1 of 28

Generalized Linear Model

2 of 28

Simple Linear Regression

  •  

3 of 28

What if the assumption is violated?

  • Response variable has boundaries
    • proportion between 0 and 1
    • count
    • positive continuous

  • A linear assumption may not apply here because the fitted values from a linear model are unbounded and can fall outside the allowable range.

 

4 of 28

Binary Outcome

  •  

Patient

Smoking (1 case 0 control)

Lung Cancer (1 case 0 control)

1

1

1

2

0

1

3

1

0

5 of 28

Count Data RNA -seq

6 of 28

Count Data RNA -seq

7 of 28

Count Data

  •  

Sample 1 (Cancer)

Sample 2 (Cancer)

Sample 3 (Normal)

Sample 4 (Normal)

Gene A

100

150

30

40

8 of 28

What is common?

 

 

 

 

 

 

Linear regression

Logistic regression

Poisson regression

 

 

 

9 of 28

Three Components of Generalized Linear Model

  •  

 

10 of 28

Exponential Family Distribution (Canonical Form)

  •  

11 of 28

Estimation

  •  

12 of 28

Estimation

  •  

13 of 28

Score Function of Canonical Exponential Family�

  •  

14 of 28

Newton’s Method (single parameter)

  •  

 

 

 

 

 

 

 

 

15 of 28

Newton’s Method (Generalization to Vector)

  •  

16 of 28

Observed Information (single parameter)

  •  

 

 

 

17 of 28

Observed Information of Canonical Exponential Family

  •  

18 of 28

Fisher’s Information

  •  

19 of 28

Newton’s Method with Fisher scoring�(Iteratively Reweighted Least Square)

  •  

20 of 28

 

  • Likelihood ratio test

  • Wald test

  • Score test

21 of 28

Likelihood ratio test�

  •  

 

 

 

 

22 of 28

Logistic Regression

  •  

23 of 28

Wald test

  •  

 

 

24 of 28

Wald Test for R’s glm

25 of 28

Rao’s Score test

  •  

 

 

26 of 28

Rao’s Score test

  •  

27 of 28

28 of 28

Comparison

Likelihood Ratio Test

Wald

Score

Basic idea

Compare maximized likelihood of reduced and full model

Check how far the MLE is from the null

Check slope of log-likelihood at null

Computation

Both reduced and full model

Full model only

Reduced model only

Large sample

Small sample

Your homework