Maximum likelihood estimation

Announcements

• HW 03 due March 20 at 11:59pm

• Project exploratory data analysis due March 20 at 11:59pm

  • Next milestone: Project presentations in lab March 28

• Statistics experience due April 22

Topics

• Likelihood

• Maximum likelihood estimation (MLE)

• MLE for linear regression

Motivation

• We’ve discussed how to find the estimators of $\mathit{\beta }$ and ${\sigma }_{ϵ}^2$ for the model

$$\mathbf{y}=\mathbf{X}\mathit{\beta }+\mathit{ϵ},\mathit{ϵ}\sim N(0,{\sigma }_{ϵ}^2\mathbf{I})$$using least-squares estimation

• Today we will introduce another way to find these estimators - maximum likelihood estimation.

• We will see the least-squares estimator is equal to the maximum likelihood estimator when certain assumptions hold

Maximum likelihood estimation

Example: Basketball shots

Suppose the a basketball player shoots the ball, such that the probability of making the basket (successfully making the shot) is $p$

. . .

• What is the probability distribution for this random phenomenon?

• Suppose the probability is $p=0.5$. What is the probability the player makes a single basket, given this value of $p$?

• Suppose the probability is $p=0.8$. What is the probability the player makes a single basket, given this value of $p$?

Shooting the ball three times

Suppose the player shoots the ball three times. They are all independent and the player has the same probability $p$ of making each basket.

Let $B$ represent a made basket, and $M$ represent a missed basket. The player shoots the ball three times with the outcome $BBM$.

. . .

• Suppose the probability is $p=0.5$. What is the probability of observing the data $BBM$, given this value of $p$?

• Suppose the probability is $p=0.3$. What is the probability of observing the data $BBM$, given this value of $p$ ?

Shooting the ball three times

Suppose the player shoots the ball three times. They are all independent and the player has the same probability $p$ of making each basket.

The player shoots the ball three times with the outcome $BBM$.

. . .

• New question: What parameter value of $p$ do you think maximizes the probability of observing this data?

• We will use a likelihood function to answer this question.

Likelihood

• A likelihood function is a measure of how likely we are to observe our data under each possible value of the parameter(s)

• Note that this is not the same as the probability function.

• Probability function: Fixed parameter value(s) + input possible outcomes

  • Given $p=0.8$ , what is the probability of observing $BBM$ in three basketball shots?

• Likelihood function: Fixed data + input possible parameter values

  • Given we’ve observed $BBM$, what is the most plausible value of $p$?

Likelihood: Three basketball shots

The likelihood function for the probability of a basket $p$ given we observed $BBM$ when shooting the ball three independent times $$L(p|BBM)=p\times p\times (1-p)$$

. . .

Thus, if the likelihood for $p=0.8$ is

$$L(p=0.8|BBM)=0.8\times 0.8\times (1-0.8)=0.128$$

Likelihood: Three basketball shots

• What is the general formula for the likelihood function for $p$ given the observed data $BBM$?

• How does assuming independence simplify things?

• How does having identically distributed data simplify things?

Likelihood: Three basketball shots

The likelihood function for $p$ given the data $BBM$ is

$$L(p|BBM)=p\times p\times (1-p)=p^2\times (1-p)$$

. . .

• We want of the value of $p$ that maximizes this likelihood function, i.e., the value of $p$ that is most likely given the observed data.

• The process of finding this value is maximum likelihood estimation.

• There are three primary ways to find the maximum likelihood estimator

  • Approximate using a graph

  • Using calculus

  • Numerical approximation

Finding the MLE using graphs

What do you think is the approximate value of the MLE of $p$ given the data?

Finding the MLE using calculus

• Find the MLE using the first derivative of the likelihood function.

• This can be tricky because of the product rule, so we can maximize the log(Likelihood) instead. The same value maximizes the likelihood and log(Likelihood).

Use calculus to find the MLE of $p$ given the data $BBM$.

Shooting the ball $n$ times

Suppose the player shoots the ball $n$ times. They are all independent and the player has the same probability $p$ of making each one.

Suppose the player makes $k$ baskets out of the $n$ shots. This is the observed data.

. . .

• What is the formula for the probability distribution to describe this random phenomenon?
• What is the formula for the likelihood function for $p$ given the observed data?
• For what value of $p$ do we maximize the likelihood given the observed data? Use calculus to find the response.

MLE in linear regression

Why maximum likelihood estimation?

