STK4900/ Lecture 7. Program

Transcription

1 STK4900/ Lecture 7 Program 1. Logistic regression with one redictor 2. Maximum likelihood estimation 3. Logistic regression with several redictors 4. Deviance and likelihood ratio tests 5. A comment on model fit Sections 5.1, 5.2 (excet 5.2.6), and 5.6 Sulementary material on likelihood and deviance 1

2 Logistic regression with one redictor We have data (x 1,y 1 ),, (x n,y n ) y Here i is a binary outcome (0 or 1) for subject i and i is a redictor for the subject x We let x ( ) = Ey ( x) = Py ( = 1 x) The logistic regression models take the form: x ( ) = ex( b0 + b1x) 1 + ex( b + b x) 0 1 This gives a "S-shaed" relation between (x) and x and ensures that (x) stays between 0 and 1 2

3 The logistic model may alternatively be given in terms of the odds: x ( ) 1 - x ( ) = ex( b + b x) (*) 0 1 x + D If we consider two subjects with covariate values and x, resectively, their odds ratio becomes [ x ] [ - x] x ( + D) 1 - ( + D) x ( ) 1 ( ) = ( b + b x +D ) ex ( ) 0 1 ex( b + b x) 0 1 = ex( b D) 1 e b 1 In articular is the odds ratio corresonding to one unit's increase in the value of the covariate By (*) the logistic regression model may also be given as: é x ( ) ù log ê = b0 + b1x (**) 1 - x ( ) ú ë û Thus the logistic regression model is linear in the log-odds 3

4 Consider the WCGS study with CHD as outcome and age as redictor (individual age, not groued age as we considered in Lecture 6) R commands: wcgs=read.table("htt:// se="\t",header=t,na.strings=".") fit=glm(chd69~age, data=wcgs,family=binomial) summary(fit) R outut (edited): Estimate Std. Error z value Pr(> z ) (Intercet) < 2e-16 age e-11 The odds ratio for one year increase in age is odds ratio for a ten-year increase is e = while the (The numbers deviate slightly from those on slide 25 from Lecture 6, since there we used mean age for each age grou while here we use the individual ages) e = How is the estimation erformed for the logistic regression model? 4

5 Maximum likelihood estimation Estimation in the logistic model is erformed using maximum likelihood estimation We first describe maximum likelihood estimation for the linear regression model: For ease of resentation, we assume that σ 2 is known The density of y i takes the form (cf slide 12 from Lecture 1): 5

6 The likelihood is the simultaneous density b 1 considered as a function of the arameters and for the 0 observed values of the y i We estimate the arameters by maximizing the likelihood. This corresonds to finding the arameters that make the observed y i as likely as ossible Maximizing the likelihood L is the same as maximizing b which is the same as minimizing For the linear regression model, maximum likelihood estimation coincides with least squares estimation 6

7 We then consider the situation for logistic regression We have data (x 1,y 1 ),, (x n,y n ), where outcome (0 or 1) for subject i and Here we have Py ( = 1 x) = i i i Py ( = 0 x) = 1- i i i x i y i is a redictor is a binary where i = ex( b0 + b1xi ) 1 + ex( b + b x ) 0 1 i Thus the distribution of y i may be written as yi 1 y P( yi xi) = i (1- i) - i 7

8 The likelihood becomes n L= Õ P( yi xi) i= 1 n Õ = - i= 1 yi 1 y i (1 i) - i Since i = ex( b0 + b1xi ) 1 + ex( b + b x ) 0 1 i the likelihood is, for given observations, a function of the unknown arameters b and 0 b1 b0 1 b We estimate and by the values of these arameters that maximize the likelihood These estimates are called the maximum likelihood estimates (MLE) and are denoted ˆ b and ˆb 1 0 8

