SEARCH

SEARCH BY CITATION

Keywords:

  • Bayesian Belief Networks;
  • Development;
  • Environment;
  • Probabilistic models;
  • River basin management

Abstract

  1. Top of page
  2. Abstract
  3. EDITOR'S NOTE
  4. INTRODUCTION
  5. THE MODELING PROCEDURE: THE OUTLINE AND OUTSET
  6. VARIABLES FOR THE MODEL
  7. MODEL STRUCTURE
  8. SCENARIO ANALYSIS
  9. DISCUSSION AND CONCLUSIONS
  10. Acknowledgements
  11. REFERENCES

The use of Bayesian Belief Networks (BBNs) in modeling of environmental and natural resources systems has gradually grown, and they have become one of the mainstream approaches in the field. They are typically used in modeling complex systems in which policy or management decisions must be made under high uncertainties. This article documents an approach to constructing large and highly complex BBNs using a matrix representation of the model structure. This approach allows smooth construction of highly complicated models with intricate likelihood structures. A case study of the Ganges river basin, the most populated river basin of the planet, is presented. Four different development scenarios were investigated with the purpose of reaching the Millennium Development Goals and Integrated Water Resources Management goals, both promoted by the United Nations Agencies. The model results warned against the promotion of economic development policies that do not place strong emphasis on social and environmental concerns. Integr Environ Assess Manag 2012; 8: 491–502. © SETAC


EDITOR'S NOTE

  1. Top of page
  2. Abstract
  3. EDITOR'S NOTE
  4. INTRODUCTION
  5. THE MODELING PROCEDURE: THE OUTLINE AND OUTSET
  6. VARIABLES FOR THE MODEL
  7. MODEL STRUCTURE
  8. SCENARIO ANALYSIS
  9. DISCUSSION AND CONCLUSIONS
  10. Acknowledgements
  11. REFERENCES

This paper represents 1 of 7 review and case study articles generated as a result of a workshop entitled “Scenario and decision analysis in environmental management using Bayesian Belief Networks” (1–2 October 2009, Oslo, Norway) hosted by the Norwegian Institute for Nature Research (NINA) and the Strategic Institute Project “Nature 2020+” and funded by the Research Council of Norway. The main aim of the workshop was to compare Bayesian network applications to different environmental and resource management problems from around the world, identifying common modeling strategies and questions for further research.

INTRODUCTION

  1. Top of page
  2. Abstract
  3. EDITOR'S NOTE
  4. INTRODUCTION
  5. THE MODELING PROCEDURE: THE OUTLINE AND OUTSET
  6. VARIABLES FOR THE MODEL
  7. MODEL STRUCTURE
  8. SCENARIO ANALYSIS
  9. DISCUSSION AND CONCLUSIONS
  10. Acknowledgements
  11. REFERENCES

Many of the basic ideas of Bayesian decision theory are rooted in the frequentist probability theory that prevailed in the 1930s and 1940s. Von Neumann and Morgenstern (1944) visualized decision problems using a decision tree, and developed the axiomatization of many basic concepts of utility theory that have become cornerstones of Bayesian decision theory. These concepts became truly Bayesian somewhat later through advancements in mathematical and statistical theory by Wald (1950), Blackwell and Girshick (1954), Savage (1954), and Luce and Raiffa (1957). Many of these concepts have been summarized and popularized by Howard (1968), North (1968), and Raiffa (1968). Statistics is seen to be persistently split into 2 camps—frequentists and Bayesians—and this split still largely prevails and has a history of more than 250 years (Efron 2005; Little 2006). The evolution of Bayesian concepts in the 1940s and 1950s further reinforced the split of statistical decision theory (Shafer 1990).

In the past several decades, most Bayesian decision theory applications have emerged in automated decision making in which large amounts of observation data are being used on-line to adjust model predictions and optimizing systems in the fields of navigation, chemical engineering, manufacturing processes, and pattern recognition (Shafer 1990). A specific cluster of decision theory applications is directed at the acquisition of knowledge and information from experts in strategic decision making on complex systems where observational data are not as readily available as in the operational settings mentioned above, and where the reasoning leading to decisions has a strong social and human dimension and context. Such systems—also addressed in this article—have most typically been used to support consensus making and to facilitate learning processes in the identification of complex, social and economic systems (Shachter 1988; Shafer 1990; Varis and Fraboulet-Jussila 2002; Borsuk et al. this issue2012; Henriksen et al. this issue2012; Johnson et al. this issue2011; Uusitalo et al. this issue2011).

The computer implementation of Bayesian knowledge-acquisition tools started in the late 1970s (e.g., Leal and Pearl 1977; Humphreys and McFadden 1980). Facilitated by the development of computational algorithms and architectures (such as Shachter 1986; Pearl 1986; Lauritzen and Spiegelhalter 1988), the application software has been evolving and represents concurrently a wide selection of tools for various purposes from statistical-mathematical package add-ins to stand-alone software, and from packages of source codes on languages such as Java and C++ to web-based toolsets (Korb and Nicholson 2004; Murphy 2007).

The Bayesian Belief Network (BBN) methodology belongs to the branch of Bayesian decision theory approaches with a strong focus on knowledge acquisition. This methodology has gradually become one of the mainstream approaches in environmental modeling and assessment (McCann et al. 2006; Nyberg et al. 2006; Uusitalo 2007). The methodology has been particularly popular in modeling complex interrelationships between socioeconomic and biophysical systems in which the uncertainties are high, and the structure of the reasoning is of particular interest (Varis and Kuikka 1999; McCann et al. 2006; Kragt et al. 2011).

