A review of model-based scenario analysis of poverty for informing

15 Ending poverty in all its forms everywhere is the first goal being targeted by the United Nations 16 2030 Agenda for Sustainable Development. Poverty eradication is a long-term process that 17 faces the challenges of many uncertainties and complex interactions with other Sustainable 18 Development Goals (SDGs). In order to better understand poverty and contribute to addressing 19 poverty in a sustainable manner, this paper aims to conduct a systematic review of model-based 20 analysis for poverty scenario in the context of SDGs. We first review 144 studies from the 21 perspectives of bibliometric information (i.e., publication types, research topics for poverty, 22 research objects, research scales and geographic locations) and models information for poverty 23 scenario analysis (i.e., model types, purposes, states, temporal and spatial range, sectors 24 considered, poverty and other SDGs indicators). Second, we discuss the pros and cons of 25 different types of models and identify seven representative models. We also discuss the 26 synergies and trade-offs between poverty and other SDGs. Finally, we identify four potential 27 research gaps in model-based poverty scenario analysis and provide suggestions for future 28 research. The review shows that poverty scenario analysis was carried out mainly from a single 29 perspective, such as economic, ecological, and agricultural. Few studies used effective models 30 to analyze poverty under an integrated interactions analysis of multiple sectors. Comprehensive 31 multi-sector models are needed for global and regional poverty scenario analysis over the 32 mediumor long-term to enhance the ability of analyzing the combined effects, synergies, and 33 trade-offs between poverty and a variety of other SDGs. 34


21
perspectives of bibliometric information (i.e., publication types, research topics for poverty, 22 research objects, research scales and geographic locations) and models information for poverty 23 scenario analysis (i.e., model types, purposes, states, temporal and spatial range, sectors 24 considered, poverty and other SDGs indicators). Second, we discuss the pros and cons of 25 different types of models and identify seven representative models. We also discuss the 26 synergies and trade-offs between poverty and other SDGs. Finally, we identify four potential 27 research gaps in model-based poverty scenario analysis and provide suggestions for future 28 research. The review shows that poverty scenario analysis was carried out mainly from a single 29 perspective, such as economic, ecological, and agricultural. Few studies used effective models 30 to analyze poverty under an integrated interactions analysis of multiple sectors. Comprehensive 31 multi-sector models are needed for global and regional poverty scenario analysis over the 32 medium-or long-term to enhance the ability of analyzing the combined effects, synergies, and 33 trade-offs between poverty and a variety of other SDGs.

Introduction 42
The United Nations 2030 Agenda for Sustainable Development commonly 43 known as the Sustainable Development Goals (SDGs) is committed to eradicating 44 poverty, protecting the planet, and ensuring peace and prosperity for humanity through 45 concerted actions (Cf, 2015). These SDGs are interdependent and interconnected, and 46 together state the shared aspirations for a more sustainable future. The first goal (SDG 47 1) of the 17 SDGs, ending poverty in all its forms everywhere, is strongly associated 48 with the well-being of every individual (United Nations, 2019). Although the 49 proportion of people living in extreme poverty (less than $1.9 a day based on 2011 50 Purchasing Power Parities (United Nations, 2019)) has fallen from 36% in 1990 to 10% 51 in 2015 globally, there are still more than 700 million people living in extreme poverty 52 (United Nations, 2019) where their essential living needs (e.g., water, sanitation, health 53 services, education) cannot be guaranteed. Poverty is still one of the most intractable 54 social problems and the most important livelihood problems faced by humanity (United 55 Nations, 2020). 56 To better understand poverty and evaluate progress towards SDG 1, researchers 57 have conducted both qualitative and quantitative analyses aiming to identify poverty 58 causes (B. W. Wang et al., 2019), measure the progress towards a set target (Vyas-59 Doorgapersad, 2018), understand linkages between poverty and other relevant factors 60 (Suich et al., 2015), and formulate or evaluate the effects of poverty reduction policies 61 (Alwang et al., 2019). However, poverty analysis, as in every other human-natural 62 system analysis (Moallemi et al., 2020), is fraught with challenges of uncertainty (i.e., 63 achieving SDG 1 is a long-term process that is vulnerable to external surprises and 64 shocks) and complexity (interconnections between poverty and other economic, social, 65 and environmental SDGs). 66 Model-based scenario analysis has been used to tackle these challenges in 67 research on poverty. Regarded as a powerful analytical method to support sustainable 68 development research (Swart et al., 2004), model-based quantitative scenario analysis 69 aims to project possible future trends or consequences under the premise that a 70 phenomenon could occur in the future with a certain likelihood, allowing policymakers 71 to explore alternative futures and to take into account their consequences for decision-72 making (Kosow and Gaßner, 2008). Different from traditional forecasting methods, it 73 emphasizes uncertainty instead of forecasting and works on the premise that there are 74 a variety of possible trends in the future, hence diverse results will be obtained. Scenario 75 analysis uses various sources of information and knowledge (e.g., experience and 76 knowledge of experts, uncertain future trends, and human behaviors) to generate a 77 series of internally consistent future scenarios, which involves highly uncertain long-78 term driving factors (e.g., demographics, climate change, and technological 79 development) and includes trends or non-linear interactions that may differ 80 significantly from past experiences. 81 Despite growing interest in model-based scenario analysis in dealing with 82 poverty (Allen et al., 2021;Laborde et al., 2021), the depth and breadth of this area and 83 opportunities for further studies have not been scoped so far. Here, we aim to fill this 84 gap by conducting a systematic review of model-based poverty scenario analysis, 85 mapping: (1) the topics addressed; (2) cataloging the quantitative models that have been 86 developed; (3) the indicators used to measure poverty; as well as identifying 87 representative models and research gaps in model-based quantitative poverty scenario 88 analysis. Based on this systematic review we synthesize the field of scenario analysis 89 for assessing poverty and chart a new research agenda for better integrating and 90 mainstreaming this critically important aspect of sustainability into modelling studies. The Scopus database is adopted for the literature search because of its broad 99 coverage in related research of poverty, SDGs, system dynamics, and scenario analysis, 100 and internet-accessible full-text resources available in related journals (Mio et Based on the  111 information above, we found 482 papers by the following search string. 112 • TITLE-ABS-KEY ("poverty" OR "sustainable development goal 1" OR "SDG 1") 113 AND TITLE-ABS-KEY ("scenario analysis" OR "scenario modeling" OR 114 (scenario AND model) OR (scenario AND modeling) OR "system dynamics" OR 115 CGE OR "computable general equilibrium" OR IAM OR "integrated assessment 116 model" OR "input-output model" OR "econometric model" OR "multi-agent 117 model"); SOURCE TYPE: (Journal). 118 As some comprehensive models that analyze the SDGs contain poverty modules 119 but were not found by the keywords and search fields above, we used the following 120 search string which returned an additional 54 papers: 121 • TITLE-ABS-KEY ("sustainable development goals" OR SDGs) AND TITLE-122 ABS-KEY ("scenario modeling" OR "scenario model" OR "scenario analysis"); 123 SOURCE TYPE: (Journal). 124

