Summary: | <p>A common complaint about structural work in applied microeconomics is that its methods
are insufficiently transparent and its findings are not robust enough (see e.g. Angrist and Pischke, 2010). This critique seems especially on target in the context of complex dynamic models
with large numbers of parameters, which are the subject of this thesis. In this literature, the
important consequences of seemingly technical assumptions are often opaque even to the researchers themselves. Any two researchers may make different choices about how to specify
and estimate a model, leading to different results.</p>
<p>For the most part, the target of this critique is the specification of the structural models
themselves. Critics complain that models are likely to be misspecified; that they rely on unrealistic assumptions about agent knowledge and rationality; or that the model parameters are
not identified given the data, even assuming that the model captures the true structure of the
data generating process. These concerns have received a lot of attention in recent work: many
structural models now feature deviations from full information or rationality (e.g. Abaluck
and Adams-Prassl, 2021), new econometric techniques promise more transparency about the
effects of identifying assumptions (e.g. Andrews, Gentzkow and Shapiro, 2017, 2020), and it is
now much more common in applied work to see explicit (formal or informal) arguments about
identification than it was in the 1990s.</p>
<p>A separate issue, which has received less attention, is a lack of transparency and insufficient
robustness of parameter estimates stemming from estimation procedures. Even if a model is
well-specified and (point-)identified, parameter estimates may be far away from the true parameters if estimators are biased or inefficient. Furthermore, the limits of computational precision can lead to estimation ‘failures’ even when estimators would be efficient with unlimited
precision. These issues are especially prominent for dynamic discrete choice models, which
are particularly challenging to estimate.</p>
<p>Achieving reasonable efficiency and robustness in the estimation of complex dynamic discrete choice models often requires substantial computational resources. This creates sharp
trade-offs for applied researchers between computational costs, the accuracy of parameter estimates, and the complexity of models that can be estimated. Despite large advances in computing power, it is still common to estimate models relying on imprecise approximations, which
can introduce economically meaningful bias into the parameter estimates. As a consequence,
the results of counterfactual policy experiments can also be misleading.</p>
<p>In this thesis, I propose new tools for estimating these models with a higher degree of accuracy at any given level of computational resources and model complexity. The thesis consists
of this introduction, three separate research papers, and a conclusion. The first paper proposes
a new technique for simulating the likelihood in finite-horizon dynamic discrete choice models
with observed payoffs that makes better use of the information contained in observed payoffs
to guide the simulation of choice probabilities. The second paper (with Jack Britton) proposes
an extension to the Generalized Indirect Inference (GII) method of Bruins et al. (2018), which
makes it practicable to estimate complex dynamic discrete choice models using GII. We illustrate our method by employing it to estimate a model of the career choices of young people in
the UK, which we use to assess the long-run effects of a conditional cash transfer scheme. The
third paper (also with Jack Britton) proposes three improvements to the approximate solution
method of Keane and Wolpin (1994), which lead to much higher accuracy without increasing
the required computation time. Unfortunately, the presentation as separate research papers
means that a small amount of material is repeated across papers/chapters.</p>
|