This article documents 1 specific approach to constructing BBNs for the analysis of highly complex natural resources and environmental management problems. The theoretical basis for the Fully Connected Belief Networks (FC BeNe) approach lies in the algorithms of Pearl (1986, 1988), and specifically in the extensions to that methodology by Varis (1998). The approach, as well as the software implementation, has further been developed through a number of case studies in river basin management (Varis and Fraboulet-Jussila 2002; Varis and Lahtela 2002; Varis and Keskinen 2006), fisheries and game management (Varis and Kuikka 1997b; Pellikka et al. 2005), and climate change impact assessment (Kuikka and Varis 1997; Varis and Kuikka 1997a). Instead of a graph interface, the model structure is being constructed by filling up a spreadsheet interfaced matrix of interlinkages.

Here the computational details are not documented because they have been published earlier (Varis 1998). Instead, we present an itemized description of the modeling procedure in a specific case study, followed by a Discussion and Conclusion section on the application of BBNs in integrated environmental assessment and management, and on the proposed approach in particular.

THE MODELING PROCEDURE: THE OUTLINE AND OUTSET

  1. Top of page
  2. Abstract
  3. EDITOR'S NOTE
  4. INTRODUCTION
  5. THE MODELING PROCEDURE: THE OUTLINE AND OUTSET
  6. VARIABLES FOR THE MODEL
  7. MODEL STRUCTURE
  8. SCENARIO ANALYSIS
  9. DISCUSSION AND CONCLUSIONS
  10. Acknowledgements
  11. REFERENCES

Technically, the FC BeNe modeling approach differs substantially from those conventional BBN models, which are operated through a graphical interface consisting of nodes and arcs. A modeling exercise using the approach applied here starts from a “tabula rasa” (blank slate) model, in which all model variables are technically connected with all other variables, but these connections are noninformative, i.e., a change in 1 variable does not influence the other variables. Also the variables are, at the outset, empty of all information, in Bayesian terms they are represented by a uniform, noninformative probability distribution. Within the modeling procedure, the tabula rasa is being filled with information on the variables and their interconnections, typically during expert panel sessions.

In addition to basic analyses, comprehensive sensitivity analyses can be carried out, including studies of the roles of the information sets obtained for the variables under study, as well as the roles of the causalities between the variables. The steps of the modeling procedure are shown in Figure 1.

thumbnail image

Figure 1. The proposed modeling procedure.

Download figure to PowerPoint

The authors of this article acted as the basis of the expert panel and worked out the model in numerous brainstorming sessions that took place over a period of 2 years. In several sessions, other topical experts also were invited to offer their views on certain specific topics. The total time allocation for this study—excluding the background studies—was the equivalent of approximately 10 working months.

The Ganges Basin and goals for its development

The goal of the case study was to analyze the policy options for the Ganges river basin in South Asia to meet 2 sets of internationally agreed development goals. The basin is by any measure one of the planet's most challenging and problematic river basins (Biswas et al. 2005; Varis et al. 2011). This world's most populated river basin is shared by India, Nepal, Bangladesh, and to a marginal extent by China. The magnitude of the development challenges faced by these countries (Table 1) reveals that decisive and comprehensive steps need to be taken toward achieving overall development in the region.

Table 1. Selected social, economic and environmental indicators for the Ganges basin riparian countriesa
Relevant to MDGIndicatorCountry
BangladeshIndiaNepal
  • a

    Data taken from UNDP (2006; 2009).

  • b

    GDP = gross domestic product; MDG = Millennium Development Goals; PPP = purchasing power parity.

 Population (millions, 2007)157.8116428.3
 Expected population (millions, 2015)168.21260.432.7
 Surface area (1000 km2)1443287147
 GDP per capita (PPP US $) (2007)124127531049
 Urban population (%) (2004)24.728.515.3
1Population below income poverty line (%) (1 US $/d) (2004)3634.724.1
2Children reaching grade 5 (%) (2003)657967
3Ratio of female to male enrollments in primary, secondary and tertiary schools (%) (2004)585852
4–6Life expectancy at birth (2004)63.663.362.1
4–7Access to improved water source (% of population) (2004)748690
4–7Access to improved sanitation facilities (% of population) (2004)393335
7CO2 emissions per capita (metric tons) (2003)0.31.20.1
8Aid % of GDPb (2004)2.50.16.4

As water is intertwined with almost all conceivable parts of life in the Ganges basin, and because water is a highly political issue in the region, it is highly justified to approach water management targets in such a situation by using criteria that are as widely accepted as possible. Therefore, the United Nations Millennium Development Goals (MDGs) and the United Nations promoted Integrated Water Resources Management (IWRM) objectives were chosen to constitute the goal-set of this analysis, as they are relevant within the water management and development context of the Ganges basin. These goals are globally accepted, and international communities have pledged to accomplish them worldwide. They are explained in brief below.

In the year 2000, the United Nations adopted the 8 MDGs (Table 2) that should be met by the year 2015. Accordingly, the leaders from virtually all 191 UN member states vowed to cooperate to achieve these goals (UN 2000, 2011).

