Modelling collective decision-making in policy networks

[Accepted to the 5th International Conference on Computational Social Science IC2S2, July 17-20, 2019, University of Amsterdam, The Netherlands]

Keywords: policy-making; group decision-making; bounded-rationality; agent-based model; empirical data.

Context & research questions

Public policy-making - the set of intentions, decisions, and processes that lead to the creation, implementation and recycling of policies within governments and international organizations - takes place in a complex system (Geyer and Cairney 2015). This system is characterised by heterogeneous actors pursuing their idiosyncratic goals in an institutional context and a changing environment that rhythms the course of and affects the content of policy decisions (Geyer and Rihani 2010, Morçöl 2013, Koliba et al. 2018). Zooming in on the micro-level, one can observe policy actors’ cognitive complexity, usually conceptualised as ‘bounded- rationality’ (Simon and March 1976; Dente 2014; Hertwig and Pedersen 2015; Zahariadis and Herweg 2017). In the aggregate, policy dynamics satisfy the hallmarks of complex systems: non-linear interactions between policy stages (Cairney 2016); disproportional re-allocation of attention in response to media coverage and external perturbations (Jones and Wolfe 2010); and power-law relationships between the frequency and the importance of budget changes (Jones et al. 2009, Sebok and Berki 2017). There is a clear conceptual link between the micro-level decisions and the macro-level policy dynamics. So far, both levels have only been studied through separate analyses. We propose to reconcile micro and macro dynamics with an empirically-grounded agent-based model that focuses on collective decision-making processes in agenda-setting, policy design and policy adoption. We explore the following research questions:

How does decision-making scale? Public policy and political science studies shifted from an institutional to an individual focus and progressed in understanding cognition and behaviour. By doing so, the aggregate aspect got lost. The way that individual behaviours - or types of decision-making - scale to group or institution-level is unclear and not addressed by the literature. The proposed agent-based model would serve this analysis by iterating group behaviour and allowing to observe how collective decisions emerge from the self-organisation of individual agents.

How does the policy-making process change in reaction to so-called “focusing events”? While most of the public policy literature mentions the importance of external perturbations to create or disrupt consensus, the nonlinear dynamics underlying these changes are either theoretically described or empirically observed, but not formalised. Our model explores the underpinnings of such formalization.

What are the benefits and drawbacks of bounded-rationality? The literature on bounded-rationality suggests that agents rely on heuristics which might lead them to make errors, but also allow them to make quick decisions in the face of uncertainty.  Our model counterfactually tests various levels of bounded-rationality to see their effect on decision-making.

What initiates a window of opportunity? Most of the public policy literature mentions the importance of windows of opportunity; the time during which changes can happen or when progress can be made. Our agent-based model generates dynamics and windows of opportunity one could reverse-engineer to analyse what happened before and, potentially, what could have led to their occurrence.


The agent-based model is implemented in the Julia language and uses bit-strings to encode information which allows us to rely on the fundamentals of evolutionary biology to model agents’ learning abilities (Mitchell 1998). We use Dente’s policy actor framework (Dente 2014) to conceptualise our agents, j. An agent is characterised by:

  • Attributes (A) including their type (t), interest (i), role (r)  and resources (rs).

  • Cognitive rules (CR) including current memory (cm) and attention (ca); the components of bounded rationality modelled as discount factors: limited knowledge (k), intellectual capacity (ic), memory (m), and attention (at); and learning strategies (l) such as imitation, innovation and trial and error.

  • Strategies (S) including a set of framing strategies to alter other agents’ interests, roles and attention.

  • Outcomes of interest (O) that agents pursue through the implementation of their strategies.

  • A network structure that evolves over time as a function of agents pursuing central positions or forming coalitions.

The second part of the model is characterised by a changing environment with focusing events (e) (i.e. a catastrophe or media coverage) that alter agents’ attention. Problems and solutions are context-dependent and characterise the policy situation agents work on. We include three policy stages - agenda-setting, policy design, adoption - to rhythm the process. Over iterations, we keep track of strategy fitness, changes in agendas and policy designs, and current foci of attention.

Empirical settings

Beyond hypothetical scenario exploration and theory-building experiments, we seed the model with empirical data from two policy networks (j, k, ... 20) in a government and an international context. Data include size of network; the agent characteristics; the hierarchy structure; the decision-making process; the set of problems and solutions; and possible external events. Figure 1 depicts an initial network structure with j1,  j2 = 12.

Figure 1: network structure

Figure 1: network structure


Figure 2 shows the hypothetical results of strategy fitness, comparing framing strategies (in green) and network progression (in black) in response to two external events (red).


The impact of our research is fivefold: (1) quantitative, inductive research in data-scarce micro-level policy contexts; (2) formalisation of previously informal theory; (3) linking micro-level motives and macrostructures through computational modelling using Julia; (4) anchoring scenarios in empirical data; (5) providing an in-silico laboratory to run counterfactual experiments for further analyses of collective decision-making.

Figure 2: fitness strategy

Figure 2: fitness strategy


Cairney, Paul. 2016. The Politics of Evidence-Based Policy Making. Springer.

Dente, Bruno. 2014. ‘Understanding Policy Decisions’. In Understanding Policy Decisions, 1–27. Springer.

Geyer, Robert, and Paul Cairney. 2015. Handbook on Complexity and Public Policy. Edward Elgar Publishing.

Geyer, Robert, and Samir Rihani. 2010. ‘Complexity and Public Policy: A New Approach to Twenty-First Century Politics’. Policy and Society, London, New York: Routledge.

Hertwig, Ralph, and Arthur Paul Pedersen. 2016. ‘Finding Foundations for Bounded and Adaptive Rationality’.

Jones, Bryan D, Frank R Baumgartner, Christian Breunig, Christopher Wlezien, Stuart Soroka, Martial Foucault, Abel François, et al. 2009. ‘A General Empirical Law of Public Budgets: A Comparative Analysis’. American Journal of Political Science 53 (4): 855–73.

Jones, Bryan D, and Michelle Wolfe. 2010. ‘Public Policy and the Mass Media: An Information Processing Approach’. In Public Policy and the Mass Media, 35–61. Routledge.

Koliba, Christopher J, Jack W Meek, Asim Zia, and Russell W Mills. 2018. Governance Networks in Public Administration and Public Policy. Routledge.

Mitchell, Melanie. 1998. An Introduction to Genetic Algorithms. MIT press.

Morçöl, Göktŭg. 2013. A Complexity Theory for Public Policy. Routledge.

Sebók, Miklós, and Tamás Berki. 2017. ‘Incrementalism and Punctuated Equilibrium in Hungarian Budgeting (1991-2013)’. Journal of Public Budgeting, Accounting & Financial Management 29 (2): 151–80.

Simon, Herbert, and James March. 1976. Administrative Behaviour and Organisations. New York: Free Press.

Zahariadis, Nikolaos, and Nicole Herweg. 2017. ‘The Multiple Streams Approach’. In The Routledge Handbook of European Public Policy, 54–63. Routledge.