Policy Research Working Paper 8345 Water When It Counts Reducing Scarcity through Irrigation Monitoring in Central Mozambique Paul Christian Florence Kondylis Valerie Mueller Astrid Zwager Tobias Siegfried Development Research Group A verified reproducibility package for this paper is Impact Evaluation Team available at http://reproducibility.worldbank.org, February 2018 click here for direct access. Policy Research Working Paper 8345 Abstract Management of common-pool resources in the absence of requirements for common crops, and ii) individualized individual pricing can lead to suboptimal allocation. In the information, comparing water requirements with each farm- context of irrigation schemes, this can create water scarcity er’s water use in the same season of the previous year. Both even when there is sufficient water to meet the total require- types of feedback tools lead to higher reported and observed ments. High-frequency data from three irrigation schemes sufficiency of water relative to recommendations, and nearly in Mozambique reveal patterns consistent with inefficiency eliminate reports of conflicts over water. The experiment in allocations. A randomized control trial compares two fails to detect an additional effect of individualized compar- feedback tools: i) general information, charting the water ative feedback relative to a general information treatment. This paper is a product of the Impact Evaluation Team, Development Research Group. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The authors may be contacted at pchristian@worldbank.org, fkondylis@worldbank.org, and vmuelle1@asu.edu. A verified reproducibility package for this paper is available at http://reproducibility.worldbank.org, click here for direct access. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Water When It Counts: Reducing Scarcity through Irrigation Monitoring in Central Mozambique By PAUL C HRISTIAN F LORENCE KONDYLIS VALERIE M UELLER A STRID Z WAGER T OBIAS S IEGFRIED* Keywords: water, irrigation, information, experiment, productivity, Africa JEL Codes: O13, Q15, Q16, Q25, D83 I. Introduction Irrigation systems are a common pool resource. When farmers fail to internalize the cost of their use on others, the allocation of water within the irrigation scheme may not only be inefficient but also affect water availability in the long term. Collective action over the management of a scheme is then necessary to ensure the sustainability of irriga- tion supply. Decentralized models of governance have evolved to formalize the delivery of water resources (Ostrom, 1990; Ostrom and Schlager, 1992). Yet, local institutions like water user associations may not yield efficient resource sharing if the main constraint users face is limited knowledge of water management (Plusquellec, 2002; Meinzen-Dick, 2007). * Author affiliations are: Christian: World Bank; Kondylis: World Bank; Mueller: Arizona State University, International Food Policy Research Institute; Zwager: World Bank; Siegfried: Hydrosolutions, Ltd. Correspond- ing authors are: Christian, pchristian@worldbank.org; Kondylis, fkondylis@worldbank.org; and Mueller, Valerie.Mueller.1@asu.edu. Luiza Cardoso de Andrade provided superb research assistance. We thank Mario Chilundo, Steven Glover, Robert Naudascher, and Euclides Tiago for excellent field assistance. We appreciate valuable comments received from Dawit Mekonnen and seminar audiences at IFPRI and the AAEA meetings. The team is grateful for the support of Pedro Arlindo, Mark Austin, Aniceto Bila, Mark Lundell, Paiva Munguambe, Eugenio Nhone, and Antonio Samuel. This project benefited from funding provided by the USAID mission and World Bank country offices in Mozambique, the CGIAR Research Program on Policies, Institutions, and Markets (PIM) led by the International Food Policy Research Institute (IFPRI), as well as the i2i trust fund financed by UKAID. 1 2 WATER WHEN IT COUNTS To understand the origins of irrigation management problems in central Mozambique, we designed an intensive user-based water monitoring protocol. Plot-level water use data was collected three times per day for 146 consecutive weeks covering 159 plots in three irrigation schemes. We paired this high-frequency time series with household surveys on agricultural practices and production to understand sources of inefficiency in water allocation in a region suffering from low agricultural productivity and periodic droughts. This data system adds to previous demonstrations of the feasibility of user-based systems for monitoring water management (Fraternali et al., 2012; Lowry and Fienen, 2013; Buy- taert and et al., 2014; Little, Hayashi and Liang, 2016). Two stylized facts emerge from the monitoring phase of our study. First, water alloca- tions are inefficient in a context lacking penalties for overuse. The recommended water requirements of crops on 20-40% of the plots were not met in most weeks over the course of one year. This occurs despite all three schemes having sufficient water to meet each crop’s weekly water requirement in the same weeks. Scarcity in this context is, therefore, purely an allocation problem. If users could be induced to more efficiently allocate water, scarcity could be eliminated under a water neutral manner. Second, we observe that wa- ter scarcity arises from a particular allocation pattern. Although water requirements vary widely across the crop cycle, water applications are inflexible. As farmers have staggered planting times within a given scheme, this inflexible water application fails to take ad- vantage of efficiency gains available through a simple reallocation of water across plots and planting stages. These stylized facts are consistent with a model of behavior where farmers rely on inefficient rules of thumb. They choose a fixed level of water to deliver to their plots upon planting and stick to this rule until harvesting. In such a model, scarcity can be WATER WHEN IT COUNTS 3 reduced by simple re-allocations. If farmers can be convinced to use smaller quantities of water during less water-intensive growth stages, more water is made available for farmers whose crops are in water-intensive stages without making anyone worse off. In this setting, we ask whether informing farmers of their water requirements can yield efficiency gains and reduce observed scarcity and conflicts over water. A growing liter- ature examines supply-side constraints in agricultural extension services (Beaman et al., 2015; Cole and Fernando, 2016; Kondylis, Mueller and Zhu, 2017; BenYishay and Mo- barak, Forthcoming), as well as demand-side constraints that may affect a farmer’s re- sponsiveness to an extension message (Aker, 2011; Beaman et al., 2015; Aker, Ghosh and Burrell, 2016). In particular, evidence on the role of tailored vs general information in changing adoption behavior emphasizes a need to balance two trade-offs. On the one hand, recommendations that are tailored to each user may have the greatest impact. Recent evidence from utilities in developed countries demonstrates that deliver- ing information to individuals on their consumption relative to others can significantly enhance conservation behavior (Ferraro and Price, 2013; Alcott and Rogers, 2014). In a context of low productivity agriculture, Hanna, Mullainathan and Schwartzstein (2014) show that due to the difficulty of learning about optimal strategies in agricultural en- vironments, seaweed farmers in Indonesia only respond to extension information when that information includes feedback on their own practices. This would suggest that, in the context of irrigation, providing water users with metered information on their usage relative to the optimal water requirement for their crop at each growth cycle may be necessary to curb inefficiencies in water use. However, the cost of tailoring recommendations to observations of individual behav- 0 This is echoed in other settings (Aker, Ghosh and Burrell, 2016; Deichmann, Goyal and Mishra, 2016). 4 WATER WHEN IT COUNTS ior can be substantial. In our context, tailored recommendations rely on the collection of high-frequency behavioral and water allocation data. Technical expertise is required to translate observed behavior into recommendations. Both factors may preclude the scala- bility of this type of feedback mechanism as an extension modality. In contrast, highlight- ing the relative water requirements over the life cycle of primary crops has the benefit of reaching a broader base of beneficiaries, since the main cost of producing the information is deciphering the primary crops grown in the region. The competing demands of scalability and direct relevance for individual farmers cre- ate conflicting pressures on agricultural extension systems. Information disseminated through public extension must be cheap enough to generate that it can be scaled up and made available to farmers across remote areas. Yet, if generalized information is not be- ing used by farmers for lack of relevance, providing general information alone may not be a panacea. To test these ideas, we run a field experiment in which we assess whether individual feedback on usage of irrigation water is necessary to influence behavior beyond general, lower cost reminders. In practice, we examine the role of (1) providing farmers with general recommendations on water requirements, and (2) juxtaposing individual use rel- ative to the general recommendations in curbing the