Improving development efficiency through decision analysis: Reservoir protection in Burkina Faso

Accepted Manuscript Improving development efficiency through decision analysis: Reservoir protection in Burkina Faso Denis Lanzanova, Cory Whitney, Keith Shepherd, Eike Luedeling PII:

S1364-8152(18)30112-9

DOI:

https://doi.org/10.1016/j.envsoft.2019.01.016

Reference:

ENSO 4380

To appear in:

Environmental Modelling and Software

Received Date: 10 February 2018 Revised Date:

10 October 2018

Accepted Date: 22 January 2019

Please cite this article as: Lanzanova, D., Whitney, C., Shepherd, K., Luedeling, E., Improving development efficiency through decision analysis: Reservoir protection in Burkina Faso, Environmental Modelling and Software (2019), doi: https://doi.org/10.1016/j.envsoft.2019.01.016. This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

ACCEPTED MANUSCRIPT

Improving development efficiency through decision analysis:

2

reservoir protection in Burkina Faso

RI PT

1

3

Denis Lanzanova1,2*, Cory Whitney2, Keith Shepherd3, Eike Luedeling3,4,

4

1

5

66, 1205 Geneva, Switzerland;

6

2

7

Bonn, Germany;

8

3

9

4

SC

University of Geneva – Institute for Environmental Sciences, Boulevard Carl-Vogt

M AN U

University of Bonn – Center for Development Research, Genscherallee 3, 53113

World Agroforestry Centre – United Nations Avenue, Gigiri, Nairobi, Kenya University of Bonn, INRES Horticultural Sciences, Auf dem Hügel 6, 53121 Bonn

*Corresponding author: [email protected]

11

Declaration of interest: none

13

Abstract

14

AC C

12

EP

TE D

10

15

In the arid areas of Sub-Saharan Africa, perennial challenges of water scarcity and food insecurity are exacerbated by climate change and variability. The

16

development of robust strategies to cope with the region’s climatic challenges

17

requires thorough consideration of uncertainty and risk in decision making.

18

We demonstrate the use of probabilistic decision analysis to compare 1

intervention options to prevent reservoir sedimentation in Burkina Faso. To

20

illustrate this approach, we developed a causal impact pathway model based

21

on the local knowledge of expert stakeholders. Input parameters were

22

described by probability distributions derived from estimated confidence

23

intervals. The model was run in a Monte Carlo simulation to generate the

24

range of plausible decision outcomes, quantified as the net present value and

25

the annual cash flow. We used Partial Least Squares regression analysis to

26

identify the parameters that most affected projected intervention outcomes

27

and we computed the Expected Value of Perfect Information (EVPI) to

28

highlight critical uncertainties. Numerical results show that the preferred

29

intervention to secure agricultural production is a combination of dredging,

30

rock dams and a buffer scheme around the reservoir. The EVPI calculation

31

reveals an information value for the profit per ton of vegetables, indicating

32

that more information on this variable would be useful for supporting the

33

decision. However, without the need for follow-up analysis, the results show

34

high probability of benefits given the combined interventions, which, given the

35

current state of information, should be preferred over inaction.

36

37 38

sedimentation, Burkina-Faso

EP

TE D

M AN U

SC

RI PT

19

AC C

ACCEPTED MANUSCRIPT

Keywords

Monte-Carlo simulation, Decision analysis, Uncertainty assessment, Reservoir

2

1 Introduction

40

Harsh climatic conditions and population growth are major challenges to food

41

security in sub-Saharan Africa (SSA)(Omisore, 2017). In arid and semi-arid

42

regions across SSA, farmers face highly fluctuating water supply, a situation

43

that will likely be exacerbated by climate change (Wood et al. 2014). At the

44

same time, water demand for food production is expected to rise as

45

populations grow (Amisigo et al. 2015; World Bank 2017). These issues are of

46

particular concern to the Upper Volta river basin. The region is considered

47

highly sensitive to environmental changes and rainfall variability (Amisigo et

48

al. 2015), and agriculture is dominated by small-scale, rainfed production,

49

with very few farmers having access to irrigation (less than 1% of agricultural

50

land is irrigated; Bharati et al. 2008). As a consequence, agricultural

51

productivity is very low.

52

In Burkina Faso, these conditions are the root cause of widespread rural

53

poverty (Sanfo and Gérard 2012). To cope with the local challenges, many

54

communities, governments, and NGOs have constructed small-scale reservoirs

55 56 57

EP

TE D

M AN U

SC

RI PT

39

AC C

ACCEPTED MANUSCRIPT

for irrigation in rural areas. More than 2000 of these structures have been

built in the Upper Volta basin (Cecchi et al. 2008). They provide seasonal

water storage for small-scale irrigation during the cropping season (Bharati et

58

al. 2008), water for livestock and an opportunity for fish farming (Venot and

59

Cecchi 2011). They act as a buffer against extreme weather events (Boelee et

3

ACCEPTED MANUSCRIPT

al. 2013) and are considered instrumental for local food security and

61

livelihoods (Palmieri et al. 2001; Wisser et al. 2010; Poussin et al. 2015).

62

However, the performance of many of these reservoirs has fallen short of

63

expectations (Barbier et al. 2011), and many are subject to degradation and

64

poor maintenance (Venot and Hirvonen 2013). One of the major problems is

65

sedimentation caused by soil erosion within the reservoir catchment, which

66

can damage the irrigation system in the short term (Kondolf et al. 2014) and

67

eventually fill in the reservoir completely, rendering the infrastructure useless

68

(Schmengler 2011; Chitata et al. 2014; Kondolf et al. 2014). Rehabilitation of a

69

reservoir damaged by sedimentation is costly at best, and sometimes not

70

possible at all (Kondolf et al. 2014). Possible intervention points for managing

71

sedimentation are the initial design of the reservoir, management of the

72

agricultural activities and the establishment of structures in the riparian areas

73

(Kondolf et al. 2014).

74

Sedimentation is a major investment risk, since it limits the time horizon, over

75

which communities can benefit from a reservoir. This highlights the

76

importance of appropriate management strategies, through which the

78

SC

M AN U

TE D

EP

AC C

77

RI PT

60

productive lifetime of these structures can be greatly extended (Palmieri et al.

2001). However, the best way to manage sedimentation is often not directly

79

apparent to communities and governments, due to uncertainty regarding the

80

costs and benefits and the poor understanding of sediment generating

81

processes and their response to management actions (Schleiss et al, 2016). To 4

ACCEPTED MANUSCRIPT

choose appropriate strategies, many policy-makers, development

83

practitioners and NGOs need scientific support.

84

Predicting sediment yield and accumulation through time is extremely

85

difficult (Morris and Fan, 1998). The Universal Soil Loss Equation (USLE)

86

(Wischmeier and Smith, 1978), later modified to the Revised Universal Soil

87

Loss Equation (RUSLE) (Renard et al, 1997) has been widely applied for this

88

purpose at the watershed scale (Griffin et al., 1988; Jain et al., 2001) but

89

suffers from high uncertainty in factors influencing reservoir sedimentation

90

(Salas and Shin, 1999). Limitations in data availability, among other factors,

91

can lead to high uncertainty in model results.

92

Bayesian frameworks have been used to link model calibration and

93

uncertainty assessment. The most common approaches include the Bayesian

94

Monte Carlo method (Qian et al., 2003), Markov chain Monte Carlo

95

(Kattwinkel and Reichert, 2017) and the generalized Likelihood Uncertainty

96

Estimation (GLUE) pseudo-Bayesian method (Zheng and Keller, 2007), which

97

are used, among others applications, to establish uncertainty bounds for

98

simulated flows (Fonseca et al., 2014). Chaudhary and Hantush (2017)

100

SC

M AN U

TE D

EP

AC C

99

RI PT

82

presented a novel approach that combines a Bayesian Monte Carlo simulation with a maximum likelihood estimation. Yet such advanced treatments of

101

uncertainty have largely been restricted to uncertainties about classic

102

hydrological parameters, whereas the success of watershed interventions

103

often depends on a host of additional factors, e.g. in the social and economic 5

ACCEPTED MANUSCRIPT

domains, which can have a major impact on success or failure of innovations.

105

Comprehensive consideration of all relevant factors is required.

106

Here we present the use of decision-centered models to address risks and

107

uncertainties in sedimentation management decisions. The approach

108

embraces complexity, makes recommendations that account for the imperfect

109

state of available knowledge and identifies critical uncertainties that decision-

110

supporting research should address (Luedeling and Shepherd 2016).

111

Decision-centered models can be collaboratively developed between analysts,

112

stakeholders and decision-makers through participatory processes, where all

113

important factors involved in a decision are gathered and synthesized into an

114

ex-ante impact projection model (Luedeling et al. 2015). Tools that determine

115

the value of information can be used to identify the most critical knowledge

116

gaps, from the perspective of a decision-maker aiming to optimize overall

117

desired outcomes (Luedeling and Goehring 2017). Such modeling techniques

118

can prioritize the knowledge gaps that should most urgently be narrowed in

119

order to reduce uncertainty about the decision or inform the design and

120

prioritization of future research (Hubbard 2014; Rosenstock et al. 2014;

122

SC

M AN U

TE D

EP

AC C

121

RI PT

104

Strong et al. 2014; Luedeling et al. 2015). Collecting additional information on

such high-value variables and using this information to update the decision

123

model allows decision-makers to iteratively improve their ability to anticipate

124

decision outcomes and identify the preferred option. When sufficient data is

6

ACCEPTED MANUSCRIPT

available, the coupling with a watershed hydrological transport model might

126

be considered.

