Monday, May 14, 2012

Example 9.31: Exploring multiple testing procedures

In example 9.30 we explored the effects of adjusting for multiple testing using the Bonferroni and Benjamini-Hochberg (or false discovery rate, FDR) procedures. At the time we claimed that it would probably be inappropriate to extract the adjusted p-values from the FDR method from their context. In this entry we attempt to explain our misgivings about this practice.

The FDR procedure is described in Benjamini and Hochberg (JRSSB, 1995) as a "step-down" procedure. Put simply, the procedure has the following steps:

0. Choose the familywise alpha
1. Rank order the unadjusted p-values
2. Beginning with the Mth of the ordered p-values p(m),
2a. if p(m) < alpha*(m/M), then reject all tests 1 ... m,
2b. if not, m = m-1
3. Repeat steps 2a and 2b until the condition is met
or p(1) > alpha/M

where M is the number of tests. The "adjusted p-value" based on this procedure is the smallest familywise alpha under which the current test would have been rejected. To calculate this, we can modify the routine above:

1. Rank order the unadjusted p-values
2. For ordered p-values p(m) M to 1,
2a. candidate ap(m) = p(m) *(M/m)
2b. if candidate ap(m) > ap(m+1) then ap(m) = ap(m+1)
2c. else ap(m) = candidate ap(m)

where ap(m) refers to the adjusted p-value corresponding to the mth ordered unadjusted p-value. It's interesting to note that the adjusted p-value for the Mth ordered test is the same as the unadjusted p-value, while the candidate adjusted p-value for the smallest test is the Bonferroni adjusted p-value. The primary difficulty with taking these p-values (as opposed to the test results) out of context is captured in steps 2b and 2c. They imply that the p-value for a given test may be lowered by other observed p-values in the family of tests. It's also true that the adjusted p-value depends on the number of tests included in the family, but this seems somewhat less troubling.

To examine the impact of the procedure on the adjusted p-values for the individual tests, we'll compare the candidate ap(m) from step 2a against the actual ap(m). Our sense is that to the degree these are different, the adjusted p-value should not be extracted from the context of the observed family of tests.

Our SAS code relies heavily on the array statement (section 1.11.5). We loop through the p-values from largest to smallest, calculating the candidate fdr p-value as above, before arriving at the final adjusted p-value. To compare the values conveniently, we make a new data set with two copies of the original data set, renaming first the candidate and then the adjusted p-values to have the same names. The in = data set option creates a temporary variable which identifies which data set an observation was read from; here it denotes which version of the same data set (and which set of p-values) was used.

data fdr;
array pvals [10] pval1 - pval10
(.001 .001 .001 .001 .001 .03 .035 .04 .05 .05);
array cfdrpvals [10] cfdr1 - cfdr10;
array fdrpvals [10] fdr1 - fdr10;
fdrpvals[10] = pvals[10];
do i = 9 to 1 by -1;
cfdrpvals[i] = pvals[i] * 10/i;
if cfdrpvals[i] > fdrpvals[i+1] then fdrpvals[i] = fdrpvals[i+1];
else fdrpvals[i] = cfdrpvals[i];

data compare;
set fdr (in = cfdr rename=(cfdr1=c1 cfdr2=c2 cfdr3=c3 cfdr4=c4
cfdr5=c5 cfdr6=c6 cfdr7=c7 cfdr8=c8 cfdr9=c9))
fdr (in = fdr rename=(fdr1=c1 fdr2=c2 fdr3=c3 fdr4=c4 fdr5=c5
fdr6=c6 fdr7=c7 fdr8=c8 fdr9=c9));
if cfdr then adjustment = "Candidate fdr";
if fdr then adjustment = "Final fdr";

proc print data = compare; var adjustment c1-c9; run;

adjustment c1 c2 c3 c4 c5 c6 c7 c8 c9

Candidate fdr 0.010 .005 .0033 .0025 .002 .05 .05 .05 .055
Final fdr 0.002 .002 .0020 .0020 .002 .05 .05 .05 .050

(We omit the last p-value because the adjustment does not affect it.) The result shows that for many of the tests in this family, a substantially smaller p-value is obtained with the final FDR p-value than the candidate. To this degree, the FDR p-value is dependent on the observed values of the p-values in the tests in the family, and ought not to be removed from the context of these other tests. We would recommend caution in displaying the FDR p-values in such settings, given readers' propensity to use them as if they were ordinary p-values, safely adjusted for multiple testing.

