ORA (Round 5) Towards realistic computational models of social influence dynamics

Lead Research Organisation: Manchester Metropolitan University
Department Name: Centre for Policy Modelling


Recently, many societies shifted towards more polarization and volatility in opinions, for example in attitudes about immigration or climate policy, but underlying reasons are poorly understood. A key obstacle is that opinion dynamics in society involve a complex micro-macro interaction between fundamental interpersonal processes of social influence, meso-level conditions like network structures, and macro-level outcomes, like polarization in opinion distributions. Agent-based simulation models (ABM) offer powerful tools to bridge theoretically micro-level processes and macro-level dynamics. However, ABM of social influence hitherto fails to inform empirical research and policy. Two major roadblocks are 1) a lack of theoretical integration of the large variety of models and 2) a lack of empirical tests of model assumptions and predictions.

ToRealSim is a unique effort by a consortium of leading experts in ABM and empirical research to overcome both roadblocks. Combining expertise on prominent modelling approaches, we develop a framework to systematically compare models and their predictions. Drawing on this, a mixed-method approach will link models to empirical evidence, utilizing our expertise on qualitative research, experiments, social network analysis and survey research. We test models empirically, both in terms of micro-level assumptions and macro-level predictions, proposing a new generation of social influence models that synthesizes insights from ToRealSim.

Planned Impact

Models developed in ToRealSim will be applied to address the highly debated question whether increasing opinion diversity is related to a decline in social cohesion. Data obtained with a recent "cohesion radar" developed by members of our consortium can be used to relate observed changes in social cohesion to, first, measured changes in diversity, polarization, consensuality, and extremity in key issues identified in WP4. Second, the agent-based models that best match those data can be used to assess whether there is a potential for structural changes of the opinion landscape towards polarization, and how this could be expected to affect societal cohesion.

A central problem of the current state-of-the-art is that practitioners are confronted with a huge number of different models but no guidance in the selection of the best model for their specific purpose. Based on the theoretical insights from WP1 and WP2 and their respective empirical tests in WP3 and WP4, we will provide a list of (i) critical decisions that modelers of processes in a given setting need to make, (ii) an overview over important modeling options, and (iii) empirically informed guidance with regard to the characteristics of the setting that should inform the selection of modeling options. Finally, we will point practitioners to context-specific intervention methods that can help avoiding undesired outcomes. Examples can be found in our own work about effects of the so-called "timing of contacts", effects of network structure, and the demographic composition of groups on polarization dynamics

A further focus is dissemination to non-academic audiences. For instance, the French team tests opinion-dynamics models about conversion of farmers to organic farming and will disseminate insights to staff of "chambres d'agriculture", involved in an already funded project. The UK team will work with citizens.is, an online platform for public debate of policy issues that is now used in many participatory and consultative exercises around Europe. In particular it will discuss project results for features of platform design that help prevent polarization or opinion clustering in settings where we can know that this outcome would be clearly unintended and undesired by the participants. The Dutch team is involved in SCOOP (NWO gravity), a recently financed major effort to study sustainable cooperation in families, organizations and communities (https://www.scoop-program.org/). We discuss insights from ToRealSim about dynamics of attitudes about immigration with SCOOP-stakeholders (e.g. Scientific Council for Government Policy, WRR, or Netherlands Institute for Social Research, SCP). The German team is involved in the "Social Cohesion Radar" of the Bertelsmann Foundation, in which insights from ToRealSim about degree and trend of polarization will be discussed with policy makers.
Description Experiments and observations of groups seeking agreement (including pairs) use a variety of strategies. Groups may adopt an argumentation approach where members add their knowledge additively in the form of causal observations and arguments, they may polarise into sub-groups (or individuals) that so that agreement is very hard, or they may enter a phase of "knowledge haggling" where contributions are traded according to principles of (a) judging the expertise of the contributor or (b) quid pro quo (you accept this if I accept that). Sometimes groups get stuck in one mode or another which can impact upon the quality of the collective decision making.