inefficient water allocation patterns observed at baseline. Use of high-frequency water data permits an event-study design to quantify the effect of any feedback on water sufficiency. Within this event-study de- sign, we randomize a subset of farmers to receive individualized feedback to assess any additional benefit of tailoring recommendations on water allocations. Two months after feedback provision, we observe a marked drop in both water-related conflicts and plot-level water scarcity. Specifically, the share of plots receiving insuffi- WATER WHEN IT COUNTS 5 cient water relative to their crop requirements decreased by 14.0 percentage points (from a pre-information weekly average of 27.48% of plots). High-frequency data on plot-level water availability suggests that these efficiency gains are achieved as farmers dynami- cally adjust their use to closely fit the water requirements over the crop cycle. Farmers who were randomized into receiving additional, individualized information on water allocations over the crop life cycle were no more likely to change their be- havior. Taken together, the findings suggest that providing relatively cheap information on general water requirements leads to water-conserving practices, with no additional benefits to providing more expensive individual feedback. The remainder of the paper is organized as follows. Section II summarizes the context, data innovations, and establishes stylized facts from the one-year water monitoring phase of the study. Section III provides a theoretical framework for inefficiency arising from rules of thumb motivated by the baseline stylized facts. Section IV describes a field experiment that tests the impact of generalized vs. individualized feedback tools on the efficiency of water allocation, while Section V outlines the empirical methodology and main experimental results. The final section concludes. II. Irrigation water use in Central Mozambique A. Study background Mozambique’s population primarily lives off low-productivity agriculture: 80% of workers are engaged in the agricultural sector, while its contribution to GDP is merely 25% (World Bank, 2011). In an effort to transform the agricultural sector, government and aid resources have been designated to support diversification of production into hor- ticulture and other export-oriented crops. One of the main challenges farmers face is 6 WATER WHEN IT COUNTS the reliance on rainfall for production. Sixty to eighty percent of annual precipitation falls during the region’s single rainy season, meaning that farming is not viable during most of the year (World Bank, 2007). In recognition of this constraint, the Government of Mozambique (GoM) has dramatically increased its investment in the water sector, spending 2.3 billion USD in 2008 (World Bank, 2010). However, given the high incidence of droughts in recent years, any planned expansion of irrigation risks exacerbating water stress. Among African countries, Mozambique is highly vulnerable to natural disasters and extreme weather (GFDRR, 2017), including both droughts and floods. The water scarcity created by droughts has been estimated to impose a drag on GDP growth of 1.1% on average (World Bank, 2007). In 2016, a drought-induced water scarcity led to food shortages affecting an estimated 400,000-1.8 million people (Vidal, 2016). These periods of extreme water scarcity imply that water overuse under ‘normal’ climate conditions may come at a high cost during times of water scarcity. Improving the performance and sustainability of irrigation schemes requires overcom- ing weak management by local institutions. Despite substantial investment, Mozam- bique’s coverage of irrigation has actually declined in recent decades. At the start of the 1980s, 120,000 hectares were irrigated, and 50,000 hectares have since been rehabilitated (FAO, 2016). These figures suggest a large margin for improvement in the sustainability of irrigation schemes. We conducted a study of three irrigation sites in Central Mozambique to i) understand whether water is currently used efficiently in a context of weak local oversight, and ii) to test strategies for alleviating these inefficiencies. Two of these schemes consist of canal irrigation, while a third, employs both canals and sprinklers. Water for all three schemes WATER WHEN IT COUNTS 7 is sourced from rivers or springs. This type of gravity-fed surface irrigation is the most common type of irrigation in Mozambique (FAO, 2016). All three schemes were re- habilitated as part of a World Bank-supported project called the Sustainable Irrigation Project (hereafter, PROIRRI), and are primarily devoted to production of horticultural crops through a contract farming arrangement with a local buyer. The PROIRRI project focused on horticultural production given its perceived commercial value. The crop port- folios in these schemes are representative of those that will be practiced under future irrigation expansions. Farmers benefiting from schemes constructed under direct support from the GoM are required to organize themselves into water users associations. Water pricing has not been established to regulate use. Given the lack of incentives to monitor or conserve water in this context, it was not previously known whether farmers waste water and whether there are potential gains from reallocating the quantities of water currently used. Collecting water use data at high temporal and spatial resolution is therefore essential to shed light on the pattern of water use and formulate viable policy actions to improve the efficiency of water allocation. B. Baseline data on water use At the heart of our contribution is a high temporal and spatial resolution water mea- surement system, which tracks weekly water availability at the plot level. We combined these with agricultural production data collected every four months. Water measurement and production surveys were collected between August, 2015 and July, 2017. We now describe the sampling strategy and detail the data collection protocols. Study Sample 8 WATER WHEN IT COUNTS To set up the measurement system, we fully mapped all plots within the three study schemes. During this exercise, we identified all plots that draw water from sprinklers, canals, or hydrants connected to the scheme; the source of water for these plots; and information on the household of the person in charge of cultivating them. During the initial contact with the water users, we identified 153 plots belonging to 120 households across the three schemes. Irrigation water availability We employ a low-cost, community-centered approach to monitor water supply at the plot level. As part of the initial scheme listing, all water diversion points and their water catchment areas were mapped, and water measurement points were identified and cali- brated. This calibration allowed us to convert simple measurements, such as water depth in a canal (measured with a ruler) or number of sprinklers on a plot, into volume of water diverted to the plots from that point. Each scheme was assigned a dedicated data collector who took the established measurements at each point three times a day.1 These measures were then converted to flows through canals, sprinklers, and hydrants. From these, we estimated the volume of water delivered to each plot in a week. We obtain our measure of water availability by dividing that volume by the area of the plot. Water availability is therefore reported as a a depth measured in mm/week.2 , 3 Agriculture surveys We additionally conduct household surveys every four months to capture detailed 1 These data collectors were farmers from the schemes who received training and a nominal payment to collect measurements. 2 This depth is interpretable as the maximum depth of water that could cover a plot if the full amount of water delivered in that week is distributed equally across the area of a cultivated plot. 3 A detailed description of the monitoring methods and imputation protocols for translating measurements into es- timates of volumetric flows is included in appendix A.A2. A full description of measurement methods being developed and tested around the world by technical partners contributing to the study is available at: http://www.imomohub.org/. WATER WHEN IT COUNTS 9 information on agricultural production and irrigation practices in the study area. Each household is asked about the plots they cultivate in the scheme, as per the listing.4 Table 1 shows the number of plots and households surveyed at every round. 