Correlation is the first relationship most analysts learn to measure and the one most often over-interpreted. The number itself is simple. What it means, and what it does not, takes a bit more care.
What the coefficient measures
The Pearson correlation coefficient, written r, measures how closely two numeric variables follow a straight-line relationship. It ranges from minus 1 to plus 1.
- +1: every point sits exactly on a rising straight line. Knowing one variable tells you the other precisely. Temperature in Celsius and in Fahrenheit is an example.
- 0: no straight-line relationship. It does not necessarily mean no relationship of any kind.
- -1: every point sits exactly on a falling straight line. Remaining balance on a fixed-amount prepaid card against amount spent is an example.
Real data almost never produces these extremes. Under the hood, r is the covariance of the two variables divided by the product of their standard deviations. Dividing by the standard deviations makes it unit-free, so it does not matter whether you measure spend in rupees or in lakh.
The sign gives direction and the magnitude gives tightness, not steepness. A correlation of 0.9 between ad spend and leads does not mean leads rise quickly with spend; it means the points hug a line closely. The line might be nearly flat. Slope is a regression question.
Squaring r gives the share of variation in one variable that a straight line on the other accounts for. An r of 0.6 sounds impressive, but r squared is 0.36: the line explains about a third of the variation, and two-thirds is due to other things. An r of 0.3, which is often reported as a meaningful finding, explains about 9%.
When the number lies
Curves. Take x values from minus 5 to plus 5 and let y be x squared. The relationship is perfect, yet r is exactly zero because the rising and falling halves cancel out. Customer satisfaction against delivery speed, or productivity against hours worked, can bend like this. Always plot the data before trusting r.
Outliers. One extreme point can create or destroy a correlation. A single enterprise customer with enormous revenue and enormous support usage can make “revenue versus support tickets” look strongly correlated across a hundred customers who otherwise show no pattern. Spearman’s rank correlation, which correlates the ranks rather than the raw values, is less sensitive to such points and also captures relationships that are consistently rising or falling but not straight.
Mixed groups. Combining groups with different baselines can produce a correlation that exists in no individual group, or reverse one that does. Francis Anscombe’s well-known quartet of 1973 makes the point visually: four small datasets with nearly identical means, variances and correlations look completely different when plotted.
Restricted range. If you only look at employees who passed a hiring test, the correlation between test score and job performance will appear weaker than it really is, because you have cut off the low scorers.
Why correlation is not causation
A strong correlation between A and B is consistent with several different realities, and only one of them is “A causes B”.
A common cause. Ice cream sales and drowning incidents rise together, because hot weather drives both. In business: stores with more staff have higher sales, but both may be driven by footfall in that location. Adding staff to a quiet store will not create footfall. The hidden driver is called a confounder.
Reverse causality. Customers who use the loyalty app spend more. Does the app make them spend, or do high spenders bother to install the app? Probably a lot of the latter.
Selection. If you only analyse customers who are still active, any trait linked to staying around will appear linked to success, even if it does nothing. This is survivorship bias in another form.
Coincidence. Search enough pairs of time series and some will move together by chance. Two series that both trend upward over time, such as smartphone users in India and the price of a kilogram of rice, will correlate strongly without any connection beyond time itself. Correlating changes rather than levels removes much of this effect.
Getting from correlation towards cause
The gold standard is a randomised experiment. If you randomly choose which customers get a discount, the only systematic difference between the groups is the discount, so a difference in outcomes can be attributed to it. A/B tests are exactly this.
When experiments are impossible or unethical, analysts use designs that approximate them. Difference-in-differences compares the change in a treated group with the change in a similar untreated group over the same period, for example cities that got a new delivery hub against cities that did not. Regression discontinuity looks at people just either side of a cut-off, like customers just above and just below a loyalty tier threshold. Controlling for known confounders in a regression helps, but cannot account for confounders you have not measured.
Without such a design, the honest phrasing is associational: “customers who used feature X retained at a higher rate” rather than “feature X increases retention”. The first sentence is true and useful for generating hypotheses. The second is a claim the data has not earned, and acting on it, by pushing every customer into feature X, can waste a quarter’s effort on a lever that does not move anything.
Written by the RP Strativa Team. Examples are illustrative.
Leave a Reply