Applied productivity analysis · November 2026
Why electricity networks in Australia and New Zealand got less productive, 2010 to 2024
I measured the total factor productivity (TFP) of 42 electricity distribution networks (13 in Australia, 29 in New Zealand) with proper Lowe indexes, and used data envelopment analysis to split it into technology, operating environment, and three kinds of efficiency. Reliability enters as a bad output, so a network that cuts costs by letting outages rise is not rewarded.
The headline: the best-practice frontier barely moved, so falling productivity is not a technology story. New Zealand networks drifted away from the frontier, and that result holds in every robustness check. Australian productivity fell sharply from 2010 to 2012, recovered almost all of that by 2021, then slipped again. Dense urban networks held their productivity far better than rural ones.
Context
Electricity distribution networks are the poles, wires and substations that carry power from the transmission grid to homes and businesses. They are natural monopolies, so regulators set how much revenue they may recover from customers, and those network charges are a material part of every household power bill.
Because there is no competition to push costs down, regulators lean on benchmarking. In Australia the Australian Energy Regulator (AER) publishes annual benchmarking reports for its 13 distribution network service providers, using multilateral TFP indexes prepared by its consultants, Quantonomics. In New Zealand the Commerce Commission requires all 29 electricity distribution businesses to publish detailed information disclosures, and commissioned a productivity study of them in 2024. In Australia, benchmarking results feed into how the AER assesses whether a network's proposed operating costs are efficient.
Two questions keep coming up in these debates. First, is a network that looks expensive actually inefficient, or does it simply serve a harder environment, such as a sparse rural area? Second, how should reliability count? A network that spends less on maintenance may look more productive until its outages are counted.
What this study adds
- Proper indexes. Regulators use multilateral Törnqvist-type indexes. I use Lowe indexes, which O'Donnell shows satisfy the identity, proportionality and transitivity axioms, so every network can be compared with every other network in every year.
- The environment is modelled, not adjusted after the fact. Customer density enters the frontier, so the decomposition reports how much of each network's productivity comes from its operating environment.
- Reliability as a bad output, consistently in both the index and the frontier model.
- A pooled Australia and New Zealand frontier. The AER already uses New Zealand data to help estimate its opex cost models, but in a scan of regulator reports and the academic literature I found no published study that reports productivity or efficiency for both countries' networks from one pooled frontier. The scan was not a systematic database search.
Method
The approach follows Christopher O'Donnell's work on proper productivity indexes and on accounting for environmental variables and bad outputs in data envelopment analysis (DEA).
- Measure output and input. Each network produces three good outputs (customers served, energy delivered, and ratcheted peak demand, which is the highest peak the network has had to be ready for) and one bad output (customer minutes off supply). It uses four inputs: real operating expenditure, overhead line length, underground cable length, and transformer capacity.
- Aggregate with fixed weights. Outputs and inputs are added up using the same representative prices for every network and year (a Lowe index). Productivity is net output divided by total input. Because the weights never change, a TFP ratio between any two observations means the same thing everywhere in the panel.
- Find the frontier. DEA finds the highest productivity that was achievable in each year, given the technology available up to that year and a network's customer density.
- Decompose. Each network's TFP is split exactly into five parts: technology, environment, technical efficiency, output-mix efficiency, and scale efficiency.
| Component | What it measures |
|---|---|
| Technology | Maximum achievable TFP at the median customer density. It can only rise or stay flat, because the frontier is built from all past observations (a sequential frontier, so no technical regress). |
| Environment | How much a network's own density raises or lowers maximum achievable TFP relative to the median. Dense city networks score above 1, sparse rural networks below 1. |
| Technical efficiency | The network's net output as a share of the most it could produce by expanding all its good outputs in proportion, holding its inputs, outages and density fixed. It is measured in net-output terms rather than as the raw DEA expansion factor. |
| Output-mix efficiency | Productivity lost because of the particular mix of outputs the network produces. |
| Scale efficiency | The remaining gap to maximum productivity, mainly from operating at a size or input mix that is not productivity-maximising. It is calculated as a residual. |
The maths
For network i in year t, with good outputs g, bad output b, inputs x and environment z:
Qit = p′git − pb bit Xit = w′xit TFPit = Qit / Xit
The Lowe TFP index comparing observation it with hs is TFPit / TFPhs. Because p, pb and w are fixed across the sample, the index is transitive and satisfies the identity and proportionality axioms.
