Please wait while the transcript is being prepared...
FRANK DUDBRIDGE: "Winner's Curse,
Replication and Meta-Analysis."
Frank Dudbridge, London School
of Hygiene and Tropical Medicine.
This talk covers
three issues that often
arise when following up the results
of a Genome Wide Association Study.
The winner's curse is
the tendency to see
a larger effect than
you would usually see
when first discovering this effect.
how genetic associations
are confirmed in further studies.
And meta-analysis concerns combining
multiple Genome Wide Association
Studies into a single
So to understand
the winner's curse, we can
consider a situation of an auction.
So supposing that
an item is up for auction.
And in this auction,
each bidder will
submit a sealed bid for the item.
Now, suppose that
the true value of the item
could be defined as
the average of all of these bids.
Then because of the way an auction
works, the winner of the auction
must pay more than
the true value of the item.
Because the winner is the person
who's made the highest
bid for the item, then
the highest bid must be higher
than the average of the bids.
So the winner has paid more
than the true value of the item.
So this is what is known
as the winner's curse.
The winner is paying
more than the true value.
Now this effect actually
occurs in many settings.
In any situation in which
there are many items
and the observation
of each item depends
both on its true effect
plus a bit of noise,
you would see the winner's curse.
So in the examples
we're looking at here,
we have many SNPs in
a Genome Wide Association Scan.
Or the many items could be
many studies of a single SNP.
We've just talked about
many bidders in an auction.
And you can have examples
from the world of sports,
where you can have
many sports players
and we make an observation
on each player.
So if we're looking at many SNPs
in a Genome Wide Association Study,
then the observation could
the odds ratio for the disease.
of each of these many SNPs.
In an auction, there could
be many bids on an item.
And in studies of sports players,
we could be looking at many players
and, for example, measuring how
many points each player scored
over the course of the season.
So for each of these quantities,
we can imagine there's a true effect,
so a true odds ratio for a SNP.
But what we observe is the true
effect plus some noise added
onto it due to random sampling.