Last modified: Thu Oct 20 11:26:25 EDT 2005 by Dean Foster

Statistical Data mining: Extreme tail probabilities

Admistrivia

• Homework 3 is due in 10 days

White estimator

• beta-hat = cov(X - xhat,Y - yhat)/var(X - xhat)
• Nice sum formation
• Estimate variance of beta-hat
• Usual method: assume Y's are IID (or at least homoskadastic)
• White estimator is more robust. Doesn't need anything except independence.
• Sandwich estimator for variance. (Name comes from matrix format.)
• Called GEE's in biostatistics.

Putting them together

• Assumption: Y - yhat is symetric
• Lemma: |Y| and sgn(Y) are independent
• Use Chernoff to compute P(beta-hat > c| |Y-yhat|'s)