TCNZ Efficiency Study Based on Stochastic Frontier Analysis (SFA)

Transcription

1 TCNZ Efficiency Study Based on Stochastic Frontier Analysis (SFA) Public version September 2002

2 Management summary PwC has been commissioned by Telecom New Zealand (TCNZ) to produce an efficiency study that measures TCNZ s level of efficiency of operation. The purpose of the analysis is to determine whether an efficiency adjustment factor should be incorporated into our (PwCC s) assessment of the costs incurred by TCNZ in meeting its Telecom Services Obligations. The approach adopted for the assessment of TCNZ s efficiency is a robust and widely recognised method (Stochastic Frontier Analysis, SFA). This method relies on the comparison of the operator under consideration with a significant number of other operators. It has been used by telecommunications (and other) regulators, as well as by operators in various countries to measure efficiency. This approach has a number of advantages: In contrast to simple efficiency indicators which focus on a small number of inputs and / or outputs, the SFA technique takes into account the total input and the total output of each operator and therefore provides a much more comprehensive picture. All measurable factors (environmental, geographical, et cetera) that have an impact on costs can be taken into account and adjusted for, so that each company s efficiency level ( frontier ) is measured against the greatest achievable efficiency level in its own country. Econometric techniques are used to ensure full comparability of operators despite the differences between their operating environments. Observation errors that may occur in the data collection process can be filtered out (this is not possible with deterministic methods). The SFA method separates genuine inefficiency from observation errors. For the purpose of this study, input and output data have been collected for TCNZ and for over 50 US LEC operators. In a first step, a number of adjustments are made to ensure comparability of this data, e.g. adjustments for exchange rates, differences in accounting conventions, etc. Based on the resulting adjusted cost data, we have tested the impact of factors that could potentially lead to cost differences between the various operators environments, e.g. differences in population density etc. Using regression analysis, the significance and quantitative impact of a variety of potential factors are tested. The regressions show that the only statistically significant environmental factor impacting on cost levels was the length of sheath per line, which is an indicator of population dispersion. 2

3 The results of the efficiency study suggest that TCNZ ranks 8th in the sample of over 50 operators. If the top decile of operators is taken as the most appropriate benchmark, then the degree of inefficiency if TCNZ s operations is calculated to be: 1.5% using the mean estimation method, which determines the average value of the inefficiency, given the residuals; or 3.0% using the mode estimation method, which calculates the most likely value of the inefficiency, given the residuals. More information on the two methods of estimating the degree of inefficiency is given in Appendix C. With a view to ensuring that the TSO costs incurred by TCNZ are not overstated, a 3% efficiency adjustment factor has been used in the TSO cost analysis. 3

4 Table of contents Management summary Introduction Measurement of Efficiency Outline of the approach Results per phase and the efficiency benchmark Conclusions Appendices...36 Appendix A: Companies in the efficiency benchmark...37 Appendix B: FCC asset and expenses categories used in the study...40 Appendix C: Mathematical background on the stochastic frontier analysis...43 Appendix D: Details of the estimated model

5 1 Introduction 1.1 Background PwC Consulting has conducted an efficiency study to establish the level of efficiency of TCNZ, relative to other operators and relative to a theoretical measure of full 100% efficiency ( efficiency frontier ). The assessment of TCNZ s efficiency is relevant to a number of issues affecting TCNZ and the industry, including the assessment of interconnection charges and the costs of Telecommunications Service Obligation (TSO) provision. The latter issue provided the main reason for carrying out this study The Commerce Commission requires a calculation of the net costs to TCNZ of providing the service required by the TSO Deed. PwC Consulting, on behalf of TCNZ, has developed a model to quantify the net costs incurred by TCNZ in meeting its obligations under the TSO. The costs included in this model may include potential inefficiencies in TCNZ s operations to the extent 1 that the cost information is based on costs actually incurred by TCNZ. PwC Consulting was commissioned to assess TCNZ s efficiency, so that the TSO cost calculation could be adjusted as appropriate to reflect the costs that would be incurred by an efficient operator Generally, efficiency is defined as the ratio of output to the input used in its production. Telecommunications operators produce a number of different outputs (lines, minutes, calls, leased lines, etc) and, in the process of providing these services, use a large number of inputs, including different types of labour, buildings, network assets, externally provided services, etc. Some efficiency studies only consider one input and/or one output, the reason being that simple indicators (such as employees per line) are used and taken to give an indication of the overall level of efficiency. Some of the limitations of this type of approach are discussed in Chapter The approach adopted in this study for the assessment of TCNZ s efficiency is a robust and widely recognised method (Stochastic Frontier Analysis, SFA). This method relies on the comparison of the operator under consideration with a significant number of other operators. It has been used by telecommunications (and other) regulators, as well as by operators in various countries to measure efficiency The result of the efficiency study is the efficiency of TCNZ, and the relative score of TCNZ within a set of 51 operators (50 LECs and TCNZ). 1 Costs of the access network are based on a bottom-up model of an efficiency operator and, as such, reflect the costs of efficiency operations. Other costs, such as retail costs etc. are based on actual accounting data and would therefore include potential inefficiencies, if any. 5

