topic 6: estimation and prediction of y h

31
Topic 6: Estimation and Prediction of Y h

Upload: shadi

Post on 05-Feb-2016

29 views

Category:

Documents


0 download

DESCRIPTION

Topic 6: Estimation and Prediction of Y h. Outline. Estimation and inference of E(Y h ) Prediction of a new observation Construction of a confidence band for the entire regression line. Estimation of E(Y h ). - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Topic 6: Estimation and Prediction of Y h

Topic 6: Estimation and Prediction of Yh

Page 2: Topic 6: Estimation and Prediction of Y h

Outline

• Estimation and inference of E(Yh)

• Prediction of a new observation

• Construction of a confidence band for the entire regression line

Page 3: Topic 6: Estimation and Prediction of Y h

Estimation of E(Yh)

• E(Yh) = μh = β0 + β1Xh, the mean value of Y for the subpopulation with X=Xh

• We will estimate E(Yh) by

• KNNL use for this estimate, see equation (2.28) on pp 52

^

h 0 1ˆ hb b X hY

Page 4: Topic 6: Estimation and Prediction of Y h

Theory for Estimation of E(Yh)

• is Normal with mean μh and variance

• The Normality is a consequence of the fact that b0 + b1Xh is a linear combination of Yi’s

• See KNNL pp 52-54 for details

2

h2 22

i

X X1ˆ

n X Xh

h

Page 5: Topic 6: Estimation and Prediction of Y h

Application of the Theory• We estimate σ2( ) by

• It then follows that

• Details for confidence intervals and significance tests are consequences

h2

2 2 hh 2

i

X X1ˆ( )

n X X

( )( )

s s

h h

h

ˆ E(Y )~ t n 2

ˆ( )( )

s

Page 6: Topic 6: Estimation and Prediction of Y h

95% Confidence Interval for E(Yh)

• ± tcs( )

where tc = t(.975, n-2)

• NOTE: significance tests can be constructed but they are rarely used in practice

h h

Page 7: Topic 6: Estimation and Prediction of Y h

Toluca Company Example (pg 19)

• Manufactures refrigeration equipment

• One replacement part manufactured in lots of varying sizes

• Company wants to determine the optimum lot size

• To do this, company needs to first describe the relationship between work hours and lot size

Page 8: Topic 6: Estimation and Prediction of Y h

hours

100

200

300

400

500

600

lotsize

20 30 40 50 60 70 80 90 100 110 120

Scatterplot w/ regr line

Page 9: Topic 6: Estimation and Prediction of Y h

***Generating the data set***;data toluca; infile ‘../data/CH01TA01.txt'; input lotsize hours;data other; size=65; output; size=100; output;data toluca1; set toluca other;proc print data=toluca1; run;

SAS CODE

Page 10: Topic 6: Estimation and Prediction of Y h

***Generating the confidence intervals for all values of X in the data set***;

proc reg data=toluca1; model hours=size/clm; id lotsize;run;

SAS CODE

clm option generates confidence intervals for the

mean

Page 11: Topic 6: Estimation and Prediction of Y h

Parameter Estimates

Variable DFParameter

EstimateStandard

Error t Value Pr > |t|Intercept 1 62.36586 26.17743 2.38 0.0259

lotsize 1 3.57020 0.34697 10.29 <.0001

Output Statistics

Obs lotsizeDependent

VariablePredicted

ValueStd Error

Mean Predict 95% CL Mean1 80 399.0000 347.9820 10.3628 326.5449 369.4191

25 70 323.0000 312.2800 9.7647 292.0803 332.479726 65 . 294.4290 9.9176 273.9129 314.945127 100 . 419.3861 14.2723 389.8615 448.9106

Page 12: Topic 6: Estimation and Prediction of Y h

Notes

• Standard error affected by how far Xh is from (see Figure 2.6)

• Recall teeter-totter idea…a change in the slope has bigger impact on Y as you move away from

X

X

Page 13: Topic 6: Estimation and Prediction of Y h

Prediction of Yh(new)

• Want to predict value for a new observation at X=Xh

• Model: Yh(new) = β0 + β1Xh +

• Since E(e)=0 same value as for E(Yh)

• Prediction interval, however, relies heavily on assumption that e are Normally distributed

Note!!