127

Research into sedimentation control in small reservoirs has largely been

128

based on disciplinary analyses, but it has not yet been able to capture many

129

important uncertainties related to the social and natural systems on which

130

reservoirs depend. We use the specific example of a small reservoir that

131

serves the communities of Lagdwenda in the Northern Volta basin, Tenkodogo

132

district, Bougou province, Burkina Faso, to demonstrate the application of

133

Decision Analysis tools for sustainable management of small reservoirs.

134

Sedimentation is the main operational concern affecting the reservoir. It

135

impacts the reliability of irrigation systems and is a major threat for the

136

resilience of the local communities to all types of climatic shocks.

137

Sedimentation in the reservoir results from a number of known factors, such

138

as, among others, poorly planned grazing of river banks or conversion of

139

natural vegetation. We demonstrate tools that can support the difficult task of

140

deciding which interventions to choose, if any, to mitigate the problems of

141

sedimentation due to the absence of riparian vegetation. We make use of a

144

AC C

EP

TE D

M AN U

SC

RI PT

125

145

interventions (buffer strips, rock dams, dredging and combinations of these).

146

Such an exercise typically involves synergies and trade-offs between

147

interventions, which have been shown to affect the achievement of

142 143

causal impact pathway model, constructed and parameterized based on the

knowledge of local expert stakeholders. We use Monte Carlo simulations to compare the ranges of plausible outcomes for several locally recognized

7

ACCEPTED MANUSCRIPT

environmental targets at large scales (Gao and Bryan, 2017). By evaluating

149

model outputs, we seek to identify the parameters that most affect projected

150

intervention outcomes and to highlight critical uncertainties.

151

2 Materials and methods

152

2.1 The reservoir of Lagdwenda

153

The reservoir of Lagdwenda is located in the Northern Volta basin, in

154

Tenkodogo district, Bougou province, Burkina Faso. The site is semi-arid, with

155

a dry season from October to mid-May and a rainy season from mid-May to

156

October. Average annual rainfall varies between 800 and 900 mm. The

157

warmest period is from March to May and a relatively cooler period from June

158

to February, with an annual average temperature of 29°C. The reservoir of

159

Lagdwenda (Figure 1) had a capacity of 63,000 cubic meters in 2002 (year of

160

construction) and benefited more than 7,000 people in 2005.

AC C

EP

TE D

M AN U

SC

RI PT

148

161 8

ACCEPTED MANUSCRIPT

Figure 1. Lagdwenda reservoir in the Northern Volta basin of Burkina Faso. a. Map of the

163

reservoir. The blue polygon at the bottom center represents the downstream irrigation

164

scheme, while the top red polygon represents the upstream cropping area. b. Variation in

165

reservoir surface area. Dark blue indicates year-round water (April 2016, 18 ha), light blue

166

indicates low-wet season water (January 2016, 34 ha), and the outer blue layer indicates

167

high wet-season water (October 2015, 55 ha). Data from Landsat 8 (MNDWI).

168

The main use of the reservoir is irrigation for rice and vegetable production.

169

Downstream from the reservoir, a formal irrigation scheme has been

170

developed (Figure 1a). In addition, farmers have established an informal

171

cropping area upstream of the reservoir (Figure 1a), which violates key

172

recommendations on reservoir protection (Schleiss et al. 2016) and

173

contributes to sedimentation. An important secondary use of the reservoir is

174

water provision for livestock and animals in the riparian area, which also

175

contributes to sedimentation.

176

Reservoir size, as well as the types of crops that are cultivated, vary markedly

177

between the wet and the dry season (Figure 1b), with the reservoir reaching

178

its maximum extent of 55 ha at the end of the wet season, when farmers

179

practice paddy rice cultivation. During the dry season, farmers grow mixed

181 182

SC

M AN U

TE D

EP

AC C

180

RI PT

162

vegetables and some cereals. Sedimentation control interventions are urgently

needed and local decision-makers are looking for cost-effective ways to

restore reservoir functions and ensure their provision over the long term.

9

ACCEPTED MANUSCRIPT

2.2 Overview of the approach

184

The method proposed in this paper provides a new approach to support

185

practical decisions on agricultural systems in the face of risk and imperfect

186

information. It is inspired by the Applied Information Economics (AIE)

187

approach developed by Hubbard Decision Research (Hubbard 2014). This

188

decision analysis approach, which has been widely used in business decision

189

support and a number of other contexts (e.g. Luedeling et al., 2015; Wafula et

190

al., 2018), employs participatory processes to explore in detail the

191

consequences of a particular decision. Rather than aiming to precisely predict

192

results for all available decision options, which is usually impossible for even

193

moderately complex systems, AIE attempts to capture the state of knowledge

194

on all processes and input variables, e.g. in the form of probability

195

distributions, and translate these into probabilistic simulations that predict

196

the full range of plausible outcomes. From these outputs, critical knowledge

197

gaps can be derived, measurements can be undertaken to narrow these gaps,

198

and the model can successively be updated until confident decision support is

199

possible.

201

SC

M AN U

TE D

EP

AC C

200

RI PT

183

Building on the AIE approach, we conducted quantitative ex-ante impact analyses for several decision options, using Monte Carlo simulations to

202

account for risks and uncertainties. The methodology combines participatory

203

approaches and modeling techniques. As a first step, a decision-centered

204

model is collaboratively developed between analysts, main stakeholders and 10

ACCEPTED MANUSCRIPT

decision-makers during a workshop. Model development seeks to capture all

206

important factors for the decision, regardless of whether they can easily be

207

measured or modelled. After synthesizing these variables into a model, the

208

state of knowledge on all variables is quantified through the use of probability

209

distributions. When no information is available on particular variables, values

210

are elicited from participants. Before providing these estimates, participants

211

are subjected to calibration training (Hubbard 2014) to improve their ability

212

to estimate their state of uncertainty. Such training has been shown to

213

increase people’s capacity to provide accurate estimates by reducing errors of

214

judgment (Lichtenstein et al. 1982). All estimates are consolidated into one

215

single probability distribution for each model parameter (Luedeling et al.

216

2015). In this way, uncertainty is explicitly represented as probabilities of

217

different possible states of the world (Pannell and Glenn 2000). A description

218

of the process can be found in Hubbard and Millar (2014). Basically, the

219

principle is to combine methods from decision theory, economics and

220

actuarial science in order to improve on human expert judgments.

221

In a second step, once numbers are available, the model is run in order to

224

AC C

EP

TE D

M AN U

SC

RI PT

205

225

knowledge gaps. The variables with the highest information value can be

226

interpreted as priorities for measurements to be undertaken to reduce

227

uncertainty around the decision (Rosenstock et al. 2014). The EVPI can thus

222 223

convert probabilistic inputs into probabilistic outputs, which express the

range of plausible decision results that can be expected. We then use the

expected value of perfect information (EVPI) to determine the most critical

11

ACCEPTED MANUSCRIPT

be used to inform the design and prioritization of future research (Strong et al.

229

2014) and may help reduce the range of plausible outcomes.

230

2.3 Analysis protocol

231

2.3.1 Step 1: Selecting experts

232

The design of an efficient sedimentation management intervention for the

233

reservoir of Lagdwenda requires assessment of multiple uncertain quantities

234

and risks. Beyond the lack of well-established knowledge on the natural and

235

anthropogenic origins of sedimentation, the issue is very often context-

236

specific. To gather appropriate information, we relied on experts’ knowledge

237

about various ways to manage sedimentation and for identifying parameters

238

of interest (such as benefits, costs and risks).

239

Therefore, the first stage of the protocol focuses on the delicate process of

240

selecting the most relevant experts. The selection process started five months

241

before the decision workshop, with a field visit of the area, in which we met

242

the local communities and their representatives. We also participated in a

243

stakeholder workshop organized in the province by the local office of SNV

245 246

SC

M AN U

TE D

EP

AC C

244

RI PT

228

(Netherlands Development Organization). We used the event as an

opportunity to establish direct contacts with officers from the ministries of

agriculture and environment, as well as local NGOs, and to get a better

247

overview of local experts and relevant stakeholders. We collected names and

248

details of potential participants and worked closely with SNV and the 12

ACCEPTED MANUSCRIPT

agricultural officer of Tenkodogo Province to maintain connections with the

250

local community representatives.

251

The list of potential invitees to a decision analysis workshop was then reduced

252

using selection criteria, e.g. those who have relevant expertise for the specific

253

context of the Lagdwenda reservoir. From these, we selected a group of eight

254

national-level and local experts, who were agricultural specialists, donors,

255

policy-makers and the local communities.