5 TABLE 1—N UMBER OF O BSERVATIONS PER ROUND Round 1 2 3 4 5 6 7 Number of listed households 72 120 121 123 124 135 138 Number of listed plots 90 153 156 158 158 158 157 Average number of plots farmed per household 1.25 1.27 1.34 1.39 1.45 1.41 1.43 Share of listed households surveyed 90.3% 86.7% 91.7% 84.6% 86.3% 86.7% 74.6% Notes: Two of the three schemes were included in round one, with the third added for round 2, causing the substantial increase in households and plots from round 1 to 2. During each round, households that weren’t surveyed in any previous round are also asked about missed rounds retroactively, so reponse rates are expected to be higher for earlier rounds, as there were multiple occasions for households to be asked about their respective periods. We collect detailed information on the cultivation patterns on all plots farmed by that family. The frequent surveys are timed roughly with the length of cropping cycles for typical crops. We use the planting and harvest dates to estimate the growth stage in which the crops are each week, using information on average growth stage length for each crop in the region. Combining this with information on weekly water requirements in each growth stage allows us to calculate an estimate of the weekly water requirements (mm/week) for all plots. We further exploit questions in which we ask farmers to de- clare whether they perceive water to be sufficient for production and whether they had a dispute concerning water for each of their plots with another farmer over the period of months between the survey rounds. A total of seven rounds of survey data were collected 4 Prior to each round of agriculture surveys, the listing is updated to track any changes of plot ownership within the scheme as well as in and out migration. In limited occasions plots are divided up or merged leading to small changes in total number of plots across survey rounds. The numbers listed refer to the number of households and plots in round 2, the first round that included all schemes. 5 We addressed non-response in a given round by re-surveying these same households in subsequent survey rounds and collecting retrospective data corresponding to the previously missed interview. The main reason for non-response are people farming their rain-fed plots during the rainy season, which can be far from the irrigation schemes. 10 WATER WHEN IT COUNTS between August 2015 and July 2017.6 Precipitation and water requirements Finally, weekly precipitation data (mm/week) were extracted from the NOAAs Cli- mate Prediction Centre (CPC) CMORPH product (Climate Prediction Center, 2015).7 Crop specific water requirements are from the FAO Cropwat 8.0 tool (Doorenbos and Kassam, 1979; Allen et al., 1998; FAO, 201). These data allow us to estimate the quan- tity of rain that falls on these schemes in every day, so that we can assess total water available to a plot on a given day as the depth of rain that falls on a plot plus the depth of water that could be applied from irrigation. C. Descriptive Statistics The first step in our study is to use our novel high-frequency irrigation monitoring system to assess irrigation and cultivation strategies in the absence of pricing or feed- back interventions for one full year with the goal of uncovering whether farmers in these schemes are using water efficiently. If we find that farmers are using water inefficiently, the next step is to identify ways in which water could be more effectively allocated. These strategies underlie the content for the extension tools we use to test whether individual feedback is necessary to improve the efficiency of allocations. This sub-section summa- rizes findings from one full year of water monitoring to uncover what we believe to be a driving source of inefficiency: a particular type of dynamic inflexibility of allocations to time-specific requirements that we call rules of thumb. Cultivation Strategies 6A detailed time-line of data collection activities can be found in the appendix in Figure A1. 7 Rainfalldata were extracted for the smallest geographical unit available in these data. Since the schemes are near to each other, this unit covers all three schemes, so precipitation data do not vary across schemes within a time period. WATER WHEN IT COUNTS 11 Table 2 presents basic baseline descriptives of the plots collected in our sample.8 Of the 120 farmers originally listed at the start of our study, the average household in a given scheme cultivated between 1.1 and 1.5 plots, with average plot sizes ranging across schemes from 0.4 to 0.9 hectares. Most of these plots were under active cultivation at baseline (67.2% to 77.4% across schemes). Typical crops include maize and horticultural crops such as cabbage, collard greens, and piri-piri, a variety of red chili pepper. Baby corn, a miniature ear of corn cultivated for export, is the most frequently cultivated crop. Between 53% and 81% of farmers cultivated baby corn in the irrigation schemes. Timing of Rainfall and Role for Irrigation Among rainfed plots in our study region, the main production cycle takes place during the wet season (October through April). Irrigation allows farmers to reliably cultivate crops during both the wet and dry seasons, with the dry season occurring in May through September. In this context, the primary function of irrigation is to smooth water avail- ability between rain events. The predictable schedule of water availability facilitated by irrigation allows farmers to stagger planting relative to the rainy season.9 Self-reported Water Sufficiency and Conflicts Over Water To motivate the value of information related to crop water requirements, Figure 1 high- lights the water stress reported by farmers in these schemes prior to the intervention. The bars show the proportion of plots whose cultivators reported disputes over water (blue bars) and the proportion of plots whose owners reported that water was sufficient to meet their crops’ requirements (red bars). Unsurprisingly, farmers clearly feel they do not have sufficient access to water by the end of the dry season. It is important to emphasize, how- 8 Notethe scheme that entered the sample in round 2 uses retroactive data to formulate these descriptive statistics. 9 Theannual distribution of precipitation shown in Figure A6 highlights the role for irrigation. Clearly the months of May through September (the dry season) reflect a period of water scarcity. Water availability spikes during rain events as shown in Figure A7, but irrigation maintains a relatively smooth average level of water availability between events. 12 WATER WHEN IT COUNTS TABLE 2—D ESCRIPTIVE STATISTICS Scheme 1 Scheme 2 Scheme 3 Sample sizes in baseline Number of listed households 47 31 53 Number of listed plots 62 31 63 Average number of plots farmed per household 1.472 1.093 1.245 Share of listed households surveyed 92.0% 88.4% 87.3% General scheme characteristics in baseline Irrigation type Canal Sprinkler Canal Average household landholding (ha) 1.528 1.642 1.832 Average plot size (ha) 0.422 0.622 0.890 Share of cultivated plots 67.2% 77.4% 72.2% Share of households that planted common crops in baseline Baby corn 53.2% 80.6% 66.0% Collard greens 23.4% 3.2% 17.0% Maize 48.9% 22.6% 17.0% Piri-Piri 29.8% 3.2% 13.2% Cabbage 23.4% 32.3% 39.6% Average share of plot area dedicated to most common crops in baseline Baby corn 43.6% 43.4% 22.0% Collard greens 23.0% 4.0% 8.3% Maize 64.6% 53.5% 26.8% Piri-Piri 51.9% 18.0% 23.0% Cabbage 43.8% 17.8% 22.4% Notes: Data refers to rounds 1-3 (rounds pre-feedback experiment). The number of listed households and plots indicates the total number of unique observations in those rounds. All other variables are averaged across the 3 rounds. ever, that water stress is not unique to the dry season. Insufficiency is reported for some plots in every month, and disputes are common in most months of the year. Reports of conflict coincide with reports of water insufficiency; the slope of a regression of monthly WATER WHEN IT COUNTS 13 conflict reports on monthly reports of sufficiency is -.473. F IGURE 1. WATER S CARCITY AND C ONFLICT D. Stylized Facts of Water Monitoring The main objective is to understand the sources of water stress in order to generate and test the key components of feedback intended to improve efficiency. Data collected from the high-frequency water monitoring system can be used to quantify the water available for each plot, each crop, at each growth stage in our three study schemes. As defined in Section II.B water availability is computed by estimating the maximum volume of water that could be delivered to a plot in a particular week given measurements taken thrice daily of water flows through the canals, hydrants, or sprinklers serving the plot. When 14 WATER WHEN IT COUNTS this volume is divided by area, our measure of water availability is interpretable as the depth of water that could be applied to a plot if the volume of water is evenly distributed across a plot. This measure allows us to benchmark the degree to which a plot has enough water to meet recommended requirements. To summarize the degree to which a plot is being allocated sufficient water to meet the cultivated crop’s requirements, we compute the water gap as water availability minus growth stage specific water requirements. Growth stages are defined by the planting dates reported by farmers in the follow-up surveys, and the average length of the growth stages according to the FAO Cropwat 8.0.10 The gap is defined in mm/day for each crop-specific growth stage. A negative water gap means that there is not enough water available to the plot to meet the recommended watering requirements given its growth stage. A positive gap means that sufficient water is available for the plot to meet its requirements. A plot whose water gap is larger in absolute value is applying a level of water that is further away from the recommendation. Using the water gap to measure proximity to the water requirements reveals interesting patterns of misallocation. Figure 2 shows the average water available to plots observed with crops in growth stages 1, 2, 3, or 4 (gray bars) along with the recommended re- quirements for the three most common crops (lines). What emerges is a striking pattern between water availability and requirements across crop growth stage. While