Page 14: Topic 6: Estimation and Prediction of Y h

Prediction of Yh(new)

• Var(Yh(new))=Var( )+Var(ξ )

• Then follows that

2

h2 22

i

X X1pred 1

n X X( )s s

h

( ) hˆ( ) / (pred) ~ ( 2)h newY s t n

Page 15: Topic 6: Estimation and Prediction of Y h

Notes

• Procedure can be modified for the mean of m observations at X=Xh (see 2.39a and 239b on page 60)

• Standard error affected by how far Xh is from (see Figure 2.6)X

Page 16: Topic 6: Estimation and Prediction of Y h

***Generating the prediction intervals for all values of X in data set***;

proc reg data=toluca1; model hours=lotsize/cli; id lotsize;run;

SAS CODE

cli option generates prediction interval for a

new observation

Page 17: Topic 6: Estimation and Prediction of Y h

Output Statistics

Obs lotsizeDependent

VariablePredicted

ValueStd Error

Mean Predict 95% CL Predict1 80 399.0000 347.9820 10.3628 244.7333 451.2307

25 70 323.0000 312.2800 9.7647 209.2811 415.278926 65 . 294.4290 9.9176 191.3676 397.490427 100 . 419.3861 14.2723 314.1604 524.6117

These are wrong…same as before. Does not include variability about regression line

Page 18: Topic 6: Estimation and Prediction of Y h

Notes

• The standard error (Std Error Mean Predict)given in this output is the standard error of not s(pred)

• The prediction interval is correct and wider than the previous confidence interval

h

Page 19: Topic 6: Estimation and Prediction of Y h

Notes

• To get correct standard error need to add the variance about the regression line

2 ˆ( ) ( )hs pred s MSE

Page 20: Topic 6: Estimation and Prediction of Y h

Confidence band for regression line

• ± Ws( )

where W2=2F(1-α; 2, n-2) • This gives combined “confidence intervals”

for all Xh

• Boundary values of confidence bands define a hyperbola

• Will be wider at Xh than single CI

h h

Page 21: Topic 6: Estimation and Prediction of Y h

Confidence band for regression line

• Theory comes from the joint confidence region for (β0, β1 ) which is an ellipse (Stat 524)

• We can find an alpha for tc that gives the same results

• We find W2 and then find the alpha for tc

that will give W = tc

Page 22: Topic 6: Estimation and Prediction of Y h

data a1; n=25; alpha=.10; dfn=2; dfd=n-2; tsingle=tinv(1-alpha/2,dfd); w2=2*finv(1-alpha,dfn,dfd); w=sqrt(w2); alphat=2*(1-probt(w,dfd)); t_c=tinv(1-alphat/2,dfd); output;proc print data=a1;run;

SAS CODE

Page 23: Topic 6: Estimation and Prediction of Y h

n alpha dfn dfd tsingle w2 w alphat t_c25 0.1 2 23 1.71387 5.09858 2.25800 0.033740 2.25800

SAS OUTPUT

Used for single 90% CI

Used for 90%

confidence band

Page 24: Topic 6: Estimation and Prediction of Y h

symbol1 v=circle i=rlclm97;

proc gplot data=toluca; plot hours*lotsize; run;

SAS CODE

Page 25: Topic 6: Estimation and Prediction of Y h

hours

100

200

300

400

500

600

lotsize

20 30 40 50 60 70 80 90 100 110 120

Page 26: Topic 6: Estimation and Prediction of Y h

Estimation of E(Yh) and Prediction of Yh

• = b0 + b1Xh

h

2i

2h22

2i

2h2

h2

)()(

)()(

XX

XX

n

11pred)

XX

XX

n

1)(

(s

ˆs

Page 27: Topic 6: Estimation and Prediction of Y h

symbol1 v=circle i=rlclm95; proc gplot data=toluca; plot hours*lotsize;

symbol1 v=circle i=rlcli95; proc gplot data=toluca; plot hours*lotsize;

run;

SAS CODE

Confidence intervals

Prediction intervals

Page 28: Topic 6: Estimation and Prediction of Y h

hours

100

200

300

400

500

600

lotsize

20 30 40 50 60 70 80 90 100 110 120

Confidence band

Page 29: Topic 6: Estimation and Prediction of Y h

hours

100

200

300

400

500

600

lotsize

20 30 40 50 60 70 80 90 100 110 120

Confidence intervals

Page 30: Topic 6: Estimation and Prediction of Y h

hours

100

200

300

400

500

600

lotsize

20 30 40 50 60 70 80 90 100 110 120

Prediction intervals

Page 31: Topic 6: Estimation and Prediction of Y h

Background Reading

• Program topic6.sas has the code for the various plots and calculations;

• Sections 2.7, 2.8, and 2.9