COMPUTATIONAL SOCIAL SCIENCE

Department of Computational Social Science Seminar Abstract

Tuesday, May 1 - 1:00 p.m.
Research Hall, Room 162

Seyed Mohammad Mussavi Rizi
Masters of Arts in Law & Diplomacy
Tufts University

MULTI-AGENT MODEL OF DEVELOPMENT AID PROJECTS IN THE AFGHAN COUNTERINSURGENCY

In this dissertation I examine why some development projects have succeeded while others failed to complete during the post-2001 Afghan insurgency. In the first chapter I deliver the first disaggregated study of the implementation of development projects in Southern Afghanistan during 2004-2010. The analysis refutes prevalent greed, grievance and opportunity hypotheses, revealing instead non-monotonicity and threshold effects of independent variables on project success.

In the second chapter I develop a novel methodology that recasts population synthesis as a network growth problem, uses surveys and remote sensing data to synthesize rural populations in Afghanistan where non-binding constraints on data render existing algorithms ineffective, and succeeds in reproducing small-variance population samples. I propose a general-purpose method to measure the uncertainty of synthesized populations through k-nearest neighbor networks and by a nonparametric estimator of entropy, illustrating how tighter constraints on population totals reduce algorithm uncertainty.

In the third chapter I create a multiagent model of rural households in Afghanistan that represents how land properties, transportation infrastructure, fuel cost and labor markets influence agents’ decisions. Agents consume, produce and trade in local interactions, with limited information and without utility maximization while regulating their relations with powerbrokers with different tribal affiliations, wealth and man power. The model is empirically validated and wheat and poppy harvests and prices can be explained under a large range of parameter combinations.

In the fourth chapter I develop models of attitude and power within a cognitive architecture made up of an action selection heuristic, and communication and individual learning. The cognitive framework provides a blueprint to enable agents to make sense of long histories of interactions, emit messages filtered by agents’ material benefits and social identity, and anticipate the outcomes of their behaviors. Farmers and traders form beliefs on the consequences of choosing specific powerbrokers as patrons, and boundedly rational powerbrokers evolve attitudes toward development spending, manage patronage relations with farmers and traders, join and leave alliances and initiate violence.

In the fifth chapter I link real-life agricultural production and consumption in Uruzgan during 2004-2010 to geography, household demographics and prevailing security conditions, showing how competing power networks that embody longstanding, tribal and personal rivalries have influenced spatial patterns of development spending in the province. Turning to the model, I show that it reaches steady state in about eight years. Introducing historical weather, security events and development spending once the model has achieved a steady state generates historical data on wheat and poppy prices and land allocations, and project failure rates that are empirically validated against real-life data. I then compare the impact of development projects on the livelihood of farmers, traders and powerbrokers under counterfactual scenarios when either all or none of development projects is implemented, showing non-uniform distributions of changes in income over different communities. Finally, I examine the predictive power of the model by turning it into a data generating function and show that including variables such as the number of anti-status quo powerbrokers and mean alliance size along with the local control of development projects reduces classification error considerably.