Water markets’ promise: the Murray–Darling Basin

The very advanced and well-designed water markets in the Murray–Darling Basin are the result of long and complex reforms. Using state-level export data for agricultural and manufacturing sectors, we study the impact of water markets on the allocation of water through the first (1994–2006) and second reform periods (2007–2015), relative to when the markets’ foundations were laid (1988–1993). We find water markets trigger a shift away from the water-intensive (less water-productive) sectors that is most pronounced during droughts in the first reform period. However, improvements in technology and techniques that reduce water intensity (increase water productivity) partially offset such a shift. We also document an inter-sectoral shift of activity between agriculture and manufacturing, as well as address some recent criticisms of water markets’ effectiveness.

The human appropriation of freshwater has reached the limits of the natural availability in many places. Recent crises in India, Australia, and California underscore the time of water abundance is over (Vörösmarty et al 2000, Zetland 2011, Jaeger et al 2017. About 3.1 billion people are projected to suffer from water scarcity by 2050 (Gosling and Arnell 2016). Excessive depletion of rivers, lakes, wetlands, and aquifers is undermining the hydrologic conditions to sustain ecosystems (Richter et al 2012, Brauman et al 2016. We need better ways to reconcile ecosystems' water needs with those of society. Inevitably, to sustain economic growth while limiting overall water use, requires increased economic productivity of water (Debaere and Kuerzendoerfer 2017, Debaere 2014, Gleick 2018, Marston et al 2018. To notch up water productivity and realize waterneutral growth, societies need to improve water management with new policy tools. Water markets are an increasingly popular tool in water-scarce Australia, Chile, the western United States, and China 4 . We study the impact of the world's most advanced water markets in Australia's Murray-Darling Basin (MDB) on water allocation across sectors, and hone in on changing techniques/technology induced by water markets, which is largely absent from the empirical literature.
Water has both private and public dimensions. It is part of an ecosystem and contributing beyond private consumption (for economic perspectives on water, see Hanemann 2005, Garrick et al 2020. Water markets are hence especially defined by policy and political contexts (Teyetlboym 2019, Debaere 2020, Garrick et al 2020 and require constant design adaptations and improvements: more homogenous rights, less transactions costs, and limited externalities are critical for their proper functioning. Different types of water rights are traded as leases, sales, options, etc, and transactions are governed by (changing) regulations, with varying effectiveness. History illustrates how political and hydrological contexts and our evolving understanding codetermine markets' establishment and design. Operationalizing MDB's markets was a major policy intervention and a painstaking, unpredictable political process involving states and the central government, with compromises and setbacks.
Water markets are cap-and-trade systems for renewable natural capital through which water rights are bought and sold independent of land titles, see Teyetlboym (2019). With an explicit cap on overall water withdrawal, they have the potential to protect rivers' environmental flows 5 . Water markets promise resilience, and flexible and decentralized responses to climate and drought shocks. Water is often priced too low, which induces overuse and discourages innovation and (private) investment in water-saving technology 6 . By putting a price on water at the upstream extraction stage, water markets ensure more realistic prices downstream.
Most importantly, water markets can improve water's allocation and overall productivity. We investigate this critical hypothesis. Market exchanges are expected to shift production to higher-value (less water-intensive) uses, since more productive users can pay a higher price 7 . Moreover, water prices incentivize water saving and more efficient irrigation systems. While droughts or command and control measures encourage more productive use of water as well, markets should amplify this productivity benefit since trades additionally make reallocating water across agents possible. To be explicit, since water markets typically operate in water scarce areas, the increase in more productive uses of water they bring about, does not necessarily imply a net reduction in water use, unless the (binding) cap on water use is explicitly reduced.
The MDB provides a unique setting to study the productivity impact of water markets: • Water markets are complex. Participation in their operation is impeded by uncertainty, information asymmetry (Chokri and Khana 2005), noncompetitive behavior (Hantke-Domas 2017), local protectionism (Hagerty 2019), transaction costs (Regnacq et al 2016), and credit constraints (Donna and Espin-Sanchez 2021). Even the harder to measure hydrological, regulatory, and socioeconomic or political context may explain 5 Water markets provide a mechanism to buy water to be left in the river, and increase river flow irrespective of the cap. They sometimes allow for a supervising authority to set the amount of water available for use within the cap in a given year or season, depending on the prevailing climate conditions. 6 For surveys of water-pricing policies, see Schoengold and Zilberman (2007), Dinar et al (2015), Quentin et al (2020). Olmstead (2010) emphasizes price adjustments. 7 For water misallocation under water rights regimes, see Libecap (2011). Debaere and Tianshu (2020)'s theoretical setup and assumptions show how water markets can shift water toward, on average, less water-intensive activities.
why markets do not emerge everywhere (Garrick et al 2013, McCann andGarrick 2014). Markets' particular design features are critical (Teyetlboym 2019). Since the MDB markets are among the world's best-managed with well-defined property rights, low transaction costs, ever-improving trading platforms and governance, they are an ideal testing ground to study whether shifting away from more water-intensive activities is in fact possible. • Australia has faced droughts with drastically reduced water availability and skyrocketing water prices. Droughts should make agents with multiple operations apply scarce water to the higher-value (less water intensive) portion of their activities. Studying the MDB before and after its water market reforms lets us separately identify the impact of markets that additionally promote agents to trade. • Australia's MDB markets are some of the largest.