• “Maximum likelihood estimation is, by far, the most popular technique for deriving estimators.” (Casella and Berger 2024, 315)

• MLEs have nice statistical properties (more on this next class)

  • Consistent

  • Efficient

  • Asymptotically normal

. . .

Note

If the normality assumption holds, the least squares estimator is the maximum likelihood estimator for $\beta$. Therefore, it has all the properties of the MLE.

Linear regression

Recall the linear model

$$\mathbf{y}=\mathbf{X}\mathit{\beta }+\mathit{ϵ},\mathit{ϵ}\sim N(\mathbf{0},{\sigma }_{ϵ}^2\mathbf{I})$$

. . .

• We have discussed least-squares estimation to find $\hat{\mathit{\beta }}$ and ${\hat{\sigma }}_{ϵ}^2$
• We have used the fact that $\hat{\mathit{\beta }}\sim N(\mathit{\beta },{\sigma }_{ϵ}^2({\mathbf{X}}^{\mathsf{T}}\mathbf{X}{)}^{-1})$ when doing hypothesis testing and confidence intervals.
• Now we will discuss how we know $\hat{\mathit{\beta }}$ is normally distributed, as we introduce MLE for linear regression

Simple linear regression model

Suppose we have the simple linear regression (SLR) model

$$y_i={\beta }_0+{\beta }_1{x}_i+{ϵ}_i,{ϵ}_i\sim N(0,{\sigma }_{ϵ}^2)$$

such that ${ϵ}_i$ are independently and identically distributed.

. . .

We can write this model in the form below and use this to find the MLE

$$y_i|x_i\sim N({\beta }_0+{\beta }_1{x}_i,{\sigma }_{ϵ}^2)$$

Side note: Normal distribution

Let $Z$ be a random variable, such that $Z\sim N(\mu ,{\sigma }^2)$. Then the probability function is

$$P(Z=z|\mu ,{\sigma }^2)=\frac{1}{\sqrt{2\pi {\sigma }^2}}\exp \{-\frac{1}{2{\sigma }^2}(z-\mu {)}^2\}$$

SLR: Likelihood for ${\beta }_0,{\beta }_1,{\sigma }_{ϵ}^2$

The likelihood function for ${\beta }_0,{\beta }_1,{\sigma }_{ϵ}^2$ is

$$\begin{array}{rl}L({\beta }_0,{\beta }_1,{\sigma }_{ϵ}^2& |x_1,\dots ,x_n,y_1,\dots ,y_n)\\ & =p(y_1|x_1,{\beta }_0,{\beta }_1,{\sigma }_{ϵ}^2)\dots p(y_n|x_n,{\beta }_0,{\beta }_1,{\sigma }_{ϵ}^2)\\ & =\prod _{i=1}^{n}p(y_i|x_i,{\beta }_0,{\beta }_1,{\sigma }_{ϵ}^2)\\ & =\prod _{i=1}^n\frac{1}{\sqrt{2\pi {\sigma }_{ϵ}^2}}\exp \{-\frac{1}{2{\sigma }_{ϵ}^2}(y_i-[{\beta }_0+{\beta }_1{x}_i]{)}^2\}\\ & =(2\pi {\sigma }_{ϵ}^2{)}^{-\frac{n}{2}}\exp \{-\frac{1}{2{\sigma }_{ϵ}^2}\sum _{i=1}^n(y_i-{\beta }_0-{\beta }_1{x}_i{)}^2\}\end{array}$$

Log-Likelihood for ${\beta }_0,{\beta }_1,{\sigma }_{ϵ}^2$

The log-likelihood function for ${\beta }_0,{\beta }_1,{\sigma }_{ϵ}^2$ is

$$\begin{array}{rl}\log & L({\beta }_0,{\beta }_1,{\sigma }_{ϵ}^2|x_1,\dots ,x_n,y_1,\dots ,y_n)\\ & =\log ((2\pi {\sigma }_{ϵ}^2{)}^{-\frac{n}{2}}\exp \{-\frac{1}{2{\sigma }_{ϵ}^2}\sum _{i=1}^n(y_i-{\beta }_0-{\beta }_1{x}_i{)}^2\})\\ & =-\frac{n}{2}\log (2\pi {\sigma }_{ϵ}^2)-\frac{1}{2{\sigma }_{ϵ}^2}\sum _{i=1}^n(y_i-{\beta }_0-{\beta }_1{x}_i{)}^2\end{array}$$

MLE for ${\beta }_0$

1️⃣ Take derivative of $\log L$ with respect to ${\beta }_0$ and set it equal to 0

$$\frac{\partial \log L}{\partial {\beta }_0}=-\frac{2}{2{\sigma }_{ϵ}^2}\sum _{i=1}^n(y_i-{\beta }_0-{\beta }_1{x}_i)(-1)=0$$

MLE for ${\beta }_0$

2️⃣ Find the ${\tilde{\beta }}_0$ that satisfies the equality on the previous slide

. . .