Table 2. The United Nations Millennium Development Goals
NumberTitle
1Eradicate extreme poverty and hunger: reduce by half the proportion of people living on less than 1 dollar per day.
2Achieve universal primary education: ensure that all boys and girls complete a full course of primary schooling.
3Promote gender equality and empower women: eliminate gender disparity in primary and secondary education, preferably by 2005, and at all levels by 2015.
4Reduce child mortality: reduce by two-thirds the mortality among children under 5.
5Improve maternal health: reduce by 75% the maternal mortality ratio.
6Combat HIV/AIDS, malaria and other diseases: halt and begin to reverse the spread of HIV/AIDS, the incidence of malaria and other major diseases.
7Ensure environmental sustainability: integrate the principles of sustainable development into country policies and programs, reverse loss of forests, halve the proportion of people without access to improved drinking water, and improve the lives of slum dwellers.
8Develop a global partnership for development. Develop further an open trading and financial system with a commitment to good governance, address the special needs of the least developed and land-locked countries, deal comprehensively with the developing countries' debt problems, ensure employment for youth and share the benefits of new technologies, and so forth.

The United Nations has equally strongly promoted the implementation of IWRM over several decades (Rahaman and Varis 2005). The GWP (2000) defined IWRM as “…a process that promotes the coordinated development and management of water, land and related resources to maximize the resultant economic and social welfare in an equitable manner without compromising the sustainability of vital ecosystems.” It also emphasized that water should be managed in a basinwide context, under the prevalence of good governance and public participation (Rahaman and Varis 2005). The 3 key strategic objectives of IWRM are as follows:

  • 1)
    Economic development: waters should be used to provide economic well-being.
  • 2)
    Social development: allocation of water across different social and economic groups in an equitable manner to ensure social development.
  • 3)
    Environmental sustainability: protect the water resources base and associated ecosystems.

Background studies

A comprehensive set of background analyses and reviews of the status and future development options of the Ganges basin was carried out.

First, a comprehensive, regional analysis of the implementation of IWRM was performed for the entire South and Southeast Asia (Biswas et al. 2005). This included an analysis of a series of country studies as well as river basin studies. The critical drivers for water resources management in the region have been summarized by Varis (2005). Rahaman (2009a, 2009b) analyzed the role of the Ganges River in the regional development of South Asia, as well as the international collaboration, conflicts, and treaties related to the river. Rahaman and Varis (2009) extended the previous studies to include another giant river of South Asia—the Brahmaputra—that actually joins the Ganges before emptying into the Bay of Bengal and constitutes along with the Meghna basin the enormous Ganges-Brahmaputra-Meghna system, home to more than 650 million people. For these background studies, data were collected from both primary and secondary sources. Primary data and information were collected from relevant organizations and experts in the study area (Nepal, Bangladesh, India, and Bhutan).

These analyses revealed that water management is heavily interlinked with institutional development as well as with different sectors, and the integration between the different sectoral policies is key to achieving overall development in the region. In addition, social, political, environmental, and economic challenges were scrutinized, and they appeared relatively similar in all the riparian countries. Therefore, the Ganges basin is hereafter regarded as 1 unit, according to the philosophy of IWRM.

VARIABLES FOR THE MODEL

  1. Top of page
  2. Abstract
  3. EDITOR'S NOTE
  4. INTRODUCTION
  5. THE MODELING PROCEDURE: THE OUTLINE AND OUTSET
  6. VARIABLES FOR THE MODEL
  7. MODEL STRUCTURE
  8. SCENARIO ANALYSIS
  9. DISCUSSION AND CONCLUSIONS
  10. Acknowledgements
  11. REFERENCES

The first part of the model construction was to identify a plausible set of variables to describe the Ganges basin and its regional development context. This was a long exercise, with plenty of trials and errors. The variable set should be comprehensive yet compact, whereas the variables should be intuitively clear and analytically unambiguous.

Four variable clusters—driving forces, policy tools, impacts, and goals—were identified. Driving forces are factors that are under a low level of control and that create the settings for the water and sector policies. Policy tools are active measures that can be used to affect the system. Together, these clusters have impacts—either positive or negative—on the environment and socioeconomy. The goals, namely the IWRM goals and the MDGs, are at the end of the network and mainly affected by the impacts.

In total, there are 47 variables (Table 3). They consist of 6 driving forces, 22 policy tools, 10 impact variables, and 9 development goals. MDGs 4 to 6 are considered 1 variable, because all of them deal with health issues.

Table 3. The variables of the Ganges Water Policy Modela
Variable groupVariable sub-groupVariable
  • a

    The variables are listed in the causal sequence as used in the model.

Driving forces Global climate change
  Human development
  Economy and globalization
  Population growth
  Urbanization
  Political instability and vulnerability
Policy tools: GeneralHuman dimensionEducation
  Population policy
  Poverty reduction
  Empowerment
 InstitutionsDevelopment of formal institutions
  Development of informal institutions
  International commitments, regional cooperation, opening for globalization
  Decentralization
 IndustriesAgriculture, forestry, and fisheries (market)
  Agriculture, forestry, and fisheries (subsistence)
  Industry
  Services
Policy tools: WaterWater demandInstitutional and managerial means
  Economical means
  Information, education, and communication
  Regulatory and legislative means
  Technical means
 Water supplyWater storage and hydropower
  Water transfer systems
  Supply, sanitation, treatment
  Nonconventional water supply
  Catchment management
ImpactsEnvironmentConsumption, production, opening material cycles
  Loss of ecosystems and biodiversity
  Land degradation
  Surface water degradation
  Groundwater degradation
 SocioeconomyGender equality
  Food security
  Public health problems
  Poverty
  National economy
GoalsMillennium development goals1. Eradicate extreme poverty and hunger
  2. Achieve universal primary education
  3. Promote gender equity and women empowerment
  4–6. Health issues
  7. Ensure environmental sustainability
  8. Develop a global partnership for development
 IWRM objectivesEconomic growth
  Social development
  Environmental sustainability

