Assesing the influence of environmental variables on the performance of water companies: An efficiency analysis tree approach

Journal Pre-proofs Assesing The Influence Of Environmental Variables On The Performance Of Water Companies: An Efficiency Analysis Tree Approach María Molinos-Senante, Alexandros Maziotis, Ramon Sala-Garrido, Manuel Mocholi-Arce PII: S0957-4174(22)01862-0 DOI: https://doi.org/10.1016/j.eswa.2022.118844 Reference: ESWA 118844 To appear in: Expert Systems with Applications Received Date: 20 April 2022 Revised Date: 6 September 2022 Accepted Date: 13 September 2022 Please cite this article as: Molinos-Senante, M., Maziotis, A., Sala-Garrido, R., Mocholi-Arce, M., Assesing The Influence Of Environmental Variables On The Performance Of Water Companies: An Efficiency Analysis Tree Approach, Expert Systems with Applications (2022), doi: https://doi.org/10.1016/j.eswa.2022.118844 This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition of a cover page and metadata, and formatting for readability, but it is not yet the definitive version of record. This version will undergo additional copyediting, typesetting and review before it is published in its final form, but we are providing this version to give early visibility of the article. Please note that, during the production process, errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain. © 2022 The Author(s). Published by Elsevier Ltd. 1 ASSESING THE INFLUENCE OF ENVIRONMENTAL VARIABLES ON THE PERFORMANCE OF WATER COMPANIES: AN EFFICIENCY ANALYSIS TREE APPROACH María Molinos-Senante1,2,*, Alexandros Maziotis1, Ramon Sala-Garrido3, Manuel Mocholi-Arce3 1Departamento de Ingeniería Hidraulica y Ambiental, Pontificia Universidad Católica de Chile, Avda. Vicuña Mackenna 4860, Santiago de Chile, Chile (e-mail: [email protected]; [email protected]) 2Insitute of Sustainable Processes, University of Valladolid, C/ Mergelina S/N, Valladolid, Spain. 3Departamento de Matemáticas para la Economía y la Empresa, Universidad de Valencia, Avda. Tarongers S/N, Valencia, Spain (e-mail: [email protected]; [email protected]) * Corresponding Author 2 Abstract Efficiency assessment is a valuable tool for industries that are regulated, such as the provision of drinking water. Hence, past research on this topic is wide. However, current, widely used approaches such as parametric, non-parametric and partial frontier methods present several limitations and pitfalls. Thus, here, the Efficiency Analysis Tree (EAT) method was trialled on a sample of water companies. This method overcomes overfitting issues, because it employs a combination of classification, regression tree methods, and non-parametric analyses. For comparative purposes, efficiency was also estimated using Data Envelopment Analysis (DEA) and Free Disposal Hull (FDH) non-parametric methods. The approach was applied empirically using a sample of English and Welsh water companies during 1991–2020. Average efficiency was estimated at 0.489, showing that water companies could save 51.1% of their costs if efficient. Except for the 2011–2015 period, efficiency increased over time, indicating that price reviews by the English and Welsh water regulator contributed to improving water company performance. The application of bootstrap regression analysis techniques showed that the main source of raw water, percentage of metered properties, population density, and percentage of water leakage represented environmental variables that significantly influenced the efficiency scores of water companies. The approach introduced here could be of use to water regulators, as it overcomes the existing limitations of traditional approaches employed to assess the performance of water companies, facilitating sound decisionmaking. 3 Keywords: regression trees; efficiency analysis; performance; water utilities; environmental variables; water services. 4 1. INTRODUCTION Measuring the efficiency of production processes is valuable for decision making units (DMUs). Such measurements show how inputs are used to generate outputs “i.e.” production technology, and evaluate the efficiency of processes. Efficiency measures the maximal (minimal) contraction of inputs (outputs) to generate the same level of output (input) (Farrell et al., 1957). The concept of efficiency has been widely used for monitoring the performance of several sectors of the economy such as education, health services, airports, e-commerce enterprises and banking (Iyer and Jain, 2019; Zakowska and Godycki-Cwirko, 2020; Pratap et al., 2022). Assessing efficiency in regulated industries (such as water, gas, and electricity) is of particular interest to researchers and policy makers (Berg and Marques, 2011; Daraio et al., 