Transcript VT PowerPoint Template
Logistic and Poisson Regression: Modeling Binary and Count Data LISA Short Course Series Mark Seiss, Dept. of Statistics November 5, 2008
Presentation Outline 1.
2.
3.
Introduction to Generalized Linear Models Binary Response Data Logistic Regression Model Count Response Data Poisson Regression Model
Reference Material
Categorical Data Analysis – Alan Agresti
Examples found with SAS Code at www.stat.ufl.edu/~aa/cda/cda.html
Presentation and Data from Examples
www.stat.vt.edu/consult/short_courses.html
Generalized Linear Models
• • • Generalized linear models (GLM) extend ordinary regression to non-normal response distributions.
3 Components • • • Random – identifies response Y and its probability distribution Systematic – explanatory variables in a linear predictor function (X β) Link function – function (g(.)) that links the mean of the response (E[Y i ]= μ i ) to the systematic component.
Model • g i
j
j x ij
for i = 1 to n
Generalized Linear Models
• Why do we use GLM’s?
• • Linear regression assumes that the response is distributed normally GLM’s allow us to analyze the linear relationship between predictor variables and the mean of the response variable when it is not reasonable to assume the data is distributed normally.
Generalized Linear Models
• Predictor Variables • Two Types: Continuous and Categorical • • Continuous Predictor Variables • • Examples – Time, Grade Point Average, Test Score, etc.
Coded with one parameter – β i x i Categorical Predictor Variables • • Examples – Sex, Political Affiliation, Marital Status, etc.
Actual value assigned to Category not important • • Ex) Sex - Male/Female, M/F, 1/2, 0/1, etc.
Coded Differently than continuous variables
Generalized Linear Models
• Categorical Predictor Variables cont.
• Consider a categorical predictor variable with L categories • One category selected as reference category • Assignment of Reference Category is arbitrary • • Variable represented by L-1 dummy variables • Model Identifiability Two types of coding – Dummy and Effect
Generalized Linear Models
• Categorical Predictor Variables cont.
• Dummy Coding (Used in R) • • x x k k = 1 if predictor variable is equal to category k 0 otherwise = 0 for all k if predictor variable equals category I • Effect Coding (Used in JMP) • x k = 1 if predictor variable is equal to category k 0 otherwise • x k = -1 for all k if predictor variable equals category I
Generalized Linear Models
•
Saturated Model
• • • Contains a separate indicator parameter for each observation Perfect fit μ = y Not useful since there is no data reduction, i.e. number of parameters equals number of observations.
• Maximum achievable log likelihood – baseline for comparison to other model fits
Generalized Linear Models
• Deviance • Let L( μ|y) = maximum of the log likelihood for the model • • L(y|y) = maximum of the log likelihood for the saturated model Deviance = D(y| μ) = -2 [L(μ|y) - L(y|y) ] • • Likelihood Ratio Statistic for testing the null hypothesis that the model is a good alternative to the saturated model Likelihood ratio statistic has an asymptotic chi-squared distribution with N – p degrees of freedom, where p is the number of parameters in the model.
Allows for the comparison of one model to another using the likelihood ratio test.
Generalized Linear Models
• Nested Models • • • Model 1 - model with p predictor variables {X 1 , X 2 , X 3 ,….,X p } and vector of fitted values μ 1 Model 2 - model with q
X
1 + … + p
X
p + 0
X
p 1 + 0
X
p 2 + … 0
X
q
Generalized Linear Models
• Likelihood Ratio Test • Null Hypothesis: There is not a significant difference between the fit of two models.
• Null Hypothesis for Nested Models: The predictor variables in Model 1 that are not found in Model 2 are not significant to the model fit.
• Alternate Hypothesis for Nested Models - The predictor variables in Model 1 that are not found in Model 2 are significant to the model fit.
• Likelihood Ratio Statistic = -2* [L(y,u 2 )-L(y,u 1 )] = D(y, μ 2 ) - D(y, μ 1 ) • • Difference of the deviances of the two models Always D(y, μ 2 ) > D(y, μ 1 ) implies LRT > 0 LRT is distributed Chi-Squared with p-q degrees of freedom
Generalized Linear Models
• Likelihood Ratio Test cont.
• Later, we will use the Likelihood Ratio Test to test the significance of variables in Logistic and Poisson regression models.
