| Abstrakt: |
Consider a regression model, Y | x ∼ D (x) , where D is a parametric distribution that depends on the p × 1 vector of predictors x. Generalized linear models, generalized additive models, and some survival regression models have this form. To obtain a prediction interval for a future value of the response variable Yf given a vector of predictors x f , apply the non parametric shorth prediction interval to Y 1 * , ... , Y B * where the Y i * are independent and identically distributed from the distribution D ̂ (x f) which is a consistent estimator of D (x f). A second prediction interval modifies the shorth prediction interval to work after variable selection and if p > n where n is the sample size. Competing prediction intervals, when they exist, tend to be for one family of D (such as Poisson regression), tend to need n ≥ 10 p , and usually have not been proven to work after variable selection. [ABSTRACT FROM AUTHOR] |