2020; Mergoni and De Witte, 2022), as it allows the impact of regulatory reforms and policies to be evaluated. Furthermore, it can be used to determine future cost allowances and tariffs for customers (Cetrulo et al., 2019; Goh and See, 2021). Efficiency has been traditionally measured using parametric and non-parametric techniques. Parametric techniques use econometrics, such as Stochastic Frontier Analysis (SFA), to compare the inputs and outputs of units. SFA incorporates both noise and inefficiency “i.e.” it is stochastic. To do this, a functional form must be specified for the production technology “e.g.” Cobb-Douglas, translog, which makes different assumptions regarding the distribution of inefficiency “e.g.” half-normal, exponential (Wang et al., 2017). In contrast, non-parametric techniques do not have these requirements. Non-parametric methods build on linear programming models, such as Data Envelopment Analysis (DEA) and Free Disposal Hull (FDH). In this case, 5 the production frontier is not estimated econometrically, but is constructed using observed data on inputs and outputs. The production frontier under DEA is piecewise linear and convex, whereas under FDH the frontier is a step function (O´Donnell, 2018). Non-parametric approaches assume that any deviation from the efficient frontier is only caused by inefficiency “i.e.” they are deterministic (Dyckhoff, 2018). To overcome the deterministic nature of this approach, other techniques have been proposed, such as the bootstrap DEA procedure and partial frontier techniques “e.g.” order-m(Simar and Wilson, 2007; Ferreira and Marques, 2017). However, it is challenging to select the number of bootstrap replications and optimal number of m (Villegas et al., 2019). Consequently, Esteve et al. (2020, 2021a) developed a new technique, called Efficiency Analysis Trees (EAT). This technique combines the Classification and Regression Trees (CART) proposed by Breiman et al. (1984) with non-parametric analysis to measure efficiency. By using regression trees, DMUs are separated into several regions using a set of different thresholds (Esteve et al., 2021b). The EAT approach adjusts the regression tree to estimate production frontiers and efficiency. Specifically, the free disposability assumption is imposed where the estimated value of the response (output) variable refers to its maximum (value), and not the average value. As a result, the estimated frontier utilizes a step function that allows efficiency scores to be measured. Esteve et al. (2020, 2021a) demonstrated that the EAT technique outperformed other non-parametric techniques (such as DEA and FDH), and improved the accuracy of efficiency measurements, because values are not overfitted. Thus, this study aimed to evaluate the efficiency of water utilities using EAT, the newly developed technique. The regression tree allowed the maximum (frontier) 6 expenditure required to provide water services to be visualised at different thresholds (rules). EAT allows the efficiency scores for each water utility to be estimated. To allow comparison, our study also estimated efficiency using non-parametric techniques (such as DEA and FDH). In parallel, we explored how several environmental variables (operational characteristics) influenced the efficiency of water utilities. Bootstrap regression analysis techniques were used, in which the EAT efficiency score was regressed against a set of factors associated with network quality, source of raw water, and population density. Literature reviews conducted by Cetrulo et al. (2019) and Goh and See (2021) demonstrated that many studies have evaluated the efficiency of water companies. The bibliometric analysis conducted by Goh and See (2021) identified 142 articles on benchmarking the performance of water companies during the years 2000-2019. Moreover, Cetrulo et al. (2019) identified that DEA was the most commonly used method to evaluate the efficiency of water companies. The aim of these previous studies was diverse. Some studies focused on comparing the efficiency of public and private companies (Estache and Trujillo, 2003; Molinos-Senante et al., 2016). It is also possible to observe benchmarking studies linked to the implementation of regulation processes (Berg and Lin, 2008; Drusiani et al., 2013). Other studies explored the impact of economies of scale, scope and density on the performance of water companies (Guerrini et al., 2015; Lo Storto, 2020). Several studies (Marques et al., 2014; Pinto et al., 2017) examined the influence of exogenous variables on the performance of water companies. However, all previous studies on this topic used traditional parametric and non-parametric methods. Despite the advantages of EAT, it has not been previously used to assess the efficiency of water companies. Hence, this study extends the 7 literature on this subject by employing a newly developed technique that combines decision tree analysis and production economics to improve the accuracy in evaluating the efficiency of water utilities. The paper unfolds as follows. Section 2 presents the methodologies employed in this study to estimate efficiency scores and the impact of environmental variables on efficiency. Section 3 describes the case study and sample data. Section 4 presents and discusses the main findings, whereas the final section concludes. 