Generalized Linear Models
• Theoretical Example of Likelihood Ratio Test • 3 predictor variables – 1 Continuous (X 1 ), 1 Categorical with 4 Categories (X 2 , X 3 , X 4 ), 1 Categorical with 1 Category (X 5 ) • Model 1 - predictor variables {X 1 , X 2 , X 3 , X 4 , X 5 } • • Model 2 - predictor variables {X 1 , X 5 } Null Hypothesis – Variables with 4 categories is not significant to the model ( β 2 = β 3 = β 4 = 0) • Alternate Hypothesis - Variable with 4 categories is significant • Likelihood Ratio Statistic = D(y, μ 2 ) - D(y, • μ 1 ) Difference of the deviance statistics from the two models • Chi-Squared Distribution with 5-2=3 degrees of freedom
Generalized Linear Models
• Model Selection • 2 Goals: Complex enough to fit the data well Simple to interpret, does not overfit the data • • Study the effect of each predictor on the response Y • • Continuous Predictor – Graph P[Y=1] versus X Discrete Predictor - Contingency Table of P[Y=1] versus categories of X Unbalance Data – Few responses of one type • Guideline – 10 outcomes of each type for each X terms • Example – Y=1 for only 30 observations out of 1000 Model should contain no more than 3 X terms
Generalized Linear Models
• Model Selection cont.
• Multicollinearity • • • Correlations among predictors resulting in an increase in variance Reduces the significance value of the variable Occurs when several predictor variables are used in the model • Determining Model Fit • Other criteria besides significance tests (i.e. Likelihood Ratio Test) can be used to select a model
Generalized Linear Models
• Model Selection cont.
• Determining Model Fit cont.
• Akaike Information Criterion (AIC) – Penalizes model for having many parameters – AIC = Deviance+2*p where p is the number of parameters in model • Bayesian Information Criterion (BIC) – BIC = -2 Log L + ln(n)*p where p is the number of parameters in model and n is the number of observations
Generalized Linear Models
• Model Selection cont.
• Selection Algorithms • • Best subset – Tests all combinations of predictor variables to find best subset Algorithmic – Forward, Backward and Stepwise Procedures
Generalized Linear Models
• Best Subsets Procedure • Run model with all possible combinations of the predictor variables • Number of possible models equal to 2 predictor variables p where p is the number of • • Dummy Variables for categorical predictors considered together Ex) For a set of predictors {X 1 , X 2 , X 3 } • runs models with sets of predictors {X 1 , X 2 , X 3 }, {X 1 , X 2 }, • {X 2 , X 3 }, {X 1 , X 3 }, {X 1 }, {X 2 }, {X 3 }, and no predictor variables.
2 3 = 8 possible models • Most programs only allow for a small set of predictor variables • • Cannot be run in a reasonable amount of time 2 10 = 1024 models run for a set of 10 predictor variables
Generalized Linear Models
• Forward Selection • Idea: Start with no variables in the model and add one at a • • • • time Step One: Step Two: Step Three: Add each variable to the model one at a Step Four: Fit model with single predictor variable and determine fit Select predictor variable with best fit and add to model time and determine fit If at least one variable produces better fit, return to step two • If no variables produce better fit, use model Drawback: Variables Added to the model cannot be taken out.
Generalized Linear Models
• • Backward Selection • Idea: Start with all variables in the model and take out one at a time • Step One: Fit all predictor variables in model and determine fit • Step Two: Delete one variable at a time and determine fit • Step Three: If the deletion of at least one variable produces better fit, remove variable that produces best fit when deleted and return to step 2 If the deletion of a variable does not produce a better fit, use model Drawback: Variables taken out of model cannot be added back in.
Generalized Linear Models
• Stepwise Selection • Idea: Combination of forward and backward selection • Forward Step then backward step • • • • • • • • Step One: Fit each predictor variable as a single predictor variable and determine fit Step Two: Select variable that produces best fit and add to model.
Step Three: Add each predictor variable one at a time to the model and determine fit Step Four: Select variable that produces best fit and add to the model Step Five: Delete each variable in the model one at a time and determine fit Step Six: Remove variable that produces best fit when deleted Step Seven: Return to Step Two Loop until no variables added or deleted improve the fit.
Generalized Linear Models
• Summary • 3 Components of the GLM • • • Random (Y) Link Function (g(E[Y])) Systematic (x t β) • Continuous and Categorical Predictor Variables • Coding Categorical Variables – Effect and Dummy Coding • • Likelihood Ratio Test for Nested Models • Test the significance of a predictor variable or set of predictor variables in the model.
Model Selection – Best Subset, Forward, Backward, Stepwise
Generalized Linear Models
•
Questions/Comments
Logistic Regression
• Consider a binary response variable.
• Variable with two outcomes • One outcome represented by a 1 and the other represented by a 0 • Examples: Does the person have a disease? Yes or No Who is the person voting for?
