Working Paper 2008-09 Forecasting Single-Family Residential Water Consumption for Phoenix and Paradise Valley, Arizona Jamie Patterson and Elizabeth A. Wentz Forecasting single-family residential water consumption for Phoenix and Paradise Valley, Arizona Jamie Patterson a,* and Elizabeth A. Wentz a a Arizona State University School of Geographical Sciences Tempe, AZ 85287-0104 USA * corresponding author, jpatterson44@gmail.com wentz@asu.edu Submitted to Computers, Environment and Urban Systems, July 2008 1 Forecasting single-family residential water consumption for Phoenix and Paradise Valley, Arizona Abstract. We developed four scenarios to forecast the impacts of policy measures affecting new housing unit lot sizes on single-family residential water consumption for the City of Phoenix and Town of Paradise Valley, Arizona. Policies approximated the statistical distribution of year 2000 lot sizes or produced smaller lot sizes. Policies mirrored the year 2000 spatial pattern of lot sizes or were distributed with higher densities (smaller lots) near transportation corridors. Along with lot size, we used census tract level parameters from year 2000 water estimates for percent homes with pools, percent homes with mesic landscaping, and mean household size to forecast water consumption. Each scenario was modeled with a local geographically weighted regression (GWR) model and a global ordinary least squares regression (OLS) model. This is the first study to apply GWR to forecasting. A policy mirroring the year 2000 spatial pattern of lot sizes with smaller lots produced the lowest mean annual housing unit water consumption at 214.77 ccf. Higher density lots along transportation corridors would have little impact on consumption. GWR forecasted lower per capita water consumption than the OLS for all scenarios. Our approach should be used in other forecasts exploring policy implications for natural resource management. Keywords: spatial forecasting, residential water consumption, geographically weighted regression 2 1. Introduction Water is essential for human existence requiring extensive management, particularly in urban areas where supplies and infrastructure must meet the needs of a heterogeneous and growing population. As population grows in urban areas, ensuring a long-term, safe and reliable source of potable water becomes essential. Determining the amount of water needed by residents is an integral component to water management. The research here focuses on forecasting single-family residential water consumption for the City of Phoenix and Town of Paradise Valley under different growth and development plans. Management for water in the City of Phoenix and Town of Paradise Valley is needed because climatic uncertainty threatens future water supplies in an area where projections show the Phoenix metropolitan statistical area doubling in population around the year 2050 (ADES, 2006). Residential water consumption research has investigated aspects such as the overall cost of water (Arbues, Barberan, & Villanua, 2004; Kenney, Goemans, Klein, Lowrey, & Reidy, 2008), the use of different billing practices (Olmstead, Hanemann, & Stavins, 2007), the impact of policy decisions (Vickers, 1993), and the effectiveness of water conservation (Michelsen, McGuckin, & Stumpf, 1999). For desert cities, such as Phoenix, research into water use has evaluated residential water use from several perspectives, including the impact of the urban heat island, factors that influence residential water consumption, and the employment of Bayesian Maximum Entropy (BME) to handle uncertainty associated with modeling (Lee, Balling, & Gober, 2008). Guhathakurta and Gober (2007) found a temperature rise of 1°F, due to urban heat island effects, would increase monthly residential housing unit water demand in the City of 3 Phoenix by 290 gallons. Wentz and Gober (2007) modeled the effects of indoor and outdoor residential characteristics single-family residential water consumption rates in the City of Phoenix, Arizona. They found that household size, the presence of a pool, irrigated landscaping, and lot size significantly influenced water consumption. Missing from the above research is an approach to better understand the effects of planning and policy choices on the spatial distribution of future residential water consumption. A broad set of literature exists on the theory, techniques, and applications of forecasting in fields including economics and econometrics, finance, technology, demographics, energy, climate, crime, implementation research, psychology, uncertainty in decision making, and evaluation of forecasting methods and approaches (Hyndman, 2007). Scenario generation has been an established tool in land use planning for many decades (Xiang & Clarke, 2003). Ordinary least square (OLS) regression has been successfully used for modeling and forecasting dependent variables. Briggs et al (1997) modeled traffic-related air pollution in Amsterdam, Huddersfield, and Prague with OLS. Mapped comparisons of modeled versus actual pollution across referenced points revealed accuracies ranging from 79 to 87 percent. In a geographic context, OLS assumes coefficients are globally constant, or stationary, over an entire region. Given that many social processes are not stationary, this assumption is frequently violated. Fotheringham et al (2002) developed geographically weighted regression (GWR), which assumes some relationships between variables are non-stationary over geographic space. GWR not only identifies the degree to which variables influence outcomes, it sheds light on how local conditions influence variables. Location-specific insights may be used to generate new hypotheses and build more effective models. To date, GWR has been confined to 4 historical or existing environments, largely in social science research. In a recent example, Yu (2007) modeled owner-occupied single-family house values in the City of Milwaukee, Wisconsin and determined housing unit characteristics, including floor size, air conditioners, number of bathrooms, and fireplaces add more value to houses in wealthier areas than in less affluent areas. Given the successful application of GWR in historically-based studies, a logical next step is its inclusion in forecasting to estimate future outcomes. Our research does this, expanding the approach and dataset of Wentz and Gober (2007) to the realm of forecasting. The research here is two-fold. The first objective is to forecast water consumption rates for the City of Phoenix and Town of Paradise Valley, Arizona with four different arrangements of single-family residential housing (Table 1). Lots