Comparison of the R and SAS code may make SAS programmers weep. The candidate values are easily calculated, and can be presented with the final p-values in one step using the p.adjust() function. Three lines of code, albeit incorporating multiple functions in each line. (And it could sensibly be done in two, calculating the candidate p-values within the rbind() function call.) Note especially the line calculating the candidate p-values, in which vectorization allows a for loop to be avoided in a very natural fashion.

fakeps = c(rep(.2, 5), 6, 7, 8, 10, 10)/200
cfdr = fakeps * 10/(1:10)
rbind(cfdr, fdr=p.adjust(fakeps, "fdr"))[,1:9]

[,1] [,2] [,3] [,4] [,5] [,6] [,7] [,8] [,9] [,10]
cfdr 0.010 0.005 0.0033 0.0025 0.002 0.05 0.05 0.05 0.0556 0.05
fdr 0.002 0.002 0.0020 0.0020 0.002 0.05 0.05 0.05 0.0500 0.05

An unrelated note about aggregatorsWe love aggregators! Aggregators collect blogs that have similar coverage for the convenience of readers, and for blog authors they offer a way to reach new audiences. SAS and R is aggregated by R-bloggers and PROC-X with our permission, and by at least 2 other aggregating services which have never contacted us. If you read this on an aggregator that does not credit the blogs it incorporates, please come visit us at SAS and R. We answer comments there and offer direct subscriptions if you like our content. In addition, no one is allowed to profit by this work under our license; if you see advertisements on this page, the aggregator is violating the terms by which we publish our work.


Joset A. Etzel said...

I use FDR differently: to obtain a significance threshold for a particular dataset; e.g. what is the FDR-corrected threshold corresponding to p=0.05?

This R code is not optimized but closely follows the pseudocode description.

ps <- sort(cfdr);
V <- length(ps);
q <- 0.05; # desired corrected p-value
cV <- 0;
for (i in 1:V) { cV <- cV + 1/i; }

desc <- rep("tooBig", V);
#for (i in 1:V) { if (ps[i] <= (i/V) * (q/cV)) { desc[i] <- "ok"; } else { break; } } # conservative method
for (i in 1:V) { if (ps[i] <= (i/V) * (q/1)) { desc[i] <- "ok"; } else { break; } } # more lenient method
if (length(which(desc == "ok")) > 0) { cutoff <- ps[max(which(desc == "ok"))]; } else { cutoff <- "none significant"; }

Ken Kleinman said...

Hi, Joset--

I'm having lot of trouble following your logic. Can you explain what you're trying to do?

Joset A. Etzel said...

I wasn't asking a question; I was just surprised to see FDR used to obtain adjusted p-values like you did in this example. I'm more familiar with having the outcome of a FDR calculation being a single value (used as a statistical threshold).

This paper describes the application (in neuroimaging):
Genovese, C.R., et al. (2002) Thresholding of Statistical Maps in Functional Neuroimaging Using the False Discovery Rate. Neuroimage 15, 870-878

Nick Horton said...

Thanks for the helpful reference. I think that we're saying the same thing however: the FDR values shouldn't be used as if they were adjusted p-values.

Joset A. Etzel said...

I definitely agree - FDR is different. I put together that bit of code to compare the output of my routine and p.adjust (they agree) and thought it might be of interest to others.