Causal sequence of variables

Because the used computational algorithm is based on a directed graph approach, the variables must be organized to form a sequence (Pearl 1988). This usually requires some elaboration to assure that as little information as possible is lost in the model for 2 reasons. First, the algorithm cannot evaluate loops; and second, propagating information back and forth (i.e., Node 4 [RIGHTWARDS ARROW] Node 5 [RIGHTWARDS ARROW] Node 3) cannot be performed precisely (instead, the information from Node 3 [RIGHTWARDS ARROW] Node 4 [RIGHTWARDS ARROW] Node 5 or from Node 5 [RIGHTWARDS ARROW] Node 4 [RIGHTWARDS ARROW] Node 3 is propagated precisely because the direction of the message does not change).

Prior probabilities

The state variables and driving forces typically have a certain prior condition and tendency to change in a certain direction. The present situation was used as a reference situation: whether the issue will grow, decrease, stay unchanged, and to what extent. Consequently, the variables of the model were defined as rates of change. A prior probability distribution was assigned to the rate of change of each variable, which is sharp if we know this variable well, but flat if it is poorly known, or is subjected to variation that is beyond the resolution of the model.

For policy tools, mathematically similar prior probability distributions are being used for decision outcomes in this approach. In such a situation, a sharp probability distribution implies that the corresponding decision variable is well under control, whereas a flat distribution tells that it is poorly controllable.

In the Ganges model, the prior probability distributions are calculated from the input marks shown in Table 5, knowing the interpretation shown in Table 4.

Table 4. The 4 development scenarios and general sector policies investigated with the Ganges Water Policy Modela
Ganges BasinEnvironmentEconomic growthHuman developmentIntegrated
MeanAccuracyMeanAccuracyMeanAccuracyMeanAccuracy
  • a

    See Table 5.

Education>***>***>>**>***
Population policy>*******>**>>**
Poverty reduction>>***<***>>>***>***
Empowerment>>***<***>>***>>**
Development of formal institutions>***>>***>>**
Development of informal institutions>**>**
International commitments, regional cooperation and opening for globalization>>***>***>**>***
Decentralization>****>>>**>>**
Agriculture, forestry, and fisheries (market)<<**>>>***>**>>***
Agriculture, forestry, and fisheries (subsistence)>***<<***>>>***>****
Industry<<**>>>***>***>***
Services**>>***>**

MODEL STRUCTURE

  1. Top of page
  2. Abstract
  3. EDITOR'S NOTE
  4. INTRODUCTION
  5. THE MODELING PROCEDURE: THE OUTLINE AND OUTSET
  6. VARIABLES FOR THE MODEL
  7. MODEL STRUCTURE
  8. SCENARIO ANALYSIS
  9. DISCUSSION AND CONCLUSIONS
  10. Acknowledgements
  11. REFERENCES

The second part of the model construction is to define the structure of the model. In this approach, a numerical value, ranging from −1 to +1 is assigned to describe the strength of the link between each pair of 2 nodes. These link strength parameters indicate how much a change in one variable influences another variable. As interpreted through a statistical analogy, it implies the proportion of the variance of each variable explained by the variance of the conditioning variable (analogous to Pearson's R2, which is additive, unlike, for instance, R), yet with causal interpretation of direction of influence. Accordingly, the sum of absolute values of all the link strengths heading to a variable may not exceed 1. The entire link matrix of the Ganges model contains 1081 (= (47 × 46)/2) elements. Therefore, it is not worthwhile to present the matrix and describe all its elements in detail in a journal article. However, as an illustration, a piece of the link matrix is shown in Figure 2. With this formulation of links and prior probability distributions, normality is assumed, but the computational approach is by no means restricted to normality. Other formulations are also possible (Varis 1998).

thumbnail image

Figure 2. The principle of the link matrix with an example draw out with selected comments attached to specific link strength values.

Download figure to PowerPoint

Consistency iteration

The posterior probability of a variable should not be much flatter than its prior probability; if it is, it might indicate inconsistency in the assessment. The states of the other variables and their connections are not fully in accord with the assumption of the state of the variable concerned. A sharper posterior distribution indicates that the information content of the entire model allows a decrease in the uncertainty of the variable under consideration. In general, an inconsistent link matrix flats out the posterior distributions quite easily.

Although being entirely different, the link matrix and its inconsistency analysis properties are sometimes mixed with some multiple criteria decision-making approaches (e.g., Saaty 1980), in which 2 variables are subjectively compared with a numerical scale, and the consistency of the assessment of a set of multiple variables can be analyzed by using eigenvalues of the resulting intercomparison matrix.

SCENARIO ANALYSIS

  1. Top of page
  2. Abstract
  3. EDITOR'S NOTE
  4. INTRODUCTION
  5. THE MODELING PROCEDURE: THE OUTLINE AND OUTSET
  6. VARIABLES FOR THE MODEL
  7. MODEL STRUCTURE
  8. SCENARIO ANALYSIS
  9. DISCUSSION AND CONCLUSIONS
  10. Acknowledgements
  11. REFERENCES

Scenarios

The Ganges model was used to analyze 3 characteristic development scenarios and for constructing a feasible compromise between them, aimed at better meeting the development goals set for the exercise than any of the original characteristic scenarios. These scenarios consisted of 3 classical corners of sustainable development and of their combination, i.e., they were the ones in which either economic development, human development, environmental concerns, or their “integrated” mix would be prioritized (Table 5). All scenarios were again constructed by the expert panel and were essentially based on the aforementioned background studies.

