Describe a relationship before fitting it

Try

Plot the original synthetic pairs (1,8), (2,6), (3,4), (4,2). Each x increase of one accompanies a y decrease of two. Sketch the decreasing straight line before calculating anything. Now imagine shuffling the y values among the x values: the separate x and y distributions stay the same, but their relationship can change completely.

Predict

Predict the sign of correlation for the plotted pairs. Then ask a harder question: can a perfectly predictable relationship have zero correlation? Think of a U-shaped pattern with equally balanced left and right sides. Correlation measures linear association, not every possible form of dependence.

Experiment

Construct a second scatterplot from (−1,1), (0,0), (1,1). These points follow y=x². Mean x is zero and mean y is 2/3. Cross-deviation products are −1/3, 0 and 1/3, totaling zero. Both variables have nonzero spread, so Pearson correlation is zero.

Contrast the two sketches in complete sentences: the first has perfect negative linear association with r=−1; the second has a deterministic nonlinear relationship but r=0. Add a fourth point to either plot and predict its effect before recomputing. The point’s horizontal position and vertical departure both matter.

Understand

A scatterplot displays two quantitative measurements on the same units. Describe direction, form, strength and unusual observations in context. Direction can be positive or negative for a broad increasing or decreasing trend. Form distinguishes a roughly straight relationship from a curve, clusters or other structures. Strength concerns how tightly observations follow that form, not how steep the graph looks.

For n pairs with positive sample standard deviations sx and sy, Pearson r is the sum of products of standardized deviations divided by n−1. Equivalently it is the cross-deviation sum divided by the square root of the product of the two squared-deviation sums. It lies between −1 and 1. It is undefined when either variable is constant because the denominator vanishes.

Correlation has no units. Changing metres to centimetres multiplies deviations and spread by the same positive factor, leaving r unchanged. A negative rescaling reverses its sign. Swapping the variables leaves r unchanged, though the regression line predicting y from x generally differs from the reversed prediction model.

A large magnitude of r does not establish causality. Seasonal temperature can influence both ice cream sales and swimming activity. An observational association can arise from confounding, selection, reverse causation or a mixture of mechanisms. The study design must support the causal interpretation; the correlation calculation cannot supply it afterward.

An unusual point can greatly change r. A point far from the main x range may strengthen an existing linear pattern if its y follows the trend, or weaken and even reverse the association if it does not. Inspect the paired data and investigate measurement provenance before excluding anything. Deletion merely to obtain a desired correlation is not a defensible analysis.

Apply

Explain the zero-correlation example to someone who interprets zero as “nothing is related.” Use both the cancellation calculation and the rule y=x². For a workplace application, identify two dashboard quantities that might move together because of a third factor, and state what additional design or information would be needed to investigate a causal claim.

Continue

Next fit a predictive line and inspect residuals. A line is useful only insofar as its errors and intended use have been examined.