Modeling Interference for Policy Evaluation in Stochastic Systems
Interference is a phenomenon where the treatment of one unit may affect the outcomes of other units. It is a major consideration for accurate policy evaluation in stochastic systems. Although this may seem intractable in a non-parametric causal inference setting, I will demonstrate that in many problems, some lightweight modeling can significantly aid in capturing and quantifying these interference effects. In particular, I will discuss two different forms of interference:
(1) Network interference. In the network interference model, units are represented as vertices on an exposure graph (for example, a social network). In this model, the treatment assigned to one unit may affect the outcomes of other units connected to it through edges in the graph. I will discuss large-sample asymptotics for treatment effect estimation under network interference, where the exposure graph is a random draw from a graphon.
(2) Congestion induced interference. In service systems, stochastic congestion can arise from temporarily limited supply and/or demand. Such congestion gives rise to interference between the waiting customers, and analytic strategies that do not account for this interference may be biased. I will discuss the potential of using knowledge about the congestion mechanism to design and analyze experiments in the presence of stochastic congestion.
I am currently a postdoctoral fellow working with Professor Susan Murphy in the Department of Statistics at Harvard University. Prior to this, I earned my Ph.D. from the Department of Statistics at Stanford University, where I was advised by Professors Emmanuel Candès and Stefan Wager.
]]>
Interference is a phenomenon where the treatment of one unit may affect the outcomes of other units. It is a major consideration for accurate policy evaluation in stochastic systems. Although this may seem intractable in a non-parametric causal inference setting, I will demonstrate that in many problems, some lightweight modeling can significantly aid in capturing and quantifying these interference effects. In particular, I will discuss two different forms of interference:
(1) Network interference. In the network interference model, units are represented as vertices on an exposure graph (for example, a social network). In this model, the treatment assigned to one unit may affect the outcomes of other units connected to it through edges in the graph. I will discuss large-sample asymptotics for treatment effect estimation under network interference, where the exposure graph is a random draw from a graphon.
(2) Congestion induced interference. In service systems, stochastic congestion can arise from temporarily limited supply and/or demand. Such congestion gives rise to interference between the waiting customers, and analytic strategies that do not account for this interference may be biased. I will discuss the potential of using knowledge about the congestion mechanism to design and analyze experiments in the presence of stochastic congestion.
]]>