water re- quirements vary greatly by growth stage, water availability does not. When crops are at growth stage 3, the available volume of water to the average plot is close to (just above) the requirement for crops. But in growth stages 1 and 2, we do not observe water quanti- 10 For the remainder of the paper, we will refer to four main growth stages, labeled 1-4, which correspond to these stages as follows: 1-Initial stage, 2-Vegetative Stage, 3-Reproduction Stage, and 4-Maturation Stage. The implications of water stress on total production vary depending on the timing of the stress in the crop cycle. Water shortages during the initial and reproduction stages typically can lead to irreversible damages (Varzi, 2016). WATER WHEN IT COUNTS 15 F IGURE 2. AVERAGE AVAILABILITY VS . REQUIREMENTS ties close to recommended levels. In fact, the water availability exceeds what is needed. Water scarcity appears to stem primarily from mis-allocation across plots within a scheme. Figure 3 displays the distribution of the water gap in each growth stage. We observe that the water gap distributions for growth stages 3 and 4 are centered around 0. Hence, enough water was available to meet the water recommendations for only half of the cultivated plots. In contrast, in earlier growth stages, virtually all plot-crop observa- tions fall to the right of 0, indicating that there is sufficient water available to meet the recommended water allocations for those plots and crops. These patterns are consistent with the idea that farmers follow rules of thumb for their plots, where they try to system- atically allocate a fixed volume of water to a given plot across the crop cycle. Yet, there 16 WATER WHEN IT COUNTS F IGURE 3. D ISTRIBUTION OF WATER G AP is not enough water for every farmer to follow the rule of thumb in every period. The broader consequences of this individual behavior are shown in Figure 4, in which the water gap is aggregated across plots and shown over time. The line shows the sum of all water gaps on all plots weighted by area.11 The fact that the line never crosses zero indicates that at a scheme level, there is sufficient water available in every week to meet scheme-wide requirements. Yet, a large percentage of plots does not have access to enough water to meet requirements, as indicated in the overlaid bar graph (Figure 4). The above evidence offers two stylized facts. First, water allocations are inefficient. Second, water scarcity arises from basing water applications on fixed quantities rather 11 In other words, this displays the total volume of water available at the scheme level minus the total amount of water required to meet the requirements for all cultivated crops in a given week. WATER WHEN IT COUNTS 17 F IGURE 4. D ISTRIBUTION OF WATER G AP than dynamic crop requirements. In addition, we notice substantial variation in within scheme planting times. Hence, some farmers have crops in early growth stages while others have crops in late growth stages.12 Taken together, these facts suggest one obvious way to improve the efficiency of water allocation. If farmers could be convinced to use less water in the first and second growth stages when water requirements are lower, there would be more water available to farmers who are in the third and fourth growth stages when requirements are higher. In the next section, we present a stylized model to demonstrate how using rules of 12 The dates of planting for the full year period in which all planting and harvesting decisions are observed prior to the feedback experiment are shown in Appendix A.A4. While July seems to be a slightly more popular month for planting than other months, there is no month which accounts for less than 4 percent of annual planting activity and no month that accounts for more than 13 percent. 18 WATER WHEN IT COUNTS thumb interact with dynamic requirements to create inefficiencies. We use this model to motivate what mechanism can be used to improve overall water efficiency. Ultimately, we cannot disentangle what drives current water applications.13 We then use this model to inform the design of an information experiment to combat these observed inefficiencies. III. Conceptual Framework for Efficiency Improvements We build a simple model of water balance to show that following fixed rules without taking advantage of crop dynamics can create a simultaneous situation of wasted water at a scheme level and water scarcity at a plot level. This model helps us to conceptualize where the gains from water allocation might come from and how information can achieve improvements. For each plot i in a scheme, a farmer chooses how much water to use in each growth stage and the investment of other inputs so as to maximize her production, subject to a water balance constraint at the scheme level. We can write this constraint as follows: N (1) max πi (w1i , ..., wGi ; xi ) s.t . wgi ;xi ∀g ∑ wig ≤ Wt ∀ t i=1 In equation (1), π () is a concave crop production (or revenue14 ) function concave in water delivered per growth stage, wig , and in other inputs, xi . The water balance constraint requires that the sum of all water allocations across N plots is less than or equal to the 13 It should be noted that we cannot disentangle precisely what motivates these rules of thumb. A few hypotheses come to mind. One possibility is that because outcomes from different watering levels are noisy, farmers are imperfectly informed about the true relationship between water and yields. If crops are especially sensitive to water in growth stage 3, it may be that these rules of thumb arise from testing different quantities of water and settling on the level that is required at the most intensive stage. Alternatively, the rules could arise from coordination costs, wherein farmers know that revising water allocation requests upward is costly. Thus, they request the maximum quantity they need for that crop over the life cycle, at planting, in order to avoid having to request an increased allotment as the crop develops. 14 We refer to farmers maximizing production. Realistically, they maximize profits, accounting for the cost of inputs. Since water is free, we abstract away from other inputs by assuming that the π () function includes the net cost of inputs that would be optimally chosen for the given set of water choices. Farmers then simply choose the level of water that maximizes their harvest given an optimally chosen set of complementary inputs on a given plot. WATER WHEN IT COUNTS 19 total availability of water in the system in week t , given by Wt . If a scheme only has one plot, then water allocation for each growth stage solves: ∂ πigt (2) =0 ∂ wigt For plot i under cultivation in growth stage g during week t , call the level of water that solves equation (2) w∗ igt . This is the optimal water allocation given perfect information. If a farmer delivers more water than w∗ igt to her plot, overwatering causes yields to be lower than potential given a choice of inputs xi through waterlogging. If a farmer delivers less than w∗ igt to her plot, water stress causes yields to be below potential. A parameter of primary interest then is the water gap for a given growth stage as wigt − w∗ igt . A positive water gap in a growth stage g observed in week t means too much water is being delivered to a plot, while a negative one means too little water is being delivered. In a single-plot scheme with no cost to water and perfect information, a positive water gap would never be observed, because excess water would be allowed to flow out of the scheme rather than being allocated as availability to the plot, and a negative one would be observed only when wigt < Wt . Introducing multiple farmers under perfect information yields a water availability con- straint of ∑N ∗ ∗ N ∗ i=1 wigt ≤ Wt . All farmers choose wigt = wigt . If ∑i=1 wigt > Wt , so that it is not possible to fully satisfy every farmer’s desired water demand, then the cooperative solution will be the set of water allocations per growth stage that solves: ∂ πigt ∂ π jgt (3) = ∀i, j, g, t . ∂ wigt ∂ w jgt 20 WATER WHEN IT COUNTS If π () is concave with regard to wg , the set of water availability allocations with full information that solves (3) will always have wigt ≤ w∗ igt . Thus, under perfect information and coordination, no one overwaters their crops, and underwatering will only occur when there is not sufficient water to meet every farmer’s demand in all growth stages. As a result, is not possible to improve any farmer’s water usage without making another farmer worse off given overall constraints on the total quantity of water available at the scheme level. Farmers may be misinformed about the true return of water on production. If they are using rules of thumb in which they believe that crop requirements do not vary by growth stage, then farmers act as though the following is true: ∂ πg ∂π (4) = ∀g ∂ wg ∂ w Under these beliefs, farmers will choose a water allocation as for all growth stages to solve: ∂π (5) =0 ∂w Since equation (5) is fixed for all growth stages, water use under rules of thumb will ¯ for all i, g, where w also not vary by growth stage, implying that wigt = w ¯ is a fixed constant. If true water requirements for plants do vary by growth stage, w∗ ¯ 15 for i,g = w 15 In fact we observe substantial variation in water availability across plots for plots with crops in the same growth stage. This could arise from each farmer setting