6 1.2 Outline of the report Chapter 2 discusses the definition and measurement of efficiency, as well as a detailed discussion of the method used (SFA) to measure efficiency and on the relative merits of this approach compared to other approaches. Chapter 3 outlines a five-phase approach of the efficiency study, and the result per phase are presented in chapter 4. Chapter 5 presents the conclusions of the study. 6

7 2 Measurement of Efficiency 2.1 Definition of efficiency In this study, the efficiency of an operator is defined by the ratio of output provided to the input used in producing the output. Examples of input or production factors are the amounts of labour and capital required to produce and deliver a certain amount of products and services. The products and services provided are the output. Examples of products and services are the number of calls, the number of lines, and the number of call minutes. Figure 1 illustrates the input-output process. Production factors Products/services Buildings Systems Network expenses Capital costs of network Operating expenses O P E R A T O R Number of calls Number of minutes Number of lines Figure 1 Production factors vs. products and services In order to construct the input/output ratio, the different components of input and output, respectively, have to be aggregated into a single number. Various methods of measuring efficiency based on comprehensive measures of input and output are discussed in the following section. 2.2 Approaches to measuring efficiency There are a number of methods that have been used in the past to assess different operators efficiency levels. These approaches can be classified as follows: Index numbers (e.g. Total Factor Productivity (TFP)) 7

8 Mathematical approaches (e.g. simple ratios or Data Envelope Analysis (DEA)) Parametric approaches (e.g. Stochastic Frontier Analysis (SFA)) We provide a brief discussion of the main methods in the next few paragraphs as background to compare current best practice efficiency analysis to alternative methods. (i) Index Numbers / Total Factor Productivity (TFP) TFP is a method which uses weighted sums of inputs in order to obtain a measure for the input required. Obviously, this procedure faces particular difficulties when there are no appropriate unit definitions for some of the inputs (such as capital goods). For this reason, comparability between different operators using different production techniques (say network architectures) is impeded. Therefore the TFP method is best suited to intertemporal (dynamic) analyses relating to the same operator in which case the same definition of aggregate input can be used for each measurement The weights attached to different inputs are normally chosen with a view to reflecting the particular input s share of total input costs. However, the weights are taken to be constant for different operators or different periods. One implication is that the TFP does not penalise operators who pay too high a price for their inputs Alternatively, input amounts can be measured - or approximated - by the costs incurred in providing the inputs. The advantage of this method based on unit costs is that the resulting measure of an operator s efficiency includes its ability to procure inputs at low prices. On the other hand, this might introduce a bias in favour of operators in whose countries inputs are cheaper due to exogenous factors which are beyond the control of the operator. Where such factors are identified appropriate adjustments need to be made. (ii) Mathematical Approaches (a) Simple ratios Frequently, telecoms operators performance has been measured with the help of simple indicators or indices such as the number of employees per line. Such measures, however, can lead to distorted results. Take, for instance, the case of an operator outsourcing various stages of the production process. The resulting ratio of employees per line will tend to be very low. However, this should not be taken to imply that the operator is more efficient than others as all remaining inputs are neglected. A similar criticism applies to the ratio of the number of employees per call minute Less obvious are the drawbacks of some other simple indicators such as costs per subscriber line. Different operators may, for instance, focus on different market segments (e.g. residential and business) with markedly different calling patterns. In this case, one 8

9 operator s traffic per line may exceed another s substantially. The costs of dimensioning a network for different traffic levels can hardly be regarded as an indication of inefficiency. (b) Data Envelope Analysis Historically, production and cost functions have been estimated using OLS regression models. For the purposes of efficiency measurement the resulting average function is a misleading indicator of efficient production possibilities in both theory and practice. In practice, an average performance standard will tend to institutionalise inefficiency. This can occur because, in reducing what appears to be attainable, average standards act as a disincentive to further improvements in performance. Furthermore, an average production function is inconsistent with the theoretical notion of a boundary function which reflects maximising behaviour Frontier performance comparisons (such as DEA), on the other hand, flow directly from the definition of the production function itself. Broadly speaking, production is a process of physical transformation in which inputs are combined to generate output. DEA uses liner programming techniques in order to establish the location of the efficiency frontier DEA can be illustrated by plotting input and output levels of different operators in a scatter diagram. The data points at the lower end of the scatter plot are connected by straight lines ( convex hull ). An assumption is made that convex combinations of inputs result in the corresponding convex combination of outputs. Intuitively speaking, this amounts to saying that a company that produces lines very efficiently and another company producing minutes very efficiently are combined (conceptually) to establish the costs an efficient operator would incur in producing a given amount of both outputs. unit cost efficiency frontier activity e.g. # of lines Figure 2 DEA Analysis 9

