Population Stratification Analysis in Genome-Wide Association Studies - Mathematical Approaches to Polymer Sequence Analysis and Related Problems - page 190

Tabl e 9. 5 Top significantly associated (p<10 4 ) SNPs with schizophrenia applying genomic
control (GC) on the basic allelic chi-squared statistics (PLINK) and on the Armitage chi-squared
statistics (EIGENSTRAT), in UCI and CATIE-NIMH samples

#TOTtop
significantly

A NS TP A S FP associated SNPs

Basic allelic 2 statistics (PLINK) UCI 16 (29.57) 40 (71.43) 56

CATIE 161 (8.97) 1,633 (91.03) 1; 794

Armitage 2 statistics (EIGENSTRAT) a UCI 13 (46.43) 15 (53.57) 28

CATIE 6 (2.18) 269 (97.82) 275

COMMON UCI 9 (10.71) 7 (8.33) 84

CATIE 5 (0.25) 138 (6.67) 2; 069

a The values refer to the results after outliers removal

Tabl e 9. 6 Top significantly associated (p<10 4 ) SNPs with schizophrenia applying genomic
control (GC) on the basic allelic chi-squared statistics (PLINK) and on the Armitage chi-squared
statistics (EIGENSTRAT), in UCI and CATIE-NIMH samples
# top significant SNPs
after GC correction on

# COMMON

allelic chi-squared # top significant SNPs significantly associated

statistic after CMH correction SNPs

UCI 16 24 10

CATIE 161 5; 926 161

The values refer to the results after outliers removal

We observe that the GC applied on chi-squared statistics of EIGENSTRAT
(Armitage chisquare statistics) and PLINK (basic allelic test chi-square) produces
datasets composed of different top SNPs: there are only nine true positive and seven
false positive across 84 significantly associated SNPs that are common between the
two statistics from both methods. This may depend also from the specific association
statistics on which GC is applied, because the statistics of Armitage test (EIGEN-
STRAT) and basic chi-square (PLINK) are obtained by different algorithms.
To evaluate whether the GC correction shows similar results with CMH test
(Table 9.6 ) and EIGENSTRAT (Table 9.7 ), we compared the results: we observe
that GC applied on the allelic chi-square shares with CMH test 10 SNPs (33.3%) in
UCI sample (Table 9.6 ), while in the CATIE sample, all the 161 top SNPs (2.7%)
obtained with the GC correction are top SNPs also using the CMH correction. On
the other hand, GC and EIGENSTRAT correction on Armitage chi-square statis-
tics present eight common SNPs (29.6%) in the UCI sample and four (6.8%) in the
CATIE sample that are common between the two methods.
To assess the power of EIGENSTRAT and PLINK to correct for substructure and
to verify the results, we calculated the genomic inflation factor on the association
results for both samples (Table 9.8 ).
The value above 1 indicates inflation in chi-squared statistics and, by definition,
is not allowed to be more than 1 in a homogeneous sample. In the UCI sample,