Literature screening 125
Literature screening was then undertaken to process the collected 536 papers 126 based on their relevance and accessibility. From the 482 papers obtained from the first 127 search string, we selected 152 papers by browsing the title, abstract, and keywords and 128 excluding irrelevant papers. Excluded papers include art-, psychology-, or medicine-129 related papers that were incorrectly captured; papers that only focused on energy 130 poverty, fuel poverty, or food poverty and had no connections with SDG 1; papers that 131 had little connection with poverty (e.g., "poverty" only appear in abstracts as future 132 research). Moreover, we further excluded 18 papers because their full texts could not 133 be accessed online, or the scenario analysis method or model presented was not used or 134 could not be used for poverty analysis. From the 54 papers obtained from the second 135 search string, 10 papers were selected by excluding duplicate and inaccessible papers 136 and papers that did not mention poverty or SDG 1 in the full text. As a result, a total of 137 144 papers were retained for detailed review. 138

Key information extraction 139
By carefully reading each paper, the key information of each paper is recorded 140 from bibliometric and model information and as shown in Table 1. From the perspective 141 of bibliometric information, the research object in a paper represents the population or 142 community studied in each paper. Research scales involve global, regional, national, 143 and local, which cover almost all countries, multiple countries or economies, one 144 country, and a part (e.g., one or more states, cities) of a country, respectively. 145 Geographic locations of research areas are differentiated by country. 146 (2) econometric 150 models (Intriligator, 1983); (3) system dynamics (SD) models (Sterman, 2000); (4) 151 microsimulation models; (5) input-output models (Ten Raa, 2009); (6) Bayesian belief 152 network (BBN) models (Darwiche, 2009); and (7) hybrid models. 153 Each model targets one of the following three model purposes: ex-ante scenario 154 analysis (i.e., estimation of future trends under different scenarios), ex-post scenario 155 analysis (i.e., ex-post assessment of an event, policy, or behavior to analyze its 156 influence), and relationships exploration (i.e., exploration of quantitative relationships 157 between poverty and other factors under different scenarios). A model is considered to 158 be static if it doesn't consider temporal factors and the process experienced, and 159 dynamic if it can be used to examine the dynamic interactions in the system modeled 160 and analyze the evolutionary process of these relationships over a time period. 161 According to the maximum year (t) simulated by dynamic models, temporal scales of 162 models can be classified as short-term, medium-term, and long-term. 163   showed that eco-environmental deterioration increased poverty via increased food 212 prices, decreased agricultural production and farmers' incomes. Moreover, some 213 measures that could improve the environmental sustainability and enhance farmers' 214 adaptability to climate change greatly reduced poverty, such as rational distribution of 215 land, soil erosion management, and sewage treatment (X. . 216 The relationship between poverty and agriculture was also explored since the 217 poorest households were thought to be more concentrated in agriculture (FAO, 2017). 218 More than 16% of existing studies investigated the relationship and impacts of 219 agriculture-related factors on poverty, which involve agricultural productivity 220 variations ( In the selected studies, 138 papers presented models for poverty scenario 238 analysis, while the remaining 6 papers were literature reviews or only introduced a 239 conceptual model or framework. For these 138 papers, more than half of them (54.35%) 240 used national-scale models, while 23.19%, 12.32%, 10.14% applied local, regional, and 241 global scale models, respectively ( Figure 4c). 242 The most widely used model type is hybrid (55 in total) which integrate at least 243 two model types, followed by CGE models (Figure 4a). The majority (46) of the hybrid 244 models are the combination of CGE and microsimulation models. Both hybrid and CGE 245 models were used mainly for ex-ante scenario analysis. There were 24 econometric 246 models, most of which were developed for relationship analysis. The remaining models 247 were all used for ex-ante scenario analysis, including 16 system dynamics models, 10 248 microsimulation models, 2 input-output models and 2 BBN models. In terms of model 249 states, dynamic models were slightly more widely used than static models ( Figure 4b). 250 All SD and BBN models were dynamic. 251 For studies presenting dynamic models that explicitly defined a simulation 252 period, 58% were used for short-term