Table 5. Semiquantitative scale used for scenario analysis and future tendencies
MeansAll variablesAccuracyPolicy variables: level of controlDriving forces and impacts: level of knowledge
 No change No controlUnknown
>Small increase*20% under control20% known
>>Modest increase**40% under control40% known
>>>Large increase***60% under control60% known
<Small decrease****80% under control80% known
<<Modest decrease*****Fully under controlFully known
<<<Large decrease

The integrated scenario is not an absolute mean of the other 3 scenarios but rather a compromise between them, seeking solutions in which none of the development goals are ignored (Varis and Lahtela 2002; Varis and Keskinen 2006). The amount and direction of change were chosen by iteration, whereas the accuracy reveals the level of control of each sector policy that such a change would actually take place. The results of the model are summarized in Tables 6 and 7.

Table 6. Results of the 4 development scenarios analyses with respect to the MDGsa
MDGTitleResults
  • a

    MDG = Millennium Development Goals.

1Eradicate extreme poverty and hungerIntegrated scenario offers best improvement with the least uncertainty. The economic development scenario is giving the worst results both in terms of effectiveness and uncertainties.
2Universal primary educationThe difference between human development scenario and integrated scenario is minor. Integrated scenario offers less uncertainty.
3Promote gender equality and empower womenDifferences between environmental, human development and integrated scenarios are minor. All these scenarios offer positive development in the achievement of MDG 3. Economic growth scenario offers the least development with highest uncertainty. The uncertainty in the integrated scenario is the lowest and is therefore considered the most desirable option.
4–6Health issuesThe environmental, human development, and integrated scenarios have an almost equal effect. However, again, the economic scenario offers the lowest improvement and highest uncertainty. The integrated scenario offers the lowest uncertainty, which implies that integrated scenario is less risky and would be a better choice in relation to these goals.
7Ensure environmental sustainabilityThe 4 scenarios indicate no major differences. However, the environmental scenario offers only a minor improvement. The lesson learned from this output is that ensuring environmental sustainability is quite difficult, if not impossible. Overallsocio economic development is needed to even maintain the current status.
8Develop global partnership for developmentAll the 4 scenarios offer positive development. The integrated scenario offers better results than any of the 3 basic scenarios both in terms of effectiveness and certainties.
Table 7. Results of the 4 development scenarios with respect to the 3 IWRMa goals
Goals/scenarioEnvironmental scenarioEconomic growth scenarioHuman development scenarioIntegrated scenario
  • a

    IWRM = Integrated Water Resources Management.

Economic growthLowest average and highest uncertainty. It reveals that environmental sustainability and economic growth are not going hand in hand.Higher improvement in relation to other scenarios. But uncertainty is higher in comparison to human development and integrated scenarios.Showing slightly lower improvement in relation to economic growth scenario. But uncertainty is lower than economic growth scenarioAlthough improvement is slightly lower than economic growth scenario, the uncertainty is lowest in relation to the 3 basic scenarios.
Social developmentEnvironmental scenario shows an almost equal average to the human development scenario.Offering lowest average and highest uncertainty.Offering best improvement in comparison to the 2 other basic scenarios.Highest average and lowest uncertainty in comparison with the 3 basic scenarios.
Environmental sustainabilityAll 4 scenarios give negative results and uncertainty is very high. Naturally, the environmental scenario implies less environmental deterioration.

Sensitivity analysis

The model was constructed to include a possibility to carry out an interactive sensitivity analysis. In such an analysis, a small perturbation is imposed on the prior probability distribution of each of the sector policies, and the response can be observed graphically. We introduced a 1% perturbation on the average of a probability distribution. This approach is analogous to the Hessian matrix from optimization theory.

The sensitivity plots show the sensitivity of the MDGs and IWRM goals to changes in sector policies. They also show how much the realization of these development goals responds to small changes in different sector policies. The examples shown in Figures 3 and 4 reveal the sensitivity of the IWRM goals to the driving forces and sector policies for the integrated scenario. The sensitivity of each development goal to the sector policies can be observed graphically for all 4 scenarios analyzed with this model.

thumbnail image

Figure 3. An illustration of the sensitivity analysis approach. Sensitivity of the 3 IWRM goals to the driving forces and general sector policies (integrated scenario). The further the line is away from zero (the circle notated with 0.0%), the more impact the policy has on the respective development goals. The impact is positive if the line goes outside the circle and vice versa.

Download figure to PowerPoint

thumbnail image

Figure 4. An illustration of the sensitivity analysis approach. Sensitivity of the 3 IWRM goals to the water sector policies (integrated scenario).