256

2.3.2 Step 2: Eliciting model structure

257

The process to elicit model structures through experts has several steps. First,

258

experts are invited to participate in a decision workshop. The information

259

collected in the participatory analysis of the decision problem is assembled

260

into a conceptual, graphical model. This is constructed as a decision’s impact

261

pathway, with causal relationships based on experts’ expectations, gathered

262

during brainstorming sessions. The conceptual model development aims to

263

capture the ‘big picture’ of the decision by gathering all system dynamics and

264

relevant issues, without taking constraints of measurement into account.

AC C

EP

TE D

M AN U

SC

RI PT

249

13

265

M AN U

SC

RI PT

ACCEPTED MANUSCRIPT

Figure 2. Diagram of a four-stage protocol for eliciting expert knowledge when designing a

267

decision model of different interventions to reduce sedimentation in the Lagdwenda

268

reservoir in Burkina Faso.

269

To elicit more details, we followed the four-stage procedure (Figure 2)

270

outlined by Whitney et al (2018). In the first stage of the procedure the

271

decision about sedimentation control was broken down by the entire group

272

into specific sub-questions. In the second stage, participants were split into

273

working groups to address the different questions both individually and

276

AC C

EP

TE D

266

277

the outputs were revised until a consensus was reached. In the third stage the

278

models produced in working groups were unified into consolidated models,

279

one for each of the initial sub-questions. In the fourth stage the consolidated

274 275

within their groups. We used participatory techniques to avoid problems of

variation among experts (Bolger and Wright, 2017). For example, individual

outputs were peer-reviewed by the other members of the group, after which

14

ACCEPTED MANUSCRIPT

sub-models were combined into one single conceptual model. The final

281

conceptual model represented the impact pathway of the management

282

decision that could be formally modeled and simulated.

283

2.3.3 Step 3: Calibrating experts

284

Parameters of the model are catalogued and grouped into two categories (cf.

285

Figure 3). The first category gathers parameters that can be documented by

286

existing academic or technical sources (e.g. databases, reports, literature). The

287

second category consists of all parameters for which no such sources exist and

288

that should be estimated. These data are simply too costly (and could take too

289

much time) to obtain through survey or field studies. To account for this in the

290

Lagdwenda case, we relied on experts’ knowledge to assess the value of these

291

parameters, but also the uncertainty around these values. It should be noted

292

here that this methodology can be applied regardless of whether these

293

uncertainties result from lack of theoretically attainable knowledge or

294

parameter uncertainties, such as future rainfall amounts, that cannot currently

295

be known precisely.

298

AC C

EP

TE D

M AN U

SC

RI PT

280

299

been adopted and tested for application in conservation science (Martin et al.

300

2012). Expert knowledge can be valuable for decision support, provided it is

301

combined with explicit consideration of uncertainty (Fred et al. 2017). If

296 297

Elicitation techniques are a well-recognized form of knowledge generation in

situations where sampling efforts would be impractical or too expensive

(Samantha et al. 2009), and formal procedures for eliciting and encoding have

15

ACCEPTED MANUSCRIPT

uncertainty is ignored, however, expert knowledge can manifest as spurious

303

opinions, which can undermine any well-intentioned decision analysis process

304

(Morgan, 2014). It is therefore crucial to elicit this information rigorously in

305

order to get valid estimates. This requires recognition and correction for

306

several biases that can affect people’s judgement (Lichtenstein et al. 1982, Soll

307

and Klayman, 2004). In order to avoid such biases, we train experts in

308

techniques that have been shown to improve people’s ability to estimate their

309

own uncertainty and thereby reduce errors of judgement. This is a crucial step

310

in the elicitation process (Figure 3).

311

All experts are required to undergo ‘calibration training’, which teaches them

312

how to make estimates as reliably as possible. The training consists of a series

313

of procedures, grounded on research findings in cognitive psychology.

314

Through these procedures participants learn how to assess their state of

315

uncertainty and reduce errors of judgement through exercises that reveal to

316

them their personal biases (overconfidence or underconfidence). To this end,

317

they compare their performance in responding to trivia questions to the

318

correct answers to these questions. Rather than providing ‘best guesses’,

321

AC C

EP

TE D

M AN U

SC

RI PT

302

322

their biases (most people are initially overconfident), experts are instructed in

323

a set of mental techniques that has been shown to improve people’s ability to

319 320

participants are requested to provide two numbers, for which they are 90%

sure that the correct answer is between these numbers. Perfectly ‘calibrated’

estimators should get 90% of their range estimates correct. Once exposed to

16

ACCEPTED MANUSCRIPT

provide accurate estimates. More information on these procedures can be

325

found in Hubbard (2014).

M AN U

SC

RI PT

324

326

Figure 3. Diagram of the process of expert calibration training for model parameter

328

estimation when parameterizing a decision model of different interventions to reduce

329

sedimentation in the Lagdwenda reservoir in Burkina Faso.

330

TE D

327

2.3.4 Step 4: Estimation and simulation

332

Calibrated experts were trained to use conscious estimation procedures to

333

provide subjective probability distributions for uncertain variables in the

336

AC C

EP

331

337

often default to a normal, uniform or triangular distribution. To help with this,

338

common distribution shapes were displayed during the estimation process.

334 335

decision model. This is usually done by eliciting confidence intervals (defined

by their upper and lower bounds) that have a 90% chance of containing the

value of interest. Selecting a distribution is not always easy for experts, who

17

ACCEPTED MANUSCRIPT

When multiple experts estimate the same quantities, a subsequent process to

340

reconcile different estimates is needed. A general conclusion from the

341

literature on how to combine the diverse elicited values is that averaging is

342

often a preferred strategy (Aidan et al. 2015). Since we had a small expert

343

group, we were able to aggregate individual assessments by consensus.

344

The resulting conceptual model was then reformulated as a set of equations

345

that reflected as much as possible the experts’ and analysts’ understanding of

346

the decision. This mathematical model was coded as a function in R

347

programming language (R Core Team, 2017). All formulas and scripts are

348

available in the supplementary materials as well as in a separate data

349

repository (Luedeling et al., 2018). The model was then parameterized (either

350

with hard data or “calibrated” estimates) and run 10,000 times as a

351

probabilistic Monte Carlo simulation. This number of runs was sufficient for

352

generating smooth outcome distributions for all cases, which was verified by

353

visual inspection. Each run provided one possible outcome. The totality of all

354

model runs generated a probability distribution that illustrates the plausible

355

outcomes given the experts’ current state of uncertainty.

357

AC C

EP

TE D

M AN U

SC

RI PT

339

358

preferable option (e.g. a specific intervention in a group of possible

359

interventions). However, the value of expected outcomes can remain unclear,

360

when uncertainty about input values is high. Sensitivity analysis can identify

356

2.3.5 Step 5: Sensitivity analysis and refinement of the model

The output of a Monte Carlo simulation often directly reveals a clearly

18

ACCEPTED MANUSCRIPT

variables that outcome projections respond to. We used Partial Least Squares

362

(PLS) regression analysis, in particular its Variable Importance in the

363

Projection (VIP) metric, for identifying the variables that most affected the

364

decision outcomes projected by the simulation (Wold, 1995; Luedeling and

365

Gassner 2012). We preferred this method over more systematic sensitivity

366

analysis methods such as the eFAST (Gao, et al., 2016) or Morris (Gao and

367

Bryan, 2016) methods, because it determines sensitivity to input variables

368

based on the outputs of the Monte Carlo analysis, rather than requiring

369

computationally expensive additional simulation runs. For Monte Carlo

370

simulation and sensitivity analysis, we used the ‘decisionSupport’ package

371

(Luedeling and Göhring, 2017) in R.

372

In cases where the outputs of the Monte Carlo simulation clearly identify one

373

decision option as preferable over the alternatives, the current state of

374

knowledge is sufficient for issuing a recommendation on how the decision

375

should be taken. When no option emerges as clearly preferable, decision-

376

critical uncertainties can be identified using the Expected Value of Perfect

377

Information (EVPI). The EVPI is the difference between the value of a decision

380

AC C

EP

TE D

M AN U

SC

RI PT

361

381

to the absolute value of the decision outcome, the EVPI is concerned only with

382

whether its value is positive or equal to zero, because this is the only criterion

383

that decides if more funds should be invested to collect better information.

378 379

made with perfect information and the value of a decision made with information that is currently available. It can be interpreted as a rational

willingness-to-pay to gain access to perfect information. Rather than referring

19

ACCEPTED MANUSCRIPT

As a first step in the EVPI computation procedure, we used Spearman’s rank

385

correlation test to check whether each of the input variables was correlated

386

with projected outcomes. If this was not the case, the EVPI for such variables

387

was set to zero. For all other variables, the relationship between input and

388

output variables was identified by first sorting the array of output values

389

produced by the Monte Carlo simulation according to values of the respective

390

input variable. The resulting set of values was then smoothed using a second-

391

order low-pass Butterworth filter (Proakis and Manolakis, 1992), with a

392

critical frequency of one divided by one tenth of the number of values in the

393

Monte Carlo output. Smoothing is necessary at this stage to separate the signal

394

emerging from the variable under scrutiny from the substantial noise caused

395

by variation in all other variables, which vary randomly within the Monte

396

Carlo simulation. The EVPI was then calculated as the sum of all outputs with a

397

sign that did not correspond to the sign of the expected value, multiplied by

398

the probability assigned to this outcome (Wafula et al., 2018).