Annually, 2 billion Australian dollars are traded, with up to 60% of irrigators participating and redistributing up to a quarter of water for human use (Quentin et al 2011, Department of Agriculture andWater Resources 2016). • Finally, MDB water markets extend beyond agriculture as cities and utilities participate (Wheeler et al 2014a, 2014b, Horne and Quentin Grafton 2019. The MDB can reveal whether water markets indeed facilitate inter-sectoral shifts away from water-intensive agriculture to industry generating far more dollars per unit of water. Inter-sectoral water allocations are critical to limit water use in growing economies (Debaere and Kuerzendoerfer 2017).
Rather than use production data, we investigate with export data whether the 18 agricultural and manufacturing sectors in Australia's eight states and territories show a shift away from the more waterintensive activities since 1988. We prefer export data since they are more granular for state manufacturing subsectors, available for longer periods, and easily combined with our water-intensities. Needless to say, we control for the determinants of exports.
Our analysis concentrates on four core MDB states and territories with active markets. We track their changing exports in terms of water intensity as markets are rolled out relative to the initial benchmark period, 1988-1993. Comparing MDB states with non-MDB states confirms our key findings, see appendix A. 8 We are mindful of the potential endogeneity of sectors' water-intensity as higher 8 While non-MDB states are not randomized controls for MDB states, it is important to emphasize that when including non-MDB states, our identification hinges on the reasonable assumption of a common trend before water markets are established across MDB and non-MDB states for sectors with similar water intensities, for which we provide support see appendix A. water prices will induce water saving technology, an issue we address with instrumental variables (IVs) below. It will allow us to gauge how market-induced changes in water-use techniques/technologies alter water's allocation. For more details, see appendix B.
To do justice to the complex design, establishment, and operation of water markets and their broader impact, we choose a long period  and study their impact on economic activity (exports), rather than study water market transactions themselves or narrower policy interventions in those markets. We distinguish three phases with key legislation, while using the earliest one as benchmark. Following the National Water Commission's (NWC) timeline, the years up to 1993 lay the institutional groundwork; the first reform period (1994)(1995)(1996)(1997)(1998)(1999)(2000)(2001)(2002)(2003)(2004)(2005)(2006) establishes the basic market infrastructure as markets emerge and expand; finally, the second reform period (since 2007) pushes sustainability. We want to capture the overall impact of the major policy interventions that establish water markets. This will include conservation efforts, water pricing policies by utilities or other downstream adjustments that emerge with water markets, and even governmental water rights purchases to increase river flow. Figure 1 shows how different the reform periods are in terms of transaction volumes of temporary and permanent rights in the southern MDB markets.
Our results reflect a nuanced assessment of the MDB water markets' impact.
• Water markets induce a shift toward less waterintensive (more productive) activities that is pronounced during droughts in the first reform. For every decrease in water use of 100 l per Australian dollar (increase in productivity), there is with water markets an increase of a sector's export share of 0.11 percentage point in droughts. On an annual basis, and relative to the pre-reform dry baseline, markets reduce water use per dollar by 8.25%. • As we instrument the water intensities with US data, we document how lower water intensity (higher productivity) induced by water markets reduces the need for sectoral reallocations by three quarters. • A two-sector specification for aggregate agricultural share confirms an inter-sectoral reallocation from agriculture to higher-value manufacturing, irrespective of droughts. • Estimates within agriculture in the MDB show water markets increase the share of less waterintensive crops exported during droughts. However, in wetter years more water-intensive exports tend to increase. This finding dovetails with ongoing concerns about the efficacy of water-market reforms, as voiced by Young and McColl (2005), Young (2014), and Horne and Quentin Grafton (2019). Markets that price water change people's behavior and increase water use in spite of a cap. Water may be intercepted, dams built, and previously unused water rights (the so-called 'sleeper rights') are being used.
There is a rich institutional and governance literature on water markets in Australia and beyond, that sometimes compares markets across countries (Grafton et al 2011, Grafton et al 2012. 9 Our contribution is to econometrically investigate the impact of water markets on the structure and water intensity of exports, as well as the mitigation through the changing techniques/technology they induce. Our data track on the ground what changes after watermarket reforms. Our endeavor takes as a starting point, Debaere and Li (2020), who studied the smaller Rio Grande. We confirm water markets' shift toward more productive activities in the more sophisticated and larger MDB setting, and explore the impact of drought and endogenous technology responses. In addition, we differentiate between manufacturing and agriculture.
Our approach complements water market transaction analyses, as well as computable general equilibrium models or simulations. Chang and Griffin (1992) use transactions data, and Hearne and Easter (1997) rely on survey data of past transactions. Donna and Espin-Sanchez (2021) use data from historical water auction transactions in Spain. Wheeler (2014) summarizes key computable general equilibrium and simulation studies for Australia (see Qureshi et al 2009, NWC 2011. Rafey (2020) estimates the transaction surplus associated with water markets within agriculture with transactions and production data in southern MDB since 2007. Different from Rafey (2020), we compare the effect of water markets to early benchmark years when there was no established market. In addition, we explore technological change. Our econometrics underscore how markets shift away from the most water-intensive crops, extend into non-agricultural sectors and induce efficiency gains (technological progress) at the sectoral level.
