Conserving the Ogallala Aquifer in southwestern Kansas: from the wells to people, a holistic coupled natural–human model

The impact of water policy on conserving the Ogallala Aquifer in Groundwater Management District 3 (GMD3) in southwestern Kansas is analyzed using a system-level theoretical approach integrating agricultural water and land use patterns, changing climate, economic trends, and population dynamics. In so doing, we (1) model the current hyper-extractive coupled natural–human (CNH) system, (2) forecast outcomes of policy scenarios transitioning the current groundwater-based economic system toward more sustainable paths for the social, economic, and natural components of the integrated system, and (3) develop public policy options for enhanced conservation while minimizing the economic costs for the region’s communities. The findings corroborate previous studies showing that conservation often leads initially to an expansion of irrigation activities. However, we also find that the expanded presence of irrigated acreage reduces the impact of an increasingly drier climate on the region’s economy and creates greater long-term stability in the farming sector along with increased employment and population in the region. On the negative side, conservation lowers the net present value of farmers’ current investments and there is not a policy scenario that achieves a truly sustainable solution as defined by Peter H. Gleick. This study reinforces the salience of interdisciplinary linked CNH models to provide policy prescriptions to untangle and address significant environmental policy issues.


Introduction
Our world faces a public policy conundrum. Crop yields on many varieties have tripled over the past 50 years, with irrigated cropping practices accounting for 40 % of the total increased level of production (United Nations, 2011, p. 3). Even so, food deserts, often created by market inequities, leave over 1 billion people worldwide malnourished (United Nations, 2011, p. ix). If projections are correct, the situation in the future does not improve. By 2050, the UN estimates that the world's population will have grown by another 2.2 billion people, most of whom will live in impoverished countries (United Nations, 2017), while demand for crops and meat products will soar by 70 % (United Nations, 2011, p. 7). Irrigated crops will continue to be crucial for meeting the world's future demands for nutrition. Unfortunately, many irrigated croplands are in regions that are, or are becoming, freshwater challenged (United Nations, 2011, 7-12). This is the case for the High Plains Aquifer in the US, also referred to as the "Ogallala Aquifer" (see Fig. 1). The entire aquifer spans 450 000 km 2 and underlies 27 % of the irrigated land in the US (Dennehy, 2000). Powell (1879) classifies this semiarid grassland ecosystem as an arid land lying west of the 100th meridian with a mean annual precipitation less than 500 mm (20 in). The aquifer is 1 of 4 "critical areas" for "annual renewable water" in the Western Hemisphere and 1 of 22 worldwide (Montaigne, 2002).
Thus far, the scientific community has developed hydrological models that confirm what we already know; that this natural system aquifer is already past its peak groundwater depletion (Steward and Allen, 2016) and will soon be so diminished that it can no longer sustain its current irrigation farming practices (McGuire, 2014;Scanlon et al., 2012;. This policy study uses a coupled natural-human (CNH) system approach focused on Groundwater Management District 3 in southwestern Kansas. The 12 counties of GMD3 are classified as hyper-extractive (Aistrup et al., 2013), with similar water resource extraction patterns to other Ogallala counties. To study GMD3, we develop an integrated, cross-disciplinary, system-level, theoretical approach, linking agricultural land and water-use practices, changing climate patterns, economic trends, and population dynamics to issues of groundwater sustainability. In so doing, we (1) accurately model the current CNH system, (2) forecast the outcomes of policy scenarios to transition the current groundwater-based economic system toward avenues that are more sustainable for the social, economic, and natural systems, and (3) develop public policy options that will conserve the aquifer while minimizing the economic cost for the region's communities.