Poverty and other SDG indicators 263
A total of 11 indicators were defined to measure poverty in model-based 264 scenario analysis. More than two-thirds of studies only adopted one indicator, and the 265 remaining used multiple indicators. These indicators are classified into direct and 266 indirect indicators, and their usage counts are shown in Table 2. 267 The most commonly used indicator is the poverty rate, which is defined as the 268 ratio of the number of people living below a given poverty line to the overall population. 269 The ratio of people living below the poverty line has been calculated by income    For econometric models, model verification is relatively easy, because it is usually 490 carried out together with the parameter estimation to maximize the goodness of fit of 491 the model. However, they are only suitable for short-term poverty projections and the 492 situation of which the future socioeconomic trends are in line with past experience. In 493 the case of rapid socioeconomic, the model effectiveness in the projection of poverty 494 indicators will be seriously affected (Rey, 2000). SD models can track cause and effect, 495 allowing the exploration of complex systems with poverty feedback loops and 496 promoting the understanding of the causes and influences of poverty. SD models can 497 be used for poverty scenario analysis outside of the experience of historical data, but 498 they have some parameters and functional forms that are difficult to estimate. Their 499 verification is also complicated, and not only involves assessing the quality of 500 parameter estimations using a variety of data, but also evaluates the effectiveness of 501 model structure (Jin et al., 2017). Microsimulation models can effectively simulate the 502 impact of different poverty alleviation policies on different groups or individuals, but 503 they require more behavioral assumptions and more accurate microeconomic data 504 compared with traditional macroeconomic models (Ballas et al., 2013). 505 Input-output models can reflect the structural relationships of industries via 506 detailed industry information, and data are required to show the income and expenditure 507 of each economic sector to support poverty analysis. However, they are difficult to split 508 and integrate relevant data reflecting the industrial linkages among regions and 509 countries under some circumstances. Similar to other model types that rely heavily on 510 historical data, they cannot effectively respond to future uncertainties (Rey, 2000). BBN 511 models use conditional probability to express the causal and conditional relationships 512 between poverty and various elements, which can learn and deduce the probability of 513 occurrence of some outcomes under conditions of limited, incomplete, and uncertain 514 information. However, they are constructed based on the assumption of sample attribute 515 independence, and the model effectiveness gets worse if the sample data violate this 516 assumption (Oladokun, 2014). Hybrid models encompass combinations of a variety of 517 models and thus can conduct both macro and micro poverty scenario analysis, cover 518 wider sectors and have higher applicability for poverty in more complicated systems. 519 However, using hybrid models have to face the difficulties of complicated model 520 development and evaluation as well as the higher unavailability of historical data. 521 Table 5. Advantages and disadvantages of various models commonly used for poverty 522 scenario analysis. 523

CGE models Link various economic sectors and industries.
Relying on the assumption of equilibrium; unable to respond effectively to future uncertainties; difficult to verify the global model and organize the data; Econometric models Easy to verify the model by fitting historical data.
Suitable for short-term development research instead of long-term research; unable to respond effectively to future uncertainties.

SD models
Exploration of causal mechanism and dynamic complex relationships; can be used for scenario analysis beyond the trend of historical data.
Difficult to obtain values of some parameters; difficult to evaluate models' effectiveness.

