j*****e
发帖数: 182 | 1 In bootstrap, you use the sample to mimic the population. When you resample,
the size of your resample should be the same as the size of your observed
sample. The thousands of resamples generate a distribution of your test
statistic/estimator. This will allow you to compute a p-value or construct a
CI. Depending on whether you are doing a hypothesis test or parameter
estimation, there are different ways to resample. For example, you can
resample by permuting your original sample, you can resampl |
|
b********n
发帖数: 95 | 2 So, what is the bootstrapping really used in the reality?
At first i just thought it was used in some cases which are lack of
information. Like sample pool is too small to estimate the statistics like
SD, C.I.
But up to you guys, it seems just a method to calculate the statistics with
the plengy population. If that is the truth, why we do not use the
traditional statistical method to do that? why we need to do the extra work
for the resampling?
a |
|
A*******s
发帖数: 3942 | 3 no, not my idea. I don't see any benefit of using bootstrapping when you
have the access to a large sample.
with
work |
|
b********n
发帖数: 95 | 4 So, what is the bootstrapping really used in the reality?
At first i just thought it was used in some cases which are lack of
information. Like sample pool is too small to estimate the statistics like
SD, C.I.
But up to you guys, it seems just a method to calculate the statistics with
the plengy population. If that is the truth, why we do not use the
traditional statistical method to do that? why we need to do the extra work
for the resampling? |
|
s*r
发帖数: 2757 | 5 check a few population application of bootstrap
1) bagging
2) confidence on the fork of genealogy/molecular evolution tree
read the famous book from efron and tell us the results |
|
g********r
发帖数: 8017 | 6 很多时候问题复杂，经典统计根本不提供CI呀。或者数据维数高，用景点方法必
须做很多假定，还不如直接从数据里学了。
bootstrap有时候也是乱用，前几年有人比较过很多个生物芯片数据的resampling
方法，结论是多数方法都是bias的。
with
work |
|
p********a
发帖数: 5352 | 7 ☆─────────────────────────────────────☆
beandragon (beandragon) 于 (Wed Mar 24 12:44:51 2010, 美东) 提到:
When using bootstrapping, we randomly extract a new sample of n heights out
of the N sampled data, where each person can be selected at most t times. By
doing this several times, we create a large number of datasets that we
might have seen and compute the statistic for each of these datasets.
在上面这段描述中，t,n,N之间应该有些什么关系，我觉得t应该是个比较大的值，我们
应该尽量使得挑选的case不会到达t这个极限值，而n也应该比t*N小很多，这样才达到
resampling的目的。但 |
|
t******g
发帖数: 2253 | 8 推荐efron的。efron是bootstrap的始祖，肯定要看他的书。 |
|
t**u
发帖数: 1572 | 9 Bootstrap Methods and Their Application (Cambridge Series in Statistical and
Probabilistic Mathematics , No 1) [Paperback]
A. C. Davison
A. C. Davison (Author)
巨包酬谢, THANK YOU VERY MUCH |
|
t**u
发帖数: 1572 | 10 Bootstrap Methods and Their Application (Cambridge Series in Statistical and
Probabilistic Mathematics , No 1) [Paperback]
A. C. Davison
A. C. Davison (Author)
巨包酬谢, THANK YOU VERY MUCH |
|
l*********s
发帖数: 5409 | 11 Surely there will be variation, unless the sample to begin with has zero
variance.
Its feasibility does not reply on normality or CLT, so it is more general.
可因为只有一个样本，怎么bootstrap显然都找不到variation。 |
|
j*******y
发帖数: 58 | 12 FT, apprently you don't know how to do bootstrap.
比如我从无限整
可因为只有一个样 |
|
P****D
发帖数: 11146 | 13 You cannot draw ONE sample and then bootstrap on it! |
|
l***o
发帖数: 5337 | 14 我当然知道这不行。我只是举一个极端的例子。就是要问bootstrap可行的条件。谢谢! |
|
P****D
发帖数: 11146 | 15 What do you mean by "可行的条件"? You want to know the minimum/maximum
sample size that you can apply bootstrapping on? I never thought of that
myself... I simply always use the original sample size.