Methods
The scope of this study is the 12 counties of Groundwater Management District no. 3 (GMD3) in southwestern Kansas (Fig. 1). Farmers in this region of the Ogallala, similar to other regions of this aquifer, tap groundwater to raise corn, sorghum, soybeans, wheat, and alfalfa. In turn, the irrigated grains and alfalfa supply inputs for large-scale confined feedlots that deliver finished cattle for several of the world's largest meatpacking factories (Broadway and Stull, 2005), making GMD3 one of the most productive value-added agricultural regions in the US and the world . However, as illustrated in Fig. 1, the emergence of this value-added agricultural economy has imposed a heavy toll on the aquifer, reducing its saturated thickness and altering its recharge (Custodio, 2002). With irrigated crops accounting for 97 % of the water withdrawals from the Ogallala Aquifer in western Kansas (United States Department of Agriculture, 2013), developing agricultural policies that conserve the life of the aquifer is essential for maintaining the long-term economic health and vitality of the region and Kansas.

Integration methodology
The integrated model is composed of independent disciplinary models that each conform to and are linked within the Open Modeling Interface (OpenMI) Standard (Gregersen

Socioeconomic impact model
The socioeconomic impact model uses cash flow at time t (CF t ), also referred to as "gross profit". CF t = revenue (i.e. price · yield) − costs (fixed costs + variable costs + lift costs). Based on CF t , the net present value (NPV) is the discounted CF t , calculated as (CF t /(1 + K t )) t where K is the discount rate on investments (set at 4 %) and t is a counter for the year, starting with 1 in 2013. In this model and in the crop choice model, all monetary variables are in inflation-adjusted 2013 dollars. Many economic impact models also estimate economic multiplier effects using Impact Analysis for Planning (IMPLAN) (Group, 2010). IMPLAN produces direct, indirect, and induced impacts on total economic activity, value-added activity, and employment. Our linked CNH model focuses on community impacts that yield employment opportunities to support a stable population base. Over the years, as agricultural production has become increasingly mechanized and technologically based, farms have consolidated, creating larger farms with fewer employees (Flora et al., 1992). This has led to considerable population decline in most rural-farming-dependent communities in the Great Plains region. Even though wealth creation from farming is important, from a community development standpoint, this wealth is most important when it translates to jobs and population. In western Kansas, IMPLAN estimates that it takes about USD 1 million CF t to produce 1.2 full-time equivalent (FTE) employees, of which 0.88 FTE are directly tied to crop production agriculture and 0.32 FTE represent indirect and induced employment. There are slight variations between CF t for irrigated cropland and nonirrigated cropland, which are detailed in Table A1 in Appendix A2.
To estimate the population impact of this employment, we conducted a cross sectional time series analysis with panelcorrected standard errors (TSCS) Katz, 1995, 2011) from 1970 to 2010, where the time increments are every 5 years (1970, 1975, 1980, etc.), the cross sections are the 12 counties of GMD3, and the error terms are both heteroskedastic and serially correlated (AR1).
where i,t = υ i,t + ρ i,t−1 . In this 40-year TSCS regression model, the independent variables are each county's employment levels (no. of employees) in agriculture, manufacturing, construction, health care, government services, and education, while the dependent variable is each county's total population. We obtained US Census population data and Bureau of Economic Analysis employment data from Woods and Poole Economics (2012). This equation explained 92 % of the variance in total population in these counties during this time period (see Appendix A1 for model results). We multiplied the regression coefficient for agricultural employment (2.15) by the IMPLAN's employment multiplier to calculate the impact on population (2.15 · (.88) = 1.88 people). Thus, each USD million CF t from crop production in GMD3, supports an additional 0.88 FTE and 1.88 residents in region. Given the connection between agriculture and value-added meat production, we assign the other 0.32 FTE emanating from indirect and induced employment impacts to the coefficient for manufacturing (.32 · 1.33 = .43 people). Taken together this suggests that each USD 1 million in CF t from crop production supports 1.2 additional jobs and 2.3 more people living in the region.

