This project is about the empirical identification and estimation of the impact of interventions, e.g., financial regulations, public area smoking bans, extensions of compulsory schooling, or additional benefits in mandatory health insurance. The model combines the potential outcomes framework with structural model assumptions to derive bounds on the expected policy effects.
The evaluation of interventions is an integral part of active policy making. For example, one might be interested in the stabilizing effect of a financial regulation, in the health and economic consequences of a smoking ban in public areas, in the effect of a harmonized education system on student performance, or in the health impacts of extended benefits in the mandatory health insurance, to name just a few interventions relevant to the public debate in Switzerland over the last years.
The potential outcomes framework is a powerful tool to address such questions. Agents are characterized by a set of potential outcomes, i.e., outcomes that describe what the agent would achieve in each possible state. The effectiveness of the policy may then be evaluated by the individual level treatment effect which is given by the comparison of the agent's potential outcomes. The fundamental evaluation problem is that for each agent the potential outcome can only be observed in the actual state (the observed outcome), the outcomes in the other states are logically unobserved.
The aim of this project is to explore structural model assumptions with many treatment states and continuous or discrete outcomes. A sensible way to look at the policy effect is the implied shift in the population distribution of potential outcomes. Without imposing assumptions on the data-generating process such treatment effects can only be bounded, but the bounds receive the widest consensus among researchers about the policy effect of interest. Since the no-assumptions bounds are often not informative, the project proceeds by imposing weak restrictions on individual and average choice behavior. In particular, the project attempts to answer the following three questions:
1. What can be learned about distributional treatment effects if treatment response and treatment selection are weakly or piecewise monotone?
2. What are the consequences of decreasing marginal returns and decreasing marginal costs for the identification of treatment effects?
3. What is the informational content of a discrete choice random utility structure in a treatment effect model with a multinomial response?
The underlying structural models have well-established foundations in decision theory, and the maintained assumptions are as weak as to allow for arbitrary error term distributions and influence of covariates. In each case it will be shown that the structural model carries information that can be gainfully employed to narrow the no-assumptions bounds.
The project is also concerned with inference. Two approaches will be investigated: First, a classical frequentist approach in which population features are replaced by sample counterparts to construct confidence intervals and conduct hypothesis tests on the treatment effects. Second, a modern Bayesian approach in which the posterior distribution of the policy effect is approximated under a semiparametric Dirichlet specification and different priors.