Our findings support descriptive evidence of shifting agricultural crops and trend analysis of water use and agricultural output (NWC 2011, Kirby et al 2014 there is also an emerging experimental literature (Dinar et al 1998, Cummings et al 2004, Garrick et al 2020). We focus especially on water markets' economic reallocation and efficiency gains. However, some of our findings are consistent with environmentalist criticisms of water markets that point to increased water use, in spite of a cap, which should be detrimental to environmental flows.
In the following sections, we first describe the MDB markets, their history, and the three phases. Next, we lay out the data, before discussing the econometric specifications. We conclude after discussing the results.

MDB water-market reforms
Australia's water markets story is primarily that of the MDB, where 90% of trading takes place (NWC 2011). 10 The history of this water-scarce continent is intimately linked to water management. With a climate prone to droughts, Australians realized early on that storage and irrigation were key for regional development. Governments focused primarily on the supply side, investing in dams, irrigation systems, weirs, etc. MDB's water markets moved away from engineering and water supply interventions. According to the NWC, they were meant to facilitate 'the economically efficient allocation of water while ensuring environmental sustainability' (NWC 2011, p 8). The MDB fits key preconditions for a successful market: • MDB's water is fully developed, which should support reallocating valuable water to higher-value uses. • Water availability varies seasonally. Water users have varying water demands and varying flexibility to respond to water shortage, making reallocating water desirable. • Increasing demand for urban and environmental water keep up the pressure to reform and facilitate water reallocations (NWC 2011, p 108).
The MDB covers 14% of Australia's landmass (see figure 2), and spans over 1000 km of the east coast. Irrigation water is directly extracted along the River Darling, whereas irrigation systems dependent on the Murray are fed by water from behind storage dams (Horne andQuentin Grafton 2019, Hanemann andYoung 2020). The basin covers New South Wales, Victoria, South Australia, Australian Capital Territory, and a small fraction of Queensland.
Until the 1970s, states issued water rights (nonpriority licenses) virtually on demand; unlike the United States, Australia abandoned riparian and prior appropriation (by seniority) allocations in favor of a non-priority permitting system (Tisdell 2014, Hanemann andYoung 2020). Initially water rights were attached to land titles and based on the area of irrigable land. With limited metering and enforcement, irrigators used water generously. By the 1980s, surface and groundwater had been overdeveloped, and maintaining and building dam and irrigation projects became prohibitively expensive. Toxic algal blooms and increasing salinity testified to the deteriorating water quality, as did the loss of native aquatic plants and animal species. The tragedy of the commons was 10 The MDB Authority website www.mdba.gov.au/watermanagement/managing-water/water-markets-trade. The Dept of Agriculture, Water and Environment give current information via www.agriculture.gov.au/water/markets. Overviews are found in Herberger (2011)  on full display. The MDB started the difficult process of introducing water markets. Water withdrawal had to be limited, water rights redefined, and a trading infrastructure with rules, enforcement, and more uniform governance put in place. The MDB water markets result from a long process and major regional and central initiatives to design an integrated and sustainable water management system. The NWC distinguishes three critical market design phases. The time line and time span illustrate markets are dynamic, and at the confluence of technical, legal, governance, and political issues that are hard to disentangle. Hence, we assess markets' impact on the economy (exports) by comparing the three phases: • Benchmark: emerging water markets (till FY 1993).
Tentative steps made water markets possible without being part of a grand plan to develop the markets we now have. A moratorium on new water licenses was put in place, and existing permits were redefined in terms of water extraction instead of area, which effectively capped water use, albeit at prevailing levels. There was limited intra-water district trading that allowed access to water for new users without new licenses. Water and land rights were split, lowering transaction costs (NWC 2011, pp 33-42, Horne and Quentin Grafton 2019, pp 177-8). • First reform period: expanding water markets (FY 1994(FY -2006. The 1994 Council of Australian Governments (COAG) Water Reform Framework created a comprehensive system of well-defined water rights, called water entitlements (permanent rights) and water allocations (temporary rights depending on annual assignment of available water to entitlements) 11  The MDB markets are a major accomplishment. Looking back, it would indeed have been better if all design features had been clear from the outset (Young 2014). Unresolved issues remain affecting the ecosystem and socio-political context (Horne and Quentin Grafton 2019). They must be addressed to avoid compromising markets' proper functioning: water rights for indigenous people, sleeper rights (dormant water rights), occasional lack of enforcement, and incentives to intercept water. We focus on markets' economic impact. While we expect reforms push economic activity toward less water-intensive uses, some changes (i.e. improved irrigation efficiency and sleeper rights) may pull in the opposite direction.