10 The above diagram shows an example with one output only. As a linear programming approach, the DEA method can also handle multi dimensional problems In order to carry out a DEA analysis one has to collect data describing input - output combinations (output defined as activities and input defined as cost of the input bundle required to carry out the activity) for a number of different operators It should be noted that DEA analysis has a number of limitations: No framework to allow for data errors (or incompatibilities in the observed data which cannot be related to explanatory variables) Very sensitive to outliers. The calculated frontier may be warped if the data are contaminated by statistical noise. Systematic over-estimation of inefficiency (iii) Parametric Approaches: Stochastic Frontier Analysis One of the requirements of meaningful studies based on benchmarking, naturally, is that one has to compare like with like. In the case of telecoms operators and their respective performance a number of adjustments may have to be undertaken in order to ensure comparability by taking into account country and/or operator specific differences, such as: different price levels. different wage rates (in contrast to capital goods which are freely traded on international markets, some input prices differ considerably from country to country due to limited mobility. In particular, this applies to labour). different accounting conventions (depreciation rates, treatment of pension costs, redundancy costs, capitalisation of interest, etc.). different operating environments (geographical factors). economies of scale and scope (due to larger subscriber numbers, higher line density or higher traffic levels per line, operators unit costs may differ) In order to make adjustments for some of the above differences it may be necessary to run regressions in order to establish a relationship between costs and different levels of the country specific variable (such as output levels, population density etc.) SFA is an advanced econometric approach which allows these adjustments to be made In a first stage of the analysis, the comparability of different operators data is ensured through adjustments for those factors where the impact is known (e.g. different exchange rates, wage rates, etc). In the second step of the SFA, regressions are run in order to quantify the impact of other factors on cost levels in different countries. Finally, observation errors are 10

11 filtered out in order to ensure that the outcome is not affected by inaccuracies in the data sources SFA is well known in the literature and has been used by telecom regulators and operators. In the UK, Oftel has applied this methodology to BT. Subsequently, it has been used by telecoms operators in regulatory negotiations in the Netherlands, Ireland and in the Caribbean The SFA methodology has a number of important advantages, in particular: 1. In contrast to simple efficiency indicators which focus on a small number of inputs and / or outputs, the SFA technique takes into account the total input and the total output of each operator and therefore provides a much more comprehensive picture. 2. All measurable factors (environmental, geographical, et cetera) that have an impact on costs can be taken into account and adjusted for, so that each company s efficiency level ( frontier ) is measured against the best available efficiency level in its own country. Econometric techniques are used to ensure full comparability of operators despite the differences between their operating environments. 3. Observation errors that may occur in the data collection process can be filtered out (this is not possible with deterministic methods). The SFA method separates genuine inefficiency from observation errors. In the light of these points, as well as international precedent, we have adopted the SFA approach for this study. 2.3 Benchmark against Local Exchange Carriers In order to conduct an efficiency study based on Stochastic Frontier Analysis (SFA), extensive operational and financial data is required. Most telecom operators in Europe (and beyond) do not publish detailed cost data and/or operational data required for an efficiency study based on SFA. However, data for 52 Local Exchange Carriers (LECs) in the US is publicly available (from the Federal Communications Commission (FCC)), and has been used for this study (2 carriers were later removed from the sample) The choice of the US LECs as an efficiency comparator is based on a number of factors: The local and inter-exchange telephony markets in which the LECs operate are relatively competitive, as against those in other countries, and the incentives to improve efficiency have consequently been relatively high; 11

12 The incentives for efficiency improvement have also been improved over the past 15 years by a marked shift away from rate-of-return regulation towards price cap regulation; Unlike their counterparts in some other counties, the LECs are privately owned and financed, and are therefore subject to competitive pressure from the capital markets to improve their financial performance; LECs operational and financial data is publicly available, which allows a transparent in-depth comparative analysis; and The LECs have been used as an efficiency benchmark by telecommunications regulators in several other countries e.g. efficiency study of British Telecom conducted by NERA for Oftel, and the study carried out by PwC on behalf of KPN. The latter was produced for submission to OPTA, the Dutch regulatory authority. The same approach has been used by NERA on behalf of the ODTR (Ireland) and by PwC on behalf of eircom, the Irish incumbent operator. 2.4 Time frame of the study The financial and operational data of the LECs is based on the most recent financial year for which data is available for the LEC operators, which corresponds to the calendar year As TCNZ s financial book year runs from July to June an adjustment has been made by taking the average of financial year and financial year This is expressed in the following formula: Value 31 /12 / 2000 = 0.5* Value financialy ear * Value financialyear Although the efficiency factor is calculated using data for the year 2000 the results are likely to be valid when applying this efficiency to data from the first half of year 2002, as the ranking (according to efficiency levels) of the US operators and TCNZ is unlikely to have changed significantly over the past 18 months Data for the calendar year 2001 is likely to be published by the FCC in December 2002 (the exact date varies from year to year) at which point the study could be updated to compare TCNZ financial years 2000/1 2001/2 with LEC data for the year

