My Avatar

LanternD's Castle

An electronics enthusiast - survive technically

STT 861 Theory of Prob and STT I Lecture Note - 5

2017-10-04

Sample mean and sample variance, biased and unbiased estimation; covariance, Hypergeometric distribution and its example; correlation coefficients; discrete distribution, Poisson distribution, Poisson approximation for the Binomial distribution.

Portal to all the other notes

Lecture 05 - Oct 05 2017

Sample mean and sample variance

Recall: The proposition $E((X-c)^2)$.

Now consider some data $x_i$, $i=1,2,3,…,n$. We imagine that this data comes from an experiment which is repeated $n$ times independently. This means that $x_i$ represents a r.v. $x_i$, where the $x_i$s are i.i.d.

We are accustomed to using the notation

\[\bar{x}=\frac{1}{n}\sum_{i=1}^{n}x_i\]

This is called “sample mean”.

\[\hat{\sigma}^2=\frac{1}{n}\sum_{i=1}^{n}(x_i-\bar{x})^2\]

This is called “sample variance”.

Now investigate the statistical properties of the two “estimators”. Replace $x_i$ by $X_i$ and try this.

Notation: $\bar{x}$ is for data points, while $\bar{X}$ is for the model notation.

Find the $E(\bar{X})$. If $E(\bar{X})=\mu$, then we say $\bar{X}$ is unbiased.

Find $E(\hat{\sigma}^2)$. Is it $Var(X)$? It might be biased.

\[E(\bar{X})=E(\frac{1}{n}\sum_{i=1}^{n}X_i)=\frac{1}{n}\sum_{i=1}^{n}E(X_i)=\mu\] \[E(\hat{\sigma}^2)=E(\frac{1}{n}\sum_{i=1}^{n}(X_i-\bar{X})^2)\]

The left-hand side of the formula in the previous proposition applied to a r.v. $X$, which is equal to $X_i$ with prob $=\frac{1}{n}$.

The stuff inside the parenthesis is actually the expectation of a r.v. equal to $X_i-\bar{X}$.

\[\begin{align*} \frac{1}{n}\sum_{i=1}^{n}(X_i-\bar{X})^2 &= \frac{1}{n}\sum_{i=1}^{n}(X_i-c)^2-(\bar{X}-c)^2\\ &=\frac{1}{n}\sum_{i=1}^{n}(X_i-\mu)^2-(\bar{X}-\mu)^2\\ \Rightarrow E[\hat{\sigma}^2]&= \frac{1}{n}\sum E(X_i-\mu)-E((\bar{X}-\mu)^2) \\ &=\frac{1}{n}nVar(X)-E((\frac{1}{n}\sum x_i-\mu)^2) \\ &=Var(X)-E(\frac{1}{n}\sum(x_i-\mu)^2) \\ &= -E((\sum x_i-\mu)^2)\\ &= Var(X)- \frac{1}{n}Var(X) \\ &= (1-\frac{1}{n})Var(X) \end{align*}\]

As a result, this is not exactly $=Var(X)$. Thus, $\hat{\sigma}^2$ is biased.

Let’s define an unbiased estimator for $Var(X)$, We just need to take

\[S^2=\frac{1}{n-1}\sum_{i=1}^{n}(x_i-\bar{X})^2\]

It is unbiased estimation of $Var(X)$.


Covariance (Chapter 1.7)

Definition: Let $X$ & $Y$ be two r.v.s living on the same prob space.

\[cov(X,Y)=E((X-E(X))(Y-E(Y))\]

Property: If $X$ & $Y$ are independent, then $cov(X,Y)=0$. Be aware, The statement is usually false, i.e. $cov(X,Y)\neq0$.

Note: if $X=Y$, $cov(X,Y)=Var(X)$.

Property: Let $X_i$, $i=1,2,…,n$ be r.v.’s.

\[\begin{align*} Var(\sum_{i=1}^{n}X_i) &= \sum_{i=1}^{n}\sum_{j=1}^{n}cov(X_i,X_j) \\ &=\sum_{i=1}^{n}Var(X_i)+\sum_{i=1}^{n}\sum_{j=1\neq i}^{n}cov(X_i,X_j) \\ \end{align*}\]

Hypergeometric distribution

Application of the previous formula: The variance of the Hypergeometric distribution (no details here, see the book).

Definition: The hypergeometric distribution with parameter $(n, N)$ is the distribution of the r.v. $X$ of the number of elements from a distinguish subset of size $n$, when one picks a sample of size $k$ without replacement from the $N$ elements.

Example 1

The number $X$ of women in a sample of size $k=5$ taken without replacement from a group with 8 women & 12 men has this hypergeometric distribution with $N=8+12=20$ and $n=8$.

