Modelling Social Evolutionary Processes and Peer Effects in Agricultural Trade Networks: the Rubber Value Chain in Indonesia

Modelling Social Evolutionary Processes and Peer Effects in Agricultural Trade Networks: the Rubber Value Chain in Indonesia

Thomas Kopp, +49 551 394821 (Corresponding author). Georg August University, Platz der Göttinger 7, 37073 Göttingen, Germany Jan Salecker Georg August University, Platz der Göttinger 7, 37073 Göttingen, Germany

Understanding market participants’ channel choices is important to policy makers because it yields information on which channels are effective in transmitting information. These channel choices are the result of a recursive process of social interactions and determine the observable trading networks. They are characterized by feedback mechanisms due to peer interaction and therefore need to be understood as complex adaptive systems (CAS).

When modeling CAS, conventional approaches like regression analyses face severe drawbacks since endogeneity is omnipresent. As an alternative, process-based analyses allow researchers to capture these endogenous processes and multiple feedback loops. This paper applies an agent-based modeling approach (ABM) to the empirical example of the Indonesian rubber trade. The feedback mechanisms are modeled via an innovative approach of a social matrix, which allows decisions made in a specific period to feed back into the decision processes in subsequent periods, and allows agents to systematically assign different weights to the decision parameters based on their individual characteristics. In the validation against the observed network, uncertainty in the found estimates, as well as under determination of the model, are dealt with via an approach of evolutionary calibration: a genetic algorithm finds the combination of parameters that maximizes the similarity between the simulated and the observed network.

Results indicate that the sellers’ channel choice decisions are mostly driven by physical distance and debt obligations, as well as peer-interaction. Within the social matrix, the most influential individuals are sellers that live close by to other traders, are active in social groups and belong to the ethnic majority in their village.


references/ReferencesNetworks.bib \addbibresourcereferences/RUBNET.bib \pdfstringdefDisableCommands

Keywords: agent-based modelling; inverse modelling; complex adaptive systems; networks; rubber; Indonesia; agricultural trade.

1 Trade networks are complex, adaptive systems

Understanding market participants’ channel choices can provide critical insights into which networks are effective in transmitting information. This is important for policy makers who prefer making use of existing structures over introducing a new network of extension workers. Marketing decisions which materialize in the observed trading network are often characterized by feedback loops, especially in rural environments and small communities. Previous research finds evidence for “higher response to […] conservation covenanting programs when agents are part of a local uniform matching network” \citep[p. 1]Iftekhar2016. This paper suggests an agent-based modelling (ABM) approach to test for the importance of social closeness and peer interaction in market participants’ trading decisions.

These decision-making processes can be referred to as complex adaptive systems (CAS) as defined by \citetHolland1996, as they are influenced by multiple and complex interactions between stakeholders and the recursive nature of the decisions. \citet[p. 10]Rammel2007 describe CAS as leading to “large macroscopic patterns which emerge out of local, small-scale interactions”. These evolutionary dynamics arise because CAS result from a) interactions amongst agents, b) interactions between agents and the environment, and c) a learning process through repetitions of these interactions. The whole system of agents therefore adapts to the environment \citepPotgieter2005. So every action and its respective result is fed back into the decision-making processes in future periods which is referred to by \citet[p. 10]Rammel2007 as “co-evolutionary processes and dynamic patterns”. While the CAS approach is most often applied to biological processes, \citetMarkose2005 argues that socio-economic systems like markets may well be understood as CAS, too.

Feedback loops are a source of high-degree complexity in CAS \citepVandenBergh2003 and do not settle to static equilibria \citepRammel2007. Conventional approaches such as regression analysis face severe drawbacks in modelling these processes since endogeneity is omnipresent \citepHolland2006.

Process-based approaches are better suited to study CAS since they can capture endogenous processes and multiple feedback loops. One promising alternative following this logic is analysis via agent-based modelling (ABM).

A prime example of CAS at work in the field of economics is in the marketing network for natural rubber in the Jambi Province in Indonesia. In Jambi Province, rubber is produced predominately by smallholders whose output is distributed via a dense network of small agricultural traders to domestic processors. These are crumb rubber factories. Since traders vary in size and capacity, smaller village traders sell the rubber to larger district traders, who then sell either to the processor or to still another trader \citepKopp2017a. The questions that this paper addresses are: why do agricultural traders sell to a specific buyer? And, on an aggregated level, what are determinants of a trading network’s structure? To avoid confusion, when differentiating between selling traders and their buyers, including other traders, this paper refers simply to “buyers” and “sellers” throughout.