13 3 Outline of the approach 3.1 The approach Before the SFA method can be applied, the financial and operational data must be collected for all companies in the sample, and several phases are required to ensure full comparability of different operator s cost and operational data. The SFA methodology comprises the following phases: Phase 1: Collect operational and financial data of TCNZ and the LECs, Phase 2: Adjust financial data to ensure comparability, Phase 3: Determine the volume measurements to be used and define an output function, Phase 4: Determine the environmental factors to be considered, Phase 5: Conduct regression analysis to estimate parameters, Phase 6: Apply Stochastic Frontier Analysis to filter the results for observation errors The final phase yields both the efficiency of TCNZ and TCNZ s rank in the sample. The approach is illustrated in the following diagram. TCNZ output volume & operational data LECs output volume & operational data LECs accounts TCNZ s Network related accounts Make data sets comparable through several adjustments Identify parameters for inclusion in output function Identify environmental factors for adjustment Conduct regression analysis to estimate parameters Split genuine inefficiency and observation errors, and establish the efficiency of all operators Phase 1 Phase 2 Phase 3 Phase 4 Phase 5 Phase 6 Figure 3 Overview of the approach The following subsections discuss each phase in turn. 13

14 3.2 Phase 1: Collect operational and financial data of TCNZ and the LECs The financial and operational data of the LECs that is used in the efficiency study, i.e. the data of 2000, was made publicly available by the Federal Communications Commission (FCC). TCNZ has provided its operational and financial data for financial years 1999/2000 and 2000/2001. The time difference in financial book year between TCNZ and the LECs is accounted for as described in section 2.4. Requirements for operational data Operational data is used in several phases during the study: 1. In phase 3, the output of an operator is calculated with the help of operational data. The total produced output of an operator in this study is expressed as a function of the number of calls (local, long distance, and international), the number of minutes (local, long distance, and international), and the number of access lines. 2. In phase 4, the unit costs are corrected for environmental factors or indicators that are based on operational data. The operational data is used to construct indicators that are associated with one of the four categories of environmental factors: geography (e.g. %age aerial cable), population dispersion (e.g. average length of local loop per access line), economies of scale (e.g. # lines), and calling patterns, (e.g. %age local calls). Requirements for financial data The following types of costs have been considered in this study: Asset costs, i.e. depreciation and cost of capital. The cost of depreciation and the cost of capital are derived from the Gross and Net Book Values of the assets, respectively. Operational expenses, e.g. costs of network support The total costs incurred by an operator in producing its output can be calculated in a straightforward way: Total Costs = operational expenses + depreciation + cost of capital The following categories of operational expenses of LECs are included in the study (a more detailed explanation can be found in Appendix B): Asset Expenses 1. Land and support assets; 14

15 2. Central office switching assets; 3. Operator Systems assets; 4. Central Office Transmission assets; 5. Cable and Wire Facilities assets. Operating Cost Expenses 1. Network Support expenses 2. General expenses 3. Central office switching expenses 4. Operator Systems expenses 5. Central office transmission expenses 6. Cable and wire facilities expenses 7. Network operations expenses 8. Marketing expenses 9. Services expenses 10. Corporate operations expenses 11. General and administration expenses 12. Public telephone terminal equipment 13. Other terminal equipment 14. Other property plant and equipment expenses 15. Access expenses 3.3 Phase 2: Adjust financial data to ensure comparability, To make the financial data comparable some straightforward adjustments are made through: Corrections for the exchange rate; TCNZ s financial information is expressed in New Zealand Dollars and must be converted to US$. Corrections for the difference in depreciation between TCNZ and the LECs, by applying the depreciation rates of TCNZ to the LECs. Depreciation policies in the US differ from those in New Zealand. For this reason, the same average asset life is taken, per asset category, for each operator. Applying the Costs of Capital rates of TCNZ to the LECs As it is the current efficiency level that in important for the purpose of the study all asset data, for the LECS and for TCNZ, was converted from HCA to CCA values using the following formula: CCA HCA = ((1 + I a ) *(1 + I g )) [(1 NBV / GBV ) / D] 15

16 Where: CCA = Value at Current Cost Accounting HCA = Value at Historic Cost Accounting NBV = Net Book Value GBV = Gross Book Value I a = Telecom specific inflation % I g = General inflation rate of New Zealand or the US D = weighed average of depreciation percentage over the asset categories Also to ensure that the same areas of operations are being considered the following costs are removed from TCNZ costs as these would be absent from the LECs. Assets In order to ensure comparability with the US LEC operators, the following assets are removed from the analysis and these adjustments feed into the capital cost and depreciation calculations: Assets associated with mobile services Assets associated with data services Customer premises equipment (we also exclude these from the US LEC data) Assets associated with the international conveyancing i.e. international satellite and earth station facilities and international undersea cable A proportion of operating systems costs to account for emergency calls. Operating costs The following operating costs are removed from TCNZ s total operating costs: Payments to other mobile operators i.e. Vodafone (these payments are not made by LEC operators in the US due to called part pays arrangements) Payments to TCNZ mobile arm (these payments are not made by LEC operators in the US due to called part pays arrangements) Costs associated with mobile services (not part of LEC activities) Costs associated with data services (except leased lines) - (not part of LEC activities) Costs associated with providing customer premisise equipment (C.P.E.) - (not part of LEC activities) Costs associated with international outgoing calls (corresponding costs are incurred by international operators in the US, not by the LECs) 16