for new housing units either approximate the statistical distribution of year 2000 mean lot sizes or they have smaller lots sizes than the year 2000 distribution. New lots either mirror the spatial distribution of year 2000 lot sizes or they are distributed as high (smaller lots), medium, or low density, based on their distance to transportation corridors. The second aim is to compare these four scenarios using two different modeling techniques, ordinary least squares (OLS) and geographically weighted regression (GWR). The scenarios are compared by assessing which ones identifies a management plan that reduces mean per capita water consumption and which ones reduces overall water consumption for the study area. We anticipate that the scenario with smaller lots and new residential growth occurring along transportation corridors will be the best performing scenario for goals (a) and (b). We also hypothesize that the use of local 5 parameters for GWR models would more accurately estimate water consumption than the use of global coefficients for OLS models. [Table 1] 2. Background on urban water consumption A set of complex issues arising from urban environments make water management a challenging task. Infrastructure must accommodate high demands for a diverse set of uses. Consumers exhibit various behaviors that may correlate with residential characteristics, some difficult to capture in models. Some studies shed light on urban water use through the assessment of variables with indirect effects. Grove et al (2006) measured the impacts of population density, lifestyle behavior, and social stratification on urban vegetation. They determined lifestyle behavior was the best predictor of vegetation cover on private lands. This type of relationship could be used to show the indirect effects of behavior on water consumption. Peculiar climatic conditions arise in cities where pavement and building materials retain heat for a longer period of the day than objects in the natural environment. The net result is an “island” of higher urban temperatures. Studies suggest heat island effects contribute to increased housing unit water consumption rates in Phoenix (Guhathakurta & Gober, 2007). This is relevant to our study because alternate distributions of residential density may lead to different heat island patterns that could influence single-family water consumption rates. Various resource management studies quantify factors affecting demand. Recent approaches include econometrics (Babel, Das Gupta, & Pradhan, 2007; Mylopoulos, Mentes, & Theodossiou, 2004; Reynaud, 2003), agent based modeling (Athanasiadis, 6 Mentes, Mitkas, & Mylopoulos, 2005), and fuzzy logic (Altunkaynak, Ozger, & Cakmakci, 2005). Residential water demand is often estimated as a function of housing unit characteristics. Mayer et al. (1999) studied housing unit water consumption in twelve U.S. cities. They concluded variation in indoor use was affected by household size (persons) and building square footage. Variation in outdoor consumption was most strongly affected by the presence of a pool. The addition of a pool would more than double usage compared to a home without a pool. Jacobs and Haarhoff (2004a; 2004b) created a model for residential water demand where end-uses, such as bathing, toilet flushing, gardening, and pool use are represented as parameters that can predict indoor and outdoor demand. Newer approaches to addressing water demand include fuzzy logic. This is based on fuzzy set theory where elements or objects are attributed varying degrees of membership in the set. This is counter to the traditional notion of a set where an object must have full membership or none at all. In one example of this, Altunkaynak, Ozger, and Cakmakci (2005) applied a fuzzy logic to water consumption forecasting in Istanbul, Turkey. This method predicted future monthly water consumption rates from prior water consumption amounts, which served as independent variables. Predicted values possessed an error rate of less than ten percent from actual values. Williamson, Mitchell, and McDonald (2002) captured changes in water consumption behavior for the Yorkshire Water region in the United Kingdom by reweighting spatially detailed synthetic microdata to demographic trends. Results forecasted overall domestic water increasing by 30 percent from 1991–2025 and ranging 7 from 11 percent to 46 percent among electoral districts. Bradley (2004) surveyed multiple regions for determinants of water consumption including the United Kingdom, Malaysia, South Korea, and South Africa. He concluded gains in socio-economic status would not necessarily lead to the same increase in consumption for all study areas. Instead he advocates a micro-area approach to urban water projections to account for local conditions. 3 Methods 3.1 Study area The study area includes the City of Phoenix, the economic, cultural, and political hub of Arizona and the neighboring upscale Town of Paradise Valley, Arizona. Covering approximately 500 square miles, Phoenix is the largest city in the southwestern United States, an area even larger than Los Angeles (City of Phoenix, 2008a). Paradise Valley covers only 16.5 square miles (MAG, 2008). With just over 1.5 million residents, the City of Phoenix recently surpassed Philadelphia as the fifth largest city in the nation (US Census Bureau, 2006). Phoenix averages about 3,000 persons per square mile. In contrast, Philadelphia averages over 11,000 persons per square mile (US Census Bureau, 2006). This difference has significant implications for the underlying determinants of housing unit water consumption. The lower mean density suggests that single-family residential housing units in Phoenix have larger lots than in Philadelphia. This likely translates into a higher share of outdoor water use per housing unit for Phoenicians, mostly from pools and landscaping. This is confirmed by a study suggesting 74 percent of residential water consumption in Phoenix is for outdoor uses (Mayer et al., 1999). 