the rule of thumb at a different level so that w¯ would be indexed by plot i. Alternatively variation could arise from all farmers having the same fixed rule, but having different bargaining position to negotiate the scarcity that arises from everyone following the inefficient rule. For example, imagine that there are only two farmers, each with one plot in the scheme, and the upstream farmer freely chooses water availability subject only to the total water constraint, while the downstream farmer chooses water availability subject to the upstream farmer’s choice and the constraint. Then if 2 ∗ w¯ > Wt for a period t , the upstream farmer will choose wit = w ¯ and the downstream farmer will choose w jt = Wt − w¯ . This bargaining would also generate variation in observed allocations even if everyone is trying WATER WHEN IT COUNTS 21 some growth stages. With this behavior there is scope for improving farmers’ water use without making other farmers worse off, because some farmers will have a non-zero water gap even if ∑N ∗ i=1 wigt ≤ Wt . Improvements in water allocation are possible without changes in total water at the scheme level if different farmers are in different growth stages at the same time. The following holds for some farmer i in growth stage g and another farmer j in growth stage h the following for at least one week t : (6) ¯ − w∗ w ¯ − w∗ igt < 0 < w jh Equation (6) indicates that if there are growth stages in which the requirements are less than the rules of thumb and growth stages in which the requirements are greater than the rules of thumb, there is scope for efficiency improving reallocation. By taking ¯ − w∗ water of the quantity w ¯ − w∗ jh and giving it to w ig , farmer i can be made better off without making farmer j worse off or changing the total quantity of water available to the scheme. Equation (6) shows a role for information in water management where farmers are following rules of thumb. Any reminder or feedback system that induces farmer j to ¯ to w∗ shift from a water allocation of w jh will improve overall water allocations in the scheme. Several parameters of interest for evaluating the value of information emerge from the simple model above. The first is the water gap for a plot in a given week. Given an es- timate of true water requirements, the observed water gap on a plot is defined above as to follow the same rule. Since we cannot distinguish variation from bargaining or each farmer having a different rule of thumb in the data, for readability we do not index w by plots i or model the bargaining that resolves overuse. 22 WATER WHEN IT COUNTS wit − w∗ it 16 At a scheme level, the aggregate water gap is W − N π (w∗ ).17 In weeks it ∑i=1 igt when the aggregate water gap is positive, information should weakly decrease the pro- portion of people who are experiencing a negative water gap. We will use the proportion of plot-weeks where we observe a negative water gap as our primary indicator of the value of information. If providing feedback reduces the share of plots where there is less water available than is required to meet recommendations, we will say that feedback is welfare enhancing. At a scheme level, water savings attributable to the information intervention is com- puted using the following: T N (7) ∑ ∑ [w∗ igt − w ¯] t =1 i=1 If all farmers are following rules of thumb, (7) the total quantity of water used by all ¯ over all cultivated plots in the farmers between period 1 and period T is the sum of the w period. The prediction for whether a scheme in total will use less water when feedback is provided is ambiguous. If rules of thumb are “mostly” higher than optimal, information that shifts farmers toward the optimal quantity will reduce the total quantity of water used. If rules of thumb are mostly lower than optimal quantities, farmers may respond to the information by trying to increase water uptake overall, subject to the constraint that the sum of all water used is less than the total water available to the scheme. It is not clear ex ante what types of reminders will achieve feasible improvements 16 The water gap is interpreted in the same way as the observed water gap described in section II. It is the difference between the farmer’s chosen water allocation and the recommended volume applied given inputs and the crop cultivated on that farmer’s plot. 17 The aggregate water gap cannot be changed by farmers’ behavior subject to choices of crops and planting times, since it is a function of only crop requirements and total water available in the scheme, but it is a measure indicating whether farmers are making constrained choices since it is positive if and only if there is enough water to meet each farmers optimum water demand given optimal allocations. WATER WHEN IT COUNTS 23 in water allocations. It may be sufficient to remind farmers that crop requirements are not constant, allowing them to adjust the rules of thumb accordingly. But farmers may be misinformed of how their own water allocations actually relate to requirements. The literature on energy metering suggests a key role for individualized feedback to align users with allocations. Since information on individual metered information is expensive to collect and implement, we test whether individual reminders are needed to achieve the allocation gains identified in equation (6). IV. Field Experiment Section III shows that efficiency-improving reallocations are possible given the type of practices observed in Section II.D. One hypothesis is that providing information on water requirements may lead to a redistribution of water from plots that over-watered in the early growth stages to plots that under-watered in the later growth stages. It is not clear what kind of information is crucial to achieve this redistribution process. We designed and ran a field experiment to test the impact of providing information on water recommendations on water allocation. We implement two different feedback tools, one providing general watering recommendations across growth stages, and a second that adds individualized feedback on a farmer’s own actual water use to the general watering recommendations. We randomly assign all households in our study area to one of two treatments, general recommendation (general feedback treatment) or general recommendation plus individ- ualized feedback (individualized feedback treatment). The first round of feedback tools was distributed in December 2016. For the randomization, we used the latest household listing available from the July 2016 survey. We expect that improvements in allocation 24 WATER WHEN IT COUNTS that arise from the provision of information would come from high water users redis- tributing water to low users. To ensure an even mix of treated households who are high or low users at baseline, we randomized the treatment in matched pairs. Within a scheme, each farmer was ranked by their mean water gap across all crops. The initial pair-wise randomization was done after round 3 when 66 households were assigned to the general feedback treatment and 63 to the individual feedback treatment. Table A1 reports ob- servable characteristics are balanced across treatment assignment at the 5 percent critical level. Farmers who join after this round are pooled together and randomly assigned to either treatment each round. Across all rounds, 75 households were assigned to general feedback and 72 to individual feedback. Figure 5 illustrates a sample of a chart provided to a farmer assigned to the general feedback treatment in Scheme 1. The sheet shows his primary crop is red pepper on plot 1010202, and includes a chart highlighting both the water requirements and length in days of each growth stage. The chart is meant to communicate the general pattern of relatively minimal amounts of water are required in growth stages 1 and 2. Higher applications of water are necessary at later stages of the crop cycle. Farmers assigned to the individualized feedback treatment receive information about their actual water use in the same season of the previous year in addition to the infor- mation about their primary crop’s water requirements. Figure 6 shows both the water requirement and water use information for a sample producer of tomatoes in Scheme 1. For this specific case, we witness the water applications exceed the requirements at every growth stage. Moreover, tomato production demands more water at later stages of growth. The water use pattern is the reverse of the water required. Feedback is specific WATER WHEN IT COUNTS 25 to tomato planted on plot 1010101.18 The information treatments were delivered by our enumeration team during each agri- culture survey. The feedback was provided prior to starting the survey. The information experiment began in December 2016 (Figure A1). The information treatment was re- peated for the same farmers two additional times in April 2017 and July 2017 to ensure that farmers received feedback in each of the relevant seasons. Treatment is assigned only once, so that a farmer once assigned always receives the same type of feedback, but the crop they are informed about can vary depending on which crop was most relevant in that same upcoming period last year. Water measurement and agriculture surveys continued for three rounds after the start of the treatment, following the schedule described in section 2.2. A total of 157 households were listed as part of the schemes, of which: 125 were surveyed in at least one of the pre- feedback rounds (rounds 1-4), and 129 responded to at least one of three post-feedback surveys (rounds 5-7). Finally, 115 households responded to at least one pre-feedback and one post-feedback round. Since feedback was delivered during the survey, delivery of feedback is conditional on finding the household for that survey round. Out of the 75 and 72 assigned to the general guidance and individual feedback, respectively, 66 and 62 received at least one round of feedback.19 Scheduling of water allocations is typically arranged by heads of water user associ- ations, and is fixed according to the days of the week and time of day. The fact that the generalized feedback communicates to farmers that they should vary their water use throughout the crop season runs in direct contrast to the idea of a fixed schedule. We 18 The farmers are reminded of the plot the information relates to using a map of the scheme. The feedback tools displayed here are shown with English translations. The versions used in the field were written in Portuguese. 