9 Confidence interval for b 1 and odds ratio 95% confidence interval for (based on the normal aroximation): OR = ex( b ) ˆ b ± 1.96 se( ˆ b ) b 1 is the odds ratio for one unit's increase in x We obtain a 95% confidence interval for OR by transforming the lower and uer limits of the confidence interval for b1 In the CHD examle we have 95% confidence interval for b 1 : ˆ b = and ± i.e. from to se( ˆ b ) = Estimate of odds ratio OR = ex(0.0744) = % confidence interval for OR : from ex(0.052) = to ex(0.096) =

10 R function for comuting odds ratio with 95% confidence limits excoef=function(glmobj) { regtab=summary(glmobj)$coef excoef=ex(regtab[,1]) lower=excoef*ex(-1.96*regtab[,2]) uer=excoef*ex(1.96*regtab[,2]) cbind(excoef,lower,uer) } excoef(fit) R outut (edited): excoef lower uer (Intercet) age

11 Wald test for H0: b 1 = 0 To test the null hyothesis alternative H : b ¹ 0 A z = 1 ˆ b1 se( ˆ b ) 1 H : b = We reject H 0 for large values of z versus the two-sided we often use the Wald test statistic: Under H 0 the test statistic is aroximately standard normal P-value (two-sided): P = 2 P(Z > z ) where Z is standard normal In the CHD examle we have Wald test statistic z = / = 6.58 ˆ b = and 1 se( ˆ b ) = which is highly significant (cf. slide 4) 11

12 Multile logistic regression Assume now that we for each subject have We let a binary outcome y redictors x1, x2,..., x x (, x,..., x) = Ey ( x, x,..., x) = Py ( = 1 x, x,..., x) Logistic regression model: x (, x,..., x) 1 2 ex( b + b x + b x b x ) = Alternatively the model may be written: 1 ex( b b x b x... b x ) æ x (, x,..., x) ö = b + b x + b x + + b x è ø 1 2 log ç 1 - x ( 1, x2,..., x) 12

13 The logistic model may also be given in terms of the odds: x (, x,..., x) x (, x,..., x) 1 2 = ex( b + b x + b x b x ) If we consider two subjects with values 1 and x 1, for the first covariate and the same values for all the others, their odds ratio becomes x ( 1+ D, x2,..., x) é ë1 - x ( 1+ D, x2,..., x) ù û x ( 1, x2,..., x) é ë1 - x ( 1, x2,..., x) ù û ex( b0 + b1( x1+d ) + b2 x b x) = = ex( b1d) ex( b + b x + b x b x ) x + D e b 1 In articular is the odds ratio corresonding to one unit's increase in the value of the first covariate holding all other covariates constant A similar interretation holds for the other regression coefficients 13

14 Wald tests and confidence intervals ˆ b = MLE for b j se( ˆ b ) = standard error for j j ˆ b j To test the null hyothesis z = ˆ b j se( ˆ b ) j H 0 j : b j = 0 we use the Wald test statistic: which is aroximately N(0,1)-distributed under H 0 j 95% confidence interval for : ˆ b ± 1.96 se( ˆ b ) OR j = ex( b ) j b j j is the odds ratio for one unit's increase in the value of the j-th covariate holding all other covariates constant j We obtain a 95% confidence interval for OR j by transforming the lower and uer limits of the confidence interval for b j 14

15 Consider the WCGS study with CHD as outcome and age, cholesterol (mg/dl), systolic blood ressure (mmhg), body mass index (kg/m 2 ), and smoking (yes, no) as redictors (as on age 152 in the text book we omit an individual with an unusually high cholesterol value) R commands: wcgs.mult=glm(chd69~age+chol+sb+bmi+smoke, data=wcgs, family=binomial, subset=(chol<600)) summary(wcgs.mult) R outut (edited): Estimate Std. Error z value Pr(> z ) (Intercet) < 2e-16 age e-08 chol e-12 sb e-06 bmi smoke e-06 15

16 Odds ratios with confidence intervals R command (using the function from slide 10): excoef(wcgs.mult) R outut (edited): excoef lower uer (Intercet) 4.50e e e-05 age chol sb bmi smoke

