Thoughts on Agent-Based Economic Geography Models
Unpacking the polarization of workplace skills. Ahmad Alabdulkareem, Morgan R. Frank*, Lijun Sun, Bedoor AlShebli,César Hidalgo, Iyad Rahwan (2018) Science Advances Vol. 4.
While not an agent-based model, this papers is in the field of complexity regional economics. The authors analyze O*NET data from the Department of Labor. Specifically, the authors use location quotients to analyze various skills as a fraction of occupations as compared with the the use of the skill more broadly. Then using conditional probabilities the authors show interdependence between various skills and map skills in a network where the nodes are skills and the interdependence between two skills forms the edge. Overall, the authors uncover a polarization in the set of workforce skills that are correlated with income.
Churning, power, laws, and inequality in a spatial agent-based model of social networks. Jae Beum Cho, Yuri S. Mansury, and Xinyue Ye (2016) Annals of Regional Science Vol. 57. Special Issue
This paper introduces several novel elements into the preferential attachment model to demonstrate how geography impacts the degree distribution of social networks and social capital inequality. The most important element the modelers introduce (for me) is space.
The novel elements introduced to the preferential attachment model include space, visibility, tie decay, and social capital. Space is introduced by placing agents, who have heterogeneous human capital, on a grid and adjusting the likelihood that agents attach to one another based on the distance between the agents. In addition to introducing geography into the preferential attachment model, the modelers also allow for network churning; agents in the network are allowed to re-evaluate their ties and dissolve those that are not worth the effort. In allowing for tie-reformation, the agents are also classified into two groups, introverts and extroverts. Introverts can only connect to those they can see while extroverts can attach to any agent in the model. For tie-decay, agents examine their ties by the number of ties they currently maintain, and the distance associated with each tie. Links to highly connected agents are likely to be deleted as are those links to distant agents. Finally, the modelers introduce the concept of social capital by summing the total human capital of an agent’s direct ties. The agent’s goal is to increase their social capital through fruitful connections, which is constrained by the cost of maintaining ties.
Among the results, allowing for new tie formation results in the power-law nature of the preferential attachment model to break-down. Allowing for both tie formation and tie dissolution results in the power-law nature of connections reappearing, as long as the parameters are approximately equal, or dissolution is greater than formation. A very interesting result is that churning, with both tie formation and dissolution, results in more unequal social capital than when no churning is present. Churning also diminishes the advantage of being in the system longer. Tie dissolution increases social capital inequality while tie formation decreases inequality.
Results from the spatial analysis indicate that geographic position matters. Initially, as network density increases, agents at the center of the grid have the highest social capital. However, as the network density moves toward saturation, those on the edges increase their social capital, which coincides with a reduction in inequality. These are identified as being the result of local and global positions which benefit from the costs associated with maintaining distant ties. Spatial constraints also inhibit introverts compared with extroverts who are not burdened by distance to maintain ties.
The paper is a great example of work that introduces spatial features into a well-known model. The results that are derived from the addition of space are both novel and enlightening. Moving forward, there are a number of possible directions that the model go foreseeable go. First, the agents could be allowed to migrate. This would result in something like a combination of Schelling’s segregation model and the current model. It would be interesting to see if social capital segregates in a similar manner as the original Schelling model. Second, it would be interesting to scale the model up by including many more agents. Doing this might require that the model allow for more than one agent on a single cell, which would move it in the direction of a model of city formation.
Knowledge networks in regional development: an agent-based model and its applications. Tamas Sebestyen & Attilla Varga (2019) Regional Studies Vol. 53 No. 9
The model by Sebestyen and Varga is a tool to advise regions in the European Union on how to select partners in the Framework Program to access the most extra-regional knowledge for regional economic purposes.
The agent-based model uses econometric data as well as a genetic algorithm to calibrate the model. The likelihood of any two regions interacting is based on their individual R&D expenditures, their cognitive proximity, and their social distance.
To set the model in motion, the authors provide an exogenous shock that alters the attractiveness of each region for cooperation. As I understand, the model involves the altered attractiveness (via the exogenous shock) to alter the social distance between regions, and thus their likelihood of finding it mutually beneficial to collaborate on patents and so forth.
As a policy simulation, the authors provide Central Hungary with an increase of 1 percent to its Cognitive Proximity to all other regions to determine which other region provides the greatest increase in access to extra-regional knowledge. The authors find that the 1% increase of Cognitive Proximity to Oberbayern, Germany increases its access to extra-regional knowledge by 2.1%, more than any other region. A possible policy suggestion is to increase Central Hungary’s patenting in “performing operations and transportation”, this would result in bringing the region 1% closer in Cognitive Proximity to Oberbayern, thus increasing its accessibility to extra-regional knowledge by 2.1%.
The paper is a good example of agent-based modeling working towards application. The model is explicitly designed to be applied and simplistic enough that it might be used by a wider audience. That being said, the calibration required to set the model up appears cumbersome and opaque. This may limit wide appeal.
One possible way the others could alter the model is to use a social network instead of the gravity model. Each region could be a node with linkages being represented by co-patenting or similar patenting output. The current attractiveness measure is a function of mass and distance. Instead, the authors could set up a social network and use some network measures such as centrality or betweenness as the measure of attractiveness.