McCain or Obama Outcome of a baseball game? Win or loss
Logistic Regression
• Logistic Regression Example Data Set • Response Variable –> Admission to Grad School (Admit) • 0 if admitted, 1 if not admitted • Predictor Variables • GRE Score (gre) – Continuous • University Prestige (topnotch) – 1 if prestigious, 0 otherwise • Grade Point Average (gpa) – Continuous
Logistic Regression
• First 10 Observations of the Data Set 0 1 0 1 0 0 1 0 ADMIT 1 0 GRE 380 660 800 640 520 760 560 400 540 700 0 0 0 1 1 0 0 0 TOPNOTCH 0 1 GPA 3.61
3.67
4 3.19
2.93
3 2.98
3.08
3.39
3.92
Logistic Regression
• Consider the linear probability model
E
i
P
(
Y
i
0 |
x
i
) (
x
i
)
x
i
where y i = response for observation i x i = 1x(p+1) matrix of covariates for observation i p = number of covariates • • • GLM with binomial random component and identity link g( μ) = μ Issue: π(X i ) can take on values less than 0 or greater than 0 Issue: Predicted probability for some subjects fall outside of the [0,1] range.
Logistic Regression
• Consider the logistic regression model
E
i
P
(
Y i
0 |
x i
) (
x i
) 1 exp exp
x i
x i
log
it
i
log 1
i
i
x
i
• • GLM with binomial random component and identity link g( μ) = logit( μ) Range of values for π(X i ) is 0 to 1
Logistic Regression
• Consider the logistic regression model log
it
i
*
gpa
i
And the linear probability model (
x
i
) *
gpa
i
Then the graph of the predicted probabilities for different grade point averages:
Important Note: JMP models P(Y=0) and effect coding is used for categorical variables
Logistic Regression
Logistic Regression
• Interpretation of Coefficient β – Odds Ratio • The odds ratio is a statistic that measures the odds of an event compared to the odds of another event.
• Say the probability of Event 1 is π 1 π 2 and the probability of Event 2 is . Then the odds ratio of Event 1 to Event 2 is:
Odds
_
Ratio
Odds
( 1
Odds
( 2 ) ) 1 1 1 2 1 2 • • • • Value of Odds Ratio range from 0 to Infinity Value between 0 and 1 indicate the odds of Event 2 are greater Value between 1 and infinity indicate odds of Event 1 are greater Value equal to 1 indicates events are equally likely
Logistic Regression
• Interpretation of Coefficient β – Odds Ratio cont.
• Link to Logistic Regression :
Log
(
Odds
_
Ratio
)
Log
( 1 1 1 )
Log
( 1 2 2 )
Logit
( 1 )
Logit
( 2 ) • Thus the odds ratio between two events is
Odds
_
Ratio
exp{
Logit
( 2 )
Logit
( 1 )}
Logistic Regression
• Interpretation of Coefficient β – Odds Ratio cont.
• Consider Event 1 is Y=0 given X and Event 2 is Y=0 given X+1
Log
(
Odds
_
Ratio
)
Logit
(
P
(
Y
0 |
X
1 ))
Logit
(
P
(
Y
0 |
X
)) • From our logistic regression model ( (
X
1 )) (
X
) • Thus the ratio of the odds of Y=0 for X and X+1 is
Odds
_
Ratio
exp( )
Logistic Regression
• Single Continuous Predictor Variable - GPA
Generalized Linear Model Fit
Response: Admit Modeling P(Admit=0) Distribution: Binomial Link: Logit Observations (or Sum Wgts) = 400
Whole Model Test Model -LogLikelihood
Difference Full Reduced 6.50444839
243.48381
249.988259
L-R ChiSquare
13.0089
DF
1
Goodness Of Fit Statistic
Pearson 401.1706
398 Deviance 486.9676
398
ChiSquare
0.4460
0.0015
DF
398 398
Prob>ChiSq
0.4460
0.0015
Prob>ChiSq
0.0003
Logistic Regression
• Single Continuous Predictor Variable – GPA cont.
Effect Tests Source
GPA
DF
1
L-R ChiSquare Prob>ChiSq
13.008897
0.0003
Parameter Estimates Term Estimate
Intercept GPA -4.357587
1.0511087
Std Error
1.0353175
0.2988695
L-R ChiSquare Prob>ChiSq
19.117873
<.0001
13.008897
0.0003
Lower CL
-6.433355
0.4742176
Upper CL
-2.367383
1.6479411
Interpretation of the Parameter Estimate: Exp{1.0511087} = 2.86 = odds ratio between the odds at x+1 and odds at x for all x The ratio of the odds of being admitted between a person with a 3.0 gpa and 2.0 gpa is equal to 2.86 or equivalently the odds of the person with the 3.0 is 2.86 times the odds of the person with the 2.0.
I
Logistic Regression
• Single Categorical Predictor Variable – Top Notch
Generalized Linear Model Fit
Response: Admit Modeling P(Admit=0) Distribution: Binomial Link: Logit Observations (or Sum Wgts) = 400
Whole Model Test Model -LogLikelihood
Difference Full Reduced 3.53984692
246.448412
249.988259
Goodness Of Fit Statistic
Pearson Deviance
L-R ChiSquare DF
7.0797
1
Prob>ChiSq
0.0078
ChiSquare
400.0000
492.8968
DF
398 398
Prob>ChiSq
0.4624
0.0008
Logistic Regression
• Single Categorical Predictor Variable – Top Notch cont.
