DESeq2 fold change does not match normalized counts data

[mrandel]

I am trying to understand how DESeq2 calculates Log2foldchange, because it does not match my calculations of fold change from my normalized counts.

My script:
CountTable_POS = read.table("DIF_pos_counts", header=TRUE, sep="")
Design = data.frame( row.names = colnames( CountTable_POS ),
condition = c( "untreated", "untreated", "untreated",
"treated", "treated", "treated" ),
type = c( "single-end", "single-end", "single-end",
"single-end", "single-end","single-end" ) )
dds_pos<-DESeqDataSetFromMatrix(countData = CountTable_POS,
colData = Design,
design = ~ condition)
dds_pos<-DESeq(dds_pos, fitType="local")
norm<-counts(dds_pos, normalized=T)
write.csv(as.data.frame(norm), file="norm_counts.csv")
res_pos<-results(dds_pos, contrast=c("condition", "treated", "untreated"))
write.csv(as.data.frame(res_pos), file="DESeq2_out.csv")

RESULTS
from norm_counts.csv:
Locus Control1 Control2 Control3 Treated1 Treated2 Treated3 Mean_Control Mean_Treated Basemean Fold change
2R_188654 3.8127 5.7220 5.7191 1.9351 4.9898 2.9518 5.0846 3.2922 4.1884 0.6475
2R_188736 3.8127 3.8147 6.6723 1.9351 4.9898 5.9036 4.7666 4.2762 4.5214 0.8971

From DESeq2_out:
Locus Basemean Log2foldchange P value Fold change
2R_188654 4.1884 -0.3108 0.4093 0.8062
2R_188736 4.5214 -0.0819 0.8279 0.9448

As you can see, the basemeans (mean of control and treated counts) match- which tells me there is no internal error- but the fold changes do not. Why is this???

[dpryan]

Fold-changes are shrunken, so they won't match what you calculate by hand (the shrinkage produces more reliable results). If you want the number that will more or less match what you're calculating them have it output the MLE results (use "addMLE=T" in results()).