Deriva la varianza del coefficiente di regressione nella regressione lineare semplice

Nella regressione lineare semplice, abbiamo $y = \beta_0 + \beta_1 x + u$ , dove $u \sim iid\;\mathcal N(0,\sigma^2)$ . Ho derivato lo stimatore:

\hat{β_{1}} = \frac{\sum_{i} (x_{i} - \bar{x}) (y_{i} - \bar{y})}{\sum_{i} (x_{i} - \bar{x})^{2}},

$\hat{\beta_1} = \frac{\sum_i (x_i - \bar{x})(y_i - \bar{y})}{\sum_i (x_i - \bar{x})^2}\ ,$ dove

\bar{x}

$\bar{x}$ e

\bar{y}

$\bar{y}$ sono i mezzi di campionamento di

x

$x$ ed

y

$y$ .

Ora voglio trovare la varianza di . Ho derivato qualcosa di simile al seguente: $\hat\beta_1$

Var (\hat{β_{1}}) = \frac{σ^{2} (1 - \frac{1}{n})}{\sum_{i} (x_{i} - \bar{x})^{2}} .

$\text{Var}(\hat{\beta_1}) = \frac{\sigma^2(1 - \frac{1}{n})}{\sum_i (x_i - \bar{x})^2}\ .$

La derivazione è la seguente:

\begin{aligned} Var (\hat{β_{1}}) \\ = Var (\frac{\sum_{i} (x_{i} - \bar{x}) (y_{i} - \bar{y})}{\sum_{i} (x_{i} - \bar{x})^{2}}) \\ = \frac{1}{(\sum_{i} (x_{i} - \bar{x})^{2})^{2}} Var (\sum_{i} (x_{i} - \bar{x}) (β_{0} + β_{1} x_{i} + u_{i} - \frac{1}{n} \sum_{j} (β_{0} + β_{1} x_{j} + u_{j}))) \\ = \frac{1}{(\sum_{i} (x_{i} - \bar{x})^{2})^{2}} Var (β_{1} \sum_{i} (x_{i} - \bar{x})^{2} + \sum_{i} (x_{i} - \bar{x}) (u_{i} - \sum_{j} \frac{u_{j}}{n})) \\ = \frac{1}{(\sum_{i} (x_{i} - \bar{x})^{2})^{2}} Var (\sum_{i} (x_{i} - \bar{x}) (u_{i} - \sum_{j} \frac{u_{j}}{n})) \\ = \frac{1}{(\sum_{i} (x_{i} - \bar{x})^{2})^{2}} \times \\ E [{(\sum_{i} (x_{i} - \bar{x}) (u_{i} - \sum_{j} \frac{u_{j}}{n}) - \underset{= 0}{\underset{⏟}{E [\sum_{i} (x_{i} - \bar{x}) (u_{i} - \sum_{j} \frac{u_{j}}{n})]}})}^{2}] \\ = \frac{1}{(\sum_{i} (x_{i} - \bar{x})^{2})^{2}} E [{(\sum_{i} (x_{i} - \bar{x}) (u_{i} - \sum_{j} \frac{u_{j}}{n}))}^{2}] \\ = \frac{1}{(\sum_{i} (x_{i} - \bar{x})^{2})^{2}} E [\sum_{i} (x_{i} - \bar{x})^{2} (u_{i} - \sum_{j} \frac{u_{j}}{n})^{2}], since u_{i}'s are iid \\ = \frac{1}{(\sum_{i} (x_{i} - \bar{x})^{2})^{2}} \sum_{i} (x_{i} - \bar{x})^{2} E {(u_{i} - \sum_{j} \frac{u_{j}}{n})}^{2} \\ = \frac{1}{(\sum_{i} (x_{i} - \bar{x})^{2})^{2}} \sum_{i} (x_{i} - \bar{x})^{2} (E (u_{i}^{2}) - 2 \times E (u_{i} \times (\sum_{j} \frac{u_{j}}{n})) + E {(\sum_{j} \frac{u_{j}}{n})}^{2}) \\ = \frac{1}{(\sum_{i} (x_{i} - \bar{x})^{2})^{2}} \sum_{i} (x_{i} - \bar{x})^{2} (σ^{2} - \frac{2}{n} σ^{2} + \frac{σ^{2}}{n}) \\ = \frac{σ^{2}}{\sum_{i} (x_{i} - \bar{x})^{2}} (1 - \frac{1}{n}) \end{aligned}

