Swiss Summer School 1998

Robert Luskin
The Theory of Regression

Robert Luskin

Robert C. Luskin is Associate Professor of government at the University of Texas at Austin. He holds a Ph.D. from the University of Michigan and has previously taught at the University of Alabama, at Indiana and Princeton Universities, and in the ICPSR Summer Program at the University of Michigan and the ECPR Summer School at the University of Essex. His general interests include public opinion, voting behavior, political psychology, and statistical methods, and he has long been particularly interested in the role of political information in toning the extent and direction of political participation. His published work has appeared in the American Political Science Review, the American Journal of Political Science, and other scholarly journals. Lately he has been engaged in collaborative research on the design and analysis of Deliberative Polls in both the U.S. and the U.K. He is also embarked on a study of political information in France. He is a member of the Advisory Board of the Texas Poll and of the Editorial Boards of Political Analysis and the American Political Science Review.

Workshop contents and objectives

This workshop is about regression models, roughly and broadly defined as statistical models to explain some single dependent variable. While concentrating on linear models, we shall also consider some nonlinear ones (notably including logit and probit). Computer-based exercises and lab sessions will provide practical experience, but as the title is meant to suggest, the course tries to convey more than just how to generate and read outputs. The lectures and readings will focus on more basic questions: What sorts of models imply and reflect what sorts of relationships between explanatory and dependent variables? What assumptions must we make, and what do they mean? How likely are the assumptions to be violated, and with what consequences? How can we tell when violations occur? What alternative assumptions might we make? What estimators provide statistically desirable estimates? Where several different estimators might serve, what are their advantages and disadvantages? What do the estimates tell us, and how certainly? While generally skirting proofs and derivations, we shall make heavy use of mathematical notation and mathematically phrased argument.

The topics we shall cover are:

  1. The Single-Equation Game: Model, Assumptions, and Meaning
  2. Some Basic Estimators: Ordinary Least Squares, Maximum Likelihood
  3. Forecasting, Fit, and Inference
  4. Trouble in Regression City: Heteroskedasticity and Autoregression (Introducing Generalized Least Squares)
  5. More Trouble: Collinearity and Related Ills
  6. Still More Trouble: Correlations of Regressor and Disturbance
  7. Nonlinear Models
  8. Models for Discrete Variables (Introducing Logit and Probit)
  9. Some Models for Time-Series: Lagged Variables and Distributed Lags
  10. Regression Encore: Some Matrix Notation
This is an ambitious agenda, and I am not sure we can cover it all in one week, but I do hope to get at least through models for discrete variables.

Basic text/overview
Damodar Gujarati, Basic Econometrics (3rd ed.).McGraw-Hill.

Henry J. Cassidy and A.H. Studenmund, Using Econometrics: A Practical Guide (3rd ed.). Addison-Wesley.

Remedial Reading
Jan Kmenta, Elements of Econometrics (2nd ed.), chs. 1-6.

Good facility with ordinary (scalar) algebra. At least one semester of statistics, at roughly the level of chs. 1-6 of Kmenta or above, enough to be at home with concepts like random variables, probability distributions, and expectations.

 

[Workshops]