Misconceptions about the FCR intervals in the paper "Reporting FDR analogous confidence intervals for the log fold change of differentially expressed genes".
Vered Madar, University of North Carolina at Chapel Hill, USA
4 September 2013
The paper "Reporting FDR analogous confidence intervals for the log fold change of differentially expressed genes" argues for the advantages in constructing confidence intervals for selected set of significant log fold changed genes. Bringing convincing arguments along with actual examples in favor of using the adjusted FCR confidence intervals (Benjamini and Yekutieli 2005).
Unfortunately two (and possibly undeliverable) major misconceptions concerning the FCR intervals appear in this paper. Although these inaccuracies do not interfere with the author's final conclusion and their very convincing results, I still wish to point them out because of their importance. From this point I will use the label "FCR intervals" to note the adjusted FCR confidence intervals (Benjamini and Yekutieli 2005).
The first misconception appears at the opening paragraph with the claim that FCR intervals are "based on the order of their corresponding p-values". This claim conflicts with the generality of the FCR intervals approach, ignoring completely that the FCR control, as presented in Benjamini and Yekutieli (2005), is not restricted to only one specific selection procedure (such as the Benjamini-Hochberg FDR controlling procedure,1995), being more general and very flexible procedure that easily fits to many other multiple selection procedure from Bonferroni to Storey. It is true that the adjusted FDR p-values or ordered p-values might play important role in selection of significant log fold changed genes in case one uses the Benjamini-Hochberg FDR controlling procedure. But once the selection is done, constructing the FCR intervals will provide set of confidence intervals of equal lengths which has nothing to do anymore with the order of p-values or their adjusted values.
The second misconception arises at the simulation study, and involves over-protecting a simulated positive correlated dependency structure by unnecessary division of the alpha level by the sum 1+1/2 + ...+1/m. Ignoring the fact that positive correlations are merely a simple case of PRDS dependency, and that for PRDS it is well proved that both Benjamini-Hochberg controlling FDR procedure (Yekutieli and Benjamini 2001) and the FCR intervals are conservative (FDR and FCR never reach above the desired level). Perhaps, a better simulation study would be one that contains actual negative correlations.
Misconceptions about the FCR intervals in the paper "Reporting FDR analogous confidence intervals for the log fold change of differentially expressed genes".
4 September 2013
The paper "Reporting FDR analogous confidence intervals for the log fold change of differentially expressed genes" argues for the advantages in constructing confidence intervals for selected set of significant log fold changed genes. Bringing convincing arguments along with actual examples in favor of using the adjusted FCR confidence intervals (Benjamini and Yekutieli 2005).
Unfortunately two (and possibly undeliverable) major misconceptions concerning the FCR intervals appear in this paper. Although these inaccuracies do not interfere with the author's final conclusion and their very convincing results, I still wish to point them out because of their importance. From this point I will use the label "FCR intervals" to note the adjusted FCR confidence intervals (Benjamini and Yekutieli 2005).
The first misconception appears at the opening paragraph with the claim that FCR intervals are "based on the order of their corresponding p-values". This claim conflicts with the generality of the FCR intervals approach, ignoring completely that the FCR control, as presented in Benjamini and Yekutieli (2005), is not restricted to only one specific selection procedure (such as the Benjamini-Hochberg FDR controlling procedure,1995), being more general and very flexible procedure that easily fits to many other multiple selection procedure from Bonferroni to Storey. It is true that the adjusted FDR p-values or ordered p-values might play important role in selection of significant log fold changed genes in case one uses the Benjamini-Hochberg FDR controlling procedure. But once the selection is done, constructing the FCR intervals will provide set of confidence intervals of equal lengths which has nothing to do anymore with the order of p-values or their adjusted values.
The second misconception arises at the simulation study, and involves over-protecting a simulated positive correlated dependency structure by unnecessary division of the alpha level by the sum 1+1/2 + ...+1/m. Ignoring the fact that positive correlations are merely a simple case of PRDS dependency, and that for PRDS it is well proved that both Benjamini-Hochberg controlling FDR procedure (Yekutieli and Benjamini 2001) and the FCR intervals are conservative (FDR and FCR never reach above the desired level). Perhaps, a better simulation study would be one that contains actual negative correlations.
Competing interests
None declared