399

Measurements of input variables with the highest EVPI, which can be used to

400

update the decision model, help to narrow uncertainty about how the decision

SC

M AN U

TE D

EP

AC C

401

RI PT

384

should be taken. The process is repeated until the best option is determined.

20

ACCEPTED MANUSCRIPT

3 Results

403

3.1 The Decision Model

404

3.1.1 Scoping and design

405

In July 2016, eight experts (see acknowledgments) were consulted in a four-

406

day workshop, where they collaborated to produce graphical models of

407

expected decision impact pathways.

408

The workshop began with framing the decision to be modeled. The overall

409

context was defined in plenary discussions, after which participants were split

410

into working groups. Participants were asked to consider the whole impact

411

pathway, and break it down into several stages. As a second step, the team

412

identified the most relevant interventions for sedimentation management. In

413

order to support the process, ten interventions based on recommendations

414

from international experts and lessons learned from other studies (Kondolf

415

and al. 2014), were proposed and debated by the experts (see Supplementary

416

Information). These were intended to stimulate the discussion; experts of the

417

modeling team were free to define other sediment control interventions.

419

SC

M AN U

TE D

EP

AC C

418

RI PT

402

Ultimately, they identified three possible interventions:

• Dredging along the main stream inlet, allowing water to flow to agricultural

420

fields west of the reservoir and reduce sediment build-up. The intervention

421

consists of dredging 3 meters deep for 2 km along the main stream inlet.

21

ACCEPTED MANUSCRIPT

422

• Building permeable rock dams, also known as ‘check dams’, along the reservoir’s tributaries, reducing flow velocities and the erosive potential of

424

water. These are low dams of loose stone retained by mesh wire that

425

prevent large quantities of sediment from being deposited in the reservoir.

426

These would be constructed every 5 km along the stream network

427

upstream of the reservoir, a spacing that has frequently been used for

428

similar interventions in this region.

SC

M AN U

• A buffer protection scheme around the reservoir and stream inlets, preventing sedimentation from agriculture around the reservoir and

431

reducing deposits of sediments coming from the surrounding catchment.

432

The protection is composed of 3 buffer strips of 75-100 meters each. The

433

first strip is made of stone barriers/contour bounding with stabilizing

434

plants (grasses). The second strip consists of vegetables mixed with shrubs

435

for firewood. The third zone is a mix of crops with fruit trees.

436

Graphic representations of all interventions are shown in Figure 4.

437

We carried out comprehensive investigations of costs, benefits and risks

438 439

EP

TE D

430

AC C

429

RI PT

423

associated with each intervention. Causal relationships between variables were taken into account.

22

RI PT

ACCEPTED MANUSCRIPT

SC

440

Figure 4. Intervention options to reduce sedimentation in the Lagdwenda reservoir in

442

Burkina Faso, co-designed by workshop participants: a. dredging for 2 km along the main

443

stream inlet to a depth of 3 meters; b. low dams of loose stone retained by mesh wire

444

every 5 km along the stream network upstream of the reservoir; c. three buffer strips of 75-

445

100 meters each.

M AN U

441

TE D

446

3.1.2 Benefits, costs and trade-offs

448

Formally, we defined a benefit as an economic surplus (e.g. net revenue from

449

agricultural production) generated by an intervention (or a combination of

450

intervention options) in comparison to the baseline case (current situation).

452

AC C

451

EP

447

Broadly speaking, water for irrigation and viability of the irrigation water supply system were regarded as the main benefits of sedimentation control.

453

Water for irrigation – Because sediments accumulate in the reservoir, the dam

454

gradually loses its capacity to store water. We modeled the avoided expected

455

decline in irrigable area (area in both upstream and downstream irrigation 23

ACCEPTED MANUSCRIPT

schemes that can be irrigated given the water stock in the reservoir) as a

457

benefit for the different scenarios.

458

Viability of the irrigation water supply system – The main irrigation scheme of

459

Lagdwenda is located downstream from the dam. Water supply is achieved

460

through a system of pipes underneath the dam’s barrier. Sediments regularly

461

disrupt the functioning of this system, blocking the pipes and preventing

462

water from flowing into the agricultural area. Interventions that reduce

463

sedimentation prevent the obstruction of irrigation pipes that supply this

464

downstream irrigation scheme.

465

Changes in land use – Trade-offs were expected among land uses when

466

implementing the buffer intervention. In the current situation (baseline case),

467

communities of Lagdwenda grow rice and vegetables in an informal irrigation

468

scheme, located upstream close to the shore of the reservoir, near the main

469

inlet. The buffer protection scheme intervention consists of three buffer strips

470

of 75-100 meters each. This would result in taking this area out of cultivation.

471

On the other hand, two buffers of the scheme are expected to be partially

472

cultivated (vegetables, fruits and cereals).

474

AC C

EP

TE D

M AN U

SC

RI PT

456

475

anthropogenic) that may impact the project, along its different stages. To this

476

end, two categories of risks were considered: ex-ante risks and ex-posts risks.

473

3.1.3 Risks and uncertainties The model seeks to take into account all risks (either natural or

24

Ex-ante risks impact the probability that an intervention is completed, which

478

affects intervention costs. When the project is implemented, the full cost of the

479

intervention applies, while a reduced, sunk cost applies otherwise. We

480

considered three ex-ante risks: the lack of community involvement, the lack of

481

institutional involvement and the lack of donor involvement.

482

Ex-post risks impact the effectiveness of the intervention, which affects the

483

benefits. Ex-post risks can be natural, social or technical. We considered three

484

ex-post risks: natural hazards, poor maintenance and poor design.

485

Additional risk factors, especially risks arising from knowledge limitations, are

486

implicitly captured by the model’s probabilistic inputs. For instance, the risk of

487

changes to the future sediment load of the inlet streams due to climate change

488

was not explicitly expressed but covered by experts’ providing wide

489

confidence intervals for future sedimentation rates.

490

Consideration of risks and uncertainties enables a risk-adjusted cost-benefit

491

analysis to inform decision-making. The decision model that emerged

492

integrates all benefits, costs and risks into estimates of the Net Present Value

493

EP

TE D

M AN U

SC

RI PT

477

AC C

ACCEPTED MANUSCRIPT

of each option (Figure 5).

25

494

AC C

EP

TE D

M AN U

SC

RI PT

ACCEPTED MANUSCRIPT

495

Figure 5. Diagram of the overall model structure of interventions to reduce sedimentation

496

in the Lagdwenda reservoir in Burkina Faso. NPV is the Net Present Value for all

497

interventions. D/S = downstream; U/S = upstream.

26

ACCEPTED MANUSCRIPT

3.1.4 Decision model code

499

We translated information on costs, benefits and risks obtained from the

500

expert workshop into mathematical equations. Code was developed by an

501

expert panel of three decision analysts (D. Lanzanova, C. Whitney and E.

502

Luedeling). Assumptions, model inputs, and the process definitions were

503

formulated beforehand. Formally, this phase included four steps.

504

• Step 1: Specification of input variables. Based on the conceptual model

505

(Fig. 5), we stated the appropriate input information to run the model,

506

either estimated by the experts or computer-generated, as well as the

507

output results (Fig. 6). All input variables were estimated by experts except

508

some general, rather specific parameters such as the project time horizon,

509

the coefficient of variation or the discount factor (supplementary material,

510

section 1.4). Time-series of intermediate variables were generated from

511

these input parameters in order to introduce variability in estimates and

512

randomness in risk realization. Final outputs of Monte Carlo simulation

513

were obtained using the decisonSupport package for R (see step 3 for

515

SC

M AN U

TE D

EP

AC C

514

RI PT

498

details on process functions).

27

M AN U

SC

RI PT

ACCEPTED MANUSCRIPT

516

Figure 6. Input-Output diagram of the simulation model of decisions to implement

518

dredging, rock dams and buffer scheme interventions, as well as all their combinations, to

519

reduce sedimentation in the Lagdwenda reservoir in Burkina Faso. Process functions are

520

detailed in Step 3.

521

• Step 2: Definition of assumptions

EP

TE D

517

We made several assumptions regarding costs, land management,

523

representation of the sedimentation processes and interventions

524 525 526

AC C

522

effectiveness.

-

Total spending costs for an intervention depend on the realization of exante risks, as well as the type of intervention itself. Generally, if an ex-

527

ante risk occurs, the project is not implemented and only study costs

528

apply. Specifically, some ex-ante risks could be irrelevant (e.g. risk of

28

ACCEPTED MANUSCRIPT

529

non-involvement of donors for dredging) and are not considered in that

530

case (supplementary material, section 2.1.1).