2. METHODOLOGY 2.1 Efficiency methods Efficiency scores for a sample of water utilities were estimated using EAT, in which the predicted values of the response (output) variable were visualised through a decision tree. DMUs were split into several non-overlapping regions based on a set of thresholds of predictor (input) variables (James et al., 2013; Rebai et al., 2019). With EAT, the efficient frontier was estimated using step functions that satisfied the basic properties of microeconomics, such as free disposability. Let us consider water companies to be evaluated. Assuming that the set of predictor 𝑛 variables is denoted as with , is used to predict a set of 𝑥 1 ,…, 𝑥 𝑚 𝒙 𝒋 ∈ 𝑅 + 𝑗 = 1,...,𝑚 response variables denoted as with . The EAT algorithm selects 𝑦 1 ,…, 𝑦 𝑠 𝒚 ∈ 𝑅 + predictor variable and threshold , in which denotes the set of likely 𝑗 𝒔 𝒋 ∈ 𝑆 𝑗 𝑺 𝒋 thresholds for variable to separate the data into two nodes, and (Esteve et al., 𝑗 𝑡 𝑅 𝑡 𝐿 2021a). The split is achieved by minimising the sum of the mean squared of error 8 (MSE); namely, the difference between the actual and response variables derived in that particular node. Mathematically, this is done as follows: (1) 𝑅 ( 𝑡 𝐿 ) + 𝑅 ( 𝑡 𝑅 ) = 1 𝑛 ∑ ( 𝑥 𝑖 , 𝑦 𝑖 ) ∈ 𝑡 𝐿 ( 𝑦 𝑖 ― 𝑦 ( 𝑡 𝐿 ) ) 2 + 1 𝑛 ∑ ( 𝑥 𝑖 , 𝑦 𝑖 ) ∈ 𝑡 𝑅 ( 𝑦 𝑖 ― 𝑦 ( 𝑡 𝑅 ) ) 2 where presents the node of the tree, is the MSE of each node , denotes the 𝑡 𝑅(𝑡) 𝑡 𝑛 sample size, and and represent the predicted value of the response 𝒚 ( 𝒕 𝑳 ) 𝒚 ( 𝒕 𝑹 ) variable, which is derived based on the data that belongs to nodes, and , 𝑡 𝐿 𝑡 𝑅 respectively. Nodes and denote the left and right nodes of the tree, respectively. 𝑡 𝐿 𝑡 𝑅 A regression tree is visualised graphically as shown in Figure 1. 𝑡_0 𝑡_1 𝑡_3 𝑡_4 𝑡_2 𝑡_5 𝑦 ( 𝑡 3 ) 𝑦 ( 𝑡 4 ) 𝑦 ( 𝑡 5 ) Figure 1. Example of a regression tree The regression tree obtained using the EAT algorithm terminates when further meaningful splits of data are not feasible. This arises when (Breiman 𝑛 ( 𝑡 ) ≤ 𝑛 𝑚𝑖𝑛 = 5 et al., 1984; Breiman et al., 2001; Esteve et al., 2020). EAT extends the CART approach by allowing the inclusion of two characteristics of production economics. First, it allows the frontier (maximum) variable to be estimated, rather than the average of the response variable. Second, the data from each node are split in a way that the free 15 percentage of raw water collected from rivers and reservoirs was used to indicate the source of raw water. The average pumping head was used as a proxy for the energy required to extract, treat, and deliver water to end users (Molinos-Senante and Maziotis, 2018). The percentage of metered properties was used as an indicator of efficiency (Brea-Solis et al., 2017). Population density was defined as population divided by the area supplied by the water company as an environmental variable (Sala- Garrido et al., 2021a; 2021b). Table 1 presents the descriptive statistics of the variables used in the analysis. Table 1. Descriptive statisticts of the English and Welsh water companies Variables Units Mean St.Dev. Min Max Volume of water delivered Ml/year 244105 258269 10216 1049122 Water connected properties 000s 1057 1083 37 4047 Length of mains km 14560 13754 480 47151 Total expenditure Millions of £ 159 170 5 866 Water leakage % 15 5 5 36 Water taken from rivers % 23 24 0 87 Water taken from reservoirs % 32 27 0 100 Water metered properties % 30 20 3 87 Average pumping head nr 128 36 52 224 Water population density 000s/km2 0.475 0.327 0.134 2.810 Number of DMUs: 682 4. RESULTS AND DISCUSSION 4.1 Efficiency estimation Figure 3 presents the regression tree from implementing the EAT algorithm. Each node shows the identification number, MSE, number of DMUs, predictor that the split was based on, and predicted value of the response variable, which is the frontier value. All variables “i.e.” water connected properties, length of water mains, and volume of drinking water delivered) contributed towards predicting total expenditure (Figure 2). Water connected properties (wcprop) and length of mains (mainskm) had a major 16 impact on costs, as shown in the regression tree (Figure 3). Different levels of frontier expenditure were required, as shown by the different set of thresholds for predictor variables. 