$\begin{align} &\text{Var}(\hat{\beta_1})\\ & = \text{Var} \left(\frac{\sum_i (x_i - \bar{x})(y_i - \bar{y})}{\sum_i (x_i - \bar{x})^2} \right) \\ & = \frac{1}{(\sum_i (x_i - \bar{x})^2)^2} \text{Var}\left( \sum_i (x_i - \bar{x})\left(\beta_0 + \beta_1x_i + u_i - \frac{1}{n}\sum_j(\beta_0 + \beta_1x_j + u_j) \right)\right)\\ & = \frac{1}{(\sum_i (x_i - \bar{x})^2)^2} \text{Var}\left( \beta_1 \sum_i (x_i - \bar{x})^2 + \sum_i(x_i - \bar{x}) \left(u_i - \sum_j \frac{u_j}{n}\right) \right)\\ & = \frac{1}{(\sum_i (x_i - \bar{x})^2)^2}\text{Var}\left( \sum_i(x_i - \bar{x})\left(u_i - \sum_j \frac{u_j}{n}\right)\right)\\ & = \frac{1}{(\sum_i (x_i - \bar{x})^2)^2}\;\times \\ &\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;E\left[\left( \sum_i(x_i - \bar{x})(u_i - \sum_j \frac{u_j}{n}) - \underbrace{E\left[\sum_i(x_i - \bar{x})(u_i - \sum_j \frac{u_j}{n})\right] }_{=0}\right)^2\right]\\ & = \frac{1}{(\sum_i (x_i - \bar{x})^2)^2} E\left[\left( \sum_i(x_i - \bar{x})(u_i - \sum_j \frac{u_j}{n})\right)^2 \right] \\ & = \frac{1}{(\sum_i (x_i - \bar{x})^2)^2} E\left[\sum_i(x_i - \bar{x})^2(u_i - \sum_j \frac{u_j}{n})^2 \right]\;\;\;\;\text{ , since } u_i \text{ 's are iid} \\ & = \frac{1}{(\sum_i (x_i - \bar{x})^2)^2}\sum_i(x_i - \bar{x})^2E\left(u_i - \sum_j \frac{u_j}{n}\right)^2\\ & = \frac{1}{(\sum_i (x_i - \bar{x})^2)^2}\sum_i(x_i - \bar{x})^2 \left(E(u_i^2) - 2 \times E \left(u_i \times (\sum_j \frac{u_j}{n})\right) + E\left(\sum_j \frac{u_j}{n}\right)^2\right)\\ & = \frac{1}{(\sum_i (x_i - \bar{x})^2)^2}\sum_i(x_i - \bar{x})^2 \left(\sigma^2 - \frac{2}{n}\sigma^2 + \frac{\sigma^2}{n}\right)\\ & = \frac{\sigma^2}{\sum_i (x_i - \bar{x})^2}\left(1 - \frac{1}{n}\right) \end{align}$

Did I do something wrong here?

I know if I do everything in matrix notation, I would get ${\rm Var}(\hat{\beta_1}) = \frac{\sigma^2}{\sum_i (x_i - \bar{x})^2}$ . But I am trying to derive the answer without using the matrix notation just to make sure I understand the concepts.

— mynameisJEFF
fonte

Yes, your formula from matrix notation is correct. Looking at the formula in question,

1 - \frac{1}{n} = \frac{n - 1}{n}

$1-\frac1{n}\,=\,\frac{n-1}{n}$ so it rather looks as if you might used a sample standard deviation somewhere instead of a population standard deviation? Without seeing the derivation it's hard to say any more.

— TooTone

General answers have also been posted in the duplicate thread at stats.stackexchange.com/questions/91750.

— whuber

Risposte:

At the start of your derivation you multiply out the brackets $\sum_i (x_i - \bar{x})(y_i - \bar{y})$ , in the process expanding both $y_i$ and $\bar{y}$ . The former depends on the sum variable $i$ , whereas the latter doesn't. If you leave $\bar{y}$ as is, the derivation is a lot simpler, because

\begin{aligned} \sum_{i} (x_{i} - \bar{x}) \bar{y} & = \bar{y} \sum_{i} (x_{i} - \bar{x}) \\ = \bar{y} ((\sum_{i} x_{i}) - n \bar{x}) \\ = \bar{y} (n \bar{x} - n \bar{x}) \\ = 0 \end{aligned}

$\begin{align} \sum_i (x_i - \bar{x})\bar{y} &= \bar{y}\sum_i (x_i - \bar{x})\\ &= \bar{y}\left(\left(\sum_i x_i\right) - n\bar{x}\right)\\ &= \bar{y}\left(n\bar{x} - n\bar{x}\right)\\ &= 0 \end{align}$

