Linear Regression Analysis - Part 1


Let’s talk about correlation.

General correlation is the quantification of a relationship between variables, most commonly represented as X & Y. The goal of correlation is to identify a factor, X, that we can use to make useful predictions about Y. We’ll get into the philisophical arguments about extrapolation vs interpolation at another time, but for now, it’s just something to mention.

Ideally, we want to make causative connections between variables in data sets, but in the real world this is difficult due to the wide array of interactions that truly exist in complex systems. You’ll often hear, “Correlation does not imply causation” thrown around. While this is true, it can be misleading. It does not necessarily imply causation, but it also doesn’t negate it. In general, if you find high correlation, you should continue whatever reasoning led you to that point, but refine your methods and ask further related questions.

Should you find correlation, you’ll be presented with a number ranging from -1 to +1. That number will tell you a lot. It will tell you whether the relationship is positive, with an upward sloping line, or a negative one, with a downward slope. It will tell you how strong the relationship is, with a number further from 0.0 implying a more useful predictive quality. The accepted threshold for a correlation coefficient that you should take seriously, is less than -0.50 or greater than +0.50.

Let’s say you find a correlation of 0.75, this simply means that 75% of your Y’s behaviour can theoretically be predicted by variable X with your model. We’d all like to be able to predict future outcomes, and linear correlation is one way we can. It’s not without risks, you may perhaps dump resources into a faulty correlation, but when done correctly, the rewards can be worth the investment.

The most common method for linear regression is the Pearson method, a parametric test which yields a correlation coefficient we can use for modeling.

Here’s a brief overview:

Pearson Test
Assumptions:
Linear associations
Continuous variables
Line Model - Y=mx+b
Data is normally distributed
Danger: Vulnerable to outliers (Influential outlier effect)

There’s always more theory to cover, and we will in the future. But for now, let’s figure out how to run this powerful analysis quickly, and pain-free with R.

Written on October 28, 2017