G***G 发帖数: 16778 | 1 qvalue太高了。怎么能降低qvalue呢?
大家用什么软件的qvalue? |
|
G***G 发帖数: 16778 | 2 我也是。
我的pvalue都是e-5级别的,到了qvalue,都是0.05以上了。
有什么方法能调整qvalue的计算参数来降低qvalue?
谢谢! |
|
m****c 发帖数: 244 | 3 你贴个p-value的histogram我来给你看看。
: 我也是。
: 我的pvalue都是e-5级别的,到了qvalue,都是0.05以上了。
: 有什么方法能调整qvalue的计算参数来降低qvalue?
: 谢谢!
|
|
h******1 发帖数: 94 | 4 If I run
> qobj = qvalue(p, lambda=seq(0, 0.90, 0.30), pi0.method="bootstrap")
[1] "ERROR: The estimated pi0 <= 0. Check that you have valid p-values or
use another lambda method."
> qobj = qvalue(p, lambda=seq(0, 0.90, 0.05), pi0.method="bootstrap")
[1] "ERROR: The estimated pi0 <= 0. Check that you have valid p-values or
use another lambda method."
> qobj = qvalue(p, lambda=seq(0, 0.90, 0.05), pi0.method="smoother")
[1] "ERROR: The estimated pi0 <= 0. Check that you have valid p-values or
use |
|
p********a 发帖数: 5352 | 5 ☆─────────────────────────────────────☆
Heidi911 (Heidi911) 于 (Tue Mar 30 11:18:25 2010, 美东) 提到:
If I run
> qobj = qvalue(p, lambda=seq(0, 0.90, 0.30), pi0.method="bootstrap")
[1] "ERROR: The estimated pi0 <= 0. Check that you have valid p-values or
use another lambda method."
> qobj = qvalue(p, lambda=seq(0, 0.90, 0.05), pi0.method="bootstrap")
[1] "ERROR: The estimated pi0 <= 0. Check that you have valid p-values or
use another lambda method."
> qobj = qvalue(p, lambda=seq(0, 0.90, 0.05), |
|
g********r 发帖数: 8017 | 6 If you look at the qvalue paper by Storey, it is clear in what case
you can get small qvalue - when the histogram of p-value has
a big peak near zero. |
|
G***G 发帖数: 16778 | 7 用什么方法计算qvalue,能得到比较小的qvalue? |
|
Z******5 发帖数: 435 | 8 我一直没搞明白qvalue有没有cut标准?没有的话,什么算高,什么算低? |
|
g*********r 发帖数: 281 | 9 qvalue 是 pvalue 经过 multiple test adjustment以后得到的,它和你进行的test总
数目有关。如果你确定你的test数目是这么多,你不应该人为降低这个值。它的cutoff
和pvalue一样。你把它当作pvalue好了。 |
|
g********r 发帖数: 8017 | 10 这些p-value都非常小啊.qvalue是假定你有一个mixture.
一部分pvalue应当是uniform(0,1) |
|
h******1 发帖数: 94 | 11 Yes, p.adjust is what I used before. But my supervisor asked me to use
pvalue.
I did preselection by fold change and this is why the p-values are so
small.
So I have to either qvalue() before filtering by fold change or p.adjust()
after filtering.
Thank you so much.
method= |
|
g********r 发帖数: 8017 | 12 Forget pre-selection and use qvalue on all the p-values.
If you did pre-selection, then all FDR generated from p.adjust are wrong.
() |
|
Y******Y 发帖数: 8753 | 13 我说的这些R都有现成package做的。
multtest, qvalue
基本上就是你通过hypothesis testing得到一堆pvalues之后,用一些方法(包括
permutation tests)重新搞出一些adjusted的p-values来控制family wise error
rate就是我们平常说的type I error rate.
而storey提出的qvalues是针对false discovery rate来做的,less conservative. |
|
|
|
F*******o 发帖数: 4216 | 16 多谢多谢,我去看看
qvalue in R package. |
|
G******n 发帖数: 289 | 17 1%左右差不多。你要是用qvalue选出来多少? |
|
|
h******1 发帖数: 94 | 19 You are right. If I add some big p-values to the vector, it runs correctly.
How to perform FDR correction to get q-values for small p-values? |
|
g********r 发帖数: 8017 | 20 If your problem doesn't generate a mixture, then q-value doesn't help you.
The whole business of FDR build on the assumption that there are many null
in
the large number of tests you conduct.
If it does, just feed it all p-values without pre-selection.
correctly. |
|
g********r 发帖数: 8017 | 21 If the pvalues you listed are all you got, then maybe try p.adjust(, method=
"BH")
correctly. |
|
s*r 发帖数: 2757 | 22 you just have a problem with q0 estimation.
just manually insert a range of q0 |
|
|
s*r 发帖数: 2757 | 24 read the manual?
i do not understand your question and i forgot the detail of the package. |
|
h******1 发帖数: 94 | 25 manual does not say any thing about presetting p0.
It only says
D. Why am I getting an error that says my pi_0 estimate is less than zero?
11
Usually this indicates that the p-values were incorrectly calculated.
Sometimes, π0
is very small, in which case this error can be avoided by using the
bootstrap
option rather than the smoother option. |
|
h******1 发帖数: 94 | 26 How come?
Then how do you know which tests are included in the multitest? How do you
define a group to perform a multi-test?
Could you provide any links or papers for your theory? |
|
g********r 发帖数: 8017 | 27 I don't have a theory. Just intuition. B-H procedure requires low dependency
.
I think by imposing fold change, you are selecting a subset with high
dependency.
you |
|
h******1 发帖数: 94 | 28 I cannot find the low dependency assumption for B-H method on internet.
I need to show it to my boss.
dependency |
|
s*r 发帖数: 2757 | 29 he said it is intuition |
|
h******1 发帖数: 94 | 30 If I include all genes in the study, all of the q-values turn out to be
very large.
dependency |
|
|
g********r 发帖数: 8017 | 32 If you do a histogram of all p-values, does it look like uniform ? If so,
the
q-values should all be large. Claiming any significant gene with a twisted
FDR is 忽悠.
The BH method assumes independent tests. It is in their paper. |
|
h******1 发帖数: 94 | 33 The histogram looks like the attached file.
But the genes are not independent after filtering with fold change.
What can I do? |
|
h******1 发帖数: 94 | 34 It cannot be attached.
It is quite uniform with a small peak on the front middle. |
|
h******1 发帖数: 94 | 35 Yes, I saw it.
But I don't have the ideal histogram for my data, so I probably cannot use
Storey's method.
In order to use B-H method, I have to filter the genes by fold change, in
order to get small q-values, but the preselection could give only dependent
genes in the study.
Do I have any other options? |
|
g********r 发帖数: 8017 | 36 not every experiment generates expected results you know. What if there's no
real biological difference?
Using a stringent p-value cutoff + a fold change cutoff might work, because
p-value or q-value don't cover absolute amount of change.....and in real
biology each condition may not be a single population blah blah blah ....
use
in
dependent |
|
g********r 发帖数: 8017 | 37 Could consider trying more liberal methods, such as local FDR. |
|
h******1 发帖数: 94 | 38 Thank you so much goldmember.
I think you are very knowledgeable. |
|