An Illustration of Rate Difference Estimation with SAS in

Download Report

Transcript An Illustration of Rate Difference Estimation with SAS in

An Illustration of Rate
Difference Estimation with SAS
in Logistic Regression
Yun Guo
Outline
•
•
•
•
Statistical methods for RD
RD with logistic regression
SAS realization
Conclusion
Statistical methods
• Wald/t approximation—no covariate adjust
• CMH (weighted estimate)-adjust strata
• Logistic Regression-adjust both
Statistical methods
• Logistic regression
(1)
,where, x is covariate, s is factor, t is treatment,
sub i is indexing subject
exp(   xi   si   ti )
pi 
1  exp(   xi   si   ti )
(2)
RD with logistic regression
• Odd Ratio through Logistic Regression
Simply subtract among different levels of
factor
Say t(t=1) vs. t(t=0),
p
p
log(OR)  log(
)t 1  log(
)t  0  exp( )
1 p
1 p
RD with logistic regression
• Rate through Logistic Regression (say
subject i=1)
exp(   xi   si   )
pt 1 
1  exp(   xi   si   )
exp(   xi   si )
pt  0 
1  exp(   xi   si )
RD  pt 1  pt  0
(3)
Intercept, covariate as well as factor, not gone!!!!
RD with logistic regression
• How about the variance of estimate?
pt 1
exp(   xi   si   )

1  exp(   xi   si   )
pt  0
exp(   xi   si )

1  exp(   xi   si )
var( RD)  var( pt 1  pt  0 )
Very messy!!!
(4)
RD with logistic regression
Hereafter subscript 1/0 of est p index treatment.
Statistical methods
• Marginal approach (sum over all samples)
Statistical methods
Statistical methods
• To conclude, we have following 4 steps
• 1, establish logistic regression
• 2, get estimated rates, covariance of estimates,
estimated RD
• 3, set up Delta method
• 4, get CI
SAS realization
EXAMPLE DATA
SAS realization
PREPARE DATA
SAS realization
STEP 1: Logistic regression
SAS realization
STEP 1: Logistic regression
SAS realization
STEP 1: Logistic regression
COV
EST
SAS realization
STEP 2: RD estimate
SAS realization
Step 3: Delta method
Step 4: CI
Conclusion
1, When covariates adjust needed, a choice is to use L
Regression
2, Other methods could be used too
Thank you!