Multiple Regression

2024

Multiple regression

Multiple regression works just like the simple linear regression we just covered, but with multiple predictor variables.

Regression model

For the \(i^{th}\) observation, we can write:

\[ \begin{align} y_{i} &= \beta_{0} + \beta_{1} x_{1_i} + \beta_{2} x_{2_i} + \dots + \beta_{k} x_{k_i} + \epsilon_{i} \\\\ \end{align} \]

\(y_{i}\) is the \(i^{th}\) observed outcome.
\(x_{k_{i}}\) is the \(i^{th}\) value of the \(k^{th}\) predictor variable.
\(\epsilon_{i}\) is called the residual and is the difference between the observed outcome and the predicted outcome.

\[ \epsilon_{i} \sim Normal(0, \sigma_{\epsilon}) \]

\(\beta_{0_{i}}\), \(\beta_{1_{i}}\), and \(\beta_{2_{i}}\) are parameters of the linear regression model.

Now lets extend to many observations:

\[ \begin{align} y_{1} &= \beta_{0} + \beta_{1} x_{1_1} + \beta_{2} x_{2_1} + \dots + \beta_{k} x_{k_1} + \epsilon_{1} \\ y_{2} &= \beta_{0} + \beta_{1} x_{1_2} + \beta_{2} x_{2_2} + \dots + \beta_{k} x_{k_2} + \epsilon_{2} \\ &\vdots \\ y_{n} &= \beta_{0} + \beta_{1} x_{1_n} + \beta_{2} x_{2_n} + \dots + \beta_{k} x_{k_n} + \epsilon_{n} \\ \end{align} \]

We can gather the independent observations up into vectors:

\[ \begin{align} \boldsymbol{y} &= \beta_{0} \begin{bmatrix} 1\\ 1\\ \vdots\\ 1 \end{bmatrix} + \beta_{1} \begin{bmatrix} x_{1_1}\\ x_{1_2}\\ \vdots\\ x_{1_n}\\ \end{bmatrix} + \beta_{2} \begin{bmatrix} x_{2_1}\\ x_{2_2}\\ \vdots\\ x_{2_n}\\ \end{bmatrix} + \ldots + \beta_{k} \begin{bmatrix} x_{k_1}\\ x_{k_2}\\ \vdots\\ x_{k_n}\\ \end{bmatrix} + \begin{bmatrix} \epsilon_1\\ \epsilon_2\\ \vdots\\ \epsilon_n\\ \end{bmatrix} \\\\ \end{align} \]

We can next gather the vectors up into a matrix:

\[ \begin{align} \boldsymbol{y} &= \begin{bmatrix} 1 & x_{1_1} & x_{2_1} & \ldots & x_{k_1} \\ 1 & x_{1_2} & x_{2_2} & \ldots & x_{k_2} \\ \vdots & \vdots & \dots & \dots \\ 1 & x_{1_n} & x_{2_n} & \ldots & x_{k_n} \end{bmatrix} \begin{bmatrix} \beta_{0} \\ \beta_{1} \\ \vdots \\ \beta_{k} \end{bmatrix} + \begin{bmatrix} \epsilon_1\\ \epsilon_2\\ \vdots\\ \epsilon_n\\ \end{bmatrix} \\\\ \end{align} \]

We can finally write the model in compact matrix form:

\[ \begin{align} \boldsymbol{y} &= \boldsymbol{X} \boldsymbol{\beta} + \boldsymbol{\epsilon} \\ \end{align} \]

\(\boldsymbol{y}\) is a vector of observed outcomes.
\(\boldsymbol{X}\) is a matrix of predictor variables and is called the design matrix
\(\boldsymbol{\beta}\) is a vector of \(\beta\) parameters.
\(\boldsymbol{\epsilon}\) is a vector of residuals.

How can we pick \(\boldsymbol{\beta}\) values that best fit our data?

let \(y_i\) denote observed values
let \(\widehat{y_{i}}\) denote predicted values:

\[ \widehat{y_{i}} = \beta_{0} + \beta_{1} x_{i} + \beta_{2} x_{i} + \ldots + \beta_{k} x_{i} \]

The best fitting \(\boldsymbol{\beta}\) values are those that minimise the discrepancy between \(y_{i}\) and \(\widehat{y_{i}}\).

\[ \DeclareMathOperator*{\argmin}{\arg\!\min} \argmin_{\boldsymbol{\beta}} \sum_{i=1}^{n} (y_{i} - \widehat{y_{i}})^2 \]

\[ \DeclareMathOperator*{\argmin}{\arg\!\min} \argmin_{\boldsymbol{\beta}} \sum_{i=1}^{n} (y_{i} - (\beta_{0} + \beta_{1} x_{i} + \beta_{2} x_{i} + \ldots + \beta_{k} x_{i}))^2 \]

The \(\boldsymbol{\beta}\) values that minimise error can be solved for analytically.
The method is to take the derivative with respect to \(\boldsymbol{\beta}\), and then find the \(\boldsymbol{\beta}\) values that make the resulting expression equal to zero.
I won’t do this here and won’t require you to do so either.
You should know however that this method of finding \(\beta\) values is called ordinary least squares.

Regression model terms

\(SS_{error} = SS_{residual} = \sum_{i=1}^{n} (y_{i} - \hat{y_{i}})^2\)
\(SS_{error}\) is what you get when you compare raw observations against the full model predictions.
\(SS_{total} = \sum_{i=1}^{n} (y_{i} - \bar{y_{i}})^2\)
\(SS_{total}\) is what you get when you compare raw observations against the grand mean.

\(SS_{error}\) comes from \(\sum_{i=1}^{n} (y_{i} - \hat{y_{i}})^2\) with \(\hat{y} = \beta_{0} + \beta_{1} x_{i} + \beta_{2} x_{i} + \ldots + \beta_{k} x_{i} + \epsilon\),
\(SS_{total}\) comes from \(\sum_{i=1}^{n} (y_{i} - \hat{y_{i}})^2\) with \(\hat{y} = \bar{y} + \epsilon\)
\(SS_{model} = \sum_{i=1}^{n} (\bar{y} - \hat{y_i})^2\) tells you how much the added complexity of the full model reduces the overall variability (i.e., makes better predictions).

The percent of variability accounted for above the simple model is given by:

\[R^2 = \frac{SS_{model}}{SS_{total}}\]

\(R^2\) is called the coefficient of determination and is just the sqaure of the correlation coefficient between \(x\) and \(y\).
The \(F\) ratio tells us whether or not the more complex model provides a significantly better fit to the data than the simplest model

\[ F = \frac{MS_{model}}{MS_{error}} \]

The regression \(F\)-ratio tells us how much the regression model has improved the prediction over the simple mean model relative to the overall inaccuracy in the regression model.

We can also ask questions about the best fitting \(\beta\) values (i.e., is either \(\beta\) significantly different from zero?).
The data you use in a regression comes from random variables and so the \(\beta\) values you estimate are also random.
It turns out the best fitting \(\beta\) values (i.e., \(\hat{\beta}\)) can be tested with a \(t\)-test.