The data
We focus on Australia's states and territories. We mainly consider the four MDB states and territories with developed water markets: New South Wales, South Australia, Australian Capital Territory, and Victoria. Here, water markets cover over 25% of all land use (see figure 2). Note that we group Queensland into the non-MDB group, since its water markets occupy only 3% of land use with economic activities 13 . Table 1 displays huge variation in water intensity. The data come from the Australian Bureau of Statistics (ABS) and the Food and Agriculture Organization (FAO). Water intensity is the ratio of a sector's liters of direct water use over its production value in Australian dollars. (Water intensity is not water per ton, the physical metric common in the sciences.) The inverse of our measure captures water productivity. The intensities are constructed with 12 The upstream MDB states were reluctant to reduce water use and pushed back against too much water for the environment, reducing water targets, see Horne and Quentin Grafton (2019). 13 Land use data come from Land Use of Australia, Version 3 1992/93 (the last pre-reform FY) by Australia's Bureau of Rural Sciences. The water market coverage is digitized from figure 2.1 in the NWC's Australian Water Market Report 2008-09. national (MDB and non-MDB states) production and water use data, and expressed in 1993 real values. Appendix B provides more details. The bold-faced entries give aggregate intensity measures for MDB states, non-MDB states, and overall manufacturing or agriculture. Agriculture's water intensity is an order of magnitude larger than in manufacturing. Table 1 also displays water intensities for disaggregate exports at the national level (state-specific intensities are not available). There is significant heterogeneity within agriculture. By contrast, intramanufacturing variation is minor. There is a steady decrease in water intensity (increase in water productivity). Water intensity is cut in half for manufacturing, yet reduced 4.5 times for agriculture.
Even though our water intensity measures come from national data, there is legitimate concern that because of the sheer size of MDB they may endogenously respond to water markets' higher water prices that trigger water saving efforts. We therefore rely on the comparable US water intensity data as IVs for the Australian measures. While the US data are clearly correlated with Australia's, they will not be affected by the MDB water market reforms. Doing so allows us to assess Australia's path relative to that of the US (another major agriculture exporter). See appendix B for details on US water intensities and IV procedure.
Droughts are critical to our analysis. We construct drought measures with state-level annual precipitation data. We draw on the 10 000+ active stations from the Bureau of Meteorology between 1901 and 2016. We selected 5656 stations that overlap with a region's economic land use in 1992/93 (the last prereform year with available land use data) and constructed a balanced panel to capture the effect that corresponds to the pre-reform distribution of economic activities. For those stations, we calculate the pooled state average of annual precipitation.
To identify drought in a state over time, we use a 0/1 indicator. With different climates across states and relatively high within-state standard deviation in Australia's precipitation, we assess droughts relative to a state's historical precipitation. A dry spell is a period of low rainfall when last 2 years falls within the lower 25th percentile of state precipitation throughout the entire 20th century 14 . Figure 3 visualizes droughts across all states. To separately identify water markets' 14 In terms of demeaned and standardized precipitation (P − P * )/σ, our dry conditions require a negative value that is smaller than −0.675 pooling two subsequent years. To be clear, our drought cutoff (with annual data) is both looser and tighter than the value of −1 that is sometimes used. On the one hand, we require dry conditions for two subsequent years. Since we are dealing with the Big Dry in the MDB, we need some persistence in dry conditions of the control periods and control states. On the other hand, also because of Australia's relatively high standard deviation of within-state annual precipitation (25% above within US states), we allow for a slightly less stringent cutoff (−0.675). With this cutoff we have dry spells before and after water markets are introduced, both in and outside the MDB.  impact on exports from that of droughts, we must obtain dry and wet spells in all NWC phases of the evolving MDB markets. The sectoral state export data stem from ABS and are classified by three-digit Australian and New Zealand Standard Industrial Classification (ANZ-SIC). We focus on agriculture and manufacturing that are most affected by water markets. (Mining typically has its own water supply.) For each state, we consider sectors' export share in total (manufacturing + agricultural) exports. Australia is about twice as open as the US, exporting over 20% of its GDP in recent years. About 70% of agriculture (using the majority of water) is exported (Australian Government 2021).
To capture global factors that affect Australia's exports, we include the total world-level sectoral imports (excluding Australia) from country-level UN Comtrade imports from 1988 to 2015 (https:// comtrade.un.org/Data/). We manually match the commodity codes based on the ANZSIC in our sample with Comtrade's Harmonized System. We, however, have no good measure to control for domestic demand should it deviate from international trends for certain goods. From 1995 onwards we rely on 'use' (exact name in statistics) in more aggregated industries from ABS (www.abs.gov.au/ statistics), which captures the domestic market. We linearly inter/extrapolate missing years.
Some of our regressions include state-level production factors that are common in regressions explaining states' international trade. These factors capture states' endowments. Annual state-level employment and capital stock data come from ABS.
Finally, we also use two state-level measures of heating degree days (HDDs). We use the 5656 + stations mentioned above, again calculating pooled state averages. One measure is the cumulative sum of days with average daily temperature above 8 • C, while the other of days above 32 • C; Deschênes and Greenstone (2007) define degree-days in this way based on the standard agronomic approach. We include both current and lagged precipitation and HDDs because harvested crops in a year could have been growing over the previous year. Appendix table 1A reports summary statistics.

The specification
Our basic specification explains Australia's sectoral export shares, which vary across states and territories and over time. We focus on the four MDB states and territories (and extend to comparison with other states in appendix A). Equation (1) determines whether the two reform periods changed the export structure in the MDB relative to the initial benchmark. We wonder whether the share of more water-intensive exports decreases relative to that of less water-intensive ones, while controlling for year effects, state × industry effects, and many additional variables. We cluster standard errors at the state-level: ExportShare sit captures the share of state s's sector i in the state's overall manufacturing and agricultural trade for year t. The critical coefficients are β 1 and β 2 .
