You are on page 1of 1

Method of Testing for

Abstract Methods:
Confounding, Effect Modification, Confounding and
Introduction
The Cochran Mantel Haenzel test has problems
of inconsistency with output if data pertaining to y=0 is input
before y=1 in the ‘datalines’ of the code. The output gives
and Calculating Point Estimates Effect
reciprocal answers with respect to y=1 and y=0. The CMH test does
give output when adjusting for a Nx2x2 data set but not 2xNx2
when N>2. One can use logistic regression with SAS utilizing proc Modification
logistic, but this only yields an odds ratio, no relative risks. My
technique utilizes the slope of vectors to calculate parameter
estimates which can be used to give OR’s and RR’s. Also, I
devised a new way to calculate p values for testing effect
Manoj Agravat MPH Using software ‘R’:
survreg(formula = Surv(count, smokers) ~ country, data = country1,
"rayleigh")
dist =
modification. I created a type of data input where a pattern and new
value estimates based on my new method yields a p value that can
be used to test for effect modification. USF College of Public Health Epidemiology/Biostatistics 2. survreg(formula = Surv(count, smokers) ~ country + smokers, data = country1,
dist = "rayleigh")
3. survreg(formula = Surv(count, smokers) ~ cases, data = country1, dist
= "rayleigh")

Method Results
I use ‘R’ and the
survreg(Surv(count,explatatory_variable)~outcome +confounder
Using SAS
Data passlungexp5;
pattern utilizing different distributions such as lognormal,
Rayleigh… This generates Beta estimates which I use in formula’s
I developed to give OR’s and RR’s. I also devised a new way to
Point OR1 RR1 RR2 OR2 RR1 RR2 input cases fit zxy xzy n;
datalines;
calculate p values for effect modification. I adjusted the ‘datalines’
input form in a new way: estimat (z) (z) (z) (x) (x) (x) 1 1 1 1 73
0 1 1 1 188
datalines; 1 0 1332 101 27
1 1 1 1 73
0 1 1 1 188
e 0 0 5201 307 82
1 1 1 1 19

Value 1.385 1.2 .959 1.005 2.35 .419


1 0 1332 101 27 0 1 1 1 38
0 0 5201 307 82 1 0 317 19 5
73 and 188 are raw data counts. Ist 1 is outcome status yes; 2nd 1 is 0 0 1015 60 16
for raw versus adjusted count (raw); I enter 1 twice for each
variable( counfounder & explanatory); When I enter adjusted
value for confounder and explanatory variable code the 2nd data
95 %CI (1.0003 (1.14,1. (1.248, (1.321, (.2,1.90 (.456,2. 1 1 1 1 137
0 1 1 1 363
1 0 4503 266 71
with 0 and enter new value. The fitted value for the confounder is
from raw count/Betaz*Betazx.; where as for explanatory variable,
fit value is from raw count/betayz where Betayz is from the
,1.652) 508)* 1.368)* 1.33)* 8) 195) 0 0 15793 933 249
;
interaction term from codes similar to the one above using survreg.
Next, use the proc logistic command in SAS:
proc logistic data=passlungexp5 descending;
* run;
proc logistic data=passlungexp2new descending;
class fit (ref='1');
class exp (ref='1');
weight n; CMH 1.385
1.27 .923 none none none model cases= fit zxy xzy / rsq lackfit;
run;
model cases= fit zxy xzy /rsq lackfit;
run; value
95 %CI (1.05,1. (1.036, (.864,.9 none none none Conclusion
Methods (Point estimates) (CMH) 82) 1.575) 84)
country1<-data.frame(country,smokers,cases,count) There is no confounding by counrty.
country1 There is no effect modification.
countrya<-survreg(Surv(count,smokers)~country, data=country1, dist="rayleigh") Adjusting for ‘z’:
summary(countrya) OR (z) is significant; ratio of odds of cancer
Controlling for ‘z’ with smoker as a factor is 1.385x odds
OR1=|(exp^(bz)/(exp-^(bz+bxz))))|*srqrt(sez)=1.385
RR1=|(exp^(-(bz+bxz))/(1-(exp^(bz+bxz))))|*sqrt(sez)=1.20 Parameter Country Country + Smoker of cancer without smokers adjusting for
country.
RR2=|(exp^(by+bxz)?(exp^(by)))|=.959 RR1(z) is significant showing risk of cancer
Controlling for ‘x’
OR2=|(exp^(-(bz))/(1-exp^(bz)))|*sqrt(sez)=1.0 Smoker is 20 % higher with smokers than
without smokers.
RR1=|(exp^((-by))/(1-exp^(-by)))|*sqrt(sey)=2.35 RR2(z) is significant. Adjusting for
RR2=|(exp^(by+byz)?(1-exp^(by)))|=.419
P value .077 .043** .019** country, proportion of cancer for
smokers is 1.248 to 1.368 x proportion
of cancer if the person was a non
smoker, or risk is 4.1 % less for
nonsmoker.

Effect Modification P value= .611


Adjusting for ‘x’:
OR(x) is marginally significant; ratio of
odds of cancer with country as a factor
is .005x odds of cancer without country
Note: * or ** indicates statistical significance . P value of .611 means no effect modification. as a variable adjusting for smokers.
RR1(x)= 2.35 is not significant.
RR2(x) = .41 is not significant.

You might also like