0 25 50 75 100 mainskm wcprop wdtot Variable Importance Figure 2. Variable importance for the EAT regression tree Delivering water to more than 2,000,000 connected properties per year could require an operating expenditure of ₤865 million, representing the maximum (frontier) expenditure. For connected properties of less than 2,000,000 and a network of pipes of 11,645 thousands of kilometres, maximum expenditure could reach ₤335 million. However, for smaller networks “i.e.” 2,824–11,645 thousands of kilometres, the 17 predicted required efficient total expenditure was ₤143.5 million. Lower expenditure was needed “i.e.” ₤34.3 million when the length of mains did not exceed 2,824 thousands of kilometres. Thus, the higher the number of connected properties, the higher the number of pipes that must be laid to deliver drinking water, raising company costs.  2089.65  11645.4  2824.8  2824.8  11645.4  2089.65 Id: 1 R: 528069.42 n(t): 682 wcprop y : [865.5] Id: 2 R: 53681.95 n(t): 563 mainskm y: [335] Id: 4 R: 6192.77 n(t): 378 mainskm y: [143.8] Id: 6 R: 98.13 n(t): 165 y:[34.3] Id: 7 R: 2197.25 n(t): 213 y:[143.8] Id: 5 R: 5953.97 n(t): 185 y :[335] Id: 3 R: 32897.56 n(t): 119 y:[865.5] Figure 3. Regression Tree for English and Welsh water companies from EAT algorithm Table 2 summarizes the efficiency scores obtained from the EAT algorithm, DEA methods and FDH methods. The efficiency scores were larger for DEA and FDH compared to EAT. Several water companies had an efficiency score of 1 for FDH and 18 DEA. In particular, 25% and 5% of the total number of DMUs under FDH and DEA, respectively, were fully efficient, and constructed the efficient frontier. Under EAT, six out of 682 DMUs (0.9%) had an efficiency score of 1. Thus, there is a risk of overfitting when estimating efficiency frontiers under FDH and DEA. This result is consistent with past research (Pereira et al., 2021). Table 2. Summary statistics of efficiency scores for English and Welsh water companies Method Mean St.Dev. Minimum Maximum Number of efficient DMUs FDH 0.789 0.211 0.252 1.000 171 DEA 0.715 0.153 0.159 1.000 34 All 0.489 0.236 0.139 1.000 6 WoCs 0.500 0.252 0.139 1.000 4 EAT WaSCs 0.476 0.214 0.154 1.000 2 The average efficiency score under FDH and DEA was estimated to be 0.789 and 0.715, respectively. Thus, the potential savings in costs among English and Welsh water companies under FDH and DEA were 21.1% and 28.5%, respectively. In contrast, the average efficiency score based on EAT was 0.489, meaning water companies could save 51.1% of costs if they operated like the efficient ones. Spearman's rank-order correlation (Table 3) showed that there the efficiency scores of EAT and FDH were more correlated than DEA. This difference was attributed to EAT and FDH estimates being generated via a step function efficient frontier, whereas DEA constructs a convex piecewise frontier. Nevertheless, differences among average efficiency scores computed using DEA, FDH and EAT demonstrated the importance of selecting an adequate method to estimate the performance of utilities (Valero-Carreras et al., 2021). This issue is even more relevant when efficiency scores are used for benchmarking purposes. For instance, the English and Welsh water industry are 19 regulated via price caps, in the form of RPI+K, where RPI is Retail Price Index and K consists of two parts, Q and X. The former represents the price increase to finance environmental improvements, while X is the offset in productivity (Helm, 2020). Recent price reviews showed that the efficiency of water utilities is evaluated through benchmarking. This information is used to set a cost baseline scenario for each utility, which is then compared with the forecast costs of a utility submitted in business plans. This information is used to set cost and revenue allowances, which are translated into price limits for five years (Ofwat, 2020). In this context, using a reliable and robust method, such as EAT, is extremely relevant. Table3. Spearman's rank-order correlation among efficiency scores EAT FDH DEA EAT 1.000 0.519 0.221 FDH 0.519 1.000 0.194 DEA 0.221 0.194 1.000 The computation of kernel densities for efficiency scores estimated using the EAT, DEA and FDH approaches allowed us to analyze the potential impact of each method used on the distribution of performance for the water companies evaluated. It also offered evidence of the concentration of efficiency around given scores. Figure 4 shows the kernel distributions of efficiency scores estimated using the EAT, DEA and FDH methods. It is illustrated that the mode of the distribution of efficiency scores estimated using the FDH and DEA methods is broadly similar whereas a different distribution was observed for efficiency scores based on the EAT approach. 