Effect Tests Source
TOPNOTCH
DF
1
L-R ChiSquare
7.0796939
Prob>ChiSq
0.0078
Parameter Estimates Term
Intercept TOPNOTCH[0]
Estimate
-0.525855
-0.371705
Std Error
0.138217
0.138217
L-R ChiSquare Prob>ChiSq
14.446085
0.0001
7.0796938
0.0078
Lower CL
-0.799265
-0.642635
Upper CL
-0.255667
-0.099011
Interpretation of the Parameter Estimate: Exp{2*-.371705} = 0.4755 = odds ratio between the odds of admittance for a student at a less prestigous university and the odds of admittance for a student from a more prestigous university.
The odds of being admitted from a less prestigous university is .48 times the odds of being admitted from a more prestigous university.
Logistic Regression
• • Variable Selection – Likelihood Ratio Test • Consider the model with GPA, GRE, and Top Notch as predictor variables
Generalized Linear Model Fit
Response: Admit Modeling P(Admit=0) Distribution: Binomial Link: Logit Observations (or Sum Wgts) = 400
Whole Model Test Model -LogLikelihood
Difference Full Reduced 10.9234504
239.064808
249.988259
Goodness Of Fit Statistic
Pearson Deviance
L-R ChiSquare
21.8469
DF
3
ChiSquare
396.9196
478.1296
DF
396 396
Prob>ChiSq
0.4775
0.0029
Prob>ChiSq
<.0001
Logistic Regression
• Variable Selection – Likelihood Ratio Test cont.
Effect Tests Source DF L-R ChiSquare
TOPNOTCH 1 2.2143635
GPA GRE 1 1 4.2909753
5.4555484
Prob>ChiSq
0.1367
0.0383
0.0195
Parameter Estimates Term Estimate
Intercept TOPNOTCH[0] GPA GRE -4.382202
-0.218612
0.6675556
0.0024768
Std Error
1.1352224
0.1459266
0.3252593
0.0010702
L-R ChiSquare Prob>ChiSq
15.917859
<.0001
2.2143635
4.2909753
5.4555484
0.1367
0.0383
0.0195
Lower CL
-6.657167
-0.503583
0.0356956
0.0003962
Upper CL
-2.197805
0.070142
1.3133755
0.0046006
Logistic Regression
• Model Selection – Forward
Stepwise Fit
Response: Admit
Stepwise Regression Control
Prob to Enter 0.250
Prob to Leave Direction: 0.100
Rules:
Current Estimates -LogLikelihood RSquare
239.06481
0.0437
Logistic Regression
• Model Selection – Forward cont.
Parameter
Intercept[1]
Estimate
-4.3821986
GRE GPA 0.00247683
0.66755511
TOPNOTCH{1-0} 0.21861181
1 1 1
nDF
1
Wald/Score ChiSq
0 5.356022
4.212258
2.244286
"Sig Prob"
1.0000
0.0207
0.0401
0.1341
1 2
Step History Step
3
Parameter
GRE GPA TOPNOTCH{1-0}
Action
Entered Entered Entered
L-R ChiSquare "Sig Prob"
13.92038
0.0002
5.712157
2.214363
0.0168
0.1367
RSquare
0.0278
0.0393
0.0437
3 4
p
2
Logistic Regression
• Model Selection – Backward • Start by selecting to enter all variables into the model
Stepwise Fit
Response: Admit
Stepwise Regression Control
Prob to Enter Prob to Leave 0.250
0.100
Direction: Backward Rules: Combine
Logistic Regression
• Model Selection – Backward cont.
Current Estimates -LogLikelihood RSquare
240.17199
0.0393
Parameter
Intercept[1] GRE GPA TOPNOTCH{1-0} 0
Estimate
-4.9493751
0.00269068
0.75468641
1 1
nDF
1 1
Wald/Score ChiSq
0 6.473978
5.576461
2.259729
"Sig Prob"
1.0000
0.0109
0.0182
0.1328
Step History Step
1
Parameter
TOPNOTCH{1-0}
Action
Removed
L-R ChiSquare "Sig Prob"
2.214363
0.1367
RSquare
0.0393
p
3
Logistic Regression
• Variable Selection – Stepwise
Stepwise Fit
Response: Admit
Stepwise Regression Control
Prob to Enter Prob to Leave 0.250
0.250
Direction: Mixed Rules: Combine
Current Estimates -LogLikelihood RSquare
239.06481
0.0437
Logistic Regression
• Variable Selection – Stepwise cont.
