In calculating the Bayesian estimate of the standard deviation there are two different parameters that the user must set. These 2 parameters relate to setting the Bayesian estimate of variance derived from the observed population. The t-test is used to detect significant differences between the means of two groups relative to the observed variance within groups. In a perfect world, all micro-array experiments would be highly replicated within each experimental treatment. Such replication would allow accurate estimates of the variance within experimental treatments to be obtained and the t-test would then perform well. Microarray experiments are expensive and time consuming to carry out, and there is the possibility that both control and experimental tissues will be limiting. As a result, the level of replication within experimental treatments is often low. This results in poor estimates of within-treatment variance and a corresponding poor performance of the t-test itself. This problem can be solved addressed using a Bayesian statistical approach that incorporates prior information in the estimation of within gene within treatment variances.

Although in terms of strong inference there is really no substitute for proper experimental replication, in the case of microarrays many parallel pseudo-replicated experiments are carried out on any given microarray (i.e., one experiment for each gene). In particular, it is possible to use estimates of within-treatment variance from a number of genes of similar expression level to stabilize estimates of variance for any given gene. More precisely, the variance within any given treatment is estimated by the weighted average of a prior estimate of the variance for that gene (obtained from of local weighted average of the variance of other genes) and the experimental estimate of the variance for that gene. This weighting factor is controlled by the experimenter and will depend on how confident the experimenter is that the background variance of a closely related set of genes approximates the variance of the gene under consideration.

An important property of this Bayesian approach to consider is that in the two limiting cases of complete confidence in the prior and zero confidence in the prior, the Bayesian approach is equivalent to simply looking at fold change and testing differences between treatments using the simple t-test respectively. In the Bayesian approach the weight given to the within gene variance estimate is a function of the number of observations contributing to that value. This leads to the desirable property of the Bayesian approach converging to the t-test as the experimenter carries out additional replications and thus becomes more confident of the observed estimate of within treatment variance for any given gene.

Indicates how wide you want the window surrrounding the point under consideration to be. This sample of the data provides an estimate of the average variability of gene expression for those genes that show a similar expression level. It is important to estimate this average from a wide enough level that it is accurate, but not so wide of a window so as to average in genes with too different of average expression level. A sliding window of 101 genes has been shown to be quite accurate when analyzing 2000 or more genes, with only 1000 genes a window of 51 genes may work better.

This is a number from 0 to infinity that indicates the weight give to the Bayesian prior estimate of within-treatment variance. Larger weights indicate greater confidence in the Bayesian prior; smaller weights indicate more confidence in the experimentally observed variance. We have observed reasonable performance with the following rule of thumb: set the confidence such that the number of experimental observations plus the confidence is greater than 8.

If the confidence is left blank or zero, then a simple classical t-test is performed. If there is only a single replicate in a given condition, then the standard deviation estimates are completely determined by the prior, and a warning is issued to the user. If the confidence is set to zero and there is only a single replicate, we set a default confidence of 5 and issue a warning to the user.