Hence

\begin{aligned} \sum_{i} (x_{i} - \bar{x}) (y_{i} - \bar{y}) & = \sum_{i} (x_{i} - \bar{x}) y_{i} - \sum_{i} (x_{i} - \bar{x}) \bar{y} \\ = \sum_{i} (x_{i} - \bar{x}) y_{i} \\ = \sum_{i} (x_{i} - \bar{x}) (β_{0} + β_{1} x_{i} + u_{i}) \end{aligned}

$\begin{align} \sum_i (x_i - \bar{x})(y_i - \bar{y}) &= \sum_i (x_i - \bar{x})y_i - \sum_i (x_i - \bar{x})\bar{y}\\ &= \sum_i (x_i - \bar{x})y_i\\ &= \sum_i (x_i - \bar{x})(\beta_0 + \beta_1x_i + u_i )\\ \end{align}$

and

\begin{aligned} Var (\hat{β_{1}}) & = Var (\frac{\sum_{i} (x_{i} - \bar{x}) (y_{i} - \bar{y})}{\sum_{i} (x_{i} - \bar{x})^{2}}) \\ = Var (\frac{\sum_{i} (x_{i} - \bar{x}) (β_{0} + β_{1} x_{i} + u_{i})}{\sum_{i} (x_{i} - \bar{x})^{2}}), substituting in the above \\ = Var (\frac{\sum_{i} (x_{i} - \bar{x}) u_{i}}{\sum_{i} (x_{i} - \bar{x})^{2}}), noting only u_{i} is a random variable \\ = \frac{\sum_{i} (x_{i} - \bar{x})^{2} Var (u_{i})}{{(\sum_{i} (x_{i} - \bar{x})^{2})}^{2}}, independence of u_{i} and, Var (k X) = k^{2} Var (X) \\ = \frac{σ^{2}}{\sum_{i} (x_{i} - \bar{x})^{2}} \end{aligned}

$\begin{align} \text{Var}(\hat{\beta_1}) & = \text{Var} \left(\frac{\sum_i (x_i - \bar{x})(y_i - \bar{y})}{\sum_i (x_i - \bar{x})^2} \right) \\ &= \text{Var} \left(\frac{\sum_i (x_i - \bar{x})(\beta_0 + \beta_1x_i + u_i )}{\sum_i (x_i - \bar{x})^2} \right), \;\;\;\text{substituting in the above} \\ &= \text{Var} \left(\frac{\sum_i (x_i - \bar{x})u_i}{\sum_i (x_i - \bar{x})^2} \right), \;\;\;\text{noting only $u_i$ is a random variable} \\ &= \frac{\sum_i (x_i - \bar{x})^2\text{Var}(u_i)}{\left(\sum_i (x_i - \bar{x})^2\right)^2} , \;\;\;\text{independence of } u_i \text{ and, Var}(kX)=k^2\text{Var}(X) \\ &= \frac{\sigma^2}{\sum_i (x_i - \bar{x})^2} \\ \end{align}$

which is the result you want.

As a side note, I spent a long time trying to find an error in your derivation. In the end I decided that discretion was the better part of valour and it was best to try the simpler approach. However for the record I wasn't sure that this step was justified

\begin{aligned} = . \frac{1}{(\sum_{i} (x_{i} - \bar{x})^{2})^{2}} E [{(\sum_{i} (x_{i} - \bar{x}) (u_{i} - \sum_{j} \frac{u_{j}}{n}))}^{2}] \\ = \frac{1}{(\sum_{i} (x_{i} - \bar{x})^{2})^{2}} E [\sum_{i} (x_{i} - \bar{x})^{2} (u_{i} - \sum_{j} \frac{u_{j}}{n})^{2}], since u_{i}'s are iid \end{aligned}

$\begin{align} & =. \frac{1}{(\sum_i (x_i - \bar{x})^2)^2} E\left[\left( \sum_i(x_i - \bar{x})(u_i - \sum_j \frac{u_j}{n})\right)^2 \right] \\ & = \frac{1}{(\sum_i (x_i - \bar{x})^2)^2} E\left[\sum_i(x_i - \bar{x})^2(u_i - \sum_j \frac{u_j}{n})^2 \right]\;\;\;\;\text{ , since } u_i \text{ 's are iid} \\ \end{align}$ because it misses out the cross terms due to

\sum_{j} \frac{u_{j}}{n}

$\sum_j \frac{u_j}{n}$ .