531

-

Crop land allocation in the irrigation scheme downstream is assumed to remain unchanged following an intervention. Potential behavioral

533

adjustments by farmers (e.g. new cropping choice in reaction to a

534

change in the water resource available) are not considered. -

The effect of sedimentation on water storage capacity is modeled as a

SC

535

RI PT

532

decline in irrigable area over time. The total irrigable area (area that can

537

be irrigated given the water stock in the reservoir) follows a sigmoid

538

function. It is likely to be the largest in the first two years, then

539

gradually decreases until half of the area is able to receive irrigation

540

water (supplementary material, section 2.3.5) -

The effect of sedimentation on the functioning of the irrigation system is

TE D

541

M AN U

536

modeled as a reduction in the irrigated area due to an obstruction of

543

pipes by sediments that prevent the scheme from being watered. As for

544

the water storage, the decline in the irrigated area follows a sigmoid

545

function: blockages are likely to be rare for a few years but gradually

547 548

AC C

546

EP

542

become more frequent until the irrigation pipes are no longer operational (supplementary material, section 2.3.6).

-

Interventions mitigate the sedimentation effects both regarding the

549

water storage capacity and the functioning of the irrigation system.

550

Formally, the impact of an intervention is modeled as delays in the

551

decrease of the reservoir capacity to store water or in the decline of the 29

ACCEPTED MANUSCRIPT

552

irrigated area because of clogged pipes, as well as in the chance for the

553

pipes to be cleared (supplementary material, sections 2.3.5 and 2.3.6).

555

• Step 3: Selection of functional specifications

RI PT

554

556

As introduced in Figure 6, we used several functions from the R

557

decisionSupport package (Luedeling & Goehring, 2017).

558

-

SC

The chance_event() function was used for random simulation of ex-ante risks as impacts on the implementation of interventions and of ex-post

560

risks as impacts on the benefits (supplementary material, section 2.1)

561

-

M AN U

559

The value varier vv() function was used to introduce variation in timeseries of variables that are assumed to vary over time. The function was

563

applied to ex-ante risks and for simulation of common random draws

564

for all intervention model runs (supplementary material, section 2.2)

565

-

TE D

562

The Gompertz_yield() function was used to simulate the loss in water storage capacity (“irrigable area”), the loss of irrigated area (because of

567

the obstruction of pipes), as well as the delays in these two

568

sedimentation outcomes as a result of the implementation of an

569

intervention. It was also applied for other minor benefits such as

571 572

AC C

570

EP

566

mitigation of the decline in fish in the reservoir (supplementary material, sections 2.3.5 and 2.3.6).

-

The discount() function was used to calculate the net present value

573

which is the sum of the discounted values of the time series of net

574

benefits (supplementary material, sections 2.3.6.1). 30

ACCEPTED MANUSCRIPT

575

-

The decisionSupport() function was used to perform the welfare decision analysis via a Monte Carlo simulation from input variables and

577

to analyze the value of information from these variables

578

(supplementary material, sections 3.1).

RI PT

576

Step 4: Validation of the model. Acceptance of the model assumptions,

580

process flow, choice of specifications and input parameters implies a

581

validation of the decision analysts’ model. To the extent possible, feedback on

582

the emerging code was elicited from the experts involved in building the

583

conceptual model.

584

The complete model code, as well as a detailed explanation of its components

585

can be found in the supplementary materials.

586

3.2 Simulation results

587

3.2.1 Projected intervention outcomes and synergies

588

Net Present Values projected for the various options varied widely (Figure 7),

589

reflecting high uncertainty about many input variables. However, some

592

AC C

EP

TE D

M AN U

SC

579

593

though all options also incurred a risk of negative net outcomes.

594

Among single interventions, dredging appeared to have the lowest potential

595

for benefits, but also the lowest risk of negative outcomes, followed by rock

590 591

indications about overall risks and expected benefit levels could be obtained.

All intervention options, and all their combinations, had positive expected values, indicating that all options promised greater benefits than inaction,

31

ACCEPTED MANUSCRIPT

dams (Figure 7). Due to relatively high up-front costs, buffer strips had the

597

greatest potential for net losses, but they also promised the highest returns

598

among the three options. An intervention that combines all three options

599

looks most promising, in terms of expected value, risk of losses and potential

600

returns.

601

TE D

M AN U

SC

RI PT

596

Figure 7. Distributions of projected outcomes of decisions to implement dredging, rock

603

dams and buffer scheme interventions, as well as all their combinations, to reduce

604

sedimentation in the Lagdwenda reservoir in Burkina Faso.

606 607 608

AC C

605

EP

602

A careful examination of the simulation values (Table 1) provides more

detailed insights. Even though dredging alone was found to be a viable low risk/low return option, it was inferior in all respects (all NPV quantiles, as well

609

as the risk of losses) to an intervention that combined dredging with rock

610

dams. A similar picture emerged for rock dams implemented in isolation, 32

ACCEPTED MANUSCRIPT

which promised lower returns and greater risks than a combination of rock

612

dams and dredging. Both rock dams and dredging implemented alone can thus

613

be excluded as candidates for the optimal sediment management strategy,

614

regardless of the decision-maker’s perception of risk. However, combining

615

interventions, rather than implementing only single ones, may generate

616

synergies and thus promise greater benefits and lower risks than single

617

interventions (Table 1).

SC

RI PT

611

M AN U

618 619

Table 1: Lower and upper bounds of 90% confidence intervals (i.e. 5% and 95%

620

quantiles), median and mean (in USD) or the NPV distribution for different combinations of

621

reservoir protection interventions and probability of loss (in USD) for the different

622

interventions to reduce sedimentation in the Lagdwenda reservoir in Burkina Faso.

Dredging Rock dams

5% quantile

95% quantile

Median

mean

Probability of loss

-112373

284780

42458

46165

0.1306

TE D

Intervention options

-150328

362929

39748

46576

0.2545

-205929

833868

52285

84972

0.3611

-102018

565771

64233

76142

0.1070

Dredging and buffer strips

-232082

842954

74432

104937

0.1819

Rock dams and buffer strips

-244915

826407

67109

90969

0.2449

Dredging, rock dams and buffer strips

-248294

886126

80124

101924

0.2004

Buffer strips

624 625

AC C

623

EP

Dredging and rock dams

A decision maker’s perception of an investment depends both on its risk/return characteristics and on the investor’s degree of risk aversion. The

626

selection of an optimal intervention is therefore related to the preferences of

627

the decision-maker. The exception to this rule is when an option is strongly

628

dominant in all aspects, but this does not apply here. Among all intervention 33

ACCEPTED MANUSCRIPT

options, the combination of dredging and buffer strips and the combination of

630

all three options appear to be most efficient. Associating dredging and buffer

631

strips looks slightly less risky, but the combination of the three options

632

promises a higher median return, and a greater expected value. The group of

633

experts also preferred this option. Therefore, we will only present results for

634

this option below.

635

3.2.2 Projected intervention outcomes and value of information

636

Our decision analysis approach produces two outputs related to the return on

637

investment – the probability distribution for the net present value (NPV) and

638

the annual cash flow (Figures 8a and 8b). We also provide information on the

639

most influential uncertainties regarding the overall magnitude of the NPV –

640

the VIP statistic of Partial Least Squares (PLS) regression – and regarding the

641

emerging decision recommendation – the EVPI. These outputs are presented

642

for the combined intervention consisting of dredging, rock dams and buffer

643

strips (Figure 8).

SC

M AN U

TE D

EP

AC C

644

RI PT

629

34

TE D

M AN U

SC

RI PT

ACCEPTED MANUSCRIPT

Figure 8. Outputs of a Monte Carlo simulation (with 10,000 model runs) for ex-ante

647

analysis of the decision to implement an intervention consisting of dredging, rock dams

648

and a buffer scheme to reduce sedimentation in the Lagdwenda reservoir in Burkina Faso.

649

a) projected distribution of net present value (NPV) for the intervention (with 100%, 90%,

650

50% and 10% confidence intervals shown by shades of purple); b) probability distribution

652 653 654

AC C

651

EP

645 646

of annual cash flow; c) sensitivity of projected outcomes to uncertain input variables, quantified by the VIP statistic of PLS regression (colors indicate positive (green) and negative (red) input/output relationships); d) Expected Value of Perfect Information (EVPI) for all variables with non-zero EVPI (only one variable in this case).

655

35

ACCEPTED MANUSCRIPT

Monte Carlo simulation revealed a wide range of plausible outcomes for the

657

intervention’s NPV. Positive outcomes are likely (80.0%), but results also

658

showed a significant chance of loss (20.0%). The total net present value (over

659

30 years) is likely to be between -64 thousand and 338 thousand USD (the 5th

660

and 95th percentiles of the distribution), with a mean value of 102 thousand

661

USD (Figure 8a).

662

The cash flow analysis illustrated that substantial initial investments are

663

needed (Figure 8b) for implementing the intervention. Accordingly, the

664

expected cash flow in year 1 was negative, with a 90% confidence interval

665

between -300 and -53 thousand USD. From year three, cash flow analysis

666

shows predominantly positive annual net revenue prospects, ranging between

667

-5 and 24 thousand USD. The cash flow stabilized around its highest level

668

(median of 17 thousand USD) towards the end of the time series, emphasizing

669

the viability of this intervention as an effective long-term strategy. Annual

670

benefits appeared quite predictable, with a fairly narrow distribution and little

671

variation over time.