A negative sign indicates reforms decrease the share of more water-intensive goods exports in the MDB. Since there are zero-value observations for exports, we avoid logs. We insert a whole series of controls under X nsit , including state precipitation, HDDs, the logarithm of state labor force and capital stock. We also include the shares of excluded sectors 15 . The fixed effects capture many factors that are difficult to specify but stay constant. Year fixed effect absorb common events across MDB states, whereas industrystate effects take into account state-specific industry components. With the year-fixed effects, we cannot assess the impact of changing the fraction of a state's water that is allocated within the cap due to changing water availability conditions. Since industry water intensities vary at the national (not state) level, there is not enough variation to include industry-year-fixed effects that would capture sectoral movements, especially in world markets. We therefore include sectoral global exports that change over time. Since MDB states are relatively similar (and water intensities do not vary by state), our preferred specification has no state-year-fixed effects and exploits the temporal variation of water-intensive versus less waterintensive sectors across the MDB 16 . Since the overall size of exports varies significantly across states, we weight our estimates by state-level average total annual export values pooling all pre-reform years.
We investigate whether water markets respond flexibly to shocks. Therefore, we run additional regressions with interactions with our drought indicator: β 4 Reform_I t × WaterIntensity it × drought st and β 5 Reform_II t × WaterIntensity it × drought st . We include the drought indicator and other corresponding double interaction terms of drought separately. We also consider similar specifications for the agricultural sector only.
Given the complexity and unpredictable nature of establishing water markets, the endogeneity of the reform periods is not a concern. Establishing water markets has been a very slow, time-consuming, political and unpredictable process. There is, however, some concern the national water intensity measures may respond to water markets, since the MDB is an important water user. Indeed, higher water prices should incentivize decreasing the water intensities. We instrument Australian water intensities with their US counterparts; see below.

Discussion
The columns of table 2 capture the coefficient estimates of our regressions. The first four rows of estimated coefficients let us compare the water  1988-93, 1994-2006, 2007-15). b 'Drought' is 1 if the total precipitation in is part of the lower 25th percentile of the state's precipitation over the entire 20th century, and 0 otherwise. c 'Industry-year control variables' includes water intensity and the logarithmic value of world total import (except Australia). d 'State-year control variables' include export share of mining, logarithmic values of employment and capital stock, the drought dummy, and weather controls for the current and previous years including precipitation, HDDs above 8 • C and above 32 • C, respectively. e 'Other interaction terms' include 'Period (94-06) × Drought' and 'Period (07-15) × Drought' for Columns (2) and (4). f Weighted by the average total annual export by states in pre-reform years. Standard errors clustered by states in parentheses, * * * p < 0.01, * * p < 0.05, * p < 0.1. market impact of the first and second reforms relative to the benchmark period across the various regressions. (The other rows refer to other controls and fixed effects that are (not) included in the specifications). The first two columns show the IVs results with US intensities instrumenting for Australia's. Columns (3) and (4) give the ordinary least squares (OLSs) estimates, which allow for endogenous responses in water-saving technology\techniques following market reforms. A few observations stand out. Consider the IV estimates. We obtain negative and statistically significant coefficients for both reform periods in the specification without drought. With a water market in place, there is a 0.15-0.16 percentage point decrease in sectors' export share for every 100 l increase per dollar. When we bring in the interaction with drought, a negative, statistically significant coefficient is estimated only for the dry years in both periods, leaving wet periods insignificant. The establishment of the water markets in the MDB (with IVs) thus would have induced a shift away from the more waterintensive exports, especially in dry years, if the MDB had not adapted its water-use technology\techniques following the market's establishment.
Comparing IV and OLS coefficients is instructive. With or without drought interactions, we obtain negative coefficients for both periods. Interestingly, and this is consistent for our drought estimates, the negative OLS coefficients are smaller than the IV estimates. As we report later, some OLS estimates may even be of the opposite sign (positive). An interesting interpretation offers itself. The endogenous adaptation to water markets (higher water prices), mitigates the need to shift export shares away from water-intensive sectors. This result is intuitive, since changing the sectoral composition entails frictions. Especially more waterintensive sectors reduce their water use. Our estimates corroborate table 1's raw data.
Our results so far confirm our earlier findings of Texas water markets' shift toward less water-intensive activities (Debaere and Li 2020). In addition, we now account for technological adaptation, because our water use measures are not based on biophysical water absorption of crops. Our present analysis casts a richer, more nuanced light on water markets' impact. Figure 4 complements table 2, and checks whether our estimates feed off of pre-existing differential trends before the water markets. Figure 4 provides annual estimates in MDB states of water markets' impact. We plot the coefficients of the double interaction of water intensity and yearly dummies (instead of β 1 and β 2 of specification (1)). Since the graphs should document water markets' impact on sectoral shifts, we present yearly estimates with US water intensity measures that do not endogenously respond to Australia's reforms. As we stretch the data with annual estimates, we obtain more insignificant coefficients. Even so, the negative impact of the reforms compared to the benchmark year (1988) is clear, especially for dry years (marked for MDB states). In table 3, we analyze manufacturing's role (a non-agricultural sector), before discussing the impact of water markets within agriculture and addressing lingering environmental criticisms.