Download figure to PowerPoint

The sensitivity plots change interactively when the scenario is modified. This provides important guidance on how the scenario could be modified to make it better with respect to any of the 8 MDGs and 3 IWRM goals. Some comments are listed below on the basis of sensitivity analyses (Figures 3 and 4) with respect to the policy combinations in terms of their efficacy in achieving the 3 IWRM goals:

  • 1)
    Economic growth: this objective would be best achieved by emphasizing policies such as industries, services, agriculture, forestry and market fisheries, water storage and hydropower development, and development of formal institutions. To avoid possible conflict with the environmental sustainability objective, careful consideration should be given to some of these policies, e.g., water storage and hydropower, as well as industries.
  • 2)
    Social development: investing in education, population policy, poverty reduction, empowerment and water supply, sanitation and treatment policies would bring positive development. However, political stability and human development are fundamental to implementing these policies.
  • 3)
    Environmental sustainability: the important policies in relation to this objective would be education, population policy, poverty reduction, catchment management, nonconventional water supply, water supply, sanitation and treatment, development of formal institutions, international commitments, and regional cooperation.

DISCUSSION AND CONCLUSIONS

  1. Top of page
  2. Abstract
  3. EDITOR'S NOTE
  4. INTRODUCTION
  5. THE MODELING PROCEDURE: THE OUTLINE AND OUTSET
  6. VARIABLES FOR THE MODEL
  7. MODEL STRUCTURE
  8. SCENARIO ANALYSIS
  9. DISCUSSION AND CONCLUSIONS
  10. Acknowledgements
  11. REFERENCES

The development of the Ganges Water Policy Model aimed at providing systematization of the various factors that are influential in water policy making toward meeting the MDGs and IWRM objectives in the Ganges basin. Reality is always far more complicated than computer models, however, and the models can only foster policy dialogue with systematic structuring and analysis of excessively complicated entities such as the one presented here. In particular, the model allows an approach that takes greater account of the broader issues and constraints faced by water managers due to the multidimensional, multifaceted task of managing water. Equally important is the capability of handling very high uncertainties and communicating these to policy makers. Such a capability is missing from many contemporary scenario modeling techniques.

Decision support and policy advantages of explicit modeling of uncertainty

Multidisciplinary and holistic approaches are needed to analyze complex policy making tasks in environmental and natural resources management in a transparent and most equitable way. An approach to using the BBN methodology in a systematic trade-off and scenario analysis in such situations was examined. It was used as part of a complex scrutiny that yielded a set of policy recommendations.

The analysis showed clearly the advantages of explicit modeling of uncertainty in a highly complicated case for the management of environment and natural resources. Awareness and understanding of the interconnectedness and high level of uncertainty of the various factors relevant to water policy making in the gigantic river basin in question can be improved through a systematic BBN analysis such as the present one. An approach such as this can be useful in guiding experts and policy makers to consider the complexity of the issues systematically and in a coherent fashion, and it yields policy-relevant results. The models are among the various means to reach the ends, which are a set of balanced results of an integrated analysis, aiming at benefitting water resources development in as sustainable a manner as possible.

Merits of the proposed approach

When analyzing highly complex and multidisciplinary systems such as the development of the Ganges river basin, the use of a Bayesian model fosters consistent thinking and argumentation and serves as a plausible platform for constructing a coherent view of a problem which is conceptually quite difficult to comprehend due to its colossal dimension.

The applied FC BeNe approach is quite efficient in this regard and allows a systematic and versatile model development procedure for the analysis of complex models. This can be easily done as an interactive process, because the software, which is a tailored spreadsheet interfaced shell, can update the whole network of the size used in the Ganges model in only fractions of a second. The approach is thus ideal for interactive learning tasks, expert panel work, or similar applications. Moreover, inclusion of statistical information, deterministic equations, and other types of information are possible (Varis 1998).

The approach is quite flexible in comparison to the conventional graphically interfaced BBN software and does not include the task of filling up the cumbersome conditional probability tables, as assumptions for the type of the probability distribution can be used and not just discrete probability tables. The Ganges model uses the Gaussian assumption but other probability distributions can also be used.

Computationally, the FC BeNe approach is based on Pearl's algorithms (1986, 1988). Considerable extensions addressing environmental and natural resources management situations were provided by Varis (1998), and a further development of the multidimensional sensitivity analysis property was introduced by Varis and Keskinen (2006). The approach includes several innovations that set it apart from the mainstreamed graph interfaced approaches in several ways, as documented above.

Bayesian modeling strategies specific to the environmental management problem

A modeling strategy for addressing problems such as the present one would start with a profound investigation of the environmental management situation under study. In the Ganges case, this was an extensive undertaking that was reported in a number of scientific books and articles. In parallel, a body of experts is needed for the modeling procedure to be sound. The involvement of stake holders and policy makers would be important at a later stage.

Certain impact on policy dialogue has been achieved in the Ganges case, notably in Bangladesh, and the role of background studies has been particularly important. In some earlier cases of similar nature (Varis and Fraboulet-Jussila 2002; Varis and Lahtela 2002; Varis and Keskinen 2006) the policy impact has become essentially bigger. The approach and the strategy applied are not particularly specific to cases such as the present one as the experience from fisheries management (Kuikka and Varis 1997), climate change adaptation (Varis and Kuikka 1997a), and game management (Pellikka et al. 2005) reveal.

Challenges of the use of BBNs

Whereas the BBNs offer a wealth of possibilities and properties uncommon for most other modeling approaches, there are also certain factors that hinder the spread and acceptance of these approaches in the environmental and natural resources management fields. Above all, a majority of the experts tend not to be familiar with, or ready to learn the theoretical, philosophical, and computational basics of the approach. The Bayesian philosophy, the inference and propagation of information, which is not bound to model physical quantities, energy fluxes etc., and the analytical formulation of subjective and uncertain information all constitute boundaries between Bayesian and nonBayesian modelers.

