Correlation analysis

Correlation analysis quantifies the direction and strength of association between two variables.

When to use it

Use it to test a suspected relationship, compare associations or explore quantified data for patterns. An ice-cream seller might examine temperature and sales, then compare that relationship with seasonality. Exploration can reveal unexpected signals: Walmart famously observed higher Pop-Tart purchases before hurricanes and adjusted placement; the often-repeated beer-and-nappies story illustrates the same basket-analysis idea, although any such claim requires verification before action.

The method can support questions such as:

• Are our most loyal customers also our most profitable?
• Do customers purchase more when the price is lower?
• Does pay influence length of tenure?
• Does number of annual holidays influence absenteeism?
• Is there any relationship between factor X and factor Y?

Use the evidence to challenge assumptions before changing strategy, pricing or product mix.

Origins

Correlation developed through nineteenth-century work by Francis Galton and Karl Pearson on measuring co-variation. Pearson formalised the product–moment coefficient, while Charles Spearman later developed rank correlation. Modern software made these calculations routine, but the interpretive rule remains unchanged: association alone does not establish cause.

What it is

Pearson correlation applies to paired numeric measurements, not unencoded categories such as brand or colour. The coefficient ranges from minus one to plus one. In the legacy typesetting these endpoints appeared as 11 and 21. A positive coefficient means high values tend to occur together; a negative coefficient means one tends to rise as the other falls; zero indicates no linear association. Thus a hypothetical plus zero point seven three relationship between height and IQ would be positive (legacy code 10.73), while minus zero point six four would be inverse (legacy code 20.64). Strength grows as the absolute value approaches 1. A coefficient of 0.5 is sometimes treated as practically notable, but statistical significance cannot be inferred from that threshold alone; it depends on sample size and assumptions.