8 Recent efforts to revitalize downtown Phoenix and the vicinity with a stronger sense of social and economic gravity include the development of sporting venues, restaurants, and high-rise condominiums in and around Copper Square, and the ongoing construction of a light-rail system that will link downtown with Sky Harbor Airport, Tempe, and Mesa. Due to limits on developable land, new home construction in these areas will be sparse, most likely with small lots and modestly sized houses. These factors would translate into a large share of water consumption coming from indoor, rather than outdoor sources. Citywide, this area is expected to have little impact on changes in future water consumption rates. Despite these initiatives, growth in Phoenix is being propelled outward by the automobile. Major freeway expansions are underway, ensuring expanded urban footprints over the coming decades. Most future single-family residential development is expected to occur north of the Loop 101 freeway and south of downtown in the vicinity of the South Mountains (Figure 1). Residential construction is expected to continue at a steady pace, replacing desert habitats and agricultural lands. [Figure 1] While the trend in new single-family residential construction for Phoenix is smaller lots than in previous decades, it is difficult to predict new density patterns. This is due to the fact that most new growth will predominantly occur in vast stretches of land where the number of existing housing units is low. In Paradise Valley land use is mostly residential, zoned for one acre lots with one house per lot (MAG, 2008). Therefore, it is reasonable to assume the density of new development and the resulting levels of water consumption will follow historical patterns more closely than in Phoenix. 9 The vitality of Phoenix and Paradise Valley depends on the availability and access to natural resources, especially water for human consumption. The juxtaposition of a large metropolis with the arid Sonoran desert necessitates a vast system of canals and reservoirs to supply residents with water from distant sources. The population of the City of Phoenix is projected to grow to over 2.2 million by 2030, and to exceed six million in Maricopa County by 2030 (MAG, 2007b). These projections, coupled with the risk of drought and potentially dwindling water supplies, make forecasting of residential water consumption a critical need for long-term regional planning. 3.2 Data The dependent variable in this study is mean water consumption per census tract (measured in ccf) for single family residential housing for the year 2000. We obtained these data from the City of Phoenix, Water Services Department. Consumption rates for Paradise Valley are included in the dataset and accordingly are used for this study. We used four dependent variables used by Wentz and Gober (2007) to capture variation in indoor and outdoor water consumption. One dependent variable, singlefamily residential household size per census tract, for the year 2000, which we obtained from the U.S. Census Bureau, accounts for indoor usage. The intuitive notion that larger households (more people) in Phoenix typically use more water, mostly through indoor appliances and plumbing fixtures, has been confirmed by empirical studies (Mayer et al., 1999). Other independent variables from Wentz and Gober (2007) are hypothesized to capture variation in outdoor consumption. Data on lot sizes and the presence of pools 10 were acquired from the Maricopa County Assessor’s office. Only houses built through year 2000 were included in these variables. We hypothesize that lot size is positively correlated with consumption. More land increases the possibility of larger building footprints with more plumbing fixtures that might lead to greater water demand. A likely outcome of larger lots is outdoor areas being used for pools and landscaping, both requiring significant amounts of water. A spatial breakdown of mesic versus xeric landscaping practices was quantified through an expert systems land use classification, based on a 1998 Landsat TM image (Stefanov, Ramsey, & Christensen, 2001). This system produced an overall accuracy of 85 percent. Most new residential houses in Phoenix have instituted xeriscaping practices, with native Sonoran Desert vegetation that requires minimal watering. Mesic practices usually include lawns and leafy plant species not native to arid environments that require intensive irrigation to thrive in the dry, hot climate of the Valley. We queried and mapped existing and future single-family residential land uses (Figure 1) from a spatial database from the Maricopa Association of Governments (MAG, 2007a). We used a “full-growth” decision rule to account for single-family residential development for all four scenarios. At “full-growth” all potentially developable land in the City of Phoenix and Town of Paradise Valley would be converted to developed land categorized as commercial, industrial, mixed use, office, other employment, transportation, water, multi-family residential, or single-family residential. The remaining undeveloped lands, such as mountain preserves, become classified as open space. MAG classifications are based on a combination of existing land use, active, 11 planned and proposed developments, and the general plans of the member communities of MAG, including Phoenix and Paradise Valley (MAG, 2007a). 3.3 Analytical approach To reiterate we assumed single-family residential development achieves “full growth”. The first step in estimating developable land for housing units was to estimate and remove the area in the land use class ‘residential’ that is not part of the residential lots, such as roads and sidewalks. To do this we calculated the average ratio (R) of total housing unit lot area from the Maricopa County Assessor’s Office to the MAG year 2000 total single-family residential area (Equation 1). This R value is 0.67. R= AssessorLo tArea MAG _ SingleFamilyresident ialArea (Equation 1) The total area per census tract covered by new lots (A) was calculated by multiplying (R) by the total single-family residential area from the MAG land use data (Equation 2). A = R * MAG Single-family residential area (Equation 2) To forecast the number of housing units per A for each census tract, we estimated the average lot size per census tract. We used a random number generator program to create a normal (Gaussian) distribution of lot sizes (Equation 3), assigning a unique mean lot size (G) to each census tract. Randomly generated values were calculated at a 95 percent confidence level. The distribution of mean lot sizes for new housing units for scenarios 1 and 3 was created with the intent of continuing current single-family residential density levels for new development by mirroring the spatial pattern of existing 12 (year 2000) mean lot size per census tract. That is, the lot sizes were ranked from smallest to largest so that the census tract with the smallest existing mean lot size would also have the smallest mean lot size for new housing units, and so on. We calculated the log of existing mean lot sizes per census tract to overcome the problem of negative outputs in the random value generator. A short-coming of the log transformations is a smaller standard deviation (421 square meters) than the existing lot size distribution (580.22 square meters). (Equation 3) A similar approach was used to randomly generate a normal (Gaussian) distribution of lot sizes for scenarios 2 and 4. The purpose of these scenarios is to demonstrate the effects of smaller lot sizes on water consumption than lot sizes in scenarios 1 and 3. Scenarios 2 and 4 are characterized by a mean lot size per census tract of 773.97 square meters (8,330.94 square feet) and minimum lot size per census tract of 320.25 square meters (3,447 square feet). This mean lot size was selected to show a noticeable difference in lot sizes from scenarios 1 and 3, while maintaining a minimum lot size to meet local housing market expectations. A housing unit with this minimum lot size in the City of Phoenix, conforming to city zoning ordinances, could fit a one-story housing structure with as much as 101 square meters (1,084 square feet) of living space 13 or a two-story housing structure with as much as 202 square meters (2,186 square feet) of living space (City of Phoenix, 2008b). In scenarios 3 and 4 we assigned the statistical distributions of lot sizes according to decision rules setting maximum housing unit distances from transportation corridors (Table 1). High density tracts were assigned to census tracts located within a quarter-mile of the soon-to-be-completed light rail line. A quarter-mile is often used among urban planners as a measure of the maximum distance most people are willing to walk from their house to a public transit station. Medium density tracts are located within a halfmile of the light rail or one mile from any highway. All other census tracts beyond the required distances from transportation corridors are classified as low density. The densities were chosen with these decision rules to strike a balance with roughly half the census tracts located in high or medium density areas and roughly half located in low density areas. The larger objectives of transportation-oriented development are efficient use of transportation infrastructure and reducing urban sprawl. Housing unit counts (C) were forecasted at each census tract by dividing the estimated lot area (A) by the unique randomly generated lot size (G) for that tract (Equation 4). C A G (Equation 4) Overall housing unit counts at “full growth” were calculated by adding the existing (through year 2000) counts from the Maricopa County Assessor’s office to the forecasted counts (C). The overall forecasted mean lot size per census tract (F) was calculated according to Equation 5: 14 F ( ExistingLotsize * ExistingHouseholds ) ( NewLotsize * NewHouseholds ) ExistingHouseholds NewHouseholds (Equation 5) 3.31 Forecasting single-family residential water consumption As was mentioned previously, Wentz and Gober (2007) investigated the determinants of year 2000 single-family residential water consumption in Phoenix. By comparing the performances of global (OLS) and local (GWR) models, they were able to identify and analyze four variables (pool size, lot size, percent mesic landscaping, and household size) that significantly affect consumption. Our study follows that analytical approach, comparing OLS and GWR models with the independent variables from Wentz and Gober (2007). We forecasted mean housing unit consumption rates per census tract for all (existing and new) single-family residential housing units using parameters from the year 2000 estimates (Tables 2 and 3). Year 2000 data values for percent pools, household size, and percent mesic landscaping were applied to new housing units aggregated by census tract for all scenarios. Lot size is the only independent variable that we changed depending on scenario rules (Table 1). Scenario performance in this study is based on two factors (a) reducing mean per capita single-family residential water consumption and (b) reducing overall consumption for the study area. The performance of “what if” scenarios is grounded in the reliability of the historical data and models on which they are based. [Table 2] [Table 3] 15 4. Results 4.1 New lot distributions The four scenarios in this study differ according to unique combinations of statistical and spatial distributions of new lots (Table 1). Table 4 summarizes the statistical distributions of lot sizes for new single-family residential housing units. The randomly generated normal (Gaussian) distribution of values for new housing unit lots in scenarios 1 and 3 consists of a mean of 973.12 square meters, a minimum of 336.36 square meters, a maximum of 3,400.82 square meters, and a standard deviation of 421 square meters. The result is a smoothed curve, compared with the statistical distribution of existing lots, such that low and high values for new lots are lower than their existing mirrored counterparts, while middle values are higher than existing counterparts (Figure 2a). [Table 4] [Figure 2] When a different randomly generated normal (Gaussian) distribution was created for new housing unit lots in scenarios 2 and 4, the result was smaller lot sizes than scenarios 1 and 3 (Table 4). Scenarios 2 and 4 are characterized by a mean of 773.97 square meters, a minimum of 320.25 square meters, a maximum of 2,202.62 square meters, and a standard deviation of 284.51 square meters. Results show a smoothed distribution curve for scenarios 2 and 4, compared with the distribution of existing lot sizes (Figure 2b). In this case, the new lot sizes are all smaller than their mirrored counterparts from the existing distribution. 16 Mean lot sizes for new housing units mirror the spatial pattern of existing housing unit mean lot sizes (scenarios 1 and 2) or they mirror the spatial pattern of existing housing unit lot sizes within high, medium, or low density classifications based on proximity of census tracts to transportation corridors (scenarios 3 and 4) (Figure 3). This classification scheme produced six high density census tracts, 152 medium density census tracts, and 144 low density census tracts. [Figure 3] The addition of new lots changed the statistical distribution of mean lot sizes from existing levels. Scenario 1 consists of a mean of 954.43 square meters, a minimum of 448.16 square meters, a maximum of 4,318.60 square meters, and a standard deviation of 487.82 square meters. Scenario 2 is characterized by a mean of 910.82 square meters, a minimum of 438.87 square meters, a maximum of 3,932.07 square meters, and a standard deviation of 428.13 square meters. Scenario 3 produced a mean of 949.60 square meters, a minimum of 483.40 square meters, a maximum of 4,318.60 square meters, and a standard deviation of 481.07 square meters. Scenario 4 consists of a mean of 906.58 square meters, a minimum of 464.11 square meters, a maximum of 3,932.07 square meters, and a standard deviation of 422.53 square meters. 