Crop choice model
The crop choice component is an iterative positive mathematical programming (PMP) model (Howitt, 1995) that simulates farmers' allocation of arable land to different crops in each county. The model operates on an annual time step, with each execution predicting farmers' choices in a single growing season. 1 In addition to harvested crop prices and cropspecific costs of production, the model accepts as inputs the current (county average) depth to water and saturated thickness of the aquifer. Depth to water affects water extraction costs, while saturated thickness affects the pumping rate of wells, which in turn creates an upper bound on the annual extraction of irrigation water. Recent work has emphasized the role of well pumping rates on crop and water-use choices (e.g., Foster et al., 2015). The model simulates land allocations as the solution to a constrained optimization problem that represents farmers' profit-maximizing mix of land uses, given price conditions, water extraction costs, and the constraints on water and land availability. The model outputs are the predicted acres planted to each crop. A separate instance of the model was calibrated to data from each county in the study region. Each county model simulated acreages for irrigated and nonirrigated plantings of the five dominant in crops in the region: wheat, corn, sorghum, soybeans, and alfalfa. Nonirrigated production of soybean and alfalfa is unfeasible given regional hydroclimatology and so are only included as irrigated crops. Thus, eight crop categories were modeled. The models were calibrated to the 2006-2008 average of observed acreages, yields, and prices for the eight crops by county, the most recent period for which comprehensive county-level data are available from the National Agricultural Statistics Service (NASS). Expected crop yields are simulated within the model from water response functions in Martin et al. (1984), which were calibrated to yield data from NASS and weather data from the Kansas Weather Data Library. 2 There is a long history of increasing crop yields due to genetic improvements in plant varieties (United Nations, 2011, p. 46). Considering this history and the continued investment in plant genetics by industry and governmental agencies, the crop choice model assumes that yields will continue to improve into the future based on the noncompounded percentage growth rates estimated from time series regression of western Kansas yields for irrigated and nonirrigated plant varieties from 1974 to 2009 (see Appendix A3, Table A2) (Rogers and Lamm, 2012). The average annual improvements in yield range from a high of 1.28 % for irrigated corn to a low of 0.55 % for dryland wheat and 0.53 % for irrigated sorghum. Crop production costs, excluding irrigation, were obtained from Kansas State University Extension enterprise budgets and were increased over time at noncompounded percentage rates in proportion to yields based on a regression analysis of budget and yield data using 2006-2012 observations. While base-level yields and costs were calibrated to 2006-2008 data, growth percentages were applied to adjust the initial simulation year to correspond to 2013. The cost of irrigation water was calculated separately based on the energy costs from pumping lifts (Rogers and Alam, 1999). Details on the model development and calibration methods are in Clark (2009), Bulatewicz et al. (2014, and Garay Armoa (2015).

Crop production model
The crop production component projects grain yield and irrigated water by using the erosion-productivity impact calculator (EPIC) model (Williams, 1995). EPIC simulates daily crop growth by representing three major processes: (1) phenological development, (2) dry matter production and partitioning to plant tissues resulting in growth, and (3) economic yield. The model reproduces the results of irrigation, fertilization, tillage, variety selection, alternative production calendars, etc. Plant growth is estimated from intercepted solar energy and plant leaf area. Daily dry matter is accumulated for the growing season as controlled by heat units or environmental conditions (typically freeze events for summer crops), and yield is estimated using a total-biomass-to-grain ratio. EPIC is able to simulate multiple crops because it embodies a generic plant model that can be parameterized to represent different species.
We previously calibrated the model for use in western Kansas (Bulatewicz et al., 2009) and further refined the parameters to support nonirrigated cropping for this study. The model component, developed in an earlier effort (Bulatewicz et al., 2014) and implemented using the simple model wrapper (Castronova and Goodall, 2010), has an embedded set of simulated output data from EPIC collected by executing the model for all combinations of the relevant inputs (soil, crop, management, weather). The component operates over a set of (independent) polygons of variable size, accepting inputs for soil type, weather station, and crop and providing outputs for yield and water use.