Parameter
Intercept[1] GRE
Estimate
-4.3821986
0.00247683
GPA 0.66755511
TOPNOTCH{1-0} 0.21861181
1 1
nDF
1 1
Wald/Score ChiSq
0 5.356022
4.212258
2.244286
"Sig Prob"
1.0000
0.0207
0.0401
0.1341
2 3
Step History Step
1
Parameter
GRE GPA TOPNOTCH{1-0}
Action
Entered Entered Entered
L-R ChiSquare
13.92038
5.712157
2.214363
"Sig Prob"
0.0002
0.0168
0.1367
Rsquare
0.0278
0.0393
0.0437
3 4
p
2
Logistic Regression
• Summary • Introduction to the Logistic Regression Model • • Interpretation of the Parameter Estimates β – Odds Ratio Variable Significance – Likelihood Ratio Test • Model Selection • • • Forward Backward Stepwise
Logistic Regression
•
Questions/Comments
Poisson Regression
• Consider a count response variable.
• • Response variable is the number of occurrences in a given time frame.
Outcomes equal to 0, 1, 2, ….
• Examples: Number of penalties during a football game.
Number of customers shop at a store on a given day.
Number of car accidents at an intersection.
Poisson Regression
• Poisson Regression Example Data Set • Response Variable –> Number of Days Absent – Integer • Predictor Variables • • • • • • Gender- 1 if Female, 2 if Male Ethnicity School – 6 Ethnic Categories – 1 if School, 2 if School 2 Math Test Score Bilingual Status – Continuous Language Test Score – Continuous – 6 Bilingual Categories
Poisson Regression
• First 10 Observations from the Poisson Regression Example Data Set 1 GENDER ethnicity school.1.or.2 ctbs.math.nce ctbs.lang.nce bilingual.status number.days.absent
2 4 1 56.988830 42.45086 2 4 2 3 4 2 4 1 1 4 1 1 4 1 37.094160 46.82059 2 4 32.275460 43.56657 2 2 29.056720 43.56657 2 3 5 6 7 8 9 10 1 4 1 1 4 1 1 4 1 2 4 1 2 4 1 2 6 1 6.748048 27.24847 3 3 61.654280 48.41482 0 13 56.988830 40.73543 2 11 10.390490 15.35938 2 7 50.527950 52.11514 2 10 49.472050 42.45086 0 9
Poisson Regression
• Consider the model
E
i
i
x
i
where Y i = response for observation i x i = 1x(p+1) matrix of covariates for observation i p = number of covariates μ i = expected number of events given x i • • GLM with poisson random component and identity link g( μ) = μ Issue: Predicted values range from ∞ to +∞
Poisson Regression
• Consider the Poisson log-linear model
E
Y
i
|
x
i
i
exp
x
i
log
i
x i
• • • GLM with poisson random component and log link g( μ) = log(μ) Predicted response values fall between 0 and +∞ In the case of a single predictor, An increase of one unit of x results an increase of exp( β) in μ
Poisson Regression
• Consider the Poisson log-linear model log
i
*
Math
_
Score
i
And the Poisson linear model
i
x
*
Math
_
Score
i
Then a graph of the predicted values from the model:
Poisson Regression
12 10 8 6 4 2 0 0
Predicted Number of Days Absent Versus Match Score
20 40 60
Math Score
80 100 120 Predicted Days Absent - Log Link Predicted Days Absent - Identity Link
Poisson Regression
• Single Continuous Predictor Variable – Math Score > fitline<-glm(number.days.absent~ctbs.math.nce,data=poisson_data,family=poisson(link=log)) > summary(fitline) Call: glm(formula = number.days.absent ~ ctbs.math.nce, family = poisson(link = log), data = poisson_data) Deviance Residuals: Min 1Q Median 3Q Max -4.4451 -2.5583 -1.0842 0.6647 12.4431 Coefficients: (Intercept) ctbs.math.nce Estimate Std. Error z value Pr(>|z|) 2.302100 0.062776 36.671 <2e-16 *** -0.011568 0.001294 -8.939 <2e-16 *** -- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Poisson Regression
• Single Continuous Predictor Variable – Math Score (Dispersion parameter for poisson family taken to be 1) Null deviance: 2409.8 on 315 degrees of freedom Residual deviance: 2330.6 on 314 degrees of freedom AIC: 3196 Number of Fisher Scoring iterations: 6 Interpretation of the parameter estimate: Exp{-0.011568} = .98 = multiplicative effect on the expected number of days absent for an increase of 1 in the Math Score Fabricated Example – If a student is expected to miss 5 days with a math of 50, then another student with a math score of 51 is expected to miss 5*.98 = 4.9 days
Poisson Regression
• Single Continuous Predictor Variable – Gender > fitline<-glm(number.days.absent~factor(GENDER),data=poisson_data,family=poisson(link=log)) > summary(fitline) Call: glm(formula = number.days.absent ~ factor(GENDER), family = poisson(link = log), data = poisson_data) Deviance Residuals: Min 1Q Median 3Q Max -3.660 -2.755 -1.128 0.902 9.738 Coefficients: Estimate Std. Error z value Pr(>|z|) (Intercept) factor(GENDER)2 1.90174 0.03036 62.644 < 2e-16 *** -0.31729 0.04747 -6.684 2.32e-11 *** -- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Poisson Regression
• Single Continuous Predictor Variable – Gender (Dispersion parameter for poisson family taken to be 1) Null deviance: 2409.8 on 315 degrees of freedom Residual deviance: 2364.5 on 314 degrees of freedom AIC: 3229.9
Number of Fisher Scoring iterations: 5 Important Note: The function factor(categorical variable) uses the dummy coding Interpretation of the parameter estimate: Exp{-0.31729} = 0.7289 = multiplicative effect on the expected number of days absent of being male rather than female If a female student is expected to miss X days, then a male student is expected to miss 0.7289*X.