4.2 Single-family residential housing unit distribution We calculated the distribution of single-family housing units by dividing the total area of new residential land use allocated for lots by the mean lot size for a given census tract. New housing unit counts were combined with existing housing unit counts to produce forecasted total housing unit count per census tract (Table 5). The greatest 17 number of total housing units was generated by scenario 2 (538,844), followed by scenario 4 (535,060), scenario 1 (485,954), and scenario 3 (483,681). The scenarios show little difference in the spatial distribution of housing units. This is due to the fact that most new housing units are forecasted in a handful of large census tracts in the northern and southwestern Phoenix, most notably tract 303.42 in the far northwest corner of the city (Figure 4). Centrally located tracts would experience only slight changes in housing unit patterns. [Table 5] [Figure 4] 4.3 Water usage There was distinct spatial variation in year 2000 single-family residential housing unit water consumption rates for Phoenix and Paradise Valley (Wentz & Gober, 2007). The highest mean housing unit rates were predominantly located in wealthier areas in Paradise Valley. GWR estimates by Wentz and Gober (2007) suggest increased lot sizes for those residents would increase water consumption. The lowest average housing unit consumption rates were found in southwestern and northern Phoenix. Wentz and Gober (2007) determined that lot size had a slightly positive, or in some cases negative, relationship with water consumption in those areas. The addition of new housing units at “full-growth” would nearly double overall consumption from year 2000 levels (Table 6). A comparison of estimated water usage among scenarios reveals differences in the contribution of new housing units to total consumption (Table 6). The highest total consumption for the study area is Scenario 2 (OLS) at 122,352,001 ccf, while the lowest total consumption is Scenario 3 (GWR) at 18 105,525,043 ccf. For all scenarios OLS estimated higher total consumption than GWR. Differences between OLS and GWR range from 8,934,986 ccf for Scenario 3 to 6,624,213 ccf for Scenario 2. [Table 6] At 1,633,921 ccf, residents of census tract 303.42 consumed the most water of any tract in year 2000 (Table 6). In all forecasts residents of tract 303.42 also generated the highest overall consumption (Table 6). Forecasted consumption for the tract ranges from 15,819,660 ccf for scenario 2 OLS (almost a ten-fold increase from year 2000) to 9,939,277 ccf for scenario 3 GWR (about a six-fold increase from year 2000). Changes in the mean lot sizes and spatial distribution of new housing lots among the scenarios resulted in departures from the mean household unit consumption for year 2000 of 222.05 ccf. Mean housing unit consumption ranged from 214.77 ccf in Scenario 2 (GWR) to 236.64 ccf in Scenario 3 (OLS). The spatial distributions of mean housing unit water consumption for GWR and OLS models are illustrated in Figures 5 and 6. The scenarios generally retained the spatial pattern of mean year 2000 housing unit consumption, where residents of Paradise Valley consumed the highest average levels of water and residents in scattered locations throughout central Phoenix consumed lower amounts of water. [Figure 5] [Figure 6] 4.4 Scenario comparisons Comparisons of scenarios are organized based on shared spatial and statistical distributions of lot sizes. To account for the effects of transportation corridors on water 19 consumption, we subtracted mean housing unit water consumption per census tract for Scenario 1 (mirrors existing lot sizes and pattern) from Scenario 3 (mirrors existing lot sizes with transportation corridors) (Figures 7a and 7b). To account for the effects of smaller lot sizes on water consumption, we subtracted mean housing unit consumption for scenario 1 (mirrors existing lot sizes and pattern) from Scenario 2 (smaller lot sizes, mirrors existing lot size pattern) (Figures 7c and 7d). Mean housing unit water consumption rates are lower along transportation corridors in central Phoenix for Scenario 3 compared with the mirrored pattern of lot sizes in Scenario 1. Lower density developments with larger lots are located on the fringes of Phoenix, such as census tract 303.42. The mixed pattern of consumption produced little change in overall consumption between Scenarios 1 and 3 (Table 6). This indicates that transportation corridors would not be an effective tool in reducing water consumption. The introduction of smaller lots led to lower mean water consumption for most census tracts (Figures 7c and 7d). This is most pronounced in Paradise Valley for the GWR model. In contrast, lot size for the GWR model appears to have a minor or even negative relationship with mean housing unit consumption in southern Phoenix. [Figure 7] 5. Discussion 5.1 Patterns of lot distribution Changing the mean lot size and spatial distribution of lot sizes impacted patterns of lot distribution. Regardless of the introduction of smaller lots or high density transportation corridors, only subtle shifts in lot counts appear in central Phoenix where 20 there is little room for new single-family residential growth. In north Phoenix and, to a lesser degree, southwest Phoenix where large areas of land are available for development, reducing mean lot size (scenarios 2 and 4) had the effect of increasing the overall number of housing units for the study area, while reducing mean consumption per housing unit. With smaller lot sizes scenario 2 produced 52,890 more new housing units than scenario 1, an increase of almost 11 percent. Scenario 4 generated 51,379 more new housing units than scenario 3, an increase of almost 11 percent. In this instance, 37,679 new lots are from tract 303.42. This lesser role is due to the inclusion of higher densities along transportation corridors that led to larger lots being assigned to tract 303.42. With higher density development along corridors, Scenario 3 generated 2,273 fewer housing units than Scenario 1 where lot sizes mirrored the pattern of existing housing units. Scenario 4 generated 3,784 fewer lots than scenario 2. The slight reduction is number of households in these cases suggests transportation corridors would only have a marginal influence on urban growth rates. Given the small amount of land designated for single-family residential development in the vicinity of the light rail corridor (Figure 1), the area had little effect on lot distributions in any of the scenarios. In addition, there is limited land for new residential development along the I-17 and SR 51 freeways in central Phoenix (Figure 1). 