Let TFP*t(z) be the maximum TFP attainable with the period-t technology in environment z, and z̄ the median density. The output-oriented decomposition (O'Donnell 2018) is:
TFPit = TFP*t(z̄) × [TFP*t(zit) / TFP*t(z̄)] × OTEit × OMEit × ROSEit
technology × environment × technical efficiency × output-mix efficiency × residual scale efficiency. Technical efficiency is measured in net-output terms: OTEit = Qit / Q̂it, where Q̂it is net output after expanding the good outputs radially to the frontier. The identity holds exactly for every observation: the product of the five components reproduces each network's TFP.
DEA technology (O'Donnell 2025): sequential (all networks, all years up to t), variable returns to scale, inactivity allowed, good outputs and inputs strongly disposable, the bad output tied to the level of activity (an equality constraint, so outages cannot be disposed of for free), and the environment strongly disposable, so the weighted average density of a network's comparison peers can be no more favourable than its own.
Weights. Representative prices are set so that the value shares at the sample mean equal the AER's 2025 MTFP cost shares:
| Good outputs | AER weight | Share of good-output value |
|---|---|---|
| Customers | 17.5 | 20.6% |
| Energy delivered (GWh) | 12.4 | 14.6% |
| Ratcheted coincident maximum demand (MW) | 55.1 | 64.8% |
The AER's index also counts circuit length as an output (weight 30.1). Here line length is an input, so that weight is dropped and the three remaining output weights keep the AER's ratios. The bad output (customer minutes off supply) is priced so that, at the sample mean, it is worth 15.2/115.1 (13.2%) of the value of good outputs. That is close to the AER's own 2006 to 2024 outage weight (14.5/114.5); using it instead moves the New Zealand result by 0.2 percentage points. Inputs: real opex 42.3, overhead lines 18.5, underground cables 12.7, transformers 26.4.
Data
The panel covers 627 network-years from 2010 to 2024: 13 Australian networks (195 observations) and 29 New Zealand networks (432 observations). All of it is public regulatory data.
| Country | Source |
|---|---|
| Australia | AER, Annual benchmarking reports 2025: Quantonomics' consolidated DNSP benchmarking data (the MTFP input file and the economic benchmarking RIN data), published November 2025. |
| New Zealand | Commerce Commission, information disclosed by electricity distributors, plus Quantonomics' NZ workbook (for the NZ opex price index) from the AER release. |
| Currency | World Bank GDP purchasing power parities (indicator PA.NUS.PPP). NZ real opex is converted to 2006 Australian dollars at the 2006 PPP ratio. |
Making the two countries comparable
- Peak demand uses coincident maximum demand, which both countries define the same way. (The AER's own index uses non-coincident demand, which New Zealand does not report.)
- Opex is deflated with each country's own opex price index, then converted at purchasing power parity. GDP PPP is a general-purpose rate, because no utility-specific PPP is published for the two countries.
- Reliability. Australian unplanned SAIDI already excludes major events. New Zealand's unplanned SAIDI includes them: Cyclone Gabrielle took Unison's 2023 figure to 1,821 minutes. I cap each New Zealand network-year at twice that network's 2010 to 2024 median, which caps 54 of 432 New Zealand observations. The uncapped run is in the dashboard; in it, 9 network-years have negative net output and drop out of the averages.
- Transformer capacity is distribution plus zone substation capacity. For Australia this includes the first stage of two-stage transformation, which the AER's own index leaves out.
- Years. Australian years end in June and New Zealand years end in March.
Exclusions
- Orion's 2010 to 2012 observations are dropped because reliability data are missing. (Orion's 2011 disclosure was waived after the Canterbury earthquakes.)
- Three Top Energy years (2015, 2022 and 2023) have negative net output: at the fixed prices, the value of their outage minutes exceeds the value of their good outputs. They stay in the frontier but are left out of averages.
- Changes from 2010 to 2024 are geometric means over networks observed in both years (13 Australian and 28 New Zealand networks).
Results
1. New Zealand's productivity fell; Australia's fell, recovered, then slipped again
New Zealand TFP fell 17.1% between 2010 and 2024, with large swings from year to year. Comparing three-year averages (2010 to 2012 against 2022 to 2024) the change is −18.0%.
Australian TFP fell sharply from 2010 to 2012, recovered almost all of that by 2021, then slipped again, so the net fall over the period came after 2021. Its end-to-end change of −5.8% depends on the starting year: on three-year averages it is +0.3%, because 2010 was a high year.
2. The frontier stood still
Best-practice productivity at the median density rose 0.5% over the whole period, all of it between 2010 and 2011. The decline is therefore about networks falling behind a frontier that stood still, not about the frontier moving.
3. New Zealand networks fell behind the frontier
In New Zealand, technical efficiency fell 21.0%: networks moved further from best practice for their density. It is the largest negative component in every robustness scenario and every random set of weights. Rising density (+3.7%) and scale efficiency (+5.4%) offset part of the fall.
