A first linear model¶
If we suppose two random variables $X$ and $Y$ depend on each other, perhaps the simplest set-up is that $$ Y = a X + b + \epsilon, $$ where
- $a$ is the slope, i.e., how much $Y$ goes up by, on average, per unit of increase in $X$,
- $b$ is the intercept, i.e., the expected value of $Y$ when $X=0$ (if this makes sense),
- $\hat Y = a X + b$ is the predicted value of $Y$ given $X$,
- $\epsilon = Y - \hat Y$ is the residual, which for the above to be true must have mean zero.
Prediction¶
Suppose $X$ and $Y$ are two, related, measurements, which we treat as random variables having some joint distribution. Question: If we know the value of $X$, how are we to best$^*$ predict the value of $Y$?
$^*$ where how about "best" means "with smallest mean squared error" (i.e., smallest MSE)
So: we'd like to define an estimator of $Y$, which we'll call $\hat Y$, and will be a function of $X$ (and only $X$), and we'd like that estimator to minimize $$ \E[(Y - \hat Y)^2] \qquad \text{(the MSE)}.$$
First observation¶
If we're allowed to add a constant to our estimator, then the mean of the optimal estimator, $\hat Y$, must match that of $Y$: $$ \E[\hat Y] = \E[Y] . $$
Note: for a counterexample about the "if we're allowed to add a constant" note, suppose we're looking for an $a$ such that $\hat Y = \exp(a X)$; then $\hat Y + m$ is not allowed, as it doesn't take the same form.
Proof: Suppose that $\tilde Y$ is an estimator; consider the MSE for $\tilde Y + m$, where $m$ is a number, which we define to be $f(m)$: $$\begin{aligned} f(m) &= \E[(Y - (\tilde Y + m))^2] \ &= \E[(Y - \tilde Y)^2]
- 2 m \E[Y - \tilde Y]
- m^2 . \end{aligned}$$ We want to find the value of $m$ that minimizes $f(m)$, so differentiate with respect to $m$: $$\begin{aligned} f'(m) &=
- 2\E[Y - \tilde Y]
- 2 m . \end{aligned}$$ Setting this equal to zero, we find that $f'(m) = 0$ if $$ m = \E[Y - \tilde Y] = \E[Y] - \E[\tilde Y]. $$
In conclusion, we've found out that if $\tilde Y$ is an estimator, then $\hat Y = \tilde Y + \E[Y] - \E[\tilde Y]$ is a better estimator. And, notice that (by additivity of means), $\E[\hat Y] = \E[Y]$.
So, any estimator of $Y$ can be improved (in the mean-squared-error sense) by shifting it so that the mean of the estimator is equal to the mean of $Y$.
Second observation¶
If the estimator $\hat Y$ is linear, i.e., $$ \hat Y = a X + b $$ for some $a$ and $b$, then $$ a = \frac{\sd[Y]}{\sd[X]} \cor[X, Y] , $$ and $b$ is chosen so that $\E[\hat Y] = \E[Y]$: $$ b = \E[Y] - a \E[X] .$$
In words, the slope, $a$, is equal to the correlation between $X$ and $Y$, but in units of standard deviations.
Proof: We'll do just the same thing as above. First, thanks to the first observation, we can assume $\E[X] = \E[Y] = 0$, which implies that $b = 0$, and $\E[X^2] = \var[X]$ and $\E[Y^2] = \var[Y]$ and $\E[XY] = \cov[X,Y]$. Now, $$\begin{aligned} \text{(MSE)} &= \E[(Y - \hat Y)^2 ] \\ &= \E[(Y - a X)^2 ] \\ &= \E[Y^2 - 2 a X Y + a^2 X^2 ] \\ &= \var[Y] - 2 a \cov[X, Y] + a^2 \var[X] . \end{aligned}$$ If we differentiate this with respect to $a$ and set it equal to zero, we find that the MSE is minimized at $$ a = \frac{\cov[X,Y]}{\var[X]} . $$ If we now substitute in for $\cov[X,Y] = \sd[X]\sd[Y]\cor[X,Y]$, we get the form of $a$ above.
The Normal distribution and least squares¶
How good the estimation procedure above works depends on the situation, of course. Here's one concrete situation - i.e., model - that we might expect to (a) describe a lot of real-world situations pretty well, and (b) be well-suited to fitting with a linear model by least squares: $$\begin{aligned} Y &\sim \text{Normal}(\text{mean}=a X + b, \text{sd}=\sigma) . \end{aligned}$$
Another way of writing the same thing is $$\begin{aligned} Y &= a X + b + \epsilon \\ \epsilon &\sim \text{Normal}(\text{mean}=0, \text{sd}=\sigma) . \end{aligned}$$
Now suppose we have $n$ data points: these will be pairs $(x_1, y_1), \ldots, (x_n, y_n)$. We'd then like to fit the model above -- i.e., find $a$, $b$, and $\sigma$ -- by maximum likelihood.
To do this, we need the likelihood function: $$ \prod_{i=1}^n \frac{1}{\sqrt{2 \pi \sigma^2}} e^{-\frac{(y_i - a x_i - b)^2}{2 \sigma^2}} . $$ The term in the exponential is known as the residual, since it's the difference between the observed value ($y_i$) and the value predicted using $x$ under the model ($a x_i + b$): $$ y_i - a x_i - b = e_i . $$ It is our estimate of $\epsilon$ in the model we are trying to fit.