20 0.25 0.27 0.29 0.31 0.33 0.35 0.37 0.39 0.41 0.43 0.0 0.2 0.4 0.6 0.8 1.0 EAT FDH DEA Density Figure 4. Kernel densities of the efficiency scores estimated using Efficiency Analysis Tree (EAT), Free Disposal Hull (FDH) and Data Envelopment Analysis (DEA) methods. The results based on EAT estimations showed that the mean efficiency of the English and Welsh water industry was 0.489. Non-relevant differences among WoCs and WaSCs were reported in terms of average efficiency scores. On average, WoCs and WaSCs could reduce their costs by 50% and 52.4%, respectively, to provide the same level of water services. Our results indicate that the water industry was characterised by high levels of inefficiency, with capacity to improve the managerial practices of companies to become more efficient, supporting previous studies (Portela et al., 2011; Byatt, 2017; Walker et al., 2020; Mocholi-Arce et al., 2021). Evaluating the distribution of efficiency scores across companies, based on EAT algorithm method, provided insights on variation in the levels of inefficiency scores (Figure 5). Most DMUs associated with both WoCs and WaSCs reported an average efficiency score ranging between 0.21 and 0.60. Thus, over the entire study period, the potential savings in costs for most English and Welsh water companies varied 21 between 40% and 80%, on average. This finding corroborates the low efficiency levels that characterised the water industry. However, the average efficiency score was lower for WoCs compared to WaSCs on occasion. Specifically, 66 out of the 382 DMUs (17.3%) related to WoCs were considerably inefficient, as their mean efficiency score did not exceed 0.20 during the study period. In contrast, this value was 9.0% for WaSCs, as 27 out of 300 DMUs had efficiency scores lower than 0.20. DMUs with the highest efficiency scores “i.e.” higher than 0.8 were mostly WoSCs. Overall, 75 out of 382 (19.6%) DMUs with average efficiency scores higher than 0.80 were WoCs. Not many WaSCs (30 out of 300; 10.0%) were within that efficiency range during the study period. Thus, while WoCs performed very well in terms of efficiency on several occasions in 1991–2020, efficiency, in most cases, did not exceed 0.60. Therefore, the management practices of WoCs must be considerably improved to reduce costs. WaSCs rarely showed high levels of efficiency. 0 10 20 30 40 50 60 70 80 90 100 [0-0.20) [0.20-0.40) [0.40-0.60) [0.60-0.80) [0.80-1.00] WoCs WaSCs Efficiency scores Number of observations Figure 5. Histogram with the distribution of efficiency scores for English and Welsh water companies. 22 As the study period covered several price reviews, the average efficiency scores of WaSCs and WoCs, estimated using the EAT approach, were split in several sub-periods to link them with the regulatory cycle of the English and Welsh water industry (Figure 6). There was an upward trend in average efficiency for both WoCs and WaSCs over the study period. Thus, the efficiency of the English and Welsh water industry improved over time. On average, the efficiency of water companies improved by 39.5% between 1991–1995 and 2016–2020 (from 0.412 to 0.575, respectively). WaSCs achieved higher efficiency gains compared to WoCs. The efficiency of WaSCs improved by 57.5% on average over the same period (from 0.364 to 0.573). The improvement in the efficiency of WoCs was also considerable over the same period, but at a lower magnitude (25.2%; from 0.461 to 0.577, respectively). Although the efficiency scores of average WaSCs were lower compared to average WoCs in the first period (1991– 1995), they caught up with the most efficient WoCs and, in some cases, became more efficient than WoCs, on average. Therefore, the efficiency scores of WoCs and WaSCs converged over time. 0.30 0.35 0.40 0.45 0.50 0.55 0.60 1991-95 1996-00 2001-05 2006-10 2011-15 2016-20 WoCs WaSCs All Regulatory periods Average efficiency scores 23 Figure 6. Average efficiency scores by regulatory period for English and Welsh water companies During the 1991-95 period, the average efficiency score for WoCs and WaSCs remained low (0.461 and 0.364, respectively). Thus, the potential savings in costs among WoCs and WaSCs could have been of 53.9% and 63.6%, respectively. Therefore, the transition period from public to private ownership had no substantial impact on cost savings, supporting previous studies. For instance, Erbetta and Cave (2007) showed that the years after privatisation had no a major impact on the efficiency of companies. An upward trend in industry efficiency was observed in 1996– 