Climate
Each simulated year's weather is determined by a random draw from meteorological records between 1985 and 2012. We build the likelihood of climate change into our simulation about the future of GMD3. ECHAM5 climate change models for the High Plains region suggest future regional warming and a gradual increase in extreme weather events, pointing toward a less suitable climate and thus reduced yields for agricultural production (Zabel et al., 2014). To model this climatic progression, we weight the weather data from 1985 to 2012 such that years of below-average dryness will, over 100 years, gradually become 25 % more likely to occur than they are now. We find that this captures both the prolonged periods of drought that are likely to become typical, while allowing for shorter-lived periods of plentiful precipitation. 3

Groundwater model
The groundwater modeling component provides estimates of groundwater storage and the changes in storage due to pumping and natural hydrologic processes. This model is linked to the crop production and economic crop choice model using OpenMI (Gregersen et al., 2007), and operates on the common county-level scale. Conservation of mass requires that recharge minus extractions is equal to the annual change in storage: This groundwater model integrates these spatially and temporally variable components of the hydrologic cycle to provide these fluxes on the common county-level aggregation scale, which are consistent with the scales of previous studies in the study region Allen, 2013, 2016). Specific steps used to prepare groundwater data follow. Storage is obtained from groundwater observation wells, kriging across wells to give a surface of saturated thickness, multiplying by specific yield to give water content, and integrating across the aquifer area within a county (Steward and Allen, 2013). Recharge is obtained by spatially integrating results from Hansen (1987) 4 , and extraction is obtained from the crop irrigation component, which was parameterized against historical pumping rates recorded in the WRIS water-use reports. The county-level model ignores the changes in storage resulting from groundwater movement between counties, since groundwater moves with an average velocity of only 30 cm day −1 , driven consistently by the west-east sloping aquifer (Gutentag et al., 1984), and the differences between what enters and leaves each county represents a very small fraction of changes in storage. This county-level model provides groundwater availability for the crop production model, and also ignores other water use such as municipalities, industry, and feedlots, which have historically used much less than 5 % of the groundwater extractions in each county. 5 The baseline model predictions accurately reproduced the groundwater data throughout the historical period. Model results were also compared to the future predictions of a higher-resolution fishnet model of Seward County . Similarly, the results from the model resolution of this study, when aggregated to the region, reproduce longer term regional projections (see also Appendix A4, Table A3) .

