A**H
发帖数: 4797 | 1 I have data like this
a: 17 2 2 3 3
b: 0 1 0 1 0
The t-test p-value = 0.16. Say this is test one.
If I change the data to the following:
a: 3 2 2 3 3
b: 0 1 0 1 0
The t-test p-value = 0.00022 (which is much smaller than test one.
According to the calculation of t-test, I understand why test one has larger
p-value, because (x-M)^2 becomes larger for test one, therefore t score
getting smaller.
However, this is somehow counterintuitive because the difference of mean
values between two groups in test one is much larger than the second test.
Anyone help me to understand why?
P.S.
Is it because the mean of group a in test one becomes somehow not "reliable"
given that 17 is much larger than other values like 2 and 3? |
A**H
发帖数: 4797 | 2 a: 2 2 3 3
b: 1 0 1 0
If I change the data to the above, the p-value is 0.0027. Let's call this
test three.
test three is significant, why it becomes not significant in test one (while
17 vs 0 is significant enough comparing to 2s or 3s vs 1s or 0s)? Two
significant ones adding together then become insignificant?
Sorry for English input. No Chinese now.
Thanks. |
c********h
发帖数: 330 | 3 我觉得是variance的问题吧，你的t-test statistic要除以var，明显第一组的var大很
多啊，这样t-value就会变小了，所以不显著 |
h***i
发帖数: 3844 | 4 y, variance issue.
【在 c********h 的大作中提到】
: 我觉得是variance的问题吧，你的t-test statistic要除以var，明显第一组的var大很
: 多啊，这样t-value就会变小了，所以不显著
|
p*******i
发帖数: 1181 | 5 Yes, high var -> less reliable -> larger p-value |
A**H
发帖数: 4797 | 6 Got it. Thanks all for the input. |
q******n
发帖数: 272 | 7 small N, Nonnormal. Use nonparametric test, Wilcoxon test. |
t*****a
发帖数: 459 | 8 同意楼上。这个用t test就是个典型的陷阱。
p value 也不能太认真。根据情况，p value有时候只能认真前几位，比如p=0.002和p=
0.008的test结果并不能用来比较两个test谁significant。这个例子里p value连前2位
都已经不可靠，不能用来做比较。 |
A**H
发帖数: 4797 | 9 看了看wiki，好像很有道理，得出的p大约0.02，和直觉比较相符，呵呵。
【在 q******n 的大作中提到】
: small N, Nonnormal. Use nonparametric test, Wilcoxon test.
|
s****b
发帖数: 2039 | 10 group a has larger variance.
larger
【在 A**H 的大作中提到】
: I have data like this
: a: 17 2 2 3 3
: b: 0 1 0 1 0
: The t-test p-value = 0.16. Say this is test one.
: If I change the data to the following:
: a: 3 2 2 3 3
: b: 0 1 0 1 0
: The t-test p-value = 0.00022 (which is much smaller than test one.
: According to the calculation of t-test, I understand why test one has larger
: p-value, because (x-M)^2 becomes larger for test one, therefore t score
|