Some of the knowledge gained about modelling such processes has been applied in another project (an EU project: "Populism and Civic Engagement") to model social influence in Austria during "immigration crisis" (2013-2017).
Exploitation Route The outcomes of the project might be useful for those involved in designing interactive forums where people routinely discuss issues or structuring face-face discussions in order to facilitate agreement.
Sectors Creative Economy,Government, Democracy and Justice,Culture, Heritage, Museums and Collections,Other

URL https://sites.google.com/view/social-influence-wiki/home
Description (PaCE) - Populism And Civic Engagement - a fine-grained, dynamic, context-sensitive and forward-looking response to negative populist tendencies
Amount € 3,003,309 (EUR)
Funding ID 822337 
Organisation European Commission 
Sector Public
Country European Union (EU)
Start 02/2019 
End 01/2022
Title Sense-making and Opinion dynamics: Conceptual paper 
Description This conceptual paper addresses the impact of emotions and power on sense making. We suggest these are units of analysis that are instrumental to understanding the cognitive and social dynamics of sense making and how this is transferred to collective Opinion Dynamics. The paper mingles inputs from different social sciences suggesting these disciplines harbour resources that can be utilized to expand OD as a field of study. The concepts noted in the paper could be used to better understand what and why cues register as we endeavor to make sense of our experiences and how and why we enact that sense the way we do. They can be used to explore intentionality based on negative preconceptions about phenomena which creates vulnerabilities that enables agents to interfere and amplify those preconceptions. The concepts can also be linked to the construct of noise in Opinion Dynamics models, noise being an abstract unknown or an uncertainty that affects Opinion Dynamics, but which tends in many models to have little empirical support or grounding. The article proposes empirical indicators to studying and validating the micro-social assumptions of Opinion Dynamics models 
Type Of Material Improvements to research infrastructure 
Year Produced 2021 
Provided To Others? Yes  
Impact The article further expounds on the conceptual underpinnings of the ToRealSim qualitative research protocol (see research infrastructure) It introduced theoretical dimensions that have not been properly explored in the Opinion Dynamics literature 
Title ToRealSim: Opinion Dynamics: Towards a qualitative research protocal 
Description The method addresses a shortcoming in the study of opinion dynamics by proposing 5 empirical units of analysis. These are developing into a set of questions and coding instrument that are designed to identify and address different dimensions of the research participants world view. The approach addresses a shortcoming in current approaches that a) do not clearly demarcate belief from other constructs i.e. values and norms and b) tend to conflate opinions and attitude. The ToRealSim protocol adds a social cost element to the attitude dimension,in other words, attitudes are the manifestation of opinions. 1. Belief is grounded in the acceptance of and respect for moral authority. Beliefs are deeply held convictions we hold to be true without requiring proof or evidence. 2. Values represent a set of intrinsic standards by which we measure the appropriateness of options. Values underpin our understanding of right or wrong, sense of fairness and injustice and as such they are salient factors that stem from beliefs and that shape opinions, attitudes, and selection but which do not always predict or determine actual behaviour. 3. Norms are socially acceptable behavioural standards embedded within a social structure of culture, class, education, gender, and other aspects of a person's social identity. 4. Opinions are internal judgements concerning a particular object or process which we propose may or may not be expressed in an attitude. There may be risks involved in expressing or enacting an opinion so we may choose to withhold the opinion because of the perceived social cost of revealing it. 5. Attitudes are declared dispositions to react in a favourable or unfavourable way towards an object, process, people or events. Because an attitude is declared either verbally or in some way signified e.g. body language, facial expression or physical action, it carries a social cost and can be empirically assessed and in many cases measured. 
Type Of Material Improvements to research infrastructure 
Year Produced 2020 
Provided To Others? Yes  
Impact The method is currently being trialed in interview situations and experimental settings: See ToRealSim datasets, databases and models 
Title ToRealSim - Opinion dynamics: Experimental workshops - Qualitative interview data 
Description Data produced in experimental workshops. Workshops designed in 3 tiers: before experiment interviews, experimental setting exploring the impact of a reasoned argumentation on opinions, post interviews to explore the impact of the debate on the interviewees. Datasets are coded InVivo and subsequently coded to explore causal inferences. Data set is anonymized with respondents providing informed consent for the processing of the data. Respondents must be informed of intentions to use the data sets for purposes other than originally intended. 
Type Of Material Database/Collection of data 
Year Produced 2021 
Provided To Others? Yes  
Impact The data set is currently being analyzed using another ToRealSim deliverable - Improvements to research infrastructure: Towards a qualitative research protocol. This analysis involves the Beta testing of the protocol that will be applied in planned experimental workshops following necessary adjustments. 
Title A set of agent-based simulation models 