It turns out that

\[Var(X)=kn\frac{N-n}{N-1}\]

Comments: use notation $p=n/N$, then

\[Var(X)=kNp\frac{1-p}{1-\frac{1}{N}}\]

Notice: If $N$ is large, the $\frac{N}{N-1}$ is almost $=1$. So this variance is almost the variance of a binomial with success parameter $p$. This is because if $k$ is much smaller than $N$, sampling without replacement is almost like sampling with replacement.

This “binomial approximation to the hypergeometric law” works well if $k\ll N$, except if $p=n/N$ is too close to 1 or 1.

Correlation coefficients

Let $X$ and $Y$ be two r.v.’s. We standardize them let

\[Z_X=(X-\mu_X)/\sigma_X\] \[Z_Y=(Y-\mu_Y)/\sigma_Y\]

where $\mu_X=E(X)$, $\mu_Y=E(Y)$, $\sigma_X=\sqrt{Var(X)}$, $\sigma_Y=\sqrt{Var(Y)}$.

Notice that $E(Z_X)=E(Z_Y)=0$, $Var(Z_X)=Var(Z_Y)=1$.

Definition: The correlation coefficient between $X$ and $Y$ is

\[Corr(X,Y)=Cov(Z_X,Z_Y)\]

Note: The correlation between $X$ and $Y$ is a value $\in[-1,1]$.

Example 2

Let $X=Y$, then $Corr(X,Y)=1$.

What if $Y=aX+b$, where $a$ and $b$ are constants?

$Corr(X,Y)=1$ if $a>0$, and $=-1$ if $a<0$.

If $X$ and $Y$ are independent, $Corr(X,Y) = 0$.

In general, $Corr(X,Y)$ measures the linear relationship between $X$ and $Y$.

Main idea: If we have a scatter plot of $x$ and $y$ data, which lines up very well along a straight line, then $Corr(X,Y)\triangleq\rho$ will be close to 1 if the line slope up and close to -1 if slop down.

Property: Because $Corr(X,Y)$ is defined using the standardized $Z_X$ and $Z_Y$, then

\[Corr(aX+b, cY+d)=Corr(X,Y)\]

Discrete Distributions (Chapter 2)

Some distributions: $Binom(n,p)$, $Geom(p)$

Important expression:

Recall: the intuition behind the formula $E(Geom(p))=1/p$: For example, if $p=1/20$ for a success and we should expect wo wait 20 units of time until the first success.

Exercise at home

Prove the $E$ and $Var$ for $Geom(p)$.

Poisson Distribution

Definition: $X$ is Poisson distribution distributed with parameter $\lambda$ if $X$ takes the values $k=0,1,2,…$ and

\[P(X=k)=e^{-\lambda}\frac{\lambda^k}{k!}\]

Compute the expectation,

\[\begin{align*} E(X)&=\sum_{k=0}^{\infty}ke^{-\lambda}\frac{\lambda^k}{k!} \\ &=e^{-\lambda}\sum_{k=1}^{\infty}\lambda^k\frac{1}{(k-1)!}\\ &= \lambda e^{-\lambda}\sum_{k=0}^{\infty}\lambda^k\frac{1}{k!}\\ &=\lambda e^{-\lambda}\times e^\lambda = \lambda \end{align*}\]

(Recall: $e^x=\sum_{k=0}^{\infty}x^k\frac{1}{k!}$, Taylor series)

It turns out $Var(X)=\lambda$ [Prove it at home, easier to calculate $E(X(X-1))$].

Quick question: What is $E(X^2)$? $=\lambda+\lambda^2$.

Poisson approximation for the Binomial distribution

Idea: if events are rare, they usually follow a Poisson law.

Fact: Let $X$ be $Bin(n,p)$ and assume $p$ is proportional to $1/n$: $p=\lambda/n$.

Then PMF of $Binom(n, p)$ is almost the same as for $Poi(\lambda)$. Specifically we mean this:

\[\lim\limits_{n\rightarrow\infty}C_n^k(\frac{\lambda}{n})^k(1-\frac{\lambda}{n})^{n-k}=e^{-\lambda}\frac{\lambda^k}{k!}\]

If $p$ is small (of order of $1/n$), then $\#success\sim Bin(n,p)\approx Por(\lambda)$, $\lambda=E(\#success)=np$.

Because of this, Poisson distribution is a good model for number of arrival (of some phenomenon) in a fixed interval of time.

This interpreted as successive units of time (e.g. minutes) in an interval of time, also explains the next property:

Fact: let $n$ and $M$ be two independent Poisson r.v.’s with parameters $\lambda$ and $\mu$, then $X=N+M$ is Poisson too, with parameter $(\lambda+\mu)$.

Because $E(X)=E(N)+E(M)=\lambda+\mu$.

We can use Binomial distribution visualization to prove the fact that $X$ is Poisson.

Exercise

Try to prove $X$ is Poisson using only PMF.



Disqus Comment 0