1. it has 5% rotating category; not necessarily the drug store, although the drug store is very impressive due to vanilla 2. the old card mush be worse than dividend... 3. ask CSR about conversion details; ur card may not be qualified for the conversion, anyway
【在 s**********g 的大作中提到】 : 1. it has 5% rotating category; not necessarily the drug store, although the : drug store is very impressive due to vanilla : 2. the old card mush be worse than dividend... : 3. ask CSR about conversion details; ur card may not be qualified for the : conversion, anyway
h*o
6 楼
这个不能用paired t-test 如果你觉得不同批次之间测试的绝对数值相差很大,但是趋势一致的时候,一个不严谨 但是有用的方法是,都 normalize到control的平均值以后再combine 做stats 比如 第一次做, control VS treatment, n=5 per group 那么你会得到control的平均值Ca, normalize以后 control的五个replicate 分别是c 1, c5 , treatment分别是T1-T5 第二次做,control vs treatment, n=5 这一次的control平均值是Cb,normalize以后 得到C6-C10 以及T6-T10 依次类推, 最后你有C1-C15, T 1-T15,然后做T -test 还有一种方法是每次都include一个bridging sample,然后每次都normalize到那个 sample 用paired t-test 不合适, 等于认为的缩小的p值
If the difference between batches are not measurable by other variables, then accounting for the correlation within batch is probably the best you can do. If your outcome of experiment is continuous, maybe plot them by batch and just eyeball it to see if there is some difference between different batches of cells. If there is, what might be more appropriate (depending on how much sample you have) would be some sort of mixed effect model, with exposure being the treatment (experiment vs. control) and batch numbers being the random effect . I have to see the data to tell what exactly to do. If the average is influenced by the batch, you would need a random intercept, but if the batch influence how well the cells respond to the treatment, you might also need some sort of random slope of treatment, I think. Alternatively, if you have enough repeats within batches, you could do an MANOVA adjusting for both treatment and batch. It might not be totally appropriate, but could be better than just a t-test. Given you have enough samples to estimate the difference among batches.
This is a very good post. I have also thought about this problem. In short and in general, I believe the trend is that a positive underlying correlation between pairs will yield a smaller p-value, thus higher power for detecting the same difference. But I am not sure this is 100% true, because mathematically I think the degree of freedom should also play a role in determining the p-value.
【在 f******e 的大作中提到】 : This is a very good post. I have also thought about this problem. In short : and in general, I believe the trend is that a positive underlying : correlation between pairs will yield a smaller p-value, thus higher power : for detecting the same difference. But I am not sure this is 100% true, : because mathematically I think the degree of freedom should also play a role : in determining the p-value.
r*r
19 楼
A few things to think about: 1. if use typical alpha value, i.e, 5% 2. If 1., then how about your p values of your two cases (paired and unpaired). If they are close to 5%, you need to be very careful. sometimes you will get the totally different answers: P value is just greater than 5% or just less than 5%.Otherwise, different T-tests will get the same answer ( both have big p values,e.g., ~30%), then you will be fine anyway. Using google, then you'll find good examples how to define unpaired and paired T-tests.
这不就是假设培养中可能出现各种variation吗,跟你的背景假设完全180度相反,要不 然你怎么能解释第一种情况是“as bad as it can be”?
【在 c****1 的大作中提到】 : 这篇文章很好。解释的很详细。建议大家都看看。
c*1
29 楼
呵呵,你没有仔细看。我的情况是design 3.
【在 D*a 的大作中提到】 : 这不就是假设培养中可能出现各种variation吗,跟你的背景假设完全180度相反,要不 : 然你怎么能解释第一种情况是“as bad as it can be”?
M*P
30 楼
paired t test 就是一个subject 测一下,然后treatment之后再测一下。 你除非把细胞一直养在一起,然后做实验时分出两部分,一部分treat,一部分control ,然后做paired t test才合理。但是要是你treat时间远远长于一起培养的时间,那你 的差异就出现了,应该用前面说的那个mixed model。
The key is to understand the biology, not playing number. P<0.10 does not mean the hypothesis were wrong. If you have to use paired t test to improve the statistical power from non significant to significant, it may indicate the change of your gene of interest in your cell lines vary greatly at each time you perform the experiment. But if both test indicate p<0.05, why bother? You should run different types of cell lines to see if the change is concordant. Then it may make sense to use paired t test.
A*y
35 楼
Exactly,if you ever read prologue of some statistics text or program (i.e. Prism). The author literally said that if you have to use the program to tell you if you hypothesis is significant or not; it is likely NOT in the real life. For example, Gleevec phase I is 100% effective! Do you really need a program to tell you that it works?
【在 c********e 的大作中提到】 : The key is to understand the biology, not playing number. : P<0.10 does not mean the hypothesis were wrong. : If you have to use paired t test to improve the statistical power from non : significant to significant, it may indicate the change of your gene of : interest in your cell lines vary greatly at each time you perform the : experiment. : But if both test indicate p<0.05, why bother? : You should run different types of cell lines to see if the change is : concordant. Then it may make sense to use paired t test. :