19 8 households that joined the schemes after round 4 did not receive any treatment. New households that joined after the original matched pairs randomization were individually assigned a feedback treatment status. 26 WATER WHEN IT COUNTS F IGURE 5. F EEDBACK T OOLS -G ENERALIZED I NFORMATION therefore expect generalized feedback to induce farmers to request changes to the water- ing schedule, and thereby influence water use patterns. WATER WHEN IT COUNTS 27 F IGURE 6. F EEDBACK T OOLS -I NDIVIDUALIZED F EEDBACK V. Results A. Feedback Effects on Water Availability and Conflict We now present results on the impact of our two information treatments on water use. By reminding farmers that the plants they grow require less water in the early stages of 28 WATER WHEN IT COUNTS growth, we expect that farmers will choose to use less water during early growth stages, allowing them to reallocate water to the later growth stages. If this hypothesis is correct, we should observe reduced conflict over water. Since both treatments receive feedback, we start by using an event study design to draw comparisons in the incidence of water-related conflicts before and after the introduction of feedback (Figure 7). In the month immediately before the first distribution of feedback (December), the incidence of conflict was similar to that reported in the same month of the previous year. One month after farmers started receiving feedback (January), reports of conflict feel to zero and remained there from January to May. Over the dry period (July-November), conflicts are again reported, but their incidence remains substantially below that of the previous year. These findings are largely corroborated with our plot- level measurements of water sufficiency (Figure 8). Overall, these results are consistent with the idea that farmers quickly began to use water more efficiently in accordance with the recommendations they received. Figures 7 and 8 show that households were more likely to perceive water as sufficient and were consequently less likely to report conflicts over water in the months following the introduction of feedback than they were before feedback started. As Sections II.D and III suggest, the main gains from reallocation would come from reallocating water across planting stages. If farmers respond to reminders and there is any scope to increase watering, they should increase watering only in the most water-intensive (late) growth stages, when crops are not getting enough water. Instead, watering in the early growth stages should decline, as there is already more than sufficient water. This general pattern of reallocation is observed in Figure 9 among both treatment groups. In both groups, the distribution of water use in the first growth stage shifts left (or closer to requirements) WATER WHEN IT COUNTS 29 F IGURE 7. S HARE OF HOUSEHOLDS THAT REPORTED CONFLICT OVER WATER following the introduction of the information feedback. This shift allows for water to be reallocated towards the third and fourth growth stages, as we observe a greater share of the density of plots to the right of the point of sufficiency in those stages. The shift observed in Figure 9 has two consequences. First, the distribution of wa- ter use relative to the crop-specific requirement is better aligned across growth stages, suggesting that farmers are refining their rules of thumb to make watering more specific to growth stages. Second, as a result of this realignment, insufficiency as measured by water availability not meeting water requirements (the left of the vertical red line) has decreased. Relative to the pre-feedback period, the distribution of water availability in the most water-intensive growth stages falls more on the side of sufficient water than was 30 WATER WHEN IT COUNTS F IGURE 8. S HARE OF PLOTS THAT HAD ENOUGH WATER the case prior to the introduction of information. The experimental variation in our data comes from the randomized assignment of whether a farmer receives individualized feedback in addition to the general recommen- dation. The previous distributions (Figure 9) do not reveal obvious differences in how farmers responded to the additional component of feedback information, but the visual representation may mask subtle differences. In order to formally test whether the in- dividualized information changes behavior more than the general recommendation, we estimate the following difference-in-difference test of outcomes on treatment status: WATER WHEN IT COUNTS 31 F IGURE 9. WATER GAP PER GROWTH STAGE (8) Di jgt =α + β Individual f eedbacki + γ Postt + δ Individual f eedbacki ∗ Postt + Xi jgt + εi jst where Di jgt is a dummy variable indicating whether farmer i has a negative water gap (water available is less than water recommended) for crop j in growth stage g in time t . The effect of providing any feedback (general or individualized) is given by γ , which measures the proportion of farmer-crops that received insufficient water to meet the crop requirements at various growth stages following the provision of any information. If any 32 WATER WHEN IT COUNTS feedback causes farmers to redistribute water from early to later growth stages, we expect there to be a reduced proportion of farmer-crops suffering from a lack of water to meet crop requirements, or γ < 0, following the receipt of feedback. The additional effect of individualized feedback, to inform farmers of how their own usage compared to the gen- eralized recommendation in the previous year, is measured by δ . We hypothesize δ ≤ 0. While the literature on energy use provides empirical support for individualized feed- back on resource-conserving behavior, it is possible that tailored-information provides no additional insight to the farmer, in this context. We condition measurement of these two effects on Xi jgt , which includes stratified pair fixed effects and month of planting fixed effects. Standard errors are clustered at the household level to account for autocor- relation. Table 3 shows that the results of information provision accord with predictions. In columns 1 and 2, the share of plot-crop-week observations where the farmer has a neg- ative water gap falls by 14% after treatment begins. The difference is significant at the 1% level. When we interact the post and individual feedback dummies with an indicator for whether a household was below the median water gap in rounds prior to the feedback experiment, the regression results show that the post-information treatment is associated with a reduction in the number of people who have a below-zero water gap of 19.8%, which is significant at the 1% level. There is no evidence that individualized information has any additional effect.20 20 We also evaluated generalized feedback effects on the natural log of revenue per hectare. We find no evidence that improved water allocations lead to significant improvements in farmer revenue (Table A2). WATER WHEN IT COUNTS 33 TABLE 3—P LOT- CROPS MOVED FROM INSUFFICIENT TO SUFFICIENT WATER AVAILABILITY (1) (2) Post -0.140∗∗∗ -0.198∗∗∗ (0.023) (0.028) Individual feedback -0.0216 -0.0369 (0.022) (0.028) Individual feedback * post 0.00818 0.0177 (0.034) (0.043) High water gap -0.155∗∗∗ (0.038) Post * High water gap 0.136∗∗∗ (0.043) Individual feedback * High water gap 0.0455 (0.049) Individual feedback * Post * High water gap -0.0344 (0.066) Observations 3724 3709 Adjusted R2 0.201 0.208 Randomization pairs fixed-effects Yes Yes Month of planting fixed-effects Yes Yes Notes: Observations at are at household-round-plot-crop-growth stage level. Sample is restricted to plots that were cultivated. Standard errors are clustered by household to account for autocorrelation. ∗ p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01 B. Robustness In the previous section, we find that the share of people with a negative water gap falls after the initiation of information provision, but that there is no additional gain from individualized feedback. However, because everyone receives general information, we cannot rule out that a factor other than information caused the improvement in people meeting requirements. The most obvious candidate for an external factor would be in- creased rainfall. Figure 10 shows the share of plots under cultivation with a negative 34 WATER WHEN IT COUNTS water gap, along with rainfall. Information provision