{[ promptMessage ]}

Bookmark it

{[ promptMessage ]}

Chapter05

# Chapter05 - Transformations and Weighting STAT 563 Spring...

This preview shows pages 1–25. Sign up to view the full content.

Transformations and Weighting STAT 563 Spring 2007

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
Model Assumptions Common violations are: Expression for the expected value of Y is not correct The variance is not constant over the range of the data The data are not normally distributed One remedy to all these violations is to transform the data Reasonable to develop a model in terms of some function of the response Or transform the predictors
Heteroscedasticity Constancy of error variance is one of the standard assumptions ( homoscedasticity) When the error variance is not constant, the error is said to be heteroscedastic Residuals tend to have a funnel-shaped distribution, either fanning out or closing in with the values of X

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
Hypothetical Example Residuals X
Heteroscedasticity We will learn how to Detect heteroscedasticity Its effects on the analysis Remove heteroscedasticity

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
Simple Example Number of injury incidents (y) and the proportion of total flights (n) for nine major airlines in a single year are given • If f i denote the total flights for the i th airline, then the proportion of total flights n i made by the i th airline is If all airlines are equally safe, the injury incidents can be explained by the model = i i i f f n i i i n y ε β + + = 1 0

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
Comments • Residuals are seen to increase with n i Assumption of homoscedasticity seems to be violated Not surprising, injury incidents may behave as a Poisson variable which has a variance proportional to its mean Try square root transformation of the response Made the residual plot little better, still R 2 is still only 48% Consider other factors (besides proportion of total flights) for a better explanation of injury incidents

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
Transformations to stabilize variance Probability distribution of Y Var(Y) in terms of its mean μ Transformation Poisson μ Binomial μ(1-μ29/ n Negative Binomial μ+λ 2 μ 2 ) 1 ( + + Y Y or Y Y 1 sin - ) ( sinh 1 1 Y λ - - Here variance is a function of the mean response

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
Weighted Least Squares Used for stabilizing variance when the variance is a simple function of one of the predictors or is dependent on a known set of weights For example, based on some empirical evidence, standard deviation of residuals is proportional to X 0 , ) ( 2 2 = k x k Var i i ε
Weighted Least Squares For a simple linear regression model, Divide both sides by x i and get Define new set of variables i i i x y ε β + + = 1 0 i i i i i x x x y + + = 1 0 X X X X Y Y = = = = = * , , , 1 * , * 0 * 1 1 * 0

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
WLS In terms of the new variables Note for the transformed model If our assumption regarding the variance holds, then we should work with the transformed model * * * 1 * 0 * i i i x y ε β + + = t cons is Var k x k x Var x x Var Var i i i i i i i i tan ) ( ) ( 1 ) ( 1 ) ( ) ( * 2 2 2 2 2 * = = = =
Example Study of 27 industrial establishments Number of supervisors (Y) Number of workers (X)

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
Clear indication of increasing trend for residual variance with X
Fit Y/X on 1/X

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document

This preview has intentionally blurred sections. Sign up to view the full version.

View Full Document
> fitw <- lm(super~work,weights=1/ (work**2)) > e.res <- ls.diag(fitw)\$stud.res > e.fit <- ls.diag(fitw)\$fitted
This is the end of the preview. Sign up to access the rest of the document.

{[ snackBarMessage ]}

### Page1 / 72

Chapter05 - Transformations and Weighting STAT 563 Spring...

This preview shows document pages 1 - 25. Sign up to view the full document.

View Full Document
Ask a homework question - tutors are online