Biomedical Engineering Reference
In-Depth Information
2 Run the following SAS program to compute the likelihood-ratio chi-square statistic
comparing the full logistic regression model to the reduced model in a one
degree-of-freedom test for the effect of genotype:
n
data subset;
set fulldata;
if loc1
=
. then delete;
***analyze only the subset for which loc1 is non-missing;
run;
*** 1st proc logistic: base model, 1 covariate included;
proc logistic data
=
subset outest
=
base_loc1;
model ctrl
=
cov1;
title "1st proc logistic: base model";
run;
*** 2nd proc logistic: full model, 1 covariate
+
genotype;
*** genotype variable, loc1, is coded as number of copies of allele 1;
proc logistic data
=
subset outest
=
full_loc1;
model ctrl
=
cov1 loc1;
title "2nd proc logistic: full model";
run;
*** get likelihoods for each model;
data like_base;
set base_loc1
(rename
=
(_LNLIKE_
=
lnlike_base));
keep lnlike_base;
data like_full;
set full_loc1
(rename (_LNLIKE_
=
lnlike_full));
keep lnlike_full;
run;
data mrg;
merge like_base like_full;
chisq_1df
=−
2 * (lnlike_base - lnlike_full);
=
pvalue
1 - probchi(chisq_1df,1);
run;
3 Repeat the analysis for each SNP to be analyzed.
o
4 Logistic-regression-based association analysis can also be carried out in PLINK using the
'--logistic' option.
Notes
m
The protocol described here is based on one used for the analysis of a large-scale GWAS
and candidate gene study of nicotine dependence [6, 7]. A single 'proc logistic' step would
suffice to obtain the p-value for the one degree of freedom test of the single genotype term;
however the code shown here, with two 'proc logistic' steps, readily generalizes to carry out
tests with more degrees of freedom (e.g. if the genotype is coded with two degrees of freedom,
or if interactions between covariate and genotype are to be jointly tested with the genotype
term).