Poisson Regression
• • Variable Selection – Likelihood Ratio Test Model with all variables > fitline<-glm(number.days.absent~factor(GENDER)+factor(school.1.or.2)+ctbs.math.nce+ctbs.lang.nce+factor(bilingual.status)+ factor(ethnicity),data=poisson_data,family=poisson(link=log)) summary(fitline) Call: glm(formula = number.days.absent ~ factor(GENDER) + factor(school.1.or.2) + ctbs.math.nce + ctbs.lang.nce + factor(bilingual.status) + factor(ethnicity), family = poisson(link = log), data = poisson_data) Deviance Residuals: Min 1Q Median 3Q Max -4.5222 -2.1863 -0.9622 0.7454 10.4077
Poisson Regression
• • Variable Selection – Likelihood Ratio Test Model with all variables Cont > Coefficients: Estimate Std. Error z value Pr(>|z|) (Intercept) 2.972325 0.424645 7.000 2.57e-12 *** factor(GENDER)2 -0.401980 0.048954 -8.211 < 2e-16 *** factor(school.1.or.2)2 -0.582321 0.070717 -8.235 < 2e-16 *** ctbs.math.nce -0.001043 0.001845 -0.565 0.57181 ctbs.lang.nce -0.003048 0.002003 -1.521 0.12822 factor(bilingual.status)1 -0.344696 0.083754 -4.116 3.86e-05 *** factor(bilingual.status)2 -0.282194 0.070846 -3.983 6.80e-05 *** factor(bilingual.status)3 -0.053406 0.081850 -0.652 0.51409 factor(ethnicity)2 -0.131202 0.420704 -0.312 0.75515 factor(ethnicity)3 -0.434061 0.418013 -1.038 0.29909 factor(ethnicity)4 -0.326230 0.419158 -0.778 0.43639 factor(ethnicity)5 -0.876270 0.416398 -2.104 0.03534 * factor(ethnicity)6 -1.188835 0.457470 -2.599 0.00936 ** -- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Poisson Regression
• • Variable Selection – Likelihood Ratio Test Model with all variables Cont (Dispersion parameter for poisson family taken to be 1) Null deviance: 2409.8 on 315 degrees of freedom
Residual deviance: 1909.2 on 303 degrees of freedom
AIC: 2796.6
Number of Fisher Scoring iterations: 6
Poisson Regression
• • Variable Selection – Likelihood Ratio Test Model with all variables except Ethnicity >fitline
Poisson Regression
• • Variable Selection – Likelihood Ratio Test Model with all variables except Ethnicity Coefficients: (Intercept) Estimate 2.5741133 factor(GENDER)2 -0.4212841 factor(school.1.or.2)2 -0.8242109 ctbs.math.nce 0.0008193 Std. Error 0.0838754 0.0484383 0.0570241 0.0018278 ctbs.lang.nce -0.0050753 factor(bilingual.status)1 -0.3080131 factor(bilingual.status)2 -0.1815997 factor(bilingual.status)3 0.0363656 0.0019380 0.0762534 0.0581877 0.0686396 -- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 z value 30.690 -8.697 -14.454 0.448 -2.619 -4.039 -3.121 0.530 Pr(>|z|) < 2e-16 *** < 2e-16 *** < 2e-16 *** 0.65398 0.00882 ** 5.36e-05 *** 0.00180 ** 0.59625
Poisson Regression
• • Variable Selection – Likelihood Ratio Test Model with all variables except Ethnicity (Dispersion parameter for poisson family taken to be 1) Null deviance: 2409.8 on 315 degrees of freedom
Residual deviance: 1984.1 on 308 degrees of freedom
AIC: 2861.5
Number of Fisher Scoring iterations: 6
Poisson Regression
• Variable Selection – Likelihood Ratio Test • Model 1 with All Variables – Deviance = -2 Log L = 1909.2 with df = 303 • Model 2 without Ethnicity - Deviance = -2 Log L = 1984.1 with • • df = 308 Likelihood Ratio Test = Deviance (Model 2) – Deviance (Model 1) = 1984.1 – 1909.2= 74.9
Likelihood Ratio Test ~ Chi Square with 308-303 = 5 degrees of freedom • • P-Value < .0001
There is significant evidence to conclude that ethnicity is a significant predictor variable.