Baseline, point estimates, and estimates of uncertainty
This integrated model is used to develop point estimates for important variables, as well as estimates of uncertainty, by simulating a policy scenario 100 times, where each policy scenario simulates a period of 100 years and the parameter that was resampled each simulation was the weather year. Figure 3 illustrates the significant findings from the groundwater, crop choice, crop production and socioeconomic impact models. The point estimates for each component of each year is presented as the average from the 100 simulations and the estimates of uncertainty are the standard deviations around each point estimate. Our holistic CNH model predicts an unsustainable outcome for the aquifer in all counties if current conditions remain unabated, which is consistent with similar results obtained using different methodological approaches (McGuire, 2014;Scanlon et al., 2012;Steward and Allen, 2016). Results in Fig. 3 illustrate that the acreage for irrigated corn continues to increase until 2030 (ca. 328 000 ha), showing that corn will remain profitable despite increased pumping depths. However, as the average saturated thickness in 6 of 12 counties falls below 9 m (the saturated thickness necessary for full use of center pivot irriga-5 In the simulation, groundwater pumping is based upon an annual decision for when to start pumping and when to stop, where the well is traditionally left on throughout the growing season. Thus the dynamics are drawdown throughout a growing season and recovery before the next pumping cycle, where large drawdowns may occur when the wells actively pump (Mullican III, 2012), and a new elevation becomes established during a recovery period (Dugan et al., 1994, p. 23). Thus, the groundwater model exchanges data at the same annual frequency as the economic crop choice model that dictates the annual water requirements. tion; Hecox et al., 2002 6 ) and below 15 m in all but Meade and Seward, the amount of acreage planted in irrigated corn is projected to decline significantly. By 2112, there are fewer than 200 000 ha of irrigated corn. However, the acres planted to dryland corn soars to ∼ 500 000 ha, surpassing by 2070 the acreage planted in dryland wheat. As the capacity of the aquifer to support irrigation decreases over time, the average saturated thickness in each county stabilizes at less than 10 m, reflecting that, on average, farmers no longer have the capacity to consistently draw the high volumes of water necessary for center pivot irrigation. Even though the CF t from irrigated crops will increase from USD 400 million to just over USD 600 million, the CF t from dryland crops will increase from ∼ USD 50 million at the beginning of the simulation to about USD 400 million at the end of the simulation. This leads to continued positive employment impacts for irrigated and dryland agriculture, increasing from a total of 1000 workers currently to 1800 workers at the end of the simulation who owe their livelihoods to crop production. The NPV reaches USD 10 billion by 2040 and USD 13.3 billion by 2060.
However, hidden behind these point estimates is much uncertainty that comes from relying on dryland farming in the semiarid counties of GMD3. This is the most apparent by examining the outputs of the socioeconomic model. Although the CF t for all dryland production is generally positive, the variation around that average grows (whiskers represent 2σ from the average) over time relative to the greater number of acres planted to dryland crops and the increased prevalence of a drier climate. This creates a higher probability for dryland crop failure in years with lower rainfall levels.
We have two points of caution regarding these findings. First, the accuracy of any model of this nature declines as it forecasts into the future. Thus, some of this uncertainty in dryland crop production is a function of our modeling methodology. Second, farmers in the GMD3 and US mitigate the risk from drought and other weather-related calamities with federally subsidized crop insurance, administered by the Risk Management Agency of the USDA. In the event of crop failure, insured farmers receive a settlement based primarily on the price of the crop during harvest multiplied by the average yield in that county over the past 10 years (Risk Management Agency of the United States Department of Agriculture, 2016). In practice, crop insurance payments for farmers are much less than the cash flow from that crop in an average year. This suggests that as the climate becomes hotter and drier, the negative swings in employment and population will be significant even if a portion of farmers' incomes from dryland fields are insured. This also reinforces an important 6 Even though the average saturated thickness for most counties is low, there is much variation among wells in each county. Thus, some wells may have more than 9 m of saturated thickness, allowing these farmers enough water to irrigate (Hecox et al., 2002), while others may have less. The right graph depicts total water consumed in GMD3 by each irrigated crop. Middle row panels: the left graph shows the average level of saturated thickness of the Ogallala in each of the 12 counties over time. Note that only two of the 12 counties end the simulation with an average saturated thickness greater than 15 m, 6 have averages less than 9 m. Generally, more than 9 m of saturated thickness is needed to irrigate using center pivots. The right graph sums the NPV across all 12 counties over time. Bottom row panels: the left graph illustrates the predicted number of additional jobs created in a simulated year due to the direct, indirect, and induced impacts of irrigation and dryland crop production, respectively. Note that over time, the level of variation in employment due to dryland crop production is exceptionally large, reflecting the greater levels of uncertainty introduced by a gradually warmer climate. The graph on the right depicts the number of additional people who live in the GMD3 in that year because of irrigated and dryland crop production. The population impact mirrors the level of variation shown in the previous graph. The whiskers denoting the level of uncertainty (2σ ) are excluded on the top two graphs on the right to maintain the clarity of presentation. historical lesson. In a retrospective analysis of counties in the Great Plains, Hornbeck and Keskin (2014) found that new access to groundwater for irrigation mitigated the impact of drought. However, drought sensitivity in irrigated counties began to increase as farmers switched to higher-value waterintensive crops and groundwater access declined.

