuss subsequently.

15 We do not use our fitted models for future prediction, so our benchmarking of companies’ domestic retail cost performance is based on historical upper quartile performance. At PR19 Ofwat (2017a, p. 18) planned to use such an approach as part of its benchmarking of companies’ domestic retail cost performance, by using an average of historical and predicted future upper quartile performance. Ultimately, such benchmarking by Ofwat did not form the basis of the final determinations of the PR19 retail price controls. See the policy implications in section 8 for more details on this.

16 The Pearson correlation coefficient between ts and s is only 0.53. There are evidently a sufficient number of companies in the sample with sufficient dual service customers to not be unduly concerned about collinearity between ts and s in the models.

17 Noting that companies typically will face a trade-off between debt management and bad debt costs. Hence, in the face of higher transience a company could: (i) invest in more debt management activities in order to attempt to recover a higher proportion of debt than it otherwise would; or (ii) not incur the higher debt management costs, but accept that its debt costs will increase as a result of the transience. The ‘optimal’ balance will vary from company-to-company, depending on their circumstances. The point, however, is that transience must logically increase these costs, when considered as a whole.

18 This additional company is Bournemouth Water, which was purchased by the parent company of South West Water in 2015, and, as a result, is not in the dataset in the last year of the sample.

19 We report average efficiencies for these 17 companies, as these are the companies Ofwat set retail price controls for at PR19.

20 Although the VIF pertains to cross-sectional models, since based on the VIF we have no concerns about multi-collinearity in the OLS models, we can make the case that there are also no multi-collinearity issues in our random effects models. This is because the same explanatory variables are used to estimate both sets of models.