2000. This trend was mainly driven by the gains in efficiency for average WaSCs, which improved from 0.364 to 0.432, on average. In contrast, the efficiency of WoCs remained constant. The 1994 price review introduced two main policies to boost industry performance. The first policy was related to higher allowed increases in customer tariffs (and more lax cost reduction targets) to invest in maintaining and upgrading the network (Molinos-Senante and Maziotis, 2018). The second policy was associated with promoting mergers among companies. These reforms might have explained improved company efficiency. Moreover, the 1999 price review was the first in which water companies were forced to reduce the prices charged to customers. The regulator wanted to ensure that any cost savings gained in previous years were passed to customers in terms of lower prices. This price review did not appear to affect the efficiency of water companies. Further cost savings were reported that might have allowed companies to regain any losses in their profits that occurred due to reduced revenue. Cost savings were higher for WoCs compared to WaSCs, supporting previous studies. For instance, Bottasso and Conti (2009) and Portela et al. (2011) observed that 24 improving industry efficiency resulted in less efficient companies moving closer to the frontier. However, there was potential for further cost savings in daily operations and the management of assets. The efficiency of the English and Welsh water industry improved considerably during the 2006-2010 period. Regulatory incentives might positively impact efficiency, such as sharing any outperformance in expenditure with customers and financial rewards and penalties when network quality improves (Villegas et al., 2019). The efficiency of WaSCs substantially improved (from 0.439 to 0.517, on average), whereas the mean efficiency of WoCs increased slightly (from 0.508 to 0.537). Therefore, less efficient WaSCs appeared to be trying to catch-up with the most efficient WoCs in the industry. In the 2011-2016 period, where cost reduction targets were tightened, efficiency slightly reduced. The 2009 price review might have been challenging for WoCs; in contrast, the mean efficiency of WaSCs exceeded that reported by WoCs. In the last period (2017–2020), the situation was reversed. During this period, the regulator introduced several incentives for companies to improve efficiency. For instance, it introduced a set of common and bespoke performance targets to monitor economic and environmental performance, and imposed financial rewards/penalties when these targets were met/not met (Villegas et al., 2019). The average efficiency of both WoCs and WaSCs became similar; however, considerable inefficiency remained, with the potential for improvement. 4.2 Influence of environmental variables on efficiency scores To analyse the influence of environmental variables on the efficiency of English and Welsh water companies, bootstrap regression analysis was conducted (Table 4). The 31 Breiman, L., Friedman, J., Stone, C. J., Olshen, R. A. (1984). Classification and regression trees. Taylor & Francis. Carvalho, P., Marques, R.C. (2016). Estimating size and scope economies in the Portuguese water sector using the Bayesian stochastic frontier analysis. Science of the Total Environment, 544, 574-586. Castillo-Jimenez, J., Montañes, A., Picazo-Tadeo, A.J. (2019). Performance in the treatment of municipal waste: Are European Union member states so different?.Science of the Total Environment, 1305-1314. Cetrulo, T.B., Marques, R.C., Malheiros, T.F. (2019). An analytical review of the efficiency of water and sanitation utilities in developing countries. Water Research, 161, 372–380. Charnes, A., Cooper, W.W., Rhodes, E. (1978). Measuring the efficiency of decision making units. European Journal of Operational Research, 2, 429-444. Cooper, W.W., Seiford, L.M., Zhu, J. (2011). Handbook on Data Envelopment Analysis. International Series in Operations Research & Management Science. Springer. Daraio, C., Kerstens, K., Nepomuceno, T., Sickles, R.C. (2020). Empirical surveys of frontier applications: a meta-review. International Transactions in Operational Research, 27 (2), 709-738. De Witte, K., Marques, R.C. (2010). Incorporating heterogeneity in non-parametric models: a methodological comparison. International Journal of Operational Research, 9 (2), 188–204. 