17 Bad debts (these are included as an offset against revenue in the LEC data) Property rates Sales and Marketing costs for national calls (in US these costs are incurred by long distance operators rather than by the LECs) Billing costs for national calls (in US long distance operators incur these costs) Costs associated with emergency calls Other Capitalised interest costs are removed as these are not included in the US LEC GAAP figures. GAAP adjustments for restructuring and provisions these are added into TCNZ costs as they would be included in the US LEC data. 3.4 Phase 3: Determine the volume measurements to be used and define an output function Output measures used Lines The total number of lines used to calculate output includes total switched lines (business, residential PR2 and payphone for TCNZ and also special access lines for the LECs) and 64k equivalent leased lines The LECs do not report leased line numbers, but they publish leased line revenue figures. We have, therefore, used an average price per 64k equivalent leased line to generate total leased line numbers. This average price was based on information on European operators average leased line prices and converted to US leased line prices using the relative leased line price index published by Teligen The quantification of leased lines is based on the number of leased line ends rather than pure leased line numbers. For TCNZ we have assumed that all national leased lines have 2 ends and international leased lines have one end. For LECs we assume the same proportion of international leased lines as in New Zealand and that domestic lines are evenly split between intra-state leased lines (comprising of 2 ends for the LEC) and inter-state leased lines (1 end). 2 Source: Teligen T basket. National and International Leased Lines 17

18 Minutes The figures used for the number of minutes include all local, national, international and interconnect minutes For TCNZ information on the actual number of minutes was used in the model but for the LECs this minute information needed to be calculated from the number of calls using an average call duration. An average LEC call duration of 3.6 mins for all calls was assumed for all the US LECs based on the FCC data on inter-lata minutes and calls after adjusting for holding time and set up time This minute data is converted to switched minutes using the routing factors for TCNZ and routing factors from the Hatfield model for the US LECs 3 as shown. LEC routing Factors tandem local Local Intra Lata Inter Lata Table 1: LEC routing factors The same routing factors have been used in previous efficiency studies. Calls Calls for US LECs are reported in the FCC information. TCNZ call numbers were obtained and then scaled up according to the percentage of successful calls (76% from TCNZ). Total output Each operator s output is calculated with a formula based on a weighted average ( Cobb Douglas ) type function of different output indicators: the number of calls, the number of lines (switched lines plus 64k equivalent leased lines), and the number of switched minutes. Again, similar output factors were used by NERA efficiency study for Oftel, as well as in other studies (e.g. previous PwC studies). The formula that is used in this study is as follows: Output = (# lines) α x (# minutes) β x (#calls) γ returns to scale). α,β,γ 0 (α+β+γ=1 would imply constant 3 As in the anaylsis of BT conducted by NERA on behalf of Oftel. 18

19 Properties of the output function The mathematical formula, also known as a Cobb-Douglas function, has the following properties: Substitution of factors: changes in the factors are exchangeable. As an example: with a fixed output, the number of calls must decrease if the number of lines and/or the number of minutes increase. Positive relationships: an increase in the number of calls, the number of lines, or the number of minutes will yield an increase in the output. Comparability: the formula yields a single number for the output that allows a comparison between different operators output levels. Differences between operators for each of the factors, and one factor to the other are dealt with automatically. Values for the parameters in the formula The values of these parameters are estimated as part of the regression analysis using the method described in section Phase 4: Determine the environmental factors to be considered The costs are adjusted for environmental factors, i.e. such factors that may affect the input-output relationship and which are not under control the operators control - while the extent of their impact is unknown (where the impact is known, adjustments can be made without the help of regressions, see phase 2) These environmental factors are likely to significantly distort simplistic efficiency benchmarks. As an example: the mix of calls may differ between operators, some providing a much larger proportion of local calls than others. As local calls tend to have less switching stages and therefore lower unit costs, one would expect an operator with a relatively high proportion of local calls to achieve relatively low unit costs (other things being equal) This effect can be identified, measured, and adjusted for with the help of regression analysis. This example is illustrated in Figure 4. 19

