Biology Reference
In-Depth Information
design, patient inclusion/exclusion criteria, and removal of duplicated
individuals. We resolve most quality issues during SwissBrod curation.
Once the within-data set preparations are complete, genes must be
matched across datasets as a prelude to combination.
6.3. Gene Matching
Since the interpretation of study results usually makes statements about
genes, matching probe/probe set expression values at the level of genes
is desirable. UniGene identifiers or gene symbols can be used. Where
there are multiple probes (or probe sets) of the same gene, we choose a
unique one to represent the corresponding gene. We generally choose
the most variable probe across samples; other common selection rules are
to choose the probe with the highest or second-highest average signal
across samples.
Here, we note that it is not necessary to consider only the common
probes or genes common across all datasets. This unnecessarily restricts
the set of genes about which inference can be made. Missing data, such
as genes which are not measured in all studies, are readily accommodated
in this framework for data combination.
6.4. Outcome Modeling
We consider within each study a generalized linear model (GLM)
28
for
single genes or sets of genes, possibly including study covariates,
describing the association between the measured quantity (e.g. gene
expression) and an outcome of interest. The association is supposed to
be captured in one (or possibly more) coefficient of the model. This
class of models is sufficiently broad and flexible to handle most com-
monly investigated situations, including quantitative, categorical, and
survival outcomes.
The models across studies are not required to contain the same or
similar parameters, as we will not be combining parameter estimates.
Rather, there is assumed a set of (at least conceptually) related hypothe-
ses that can be connected through test statistics based on model coeffi-
cients. To get around heterogeneity in the underlying measurements, it
is important to use relative measures within a study.
