Published on December 21, 2020

Deriving the OLS Estimator

janedoe@janedoe10

johndoe@johndoe10

Introduction

Parsing and display of math equations is included in this blog template. Parsing of math is enabled by remark-math and rehype-katex. KaTeX and its associated font is included in _document.js so feel free to use it on any page. ¹

Inline math symbols can be included by enclosing the term between the $ symbol.

Math code blocks are denoted by $$.

If you intend to use the $ sign instead of math, you can escape it (\$), or specify the HTML entity ($) ²

Inline or manually enumerated footnotes are also supported. Click on the links above to see them in action.

Deriving the OLS Estimator

Using matrix notation, let $n$ denote the number of observations and $k$ denote the number of regressors.

The vector of outcome variables $\mathbf{Y}$ is a $n \times 1$ matrix,

tex$$
\mathbf{Y} = \left[\begin{array}
  {c}
  y_1 \\
  . \\
  . \\
  . \\
  y_n
\end{array}\right]
$$

The matrix of regressors $\mathbf{X}$ is a $n \times k$ matrix (or each row is a $k \times 1$ vector),

latex\mathbf{X} = \left[\begin{array}
  {ccccc}
  x_{11} & . & . & . & x_{1k} \\
  . & . & . & . & .  \\
  . & . & . & . & .  \\
  . & . & . & . & .  \\
  x_{n1} & . & . & . & x_{nn}
\end{array}\right] =
\left[\begin{array}
  {c}
  \mathbf{x}'_1 \\
  . \\
  . \\
  . \\
  \mathbf{x}'_n
\end{array}\right]

The vector of error terms $\mathbf{U}$ is also a $n \times 1$ matrix.

At times it might be easier to use vector notation. For consistency, I will use the bold small x to denote a vector and capital letters to denote a matrix. Single observations are denoted by the subscript.

Least Squares

Start: $y_i = \mathbf{x}'_i \beta + u_i$

Assumptions:

1. Linearity (given above)
2. $E(\mathbf{U}|\mathbf{X}) = 0$ (conditional independence)
3. rank($\mathbf{X}$) = $k$ (no multi-collinearity i.e. full rank)
4. $Var(\mathbf{U}|\mathbf{X}) = \sigma^2 I_n$ (Homoskedascity)

Aim: Find $\beta$ that minimises the sum of squared errors:

Solution: Hints: $Q$ is a $1 \times 1$ scalar, by symmetry $\frac{\partial b'Ab}{\partial b} = 2Ab$.

Take matrix derivative w.r.t $\beta$:

tex\begin{aligned}
  \min Q           & = \min_{\beta} \mathbf{Y}'\mathbf{Y} - 2\beta'\mathbf{X}'\mathbf{Y} +
  \beta'\mathbf{X}'\mathbf{X}\beta \\
                   & = \min_{\beta} - 2\beta'\mathbf{X}'\mathbf{Y} + \beta'\mathbf{X}'\mathbf{X}\beta \\
  \text{[FOC]}~~~0 & =  - 2\mathbf{X}'\mathbf{Y} + 2\mathbf{X}'\mathbf{X}\hat{\beta}                  \\
  \hat{\beta}      & = (\mathbf{X}'\mathbf{X})^{-1}\mathbf{X}'\mathbf{Y}                              \\
                   & = (\sum^{n} \mathbf{x}_i \mathbf{x}'_i)^{-1} \sum^{n} \mathbf{x}_i y_i
\end{aligned}

Footnotes