To answer these questions, we model each seller’s decision regarding whom to sell to as a recursive process. An individual seller’s initial selection of a buyer is decided through ranking all potential buyers based on various characteristics, which are weighted by global parameters. The decisions made by an individual seller’s peers in the previous period affect his/her decision-making process in the current period. But just how strongly do the past decisions made by other sellers affect the channel choices of the individual trader under consideration? This decision appears to be in part determined by the social closeness between the sellers. These effects are operationalized in our model via a so-called social matrix, which quantifies the social closeness between each seller and his or her peers. The resulting matrix weights the impact of all other sellers’ decisions on each individual seller. After each iteration descriptive metrics of the predicted network are saved. This process is then repeated until the metrics converge. The resulting, seller-specific lists order all potential buyers according to each individual seller’s propensity to engage in trade with them. An optimization algorithm is used to determine the values of the global parameters which maximize the number of correctly predicted trading links.

To summarize, this paper models the channel-choices of agricultural traders with an ABM approach. This is superior to approaches of regression analysis because of dynamic network effects introduced by feedback-loops in the agents’ decisions. The model includes an innovative approach of multiplying a social matrix with a weighting vector, allowing for heterogeneous effects based on individual characteristics. This enables the researcher to identify influential individuals whose decisions are disproportionally influential. The analysis builds upon a unique dataset generated during a 2012 micro-level survey of agricultural traders in Jambi province, Indonesia.

To the best of the authors’ knowledge, this is the first paper employing an ABM approach based on the theory of CAS to predict agricultural traders’ channel choices. The inclusion of a social matrix in traders’ network analysis is an innovative approach as well. The methodology developed in this paper can be employed in other areas where the identification of efficient channels for transmitting information is desired.

This paper is structured as follows: the literature review section gives an overview of ABM based approaches used in the agricultural economics literature so far. Section three presents our ABM RUBNET. The complete model description (ODD protocol: overview, design concepts, and details) is included in the appendix. Section four presents and discusses the results, and section five provides a conclusion.

2 ABM and CAS in the (agricultural) economics literature

While agent-based modelling is increasingly being applied in agricultural economics literature, virtually no empirical work has been undertaken to model marketing decisions at the micro level. The majority of studies model production decisions at the farm level, such as investment decisions \citepFeil2013,Resende-Filho2008, adaption of new technologies \citepSchreinemachers2009, participation in certification schemes \citepLatynskiy2017, farm-level climate change adaptation \citepTroost2014, or breeders’ responses to price shocks \citepZhang2010. \citetLatynskiy2017 model decision making within coffee farmers cooperatives to understand processes of collective action in voluntary sustainability certifications. On a similar track, \citetIftekhar2016 model the participation of farmers in conservation programs subject to constraints on land-use and existing social networks. \citetBoyer2013 analyse processors’ market power by including the US Livestock Mandatory Price Reporting Act into an auction-based ABM.

In a broader economic application, \citetKlos2001 analyse transaction costs with an ABM to explicitly account for often-ignored characteristics of transactions like mutual trust and heterogeneity of agents, and challenge the assumption of efficient outcomes. \citetAlfarano2009 use an ABM approach to generate evidence on macro level outcomes from the behavior of a multitude of diverse agents on a micro level on financial markets. \citet[p. 1182]Zhang2010 argue that agent-based computational economics (ACE) “can be used to study problems with behavioral assumptions that are too difficult to analyze with mathematical methods. ACE is more economical and time efficient compared with experiments with human subjects (e.g. \citealpWard1999) and is more controllable.”.

In the CAS literature, \citetMarkose2005 provides an extensive overview of ABM approaches in analysing CAS in economics. These are required in situations that deviate from the basic assumptions typically made in economics, for example when understanding processes such as “innovation, competitive co-evolution, persistent heterogeneity, increasing returns, the error-driven processes behind market equilibria, herding, crashes and extreme events such as in the business cycle or in stock markets.” \citep[p. 159]Markose2005. \citetButler2016 reviews the literature on applications of complex systems to agricultural economics. He generally concludes that nowadays (agricultural) markets are too complicated for regression analysis, which assumes the emergence of equilibria, because markets need to be understood as complex systems “in which economic agents […] continually adjust and react to market behavior of others” \citep[p. 2]Butler2016. Reasons for this perceived complexity include feedback loops and neighbor effects, which are likely to occur in the seller’s decision-making process. \citetChen2001 analyse the behavior of traders on an artificial stock-market with an ABM. They find that while traders may behave as if they do not believe in the efficient market hypothesis, their aggregate behavior results in an efficient capital market.

