Residuals and Total Squared Error

Chapter 11: Regression Analysis: Simple Linear Regression

Residuals and Total Squared Error

A regression line is the best-fitting straight line through a set of data points. Regression analysis is all about predicting values, and what makes a regression line 'best-fitting' is that it has the lowest possible amount of prediction error.

In the context of regression, the amount of prediction error is expressed in terms of residuals.
$\phantom{0}$

Residual

A residual is the vertical distance between the regression line and a data point and is denoted by $r$ .

Calculating Residuals

To calculate a residual, take a point $(X,Y)$ from the data and determine the height of the regression line at point $X$ . This point is the predicted value of $Y$ and is denoted by $\hat{Y}$ .

Next, subtract the predicted value $\hat{Y}$ from the observed value $Y$ to determine the value of the residual:

$r_i = Y_i - \hat{Y}_i$

Calculation of Residuals

Consider the regression equation $\hat{Y}=2X$ and the data points $(1,3)$ , $(3,1)$ , and $(4,3)$ . The residuals of these three data points are calculated as follows:

For the first point :
- $\purple{\hat{Y}_1}=2\cdot 1=2$
- $\blue{Y_1} = 3$
- $\orange{r_1}= Y_1-\hat{Y}_1=3-2=1$ .
For the second point :
- $\purple{\hat{Y}_2}=2\cdot 3=6$
- $\blue{Y_2}=1$
- $\orange{r_2}= Y_2-\hat{Y}_2 = 1-6 =-5$ .
For the last point :
- $\purple{\hat{Y}_3} =2\cdot 4=8$
- $\blue{Y_3}=3$
- $\orange{r_3}= Y_3-\hat{Y}_3 = 3-8=-5$ .

$\phantom{0}$

$\phantom{0}$
One of the most commonly used measures to summarize the total amount of prediction error is the Total Squared Error.
$\phantom{0}$

Total Squared Error

The Total Squared Error is the sum of the squared residuals and is often abbreviated TSE.

$\text{TSE} = \sum{r^2} = \sum{(Y-\hat{Y})^2}$

The reason for squaring the residuals before adding them together is to prevent positive and negative residuals from canceling one another. Consequently, the total squared error will always be a positive number.

Calculation of Total Squared Error

Consider the regression line and residuals from the previous example. In this case, the Total Squared Error is:

$\begin{array}{rcl} \text{TSE} &=& \sum{(Y-\hat{Y})^2}\\ &=& (Y_1-\hat{Y}_1)^2 + (Y_2-\hat{Y}_2)^2 + (Y_3-\hat{Y}_3)^2\\ &=& (3-2)^2+(1-6)^2+(3-8)^2\\ &=& 1^2 + (-5)^2 + (-5)^2\\ &=& 1 + 25 + 25\\ &=& 51 \end{array}$