Microsimulation models
Analyze the impacts on different populations and even individuals.
Need more behavioral assumptions and more accurate and true microeconomic data; difficult to evaluate models' effectiveness.

Input-output models
Reflect the structural relationships of industries by detailed industry information.
Difficult to split and integrate relevant data reflecting the industrial linkages among regions and countries; unable to respond effectively to future uncertainties.

BBN models
Causal and conditional relationships exploration.
Use the hypothesis of sample attribute independence.

Hybrid models
Macro and micro combination; wider sectoral coverage; suitable for studying complex issues.
More complicated model development; more data demand; difficult to evaluate models' effectiveness.
In summary, it is recommended to use CGE or econometric models if a study 524 focuses more on economic activities and poverty. Input-output models are more suitable 525 to explore the relationship between poverty and each single industry (e.g., agriculture, 526 forestry, fishery, manufacturing, transportation). Microsimulation models are 527 appropriate to conduct the poverty analysis at the micro level (e.g., individuals, 528 communities). SD and BBN models are the better choice if the dynamic causal 529 mechanisms covering poverty and multiple other sectors need to be explored. Hybrid 530 models can be utilized to research poverty in complex systems with dynamic causal 531 mechanisms, relationships of various sectors and industries by combining multiple 532 types of models at macro and micro levels. 533

Representative models 534
We derived seven representative models ( in the economy, society and the environment, and identifying the interrelated factors 620 and behaviors in systems, and then establishing their dynamic relationships. These 621 efforts will promote a comprehensive understanding of the evolution mechanism of 622 poverty in a complex system instead of the simple behavioral association between 623 poverty and certain factors, which ultimately help uncover better poverty reduction 624 strategies with consideration of synergies and trade-offs for other SDGs. 625

Conclusions and suggestions for future poverty scenario analysis 626
This paper reviewed 144 papers on model-based poverty scenario analysis. We 627 classified these models into seven types, including computable general equilibrium, 628 econometric models, system dynamics models, microsimulation models, input-output 629 models, Bayesian belief network models, and hybrid models. These models were used 630 for ex-ante scenario analysis, ex-post scenario analysis, and relationships exploration. 631 We also identified seven representative poverty scenario analysis models. We found the 632 following research gaps based on the review of bibliometric and model information, 633 and the discussions on different model types and interactions between poverty and other 634 SDGs. 635 (1) Around 80% of previous studies were carried out at national and local levels 636 and models that could be used for medium-and long-term poverty simulations were 637 very limited. However, in the context of increasing international cooperation and 638 integration, poverty research from global to local scales is indispensable. It is conducive 639 to understanding the evolution mechanism of poverty and their interactions with other 640 SDGs and other related international agendas (e.g., the Paris Agreement), guiding 641 global to local poverty strategies in a long-term perspective (Hughes et al., 2015). 642 (2) Poverty scenario analysis was mainly carried out from the single perspective 643 of the economy, eco-environment, and agriculture, while comprehensive analyses that 644 integrate multiple sectors (e.g., economic, social, and environmental) was seldom 645 reported. Few models can address synergies and trade-offs between SDG 1 and other 646 SDGs, but the interactions between poverty and other SDGs and their potential impacts 647 are essential for reducing poverty and the resulting negative impacts (De Neve and  648 Sachs, 2020), and poverty alleviation needs to be dealt with scientifically in a more 649 comprehensive and integrated way (Adger and Winkels, 2014). 650 (3) The hybrid models used in poverty scenario analysis were mainly the 651 integration of CGE and microsimulation models. The advantages of these models were 652 not fully reflected for modelling dynamic causal mechanisms and multiple sectors 653 relationships in complex systems. 654 (4) The poverty rate was the most widely used indicator to measure poverty in 655 previous studies. However, due to the complexity of poverty and its diverse driving 656 factors, this indicator cannot represent the diverse information of poverty, such as the 657 depth and inequality of poverty. 658 As a result of the literature review about model-based poverty scenario analysis, 659 some suggestions for future research are provided below to fill up the research gaps in 660 existing studies. 661 (1) It is desirable to develop effective scenario analysis models for more 662 medium-and long-term simulations of poverty changes under different future scenarios, 663 especially global and regional models for understanding the evolution of global or 664 regional poverty. 665 (2) The second promising direction is to develop scenario analysis models 666 covering multiple sectors and a broad range of variables for these sectors so that the 667 combined effects of multiple poverty alleviation policies can be evaluated. These 668 variables include economic growth, population, education, health, agriculture, climate 669 change, land use, water use, and energy use. 670 (3) It will be helpful to enhance the modeling of synergies and trade-offs 671 between poverty and other SDGs, particularly with the relevant SDGs that are 672 considered to have significant synergies or trade-offs (e.g., SDGs 2-3, SDGs 7-9, SDG 673 13) (