# See next slide 47 esl chapter 4 linear methods for

This preview shows page 1. Sign up to view the full content.

This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: o oo oooooo o ooooo o o oooooo oo ooo o ooooooo o o oooooo o o oo ooo oo o ooo ooo o o oooo ooo ooooo o ooo o oo oo o o ooo oo o o ooo oo ooooo ooo o ooooooo o o oooo o oo o ooo o oooooo o oo oo ooooo oo oo o oo o o oooooo o oo oooo o o oooo o o o oo ooooo o ooooooo oo o age 20 0 50 100 0.0 0.4 famhist 20 40 60 A scatterplot matrix of the South African heart disease data. Each plot shows a pair of risk factors, and the cases (160) and controls (302) are color coded (red is a case). The variable famhist (family history of heart disease) is binary (yes or no). 44 ESL Chapter 4 — Linear Methods for Classiﬁcation Trevor Hastie and Rob Tibshirani Results from a logistic regression ﬁt to the South African heart disease data. Coefﬁcient Standard Error Z Score (Intercept) −4.130 0.964 −4.285 sbp 0.006 0.006 1.023 tobacco 0.080 0.026 3.034 ldl 0.185 0.057 3.219 famhist 0.939 0.225 4.178 obesity -0.035 0.029 −1.187 alcohol 0.001 0.004 0.136 age 0.043 0.010 4.184 45 ESL Chapter 4 — Linear Methods for Classiﬁcation Trevor Hastie and Rob Tibshirani Model building ˆ • Deviance: dev(y, p) = −2 (β ) ˆ • H0 : First q components of β are non-zero • H1 : β is unrestricted • under H0 , dev(y, p0 ) − dev(y, p1 ) ∼ χ2−q asymptotically (as ˆ ˆ p N → ∞) • “Chi-square statistic”—quadratic approximation to deviance: n n wi (zi − i=1 ˆ xT β ) i = i=1 (yi − pi )2 ˆ pi (1 − pi ) ˆ ˆ ˆ • β ∼ N (β, (XT WX)−1 ) asymptotically, if the model is correct. 46 ESL Chapter 4 — Linear Methods for Classiﬁcation Trevor Hastie and Rob Tibshirani Case-control sampling and logistic regression • In South African data, there are 160 cases, 302 controls — π = 0.35 ˜ are cases. Yet the prevalence of MI in this region is π = 0.05. • With case-control samples, we can estimate the regression parameters βj accurately; the constant term β0 is incorrect. • We can correct the estimated intercept by a simple transformation ˆ∗ ˆ β0 = β0 + log π ˜ π − log 1−π 1−π ˜ • Often cases are rare and we t...
View Full Document

## This document was uploaded on 03/10/2014 for the course STATS 315A at Stanford.

Ask a homework question - tutors are online