FDR calculation about sccomp HOT 3 CLOSED

I will give reference soon.

from sccomp.

I also want to ask whether it is normal to see the FDR is smaller than the pH0.

from sccomp.

Hello @chiwwong ,

the FDR is calculated according to Bayes statistics. From the paper you read

"
The probability of the null hypothesis (i.e., no effect across conditions) for each group is obtained by estimating the posterior probability of γg,c (or any combination thereof if contrasts are specified) being larger or smaller than a fold-change threshold (0.2 by default). The false-discovery rate is obtained by sorting in ascending order the probability of the null hypothesis (for any coefficient) and calculating the cumulative average as described by Stephens (24). The existence of an association between cell-group proportions and a factor including three or more categories (analogously to one-way ANOVA) is estimated by comparing the predictive errors between the model with a three-category design and a model with a one-mean design (intercept only). This comparison is achieved through leave-one-out cross-validation (R package loo (42)) and calculating approximate SEs for estimated predictive errors.
"

I also want to ask whether it is normal to see the FDR is smaller than the pH0.

Yes it is normal, because FDR in Bayesian statistics is providing an "average" probability for a group of observations, the probability pf the effect being bigger than a minimal effect, which can be set in the sccomp function.

from sccomp.