20 Unit Carrier 1 Carrier 2 %age local calls Figure 4 Theoretical example of regression based adjustment The environmental factors that are analysed with the help of regression analysis, relate to four categories of factors that impact on the difference between cost levels in New Zealand and other countries. For each factor, the hypothesis about the relationship between the costs per unit output and the factor in question is formulated and tested. 1. Economies of scale Number of lines. Hypothesis: negative relationship between unit costs and number of lines due to economies of scale, because the required expenses and assets should increase less than proportionally if the number of lines increases. Number of calls. Hypothesis: negative relationship between unit costs and number of calls due to economies of scale, because the required expenses and assets should increase less than proportionally if the number of calls increases. 2. Calling pattern 20

21 %age local calls. Hypothesis: negative relationship between costs and the percentage of local calls, because a higher percentage implies that less national and international traffic needs to be facilitated, and thus that the operating expenses and asset costs will be lower. Calling rate per line. Hypothesis: negative relationship between unit costs and the calling rate per line, because with a higher calling rate the total costs per line can be divided over more units of outputs (as output units comprise calls as well). 3. Geography %age aerial cable. Hypothesis: relationship would depend on relative importance of capital and maintenance costs. A negative relationship between costs and the percentage of aerial cable is expected on the capital side because aerial cable capital requirements are lower than those for underground cable (due to high duct costs), but maintenance costs of aerial cable are higher (repairs required depending on weather conditions, etc). 4. Population dispersion Average length of local loop per access line. Hypothesis: Positive relationship between costs and the average length of local loop per access line, because a higher average length implies higher capital costs. Average size of the switch. Hypothesis: Negative relationship between costs and the average size of the switch, because when the population is concentrated in few locations, larger switches (and fewer) are used. Therefore costs are lower. Length of sheath 4 per line. 4 Outer jacket (usually metal or plastic) that surrounds copper and fibre cables to prevent water damage to the cables inside. 21

22 Hypothesis: Positive relationship between costs and the length of sheath per access line, because a high population dispersion requires longer cables, and increased sheath length, and thus more costs. 3.6 Phase 5: Conduct regression analysis to estimate parameters Regression analysis is used to establish a relationship between costs, output measurements and environmental factors. The regressions show whether the relationship between a factor and the scaled costs is significant, and if so, it shows how strong the relationship is. The econometric model used incorporates the above factors and is given by the following formula: Total _ Costs = con * Lines α * Calls β * Min δ * ( Sheath / Line) ρ * ( AerialCable) σ ε Where: con: Constant Lines: Number of lines (customers) Calls: Number of calls Min: Number of minutes Sheath/line: Weighted average of scaled length of sheath per line (weighted with the % of aerial and non-aerial cables) % Aerial cable: Percentage aerial cables of all cables Other environmental variables (with their parameters) α,β,δ,ρ,σ: Parameters to be estimated by regressions ε: Errors in the model (due to inefficiency and modelling errors) Based on this general model, a sequence of regressions are run, the first set including just one variable (each one in turn). The most significant variable would be selected and then combined with each of the remaining ones (separately). Again, the most significant one is selected. If both remain significant, then the remaining variables are added, each in turn. This process continues until no further significant variables can be found In the regression analysis, we also excluded a variable from the model if one of the former variables was very highly correlated with this new variable. For reasons of multicollinearity this approach is general practice (we used a minimum correlation of 0,95) to exclude these variables. As a result of this calls and minutes were excluded from the final model due to their high correlation with the number of lines. This does not impact the accuracy of the modelling and the R squared (measure of fit) for the final model remains high (see Appendix D). 22

23 3.6.4 The estimated coefficients of those factors which are found to be statistically significant are then used to adjust for the impact of each of these factors on the costs of the operators Based on the results of the regression analysis, adjustments relative to TCNZ are made for the environmental factors in question (see chapter 4). 3.7 Phase 6: Apply Stochastic Frontier Analysis to filter the results for observation errors Statistical theory (outlined Appendix C) is used to establish the efficiency optimum, the so-called efficiency frontier. The method provides the level of overall efficiency for each operator The Stochastic Frontier Analysis provides a 100% efficiency frontier that, generally, is not reached by any operator, i.e. the most efficient operator may, for instance, be 95% efficient. The Stochastic Frontier Analysis framework provides: The efficiency frontier The difference between observations and the efficiency frontier ( residuals ) The split of residuals into observation errors and inefficiencies. Thus, genuine inefficiencies can be determined Observation errors and the inefficiency cause a difference between the theoretic costfrontier and the observed costs. Observation errors are modelled as disturbances described by symmetric distribution functions (e.g. Normal distribution) because observation errors may be positive as well as negative. The inefficiencies are modelled as disturbances that have a skewed distribution function because inefficiencies are always positive In this way, the inefficiency of each operator can be calculated based on the residuals, i.e. the difference between the observed cost per unit output and the efficiency frontier. One of the following methods apply for the calculation: Mean Method: this method calculates the average value of the inefficiency, given the residual. Mode method: this method calculates the most likely value of the inefficiency, given the residual The following diagram illustrates the rationale behind the process of splitting the residuals into observation error and genuine inefficiency. 23