Sustainability and the Ogallala in GMD3
Our definition of sustainability parallels that of Peter H. Gleick, who defines sustainability in terms of using water to allow "human society to endure and flourish into the indefinite future without undermining the integrity of the hydrological cycle or the ecological systems that depend on it." (Gleick, 2000). We add, however, one proviso to this definition. The economies of the High Plains Aquifer region have already substantially depleted the aquifer (see Fig. 1). Stream flows for riparian and aquatic ecosystems have been impaired (Ahring and Steward, 2012). Reversing the impacts of the past 50 years may not be possible without ceasing all irrigation activity in the region. Even then, given the recharge rates of the aquifer, it would take 500 to 1300 years to fully recharge the aquifer in western Kansas . This is not a viable scenario. 7 Thus, our sustainability policy scenarios focus on maintaining current saturated thicknesses and stemming the current pattern of continuous depletion, while maintaining to the extent possible the employment levels, wealth generation, and population impacts in the region. 8

Scenarios 1 and 2
The first two policy scenarios use two separate Kansas water conservation statutes to model different policy approaches for achieving a 10 to 20 % reduction in irrigation. The Kansas Groundwater Management District Act contains provision K.S.A. 82a-1036, which allows the Chief Engineer to designate an "Intensive Groundwater Use Control Area" (IGUCA) to implement corrective control provisions reducing the permissible groundwater withdrawal based upon relative dates of priority of such rights (this statute also allows for a rotating schedule; Steward et al., 2008). Thus, this first scenario takes at least 20 % of fields out of irrigated crop production based on senior versus junior water rights. This is modeled by reducing by 20 % the acreage assigned to irrigation in each county. We assume that this acreage will be returned to dryland production. The second policy scenario emanates from K.S.A. 82a-1041(d)(1), which allows adjacent water users in a region to create "Local Enhanced Management Areas" (LEMAs). Under this statute, if a large consensus of irrigators in a contiguous area agree to limit water use by a prescribed percentage then that reduction becomes a legally enforceable limitation on all irrigators. Currently, there is one LEMA restricting irrigation in Kansas, located in Sheridan and Thomas counties, which are north of GMD3. LEMAs have two advantages over the IGUCA approach. First, it is a bottom-up process, where irrigators in an area agree to the restrictions through a consensus among themselves instead of regulations set centrally by the chief engineer. The work of Ostrom (2010) suggests that this is a better institutional design for managing common pool resources. Second, this represents a "shared pain" approach; an approach that is preferred by irrigators over the enforcement of junior versus senior water rights. We operationalize the LEMA policy scenario by assuming that irrigators in GMD3 agree to create a LEMA that doubles the interval between irrigation applications, thus slowing the rate of water application by about 12 % across the management district. Based on previous research, we do not anticipate that either of these scenarios will produce a sustainable outcome. Rather, the focus here is on the effects of restrictions on farmers' NPV and the local communities. Table 1 reports the results of simulations for scenarios 1 and 2 compared to the baseline analysis. Removing 21 % of fields from irrigation based on junior water rights means there is an immediate reduction in the number of irrigated acres in all the crop varieties; however, irrigated corn only decreases by 18 000 ha (300 000 ha in the baseline to 282 000 ha in junior rights scenario). The crop choice model suggests that the bulk of acres taken out of irrigated production will move to dryland corn (an increase of 106 000 ha in 2013 over the baseline) production. By contrast, doubling the irrigation interval has little appreciable impact on crop choices in GMD3 compared to the baseline. For scenario 1, the maxi-mum number of acreage for irrigated corn tops out in 2063 at 329 000 ha compared to 350 000 in 2055 in scenario 2. Interestingly, under scenario 2, the number of acres under irrigation increases incrementally from 741 to 811 000 ha in 2070 (not shown). This pattern is consistent with previous research (Ward and Pulido-Velazquez, 2008a, b; noting that farmers use their water savings on more fields to increase their capital returns. 9 Both scenarios improve the lifespan of the aquifer by 10 to 15 years, but neither comes close to achieving aquifer sustainability given the very slow rate of recharge in most of GMD3. Both approaches produce similar types of outcomes for CF t and employment. The NPV by 2060 is USD 11.3 billion for the junior rights scenario and USD 12.1 billion for the LEMA scenario, compared to USD 13.3 billion for the baseline. Interestingly, in the long term, communities in GMD3 benefit from conservation. Extending the life of the aquifer in both scenarios leads to more people being employed by the direct and indirect impacts of production farming (an average of 340 to 400 workers a year by 2112).