672

3.2.3 Important uncertainties

674

AC C

EP

TE D

M AN U

SC

RI PT

656

675

total of 15 variables had VIP scores exceeding 0.8 – a threshold that is often

676

interpreted as signifying importance (Eriksson et al., 2001). The most

673

The sensitivity analysis (implemented by PLS regression) indicated that a number of input variables had important effects on the NPV (Figure 8c). A

36

influential variable was the profit on a ton of vegetables in the downstream

678

irrigation scheme, followed by the discount rate (Figure 8c).

679

The expected value of perfect information (EVPI) can be expressed in

680

monetary terms as the decision-maker’s willingness to pay to gain access to

681

perfect information (Hubbard, 2014). Decision analysis typically reveals that

682

the decision recommendation can only be influenced by a very small number

683

of uncertain variables, with many others affecting the projected NPV, but not

684

its sign (e.g. Wafula et al., 2018). This, however, is the only important criterion

685

to a decision-maker choosing the optimal option. In the present case, only a

686

single variable – the profitability of vegetable production in the downstream

687

irrigation scheme – had a non-zero EVPI (Figure 8d). Even for this variable,

688

the EVPI was only around 1,100 USD, which is low compared to the overall

689

value of the intervention. This implies that even with the initial state of

690

knowledge, a relatively confident recommendation can be made, but that a

691

small investment in obtaining more information on the economics of

692

vegetable production downstream from the reservoir would be justified by

693

increasing certainty about the decision.

694

695

Here we have demonstrated the application of Decision Analysis techniques to

696

support practical decisions in the face of risk and uncertainty. Its principle is

697

to provide customized analysis for a particular decision context (Howard and

EP

TE D

M AN U

SC

RI PT

677

AC C

ACCEPTED MANUSCRIPT

4 Discussion

37

ACCEPTED MANUSCRIPT

Abbas 2016). This would likely mean that a different set of variables would

699

emerge as important for other reservoirs. In the case of Lagdwenda, the most

700

critical information for decision-making regards the profit per ton of vegetable

701

and the yields of the different crops. This information, especially concerning

702

yields, is highly context specific.

703

In comparison with the physical empirical models that are commonly used to

704

evaluate intervention impacts on sedimentation, which are usually

705

deterministic and restricted to consideration of physical system dimensions

706

(Uusitalo, 2015), our modeling approach has two major advantages: First, it

707

allows doing justice to the considerable importance of political and social

708

factors that are often major determinants of intervention impacts

709

(Holzkämper, 2012). Second, it recognizes that uncertainty is present at every

710

step of an environmental management analysis (Refsgaard, 2007), allowing

711

incorporation of such uncertainties into probabilistic models.

712

The extent to which a strategy is preferred over another depends on many

713

factors, including both its costs (investments, implementation) and its

714

benefits, which in the case of Lagdwenda are mainly the value of the

716

SC

M AN U

TE D

EP

AC C

715

RI PT

698

agricultural output over a period of time (in our case 30 years). Costs are generally based on technical information (such that can be provided by

717

engineers), whereas benefits depend on a number of uncertain agricultural

718

and economic factors (e.g. profit per ha, discount rate). The use of the holistic

719

modeling techniques described here can help decision-makers consider the 38

ACCEPTED MANUSCRIPT

uncertainties instead of making decisions based only on the available technical

721

information.

722

Despite being considered the most effective technical solution to prevent

723

sedimentation (Kondolf et al., 2014; Palmieri et al., 2001; Schmengler, 2011),

724

buffer strips were not the dominant option, mainly because of the high initial

725

costs of implementation. This illustrates the discrepancy between the

726

technical optimum and the best economic option for solving a problem. All

727

intervention options, and all their combinations had positive expected values,

728

indicating that all options promised greater benefits than inaction, though all

729

options also incurred a risk of negative net outcomes. In the context of

730

Lagdwenda, sedimentation can be controlled most effectively when several

731

interventions are implemented simultaneously. Synergies are generated

732

through the interactions of the different interventions, and the extra costs are

733

more than offset by the benefits of the reduction in sedimentation. This is

734

explained by the non-linear impacts of siltation on farmers’ irrigation system.

735

Sediments that originate from erosion of the reservoir shores and from the

736

main inlets interact with each other and result in amplified effects (e.g.

738

SC

M AN U

TE D

EP

AC C

737

RI PT

720

threshold effects in the clogging of irrigation pipes, which makes them

inoperable after a certain amount of sedimentation.

739

Assessment of the Value of Information was useful in identifying and

740

addressing critical uncertainties in the model. The EVPI helps to focus

741

research and decrease overall uncertainty about which decision option to 39

ACCEPTED MANUSCRIPT

choose. This capability makes the EVPI a useful addition to a wide range of

743

decision processes. In the case of Lagdwenda, uncertainties arise

744

predominantly regarding the value of benefits (Figure 8). The two most

745

important uncertainties are the profit per ton of vegetables (with a positive

746

information value) and the discount rate. Many of the important parameters,

747

such as yield and profit, depend mainly on exogenous factors, e.g. price,

748

macroeconomic policy and evolution of inflation. Consequently, it could be

749

challenging (or impossible) to gather reliable information on them, especially

750

for a longer time horizon such as the one used in this model. The uncertainties

751

exposed by the VIP analysis may offer insights into where decision makers

752

may want to conduct further technical research, but such a decision should

753

take external constraints into account.

754

Through the use of the EVPI analysis, we were able to determine decision-

755

critical uncertainties, the reduction of which could help to narrow uncertainty

756

about how the decision should be taken. EVPI (Figure 8d) values are low, and

757

only the profit per ton of vegetables has a non-zero value. This means that

758

uncertainty could be reduced by taking measurements on the profits from

761

AC C

EP

TE D

M AN U

SC

RI PT

742

762

supporting research does not have to eliminate all uncertainty, but can focus

763

on a few key pieces of information.

759 760

crop production. However, this may not even be necessary, since the low EVPI

and the low chance of loss (20%) indicate that the combined intervention should be preferred over inaction. This demonstrates how decision-

40

5 Conclusion

765

Sedimentation in Lagdwenda’s reservoir impacts the reliability of irrigation

766

and thereby the livelihoods of local people, a problem that is common for

767

small-scale reservoirs in the Upper Volta. This problem requires a sediment

768

management strategy that is appropriate to the local context. Development of

769

such a strategy is hindered by data scarcity for many important variables.

770

Decision analysis approaches allow comprehensive ex-ante assessment of the

771

effectiveness of intervention options through collaborative development of

772

impact pathway models. Data limitations are addressed through probabilistic

773

simulation, for which all variables are expressed as probability distributions,

774

and decision-critical knowledge gaps are highlighted by sensitivity analysis

775

and Value of Information analysis.

776

A combination of all three candidate interventions – dredging, rock dams and

777

buffer strips – emerged as the most effective management strategy, with a

778

relatively low chance of losses and long-lasting net benefits for local

779

communities. The magnitude of net benefits depended on several uncertain

780 781 782

EP

TE D

M AN U

SC

RI PT

764

AC C

ACCEPTED MANUSCRIPT

variables, but only a single variable – the profitability of vegetable production in the downstream irrigation scheme – had positive information value. In consequence, a relatively confident recommendation can be provided to

783

decision-makers, even in the face of multiple uncertainties. In addition to this,

784

the analysis revealed that most of the residual uncertainty about whether the

41

ACCEPTED MANUSCRIPT

intervention is worth doing can be addressed through limited measurements

786

on just one variable, which is probably quite easy to measure.

787

Coupling local knowledge systems with rigorous simulation techniques, our

788

methodology provides actionable information for robust decision-making. The

789

model is grounded on the knowledge of local experts and stakeholders who

790

provided a unique understanding of the socio-ecological system that could not

791

have been obtained with traditional disciplinary approaches.

792

6 Acknowledgements

793

This research was supported by the CGIAR Water, Land and Ecosystems

794

research project on Targeting Agricultural Innovations and Ecosystem

795

Services in the Northern Volta river basin (TAI). The authors are grateful to

796

the members of the panel of experts: Daouda Sorgho (Lagdwenda farmers’

797

association), Sorgho Salif (local water committee of Lagdwenda), Ismaël Paré

798

(Provincial department of agriculture and hydraulic infrastructures), Modi

799

Diallo (Provincial department of environment, green economy and climate

800

change), Sambyamba Séraphin Zongo (Nakambé water agency), Barnabé

802 803

SC

M AN U

TE D

EP

AC C

801

RI PT

785

Romuald Konseiga (ECR construction company), Mohamed Wendegoudi Ouédraogo (voluntary association for the promotion of leadership, health and development AJVLS) and Mansour Boundaogo (SNV World).

42

ACCEPTED MANUSCRIPT

References

805

Aidan, L., Bonnie, C.W., Mark, B., 2015. Collective wisdom: Methods of

806

confidence interval aggregation. Journal of Business Research 68, 1759–1767.

807

Allan, J., Samantha, L.C., Kerrie, M., 2010. Elicitator: An expert elicitation tool

808

for regression in ecology. Environmental Modelling & Software 25, 129–145.

809

Amisigo, B., McCluskey, A., Swanson, R., 2015. Modeling impact of climate

810

change on water resources and agriculture demand in the Volta basin and

811

other basin systems in Ghana. Sustainability 7, 6957-6975.