32 Ding, L., Yang, Y., Wang, L., Calin, A.C. (2020). Cross Efficiency Assessment of China's marine economy under environmental governance. Ocean and Coastal Management, 193,105245. Drusiani, R., Gatta, M., Gerelli, G.G. (2013). Regulation of water service and efficient use of water. Water Science and Technology: Water Supply, 13 (4), 932-938. Dyckhoff, H. (2018). Multi-criteria production theory: foundation of non-financial and sustainability performance evaluation. Journal of Business Economics, 88 (7-8), 851– 882. D'Inverno, G., Carosi, L., Romano, G. (2021). Environmental sustainability and service quality beyond economic and financial indicators: A performance evaluation of Italian water utilities. Socio-Economic Planning Sciences, 75, 100852. Elbeltagi, A., Pande, C.B., Kouadri, S., Islam, A.R.M.T. (2022). Applications of various data-driven models for the prediction of groundwater quality index in the Akot basin, Maharashtra, India. Environmental Science and Pollution Research, 29 (12), 17591- 17605. Erbetta, F., Cave, M. (2007). Regulation and Efficiency Incentives: Evidence from the England and Wales Water and Sewerage Industry. Review of Network Economics, 6 (4), 425–452. Esteve, M., Aparicio, J., Rabasa, A., Rodriguez-Sala, J. J. (2020). Efficiency analysis trees: A new methodology for estimating production frontiers through decision trees. Expert Systems with Applications, 162, 113783. 33 Esteve, M., Aparicio, J., Rodriguez-Sala, J. J., Zhu, J. (2022). Random Forests and the measurement of super-efficiency in the context of Free Disposal Hull. European Journal of Operational Research, In press. Esteve, M., España, V. J., Aparicio, J., Barber, X. (2021b). eat: Efficiency Analysis Trees. R package version 0.1.2. Available at: https://cran.r-project.org/web/packages/eat. Esteve, M., Rodriguez-Sala, J.J., Lopez-Espin, J.J., Aparicio, J. (2021a). Heuristic and Backtracking Algorithms for Improving the Performance of Efficiency Analysis Trees. IEEE Access, 9, 17421-17428. Farrell, M. J. (1957). The measurement of productive efficiency. Journal of the Royal Statistical Society: Series A (General), 120 (3), 253–281. Ferreira, D.C., Marques, R.C. (2017). A step forward on order-α robust nonparametric method: inclusion of weight restrictions, convexity and non-variable returns to scale. Operational Research, 1-36. Filippini, M., Hrovatin, N., Zorić, J. (2018). Cost efficiency of Slovenian water distribution utilities: An application of stochastic frontier methods. Journal of Productivity Analysis, 29 (2), 169-182. Goh, K.H., See, K.F. (2021). Twenty Years of Water Utility Benchmarking: A Bibliometric Analysis of Emerging Interest in Water Research and Collaboration. Journal of Cleaner Production, 284, 124711. Green, C.T., Ransom, K.M., Nolan, B.T., Liao, L., Harter, T. (2021). Machine learning predictions of mean ages of shallow well samples in the Great Lakes Basin, USA. Journal of Hydrology, 603, 126908. 34 Guerrini, A., Romano, G., Leardini, C. (2018). Economies of scale and density in the Italian water industry: A stochastic frontier approach. Utilities Policy, 52, 103-111. Guerrini, A., Romano, G., Leardini, C., Martini, M. (2015). The effects of operational and environmental variables on efficiency of Danish water and wastewater utilities. Water (Switzerland), 7 (7), 3263-3282. Helm, D. (2020). Thirty years after water privatization - Is the English model the envy of the world?.Oxford Review of Economic Policy, 36 (1), 69-85. Iyer, K.C., Jain, S. (2019). Performance measurement of airports using data envelopment analysis: A review of methods and findings. Journal of Air Transport Management, 81, 101707. James, G., Witten, D., Tibshirani, R., Hastie, T. (2013). An introduction to statistical learning with applications in R. Springer, New York. lo Storto, C. (2020). Measuring the efficiency of the urban integrated water service by parallel network DEA: The case of Italy. Journal of Cleaner Production, 276, 123170. Marques, R.C., Berg, S., Yane, S. (2014). Nonparametric benchmarking of Japanese water utilities: Institutional and environmental factors affecting efficiency. Journal of Water Resources Planning and Management, 140 (5), 562-571. Mergoni, A., De Witte, K. (2022). Policy evaluation and efficiency: a systematic literature review. International Transactions in Operational Research, 29 (3), 1337- 1359. Mocholi-Arce, M., Sala-Garrido, R., Molinos-Senante, M., Maziotis, A. (2021). Performance assessment of water companies: A metafrontier approach accounting 35 for quality of service and group heterogeneities. Socio-Economic Planning Sciences, 74, 100948. Molinos-Senante, M., Maziotis, A. (2018). Assessing the influence of exogenous and quality of service variables on water companies’ performance using a true-fixed stochastic frontier approach. Urban Water Journal, 15 (7), 682–691. Molinos-Senante, M., Maziotis, A., Sala-Garrido, R. (2014). The Luenberger productivity indicator in the water industry: an empirical analysis for England and Wales. Utilities Policy, 30, 