coincided with the start of the rainy season, and it is clear that rainfall was both more intense and consistent during the rainy season following the information treatment. However, for the last several months of the study, we observe the dry season when rain has stopped. In the dry season, the share of plots with negative water gap is higher than the rainy season, but it is much lower than in the previous year. This suggests that we cannot rule out that information contributed to the observed improvements in farmers’ ability to meet average requirements. One concern in interpreting the result that individual feedback does not have an ad- ditional effect on water use relative to general feedback is that attrition from the survey is correlated with the treatment. Attrition appears to be balanced by feedback treatment status. Only 6.67% of households assigned to individual feedback do not appear in any post-feedback round, and only 9.23% of households assigned to general feedback do not appear in any post feedback round. The p-value for the difference in these rates of attrition is .559.21 We additionally assess the relationship between water requirements and availability before and after the information treatment began to substantiate the claim made in this paper: the observed changes farmers make in response to any feedback is related to mov- ing away from rules of thumb and towards an allocation where water available to plots more closely matches growth-stage-specific requirements. If farmers are following rules of thumb with no regard to requirements, the slope of a regression of availability on re- quirements would be zero. If they were allocating perfectly efficiently the slope should be one. 21 We further investigate selective attrition by testing whether the pre-feedback values of observables used to test balance in A1 can predict attrition. We cannot reject that attrition is balanced on these variables. A joint test of their significance has an F-stat of 1.22. WATER WHEN IT COUNTS 35 F IGURE 10. S HARE OF PLOTS WITH NEGATIVE WATER GAP BY WEEK Figure 11 shows the correlation between availability and requirements before and after information treatment began. The gray line shows the correlation in the pre-information period, the blue line shows the correlation in the post-information period, and the red dashed line shows the 45-degree line where the slope is 1 for reference. Clearly avail- ability matches requirements much more closely after the information was provided than before. A test of whether the coefficient of this relationship22 equals one before the treatment strongly rejects the null hypothesis in the pre-feedback period (t statistic, 22 Formally, we test whether β = 1 in the following regression: availability cgtm = α + β ∗ requirementscgtm + raint + schemeg + monthm + εcgtm , where availabilitycgtm is the daily average volume of water per hectare delivered to the plots in a furrow-command area c in scheme g in study week t in calendar month m, and requirementscgtm is the recommended quantity of water for the mix of crops cultivated in this furrow command area according to the crop mix and growth stage of crops. Controlling for rainfall ensures that the coefficient only measures changes associated with allocation decisions made by farmers and not common level effects of rainfall experienced by everyone in the scheme. schemeg is a vector of scheme fixed effects and monthm is a vector of calendar month fixed effects to control for seasonality. 36 WATER WHEN IT COUNTS p-value=0.031). In the post information period, the coefficient of availability to require- ments is 1.727, and we reject that the slope is equal to 1 with a p-value of .023, sug- gesting that allocations as assessed by availability may have become overly responsive to requirements in response to the feedback. If requirements are not correlated with rain, as we would expect in irrigation schemes where planting decisions need not be tied to rainfall, differences in rain in the pre- and post- feedback periods would be expected to shift the axis of the post-information line in Figure 11 rather than the slope. Since it is the slope that changes and not the constant, this is further evidence that it is the informa- tion rather than an external factor, such as increased rainfall, that improves the ability of farmers to at least meet minimum requirements. This improvement is observed among farmers who receive the general reminders as well as those who receive the individual feedback, suggesting that individualized feedback based on metering is not instrumental in achieving the improvement. VI. Conclusion Using high-frequency data collected in three irrigation schemes in central Mozam- bique reveals that a major source of water inefficiency comes from applications of rules of thumb. In this context, these rules of thumb are likely established to satisfy equitable decision rules pertaining to the scheduling of extracting water within a water user group. Yet, these rules fail to consider the precise water requirements of specific crops within the scheme over time. As a result, a significant portion of farmers tend to over-water at earlier stages of the crop cycle, jeopardizing the amount of water available at later stages of the crop cycle when water requirements are at their peak. We show that providing farmers with an auxiliary rule of thumb related to the water WATER WHEN IT COUNTS 37 F IGURE 11. R ELATIONSHIP BETWEEN REQUIREMENTS AND AVAILABILITY requirements of their primary crops may offer a low-cost intervention to remedy water scarcity in schemes. Reminding farmers of the water requirements over their primary crop’s growth cycle significantly reduced the number of conflicts over water use and the proportion of farmers who self-reported having insufficient water. We show the improve- ment comes from increases in water usage only during the most water-intensive weeks, suggesting that farmers respond correctly to information that highlights inter-temporal distinctions in water requirements. In contrast, administering user-based water monitor- ing systems to provide individualized feedback on water use does not appear to merit its associated costs. Water conservation will remain crucial for the broader maintenance of water basins 38 WATER WHEN IT COUNTS in Mozambique. While generalized recommendations appear sufficient to reduce water scarcity in ways that are consistent with improved scheduling rules being the mechanism, we do not find that these improvements are associated with significantly higher revenue for farmers. Continued water monitoring will reveal whether profits may improve as farmers better optimize their cropping strategies to reflect the newly more efficient allo- cations of water. What we learn from this experiment is that information campaigns may be instrumental in achieving water savings over periodic dry spells. REFERENCES Aker, Jenny C. 2011. “Dial A for agriculture: a review of information and communi- cation technologies for agricultural extension in developing countries.” Agricultural Economics, 42(6): 631–647. Aker, Jenny C., Ishita Ghosh, and Jenna Burrell. 2016. “The promise (and pitfalls) of ICT for agriculture initiatives.” Agricultural Economics, 47(S1): 35–48. Alcott, H., and T. Rogers. 2014. “The Short-Run and Long-Run Effects of Behavioral Interventions: ExExperiment Evidence from Energy Conservation.” American Eco- nomic Review, 104(10): 3003–3037. Allen, R., L. Pereira, D. Raes, and M. Smith. 1998. “Crop Evapotranspiration: Guide- lines for Computing Crop Water Requirements.” Food and Agriculture Organization of the United Nations FAO Irrigation and Drainage Paper 56, Rome. Beaman, L., A. BenYishay, J. Magruder, and A. M. Mobarak. 2015. “Can Network Theory-based Targeting Increase Technology Adoption?” University of California, Berkeley Working Paper. BenYishay, A., and A. M. Mobarak. Forthcoming. “Social Learning and Incentives for WATER WHEN IT COUNTS 39 Experimentation and Communication.” Review of Economic Studies. Braca, Giovanni. 2008. “Stage-discharge relationships in open channels: Practices and problems.” FORALPS Technical Report, 11, , (1): 44. Buytaert, W., and et al. 2014. “Citizen Science in Hydrology and Water Resources: Opportunities for Knowledge Generation, Ecosystem Service Management, and Sus- tainable Development.” Frontiers in Earth Science, 2. Climate Prediction Center, National Centers for Environmental Prediction, Na- tional Weather Service NOAA U.S. Department of Commerce. 2015. “NOAA NCEP CPC CMORPH daily mean morphed cmorph: estimated precipitation data from 23 Feb 2005 to present.” Cole, S., and A. N. Fernando. 2016. “‘Mobile’izing Agricultural Advice: Technology Adoption, Diffusion and Sustainability.” Notre Dame University Working Paper. Deichmann, Uwe, Aparajita Goyal, and Deepak Mishra. 2016. “Will digital tech- nologies transform agriculture in developing countries?” Agricultural Economics, 47(S1): 21–33. Doorenbos, J., and A. Kassam. 1979. “Yield Response to Water.” Food and Agriculture Organization of the United Nations FAO Irrigation and Drainage Paper 33, Rome. FAO. 201. “Cropwat 8.0.” FAO. 2016. “Aquastat: Mozambique.” Ferraro, P., and M. Price. 2013. “Using Nonpecuniary Strategies to Influence Behavior: Evidence from a Large-Scale Field Experiment.” Review of Economics and Statistics, 95(1): 64–73. Fraternali, P., A. Castelletti, R. Soncini-Sessa, C. Ruiz, and A. Rizzoli. 