For Australia, how the change splits between components depends on the model. Under variable returns to scale (the main model), technical efficiency is flat (−0.9%) and the fall shows up in output-mix efficiency (−4.0%) and scale efficiency (−4.7%). Under constant returns to scale, technical efficiency falls 18.8% and scale efficiency rises 18.2%. The large Australian networks have few comparable peers, so I don't read much into the Australian split.
4. Reliability matters for New Zealand, not Australia
Leaving reliability out, New Zealand's TFP decline shrinks from 17% to 13%, while Australia's barely changes (−5.8% with, −5.8% without). Without the cap on major-event outages, New Zealand's decline would be 23%, so the size of the reliability effect depends on how major events are treated.
5. The level gap: partly density, mostly distance from the frontier
Averaged over 2020 to 2024, New Zealand TFP is 66% of Australian TFP (60% to 71% across the model variants). Lower customer density accounts for part of the gap (environment component 0.89 in New Zealand against 1.02 in Australia). The rest is distance from the frontier: TFP efficiency, the product of the three efficiency components, is 0.43 in New Zealand against 0.57 in Australia. Whether that distance shows up as technical or scale efficiency depends on the returns-to-scale assumption.
6. Density divides the winners from the losers
Urban networks (the densest third) lost only 3% of their TFP. Rural and mixed networks lost 17% and 20%, mostly through technical efficiency in the main model. City networks such as CitiPower, United Energy, Wellington Electricity and Jemena lead the 2020 to 2024 ranking.
Explore the results
Choose how reliability counts, then compare groups or pick a single network. The explanations next to each chart are written from the same verified numbers and update as you change the view. Hover or tap a chart for exact values; every view also has a table.
Change from 2010 to 2024, all components
Zoom in on one network
This network's data, year by year
Ranking as a table
Robustness and limits
I re-estimated the model under 20 named variants, including the main model (returns to scale, index weights, the value placed on outages, the cap on New Zealand major-event outages, and the first year of the frontier) and under 200 randomly drawn sets of index weights. The table shows the change from 2010 to 2024.
Which findings hold
- Holds No technical progress: the best-practice frontier barely moved. Frontier change is +0.3% to +1.9% in every named scenario. In 80% of random weight draws it is under 2%, and the largest is +18%.
- Weak Australian TFP fell about 6% from 2010 to 2024. The change is −6% in the base and barely moves with returns to scale, outage caps or sample start, but it depends on the base year and weights. Comparing 2010-12 with 2022-24 gives +0.3%, because 2010 was a high year. Across weight draws the change ranges from −20% to +5% (5th to 95th percentile), and it is negative in 86% of draws.
- Does not hold Australia's decline came from mix and scale efficiency, with technical efficiency flat. This split depends on variable returns to scale. Under constant returns, Australian technical efficiency changes −19% and scale efficiency +18%. Under VRS the large Australian networks have few peers and sit on the frontier by construction.
- Holds NZ TFP fell about 17%, driven by technical efficiency. NZ TFP changes −13% to −23% across named scenarios and −18% on 3-year averages. Technical efficiency (−15% to −26% in named scenarios) is the largest negative component in 100% of named scenarios and 100% of weight draws. NZ falls more than Australia in 90% of weight draws.
- Holds Reliability matters for NZ, not Australia. Australian TFP change stays between −5.9% and −5.8% whatever is done to reliability. NZ: −13% without reliability, −16% with a 1.5x cap, −17% in the base, −20% with outages valued twice as high, and −23% uncapped. The size of the NZ effect depends on major-event treatment, not its direction.
- Holds NZ TFP is about two-thirds of Australia's in 2020-24, and lower density explains about 12 points. The NZ/AU level ratio is 0.60 to 0.71 in named scenarios (0.49 to 0.78 across weight draws). The environment ratio stays at 0.86 to 0.90 in named scenarios.
- Does not hold Technical efficiency explains most of the rest of the level gap. Under constant returns the technical-efficiency ratio is 0.94 (vs 0.77 under VRS) and the remaining gap moves to scale efficiency (0.82). Australian networks lose their near-perfect scores under CRS (2020-24 mean 0.77), which confirms that VRS overstates their efficiency.
