|
Conference on Quantitative
Social Science Research Using R
(click on the title to download the abstract)
|
|
Andrew Gelman
|
|
Bayesian generalized linear models and an appropriate default prior
|
Many statistical methods of all sorts have tuning parameters. How can
default settings for such parameters be chosen in a general-purpose
computing environment such as R? We consider the example of prior
distributions for logistic regression.
Logistic regression is an important statistical method in its own right
and also is commonly used as a tool for classification and imputation.
The standard implementation of logistic regression in R, glm(), uses
maximum likelihood and breaks down under separation, a problem that
occurs often enough in practice to be a serious concern. Bayesian
methods can be used to regularize (stabilize) the estimates, but then
the user must choose a prior distribution. We illustrate a new idea,
the "weakly informative prior," and implement it in bayesglm(), a slight
alteration of the existing R function. We also perform a
cross-validation to compare the performance of different prior
distributions using a corpus of datasets.
|
|
|
|