Poisson Regression
• Model Selection • Forward Selection > fitline<-glm(number.days.absent~1,data=data1,family=poisson(link=log)) > step(fitline,scope = list(upper = ~factor(GENDER)+factor(school.1.or.2)+ctbs.math.nce+ctbs.lang.nce+factor(bilingual.status)+factor(ethnicity), lower = ~1),direction="forward") Start: AIC=3273.22
number.days.absent ~ 1 Df + factor(school.1.or.2) 1 + factor(ethnicity) 5 + ctbs.lang.nce 1 + ctbs.math.nce 1 + factor(bilingual.status) 3 + factor(GENDER) 1
2969.3
3177.0
3196.0
3208.6
3229.9
3273.2
-
Poisson Regression
• Model Selection • Forward Selection cont.
Step: AIC=2969.12
number.days.absent ~ factor(school.1.or.2) Df + factor(ethnicity) 5 + factor(GENDER) 1 + factor(bilingual.status) 3 + ctbs.lang.nce 1 + ctbs.math.nce 1
2896.7
2937.4
2960.1
2964.1
2969.1
Poisson Regression
• Model Selection • Forward Selection cont.
Step: AIC=2894.07
number.days.absent ~ factor(school.1.or.2) + factor(ethnicity) Df + factor(GENDER) 1 Deviance 1951.3 AIC 2828.7
+ factor(bilingual.status) 3 + ctbs.math.nce 1 + ctbs.lang.nce 1
2888.5
2889.9
2894.1
Step: AIC=2828.67
number.days.absent ~ factor(school.1.or.2) + factor(ethnicity) + factor(GENDER) Df Deviance AIC + factor(bilingual.status) 3 1915.3 2798.8
+ ctbs.lang.nce 1 + ctbs.math.nce 1
2821.7
2828.7
Poisson Regression
• Model Selection • Forward Selection cont.
Step: AIC=2798.75
number.days.absent ~ factor(school.1.or.2) + factor(ethnicity) + factor(GENDER) + factor(bilingual.status) Df + ctbs.lang.nce 1 + ctbs.math.nce 1
2796.9
2798.8
Step: AIC=2794.89
number.days.absent ~ factor(school.1.or.2) + factor(ethnicity) + factor(GENDER) + factor(bilingual.status) + ctbs.lang.nce
Df
2796.6
Poisson Regression
• Model Selection • Forward Selection cont.
Call: glm(formula = number.days.absent ~ factor(school.1.or.2) + factor(ethnicity) + factor(GENDER) + factor(bilingual.status) + ctbs.lang.nce, family = poisson(link = log), data = data1) Coefficients: (Intercept) factor(school.1.or.2)2 factor(ethnicity)2 factor(ethnicity)3 factor(ethnicity)4 2.948689 -0.586678 -0.126806 -0.423376 -0.313360 factor(ethnicity)5 factor(ethnicity)6 factor(GENDER)2 factor(bilingual.status)1 factor(bilingual.status)2 -0.862743 -1.175574 -0.404215 -0.343907 -0.284027 factor(bilingual.status)3 ctbs.lang.nce -0.051558 -0.003763 Degrees of Freedom: 315 Total (i.e. Null); 304 Residual Null Deviance: 2410
Poisson Regression
• Model Selection cont.
• Backward Selection > fitline<-glm(number.days.absent~factor(GENDER)+factor(school.1.or.2)+ctbs.math.nce+ctbs.lang.nce+factor(bilingual.status)+ factor(ethnicity),data=poisson_data,family=poisson(link=log)) > backwards<-step(fitline,direction="backward") Start: AIC=2796.57
number.days.absent ~ factor(GENDER) + factor(school.1.or.2) + ctbs.math.nce + ctbs.lang.nce + factor(bilingual.status) + factor(ethnicity) Df -
ctbs.math.nce 1
1909.5
1909.2 1911.5 1937.8 1984.1 1977.8 1983.6 AIC
2794.9
2796.6
2796.9
2819.2
2861.5
2863.2
2869.0
Poisson Regression
• Model Selection cont.
• Backward Selection cont.
Step: AIC=2794.89
number.days.absent ~ factor(GENDER) + factor(school.1.or.2) + ctbs.lang.nce + factor(bilingual.status) + factor(ethnicity) Df
2798.8
2817.8
2859.7
2862.8
2869.9
Poisson Regression
• Model Selection cont.
• Stepwise Selection cont.