2012. “Putting Humans in the Loop: Social Computing for Water Resources Management.” Environ- 40 WATER WHEN IT COUNTS mental Modelling and Software, 37(C): 68–77. Global Facility for Disaster Reduction and Recovery (GFDRR). 2017. “Country Profile: Mozambique.” Accessed online February 9, 2017 at: https://www.gfdrr.org/sites/gfdrr/files/region/MZ.pdf. Hanna, Rema, Sendhil Mullainathan, and Joshua Schwartzstein. 2014. “Learning through noticing: Theory and evidence from a field experiment.” The Quarterly Jour- nal of Economics, 129(3): 1311–1353. Kondylis, F., V. Mueller, and J. Zhu. 2017. “Seeing is Believing? Evidence from an Extension Network Experiment.” Journal of Development Economics, 125: 1–20. Little, K., M. Hayashi, and S. Liang. 2016. “Community-Based Groundwater Moni- toring Network Using a Citizen-Science Approach.” Groundwater, 54(3): 317–324. Lowry, C., and N. Fienen. 2013. “CrowdHydrology: Crowdsourcing Hydrologic Data and Engaging Citizen Scientists’.” Groundwater, 51(1): 151–156. Meinzen-Dick, R. 2007. “Beyond Panaceas in Water Institutions.” Proceedings of the National Academy of Sciences, 104(39): 15200–15205. Ostrom, E. 1990. Governing the Commons: The Evolution of Institutions for Collective Action. Cambridge:Cambridge University Press. Ostrom, E., and E. Schlager. 1992. “Property-Rights Regimes and Natural Resources: A Conceptual Analysis.” Land Economics, 68(3): 249–262. Plusquellec, H. 2002. “Is the Daunting Challenge of Irrigation Achievable?” Irrigation and Drainage, 51(3): 185–198. Varzi, M. M. 2016. “Crop Water Production Functions–A Review of Available Mathe- matical Method.” Journal of Agricultural Science, 8(4): 76–83. Vidal, John. 2016. “As Mozambique’s Rivers Dry Up, the Hopes of a Harvest Evaporate WATER WHEN IT COUNTS 41 Too.” The Guardian. World Bank. 2007. “Mozambique Country Water Resources Assistance Strategy: Mak- ing Water Work for Sustainable Growth and Poverty Reduction.” 56855, World Bank Group. World Bank. 2010. “Mozambique Public Expenditure Review for the Water Sector.” 54692-MZ, World Bank Group. World Bank. 2011. “Mozambique: Analysis of Public Expenditure in Agriculture.” 59918-MZ, World Bank Group. 42 WATER WHEN IT COUNTS A PPENDIX A1. Data Collection Timeline F IGURE A1. T IMELINE OF ACTIVITIES WATER WHEN IT COUNTS 43 A2. Estimating per plot water-availability from people-based observations F IGURE A2. M AP OF S CHEME 1: G AUGES WHERE DAILY DISCHARGE WAS MEASURED ARE MARKED IN TURQUOISE , THE RIVER PROVIDING SOURCE WATER IS BLUE , THE MAIN CANAL IS RED , LINED SECONDARY CANALS ARE BLACK , UNLINED SECONDARY AND TERTIARY CANALS ARE YELLOW, LIGHT RED AREAS FRAMED IN WHITE , AND PLOTS BELONGING TO ONE FARMER ARE DESIGNATED AS LIGHT RED AREAS FRAMED IN WHITE . The traditional way to gather information on discharge at a gauge is to measure the water level and use a previously derived stage-discharge relationship to derive the current discharge. For all monitored discharge sites in the three schemes, these relationships were established. Using a flow-meter, the discharge for at least 4 different water levels was measured and the parameter c f it of Equation A1 was calibrated. The standard procedure is further described in (Braca, 2008). During the measurement campaign, discharge in secondary canals is either measured directly or indirectly via the difference of instantaneous discharge in the main canal be- fore and after the abstraction point (see Equation A2). In the Scheme 1, daily measure- ments are taken at 21 gauges which allows to derive the discharge at 19 secondary canals (see Figure A2). For the Scheme 3, 27 gauges were calibrated to derive the discharge at 32 secondary canals (see Figure A3). The water-level was measured 3 times daily, at 8am, 12am and 16pm. An irrigation 44 WATER WHEN IT COUNTS F IGURE A3. M AP OF SCHEME 3: G AUGES WHERE DAILY MEASUREMENTS WERE PERFORMED ARE MARKED IN YELLOW, BLUE : M AIN AND SECONDARY CANALS OF IRRIGATION NETWORK , LIGHT BLUE : RIVER , PLOTS ARE FRAMED IN WHITE . duration of 12 hours per day was assumed to derive the daily volume of water that was abstracted by the respective secondary canal (Equation A3). Using an area-based parti- tioning (Equation A4) the available volume was then distributed within all plots that are supplied by the respective canal. The amount of water applied on plots via sprinkler irrigation on Scheme 2 (see Fig- ure A4) is calculated under the assumption that a) a constant sprinkler discharge QSprinkler of 0.0034m3 /s is present. It was derived through volumetric dosing and b) a constant ir- rigation duration of 4 hours. The amount of sprinklers located on each sprinkler plot on scheme 2 was monitored 3 times per day allowing to derive the water availability on a daily basis (see Equation A5). WATER WHEN IT COUNTS 45 F IGURE A4. M AP OF S CHEME 2. WHITE : SPRINKLER PLOTS , RED : DELINEATION OF PRESSURIZED PIPES TO WHICH SPRINKLER SYSTEMS ARE CONNECTED TO . 2/3 Q(t ) = c f it ∗ A ∗ Rhydraulic [m3 /s] h(t ) = Instantaneous water level at gauge [m] (A1) Rhydraulic = f (h(t )) = Hydraulic radius [m] A = f (h(t )) = Cross-sectional area [m2 ] c f it = Fit parameter [m1/3 /s] (A2) Qabstracted (t ) = Q pre (t ) − Q post (t )[m3 /s] 46 WATER WHEN IT COUNTS F IGURE A5. T HE STAGE - DISCHARGE RELATIONSHIP AT G AUGE 16 IN S CHEME 1: H = WATERLEVEL , Q = D IS - CHARGE , T HE RED - CROSSES INDICATE FOUR ON - SITE FLOW MEASUREMENTS . T HE BLACK LINE INDICATES THE ESTABLISHED STAGE - DISCHARGE CURVE USING A LEAST SQUARE FIT TO CALIBRATE THE PARAMETER c f it FROM EQUATION A1 . (A3) Vcanal = (Q8am + Q12am + Q16 pm ) ∗ 4 ∗ 602 [m3 /day] Area( ploti ) V ( ploti ) = Vcanal ∗ N [m3 /day] (A4) ∑i=1 Area( ploti ) N = Number of plots served by this secondary canal WATER WHEN IT COUNTS 47 (A5) V ( ploti ) = (Sprinkactive (8am)+ Sprinkactive (12 pm)+ Sprinkactive (18 pm)) ∗ Qsprinkler ∗ 4 ∗ 602 [m3 /day] 48 WATER WHEN IT COUNTS A3. Trends in Water Monitoring Data F IGURE A6. I NTRA - ANNUAL VARIATION IN P RECIPITATION (1998-2016) Note: Precipitation time series (1/1/1998-05/03/2016) extracted from the NOAA CPC CMORPH product, available at: http://www.cpc.ncep.noaa.gov/products/janowiak/cmorph description.html. A4. Planting Dates from Agriculture Surveys WATER WHEN IT COUNTS 49 F IGURE A7. AVAILABILITY OF WATER FROM P RECIPITATION AND I RRIGATION OVER THE P ERIOD BEFORE I NFOR - MATION T REATMENT 50 WATER WHEN IT COUNTS F IGURE A8. N UMBER OF CROPS PLANTED BY PLOT BY MONTH OF PLANTING WATER WHEN IT COUNTS 51 A5. Balance on observables from matched pairs randomization TABLE A1—BALANCE TABLE (1) (2) T-test General feedback Individual feedback Difference Variable N/[Clusters] Mean/SE N/[Clusters] Mean/SE (1)-(2) Listing Plot area (ha) 56 0.648 54 0.705 -0.056 [56] (0.076) [54] (0.061) Household average across plots Average share of plot area that was cultivated 56 0.283 54 0.320 -0.037 [56] (0.026) [54] (0.036) Share of months household reports there was conflict 56 0.229 54 0.193 0.036 [56] (0.037) [54] (0.037) Share of months household reports there was enough water 56 0.742 54 0.657 0.085* [56] (0.033) [54] (0.040) Average yield per hectare (thousands of Meticais) 53 154.043 50 123.594 30.449 [53] (19.147) [50] (13.432) F-test of joint significance (F-stat) 1.46 Household-plot-crop-growth stage observations Total water availability (log mm/day) 282 1.808 298 1.869 -0.061 [52] (0.061) [49] (0.093) Mean precipitation in that GS (log mm/day) 282 0.846 298 0.855 -0.009 [52] (0.049) [49] (0.059) Water gap (log mm/day) 278 0.602 298 0.625 -0.023 [52] (0.061) [49] (0.098) Share of observations with negative water gap 278 0.183 298 0.201 -0.018 [52] (0.030) [49] (0.033) Absolute water gap (log mm/day) 278 0.715 298 0.736 -0.021 [52] (0.048) [49] (0.087) F-test of joint significance (F-stat) 2.89** Notes: Data refers to round 2 when available, as feedback was randomized based on data from this round, and round 1 when the household wasn’t surveyed on round 2. Seven households did not cultivate any crops during this round’s period. For these households, the average share of plot area cultivated is zero, and yields are missing. The number of observations for water gap is different from the number of observations for water availability due to lack of data on water requirements for onion. The value displayed for t-tests are the differences in the means across the groups. Standard errors are clustered at the household level. The covariate variable pair is included in all estimation regressions. ***, **, and * indicate significance at the 1, 5, and 10 percent critical level. 52 WATER WHEN IT COUNTS A6. Effect of Information Provision on Yields TABLE A2—D IFF - IN -D IFF E FFECT OF I NFORMATION ON Y IELDS (1) (2) Individual feedback 0.0208 -0.00756 (0.124) (0.187) Post 0.422 0.690 (0.391) (0.462) Individual feedback * Post -0.134 -0.333 (0.198) (0.289) Average pre-feedback water gap is negative -0.0892 (0.218) Individual feedback * Average pre-feedback water gap is negative 0.0693 (0.239) Post * Average pre-feedback water gap is negative -0.548∗∗ (0.266) Individual feedback * Post * Average pre-feedback water gap is negative 0.496 (0.354) Observations 660 617 Adjusted R2 0.058 0.074 Randomization pairs fixed-effects Yes Yes Round fixed-effects Yes Yes Notes: Observations at are at household-round-plot level. Sample is restricted to plots that were cultivated. Standard errors are clustered by household. ∗ p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01