24 In this study, the both the Mean and Mode methods were used. Observation errors may be positive or negative (with equal likelihood). Therefore, they can be captured by a symmetric distribution Inefficiency always causes costs to be higher. Therefore, it can be captured by an asymmetric distribution The data only shows the sum of the two distributions. Its degree of symmetry reveals the relative importance of the observation error. 24

25 4 Results per phase and the efficiency benchmark 4.1 Phase 1: Collect operational and financial data of TCNZ and the LECs Financial and operational data is collected in this phase. An illustrative example of operational data for a LEC number is given in the table below. Local call minutes 145,696,162 LD call minutes 8,786,920 International call minutes 20,849,684 Number of minutes 175,332,766 Number of calls 48,202,380 Eq. Local loop pair km/line 1.60 Table 2 Example of operational data of a LEC An example of the financial data (assets in 000 $ US) is given in Table 3. LAND 38,854 MOTOR VEHICLES 77,499 AIRCRAFT 0 SPECIAL PURPOSE VEHICLES 0 GARAGE WORK EQUIPMENT 5,750 OTHER WORK EQUIPMENT 108,396 BUILDINGS 864,210 FURNITURE 14,082 OFFICE EQUIPMENT 20,805 GENERAL PURPOSE COMPUTERS 112,145 TOTAL LAND & SUPPORT ASSETS 1,241,741 Table 3 Example of financial data of a LEC Interconnection costs are included in the cost data (for both, the LECs and TCNZ excluding mobile interconnection costs in the case of TCNZ). 25

26 4.2 Phase 2: Define a comparable cost level for the produced output The financial data of TCNZ and the LECs differs in many ways that relate to the different price levels of New Zealand and the US. To make the data comparable some straight-forward adjustments are made: For all assets assumed to be internationally traded the exchange rate between the US$ and the New Zealand $ of (2000) was used to establish TCNZs costs in US$ 5. For other assets (land, buildings and capitalised labour) a PPP exchange rate of was used 6. For operating expenses, we also used a weighted average of PPP and exchange rates to reflect the fact that part of the operating expenses, in particular labour and accommodation, do not relate to internationally traded goods. The nominal post-tax average cost of capital used is 11.0% 7. The formula for the CCA/HCA ratio in section 3.3 is used to adjust the book value of TCNZ and the LECs assets to NBV at CCA. The CCA/HCA ratio calculated for TCNZ is as 104.3%, and the average of the LECs is 107.8% In the formula of section 3.3 the telecoms inflations rate is needed. Here we have applied inflation rates to each of the following 5 central LEC cost categories: TOTAL LAND & SUPPORT ASSETS TOTAL CENTRAL OFFICE SWITCHING OPERATOR SYSTEMS TOTAL CENTRAL OFFICE TRANSMISSION TOTAL CABLE AND WIRE FACILITIES Table 4: Cost categories for calculation of asset inflation rates The average asset inflation rate is calculated as the weighted average of the inflation rates by asset class. The latter were obtained from TCNZ LRIC model for the transmission, switching and land and support assets and from that information used in the TSO model for the access network The general inflation rate used in the equation for New Zealand is 2.6% 8. 5 Source: PwC exchange rate database average for the calendar year 6 Source: OECD PPP statistics on a GDP weighted basis 7 Source: This was subsequently revised to 11.2% after a WACC study conducted by PwC New Zealand. This change would have negligible impact on the end result. 26

27 4.2.5 To establish the average depreciation rate of the LECs the asset categories and depreciation rates in Table 6 are used. The depreciation rates are based on TCNZ s asset lives. Asset life Depr rate (%) Land and support assets % Switches % Operator systems % Transmission systems % Cables and duct % Table 5: Asset lives and depreciation rates used As a result of this phase, the financial data of the LECs and TCNZ are comparable. The result is displayed below. Total Costs 12,000,000 10,000,000 8,000,000 6,000,000 4,000,000 2,000, $ (000s) Operator Figure 5: Unscaled costs Here New Zealand is operator number 1 LECs are Source: Economic Intelligence unit inflation rate for year

28 4.2.7 The figure shows that the costs per LEC are significantly different, which is not surprising as the output levels of the LECs are not yet taken into account. The produced output of the LECs is discussed in the next section. 4.3 Phases 3,4 and 5: Identify relevant output function and environmental adjustments and estimate parameters The regressions to determine the output function and relevant environmental parameters are conducted simultaneously using the method described in section 3.6. The regression on the total costs rather than on the asset and expenses separately yielded the most reliable result. Table 6 presents the result of these regressions. Factor Category Hypothesis Results Regression No. of lines Output function Positive relationship No. of calls Output function Positive relationship No. of mines Output function Positive relationship %age local calls Calling pattern Positive/Negative relationship %age aerial cable Geography Negative/Positive relationship Length of sheath per access line Population dispersion Positive relationship Table 6: Parameters tested and regression results 9 Significant and confirms the hypothesis Highly correlated from number of lines so excluded from regression equation Highly correlated from number of lines so excluded from regression equation Not significant Not significant Weighted with the % of aerial and non-aerial cables Significant and conform the hypothesis 9 Other environmental factors as in section 3.5 were tested in similar model runs and found not to be significant. 28

