Question about finding slope from my textbook, walking through linear regression through the least squares method (pre-calculus)

33 Views Asked by At

I have question about a specific step describing how to do linear regression with the least squares method. In the textbook, it says that you need to find $\triangle x$ and $\triangle y$ (which are $x - \bar{x}$ and $y - \bar{y}$) then you multiply each one together, and find $\Sigma \triangle x \triangle y$. You then go on to divide this by $ \Sigma (\triangle x)^{2} $

$\frac {\Sigma \triangle x \triangle y}{\Sigma (\triangle x)^{2}} $

Which is the slope of the linear regression line

I'm guessing multiplying $\triangle x$ and $\triangle y$ means you're finding the area of the square the actual coordinates make

but then $(\triangle x)^{2}$ means you're finding the supposed, or standard size of what the squares should be? (based on the x values you have)

then the slope is the ratio between them, giving you the line of best fit

Am I understanding this correctly?

1

There are 1 best solutions below

0
On

Derivation

We wish to find the slope $m$ from the following linear equations. Keep in mind that each equation has its error $\epsilon$

$$ \Delta y_{i} = m\Delta x_{i}+\epsilon_{i} $$

When we do least square regression, we select $m$ that minimize the sum of squared errors:

$$ \begin{aligned} \sum_{i}\epsilon_{i}^{2}&=\sum_{i}\left(\Delta y_{i}-m\Delta x_{i}\right)^{2} \\ &= m^{2}\sum_{i}\left(\Delta x_{i}\right)^{2}-2m\sum_{i}\Delta x_{i}\Delta y_{i}+\sum_{i}\left(\Delta y_{i}\right)^{2} \end{aligned} $$

Notice that the RHS is a quadratic polynomial in $m$ with positive first coefficient. The $m$ that minimize the quadratic equation is given by:

$$ m=\frac{\sum_{i}\Delta x_{i}\Delta y_{i}}{\sum_{i}\left(\Delta x_{i}\right)^{2}} $$

Some Intuition

Multiply the initial linear equation by $\Delta x_{i}$ and sum for all values of $i$:

$$ \sum_{i}\Delta x_{i}\Delta y_{i}=m\sum_{i}\left(\Delta x_{i}\right)^{2}+\sum_{i}\Delta x_{i}\epsilon_{i} $$

When we substitute the $m$ values we obtain

$$ 0=\sum_{i}\Delta x_{i}\epsilon_{i} $$

One interpretation is: the least square regression choose the slope such that the errors $\epsilon$ are uncorrelated with the regressors $\Delta x$.