Scenarios 3-5
Given that neither of first two policy options achieves sustainability, we explore the relationship between water conservation and the associated socioeconomic consequences for the farmers and communities in GMD3. To do so we simulated the implementation of a LEMA across GMD3 that incrementally increases the interval for irrigation by 3X (a 26 % reduction compared to 2014 usage), 4X (a 35 % reduction), and 6X (a 48 % reduction). We focus on the LEMA policy approach because it is a theoretically more pleasing policy prescription (see arguments above and Ostrom's work; Ostrom, 1990). Significantly, increasing the interval by 6X stretches to the maximum limit the marginal utility of irrigation for the purpose of assuring increased crop yields. Thus, after the 6X point, the LEMA approach begins to lose its policy integrity.
Tripling the irrigation interval for irrigated corn production gradually increases the acres in production from 300 000 ha in 2013 to a peak of 376 000 ha by 2071. Dryland corn, which in the baseline analysis becomes the most predominant crop after 2080, only surpasses irrigated corn in acres planted in 2110, at the end of the 3X simulation. Given the large variation in yields and revenues associated with dryland corn production, policies that reduce dependence on this high-risk crop are desirable. The 3X scenario tends to benefit the communities of GMD3, as the number of additional people employed due to the direct and indirect impacts of production agriculture increases from fewer than 1000 in 2013 to 1940 in 2112. Similarly, the number of people living in the region because of direct and indirect economic impacts from 9 Scenarios 3-5 follow this same pattern. irrigation and dryland farming increases from 2000 in 2013 to 4200 in 2112.
Disappointingly, increasing the irrigation interval by 4X or 6X does not produce sustainable outcome for the aquifer. A total of 6 of the 12 counties under the 4X scenario and 2 of 12 counties under the 6X scenario still end the simulation with average saturation depths of 15 m or less. There is also an economic cost to irrigators to achieve this level of water savings. The NPV in 2060 shrinks to USD 9.5 billion for 4X scenario and USD 7.6 billion for the 6X scenario. On the positive side, after an initial decline in employment and population early in the simulation, both rebound to levels just above the baseline.