- Holds Urban networks lost far less TFP than rural and mixed networks. Urban networks change −1% to −7% in named scenarios, against −13% to −32% for rural and mixed ones. Urban falls least in 91% of weight draws.
| Scenario | AU TFP | NZ TFP | AU technical eff. | NZ technical eff. | NZ/AU TFP, 2020 to 2024 |
|---|---|---|---|---|---|
| Base model: variable returns to scale, AER weights, NZ outages capped at 2x network median | −5.8% | −17.1% | −0.9% | −21.0% | 66% |
| Base model without reliability (no bad output) | −5.8% | −13.1% | +0.1% | −15.2% | 71% |
| Constant returns to scale | −5.8% | −17.1% | −18.8% | −22.4% | 66% |
| Non-increasing returns to scale | −5.8% | −17.1% | −0.9% | −20.6% | 66% |
| Constant returns to scale, without reliability | −5.8% | −13.1% | −9.9% | −16.5% | 71% |
| Equal weights on customers, energy and peak demand | −5.8% | −19.0% | −0.9% | −20.8% | 68% |
| Output weights favour customers (50/20/30) | −2.0% | −18.4% | −0.9% | −20.5% | 65% |
| Output weights favour energy (20/50/30) | −9.4% | −19.9% | −0.9% | −21.0% | 69% |
| Equal weights on the four inputs | −8.7% | −14.5% | −0.9% | −21.0% | 62% |
| Opex weighted 60% of inputs | −3.9% | −20.2% | −0.9% | −21.0% | 66% |
| Opex weighted 25% of inputs | −7.9% | −14.0% | −0.9% | −21.0% | 67% |
| Outages valued at half the AER rate | −5.8% | −14.8% | −0.8% | −18.1% | 68% |
| Outages valued at twice the AER rate | −5.8% | −19.6% | −1.1% | −19.9% | 60% |
| NZ outages capped at 1.5x network median | −5.9% | −15.6% | −0.8% | −20.7% | 66% |
| NZ outages capped at 3x network median | −5.8% | −17.8% | −1.0% | −22.0% | 66% |
| NZ outages capped at 4x network median | −5.8% | −18.8% | −1.0% | −22.6% | 65% |
| NZ outages capped at each network's 90th percentile year | −5.8% | −20.1% | −1.0% | −22.4% | 66% |
| NZ outages uncapped (major events included) | −5.8% | −23.3% | −1.0% | −26.3% | 64% |
| Frontier built from 2006 for Australia (NZ still from 2010) | −6.0% | −17.2% | −1.4% | −21.6% | 66% |
| Frontier built from 2008 for both countries | −5.9% | −16.9% | −1.3% | −18.8% | 66% |
| Random weights, 5th percentile | −20.4% | −28.7% | −1.0% | −30.6% | 49% |
| Random weights, median | −8.5% | −18.7% | −0.9% | −20.5% | 60% |
| Random weights, 95th percentile | +5.3% | −5.2% | −0.8% | −18.3% | 78% |
The random-weight rows rank each column separately, so a row is not a single weight set.
Limits to keep in mind
- Australian technical efficiency is close to 1 for every network. Under variable returns to scale, the 13 large Australian networks have few comparable peers, so they largely define their own part of the frontier. Under constant returns to scale they lose those near-perfect scores, which is why the Australian component split is treated as model-dependent above.
- Scale efficiency is a residual, so for New Zealand it also absorbs year-to-year reliability shocks.
- The weights matter. Peak demand carries 64.8% of good-output value. The random-weight checks above show how much the results move when that changes.
- One environmental variable. Terrain, vegetation and bushfire exposure are not yet in the model. Adding them would move part of the technical-efficiency gap into the environment component.
- Opex coverage differs. Australian opex includes capitalised corporate overheads; New Zealand opex follows each network's own capitalisation policy. This affects levels more than changes.
- Net output can go negative. With fixed prices, the value of outages exceeds the value of good outputs in 3 Top Energy years, so those years are left out of averages. O'Donnell's sustainable productivity index cannot go negative; using it gives Australia −7.6% and New Zealand −18.8%, the same story.
- Currency conversion. GDP PPP is a general price comparison, not a network-cost one.
References
- O'Donnell, C.J. (2016). Using information about technologies, markets and firm behaviour to decompose a proper productivity index. Journal of Econometrics 190(2), 328–340.
- O'Donnell, C.J. (2018). Productivity and Efficiency Analysis: An Economic Approach to Measuring and Explaining Managerial Performance. Springer.
- O'Donnell, C.J. (2022). How to build sustainable productivity indexes. CEPA Working Paper WP10/2022, University of Queensland.
- O'Donnell, C.J. (2025). Accounting for environmental variables and bad outputs in DEA. CEPA Working Paper WP03/2025, University of Queensland.
- O'Donnell, C.J. (2026). Proper quantity indexes for coherent productivity comparisons. CEPA Working Paper WP01/2026, University of Queensland.
- Australian Energy Regulator (2025). Annual benchmarking report, electricity distribution network service providers, and supporting Quantonomics data.
About
Stefan Smith
Bachelor of Economics, majoring in Quantitative Analysis and in International & Financial Economics.