We aggregate all manufacturing and all agricultural subsectors respectively, and investigate agriculture's overall share in the state-by-year panel of table 3. To alleviate concerns about the common trend assumption between MDB and non-MDB states, we apply a synthetic control approach in Columns (3) and (4), which corroborates the robustness of Columns (1) and (2). 17 We obtain significant, negative coefficients after the first-round reform. This illustrates an inter-sectoral shift away from (aggregate) agriculture to manufacturing irrespective of drought. This is intuitive since agriculture uses per Australian dollar, 100 times more water. Based on those huge average differences, one expects also differences in marginal productivity, incentivizing inter-sectoral reallocations. The later reforms also exhibit negative, but statistically insignificant coefficients. 17 The synthetic control approach is used to construct more comparable control group when the treatment group (in our case, the MDB water market states) is small in number and it is difficult to choose comparable controls (Abadie 2021). By using this approach, we 'synthesize' a more comparable control state for each MDB state using a specific weighted average of (raw) control states based on pre-reform logarithmic values of employment and capital stock.
The inter-sectoral movement from agriculture to manufacturing is important. It confirms potential gains from including urban areas, which has not been easy. Table 4 focuses on agriculture's subsectors in the MDB. The left part reports IV instruments, the right part, OLS estimates. The IV estimates in Column (2) display the familiar, statistically significant negative coefficient for dry years. This is consistent with Rafey (2020), who also finds the effects of water markets amplified during droughts-Rafey (2020)'s shorter time period perhaps explains why he does not consider technological/technique changes. The coefficient becomes insignificant in OLS (Column (4)) with endogenous technology adoption. It is noteworthy, however, that for non-drought years, we consistently obtain positive coefficients for OLS and IV. This suggests water markets do not induce the expected shift to less water-intensive crops within agriculture compared to the benchmark in wetter years.
The positive coefficients warrant two observations: • The inter-sectoral shift from agriculture to manufacturing in table 3 counterbalances the reallocation in agriculture to more water-intensive crops in wetter years. This inter-sectoral movement helps explain why we discern a move to less   1994-2006 and 2007-15, respectively (1988-93 serves as the baseline period); 'water intensity' in 100 l per Australian dollar varies by the three periods (i.e. 1988-93, 1994-2006, 2007-15). b 'Drought' is 1 if the total precipitation in is part of the lower 25th percentile of the state's precipitation over the entire 20th century, and 0 otherwise. c 'Industry-year control variables' includes water intensity and the logarithmic value of world total import (except Australia). d 'State-year control variables' include export share of mining, logarithmic values of employment and capital stock, the drought dummy, and weather controls for the current and previous years including precipitation, HDDs above 8 • C and above 32 • C, respectively. e 'Other interaction terms' include 'Period (94-06) × Drought' and 'Period (07-15) × Drought' for Columns (2) and (4). f Weighted by the average total annual export by states in pre-reform years. Standard errors clustered by states in parentheses, * * * p < 0.01, * * p < 0.05, * p < 0.1s.
water-intensive sectors in table 2's regressions for all subsectors, at least during droughts. • The move toward more water-intensive crops supports criticism against water markets. Young (2014) and others caution Australian markets may have increased water use by changing incentives. Irrigators monetized previously un-used sleeper rights. There were attempts to capture water on lands and build small dams that increased the output of water-intensive crops. And there are concerns over the environmental context, and whether markets have lived up to the sustainability promises of the second reform.
Water markets are tools in the fight against water scarcity. They may improve water's allocation, and the production of less water-intensive sectors. Back-ofthe-envelope calculations show overall water use per Australian dollar of export dropped by 8.25% annually in droughts when allowing for endogenous technological improvements of the more water-intensive sectors 18 .

Conclusion
With climate change and ever-growing populations and economies, water-neutral growth is important. Societies have to do more with less. Less water producing more, however, calls for increasing water productivity (decreasing water intensity). Australia's MDB water markets are some of the largest and most advanced whose establishment and design was long and complex. After initial basic institutional changes, water markets took off during the first (1994)(1995)(1996)(1997)(1998)(1999)(2000)(2001)(2002)(2003)(2004)(2005)(2006) and second reform period (since 2007). We study the broader, more aggregate impact of this significant policy intervention, rather than analyze one (of many) policy interventions during this storied history. To do justice to this intricate history, we assess water markets' impact on the economic (export) structure of Australia's states through the various development phases that NWC identified. Our results reflect this complexity and adjusting design features.
Water-market reforms have increased overall water productivity by 8.25% on an annual basis during dry periods. The movement toward higher-value sectors is partly driven by the shift towards less water intensive manufacturing. To our knowledge, we highlight for the first time how water markets induce technology/technique changes that mitigate the need to shift away from water-intensive activities. The reduction in water intensity is especially pervasive for more water-intensive sectors.
Our analysis touches on ongoing environmental criticism of Australia's water markets concerning overall water use. We see in dry years, a shift in the MDB toward less water-intensive crops that is mitigated by adaptation, which comparisons with non-MDB states confirm. However, we find within MDB's agriculture a shift toward more (not less) waterintensive activities in wetter periods, a shift that does not entirely disappear when assessed in the context of all of Australia's agriculture. 18 In Column (4) of table 2, an industry's export share on average decreases by 0.11 percentage point for every additional 100 l per Australian dollar of water intensity (relative to the least waterintensive industry) in drought years after the Phase I reforms. With this estimate we predict the counterfactual export composition for drought years hypothetically in absence of the reform. Combining this counterfactual and the actual export compositions with industry water intensity measures, we can compare their water use. When comparing the water intensity measures under both scenarios, water use per Australian dollar (after Phase I reforms) is 47% of the counterfactual measure without the reform. Thus, reforms nearly double water productivity of exports in droughts. This shift occurred over 9.5 years on average, so in annual rates we get 8.25%.