812

Barbier, B., Ouedraogo, H., Dembele, Y., Yacouba, H., Boubacar, B., Jamin, J.-Y.,

813

2011. L’agriculture irriguée dans le sahel ouest-africain. Cah Agric 20.

814

Bezold, C., 2009. Aspirational futures. Journal of Futures Studies 13, 81–90.

815

Bharati, L., Rodgers, C., Erdenberger, T., Plotnikova, M., Shumilov, S., Vlek, P. et

816

al., 2008. Integration of economic and hydrologic models: Exploring

817

conjunctive irrigation water use strategies in the Volta basin. Agricultural

818

Water Management 95, 925–936.

819

Birner, R., McCarty, N., Robertson, R., Waale, D., Schiffer, E., 2010. Increasing

821

SC

M AN U

TE D

EP

AC C

820

RI PT

804

access to irrigation: lessons learned from Investing in small reservoirs in Ghana. International Food Policy Research Institute (IFPRI), Accra, Ghana.

822

Bolger, F., Wright, G., 2017. Use of expert knowledge to anticipate the future:

823

Issues, analysis and directions. International Journal of Forecasting 33(1),

824

230-243. 43

Boelee, E., Yohannes, M., Poda, J.-N., McCartney, M., Cecchi, P., Kibret, S. et al.,

826

2013. Options for water storage and rainwater harvesting to improve health

827

and resilience against climate change in Africa. Regional Environmental

828

Change 13, 509–519.

829

Cecchi, P., Meunier-Nikiema, A., Moiroux, N., 2008. Towards an atlas of lakes

830

and reservoirs in Burkina Faso. Q13 IRD.

831

Chaudhary, A., Hantush, M.M., 2017. Bayesian Monte Carlo and maximum

832

likelihood approach for uncertainty estimation and risk management:

833

Application to lake oxygen recovery model. Water Research 108, 301-311.

834

Chitata, T., Mugabe, F.T., Kashaigili, J.J., 2014. Estimation of small reservoir

835

sedimentation in semi-arid southern Zimbabwe. Journal of Water Resource

836

and Protection 6, 1017–1028.

837

Cooper, P., Dimes, J., Rao, K., Shapiro, B., Shiferaw, B., Twomlow, S., 2008.

838

Coping better with current climatic variability in the rain-fed farming systems

839

of sub-Saharan Africa: An essential first step in adapting to future climate

840

change? Agriculture, Ecosystems & Environment 126, 24–35.

841

843 844

Eriksson, L., Johansson, E., Kettaneh-Wold, N., Wold, S. (2001). Multi- and

845

Megavariate Data Analysis: Principles and Applications. Umetrics Academy,

846

Umeå, Sweden.

842

EP

TE D

M AN U

SC

RI PT

825

AC C

ACCEPTED MANUSCRIPT

David, E.M., Jeremy, E.O., John, A.C., 2014. A web-based tool for eliciting

probability distributions from experts. Environmental Modelling & Software 52, 1–4.

44

ACCEPTED MANUSCRIPT

Fonseca, A., Ames, D.P., Yang, P., Botelho, C., Boaventura, R., Vilar, V., 2014.

848

Watershed model parameter estimation and uncertainty in data-limited

849

environments. Environmental Modelling & Software 51, 84-93.

850

Fred, A.J., Brian, J.S., Mathieu, B., Julien, M., Christina, R., Frank, M. et al., 2017.

851

Expert elicitation, uncertainty, and the value of information in controlling

852

invasive species. Ecological Economics 137, 83–90.

853

Gao, L., Bryan, B.A., 2017. Finding pathways to national-scale land-sector

854

sustainability. Nature 544, 217–222.

855

Gao, L., Bryan, B.A., 2016. Incorporating deep uncertainty into the elementary

856

effects method for robust global sensitivity analysis. Ecological Modelling 321,

857

1-9.

858

Gao, L., Bryan, B.A., Nolan, M., Connor, J.D., Song, X., Zhao, G., 2016. Robust

859

global sensitivity analysis under deep uncertainty via scenario analysis.

860

Environmental Modelling & Software 76, 154-166.

861

Gigerenzer, G., Hoffrage, U., 1995. How to improve bayesian reasoning without

862

instruction: Frequency formats. Psychological Review 102, 684–704.

864

SC

M AN U

TE D

EP

AC C

863

RI PT

847

Gohar, A.A., Ward, F.A., Amer, S.A., 2013. Economic performance of water storage capacity expansion for food security. Journal of Hydrology 484, 16–25.

865

Griffin, M.L., Beasley, D.B., Fletcher, J.J., Foster, G.R., 1988. Estimating soil loss

866

on topographically non-uniform field and farm units. Journal of Soil and Water

867

Conservation 43, 326-333. 45

Holzkämper, A., Kumar, V., Surridge, B.W.J., Paetzold, A., Lerner, D.N., 2012.

869

Bringing diverse knowledge sources together – A meta-model for supporting

870

integrated catchment management. Journal of Environmental Management 96,

871

116-127

872

Howard, R.A., Abbas, A.E., 2016. Foundations of decision analysis. Prentice

873

Hall, NY.

874

Hubbard, D.W., 2014. How to measure anything: Finding the value of

875

intangibles in business. Hoboken, New Jersey: Wiley.

876

Hubbard, D., Millar, M. 2014. Modeling resilience with applied information

877

economics. Report prepared by the Technical Consortium, a project of the

878

CGIAR. Technical Report Series No 1: Measuring Resilience in the Horn of

879

Africa, 24p. Nairobi, Kenya: International Livestock Research Institute (ILRI).

880

Ilan, Y., 2004. Receiving other people’s advice: Influence and benefit.

881

Organizational Behavior and Human Decision Processes 93, 1–13.

882

Jain, S.K., Kumar, S., Varghese, J., 2001. Estimation of soil erosion for a

883

Himalayan watershed using GIS technique. Water Resources Management 15,

884 885 886

EP

TE D

M AN U

SC

RI PT

868

AC C

ACCEPTED MANUSCRIPT

41-54.

Janis, I.L., 1972. Victims of groupthink. Houghton Mifflin, New York. Jonathan, J.K., 1993. The influence of prior beliefs on scientific judgments of

887

evidence quality. Organizational Behavior and Human Decision Processes 56,

888

28–55.

46

ACCEPTED MANUSCRIPT

Katic, P., Morris, J., 2016. Targeting investments in small-scale groundwater

890

irrigation using Bayesian networks for a data-scarce river basin in Sub-

891

Saharan Africa. Environmental Modelling & Software 82, 44-72.

892

Kattwinkel, M., Reichert, P., 2017. Bayesian parameter inference for

893

individual-based models using a Particle Markov Chain Monte Carlo method.

894

Environmental Modelling & Software 87, 110-119.

895

Kondolf, G.M., Gao, Y., Annandale, G.W., Morris, G.L., Jiang, E., Zhang, J. et al.,

896

2014. Sustainable sediment management in reservoirs and regulated rivers:

897

Experiences from five continents. Earth’s Future 2, 256–280.

898

Lichtenstein, S., Fischhoff, B., Phillips, L.D., 1982. Calibration of probabilities:

899

The state of the art to 1980. Cambridge University Press, New York.

900

Luedeling, E., Goehring, L., 2017. DecisionSupport – quantitative support of

901

decision making under uncertainty. Contributed package to the r

902

programming language. Version 1.103.2.

903

Luedeling, E., Shepherd, K., 2016. Decision-focused agricultural research.

904

Solutions 7, 46–54.

SC

M AN U

TE D

AC C

EP

907 908

Frontiers in Environmental Science 3, 241.

905 906

RI PT

889

Luedeling, E., Oord, A.L., Kiteme, B., Ogalleh, S., Malesu, M., Shepherd, K.D. et al.,

2015. Fresh groundwater for Wajir ex-ante assessment of uncertain benefits for multiple stakeholders in a water supply project in northern Kenya.

47

Luedeling, E., Gassner, A., 2012. Partial Least Squares regression for analyzing

910

walnut phenology in California. Agricultural and Forest Meteorology 158, 43–

911

52.

912

Luedeling, E., Whitney, C., Lanzanova, D., 2018, Reservoir management in

913

Burkina Faso - decision analysis repository. Harvard Dataverse,

914

https://doi.org/10.7910/DVN/4RAKHX.

915

Morgan, M. G., 2014. Use (and abuse) of expert elicitation in support of

916

decision making for public policy. Proceedings of the National Academy of

917

Sciences of the United States of America 111 (20), 7176–7184.

918

Morris, G.L., Fan, J., 1998. Reservoir Sedimentation Handbook: Design and

919

Management of Dams, Reservoirs, and Watersheds for Sustainable Use.

920

McGraw-Hill Professional.

921

Maier, N.R.F., Hoffman, L.R., 1960. Using trained developmental discussion

922

leaders to improve further the quality of group decisions. Journal of Applied

923

Psychology 44, 247–251.

924

Martin, T.G., Burgman, M.A., Fidler, F., Kuhnert, P.M., Low-Choy, S., McBride, M.