17 For a numerical covariate it may be more meaningful to resent an odds ratio corresonding to a larger increase than one unit (cf. slide 13) This is easily achieved by refitting the model with a rescaled covariate If you (e.g) want to study the effect of a ten-years increase in age, you fit the model with the covariate age_10=age/10 R commands: wcgs.resc=glm(chd69~age_10+chol_50+sb_50+bmi_10+smoke, data=wcgs, family=binomial, subset=(chol<600)) summary(wcgs.resc) R outut (edited): Estimate Std. Error z value Pr(> z ) (Intercet) < 2e-16 age_ e-08 chol_ e-12 sb_ e-06 bmi_ smoke e-06 Note that values of the Wald test statistic are not changed (cf. slide 15) 17

18 Odds ratios with confidence intervals: R command (using the function from slide 10): excoef(wcgs.resc) R outut (edited): excoef lower uer (Intercet) age_ chol_ sb_ bmi_ smoke

19 An aim of the WCGS study was to study the effect on CHD of certain behavioral atterns, denoted A1, A2, B3 and B4 Behavioral attern is a categorical covariate with four levels, and must be fitted as a factor in R R commands: wcgs$behcat=factor(wcgs$behat) wcgs.beh=glm(chd69~age_10+chol_50+sb_50+bmi_10+smoke+behcat, data=wcgs, family=binomial, subset=(chol<600)) summary(wcgs.beh) R outut (edited): Estimate Std. Error z value Pr(> z ) (Intercet) < 2e-16 age_ e-07 chol_ e-12 sb_ e-05 bmi_ smoke e-05 behcat behcat behcat

20 Here we may be interested in : Testing if behavioral atterns have an effect on CHD risk Testing if it is sufficient to use two categories for behavioral attern (A and B) In general we consider a logistic regression model: x (, x,..., x) 1 2 ex( b + b x + b x b x ) = ex( b b x b x... b x ) Here we want to test the null hyothesis that q of the zero, or equivalently that there are q linear restrictions among the b j 's are equal to b j 's Examles: H H : b = b = b = b = 0 ( q= 4) : b = b and b = b ( q= 2)

21 Deviance and sum of squares For the linear regression model the sum of squares was a key quantity in connection with testing and for assessing the fit of a model We want to define a quantity for logistic regression that corresonds to the sum of squares To this end we start out by considering the relation between the log-likelihood and the sum of squares for the linear regression model For the linear regression model l=log L takes the form (cf. slide 6): The log-likelihood obtains its largest value for the saturated model, i.e. the model where there are no restrictions on the μ i 21

22 For the saturated model the μ i are estimated by log-likelihood becomes, and the For a given secification of the linear regression model the μ i are estimated by the fitted values, i.e. ˆ µ ˆ, with corresonding i = yi log-likelihood lˆ n 1 s µ n 2 = - log(2 ) - 2 å i - 2 2s i = 1 ( y ˆ ) 2 i The deviance for the model is defined as and it becomes n 1 D = y - µ å 2 s i = 1 ( ˆ ) 2 i i For the linear regression model the deviance is just the sum of squares for the fitted model divided by σ 2 22

23 Deviance for binary data We then consider logistic regression with data ( y, x, x,..., x ) i = 1,2,..., n y i i 1i 2i i where is binary resonse and the are redictors x ji We introduce log-likelihood (cf. slide 8) = P( y = 1 x, x,..., x ) i i 1i 2i i l l 1 n and note that the = (,..., ) is a function of,..., 1 n For the saturated model, i.e. the model where there are no restrictions on the i, the i are estimated by the log-likelihood takes the value and 23

24 For a fitted logistic regression model we obtain the estimated robabilities ex( ˆ b0 + ˆ b ˆ 1x1i bxi) ˆi = ˆ( x1 i, x2i,..., xi) = 1 + ex( ˆ b + ˆ b x ˆ b x ) 0 1 1i i and the corresonding value ˆ ( ˆ,..., ˆ ) l = l 1 n of the log-likelihood The deviance for the model is defined as The deviance itself is not of much use for binary data But by comaring the deviances of two models, we may check if one gives a better fit than the other. 24