Now, the logarithm of the likelihood function (i.e., the log-likelihood) is $$
\ell(a, b, \sigma)
=
- \frac{1}{2 \sigma^2} \sum_{i=1}^n e_i^2
- \frac{n}{2} \log(2 \pi \sigma^2) .
$$ Notice that the only place that $a$ and $b$ enter in to this is via the errors, $e_i$.
Therefore, to find the values of $a$ and $b$ that maximize $\ell(a, b, \sigma)$, regardless of $\sigma$, we need to minimize $\sum_i e_i^2$, the sum of squared errors.
In other words, the maximum likelihood estimates of $a$ and $b$ are the same as the the values of $a$ and $b$ that minimize the MSE!
(You can work out the MLE for $\sigma$ yourself.)
Multivariate linear models¶
Now suppose that we have a bunch of variables we want to use to predict $Y$. Call these $X = (X_1, \ldots, X_k)$. To do this, let's define coefficients $a = (a_1, \ldots, a_k)$, and using them produce the estimate $$ \hat Y = a_1 X_1 + \cdots a_k X_k , $$ and as before find $a$ to minimize the MSE, $$ \E[(Y - \hat Y)^2] . $$
Here we haven't explicitly included an intercept (before, $b$) because we can just assume that one of the $X_i$'s is always equal to 1 (and if not, stick on a dummy variable that is).
Notice that the estimate is just the inner product of $a$ and $X$, i.e., $\hat Y = a^T X$. Substituting this in to the MSE: $$\begin{aligned} \E[(Y - \hat Y)^2] &= \E[(Y - a^T X)^2] \\ &= \E[Y^2 - 2 Y a^T X + (a^T X)^2] . \end{aligned}$$ Now, we want to find the value of $a$ that makes the derivative of this equal to zero. Unlike before, $a$ is a $k$-vector, not a single number, but multivariable calculus tells us that (1) $\frac{d}{da} a^T x = x$, and (b) $\frac{d}{da} (a^T x)^2 = 2 x x^T a$. So, $$\begin{aligned} \frac{d}{da} \E[(Y - \hat Y)^2] &=
- 2 \E[XY] + 2 \E[X X^T] a .
\end{aligned}$$ Note that $\E[XY]$ is a *vector*, whose $i^\text{th}$ component is $\E[X_i Y]$, and $\E[X X^T]$ is a *matrix*, whose $(i,j)^\text{th}$ component is $\E[X_i X_j]$. Think of these as the vector of covariances of $X$ with $Y$ and the covariance matrix of $X$, respectively. The value of $a$ that makes this zero is $$ a = \E[X X^T]^{-1} \E[X Y] , $$ where $X^T$ is the transpose of $X$ and $\E[X X^T]^{-1}$ is the inverse of the matrix $\E[X X^T]$, if it exists.
What have we learned? We've found that the linear estimator of $Y$ that minimizes the mean-squared error can be found through some simple linear algebra using the covariance matrix of $X$ and the vector of correlations between $X$ and $Y$.
Multivariate linear models, take 2¶
Let's look at that again, but from the point of view of data. Now suppose we have a vector of reponses $y = (y_1, \ldots, y_n)$ and a matrix of predictors $X = (x_1, \ldots, x_n)$, where $x_i = (x_{i1}, \ldots, x_{ik})$ is the $i^\text{th}$ row of $X$. Think of $X$ and $y$ as training data, and we want to know how we can best predict (future) $y$ using the corresponding (future) values of $x$. One way to answer this is to ask: what are the coefficients $a = (a_1, \ldots, a_k)$ for which $$ \hat y_i = a_1 x_{i1} + \cdots a_k x_{ik} $$ best predicts $y_i$ in our training data?
Mean squared error, again¶
One way to answer this question is by finding the $a$ that minimizes $$ M(a) = \sum_{i=1}^n (y_i - \hat y_i)^2 $$ ... which, divided by $n$, is the mean squared error.
Writing this out in vector notation, this is $M(a) = (y - \hat y)^T (y - \hat y)$, and since $\hat y = X a$, $$\begin{aligned} M(a) &= y^T y - 2 a^T X^T y + a^T X^T X a . \end{aligned}$$ So, $$\begin{aligned} \frac{d}{da} M(a) &=
- 2 X^T y + 2 X^T X a ,
\end{aligned}$$ and the value of $a$ that makes this zero is $$ a = (X^T X)^{-1} X^T y . $$ Note the similarity to the previous theory.
Example¶
(switch to the elephant example)
With scikit-learn¶
Let's see how to do the same thing using a package, scikit-learn. First, we'll simulate up some simple data:
n = 200
k = 2
x = rng.uniform(size=(n, k))
beta = [10, -2]
y = x.dot(beta) + rng.normal(size=n)
fig, (ax1, ax2) = plt.subplots(1, 2)
ax1.scatter(x[:,0], y); ax1.set_xlabel(f"x0"); ax1.set_ylabel("y")
ax2.scatter(x[:,1], y); ax2.set_xlabel(f"x1"); ax2.set_ylabel("y");
from sklearn import linear_model
line_reg = linear_model.LinearRegression(fit_intercept=True)
line_reg.fit(x, y)
line_reg.coef_
array([ 9.54442251, -1.79347923])
y_hat = line_reg.predict(x)
fig, (ax1, ax2) = plt.subplots(1, 2)
ax1.scatter(x[:,0], y); ax1.set_xlabel(f"x0"); ax1.set_ylabel("y")
ax2.scatter(x[:,1], y); ax2.set_xlabel(f"x1"); ax2.set_ylabel("y")
ax1.scatter(x[:,0], y_hat); ax2.scatter(x[:,1], y_hat);