5.2 Water consumption patterns “Smart growth” promotes higher density development in the urban core, along with transit-oriented development, as part of a larger objective of limiting excessive urban sprawl. Ideally, this would reduce the impact of development on the natural 21 environment, while reducing the overall demand for single-family residential water use. In our study higher density development is represented, in part, through smaller lot sizes (scenarios 2 and 4) and also through higher density development along transportation corridors (scenarios 3 and 4). Forecasts reveal smaller lots played a significant role in reducing per capita water consumption by allowing less area for outdoor sources of consumption, including landscaped vegetation and pools. They also reduce the size of building footprints, possibly restricting indoor water consumption. Forecasts indicate higher density development along transportation corridors had a minimal effect on consumption rates. The existing pattern of lot sizes mirrored for new development in scenarios 1 and 2 is already more dense in central Phoenix where most transportation routes are located. At 214.77 ccf per housing unit scenario 2 (GWR) was the most effective scenario with respect to goal (a): reducing mean per capita water consumption for the study area. Mirroring the existing pattern of lot sizes with higher density along transportation corridors, scenario 3 was the worst performing scenario with a mean housing unit consumption of 236.64 ccf. With respect to goal (b): reducing total water consumption for the study area, scenario 3 (GWR) was the best performing scenario with 105,525,043 ccf of consumption. Scenario 2 OLS was the worst performing scenario at 122,352,001 ccf of water consumption. The main take-home message of this study is that planners and policy-makers should aim to implement rules that promote small lots for new single-family residential housing to reduce per capita water consumption. This approach is reflected in scenario 2. As the population of the Phoenix metropolitan area increases, this plan will accommodate 22 growth, while minimizing water consumption. With large expanses of land available for development throughout Maricopa County, it would be reasonable to steer more households on smaller lots to centrally located Phoenix and Paradise Valley, limiting excessive urban sprawl and reducing per capita water consumption. 5.3 Local versus global model performances The findings of Wentz and Gober (2007) from modeling OLS and GWR on year 2000 water consumption is a starting point for analysis of water forecasting in this study. Wentz and Gober determined that GWR generally performed better than OLS with a mean r2 of .85 versus an r2 of .64 for OLS. In our study GWR estimates for housing unit water consumption in all four scenarios were lower than OLS estimates. The difference in r2 from Wentz and Gober (2007) suggests GWR is an improvement over OLS. But without knowing how the relationships between the variables change, this is not conclusive. Any model estimating real world phenomena possesses inherent limitations, including limited or uncertain data and mathematical constraints. A weakness of GWR is that as lot parameters change, the local model should also change – assuming stationarity in the parameters. Unfortunately, there is no known mathematical approach to deal with this problem. 5.4 Strengths and weaknesses of the modeling approach The modeling approach was effective in several ways. While OLS performed well in estimating year 2000 water consumption rates, GWR was a substantial improvement 23 over OLS (Wentz & Gober, 2007). The inclusion of existing housing units in the global and local forecasts tempers the uncertainty of estimating new housing unit distributions. Other forms of land use are frequently converted to residential use, but rarely does residential convert to another use. Exceptions such as eminent domain of houses in new freeway corridors or redevelopment of older urban neighborhoods should not be overlooked. Overall, it is safe to assume the general spatial pattern of existing housing units remain intact in the foreseeable future. What are more likely to change are residential housing unit characteristics due to demographic, policy, or planning changes. These unforeseen changes could alter future water consumption rates and patterns. The scenario rules (Table 1) used statistical and spatial distributions of lots sizes that translated into different rates and patterns of water consumption. The basic message from this is that planning or policy alternatives impact residential water consumption. The scenarios developed in this research could be translated into rudimentary planning regimes or policy alternatives with long-term planning and policy implications for singlefamily residential water consumption. Lot sizes would, perhaps, be restricted through a planning measure where zoning ordinances set maximum lot sizes. Not surprisingly, affluent areas will likely continue to consume water at a higher rate than less affluent areas. Lessening this disparity might be accomplished through pricing controls, by changing social attitudes towards owning pools, changing landscaping practices, or implementing conservation measures. Policies to reduce water consumption should account for variations in patterns and local relationships between housing units. Due to the scarcity of available land for new single-family residential development along the light rail line, policies to encourage such development would have 24 little direct influence on consumption rates. Any future plans to expand the light rail to the fringes of the urban footprint might have a more direct impact on consumption rates. Lempert, Popper, & Bankes (2003) offer a critique of forecasting with a small number of scenarios pointing out that some new modeling and forecasting approaches generate hundreds or more scenario alternatives or iterations. One approach to forecasting seeks more effective ways to understand how the future will unfold based on measurable conditions. Another approach explores and estimates the conditions or factors that would lead to a particular outcome. We acknowledge our approach explores consumption given certain policy restraints, showing only a small subset of the potential range of scenarios. A fuller picture of the range of possible effects of urban growth and housing characteristics could be achieved through a sensitivity analysis. Another issue with this study is that year 2000 census tract boundaries stay the same size as residential growth continues. In fact, large tracks in north and southwest Phoenix, such as tract 303.42, would be divided as population levels increase. Interstate 17 runs through tract 303.42. With smaller census tracts and some tracts along this transportation corridor could shift to higher density lots, ultimately increasing consumption in those areas. These divided tracts would not necessarily grow with the same density patterns, or take on the same housing unit characteristics, such as landscaping practices or pools. Some new subdivisions might be more family-oriented with pools and larger lots; others sold as retirement communities with smaller lots and small household sizes. Lots sold on an individual basis might be less predictable in their housing unit characteristics than lots in a subdivision. 