# Standard errors for multiple regression coefficients?

I realize that this is a very basic question, but I can't find an answer anywhere.

I'm computing regression coefficients using either the normal equations or QR decomposition. How can I compute standard errors for each coefficient? I usually think of standard errors as being computed as:

$SE_\bar{x}\ = \frac{\sigma_{\bar x}}{\sqrt{n}}$

What is $\sigma_{\bar x}$ for each coefficient? What is the most efficient way to compute this in the context of OLS?

When doing least squares estimation (assuming a normal random component) the regression parameter estimates are normally distributed with mean equal to the true regression parameter and covariance matrix $\Sigma = s^2\cdot(X^TX)^{-1}$ where $s^2$ is the residual variance and $X^TX$ is the design matrix. $X^T$ is the transpose of $X$ and $X$ is defined by the model equation $Y=X\beta+\epsilon$ with $\beta$ the regression parameters and $\epsilon$ is the error term. The estimated standard deviation of a beta parameter is gotten by taking the corresponding term in $(X^TX)^{-1}$ multiplying it by the sample estimate of the residual variance and then taking the square root. This is not a very simple calculation but any software package will compute it for you and provide it in the output.

### Example

On page 134 of Draper and Smith (referenced in my comment), they provide the following data for fitting by least squares a model $Y = \beta_0 + \beta_1 X + \varepsilon$ where $\varepsilon \sim N(0, \mathbb{I}\sigma^2)$.

                      X                      Y                    XY
0                     -2                     0
2                      0                     0
2                      2                     4
5                      1                     5
5                      3                    15
9                      1                     9
9                      0                     0
9                      0                     0
9                      1                     9
10                     -1                   -10
---                     --                   ---
Sum                  60                      5                    32
Sum of  Squares     482                     21                   528


Looks like an example where the slope should be close to 0.

$$X^t = \pmatrix{ 1 &1 &1 &1 &1 &1 &1 &1 &1 &1 \\ 0 &2 &2 &5 &5 &9 &9 &9 &9 &10 }.$$

So

$$X^t X = \pmatrix{n &\sum X_i \\ \sum X_i &\sum X_i^2} = \pmatrix{10 &60 \\60 &482}$$

and

\eqalign{ (X^t X)^{-1} &= \pmatrix{ \frac{\sum X_i^2}{n \sum (X_i - \bar{X})^2} &\frac{-\bar{X}}{\sum (X_i-\bar{X})^2} \\ \frac{-\bar{X}}{\sum (X_i-\bar{X})^2} &\frac{1}{\sum (X_i-\bar{X})^2} } \\ &= \pmatrix{\frac{482}{10(122)} &-\frac{6}{122} \\ -\frac{6}{122} &\frac{1}{122}} \\ &= \pmatrix{0.395 &-0.049 \\ -0.049 &0.008} }

where $\bar{X} = \sum X_i/n = 60/10 = 6$.

Estimate for $β = (X^TX)^{-1} X^TY$ = ( b0 ) =(Yb-b1 Xb) b1 Sxy/Sxx

b1 = 1/61 = 0.0163 and b0 = 0.5- 0.0163(6) = 0.402

From $(X^TX)^{-1}$ above Sb1 =Se (0.008) and Sb0=Se(0.395) where Se is the estimated standard deviation for the error term. Se =√2.3085.

Sorry that the equations didn't carry subscripting and superscripting when I cut and pasted them. The table didn't reproduce well either because the spaces got ignored. The first string of 3 numbers correspond to the first values of X Y and XY and the same for the followinf strings of three. After Sum comes the sums for X Y and XY respectively and then the sum of squares for X Y and XY respectively. The 2x2 matrices got messed up too. The values after the brackets should be in brackets underneath the numbers to the left.

• Not meant as a plug for my book but i go through the computations of the least squares solution in simple linear regression (Y=aX+b) and calculate the standard errors for a and b, pp.101-103, The Essentials of Biostatistics for Physicians, Nurses, and Clinicians, Wiley 2011. a more detailed description can be found In Draper and Smith Applied Regression Analysis 3rd Edition, Wiley New York 1998 page 126-127. In my answer that follows I will take an example from Draper and Smith. – Michael R. Chernick May 7 '12 at 15:53
• When I started interacting with this site, Michael, I had similar feelings. With experience, they have changed. It's worthwhile knowing some $\TeX$ and once you do, it's (almost) as fast to type it in as it is to type in anything in English. I also learned, by studying exemplary posts (such as many replies by @chl, cardinal, and other high-reputation-per-post users), that providing references, clear illustrations, and well-thought out equations is usually highly appreciated and well received. High quality is one thing distinguishing this site from most others. – whuber May 7 '12 at 21:19
• That is all nice Bill and it is nice that so many people are dedicated to give those high quality posts. I may use Latex for other purposes, like publishing papers. But I don't have the time to go to all the effort that people expect of me on this site. i am not going to invest the time just to provide service on this site. – Michael R. Chernick May 7 '12 at 21:42
• I think the disconnect is here: "This is just one of many things about this site that requires those posting to put in extra time and effort" - @whuber and I are both saying that it, in fact, does not take extra time if you know how to do it. We don't learn $\TeX$ so that we can post on this site - we (at least I) learn $\TeX$ because it's an important skill to have as a statistician and happens to make posts much more readable on this site. – Macro May 8 '12 at 11:05
• Like many of the people on here, yes, I work as a statistician, but I also happen to find it fun - this site is recreational for me and it's a nice bonus that others find some of my posts useful. If you find marking up your equations with $\TeX$ to be work and don't think it's worth learning then so be it, but know that some of your content will be overlooked. – Macro May 8 '12 at 11:16