Conclusions
These results corroborate previous studies that show that conservation often leads initially to an expansion of irrigation activities, as farmers use their water application savings on more fields to increase their capital returns (Ward and Pulido-Velazquez, 2008a, b;. However, our coupled model extends this finding by showing that the expanded presence of irrigated acreage in GMD3 will reduce the impact of an increasingly drier climate on the region's economy and create greater stability in the farming sector along with increased employment and more people living in the region. The two policy mechanisms discussed in this study, (1) senior versus junior water rights and (2) LEMA, represent policy tools that have thus far only been used in Kansas after the impacts of groundwater depletion have manifested. Thus, they are policy tools for managing a crisis. This begs the question whether one of these conservation enforcement tools or some other policy prescription can be brought to bear to conserve the aquifer before a crisis occurs?
Our scenarios demonstrate that any form of conservation enacted today lowers the income of agricultural production in the short and long term. The differential between what the income producers would have earned under the baseline model versus what they are likely to earn if conservation measures were enacted represents an opportunity cost. This opportunity cost is the major obstacle preventing the adoption of any conservation measures.
Policy analysts working for GMD3 in the early 2000s noted this opportunity cost associated with conservation and promoted a crop subsidy to address this differential (Gilson and Aistrup, 2001). What would be the cost if a subsidy were provided for conservation? We estimate this by taking the difference between the average baseline-estimated CF t over the first 5 years (USD 420 million) and each scenario's corresponding average CF t . Thus, for example, the estimated average annual subsidy to implement the 3X scenario is USD 113 million and for the 6X scenario is USD 218 million (2013 dollars). This subsidy is not trivial, but it would allow producers to earn the income they would have if they had continued irrigating at 2013 levels. 10 Subsidies or other policy interventions need to define an outcome considered desirable to address the current situation. In our case, there are two major goals that potential policies could work towards. The first is implementing procedures to extend the current agricultural production regime as long as possible. In this case, addressing the opportunity costs requires an investment (the subsidies) with the expectation that the extended lifetime of the aquifer provides social and economic goods substantial enough to justify the investment before the inevitable pumping reductions imposed by low rates of groundwater recharge. This would provide opportunities for local communities to accumulate resources before that happens, relying on the market to determine how much can be accumulated. The second option is using policy tools to navigate the regional economy toward a different agricultural regime, such as a specific dryland agricultural system. In this case, the desired outcome would be facilitated by policies more directed toward that outcome, assuming that the region itself would be better off under those conditions. These choices reflect the age-old debate in policy making about helping communities accumulate resources to invest in any way they see fit, hoping for a sustainable regional economy to emerge, or provide assistance to move stakeholders along a specific path determined at the regional level. Regardless of which direction policy makers choose, the desired outcome must be clearly defined, as rudderless boats seldom reach their destinations no matter how low the water levels drop.
Perhaps the most important research outcome is that this study establishes the salience of interdisciplinary linked CNH models that seek to untangle and address significant environmental policy issues. Other studies of intensive wateruse regions have been insightful, but none have incorporated the breadth of our model's components or have used an OpenMI framework. Our modular and holistic model, which includes the major variables of socioeconomic impact, crop choice, crop production, and groundwater supply, points toward the policies that can be implemented today to bring a more sustainable future to this region.
Additional research is necessary to refine this CNH model to (1) model the dynamic nature of the grain commodities market, (2) take into account new efforts by agri-industry and universities to double grain production levels over the next 15 years, and (3) take advantage of improved scientific models of climate change to more accurately portray the uncertainties that irrigators face and the additional demands for water that climate change may induce in this water-challenged region. Researchers in the future can adapt this holistic model to take account of these factors to build new models of sustainability from the wells that pump the water from the aquifer to the communities where people are affected.
Appendix A: CNH model details

A1 Socioeconomic model
To estimate the population impact of this employment, we conducted a cross sectional time series analysis with panelcorrected standard errors (TSCS) controlling for autocorrelation (AR 1) (Beck and Katz, 1995, 2011) from 1970 where the time increments are every 5 years (1970, 1975, 1980, etc.), the cross sections are the 12 counties of GMD3, and the error terms are both heteroskedastic and serially correlated. We found the following: Total population i,t = 1.09 + 2.15 * * Agriculture i,t + 1.33 * * Manufacturing i,t + 1.45 * * Construction i,t + 8.14 * * Health i,t + 3.42 * * Government i,t + 6.38 + Education i,t * * p ≤ .001, + p > .05 .
Although the manufacturing impact coefficient may seem low, most of the workers in meatpacking plants are immigrants from Mexico and Central America, who are single or are married and have left their spouses and families in their home country to work in these meatpacking facilities (Broadway and Stull, 2005). Thus, each employee in manufacturing in GMD3 does not yield the type of population impact that manufacturing would in other regions.

J.
A. Aistrup et al.: Conserving the Ogallala Aquifer A2 IMPLAN multipliers for crops Table A1 shows IMPLAN multipliers for crops in western Kansas used in the socioeconomic model. Estimating employment impacts can be controversial when computed as a function of total expenditures and revenues, which tends to overestimate the employment impact. We choose instead to calculate employment impacts as a function of CF t (i.e. profits).  Table A2 shows the historic data used to estimate the increasing crop yields for corn, soybeans, sorghum, wheat, and alfalfa, for both irrigated and dryland varieties.  Table A3 compares the predictions of regional (GMD3 total) water withdrawals from coupled model simulations from 2013 to 2016, the period of overlap between simulated model results and observed data. Total withdrawals are a key summary measure that combines the results of several model components and drives regional economic impacts. Observed data for years with specific realized weather would not be expected to match mean results, which are averaged across the weather distribution. Observed data diverge from mean predictions by 0.18 to 1.13 SD (standard deviation). Observations from all 4 years are well within the 90 % confidence interval.