If I were you, I would use a realistic example to explain what I meant.
谢! |
|
r*****y
发帖数: 199 | 16 I took a course of bootstrap this semester, it seems the situation we
discussed a lot is for small sample. For large sample, you can always rely
on the asymptotic theory. |
|
j*******y
发帖数: 58 | 17 no people use bootstrap to estimate mean. you already have a sample mean.
for variance, 30 is enough. 20 may not. precision is lower.
。 |
|
l***o
发帖数: 5337 | 18 那位大牛指教一下用CLT和bootstrap估计conficence interval
的区别，
即基于什么样的样本和数据结构，那种方法更适用。多谢！
mean. |
|
|
p********r
发帖数: 1465 | 20 For example, we have 10 values of weights. It would not be precise for us to
estimate the population mean weight and its confidence interval based on
normality assumption. For bootstrap, first of all, we draw a sample of size
10 with replacement and obtain the average weight. Then we repeat this
procedure 100 times and we can get a new sample of 100 average weights. Now
we can obtain a more precise estimate and a better distribution of the
average weight. |
|
l***o
发帖数: 5337 | 21 谢谢大师出手！ 我举的一个样本的例子只是想引出问题，并不是真要考虑这么做。请
问大师，如果你抛
弃bootstrap，我完全没意见，但按照你的理论，怎么估算estimate的variance问题？
谢谢！
p.s.阁下的理论有什么应用吗？还是只是谈着开心用的？
sample
谬。 |
|
r*****y
发帖数: 199 | 22
sample
谬。
I totally disagree with you. I think you have no idea about the theory and
application of bootstrap. |
|
T*******I
发帖数: 5138 | 23 我的问题是：bootstrapped的样本可不可加？如果不可加，为什么？如果可加，为什么
？后一个问题我已经在我的09年的JSM文章里作出了肯定的回答：它们可加，因为它们
本来就来自同一总体！反之，如果说它们不可加，则整个统计学的基础将发生动摇，而
这种可能性不存在。 |
|
d******e
发帖数: 7844 | 24 如果明白了bootstrap的基本原理,用R实现也就是七八行就搞定了啊. |
|
c*2
发帖数: 24 | 25 Not an expert, but I heard bootstrap on residues may be more appropriate. |
|
r*****y
发帖数: 199 | 26 Why on residual, you don't even need distribution assumption when you use
bootstrap to perform two-sample comparison. |
|
c*2
发帖数: 24 | 27 Once I worked with a prominent statistician. He suggested the bootstrap
residual approach. You may want to consult with a statistician. |
|
h**********6
发帖数: 66 | 28 如题。本人不是统计专业，只会一些简单R，看到paper上用bootstrap获得95%CI，但不
知道怎么编程，，，若有牛人发个简单的R script例子，不胜感激（仅做学习参考之用
）
Email：t****************[email protected] 非常非常感谢 ：） |
|
h**********6
发帖数: 66 | 29 假设A，B follow一个函数关系，对于每一个固定的B（已知），有5个A值与之对应，这
5个A假设follow normal distribution，然后我要通过A的varibility去估计C的var，C
和A follow以下关系： C=∑A*B。要用bootstrap估计C的confidence interval。
谢谢~~ |
|
h**********6
发帖数: 66 | 30 这个是simulation？bootstrapping不是要从原有data（就是那些A）中做resampling么？ |
|
r*****y
发帖数: 199 | 31 随便找一根efron的bootstrap的书稍微看看啊 |
|
l*********s
发帖数: 5409 | 32 需要bootstrap 连续平方和的期望值，cpu只是双核，还是慢；俺想用显卡做并行来提
高效率，有高手能讲讲怎么做嘛？多谢 |
|
|
d********t
发帖数: 837 | 34 bootstrap的convergence问题80年代就解决了。问题你连他要估计什么都不知道。 |
|