25 Consider the WCGS study with age, cholesterol, systolic blood ressure, body mass index, smoking and behavioral attern as redictors (cf slide 19) R outut (edited): Estimate Std. Error z value Pr(> z ) (Intercet) < 2e-16 age_ e-07 chol_ e-12 sb_ e-05 bmi_ smoke e-05 behcat behcat behcat Null deviance: on 3140 degrees of freedom Residual deviance: on 3132 degrees of freedom The deviance of the fitted model is denoted "residual deviance" in the outut The "null deviance" is the deviance for the model with no covariates, i.e. for the model where all the i are assumed to be equal 25

26 Deviance and likelihood ratio tests We want to test the null hyothesis H 0 that q of the b j 's are equal to zero, or equivalently that there are q linear restrictions among the b j 's To test the null hyothesis, we use the test statistic G = D0 - D where D 0 is the deviance under the null hyothesis and D is the deviance for the fitted model (not assuming H 0 ) We reject H 0 for large values of G To comute P-values, we use that the test statistic G is chi-square distributed (c 2 ) with q degrees of freedom under H 0 26

27 We will show how we may rewrite G in terms of the likelihood ratio We have Here where lˆ= log Lˆ and lˆ = log Lˆ 0 0 Lˆ = max L and Lˆ = max L model 0 H 0 Thus G = D0 - D ˆ ˆ = -2( l0 - l) = -2log( Lˆ ˆ) 0 L Thus large values of G corresonds to small values of the likelihood ratio Lˆ the likelihood ratio test 0 Lˆ and the test based on G is equivalent to 27

28 For the model with age, cholesterol, systolic blood ressure, body mass index, smoking, and behavioral attern as redictors (cf slide 25) the deviance becomes R commands: anova(wcgs.resc,wcgs.beh,test="chisq") R outut (edited): Analysis of Deviance Table D = For the model without behavioral attern (cf slide 17) the deviance takes the value D 0 = The test statistic takes the value: G= D0 - D= = 24.8 Model 1: chd69 ~ age_10 + chol_50 + sb_50 + bmi_10 + smoke Model 2: chd69 ~ age_10 + chol_50 + sb_50 + bmi_10 + smoke + behcat Resid.Df Resid.Dev Df Deviance P(> Chi ) e-05 28

29 Model fit for linear regression (review) 1. Linearity 2. Constant variance 3. Indeendent resonses 4. Normally distributed error terms and no outliers Model fit for logistic regression 1. Linearity: Still relevant, see following slides Var( y x ) = (1- ) 2. Heteroscedastic model, i i i i, i.e. deends on. Ey ( x) = i i i However this non-constant variance is taken care of by the maximum likelihood estimation. 3. Indeendent resonses: See Lecture 10 on Friday. 4. Not relevant, data are binary, no outliers in resonses (but there could well be extreme covariates, influential observations). 29

30 Checking linearity for logistic regression We want to check if the robabilities can be adequately described by the linear exression æ x (, x,..., x) ö = b + b x + b x + + b x è ø 1 2 log ç 1 - x ( 1, x2,..., x) We will discuss 3 aroaches: 1. Grouing the covariates 2. Adding square terms or logarithmic terms to the model 3. Extending the model to generalized additive models (GAM) æ x (, x,..., x) ö = log b0 f1( x1) f2( x2)... f( x) ç 1 - x ( 1, x2,..., x) è ø 30

31 1. Grouing the variables For a simle illustration we consider the situation where age is the only covariate in the model for CHD, and we want to check if the effect of age is linear (on the log-odds scale) The rocedure will be similar if there are other covariates in addition to age We may here fit a model considering the age grou as a factor (age grous: 35-40, 41-45, 46-50, 51-55, 56-60) Or we may fit a model where the mean age in each age grou is used as numerical covariate (means: 39.5, 42.9, 47.9, 52.8, 57.3) We may then use a deviance test to check if flexible categorical model gives a better fit than the linear numerical. Here we find no imrovement, =