Description An ongoing series of simulation models that re-implement and then extend the basic opinion dynamics model (Deffuant et al. 2002). These are all written in NetLogo (Wilensky 1999), and are to allow for the testing of this family of opinion dynamics models to additional assumptions. The original model was very abstract, and designed to be the minimum possible model to show how the polarisation of opinion could occur through interaction alone (rather than all opinions coming together). Each agent has two numbers, representing (a) their opinion on an issue and (b) how certain they are about that opinion. The second affects how far another agent's opinion can be when they meet, yet still influence them. Then agents randomly meet. Each time they do influence each other their opinions and uncertainties converge a little. 1. The basic re-implementation ("ToRealSim OD variants - base v2.7.nlogo"). This simplifies the original code somewhat and adds (a) some new statistics (b) visualisations (a profile of the agents' opinions and uncertainties, and one on the clustering levels) (c) options (e.g. original distribution of moderates opinions) and (d) sources of noise. The sources of noise include: (I) a probability that any agent is randomly reset to a new random moderate (equivalent to an existing person leaving and being replaced by a new person from outside the frame) (II) a probability that another agent that is distant from the original (outside the range given by its opinion and uncertainty) still influences an agent (III) the standard deviation of some random normal noise added to opinions of each agent each simulation tick (equivalent to opinion drift for reasons outside the model scope) (IV) the standard deviation of some random normal noise added to uncertainties of each agent each simulation tick (equivalent to opinion drift for reasons outside the model scope) The original did not include these possibilities, each of which can affect the outcomes and better reflect the reality of any observed situation. The other models build upon this basic core. 2. A variant (ToRealSim OD variants - base v2.6 - locales v1.2.nlogo) that adds a series of locations that structure interaction - so influence can only happen at a location and each agent only goes to a few of the possible locations. The original was a "well mixed" model, in that any agent randomly mixes with any other (but only influences if they are of sufficiently similar opinion). The amount of movement between locations, and trying new locations etc. can be varied to see if this makes a difference to outcomes. 3. A variant (ToRealSim OD variants - base v2.6 - homphily 0.2.nlogo) that adds observable markers to agents, where agents prefer to interact with those with similar markers. Markers can be of two types, changeable ones (that might correspond to clothes, hair style etc.) and ones hard to change (in reality this could be anything that is not easy to change: sex, accent, skin colour etc.). Deffuant, G., Amblard, F., Weisbuch, G., & Faure, T. (2002). How can extremism prevail? A study based on the relative agreement interaction model. Journal of artificial societies and social simulation, 5(4). Wilensky, U. 1999. NetLogo. http://ccl.northwestern.edu/netlogo/. Center for Connected Learning and Computer-Based Modeling, Northwestern University. Evanston, IL. 
Type Of Technology Software 
Year Produced 2019 
Open Source License? Yes  
Impact Work is still in development, and this family of models will be extended in collaboration with the rest of the project. However these already throw light upon the robustness of this family of models. Model 1 shows that even tiny probabilities of reset (I) or action at a distance (II) can change the nature of outcomes. Model 2 shows that even small amounts of change in locations is equivalent to the well-mixed model. Model 3 shows a wide variety of medium-term dynamics with different sub-groups weakly interacting that the original model did not show. 
URL http://cfpm.org/ToRealSim/models
Title Data-based simulation of opinion dynamics 
Description The simulation is a variant of the "ToRealSim OD variants - base v2.7" base model, which is based on the standard DW opinion dynamics model (but with the differences that rather than one agent per tick randomly influencing another, all agents randomly influence one other per tick - this seems to make no difference to the outcomes other than to scale simulation time). Influence can be made one-way by turning off the two-way? switch Various additional variations and sources of noise are possible to test robustness of outcomes to these (compared to DW model). In this version agent opinions change following the empirical data collected in some experiments (Takács et al 2016). Thus this allows us to compare the micro-level cognitive model with the possible macro-level results if this were how the actors being represented determined their response. The model, the documentation is in the "Info" tab and some slides explaining the simulation and its derivation are freely available in the following open access archive: Bruce Edmonds (2021). "A data-informed bounded-confidence opinion dynamics model" (Version 0.3.0). CoMSES Computational Model Library. Retrieved from: https://www.comses.net/codebases/537b40cf-e300-4245-bcc6-427b53515bf6/releases/0.3.0/ 
Type Of Technology Software 
Year Produced 2020 
Open Source License? Yes  
Impact This showed the incompatibility of the experimental data with assumptions about polarization and other macro-level dynamics made in Bounded-Confidence Opinion-Dynamics models. To be more precise, it showed (a) the deviations of the data from these assumptions and (b) what extra assumptions are needed to make it consistent with those macro-level dynamics. 
URL https://sites.google.com/view/social-influence-wiki
Description Dissemination and recruitment activities 
Form Of Engagement Activity A talk or presentation
Part Of Official Scheme? No
Geographic Reach Local
Primary Audience Public/other audiences
Results and Impact ToRealSim has been in presented to various groups as part of the experimental workshop recruitment procedure. These include the general public, students, participants in projects that concern social innovation, academics involved and or with an interest in the topic.
Year(s) Of Engagement Activity 2020,2021