Data availability statement
The data that support the findings of this study are available upon reasonable request from the authors.

Appendix A. Analyses comparing the MDB states with other states
A.1. Data and empirical strategy As mentioned in section 2 of the main manuscript, summary statistics for variables that we use for the regressions are reported in appendix table 1A. Moreover, in appendix table 1B, we further compare the average of our variables for the MDB and non-MDB states. Overall, the values are similar, except for the non-MDB states' higher share of mining in total exports and our drought measure 19 .
In specification (A1), we extend equation (1) from section 3 to all of Australia. The regression includes the non-MDB states and territories as controls, and hence, the two reform periods are interacted with a dummy for the MDB states and territories. Here also, we cluster the standard error at the state level to address within-state correlations: The main coefficients of interest are β 1 and β 2 . Both should indicate whether water markets in the MDB have a discernable impact when compared to other states without water markets. As emphasized before, it is hard to argue that states are random or identical controls for the MDB. For that reason, we have included state-year effects in the regression (and drop control variables that only vary by year and state) 20 . Thus β 1 and β 2 capture the triple differences of interest in within-state variations among the more and less water-intensive sectors over time by comparing the two groups of states. To be clear, what these two coefficients of interest pick up is not double differences such as the potentially distinct between-state variations in the changes of sectoral composition, but the more comparable within-sector evolutions of similar sectors pooling all states. As a result, the coefficients of β 1 and β 2 can help indicate whether there are common factors across Australia that may conflate any assessment of water markets' impact on the relative export shares of more versus less water-intensive sectors that is singularly focused on the MDB states. Note that we also upgrade this second specification to study the specific role of droughts, and include two extra terms in regression (2): β 7 Reform_I t × WaterIntensity it × drought st × MDB s and β 8 Reform_II t × WaterIntensity it × drought st × MDB s , as well as other triple-and double-interaction terms of drought. As before, we also run regression (A1) separately for all the agricultural subsectors.

A.2. Results
In appendix table 2, we consider the impact of the water markets in the context of all of Australia. We bring in the non-MDB states, and exploit the withinstate variations among the more and less waterintensive sectors across the two groups of states. We want to see to what extent they corroborate our findings for the MDB only. As is well known, it is possible that Australia-wide shifts in activities between more or less water-intensive sectors (due to, for example, crop price movements or changing preferences) may confound the effect of the water markets, obtained in table 3 that was identified only within the water market states of the MDB.
The inclusion of the non-MDB states in appendix table 2 weakens the overall results. The scale of the coefficients is approximately half of the corresponding numbers in table 2 for the MDB states sample, and we obtain less statistical significance. We focus on our preferred specification, which separates out dry and wet spells. The estimates that stand out especially are those for the first reform period. In dry years, there is a consistent shift toward less water-intensive sectors with and without IV. (For the OLS estimates, there is-even in wet years-a statistically significant negative coefficient.) The second reform period still exhibits a negative, but marginally statistically insignificant, effect in drought years.
In this specification across all of Australia, we see again the importance of the endogenous response to water use. The estimated OLS coefficients here are also much smaller than the IV coefficients, indicating that adaptation has consistently mitigated the need to reallocate economic activity to less water-intensive activities. Without that adaptation, there would have been a much stronger swing toward the less waterintensive sectors 21 .
Appendix figure 1 complements the estimates in appendix table 2 and addresses potential concerns as to whether our estimates feed off pre-existing differential trends from before the water markets were in 21 In appendix tables 6 and 7 we provide estimates that control for domestic use with our imperfect (extrapolated) measure. Including this measure does not alter our results in an important way.
place. This figure visualizes for all of Australia the corresponding coefficients, meaning the triple interactions of the MDB dummy, the US water intensity, and yearly dummies (instead of β 1 and β 2 of specification (A1)). Similar to figure 4 of section 4, we observe mostly statistically insignificant fluctuations before the two rounds of the reforms, but movements towards the negative direction afterwards, especially for the marked-out drought years in MDB states.
Appendix table 3 finally puts the estimation results that we obtained for the agricultural sector in the context of Australia as a whole. When considering the agricultural sectors across all states and territories, there is still a significant effect of water markets in the reform periods on the allocation of agricultural exports toward less water-intensive crops for drought years. Here again, the reallocation with the IVs is attenuated in the OLS regressions that allow for technology adaptation. The estimates for the nondrought years are also of interest, however. In one of the reform periods at least, the IV estimates are no longer significant in the context of all of Australia during wet years, which opens the possibility that part of the shift toward more (not less) water-intensive crops may be a nationwide trend. Perhaps, in a country that experiences frequent droughts, water-saving as well as crop-shifting decisions happen even in the absence of a water market and an explicit price of water. It is noteworthy, however, that the OLS estimates for the first period indicate a significant shift toward more water-intensive sectors in water-market states. This could suggest that adapting irrigation techniques or technology may have pushed agriculture in the water-market states of the MDB toward more waterintensive crops. More detailed micro-analysis will be necessary to settle this debate.