This split between statistics and users of statistical models to frequentists and Bayesians (Savage 1961; Shafer 1990; Howson and Urbach 1991; Efron 2005; Little 2006) is clearly visible in our field. The boundaries are high and challenging when communicating the models and their outcomes to end users, or when incorporating them into the knowledge elicitation procedures. Many experts and end users of environmental models readily accept highly complicated and in many parts highly subjective model formulations when they are hidden behind so-called, physically based models, consisting typically of a large set of partial differential equations and having dozens of more or less ambiguous, nonidentifiable parameters. However, the same individuals tend to reject approaches in which the uncertainty and subjectivity are explicitly and transparently formulated as model components.

Interestingly, many Bayesian modelers consider physically based simulation models with the above mentioned features closer to Bayesian models than frequentist ones, and categorize, in the philosophical sense, the classical physics-based modelers such as Laplace as forerunners of Bayesian methodology (Efron 2005) as the prior information (physics-based equations, nonmeasurable parameter values, etc.) is being widely used. Efron (2005) maintains that today's great expansion of data sets, modeling approaches and so forth press the developers of methodology to cross the ancient boundary between the frequentists and Bayesians. We would like to add that the societal and political importance and role of various statistical-mathematical inference approaches pushes in the same direction.

After all, no approach is superior to others in all respects, and a toolbox of a modeler in environmental and natural resources management field should incorporate several complementary approaches, including Bayesian tools. Being too one-sided in methodological terms is a great danger (Hopple 1986) for any modeler who deals with diverse and complex matters such as faced in integrated environmental assessment and management.

Acknowledgements

  1. Top of page
  2. Abstract
  3. EDITOR'S NOTE
  4. INTRODUCTION
  5. THE MODELING PROCEDURE: THE OUTLINE AND OUTSET
  6. VARIABLES FOR THE MODEL
  7. MODEL STRUCTURE
  8. SCENARIO ANALYSIS
  9. DISCUSSION AND CONCLUSIONS
  10. Acknowledgements
  11. REFERENCES

The authors thankfully acknowledge the financial support of the Academy of Finland Project 133748, and Maa-ja vesitekniikan tuki r.y. The authors would like to thank David Barton, Marko Keskinen, Virpi Stucki, Matti Kummu, Pertti Vakkilanen, Sakari Kuikka, Iswer Raj Onta, and M. Maniruzzaman Miah for their constructive comments.

