d*******5 发帖数: 87 | 1 题目是x是服从正态,但是观测值最大只有8,问如何求mean。求教一下大牛们 |
X*7 发帖数: 40 | 2 如何求mean,这个mean是指正态的mean?能求得出来? |
v*******e 发帖数: 11604 | 3
假定有N个观测值,最大为8。那么写出最大值的pdf(这个可以直接写出,用order
statistics,Casella and Berger 书上有公式),然后用maximum likelihood方法求均
值的期望。最后应该是个关于N的函数。
【在 d*******5 的大作中提到】 : 题目是x是服从正态,但是观测值最大只有8,问如何求mean。求教一下大牛们
|
d*******5 发帖数: 87 | 4
我算了一下是
int_-inf^y_k {f(x) [(k-1)(u-x)/sigma^2-(y_k-u)/sigma^2]}dx
note: f(x) is the pdf of normal(u, sigma), y_k is the largest value
这个函数怎么求解u呢?估计是个trivial的问题,不过我不是很清楚
【在 v*******e 的大作中提到】 : : 假定有N个观测值,最大为8。那么写出最大值的pdf(这个可以直接写出,用order : statistics,Casella and Berger 书上有公式),然后用maximum likelihood方法求均 : 值的期望。最后应该是个关于N的函数。
|
v*******e 发帖数: 11604 | 5 若没办法求解析解,就用数值方法求解。用R或者别得什么,对确定的N,画个曲线,肉
眼找最大值亦可----反正就是个面试题,主要看思路。 |
H**********f 发帖数: 2978 | 6 这就是想看看你能不能说出个123吧,比如known or unknown variance,order
statistic,pdf,MLE,Newton-Raphson,prior distribution,posterior mode,就
扯呗,面试当场不太可能做得出来吧,回头也没必要再较真了
【在 d*******5 的大作中提到】 : 题目是x是服从正态,但是观测值最大只有8,问如何求mean。求教一下大牛们
|
h***i 发帖数: 3844 | 7 你平常看到一组数,要求mean 和var,也会这么算么?
【在 v*******e 的大作中提到】 : 若没办法求解析解,就用数值方法求解。用R或者别得什么,对确定的N,画个曲线,肉 : 眼找最大值亦可----反正就是个面试题,主要看思路。
|
g******2 发帖数: 234 | 8 is it possible the question is on how to compute the mean of a truncated
normal distribution E(X|X<8) ? |
d*******5 发帖数: 87 | 9 这个mean只跟最大值有关系么?其他的点的值都不用? |
v*******e 发帖数: 11604 | 10
你自己说只知道最大值为8.
【在 d*******5 的大作中提到】 : 这个mean只跟最大值有关系么?其他的点的值都不用?
|
R*****0 发帖数: 146 | 11 It depends on how you understand the question. One way is as verdelite
suggested, seeking MLE for the largest observation based on a single data
point, 8; the other way is as getdown2 said, assuming a truncated Normal
distribution. To do that, refer to the 'Moments' section for the link below:
http://en.wikipedia.org/wiki/Truncated_normal_distribution
In your case b=8 and once you know the SD(sigma), you can let E(X|X<8)=
sample mean and solve for the mean(mu). Of course, if you know the largest
value 8 deviated a lot from the real mean(mu), then it won't hurt if you let
mu=sample mean being your estimate.
I think either way to do it makes sense but I prefer the second one since it
makes use of all the data (maybe too much for that value 8).
【在 d*******5 的大作中提到】 : 这个mean只跟最大值有关系么?其他的点的值都不用?
|
g******2 发帖数: 234 | 12 if it's really a question on calculating /mu from truncated normal, it
should be very straight-forward.
1. Look at \hat{P}(X == 8)
2. if \hat{P}(X == 8) < 0.5, use the median as the estimate of /mu
if \hat{p}(x == 8) > 0.5, then we know \mu + \sigma * Z_{\hat{p}(x == 8)}
= 8. We need another equation to solve for \mu and \sigma. Just find
another quantile that's smaller than 8, say X. The second equation would be
\mu + \sigma * Z_{\hat{p}(x <= X)} = X. (I would suggest to use X such that
\hat{p}(x <= X) = 0.5 * (1 - \hat{p}(x == 8)) ) |
w*******9 发帖数: 1433 | 13 You need information on the variance to say the least. Just imagine if the
variance is VERY small,
the true mean would be somewhere close to 8. If the variance is huge, the
true mean could be way smaller than 8. |