When studying trading networks, we are observing a highly dynamic system in which each seller’s decisions have an influence on each of his or her neighbors’ decisions, which might then in turn influence other neighboring traders. Since these effects go back and forth, endogeneity is omnipresent, which prevents the use of conventional regression analysis. And whenever observations are only available for one point in time it is difficult to assess how knowledge spreads between stakeholders. The ABM approach is a way of circumventing the endogeneity problem \citepZhang2010. We therefore employ an agent-based, pattern-oriented modeling approach to generate a hypothetical outcome under certain assumptions (represented as model parameters, [Grimm2005a]). We predict trading connections of model agents based on sets of parameter values and compare the emerging network to the observed network. Global parameters are then changed systematically in order to maximize the level of resemblance between the simulated and empirical trading network.

3 Application to the rubber value chain in Indonesia

Value chains in rural areas of less developed countries tend to rely on agricultural traders and middlemen. These key agents offer crucial services, such as the transportation of farm output, the reduction of information asymmetries (for example on prices), and the lending of credit to farmers. The traders we observed in the rubber market value chain in Jambi Province, Indonesia, provide all of these services.

Unlike many other agricultural products, Rubber is not perishable. This enables the formation of long value chains consisting of many stakeholders, resulting in extended networks between traders, as can be seen in Figure 1.

Source: Own production, based on original survey data, collected in a representative survey with rubber traders in 40 villages in the Jambi Province (Sumatra, Indonesia) as well as with the downstream traders that the initial respondents named as their buyers. Further information on the sample can be found in \citepKopp2017a. Borders of Jambi and Sumatra from Center for International Forestry Research, surface of Jambi from NASA/EOSDIS.
Figure 1: Trader Network in Jambi.

3.1 Model description

The overall aim of our study is to reconstruct the channel choice behavior of selling agents within an empirical context. So the specific purpose of RUBNET is to understand channel choice behavior of rubber traders in Jambi, Sumatra, Indonesia. Testable hypotheses are:

  • : Only distance matters in the marketing decision.

  • : Sellers sell to the buyer that offers the highest price.

  • : Neighboring sellers influence each other.

  • : If a seller has a credit with a specific buyer, he/she cannot sell to another one.

We run an optimization scenario by using genetic algorithms that vary weight parameters in order to maximize the proportion of correctly predicted trading links \citepKumar2010.

RUBNET is implemented in NetLogo version 6.0.3 \citepWilensky1999. The detailed ODD (Overview, Design concepts, Details) protocol for describing individual-based models \citepGrimm2006a,Grimm2010 is provided in the appendix.

3.1.1 Behavioral assumptions

Traders are assumed to make decisions, which are mainly influenced by revenue and transaction costs of selling, in order to maximize their profits. Revenue depends on the price they receive from their buyer. Transaction costs depend on the location relative to the seller and other unobserved characteristics of the buyer. The selection of a buyer therefore crucially affects the seller’s profits. Personal relationships between stakeholders are assumed to play a central role in the decision-making process as well. They include interactions between the trader under consideration with a) potential buyers and b) with his or her peers, i.e. neighboring sellers. The relationship between the seller and potential buyers is characterized by a) stable characteristics such as the physical distance and b) time-variant characteristics such as the amount of credit the seller has taken from the buyer in previous periods. The relations between neighboring sellers include spill over and learning effects and therefore generate feedback mechanisms. “Neighborhood” is defined widely, i.e. along a number of dimensions such as physical proximity, ethnicity, and similarity in level of education. The causal chain of an agent’s decision making is displayed in Figure 3.1.1.

Figure 2: Causal chain of decision making process
Comments 0
Request Comment
You are adding the first comment!
How to quickly get a good reply:
  • Give credit where it’s due by listing out the positive aspects of a paper before getting into which changes should be made.
  • Be specific in your critique, and provide supporting evidence with appropriate references to substantiate general statements.
  • Your comment should inspire ideas to flow and help the author improves the paper.

The better we are at sharing our knowledge with each other, the faster we move forward.
The feedback must be of minimum 40 characters and the title a minimum of 5 characters
Add comment
Loading ...
This is a comment super asjknd jkasnjk adsnkj
The feedback must be of minumum 40 characters
The feedback must be of minumum 40 characters

You are asking your first question!
How to quickly get a good answer:
  • Keep your question short and to the point
  • Check for grammar or spelling errors.
  • Phrase it like a question
Test description