Example 1

This is a simple sine wave. A clear but decidedly non-linear relationship. The correlation here is 0 but the relationship obvious.

set.seed(2112)
x <- seq(0, 3*pi, .01)
y <- sin(x)
cor_xy <- cor(x, y) %>% round(2)
plot(x, y, type = "l", lwd = 3, col = wp[5])
abline(lm(y ~ x), col = wp[1], lwd = 3)
text(x = 1.5 * pi, y = .4, labels = paste("Correlation =", cor_xy))

Example 2

Here, what goes up must come down…but if we look only at the correlation, it’s like nothing happened at all.

x <- seq(0, 10, .1)
y <- ifelse(x <= 5, x, 10-x)
cor_xy <- cor(x, y) %>% round(2)
plot(x, y, type = "l", lwd = 3, col = wp[5])
abline(lm(y ~ x), col = wp[1], lwd = 3)
text(x = 5, y = 3, labels = paste("Correlation =", cor_xy))

Example 3

This is one that I borrowed from Nassim Nicholas Taleb here. He presents a more thorough (if varied and ad hominem-laden) treatment of the misuse of correlation. I could do without the tone and attacks in that piece but I thought this core demonstration was particularly relevant to human capital analytics.

In words, he shows that a non-linear relationship in the presence of noise can fool us into thinking we have a strong linear relationship.

For example, imagine you have a post-hire training assessment scored from 0-100 and a first-month performance measure between 0 and 150 (the units are arbitrary here). Let’s further assume that the true underlying relationship between the assessment and later first-month performance is non-linear such that performance increases as assessment scores increase from 0-50 and then totally flatten for those scoring above 50 (see figure below).

x <- 0:100
y <- ifelse(x <= 50, x, 50) + 30
cor_xy <- cor(x,y) %>% round(2)
plot(x, y, col = wp[5], ylim = c(0, 150),
     xlab = "Test", ylab = "Performance", pch = 19)
abline(lm(y ~ x), col = wp[1], lwd = 3)
text(50, 5, labels = paste("Correlation = ", cor_xy))

We see a super strong correlation in this noise-free case, but even here it’s deceptive. What we really have is a perfect correlation for the first half and then no relationship for those in the second half. Zero.

Now we’ll progressively add noise (with a mean of zero) to the performance scores and then watch how the correlation decreases gradually but hides the underlying non-linear relationship.

Eventually we see a result that fools us into thinking we have a great linear relationship.

for (i in seq(0, 10, 2)) {
      set.seed(42)
      x <- rep(1:100,10)
      y <- ifelse(x <= 50, x, 50) + rnorm(length(x), mean = 0, sd = 2*i) + 30
      cor_xy <- cor(x,y) %>% round(2)
      plot(x, y, col = wp[5], ylim = c(0, 150),
           xlab = "Test", ylab = "Performance", pch = 19)
      abline(lm(y ~ x), col = wp[1], lwd = 3)
      text(50, 5, labels = paste("Correlation = ", cor_xy, "\nNoise sd= ", 2*i))
      
}

In the analytics and predictive modeling sphere, we would take a .57 correlation every day. The results here should give us pause about using linear methods in the presence of possible non-linear relationships.

In a future post, I will say more about the issues and alternatives to using linear methods like correlation. For now, I encourage you to take some time for independent research, see what you learn….and please, always plot your data.

Until then…