29 4.3.2 The only significant factor (apart from lines) identified in our regression analysis was the length of sheath per access line. This is in line with results obtained for efficiency studies in other countries Based on the coefficients estimated in the regression, adjustments are made for the differences in sheath length per access line. For the regression (and for the subsequent adjustments) we used the length of sheath per line (aerial and non-aerial sheath weighted by the percentage of aerial cable) The method used to adjust for differences in output levels and for differences in environmental variables is straight forward. First, an adjustment given by the formula below is made to reflect differences in the level of production of the US operators to the level of Telecom New Zealand. Unit costs i = TotalCosts i / [(Lines i a )/ (Lines_TCNZ a ], where i stands for LEC operator i, and a is the regression parameter The result is a Unit Cost per operator where a unit of output is defined by TCNZ s total output Then we adjust the unit cost per operator for environmental factors to make the US operators comparable to the environment of Telecom New Zealand. For this adjustment, the following formula applies: Adjusted Unit Costs = Unit costs / [(WeightedScaledSheathlengthPerLine c ) /(WeightedScaledSheathlengthPerLine_TCNZ c )]. 4.4 Phase 6: Apply SFA to calculate efficiency and to filter out observation errors In the previous phases the costs per unit of output were established, and corrections for environmental factors were made. The Stochastic Frontier Analysis is used to calculate the efficiency frontier using the approach outlined in Appendix C and to establish the relative efficiency level of TCNZ and the LEC operators. 29

30 4.4.2 The SFA analysis was applied the unit cost data both before and after adjustment for environmental factors. In our view, the results obtained after environmental adjustment provide a more accurate picture of relative efficiency. For completeness, however, the results obtained before adjustment for environmental factors are shown below. Efficiency of operators before environmental adjustments are made - mean method 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0%

31 Efficiency of operators before environmental adjustments are made - mode method 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% As the diagrams illustrate, TCNZ s relative efficiency, prior to adjustment for environmental factors, is as follows: Mean method: 84.5% efficient relative to frontier and rank 22 Model methods: 87.5% efficient relative to frontier and rank When adjustments for environmental factors are included, TCNZ s relative position improves. The results obtained based from the SFA analysis using the mean regression estimation method are as follows. TCNZ reaches an efficiency level of 93.6%. This compares to an average efficiency of the LECs of 73.6%. The maximum efficiency achieved by any operator in the sample is 95.9%. Ordered by level of efficiency obtained, TCNZ is on rank The main results are summarised in the following table. Efficiency TCNZ 93.6% Average efficiency of LECs 73.6% Position TCNZ 8 31

32 4.4.6 If the more aggressive mode estimation method was used the implied efficiency level for TCNZ is 96.6%. This compares to an average for the LECs of 74.4%. Here the maximum efficiency is calculated at 100% and TCNZ is 8 th in the sample. Efficiency TCNZ 96.6% Average efficiency of LECs 74.4% Position TCNZ Efficiency levels obtained by all operators using the mean and mode methods are illustrated in the following charts. Efficiency of operators after environmental adjustments using mean method 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0%

33 Efficiency of operators after environmental adjustments using mode method 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% Figure 7a and 7b: Efficiency of operators after environmental adjustments have been made These results show that TCNZ is one of the most efficient operators in the sample, taking into account appropriate environmental adjustments (for length of sheath per line). 33

34 5 Conclusions From the efficiency study it is concluded that TCNZ s efficiency after environmental adjustments (for sheath length per line only) have been made is 93.6% if the mean method of estimation is used and 96.6% if the mode method is used. TCNZ s position in a benchmark with other 50 operators is rank 8th in both cases The results from this study are based on: An approach which has been employed by telecommunications regulators in several other jurisdictions; A comparison with US LECs, which are considered relatively efficient; Financial data which have been adjusted to provide comparability; Further adjustments to take account of differences in output mix; A method that accounts for the non-controllable differences between New Zealand and the US; and A method that is able to account for observation errors In the light of these features, we believe that results of the analysis provide a robust and reliable indicator of TCNZ s efficiency When using this efficiency level in practice it is unrealistic to compare TCNZ to the efficiency frontier (ie the theoretical 100% efficiency level) as no operator has been able to achieve this level. In the past regulators have tended to compare the operator under consideration to the best in class, or, more commonly, with the top decile of operators. Comparison to: Theoretical maximum TCNZ implied inefficiency using mean estimation method TCNZ implied inefficiency using mode estimation method 6.4% 3.4% Best in class 2.3% 3.4% Average of top decile Table 8: Implied inefficiency levels 1.5% 3.0% 34