> fitline<-glm(number.days.absent~1,data=data1,family=poisson(link=log)) > step(fitline,scope = list(upper=~factor(GENDER)+factor(school.1.or.2)+ctbs.math.nce+ctbs.lang.nce+factor(bilingual.status)+factor(ethnicity), lower = ~1),direction="both") Start: AIC=3273.22
number.days.absent ~ 1 Df + factor(school.1.or.2) 1 + factor(ethnicity) 5 + ctbs.lang.nce 1 + ctbs.math.nce 1 + factor(bilingual.status) 3 + factor(GENDER) 1
2969.3
3177.0
3196.0
3208.6
3229.9
3273.2
Poisson Regression
• Model Selection cont.
• Stepwise Selection cont.
Step: AIC=2969.12
number.days.absent ~ factor(school.1.or.2) Df + factor(ethnicity) 5 + factor(GENDER) 1 + factor(bilingual.status) 3 + ctbs.lang.nce 1 + ctbs.math.nce 1
2896.7
2937.4
2960.1
2964.1
2969.1
3273.2
Poisson Regression
• Model Selection cont.
• Stepwise Selection cont.
• • • • • • • • • • Step: AIC=2894.07
number.days.absent ~ factor(school.1.or.2) + factor(ethnicity) Df + factor(GENDER) 1 Deviance 1951.3 AIC 2828.7
+ factor(bilingual.status) + ctbs.math.nce + ctbs.lang.nce 3 1 1 1981.6 2011.1 2012.5 2863.0
2888.5
2889.9
2969.1
2969.3
Poisson Regression
• Model Selection cont.
• Stepwise Selection cont.
Step: AIC=2828.67
number.days.absent ~ factor(school.1.or.2) + factor(ethnicity) + factor(GENDER) Df Deviance AIC + factor(bilingual.status) + ctbs.lang.nce + ctbs.math.nce
2817.8
2821.7
2828.7
- factor(GENDER) - factor(ethnicity) - factor(school.1.or.2) 1 5 1 2018.7 2029.3 2050.5 2894.1
2896.7
2925.9
Poisson Regression
• Model Selection cont.
• Stepwise Selection cont.
Step: AIC=2798.75
number.days.absent ~ factor(school.1.or.2) + factor(ethnicity) + factor(GENDER) + factor(bilingual.status) Df + ctbs.lang.nce 1 Deviance 1909.5 AIC 2794.9
+ ctbs.math.nce 1
2798.8
2828.7
2863.0
2866.8
2884.8
Poisson Regression
• Model Selection cont.
• Stepwise Selection cont.
• • Step: AIC=2794.89
number.days.absent ~ factor(school.1.or.2) + factor(ethnicity) + factor(GENDER) + factor(bilingual.status) + ctbs.lang.nce
+ ctbs.math.nce 1 - ctbs.lang.nce 1 - factor(bilingual.status) 3 1909.2 1915.3 1938.5 2796.6
2798.8
2817.8
- factor(ethnicity) 5 - factor(GENDER) 1 - factor(school.1.or.2) 1 1984.3 1979.4 1986.5 2859.7
2862.8
2869.9
Poisson Regression
• Model Selection cont.
• Stepwise Selection cont.
Call: glm(formula = number.days.absent ~ factor(school.1.or.2) + factor(ethnicity) + factor(GENDER) + factor(bilingual.status) + ctbs.lang.nce, family = poisson(link = log), data = data1) Coefficients: (Intercept) factor(school.1.or.2)2 factor(ethnicity)2 factor(ethnicity)3 factor(ethnicity)4 2.948689 -0.586678 -0.126806 -0.423376 -0.313360 factor(ethnicity)5 factor(ethnicity)6 factor(GENDER)2 factor(bilingual.status)1 factor(bilingual.status)2 -0.862743 -1.175574 -0.404215 -0.343907 -0.284027 factor(bilingual.status)3 ctbs.lang.nce -0.051558 -0.003763 Degrees of Freedom: 315 Total (i.e. Null); 304 Residual Null Deviance: 2410 Residual Deviance: 1909 AIC: 2795
Poisson Regression
• Lets look back at the Poisson log-linear model log
i
*
Math
_
Score
i
• Taking the sample mean and sample variance of the response for intervals of Math Scores
Math Score
0-20 20-40 40-60 60-80 80-100
Sample Mean
11.66666667
6.453333333
5.270072993
4.324675325
9.666666667
Sample Standard Deviation
10.64397095
6.595029523
7.382913152
5.434881392
14.50861813
Poisson Regression
• Overdispersion for Poisson Regression Models • •
For Y
i
~Poisson( λ
i
), E [Y
i
] = Var [Y
i
] = λ
i
The variance of the response is much larger than the mean.
•
Larger variance known as overdispersion
•
Consequences: Parameter estimates are still consistent
•
Standard errors are inconsistent Remedy: Negative Binomial model
Poisson Regression
• Summary • Introduction to the Poisson Regression Model • • Interpretation of β Variable Significance – Likelihood Ratio Test • Model Selection • • • Forward Backward Stepwise • Overdispersion
Poisson Regression
•