Guiding Questions

• The correlation between two variables tells us how they move together

• Sample correlation is a random variable

Case study: is height hereditary?

We have access to Galton’s family height data through the HistData package. This data contains heights on several dozen families: mothers, fathers, daughters, and sons. To imitate Galton’s analysis, we will create a dataset with the heights of fathers and a randomly selected son of each family:

library(tidyverse)
library(HistData)
data("GaltonFamilies")

set.seed(1983)
galton_heights <- GaltonFamilies %>%
  filter(gender == "male") %>%
  group_by(family) %>%
  sample_n(1) %>%
  ungroup() %>%
  select(father, childHeight) %>%
  rename(son = childHeight)

In the exercises, we will look at other relationships including mothers and daughters.

Suppose we were asked to summarize the father and son data. Since both distributions are well approximated by the normal distribution, we could use the two averages and two standard deviations as summaries:

galton_heights %>%
  summarize(mean(father), sd(father), mean(son), sd(son))

## # A tibble: 1 × 4
##   `mean(father)` `sd(father)` `mean(son)` `sd(son)`
##            <dbl>        <dbl>       <dbl>     <dbl>
## 1           69.1         2.55        69.2      2.71

However, this summary fails to describe an important characteristic of the data: the trend that the taller the father, the taller the son.

galton_heights %>% ggplot(aes(father, son)) +
  geom_point(alpha = 0.5)

We will learn that the correlation coefficient is an informative summary of how two variables move together and then see how this can be used to predict one variable using the other.

The correlation coefficient

The correlation coefficient is defined for a list of pairs $(x_1,y_1),\dots ,(x_n,y_n)$ as the average of the product of the standardized values:

$$\rho =\frac{1}{n}\sum _{i=1}^n\left(\frac{x_i-{\mu }_x}{{\sigma }_x}\right)\left(\frac{y_i-{\mu }_y}{{\sigma }_y}\right)$$ with ${\mu }_x,{\mu }_y$ the averages of $x_1,\dots ,x_n$ and $y_1,\dots ,y_n$, respectively, and ${\sigma }_x,{\sigma }_y$ the standard deviations. The Greek letter $\rho$ is commonly used in statistics books to denote the correlation. The Greek letter for $r$, $\rho$, because it is the first letter of regression. Soon we learn about the connection between correlation and regression. We can represent the formula above with R code using:

rho <- mean(scale(x) * scale(y))

To understand why this equation does in fact summarize how two variables move together, consider the $i$-th entry of $x$ is $\left(\frac{x_i-{\mu }_x}{{\sigma }_x}\right)$ SDs away from the average. Similarly, the $y_i$ that is paired with $x_i$, is $\left(\frac{y_1-{\mu }_y}{{\sigma }_y}\right)$ SDs away from the average $y$. If $x$ and $y$ are unrelated, the product $\left(\frac{x_i-{\mu }_x}{{\sigma }_x}\right)\left(\frac{y_i-{\mu }_y}{{\sigma }_y}\right)$ will be positive ( $+\times +$ and $-\times -$ ) as often as negative ($+\times -$ and $-\times +$) and will average out to about 0. This correlation is the average and therefore unrelated variables will have 0 correlation. If instead the quantities vary together, then we are averaging mostly positive products ( $+\times +$ and $-\times -$) and we get a positive correlation. If they vary in opposite directions, we get a negative correlation.

The correlation coefficient is always between -1 and 1. We can show this mathematically: consider that we can’t have higher correlation than when we compare a list to itself (perfect correlation) and in this case the correlation is:

$$\rho =\frac{1}{n}\sum _{i=1}^n{\left(\frac{x_i-{\mu }_x}{{\sigma }_x}\right)}^2=\frac{1}{{\sigma }_x^2}\frac{1}{n}\sum _{i=1}^n{(x_i-{\mu }_x)}^2=\frac{1}{{\sigma }_x^2}{\sigma }_x^2=1$$

A similar derivation, but with $x$ and its exact opposite, proves the correlation has to be bigger or equal to -1.

