Replication is an important aspect of all statistical experimental design. As described in Yang and Speed (16), there can be three types of microarray replicates: biological replicates where mRNA samples are taken from different units; technical replicates, where mRNA samples are taken from the same unit and are split and hybridized onto different arrays; and within-array replicates, where probes are spotted in replicate on the same array. These types apply equally to longitudinal and cross-sectional time course experiments. The variation between gene expression measurements taken on these three types of replicates will be different, and in general is gene-specific. Replication is a good thing, as it provides estimates of variability relative to which temporal changes and/or condition differences can be assessed, making analysis much more straightforward. Biological replicates are generally best, as they permit the conclusions from the experiment to be extrapolated to the wider population of units from which the experimental units were obtained, something which is not possible with only technical replicates. With unreplicated experiments, the inference to a wider population is not possible, and the analysis is less straightforward, being more dependent on unverifiable assumptions, as there is no estimate of pure error which can be used. We suggest at least three replicates at every time point. When replicates are available, it is better to use the variation between them in any analysis, rather that just average across the replicates and proceed as with a single time course experiment. Many of the methods we describe below require replicates, and are designed to be effective with the small numbers of replicates common in this context.

0 0

Post a comment