|
|
Jeff Gill
Visiting Associate Professor (F06-S07)
|
 |
I am a Visiting Associate Professor in the Government Department,
at Harvard University (2006-2007),
and currently an associate editor of the journal Political Analysis.
Normally I am an Associate Professor in the Department of Political Science
at the University of California, Davis.
I am involved with one research group here at Harvard, the Institute for Quantitative Social Science.
(IQSS), which has an active and interesting
Social Science Statistics Blog.
My research applies Bayesian modeling and data analysis (decision theory, testing, model selection,
elicited priors) to questions in general social science quantitative methodology, American
political behavior and institutions, focusing on Congress, the bureaucracy, and voters, using
computationally intensive tools (Monte Carlo methods, MCMC, stochastic optimization, non-parametrics).
- Government 1007. Quantitative Analysis of American Politics - (New Course)
Catalog Number: 5315
Half course (fall term). M., W., at 10. EXAM GROUP: 3
This course introduces the application of basic quantitative methods to the study of American Politics. The emphasis is on
how mathematical and statistical tools help understand political behavior and institutions in the U.S. Topics include
measurement, measures of association, comparison of means, plotting, and regression. The class will provide tools for
analyzing political data and presentation of results in research papers, class essays, and theses. No prior mathematics or
government courses required.
Syllabus
- Government 2003. Hierarchical Bayesian Modeling
Catalog Number: 3713
Half course (fall term). M., 2-4, EXAM GROUP: 7, 8
Provides students a solid understanding of Bayesian inference and Markov chain Monte Carlo methods. Topics covered include:
Bayesian treatment of the linear model, Markov chain Monte Carlo methods, assessing model adequacy, and hierarchical models.
Prerequisite: Government 1000 and Government 2000, the equivalents, or permission of the instructor.
Syllabus
- Government 2002. Topics in Quantitative Methods
Catalog Number: 8168
Half course (spring term). W., 4-6, EXAM GROUP: 9
Provides students a solid understanding of statical computing issues that concern empirical
researchers in the social sciences. Topics covered include: computer architecture, Monte
Carlo simulation, bootstrapping and jackknifing, nonparametric smoothing, and Markov chain
Monte Carlo methods. Prerequisite: Government 2000 and Government 2001, the equivalents, or
permission of the instructor.
Syllabus
The following lists some of my current projects. Working papers can be furnished upon request.
- Dynamic Adaptive Markov Chains for MCMC.
Self-tuning Markov chain algorithms that respond to surface characteristics and adapt are
a relatively new approach and one that still requires theoretical defense as well
as applied demonstration. We (Jeff Gill and George Casella) have already shown that with one
scheme we can analyze previously intractable voting models in high dimensions. The ongoing
work here seeks to generalize the approach with a theoretical justification that
stability properties can be preserved. We currently have a funded {\em National Science Foundation}
(Mathematical Social and Behavioral Sciences program $350,000) grant for work to develop
a family of nonparametric priors (mixtures of Dirichlet processes) and apply this technology to the
resulting estimation challenges.
- Modeling Qualitative Information with Elicited Priors: Confidence in Judicial and
Bureaucratic Institutions in Central America.
Researchers who wish to systematically combine qualitative and quantitative information
in the same model have found few helpful procedures thus far. However, elicitation
and probabilistic interpretation of expert opinion have the potential to bring together
previously disparate approaches in political science. To bridge the divide, we (Jeff Gill
and Lee Walker) developed and explained procedures to format qualitative, descriptive, and
narrative information for inclusion into a standard empirical model. We extend
this work to other elites in Central America and demonstrate that much more can be learned
about nascent governments with the Elicited-Bayesian approach than by using strictly
quantitative methods (i.e. survey research) or strictly qualitative methods (i.e. unstructured
elite interviews).
- Is Partial-Dimension Convergence a Problem for MCMC Algorithms?
General issues of Markov chain convergence have greatly concerned statisticians since widespread
use of MCMC began after 1990. Unfortunately, no attention has been given to the problem of
evidential convergence amongst a only subset of the dimensions of the chain. Yet the structure
of at least one MCMC kernel (probably the most commonly used), the Gibbs sampler, is
built around marginal draws conditional on other dimensions. Therefore, even the
parameters that appear to be converged are conditioned on non-convergent parameters.
In this article, currently under ``revise and resubmit'' at \emph{Political Analysis},
I prove that the Gibbs sampler needs to have convergence in every dimension to assert
convergence in any dimension. The paper includes a different finding for the
Metropolis-Hastings algorithm, and concludes with three empirical examples.
- Queueing Theory Models for Political Science Data.
Queueing theory is widely used in many literatures to describe assembly-line type processes,
and services in which the completion time is indeterminant. However, there are almost no
applications of queueing theory in political science. This is curious because political
actors queue up for desired benefits under a number of circumstances. This project with
Cherie Goodenough (UC Davis graduate student) looks at the theoretical and practical basis
for applying of queueing theory to the analysis of institutional politics. Empirical
applications include the process of bills through legislatures, the scheduling and hearing
of court cases, and initiatives within international institutions.