For other pairs, the correlation is in between -1 and 1. The correlation between father and son’s heights is about 0.5:

galton_heights %>% summarize(r = cor(father, son)) %>% pull(r)

## [1] 0.4334102

To see what data looks like for different values of $\rho$, here are six examples of pairs with correlations ranging from -0.9 to 0.99:

R <- sample_n(galton_heights, 25, replace = TRUE) %>%
  summarize(r = cor(father, son)) %>% pull(r)

B <- 1000
N <- 25
R <- replicate(B, {
  sample_n(galton_heights, N, replace = TRUE) %>%
    summarize(r=cor(father, son)) %>%
    pull(r)
})
qplot(R, geom = "histogram", binwidth = 0.05, color = I("black"))

## Warning: `qplot()` was deprecated in ggplot2 3.4.0.
## This warning is displayed once every 8 hours.
## Call `lifecycle::last_lifecycle_warnings()` to see where this warning was
## generated.

mean(R)

## [1] 0.4307393

sd(R)

## [1] 0.1609393

ggplot(aes(sample=R), data = data.frame(R)) +
  stat_qq() +
  geom_abline(intercept = mean(R), slope = sqrt((1-mean(R)^2)/(N-2)))

## `geom_smooth()` using formula = 'y ~ x'

Conditional expectations

Suppose we are asked to guess the height of a randomly selected son and we don’t know his father’s height. Because the distribution of sons’ heights is approximately normal, we know the average height, 69.2, is the value with the highest proportion and would be the prediction with the highest chance of minimizing the error. But what if we are told that the father is taller than average, say 72 inches tall, do we still guess 69.2 for the son?

It turns out that if we were able to collect data from a very large number of fathers that are 72 inches, the distribution of their sons’ heights would be normally distributed. This implies that the average of the distribution computed on this subset would be our best prediction.

In general, we call this approach conditioning. The general idea is that we stratify a population into groups and compute summaries in each group. To provide a mathematical description of conditioning, consider we have a population of pairs of values $(x_1,y_1),\dots ,(x_n,y_n)$, for example all father and son heights in England. In the previous week’s content, we learned that if you take a random pair $(X,Y)$, the expected value and best predictor of $Y$ is $\text{E}(Y)={\mu }_y$, the population average $1/n\sum _{i=1}^n{y}_i$. However, we are no longer interested in the general population, instead we are interested in only the subset of a population with a specific $x_i$ value, 72 inches in our example. This subset of the population, is also a population and thus the same principles and properties we have learned apply. The $y_i$ in the subpopulation have a distribution, referred to as the conditional distribution, and this distribution has an expected value referred to as the conditional expectation. In our example, the conditional expectation is the average height of all sons in England with fathers that are 72 inches. The statistical notation for the conditional expectation is

$$\text{E}(Y\mid X=x)$$

with $x$ representing the fixed value that defines that subset, for example 72 inches. Similarly, we denote the standard deviation of the strata with

$$\text{SD}(Y\mid X=x)=\sqrt{\text{Var}(Y\mid X=x)}$$

Because the conditional expectation $E(Y\mid X=x)$ is the best predictor for the random variable $Y$ for an individual in the strata defined by $X=x$, many data science challenges reduce to estimating this quantity. The conditional standard deviation quantifies the precision of the prediction.

In the example we have been considering, we are interested in computing the average son height conditioned on the father being 72 inches tall. We want to estimate $E(Y|X=72)$ using the sample collected by Galton. We previously learned that the sample average is the preferred approach to estimating the population average. However, a challenge when using this approach to estimating conditional expectations is that for continuous data we don’t have many data points matching exactly one value in our sample. For example, we have only:

sum(galton_heights$father == 72)

## [1] 8

fathers that are exactly 72-inches. If we change the number to 72.5, we get even fewer data points:

sum(galton_heights$father == 72.5)

## [1] 1

A practical way to improve these estimates of the conditional expectations, is to define strata of with similar values of $x$. In our example, we can round father heights to the nearest inch and assume that they are all 72 inches. If we do this, we end up with the following prediction for the son of a father that is 72 inches tall:

conditional_avg <- galton_heights %>%
  filter(round(father) == 72) %>%
  summarize(avg = mean(son)) %>%
  pull(avg)
conditional_avg

## [1] 70.5

Note that a 72-inch father is taller than average – specifically, 72 - 69.1/2.5 = 1.1 standard deviations taller than the average father. Our prediction 70.5 is also taller than average, but only 0.49 standard deviations larger than the average son. The sons of 72-inch fathers have regressed some to the average height. We notice that the reduction in how many SDs taller is about 0.5, which happens to be the correlation. As we will see in a later lecture, this is not a coincidence.

If we want to make a prediction of any height, not just 72, we could apply the same approach to each strata. Stratification followed by boxplots lets us see the distribution of each group:

galton_heights %>% mutate(father_strata = factor(round(father))) %>%
  ggplot(aes(father_strata, son)) +
  geom_boxplot() +
  geom_point()

Not surprisingly, the centers of the groups are increasing with height. Furthermore, these centers appear to follow a linear relationship. Below we plot the averages of each group. If we take into account that these averages are random variables with standard errors, the data is consistent with these points following a straight line:

The fact that these conditional averages follow a line is not a coincidence. In the next section, we explain that the line these averages follow is what we call the regression line, which improves the precision of our estimates. However, it is not always appropriate to estimate conditional expectations with the regression line so we also describe Galton’s theoretical justification for using the regression line.

