Biology Reference
In-Depth Information
One approach which can start at this level is hierarchical modeling.
Using this approach successfully requires a good understanding of how
the different data types may be jointly modeled. In many applications,
this will be difficult to implement.
5.2. Pooling Adjusted Data
Correction before pooling (see, for instance, Benito
et al
.
9
) consists of
applying an adjustment to the data separately for each study, and then
combining adjusted values to analyze together as a single set. For example,
rather than including a study-specific effect in addition to covariate
adjustment in an overall model for pooled raw data, separate adjustments
for each study could be made.
This method might not work so well for very heterogeneous microar-
ray studies, where alternative technologies provide intrinsically different
measures, e.g. single- versus dual-channel array types. There is as yet no
accepted way to transform such diverse values to a single, common scale
of measurement.
In addition, separate adjustment will not guarantee that the resulting
adjusted observations (typically residuals) will be similarly distributed
across studies. Without some further modification (e.g. scaling or other
transformation), even these adjusted observations may well remain too
heterogeneous for pooling.
5.3. Combining Parameter Estimates
Meta-analysis consists of statistical methods for combining results of
independent studies addressing related questions. One aim of combin-
ing results is to obtain increased power — studies with small sample
sizes are less likely to find effects even when they exist. Putting results
together increases the effective sample size, thereby allowing more pre-
cise effect estimation and increasing power. The uncovering of a signif-
icant effect from a combined analysis, where individual studies do
not make positive findings at the same significance level, has been
referred to in the microarray meta-analysis literature as “integration-
driven discovery” (IDD).
11