- Bayesian Hierarchical Modeling for Survey Aggregation with Integrated Ecological Inferences.
Integrating survey results, population demographics, and individual-level information is
a critical and persistent problem. Obviously such cross-level inference invites a host
of methodological and substantive challenges, and various solutions have been proposed.
We (Bob Huckfeldt, Beth Simas, and Jeff Gill, all of UC Davis) seek to address this
problem with Bayesian hierarchical models that use the explicit structure of priors and
hyperpriors to reconcile differing levels of data aggregation and explicitly integrate
ecological inferences. The approach promises to separate and protect diverse sources of
aggregation and variance.
- Bayesian Methods: A Social and Behavioral Sciences Approach, Second Edition.
I am currently under contract to produce a second edition of the Bayesian book from 2002.
Publication will be in 2007. This is a field that moves rapidly, making an update important.
There will be a greater emphasis on MCMC tools and the theory behind their use as well as a
continued emphasis on regression-style models. A review of the first edition can be found
in the Journal of Politics 65(3):909-911, and two more in are available at
The Political Methodologist.
- Principal-Agent Analysis of Bureaucratic Complexity with an Application to the Department of
Homeland Security.
Principal-agent models long have been the subject of studies of political bureaucratic relations.
In this project, we (Jeff Gill and Rick Waterman) introduce structural
complexity and its implications for political control of the bureaucracy to the principal-agent model.
Using the establishment of the Department of Homeland Security as an example, we argue that complex
structures introduce another dimension which complicates principal attempts to control the bureaucracy.
More importantly, it also can introduce situations where agents within the same bureaucracy have vastly
different policy preferences. Under these circumstances the benefits of the information asymmetry can
be mitigated, as dueling agents inside the same agency have incentives to reach
out to different principals in an attempt to build political coalitions.
- Multiple Imputation for Missing Categorical Data in Political Science.
Multiple imputation is the most substantial improvement in the handling of missing data. The ease of
current software solutions the have made the process of imputing missing data quite easy. Unfortunately,
though, the underlying engine typically assumes missingness on a continuous metric. Attempts to rectify
this and provide helpful software for categorical multiple imputation with regression-style applications
has been only partially successful. Currently Skyler Cranmer (UC Davis graduate student, currently here
at Harvard as a Teaching Assistant) and I have developed a modernized hot-decking
procedure for imputing missing categorical data along the lines of multiple imputation but with a
method that preserves the structure of the discreteness as measured. We are currently in the
process of justifying the theoretical properties and developing an
R package.
- Elicited Priors for National Security Research.
In work partly funded by the government last year (2005), John Freeman (University of Minnesota) and I have
developed procedures for eliciting structured information on networks for updating Bayesian models of
connectivity through an online software system. The intention is to have a procedure for sampling and
comparing qualitative information that national security analysts possess that cannot otherwise be directly
inserted into a model specification. Updating network connectivity is an interesting question unto itself,
but the application to political actors adds challenging and fascinating complexities since information is
not only qualitative but often expressed in different terms and contexts. We have a basic software module
freely available to researchers (provided no commercial use). Despite the difficulties listed, prior
information in a Bayesian context is an ideal means of obtaining and describing levels of uncertainty that
analysts possess.
- The Rule of 30: A Different Agenda for New Political Methodology.
Chris Achen's 2003 paper is an important and widely read critique of directions in political
methodology that introduces A Rule of Three (ART): ``A statistical specification with more
than three explanatory variables is meaningless.'' The idea is that we (political scientists)
are unable to understand or justify the assumptions or assess the model fit with larger
specifications without strong theoretical grounding to help substantively. Achen's core advice
is to control for other variables in the sample rather than in the model, setting aside subgroups
that are small, inconvenient, or messy. This moves complexity in final quality assessment of
results from analysis of the model to analysis of the data. Conversely, it is my assertion that
we are much better as a discipline in sorting out good and poor models, and that our toolbox here
is extensive. What we lack is a correspondingly well-developed set of methodologies for
assessing the quality of observational data, with its concomitant measurement error, coding
error, and possibly sampling error.
- Effects of Politics and Economics on Country-Level FMD-Status.
(R.B. Garabed, A.M. Perez, W.O. Johnson, J. Gill and M.C. Thurmond).
The primary objective of the study presented in this paper is to quantify globally the effects of
governance and economic health on the country-level presence of FMD (foot-and-mouth disease). A
secondary objective is to use country-level information to estimate the probability that FMD was
present in each country in the world for each year from 1997 to 2006. By quantifying the effects
of governance and economic health on FMD presence and by providing good global FMD-presence
prediction, we hope to provide information that will improve future FMD control and eradication
plans.