Scatterplots reveal how two quantitative variables vary together. A fitted line summarizes an approximately linear relationship, while correlation describes its direction and strength. These tools support prediction, but their conclusions depend on the data range, unusual observations, and the distinction between association and causation.
Read the point cloud before fitting
Describe direction, form, strength, and unusual points. A positive direction means tends to increase with ; a negative direction means it tends to decrease. Form may be linear or curved. Strength concerns how tightly observations follow the pattern.
The correlation coefficient r lies between and 1 and measures linear association. A value near 1 or indicates a strong straight-line pattern. A value near zero does not rule out a strong curved relationship. Correlation has no physical units and is unchanged by converting either variable to another unit with a positive scale factor.
A single unusual point can substantially change a correlation or fitted line. Always inspect the scatterplot rather than interpreting r alone.
Example: interpret a prediction equation
A model for weekly sales is where is advertising spending in hundreds of dollars and is units sold. Each additional 100 dollars of advertising is associated with a predicted increase of 18 units, within the model's context.
At spending of 500 dollars, is 5, not 500, giving predicted sales The intercept predicts 120 sales at zero advertising, but that interpretation is useful only if zero is meaningful and reasonably supported by the data.
If actual sales are 198, the residual is Residual always means observed minus predicted. Its sign locates the point relative to the fitted line.
Least squares gives a particular fitted line
The least-squares regression line minimizes the sum of squared vertical residuals. It does not minimize horizontal distances or distances perpendicular to the line. Because the prediction direction is from , exchanging the variables generally produces a different regression equation.
An informal line drawn through a scatterplot should follow the central pattern, but it need not equal the least-squares line exactly. If a problem supplies a regression equation, use that equation rather than rebuilding a line from two conveniently selected data points.
For a linear model with an intercept, the least-squares line passes through This offers a consistency check when the means are provided.
Example: interpolation and extrapolation
Suppose the advertising observations range from equal to 2 through 8. Predicting at equal to 5 is interpolation. Predicting at equal to 30 is extrapolation, far outside the observed range.
The algebra can produce a number at 30, but the business may saturate its market, change prices, or encounter limited inventory. The equation's ability to be evaluated does not establish the reliability of its prediction.
A curved pattern in a residual plot indicates that a straight line systematically misses part of the relationship. Random-looking residual scatter around zero is more consistent with a suitable linear form, though other assumptions may still need checking.
Association is not a causal conclusion
An observed relationship may arise from a lurking variable, selection effects, reverse causation, or coincidence. Random assignment in a well-designed experiment supports causal interpretation; a scatterplot alone does not.
For example, cities with more firefighters may have more fire damage because larger fires require more firefighters. The positive association does not mean firefighters cause the damage.
Try it yourself
A model is At equal to 12, the observed y-value is 30. Find the prediction, residual, and direction of association represented by the line.
Answer: The prediction is 26, and the residual is placing the observation above the line. The negative slope represents decreasing predicted as increases.
What to remember
Inspect the pattern, interpret slope in context, calculate residuals in the correct order, and limit predictions to situations where the model has support. Correlation summarizes linear association, not causation.