Appendix B. Water intensity measures and the two-stage least squares (2SLSs) methodology
In this appendix we describe the construction of our water intensity measures and the 2SLSs methodology.
In general, the water intensity measures for industry i in year t follow the following equation: Water Intensity it = quantity of direct water use it /value of product it . (A2) The first two sections detail how we construct the intensity measures respectively for Australia and the United States. Due to data constraints, in our main specification, we use the pooled average for each of the pre-reform (1988-93), first-round reform (1994-2006), and second-round reform  periods to generate a full balanced panel of water intensity measures. The last section explains the 2SLSs methodology and the rationale for our IV choice.

B.1. Water intensity measures for Australia
As the numerator of equation (A2), the quantity of direct water use by industry for both agriculture and manufacturing is mainly based on the 'water consumption' measures reported in the water account data from the ABS, which is available for 1993-94 to 1996-97, 2000-01, 2004-05, and 2008-09 to 2013-14. When the water account data is not available for some of the crop categories for certain years, we rely on the 'water applied' measure reported in the gross value of agricultural commodities (GVACs) by the ABS, and rescale it proportionally based on other overlapping years to match the 'water consumption' from the water account.
As the denominator, the value of agricultural products by industry comes from the GVAC covering 1988GVAC covering -97 and 2001GVAC covering -2009 and the official data from FAO which contain food prices from 1991 to 2015, values from 1991 to 2013, and quantities from 1988 to 2014. For the few earlier years  where 'water consumption' cannot be complemented with any of the sources, we assume that the physical water intensity for industry i (quantity of direct water use i /quantity of product i ) is fixed over time, and use the price evolvement to back out value-based water intensity from later years to earlier years, i.e., For the manufacturing sector, the denominator of equation (A1) is also available at the industry level from the ABS. However, because a suitable weight measure cannot be applied for most manufacturing industrial categories as used in equation (A3), we can only generate the water intensity for 1993-97, 2001, 2005, and 2009-14. Due to the incompleteness in the time series of the water intensity measures for manufacturing, we use the pooled average of water intensity within each of the three periods (i.e. 1988-93, 1994-2006, and 2007-15) in our analysis.

B.2. Water intensity measures for the United States
The quantity of industry-level water use, as the numerator of equation (A2), comes from the 'water withdrawal' measure in the US Geological Survey (USGS), which is available every 5 years. We take the data from 1985 to 2015, and interpolate for the years that match our annual sample in 1988-2015. However, the USGS data is only available for relatively aggregated categories, such as livestock, irrigation, etc, so we disaggregate the USGS categories based   1988-93, 1994-2006, 2007-15 1988-93, 1994-2006, 2007-15). b 'Drought' is 1 if the total precipitation in is part of the lower 25th percentile of the state's precipitation over the entire 20th century, and 0 otherwise. c 'Industry-year control variables' includes water intensity and the logarithmic value of world total import (except Australia). d 'Other interaction terms' include 'MDB × Water intensity' , 'Period (94-06) × Water intensity' , and 'Period (07-15) × Water intensity' for Columns (1) and (3), and Columns (2) and (4) additionally include 'Period (94-06) × Drought × Water intensity' , 'Period (07-15) × Drought × Water intensity' , 'Drought × Water intensity' , and 'Drought × MDB × Water intensity' . e Weighted by the average total annual export by states in pre-reform years. Standard errors clustered by states in parentheses, * * * p < 0.01, * * p < 0.05, * p < 0.1.  1994-2006 and 2007-15, respectively (1988-93 serves as the baseline period); 'Water intensity' in 100 l per Australian dollar varies by the three periods (i.e. 1988-93, 1994-2006, 2007-15). b 'Drought' is 1 if the total precipitation in is part of the lower 25th percentile of the state's precipitation over the entire 20th century, and 0 otherwise. c 'Industry-year control variables' includes water intensity and the logarithmic values of world total import (except Australia). d Weighted by the average total annual export by states in pre-reform years. Standard errors clustered by states in parentheses, * * * p < 0.01, * * p < 0.05, * p < 0.1. The denominator for the output value by industry also comes from the same publication provided by the Bureau of Labor Statistics in 2017 that is mentioned above, which consistently provides the necessary statistics for our entire sample period.

the US Labor
After putting all the above parameters into equation (A2), we are able to construct a water intensity measure for the US industries covering our entire sample period.  1994-2006 and 2007-15, respectively (1988-93 serves as the baseline period); 'Water intensity' in 100 l per Australian dollar varies by the three periods (i.e. 1988-93, 1994-2006, 2007-15). b 'Drought' is 1 if the total precipitation in is part of the lower 25th percentile of the state's precipitation over the entire 20th century, and 0 otherwise. c 'Industry-year control variables' includes water intensity, the logarithmic value of world total import (except Australia), and domestic demand by industry categories with data extrapolated for 1988-94. d 'State-year control variables' include export share of mining, logarithmic values of employment and capital stock, the drought dummy, and weather controls for the current and previous years including precipitation, HDDs above 8 • C and above 32 • C, respectively. e 'Other interaction terms' include 'Period (94-06) × Drought' and 'Period (07-15) × Drought' for Columns (2) and (4). f Weighted by the average total annual export by states in pre-reform years. Standard errors clustered by states in parentheses, * * * p < 0.01, * * p < 0.05, * p < 0.1.