925

927 928

communicating and incorporating scientific uncertainty in climate decision

929

making. A Report by the Climate Change Science Program and the

926

EP

TE D

M AN U

SC

RI PT

909

AC C

ACCEPTED MANUSCRIPT

et al., 2012. Eliciting expert knowledge in conservation science. Blackwell

Publishing Inc.

Morgan, M.G., 2009. Best practice approaches for characterizing,

48

Subcommittee on Global Change Research. National Oceanic and Atmospheric

931

Administration, Washington, DC, 96 pp.

932

Oakley, J.E., Brennan, A., Tappenden, P. and Chilcott, J., 2010. Simulation

933

sample sizes for Monte Carlo partial EVPI calculations. Journal of health

934

economics, 29, 468–477.

935

Omisore, A.G., 2017. Attaining Sustainable Development Goals in sub-Saharan

936

Africa; the need to address environmental challenges. Environmental

937

Development (In Press, Accepted Manuscript).

938

Palmieri, A., Shah, F., Dinar, A., 2001. Economics of reservoir sedimentation

939

and sustainable management of dams. Journal of Environnmental

940

Management 61, 149–163.

941

Pannell, D.J., Glenn, N.A., 2000. A framework for the economic evaluation and

942

selection of sustainability indicators in agriculture. Ecological Economics 33,

943

135–149.

944

Poussin, J.-C., Renaudin, L., Adogoba, D., Sanon, A., Tazen, F., Dogbe, W. et al.,

945

2015. Performance of small reservoir irrigated schemes in the Upper Volta

946

948 949

Macmillan Publishing Company.

950

Qian, S.S., Stow, C.A., Borsuk, M.E., 2003. On Monte Carlo methods for Bayesian

951

inference. Ecol. Model 159, 269-277.

947

EP

TE D

M AN U

SC

RI PT

930

AC C

ACCEPTED MANUSCRIPT

basin: Case studies in Burkina-Faso and Ghana. Water Resources and Rural Development 6, 50–65. Proakis, J.G., Manolakis, D.G., 1992. Digital Signal Processing. New York:

49

R Core Team, 2017. R: A language and Environment for Statistical Computing.

953

R Foundation for Statistical Computing, Vienna.

954

Rayner, S., 2012. Uncomfortable knowledge: The social construction of

955

ignorance in science and environmental policy discourses. Economy and

956

Society 41, 107–125.

957

Refsgaard, J.C., van der Sluijs, J.P., Højberg, A.L., Vanrolleghem, P.A., 2007.

958

Uncertainty in the environmental modelling process e a framework and

959

guidance. Environmental Modelling & Software 22, 1543-1556.

960

Renard, K.G., Foster, G.R., Weesies, G.A., McCool, D.K., Yoder, D.C., 1997.

961

Predicting soil erosion by water: a guide to conservation planning with the

962

revised universal soil loss equation (RUSLE). Agriculture Handbook 703,

963

United States Department of Agriculture.

964

Rosenstock, T.S., Mpanda, M., Rioux, J., Aynekulu, E., Kimaro, A.A., Neufeldt, H.

965

et al., 2014. Targeting conservation agriculture in the context of livelihoods

966

and landscapes. Agriculture, Ecosystems & Environment 187, 47–51.

967

Salas, J.D., Shin, H.S., 1999. Uncertainty analysis of reservoir sedimentation.

968

970 971

90, 265–277.

969

EP

TE D

M AN U

SC

RI PT

952

AC C

ACCEPTED MANUSCRIPT

Journal of Hydraulic Engineering 125, 339-350. Samantha, L.C., Rebecca, O., Kerrie, M., 2009. Elicitation by design in ecology: Using expert opinion to inform priors for bayesian statistical models. Ecology

50

ACCEPTED MANUSCRIPT

Sanfo, S., Gérard, F., 2012. Public policies for rural poverty alleviation: The

973

case of agricultural households in the plateau central area of Burkina Faso.

974

Agricultural Systems 110, 1–9.

975

Schleiss, A.J., Franca, M.J., Juez, C., De Cesare, G., 2016. Reservoir

976

sedimentation. Journal of Hydraulic Research 54 (6), 595-610.

977

Schmengler, A.C., 2011. Modeling soil erosion and reservoir sedimentation at

978

hillslope and catchment scale in semi-arid Burkina Faso. PhD thesis.

979

University of Bonn, Bonn, Germany.

980

Soll, J.B., Klayman, J., 2004. Overconfidence in interval estimates. Journal of

981

Experimental Psychology: Learning, Memory, and Cognition 30, 299–314.

982

Stirling, A., 2010. Keep it complex. Nature 468, 1029–1031.

983

Strong, M., Oakley, J.E., Brennan, A., 2014. Estimating multiparameter partial

984

expected value of perfect information from a probabilistic sensitivity analysis

985

sample: A nonparametric regression approach. Medical Decision Making 34,

986

311–326.

987

Tigrek, S., Göbelez, O., Aras, T., 2009. Sustainable management of reservoirs

990

AC C

EP

TE D

M AN U

SC

RI PT

972

991

41–53.

992

Tversky, A., Kahneman, D., 1974. Judgment under uncertainty: Heuristics and

993

biases. American Association for the Advancement of Science.

988 989

and preservation of water quality, in: El Moujabber, M., Mandi, L., TrisorioLiuzzi, G., Martín, I., Rabi, A., Rodríguez, R. (Eds.), Technological perspectives for rational use of water resources in the Mediterranean region, CIHEAM, Bari,

51

ACCEPTED MANUSCRIPT

Uusitalo, L., Lehikoinen, A., Helle, I., Myrberg, K., 2015. An overview of

995

methods to evaluate uncertainty of deterministic models in decision support,

996

Environmental Modelling & Software 63, 24-31

997

Venot, J.-P., Cecchi, P., 2011. Valeurs d’usage ou performances techniques :

998

Comment apprécier le rôle des petits barrages en Afrique subsaharienne ? Cah

999

Agric 20, 112–117.

Venot, J.-P., Hirvonen, M., 2013. Enduring controversy: Small reservoirs in sub-

1001

Saharan Africa. Society & Natural Resources 26, 883–897.

1002

Wafula. J., Karimjee, Y., Tamba, Y., Malava, G., Muchiri, C., Koech, G., De Leeuw,

1003

J., Nyongesa, J., Shepherd, K., Luedeling, E., 2018. Probabilistic Assessment of

1004

Investment Options in Honey Value Chains in Lamu County, Kenya. Frontiers

1005

in Applied Mathematics and Statistics 4, article 6, 1-11.

1006

Wischmeier, W.H., Smith, D.D., 1978. Predicting Rainfall Erosion Losses - a

1007

Guide to Conservation Planning.

1008

Wisser, D., Frolking, S., Douglas, E.M., Fekete, B.M., Schumann, A.H.,

1009

Vörösmarty, C.J., 2010. The significance of local water resources captured in

1010

1012 1013

Krawinkel, M., Tabuti, J.R.S., Luedeling, E. (2017). Impact pathways from

1014

agricultural development policy to household nutrition in Uganda. Manuscript

1015

under review.

1011

EP

TE D

M AN U

1000

AC C

SC

RI PT

994

small reservoirs for crop production – a global-scale analysis. Journal of

Hydrology 384, 264–275. Whitney, C.W., Lanzanova, D., Muchiri, C., Shepherd, K.D., Rosenstock, T.S.,

52

ACCEPTED MANUSCRIPT

Wold, S., 1995. PLS for multivariate linear modeling. In: van der Waterbeemd,

1017

H. (Ed.), Chemometric Methods in Molecular Design: Methods and Principles

1018

in Medicinal Chemistry, Verlag-Chemie, Weinheim, Germany, 195–218.

1019

Wood, S.A., Jina, A.S., Jain, M., Kristjanson, P., DeFries, R.S., 2014. Smallholder

1020

farmer cropping decisions related to climate variability across multiple

1021

regions. Global Environmental Change 25, 163–172.

1022

World Bank, 2017. World Bank data, consulted the 5th of December 2017.

1023

.

1024

Yaniv, I., 2004. The Benefit of Additional Opinions. Current Directions in

1025

Psychological Science 13(2), 75-78.

1026

Yaniv, I., Foster, D.P., 1997. Precision and accuracy of judgmental estimation. J.

1027

Behav. Decis. Making 10, 21–32.

1028

Zheng, Y., Keller, A.A., 2007. Uncertainty assessment in watershed-scale water

1029

quality modeling and management: 1. Framework and application of

1030

generalized likelihood uncertainty estimation (GLUE) approach. Water

1031

Resour. Res. 43, W08407.

AC C

EP

TE D

M AN U

SC

RI PT

1016

53

Highlights

RI PT SC M AN U TE D

• •

EP

•

A probabilistic decision model is used to compare interventions to prevent sedimentation of a reservoir in the Volta river basin. Decision Analysis is demonstrated as an approach that can generate robust actionable knowledge without the need to eliminate all uncertainty. Results show that sedimentation interventions are likely to be beneficial compared to inaction Combining interventions produces synergies and maximizes benefits of interventions

AC C

•

ACCEPTED MANUSCRIPT