This preview has intentionally blurred sections. Sign up to view the full version.
View Full DocumentThis preview has intentionally blurred sections. Sign up to view the full version.
View Full DocumentThis preview has intentionally blurred sections. Sign up to view the full version.
View Full Document
Unformatted text preview: Logistic Regr. Model for Binomial Count Data I Bernoulli model appropriate for 0/1 response on an individual I What if data are # events out of # trials per subject? I Example: Toxicology study of the carcenogenicity of aflatoxicol. I (from Ramsey and Schaefer, The Statistical Sleuth , p 641) I Tank of trout randomly assigned to dose of aflatoxicol I 5 doses. CRD. 4 replicate tanks per dose I 8690 trout per tank I Response is # trout with liver tumor I Could use Bernoulli model for each individual fish I But, all fish in a tank have the same covariate values (dose) I easier to analyze data in summarized form (# with tumor, # in tank) c 2011 Dept. Statistics (Iowa State University) Stat 511 section 27 1 / 23 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 0.05 0.10 0.15 0.20 0.25 0.0 0.2 0.4 0.6 0.8 Dose (ppm) # tumor/# in tank c 2011 Dept. Statistics (Iowa State University) Stat 511 section 27 2 / 23 I Each response is # “events” out of # trials. I y i ∼ Binomial ( m i , π i ) , i = 1 , ..., n , where m i is a known number of trials for observation i. I π i = exp ( x i β ) 1 + exp ( x i β ) I y 1 , ..., y n are independent. I Note: two levels of independence assumed in this model I each response, y i , is independent I trials within each each response are independent  π i I y i ∼ Binomial ( m i , π i ) when I y i = ∑ m i j = 1 Z ij , where Z ij ∼ Bernoulli ( π i ) I And Z ij independent I May be an issue for both the Challenger and trout data sets I Binomial model assumes no flight (tank) effects I Data are same as ≈ 360 fish raised individually, or the same as one tank with ≈ 360 fish c 2011 Dept. Statistics (Iowa State University) Stat 511 section 27 3 / 23 I For now, assume Binomial model reasonable I Facts about Binomial distributions: If y i ∼ Binomial ( m i , π i ) I E ( y i ) = m i π i I Var ( y i ) = m i π i ( 1 π i ) I f ( y i ) = m i y i π y i i ( 1 π i ) m i y i for y i ∈ { , ..., m i } I l ( β  y ) = ∑ n i = 1 [ y i log ( π i 1 π i ) + m i log ( 1 π i )] + const = ∑ n i = 1 [ y i x i β m i log ( 1 + exp { x i β } )] + const . I l ( β  y , m , x ) for ( y 1 , m 1 , x 1 ) , ( y 2 , m 2 , x 2 ) , . .. ( y n , m n , x n ) same (apart from constant) as Bernoulli lnL: l ( β  z ) for ( z ij , x ij ) = ( , x 1 ) , ( , x 1 ) , .. . ( 1 , x 1 ) , .. . , ( , x 2 ) , .. . ( 1 , x 2 ) , .. . , ··· , ( , x n ) , .. . , ( 1 , x n ) , .. . I MLE’s ˆ β obtained by numerically maximizing l ( β  y ) over β ∈ IR p c 2011 Dept. Statistics (Iowa State University) Stat 511 section 27 4 / 23 I Results for trout data: Coefficient Estimate se z p Intercept0.867 0.07611.30 < 0.0001 dose 14.33 0.937 15.30 < 0.0001 I Looks impressive, but ......
View
Full
Document
This note was uploaded on 02/11/2012 for the course STAT 511 taught by Professor Staff during the Spring '08 term at Iowa State.
 Spring '08
 Staff
 Bernoulli, Binomial

Click to edit the document details