Title:
Assessment and application of a nonparametric confidence set approach
based on the mean test
Abstract:
In genomewide linkage analyses one needs to adjust for the number of
simultaneous hypotheses tested. Such an adjustment is very difficult due
to the complex nature of dependencies amongst data arising from linkage
studies. It is possible, though, to avoid dealing with multiple testing
issues, by constructing confidence sets of markers tightly linked to a
genetic trait through testing hypotheses that are reversal of the usual
setup.
Here, we examine the effect of several factors on the confidence set
derived using the nonparametric mean statistic based on IBD sharing of
affected sib pairs. Simulations are performed to assess the performance of
the approach in terms of its true and false discovery rates, under varying
conditions of the underlying single locus disease model, the amount of
data, the heterozygosity of the marker, and the number of family members
genotyped at the marker locus. We also provide rough guidelines for the
choice of coverage probability of the confidence set that would lead to a
false discovery rate of a desirable level. Our methods are then applied to
the simulated data from the Genetic Analysis Workshop (GAW) 13, focusing
on the high blood pressure trait. Our simulation results show that when
the assumption of a single locus trait with accurate risk estimates is
met, the confidence set is able to sufficiently localize the disease
causing gene. However, a moderate increase of about 30% in the amount of
data is necessary for the method to achieve the same power as when the
marker is 100% polymorphic. Also, inaccurate estimation of the risks may
compromise the power or inflate the false discovery rate of the method,
particularly when genetics only plays a limited role in the disease
causing mechanism. Incorrect specification of the number of disease
causing genes also reduces the power of the method. Nevertheless, the
results for the GAW data suggest that, despite potential reduction in
power due to deviation from the ideal situation, our method can still
achieve significantly higher power than standard nonparametric methods
while maintaining comparable false positive rates, especially when the
sample size is reasonably large.