18–28. Molinos-Senante, M., Porcher, S., Maziotis, A. (2017). Impact of Regulation on English and Welsh Water-Only Companies: An Input Distance Function Approach. Environmental Science and Pollution Research, 24 (20), 16994–17005. Ofwat (2020). Final determinations. Available at: https://www.ofwat.gov.uk/regulated-companies/price-review/2019-price- review/final-determinations/ Pereira, M.A., Camanho, A.S., Marques, R.C., Figueira, J.R. (2021). The convergence of the World Health Organization Member States regarding the United Nations’ Sustainable Development Goal ‘Good health and well-being’. Omega (United Kingdom), 104, 102495. Pinto, F.S., Simões, P., Marques, R.C. (2017). Water services performance: do operational environment and quality factors count?.Urban Water Journal, 14 (8), 773- 781. 36 Portela, M.C.A.S., Thanassoulis, E., Horncastle, A., Maugg, T. (2011). Productivity change in the water industry in England and Wales: application of the meta-Malmquist index. Journal of Operational Research Society, 62 (12), 2173–2188. Pratap, S., Daultani, Y., Dwivedi, A., Zhou, F. (2022). Supplier selection and evaluation in e-commerce enterprises: a data envelopment analysis approach. Benchmarking, 29 (1), 325-341. Rebai, S., Yahia, F.B., Essid, H. (2019). A graphically based machine learning approach to predict secondary schools performance in Tunisia. Socio-Economic Planning Sciences, 70, 100724. Renzetti, S., Dupont, D.P. (2009). Measuring the technical efficiency of municipal water suppliers: The role of environmental factors. Land Economics, 85 (4), 627-636. Saal, D.S., Parker, D., Weyman-Jones, T. (2007). Determining the contribution of technical change, efficiency change and scale change to productivity growth in the privatized English and Welsh water and sewerage industry: 1985-2000. Journal of Productivity Analysis, 28 (1-2), 127-139. Sala-Garrido, R., Mocholi-Arce, M., Molinos-Senante, M., Maziotis, A. (2021a). Marginal abatement cost of carbon dioxide emissions in the provision of urban drinking water. Sustainable Production and Consumption, 25, 439-449. Sala-Garrido, R., Mocholi-Arce, M., Molinos-Senante, M., Smyrnakis, M., Maziotis, A. (2021b). Eco-Efficiency of the English and Welsh Water Companies: A Cross Performance Assessment. International Journal of Environmental Research and Public Health, 18 (6), 2831, 1-19. 37 See, K.F. (2015). Exploring and analysing sources of technical efficiency in water supply services: Some evidence from Southeast Asian public water utilities. Water Resources and Economics, 9, 23-44. Simar, L., Wilson, P.W. (2007). Estimation and inference in two-stage, semiparametric models of production processes. Journal of Economics, 136 (1), 31–64. Torres, M., Morrison Paul, C.J. (2006). Driving forces for consolidation or fragmentation of the US water utility industry: A cost function approach with endogenous output. Journalof Urban Economics, 59 (1), 104-120. Valero-Carreras, D., Aparicio, J., Guerrero, N.M. (2021). Support vector frontiers: A new approach for estimating production functions through support vector machines. Omega (United Kingdom), 104, 102490. Villegas, A., Molinos-Senante, M., Maziotis, A. (2019). Impact of environmental variables on the efficiency of water companies: a double bootstrap approach. Environmental Science and Pollution Research26, 31014–31025. Walker, N.L., Williams, A.P., Styles, D. (2020). Key performance indicators to explain energy & economic efficiency across water utilities, and identifying suitable proxies. Journal of Environmental Management, 269, 110810. Wang, L., Zhou, Z., Yang, Y., Wu, J. (2020). Green efficiency evaluation and improvement of Chinese ports: A cross-efficiency model. Transportation Research Part D, 88, 102590. Wang, X., Han, L., Yin, L. (2017). Environmental Efficiency and Its Determinants for Manufacturing in China. Sustainability, 9, 47. 38 Zakowska, I., Godycki-Cwirko, M. (2020). Data envelopment analysis applications in primary health care: a systematic review. Family practice, 37 (2), 147-153. Declaration of interests ☒The authors declare that they have no known competingfinancialinterests or personal relationships that could have appeared to influence the work reportedin this paper. ☐The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: HIGHLIGHTS Overfitting issues were overcome in efficiency assessment by using efficiency analysis trees. Average efficiency of water companies over 1991 – 2000 was 0.489 39 Environmental variables influencing performance of water companies were identified. MMS: Conceptualization; Visualization; Writing – Review & Editing AM: Methodology; Sofware; Writing – Original Draft RSG: Data curation; Writing – Review & Editing MMA: Writing – Review & Editing; Supervision