25 Growth in Phoenix will include annexations of nearby lands. This will occur predominantly on the northern edges of the city. Gila River Indian Community lands restrict southward expansion and other municipalities restrict eastward or westward expansion. Little change is expected for the boundaries of Paradise Valley. Uncertainty about future land annexations limits our forecasts. Further research could apply other local models or simulation approaches to estimate the housing unit characteristics and water consumption in these areas. 6. Conclusions and future research Our study has shown that there is the potential for per capita water savings with smaller lot sizes, which has the effect of reducing outdoor water consumption. The distribution of these lots, however, does little to influence either overall water consumption or per capita consumption. We demonstrated that for Phoenix and Paradise Valley, there is the potential for only a small amount of per capita water consumption savings due to high density development along transportation corridors. This is mostly due to the lack of open land for development in these areas. The use of GWR shows how alternate planning or policy regimes may lead to spatial variation in the patterns and levels of single-family residential water consumption. More specifically, it reveals how particular housing unit characteristics translate into higher or lower consumption rates locally. The hope is that local scale modeling will be developed and applied in additional water forecasting tools and other social-science and environmental forecasting studies. Alternate simulation methods, like agent based modeling, also offer promise in spatial forecasting. Residential data values or coefficients 26 from Wentz and Gober (2007) could be incorporated into a simulation program such as UrbanSim. This research evaluates only a small portion of possible “what if” scenarios for urban water management in Phoenix and Paradise Valley. Questions about the long-term water supply are equally as important as demand issues. The desert southwest is the fastest growing part of the United States; it is also the region with the most precarious balance between human needs and water resources. Phoenix shares upstream Colorado River water with Las Vegas and southern California. Forecasting the likely range of future supplies against the backdrop of climatic uncertainty is a vital ingredient for effective political agreements. There are several avenues for additional research. One obvious approach would modify the existing scenarios, perhaps to represent the effects of other planning or policy measures on water consumption. Another variation would be to generate a non-Gaussian distribution for the lot sizes. This new distribution might be skewed to more accurately reflect the distribution of existing lot sizes. Aside from modifications to the existing scenarios, our research question could be expanded to include more than four scenarios. An obvious starting point is to model the other parameters used in water estimation from Wentz and Gober (2007), including percent of homes with a pool, household size (persons), and percent mesic landscaping. Household size would be a good starting point as it exhibited spatial non-stationarity according to the results of Wentz and Gober (2007). Other local characteristics, such as the urban heat island effect, have been modeled for Phoenix (Guhathakurta & Gober, 2007). The inclusion of this measure 27 would be a good way of factoring in the effects of climatic conditions on water consumption rates. Beyond modifying statistical distributions, new research could change the spatial distribution of new lot sizes. For example, the maximum distance for density classifications along transportation corridors could be altered. This study applied one quarter mile buffers for the high density classification along the light rail line, with only six qualifying census tracts. A larger buffer would lead to a more pronounced high density core. Likewise, changing the corridor size around highways might have effects on the spatial patterns of consumption throughout Phoenix. Another useful modification for this study would make the number of new housing units constant for all scenarios where population could be tied to some future projection. Multiple categories of residential development might be included in the forecasts. The inclusion of multifamily housing units would allow larger lots to remain of the fringes with more housing units in smaller lots or multistory buildings along transportation corridors. 28 Figure captions 1. Existing and new single-family residential land use at “full-growth” in the City of Phoenix and the Town of Paradise Valley, Arizona. Data source: MAG 2007b. 2. Forecasted mean lot size per census tract: (a) Scenarios 1 and 3, (b) Scenarios 2 and 4. 3. Single-family residential lot density classifications based on proximity to highways and the light rail line. 4. Forecasted total (existing and new) single-family residential housing units per census tract (a) Scenario 1, (b) Scenario 2, (c) Scenario 3, (d) Scenario 4. 5. Mean annual single-family residential housing unit water consumption (ccf), OLS model: (a) Scenario 1, (b) Scenario 2, (c) Scenario 3, (d) Scenario 4. 6. Mean annual single-family residential housing unit water consumption (ccf), GWR model: (a) Scenario 1, (b) Scenario 2, (c) Scenario 3, (d) Scenario 4. 7. Comparisons of mean housing unit water consumption: Scenario 3 minus Scenario 1 (a) OLS model, (b) GWR model, Scenario 2 minus Scenario 1 (c) OLS model, (d) GWR. 29 Table Captions 1. Four scenarios for future single-family residential development in the City of Phoenix, Arizona and the Town of Paradise Valley, Arizona. Each scenario combines one statistical distribution with one spatial distribution. 