d********t
发帖数: 837 | 35 bootstrap的convergence问题80年代就解决了。问题你连他要估计什么都不知道。 |
|
j*****e
发帖数: 182 | 36 Not all the asymptotics have been solved. This paper is just a proof for a
special case. What kind of bootstrap works is still on-going research. |
|
s*****1
发帖数: 9 | 37 The bootstrap is one way to use data to do statistical inference. Whether it
can reveal truth depends on the quality of your data and your model. |
|
q**j
发帖数: 10612 | 38 这玩意难道就这样说说就有道理了？不是有人证明了sample mean是population mean的
suffiicent statistic了么？sample vairance是population variance的sufficient s
tatitic 了么？这样还boot什么？
难道distribution不是normal的情况下boot就有用了？这个是否应该证明一下，什么情
况下boot比较有用。
的简
boot
bootstrap
samp
来的 |
|
t****r
发帖数: 702 | 39 当然不是这样说说就有道理了，严格的证明也有很多人做啊，只是很多人都不会去看而
已。
而且统计很多做法，最重要的还是intuition，或者大师所津津乐道的哲学（逻辑）。你
所说的sufficient statistic，只是对poupulation mean 或者population variance来
说，所有样本中关于这两个参数的信息都被summarize到sample mean或者sample varia
nce中了。但是同时如果考虑到他们又是complete statistic，也就是说他们是包含所
有这些参数信息“最小的”statistics. 很显然，他们过滤掉了很多原来样本中有用的
信息。 最好的sufficient statistic就是样本本身哈。 很简单一个例子就是，我们要
刻画一个变量的分布，显然仅仅有mean和variance是不够的。除了normal 的
distribution之外，有相同mean和variance的分布有无穷多个。
再用tree classifer来做例子。如果我们可以试图定义一个“sufficient statistic”
来进行分类，那... 阅读全帖 |
|
k********g
发帖数: 56 | 40 赞，这样严肃认真的讨论一些理论问题感觉真好。
的简
boot
bootstrap
samp
来的 |
|
|
p*******g
发帖数: 809 | 42 有没有人用bootstrap做high-dimensional data的统计分析的? 有何评价? |
|
|
T*******I
发帖数: 5138 | 44 Bootstrap法的构建、展开式讨论和应用都必须以那个命题得到了证明为前提，否则一
切都是枉然或陷入数字游戏。
be
estimate
is
all |
|
w****p
发帖数: 167 | 45 So,can I understand your explanation like this?
A and B are two estimators of a parameter. A is biased but B is not.
Bootstrap can let us know the bias of A, so (A-bias) is unbiased estimate.
But why don't we just use B as the estimator? |
|
r***c
发帖数: 398 | 46 you are expert. that is main assumption made in bootstrap
to
a
is |
|
T*******I
发帖数: 5138 | 47 那个bias的方向也可能是反的，所以，最后的估计结果也可以是
A+(B-A)
我的意思是，下一组1000次的bootstrap可能会给你一个反向的bias. |
|
A*****r
发帖数: 795 | 48 bootstrap当然是第一个了
孙天使好歹也是phd啊.......汗
replacement |
|
I*****a
发帖数: 5425 | 49 1) If you can't get the distributions of ahat and bhat without bootstrapping
, method 2 won't give that either.
2) If you can directly estimate, say the variance of ahat and bhat, by
assumptions, then method 2 tends to give you smaller variances with bias.
replacement |
|
c***z
发帖数: 6348 | 50 thanks for sharing! I am studying bootstrap too, for a totally different
purpose: adjust for sample bias
just wondering if the following would work:
if my sample has 35% male, while I know that the population has 50% male. I
am resampling with replacement, so that my new sample has 50% male. Then I
do things with the new sample. |
|