After a few steps…

$$\begin{array}{rl}& \Rightarrow \sum _{i=1}^n{y}_i-n{\tilde{\beta }}_0-{\tilde{\beta }}_1\sum _{i=1}^n{x}_i=0\\ & \Rightarrow \sum _{i=1}^n{y}_i-{\tilde{\beta }}_1\sum _{i=1}^n{x}_i=n{\tilde{\beta }}_0\\ & \Rightarrow \frac{1}{n}\sum _{i=1}^n{y}_i-\frac{1}{n}{\tilde{\beta }}_1\sum _{i=1}^n{x}_i={\tilde{\beta }}_0\end{array}$$

MLE for ${\beta }_0$

3️⃣ We can use the second derivative to show we’ve found the maximum

$$\frac{{\partial }^2\log L}{\partial {\beta }_0^2}=-\frac{n}{2{\tilde{\sigma }}_{ϵ}^2}<0$$

. . .

Therefore, we have found the maximum. Thus, MLE for ${\beta }_0$ is

$${\tilde{\beta }}_0=\overline{y}-{\tilde{\beta }}_1\overline{x}$$

. . .

Note that ${\tilde{\beta }}_0$ is equal to ${\hat{\beta }}_0$, the least-squares estimate

MLE for ${\beta }_1$ and ${\sigma }_{ϵ}^2$

We can use a similar process to find the MLEs for ${\beta }_1$ and ${\sigma }_{ϵ}^2$

$${\tilde{\beta }}_1=\frac{\sum _{i=1}^n(y_i-\overline{y})(x_i-\overline{x})}{\sum _{i=1}^n(x_i-\overline{x}{)}^2}$$

. . .

$${\tilde{\sigma }}_{ϵ}^2=\frac{\sum _{i=1}^n(y_i-{\tilde{\beta }}_0-{\tilde{\beta }}_1{x}_i{)}^2}{n}=\frac{\sum _{i=1}^n{e}_i^2}{n}$$

. . .

Note: ${\tilde{\beta }}_1={\hat{\beta }}_1$ and ${\tilde{\sigma }}_{ϵ}^2\approx {\hat{\sigma }}_{ϵ}^2$

MLE in matrix form

MLE for linear regression in matrix form

$$L(\mathit{\beta },{\sigma }_{ϵ}^2|\mathbf{X},\mathbf{y})=\frac{1}{(2\pi {)}^{n/2}{\sigma }_{ϵ}^n}\exp \{-\frac{1}{2{\sigma }_{ϵ}^2}(\mathbf{y}-\mathbf{X}\mathit{\beta }{)}^{\mathsf{T}}(\mathbf{y}-\mathbf{X}\mathit{\beta })\}$$

. . .

$$\begin{array}{rl}\log L(\mathit{\beta },{\sigma }_{ϵ}^2& |\mathbf{X},\mathbf{y})\\ & =-\frac{n}{2}\log (2\pi )-n\log ({\sigma }_{ϵ})-\frac{1}{2{\sigma }_{ϵ}^2}(\mathbf{y}-\mathbf{X}\mathit{\beta }{)}^{\mathsf{T}}(\mathbf{y}-\mathbf{X}\beta )\end{array}$$

. . .

1. For a fixed value of ${\sigma }_{ϵ}$ , we know that $\log L$ is maximized when what is true about $(\mathbf{y}-\mathbf{X}\mathit{\beta }{)}^{\mathsf{T}}(\mathbf{y}-\mathbf{X}\mathit{\beta })$ ?
2. What does this tell us about the relationship between the MLE and least-squares estimator for $\mathit{\beta }$?

Putting it all together

• The MLE $\tilde{\mathit{\beta }}$ is equivalent to the least-squares estimator $\hat{\mathit{\beta }}$ , when the errors follow independent and identical normal distributions

• MLEs have nice properties, so this means the least-squares estimator $\hat{\mathit{\beta }}$ inherits all the nice properties of MLEs

• The MLE ${\tilde{\sigma }}_{ϵ}^2$ is approximately equal to the least-squares estimator ${\hat{\sigma }}_{ϵ}^2$. When $n>>p$, the difference is trivial

References

Casella, George, and Roger Berger. 2024. Statistical Inference. CRC Press.