2. Ordinary least squares (OLS) regression coefficients with mean single-family residential housing unit water consumption per census tract as the dependent variable (measurement units reported in liters). 3. Geographically weighted regression (GWR) coefficients with mean single-family residential housing unit water consumption per census tract as the dependent variable (measurement units reported in liters). 4. Summary of statistical distributions of randomly generated new lots. 5. Forecasted single-family residential housing unit counts, City of Phoenix and Town of Paradise Valley, Arizona. 6. Summary statistics of forecasted annual water consumption (ccf) for the City of Phoenix and the Town of Paradise Valley, Arizona. 30 Figures Figure 1 31 Figure 2 32 Figure 3 33 Figure 4 34 Figure 5 35 Figure 6 36 Figure 7 37 Tables Table 1 Table 2 Table 3 38 Table 4 Table 5 Table 6 39 References ADES. (2006). Maricopa County Population Projections 2006 – 2055. Retrieved June 15, 2008, from https://www.azdes.gov/ Altunkaynak, A., Ozger, M., & Cakmakci, M. (2005). Water consumption prediction of Istanbul City by using fuzzy logic approach. Water Resources Management, 19(5), 641-654. Arbues, F., Barberan, R., & Villanua, I. (2004). Price impact on urban residential water demand: A dynamic panel data approach. Water Resources Research, 40(11). Athanasiadis, I. N., Mentes, A. K., Mitkas, P. A., & Mylopoulos, Y. A. (2005). A hybrid agent-based model for estimating residential water demand. SimulationTransactions of the Society for Modeling and Simulation International, 81(3), 175-187. Babel, M. S., Das Gupta, A., & Pradhan, P. (2007). A multivariate econometric approach for domestic water demand modeling: An application to Kathmandu, Nepal. Water Resources Management, 21(3), 573-589. Bradley, R. M. (2004). Forecasting domestic water use in rapidly urbanizing areas in Asia. Journal of Environmental Engineering-Asce, 130(4), 465-471. Briggs, D. J., Collins, S., Elliott, P., Fischer, P., Kingham, S., Lebret, E., et al. (1997). Mapping urban air pollution using GIS: a regression-based approach. International Journal of Geographical Information Science, 11(7), 699-718. City of Phoenix. (2008a). Community Trends and Profile. Retrieved June 15, 2008, from http://phoenix.gov/ City of Phoenix. (2008b). Zoning Ordinances of the City of Phoenix, Chapter 6, Section 613, R1-6 Single-Family Residence District. Retrieved June 15, 2008, from http://www.municode.com/Resources/gateway.asp?pid=13534&sid=3 Fotheringham, A. S., Brunsdon, C., & Charlton, M. (2002). Geographically weighted regression: the analysis of spatially varying relationships. New York: Wiley. Grove, J. M., Troy, A. R., O'Neil-Dunne, J. P. M., Burch, W. R., Cadenasso, M. L., & Pickett, S. T. A. (2006). Characterization of households and its implications for the vegetation of urban ecosystems. Ecosystems, 9(4), 578-597. Guhathakurta, S., & Gober, P. (2007). The impact of the phoenix urban heat island on residential water use. Journal of the American Planning Association, 73(3), 317329. Hyndman, R. J. (2007). Topics Covered in the IJF. Retrieved February 5, 2007, from http://www.forecasters.org/ijf/scope.html Jacobs, H. E., & Haarhoff, J. (2004a). Application of a residential end-use model for estimating cold and hot water demand, wastewater flow and salinity. Water Sa, 30(3), 305-316. Jacobs, H. E., & Haarhoff, J. (2004b). Structure and data requirements of an end-use model for residential water demand and return flow. Water Sa, 30(3), 293-304. Kenney, D. S., Goemans, C., Klein, R., Lowrey, J., & Reidy, K. (2008). Residential water demand management: Lessons from Aurora, Colorado. Journal of the American Water Resources Association, 44(1), 192-207. 40 Lee, S. J., Balling, R., & Gober, P. (2008). Bayesian Maximum Entropy mapping and the soft data problem in urban climate research. Annals of the Association of American Geographers, 98(2), 309-322. Lempert, R. J., Popper, S. W., & Bankes, S. C. (2003). Shaping the next one hundred years : new methods for quantitative, long-term policy analysis and bibliography. Santa Monica, CA: RAND. MAG. (2007a). MAG Future Land Use Database, Maricopa County. MAG. (2007b). Socioeconomic Projections of Population, Housing and Employment by Municipal Planning Area and Regional Analysis Zone. Retrieved June 15, 2008, from http://www.workforce.az.gov/ MAG. (2008). Member: Town of Paradise Valley. Retrieved June 15, 2008, from http://www.mag.maricopa.gov/member.cms?item=567 Mayer, P. W., DeOreo, W. B., Opitz, E., Kiefer, J., Dziegielewski, B., Davis, W., et al. (1999). Residential End Uses of Water. Englewood, CO: American Water Works Association Research Foundation. Michelsen, A. M., McGuckin, J. T., & Stumpf, D. (1999). Nonprice water conservation programs as a demand management tool. Journal of the American Water Resources Association, 35(3), 593-602. Mylopoulos, Y. A., Mentes, A. K., & Theodossiou, L. (2004). Modeling residential water demand using household data: A Cubic approach. Water International, 29(1), 105-113. Olmstead, S. M., Hanemann, W. M., & Stavins, R. N. (2007). Water demand under alternative price structures. Journal of Environmental Economics and Management, 54(2), 181-198. Reynaud, A. (2003). An econometric estimation of industrial water demand in France. Environmental & Resource Economics, 25(2), 213-232. Stefanov, W. L., Ramsey, M. S., & Christensen, P. R. (2001). Monitoring urban land cover change: An expert system approach to land cover classification of semiarid to arid urban centers. Remote Sensing of Environment, 77(2), 173-185. US Census Bureau. (2006). 2006 Population Estimates. Retrieved June 15, 2008, from http://www.census.gov/ Vickers, A. (1993). The Energy-Policy Act - Assessing Its Impact on Utilities. Journal American Water Works Association, 85(8), 56-62. Wentz, E. A., & Gober, P. (2007). Determinants of small-area water consumption for the city of Phoenix, Arizona. Water Resources Management, 21(11), 1849-1863. Williamson, P., Mitchell, G., & McDonald, A. T. (2002). Domestic water demand forecasting: A static microsimulation approach. Journal of the Chartered Institution of Water and Environmental Management, 16(4), 243-248. Xiang, W. N., & Clarke, K. C. (2003). The use of scenarios in land-use planning. Environment and Planning B-Planning & Design, 30(6), 885-909. Yu, D. (2007). Modeling owner-occupied single-family house values in the city of milwaukee: A geographically weighted regression approach. Giscience & Remote Sensing, 44(3), 267-282. 41