The regression line

If we are predicting a random variable $Y$ knowing the value of another $X=x$ using a regression line, then we predict that for every standard deviation, ${\sigma }_X$, that $x$ increases above the average ${\mu }_X$, $Y$ increase $\rho$ standard deviations ${\sigma }_Y$ above the average ${\mu }_Y$ with $\rho$ the correlation between $X$ and $Y$. The formula for the regression is therefore:

$$\left(\frac{Y-{\mu }_Y}{{\sigma }_Y}\right)=\rho \left(\frac{x-{\mu }_X}{{\sigma }_X}\right)$$

We can rewrite it like this:

$$Y={\mu }_Y+\rho \left(\frac{x-{\mu }_X}{{\sigma }_X}\right){\sigma }_Y$$

If there is perfect correlation, the regression line predicts an increase that is the same number of SDs. If there is 0 correlation, then we don’t use $x$ at all for the prediction and simply predict the average ${\mu }_Y$. For values between 0 and 1, the prediction is somewhere in between. If the correlation is negative, we predict a reduction instead of an increase.

Note that if the correlation is positive and lower than 1, our prediction is closer, in standard units, to the average height than the value used to predict, $x$, is to the average of the $x$s. This is why we call it regression: the son regresses to the average height. In fact, the title of Galton’s paper was: Regression toward mediocrity in hereditary stature. To add regression lines to plots, we will need the above formula in the form:

$$y=b+mx\text{ with slope }m=\rho \frac{{\sigma }_y}{{\sigma }_x}\text{ and intercept }b={\mu }_y-m{\mu }_x$$

Here we add the regression line to the original data:

mu_x <- mean(galton_heights$father)
mu_y <- mean(galton_heights$son)
s_x <- sd(galton_heights$father)
s_y <- sd(galton_heights$son)
r <- cor(galton_heights$father, galton_heights$son)

galton_heights %>%
  ggplot(aes(father, son)) +
  geom_point(alpha = 0.5) +
  geom_abline(slope = r * s_y/s_x, intercept = mu_y - r * s_y/s_x * mu_x)

The regression formula implies that if we first standardize the variables, that is subtract the average and divide by the standard deviation, then the regression line has intercept 0 and slope equal to the correlation $\rho$. You can make same plot, but using standard units like this:

galton_heights %>%
  ggplot(aes(scale(father), scale(son))) +
  geom_point(alpha = 0.5) +
  geom_abline(intercept = 0, slope = r)

B <- 1000
N <- 50

set.seed(1983)
conditional_avg <- replicate(B, {
  dat <- sample_n(galton_heights, N)
  dat %>% filter(round(father) == 72) %>%
    summarize(avg = mean(son)) %>%
    pull(avg)
  })

regression_prediction <- replicate(B, {
  dat <- sample_n(galton_heights, N)
  mu_x <- mean(dat$father)
  mu_y <- mean(dat$son)
  s_x <- sd(dat$father)
  s_y <- sd(dat$son)
  r <- cor(dat$father, dat$son)
  mu_y + r*(72 - mu_x)/s_x*s_y
})

mean(conditional_avg, na.rm = TRUE)

## [1] 70.49368

mean(regression_prediction)

## [1] 70.50941

sd(conditional_avg, na.rm = TRUE)

## [1] 0.9635814

sd(regression_prediction)

## [1] 0.4520833

galton_heights %>%
  mutate(z_father = round((father - mean(father)) / sd(father))) %>%
  filter(z_father %in% -2:2) %>%
  ggplot() +
  stat_qq(aes(sample = son)) +
  facet_wrap( ~ z_father)

mu_x <- mean(galton_heights$father)
mu_y <- mean(galton_heights$son)
s_x <- sd(galton_heights$father)
s_y <- sd(galton_heights$son)
r <- cor(galton_heights$father, galton_heights$son)
m_1 <-  r * s_y / s_x
b_1 <- mu_y - m_1*mu_x

m_2 <-  r * s_x / s_y
b_2 <- mu_x - m_2 * mu_y

galton_heights %>%
  ggplot(aes(father, son)) +
  geom_point(alpha = 0.5) +
  geom_abline(intercept = b_1, slope = m_1, col = "blue") +
  geom_abline(intercept = -b_2/m_2, slope = 1/m_2, col = "red")