32 R commands: fit.catage=glm(chd69~factor(agec), data=wcgs,family=binomial) wcgs$agem=39.5*(wcgs$agec==0)+42.9*(wcgs$agec==1)+47.9*(wcgs$agec==2)+ 52.8*(wcgs$agec==3)+57.3*(wcgs$agec==4) fit.linage=glm(chd69~agem, data=wcgs,family=binomial) summary(fit.catage) summary(fit.linage) anova(fit.linage, fit.catage,test="chisq") R outut (edited): Estimate Std. Error z value Pr(> z ) (Intercet) < 2e-16 factor(agec) factor(agec) factor(agec) factor(agec) e-05 Estimate Std. Error z value Pr(> z ) (Intercet) < 2e-16 agem e-10 Model 1: chd69 ~ agem Model 2: chd69 ~ factor(agec) Resid. Df Resid. Dev Df Deviance P(> Chi )

33 2. Adding square terms or log-terms The simle model can be extended to more flexible models such as or log é x ( ) ù ê = b 1 ( ) 0 + b1x - x ú ë û log é x ( ) ù ê = b 1 ( ) 0 + b1x + b2x - x ú ë û ( ) log é x ê ù = b + 0 b1x + b2log( x) 1 - x ( ) ú ë û We may then use a deviance test to check if the flexible models give a better fit than the original. Here we neither find any imrovement, =0.79 and =

34 R commands: fit=glm(chd69~age, data=wcgs,family=binomial) fita2=glm(chd69~age+i(age^2), data=wcgs,family=binomial) fitlog=glm(chd69~age+log(age), data=wcgs,family=binomial) anova(fit,fita2,test="chisq") anova(fit,fitlog,test="chisq") R outut (edited): > anova(fit,fita2,test="chisq") Model 1: chd69 ~ age Model 2: chd69 ~ age + I(age^2) Resid. Df Resid. Dev Df Deviance Pr(>Chi) > anova(fit,fitlog,test="chisq") Model 1: chd69 ~ age Model 2: chd69 ~ age + log(age) Resid. Df Resid. Dev Df Deviance Pr(>Chi)

35 3. Generalized additive model In this examle just with one covariate: where f1( x1) æ x ( ) ö = + è ø 1 log ç b0 f1( x1) 1 - x ( 1) is a smooth function estimated by the rogram. The aroach can easily be extended to several covariates. We can then (a) Plot the estimated function with confidence intervals. Will a straight line fit within the confidence limits? (a) Comare the simle and flexible model by a deviance test. 35

36 R outut (edited): > library(gam) Øfitgam=gam(chd69~s(age), data=wcgs, family=binomial) > lot(fitgam,se=t) > anova(fit,fitgam,test="chisq") Analysis of Deviance Table Model 1: chd69 ~ age Model 2: chd69 ~ s(age) Resid. Df Resid. Dev Df Deviance Pr(>Chi) * For these data (a) The informal grahical check just allows a straight line within confidence limits. (a) However, the deviance test gives a weakly significant deviation from linearity (=0.032) There may thus be some unimortant deviation from linearity. 36

37 Deviance and groued data On slides in Lecture 6 we saw that we got the same estimates and standard errors when we fitted the model with mean age in each age grou as numerical covariate using binary data and groued data R commands: summary(fit.linage) chd.groued=read.table("htt:// ", header=t) fit.groued=glm(cbind(chd,no-chd)~agem, data=chd.groued, family=binomial) summary(fit.groued) R outut (edited): Estimate Std. Error z value Pr(> z ) (Intercet) < 2e-16 agem e-10 Null deviance: on 3153 degrees of freedom Residual deviance: on 3152 degrees of freedom Estimate Std. Error z value Pr(> z ) (Intercet) < 2e-16 agem e-10 Null deviance: on 4 degrees of freedom Residual deviance: on 3 degrees of freedom 37

38 We see that the "residual deviance" and the "null deviance" are not the same when we use binary data and when we use groued data However, the difference between the two is the same in both cases As long as we look at differences between deviances, it does not matter whether we used binary or groued data 38