REFERENCES

  1. Top of page
  2. Abstract
  3. EDITOR'S NOTE
  4. INTRODUCTION
  5. THE MODELING PROCEDURE: THE OUTLINE AND OUTSET
  6. VARIABLES FOR THE MODEL
  7. MODEL STRUCTURE
  8. SCENARIO ANALYSIS
  9. DISCUSSION AND CONCLUSIONS
  10. Acknowledgements
  11. REFERENCES
  • Biswas AK, Varis O, Tortajada C, editors. 2005. Integrated water resources management in South and South-East Asia. India: Oxford Univ Press. 353 p.
  • Blackwell D, Girshick MA. 1954. Theory of games and statistical decisions. New York (NY): Wiley. 355 p.
  • Borsuk M, Schweizer S, Reichert P. Forthcoming 2012. A Bayesian network model for integrative river rehabilitation planning and management. Integr Environ Assess Manag 8: 462472.
  • Efron B. 2005. Bayesians, frequentists, and scientists. J Am Stat Assoc 100: 15.
  • [GWP] Global Water Partnership. 2000. IWRM at a glance: Technical Advisory Committee. Stockholm, Sweden: Global Water Partnership Secretariat.
  • Henriksen HJ, Brugnach M, Zorrilla P, de la Hera A. 2012. Use of Bayesian belief networks for dealing with ambiguity in integrated groundwater management. Integr Environ Assess Manag 8: 430444.
  • Hopple GW. 1986. Decision aiding dangers: The law of the hammer and other maxims. IEEE Trans Syst Man Cybern 16: 948964.
  • Howard RA. 1968. The foundations of decision analysis. IEEE Trans Syst Sci Cybern 4: 211219.
  • Howson C, Urbach P. 1991. Bayesian reasoning in science. Nature 350: 371374.
  • Humphreys PC, McFadden W. 1980. Experiences with MAUD: Aiding decision structuring versus bootstrapping the decision maker. Acta Psychol 45: 5169.
  • Johnson S, Low Choy S, Mengersen K. 2011. Integrating Bayesian networks and geographic information systems: Good practice examples. Integr Environ Assess Manag 8: 473479.
  • Korb KB, Nicholson AE. 2004. Bayesian artificial intelligence. London (UK): Chapman and Hall/CRC. 364 p.
  • Kragt ME, Newham LTH, Bennett J, Jakeman AJ. 2011. An integrated approach to linking economic valuation and catchment modeling. Environ Model Software 26: 92102.
  • Kuikka S, Varis O. 1997. Uncertainties of climatic change impacts in Finnish watersheds: A Bayesian network analysis of expert knowledge. Boreal Environ Res 2: 109128.
  • Lauritzen SL, Spiegelhalter DJ. 1988. Local computations with probabilities on graphical structures and their application to expert systems. J R Stat Soc Series B 50: 157224.
  • Leal A, Pearl J. 1977. An interactive program for conversational elicitation of decision structures. IEEE Trans Syst Man Cybern 7: 368376.
  • Little RJ. 2006. Calibrated Bayes: A Bayes/frequentist roadmap. Am Stat 60: 213223.
  • Luce RD, Raiffa H. 1957. Games and decisions: Introduction and critical survey. New York (NY): Wiley. 509 p.
  • McCann RK, Marcot BG, Ellis R. 2006. Bayesian belief networks: Applications in ecology and natural resource management. Can J For Res 36: 30533062.
  • Murphy K. 2007. Software for graphical models: A review. Int Soc Bayesian Anal Bull 14: 1315.
  • North DW. 1968. A tutorial introduction to decision theory. IEEE Trans Syst Sci Cybern 4: 200210.
  • Nyberg JB, Marcot BG, Sulyma R. 2006. Using Bayesian belief networks in adaptive management. Can J For Res 36: 31043116.
  • Pearl J. 1986. On evidential reasoning in a hierarchy of hypotheses. Artif Intel 28: 915.
  • Pearl J. 1988. Probabilistic reasoning in intelligent systems: Networks of plausible inference. San Mateo (CA): Morgan-Kaufmann. 552 p.
  • Pellikka J, Kuikka S, Lindén H, Varis O. 2005. The role of game management on wildlife populations—Uncertainty analysis of expert knowledge. Eur J Wildl Res 51: 4859.
  • Rahaman MM. 2009a. Integrated Ganges basin management: Conflicts and hope for regional development. Water Policy 11: 168190.
  • Rahaman MM. 2009b. Principles of transboundary water resources management and Ganges treaties: an analysis. Int J Water Res Dev 25: 159173.
  • Rahaman MM, Varis O. 2005. Integrated water resources management: Evolution, prospects and future challenges. SSPP 1: 1521.
  • Rahaman MM, Varis O. 2009. Integrated water management of the Brahmaputra basin: Perspectives and hope for regional development. Nat Resour Forum 33: 6075.
  • Raiffa H. 1968. Decision analysis. Reading (MA): Addison-Wesley. 309 p.
  • Saaty T. 1980. The analytic hierarchy process. New York (NY): McGraw-Hill. 287 p.
  • Savage LJ. 1954. The foundations of statistics. New York (NY): Wiley. 294 p.
  • Savage LJ. 1961. The foundations of statistics reconsidered. Proceedings of the Fourth Berkeley Symposium on Mathematics and Probability. Berkeley (CA): Univ of California. p. 575586.
  • Shafer G. 1990. Decision making. In: Shafer G, Pearl J, editors. Readings in uncertain reasoning. San Mateo (CA): Morgan Kaufmann. p 6167.
  • Shachter RD. 1986. Evaluating influence diagrams. Oper Res 34: 871882.
  • Shachter RD. 1988. Probabilistic inference and influence diagrams. Oper Res 36: 589604.
  • [UN] United Nations. 2000. United Nations Millennium Declaration, A/55/L.2. New York (NY): United Nations.
  • [UN] United Nations. 2011. United Nations Millennium Development Goals. [cited 2011 February 4]. Available from: http://www.un.org/millenniumgoals
  • [UNDP] United Nations Development Programme. 2006. Human development report. New York (NY): UNDP.
  • [UNDP] United Nations Development Programme. 2009. Human development report. New York (NY): UNDP.
  • Uusitalo L. 2007. Advantages and challenges of Bayesian networks in environmental modeling. Ecol Model 203: 312318.
  • Uusitalo L, Bäck S, Söderkultalahti P, Kauppila P, Kuikka S. 2011. Assessing the roles of environmental factors in coastal fish production in the northern Baltic Sea: A Bayesian network application. Integr Environ Assess Manag 8: 445455.
  • Varis O. 1998. A belief network approach to optimization and parameter estimation: application to resource and environmental management. Artif Intel 101: 135163.
  • Varis O. 2005. Externalities of integrated water resources management in South and South-East Asia. In: Biswas AK, Varis O, Tortajada C, editors. Integrated water resources management in South and South-East Asia. India: Oxford Univ-India. p 1–38.
  • Varis O, Fraboulet-Jussila S. 2002. Water resources development in the lower Senegal River Basin: Conflicting interests, environmental concerns and policy options. Int J Water Res Dev 18: 245260.
  • Varis O, Keskinen M. 2006. Policy analysis for the Tonle Sap Lake, Cambodia—a Bayesian network model approach. Int J Water Resour Dev 22: 417431.
  • Varis O, Kuikka S. 1997a. BeNe-EIA: A Bayesian approach to expert judgment elicitation with case studies on climatic change impacts on surface waters. Clim Change 37: 539563.
  • Varis O, Kuikka S. 1997b. Joint use of multiple environmental assessment models by a Bayesian meta-model: The Baltic salmon case. Ecol Model 102: 341351.
  • Varis O, Kuikka S. 1999. Learning Bayesian decision analysis by doing: Lessons from environmental and natural resources management. Ecol Model 119: 177195.
  • Varis O, Kummu M, Salmivaara A. Forthcoming 2011. Ten major rivers in Asia-Pacific: An assessment of vulnerability. Appl Geogr 32: 441454.
  • Varis O, Lahtela V. 2002. Integrated water resources management along the Senegal River: Introducing an analytical framework. Int J Water Resour Dev 18: 501521.
  • von Neumann J, Morgenstern O. 1944. Theory of games and economic behavior. Princeton (NJ): Princeton Univ Press. 739 p.
  • Wald A. 1950. Statistical decision functions. New York (NY): Wiley. 179 p.