— TooTone
fonte

I noticed that I could use the simpler approach long ago, but I was determined to dig deep and come up with the same answer using different approaches, in order to ensure that I understand the concepts. I realise that first

\sum_{j} \hat{u_{j}} = 0

$\sum_j \hat{u_j} = 0$ from normal equations (FOC from least square method), so

\bar{\hat{u}} = \frac{\sum_{i} u_{i}}{n} = 0

$\bar{\hat{u}} = \frac{\sum_i u_i}{n}=0$ , plus

\bar{\hat{u}} = \bar{y} - \bar{\hat{y}} = 0

$\bar{\hat{u}} = \bar{y} - \bar{\hat{y}} = 0$ , so

\bar{y} = \bar{\hat{y}}

$\bar{y} = \bar{\hat{y}}$ . So there won't be the term

\sum_{j} \frac{u_{j}}{n}

$\sum_j \frac{u_j}{n}$ in the first place.

— mynameisJEFF

ok, in your question the emphasis was on avoiding matrix notation.

— TooTone

Yes, because I was able to solve it using matrix notation. And notice from my last comment, I did not use any linear algebra. Thanks for your great answer anyway^.^

— mynameisJEFF

sorry are we talking at cross-purposes here? I didn't use any matrix notation in my answer either, and I thought that was what you were asking in your question.

— TooTone

sorry for misunderstanding haha...

— mynameisJEFF

I believe the problem in your proof is the step where you take the expected value of the square of $\sum_i (x_i - \bar{x} )\left( u_i -\sum_j \frac{u_j}{n} \right)$ . This is of the form $E \left[\left(\sum_i a_i b_i \right)^2 \right]$ , where $a_i = x_i -\bar{x}; b_i = u_i -\sum_j \frac{u_j}{n}$ . So, upon squaring, we get $E \left[ \sum_{i,j} a_i a_j b_i b_j \right] = \sum_{i,j} a_i a_j E\left[b_i b_j \right]$ . Now, from explicit computation, $E\left[b_i b_j \right] = \sigma^2 \left( \delta_{ij} -\frac{1}{n} \right)$ , so $E \left[ \sum_{i,j} a_i a_j b_i b_j \right] = \sum_{i,j} a_i a_j \sigma^2 \left( \delta_{ij} -\frac{1}{n} \right) = \sum_i a_i^2 \sigma^2$ as $\sum_i a_i = 0$ .

— Stefano
fonte

Begin from "The derivation is as follow:" The 7th "=" is wrong.

Because

$\sum_i (x_i - \bar{x})(u_i - \bar{u})$

$= \sum_i (x_i - \bar{x})u_i - \sum_i (x_i - \bar{x}) \bar{u}$

$= \sum_i (x_i - \bar{x})u_i - \bar{u} \sum_i (x_i - \bar{x})$

$= \sum_i (x_i - \bar{x})u_i - \bar{u} (\sum_i{x_i} -n \bar{x})$

$= \sum_i (x_i - \bar{x})u_i - \bar{u} (\sum_i{x_i} -\sum_i{x_i})$

$= \sum_i (x_i - \bar{x})u_i - \bar{u} 0$

$= \sum_i (x_i - \bar{x})u_i$

So after 7th "=" it should be:

$\frac {1} {(\sum_i(x_i-\bar{x})^2)^2}E\left[\left(\sum_i(x_i-\bar{x})u_i\right)^2\right]$

$=\frac {1} {(\sum_i(x_i-\bar{x})^2)^2}E\left(\sum_i(x_i-\bar{x})^2u_i^2 + 2\sum_{i\ne j}(x_i-\bar{x})(x_j-\bar{x})u_iu_j\right)$

= $\frac {1} {(\sum_i(x_i-\bar{x})^2)^2}E\left(\sum_i(x_i-\bar{x})^2u_i^2\right) + 2E\left(\sum_{i\ne j}(x_i-\bar{x})(x_j-\bar{x})u_iu_j\right)$

= $\frac {1} {(\sum_i(x_i-\bar{x})^2)^2}E\left(\sum_i(x_i-\bar{x})^2u_i^2\right)$ , because $u_i$ and $u_j$ are independent and mean 0, so $E(u_iu_j) =0$

= $\frac {1} {(\sum_i(x_i-\bar{x})^2)^2}\left(\sum_i(x_i-\bar{x})^2E(u_i^2)\right)$

$\frac {\sigma^2} {(\sum_i(x_i-\bar{x})^2)^2}$

— user158565
fonte

It might be helpful if you edited your answer to include the correct line.

— mdewey

Your answer is being automatically flagged as low quality because it's very short. Please consider expanding on your answer

— Glen_b -Reinstate Monica