What a Correlation Coefficient Summarizes
A scatterplot shows how two quantitative variables vary together across paired observations. Pearson’s sample correlation coefficient, written \(r\), gives a compact numerical summary of the direction and strength of the linear pattern in those data. It is useful when a single number is needed, but it cannot replace looking at the scatterplot: a numerical summary necessarily leaves out details of the pattern.
The sign of \(r\) indicates direction. A positive value goes with an upward linear pattern: larger values of one variable tend to occur with larger values of the other. A negative value goes with a downward linear pattern: larger values of one tend to occur with smaller values of the other. The magnitude, or absolute value, of \(r\) describes how closely the data follow a straight-line pattern. Values nearer to 1 in magnitude indicate a tighter linear pattern; values nearer to 0 indicate a weaker linear pattern.
The endpoints describe ideal cases. If \(r=1\), all the plotted points lie exactly on a straight line that rises from left to right. If \(r=-1\), they lie exactly on a straight line that falls from left to right. In most data, points do not lie perfectly on a line, so \(r\) is between these endpoints. In particular, \(r=0\) indicates no linear association in the sample; it does not, by itself, establish that the variables are unrelated in every possible way.
The coefficient is unit-free: its value is not expressed in the units of either variable. Its interpretation still depends on context, however. A correlation summarizes the observed linear pattern for the data at hand; it does not say what caused the pattern, provide a prediction in the units of the response variable, or describe every feature visible in a scatterplot.
Reading the Sign and Magnitude
Begin with the sign, then consider the magnitude. A phrase such as “negative association” is incomplete unless it identifies the variables and explains what larger values tend to accompany. Likewise, describing a correlation as “strong” or “weak” should make clear that the description concerns a linear pattern.
To compare the strength of two correlations, compare their absolute values. For example, \(r=-0.72\) has greater magnitude than \(r=0.45\), because
Thus, the first value indicates a stronger linear pattern in magnitude, while its negative sign indicates the opposite direction. The number alone does not supply universal labels such as “weak,” “moderate,” or “strong.” Such labels are qualitative, and sensible descriptions depend on the setting, the scatterplot, and the purpose of the analysis. When using them, explain what they mean in context rather than treating them as fixed categories separated by universal cutoffs.
Worked Example: A Positive Correlation Between Practice and Score
In a clearly hypothetical dataset, 24 students each report how many hours they practiced for a skills assessment, and each student receives a score out of 100. A scatterplot shows an upward pattern that is fairly close to a straight line, and the reported sample correlation is \(r=0.86\). Interpret this result.
Direction. Since \(r=0.86>0\), the sample association is positive. Among these students, more practice hours tend to occur alongside higher assessment scores.
Strength of the linear pattern. The magnitude is \(|r|=|0.86|=0.86\). This is relatively close to 1, and the scatterplot is described as fairly close to a rising straight line. Together, those facts support describing the linear association as strong in this hypothetical sample.
Context and limits. A suitable interpretation is: “For these 24 students, reported practice hours and assessment scores have a strong positive linear association: students reporting more practice hours tend to have higher scores.” The value \(0.86\) does not mean that scores rise by 0.86 points per practice hour; that would confuse correlation with a slope. Nor does it mean that 86% of students improved or that practicing caused higher scores.
A useful justification for comparing magnitudes is that sign and strength answer different questions. The sign distinguishes an upward from a downward linear pattern; the absolute value removes that sign and leaves a measure of how pronounced the linear pattern is. For instance, \(-0.72\) and \(0.72\) have equal magnitude, since \(|-0.72|=|0.72|=0.72\), even though their directions differ. Therefore, an interpretation needs both pieces: the sign for direction and the magnitude for strength.
Worked Example: A Negative Correlation Between Sleep and Response Time
Suppose a hypothetical study records usual nightly sleep, in hours, and response time, in milliseconds, for 40 adults completing the same task. The reported correlation is \(r=-0.58\). The scatterplot slopes downward, with a noticeable but not especially tight linear pattern. Describe the result.
Direction. Since \(-0.58<0\), the association is negative. In this sample, adults reporting more hours of sleep tend to have shorter response times. The negative sign describes how the two recorded variables vary together; it does not mean that either variable itself is negative.
Magnitude. Here \(|r|=|-0.58|=0.58\). The coefficient is not near either endpoint, and the plot is not especially tight around a line. It is reasonable to describe the sample’s linear association as moderate, provided that this is a qualitative description rather than a universal classification.
Units and interpretation. The hours and milliseconds belong in the contextual explanation, but not as units attached to \(r\). A careful summary is: “In these 40 adults, usual nightly sleep and response time show a moderate negative linear association; those reporting more sleep tend to have shorter response times.” This statement summarizes association in the observed sample. It does not establish that increasing sleep would cause a particular change in response time.
Why the Scatterplot Still Matters
Pearson’s \(r\) is specifically a summary of linear pattern. It is possible for a scatterplot to show a clear relationship that bends rather than following a straight line. In that situation, the points may have little overall upward or downward linear tendency, so \(r\) can be near zero even though the variables are not unrelated.
Worked Example: A Curved Pattern With Correlation Zero
Consider a hypothetical set of seven paired observations. The first variable is temperature in degrees Celsius, and the second is a constructed measure of deviation from a preferred temperature. The plotted points are:
| Temperature (°C) | Deviation measure |
|---|---|
| 1 | 9 |
| 2 | 4 |
| 3 | 1 |
| 4 | 0 |
| 5 | 1 |
| 6 | 4 |
| 7 | 9 |
The plot falls as temperature moves from 1°C to 4°C, then rises as temperature moves from 4°C to 7°C. It forms a symmetric curve, not a straight-line pattern. For these pairs, the sample correlation is \(r=0\), so \(|r|=|0|=0\): there is no overall linear association in the sample.
This result follows from the symmetry of the points around the center temperature, 4°C. For each temperature \(4-d\) with deviation measure \(d^2\), there is a matching temperature \(4+d\) with the same deviation measure \(d^2\). The paired values on the left and right therefore balance rather than producing an overall rising or falling straight-line trend. The scatterplot nevertheless shows a clear curved pattern. The correct conclusion is “no linear association in these data,” not “no relationship.”
This example illustrates why interpretation should begin with the plotted data, not stop at a reported coefficient. A scatterplot can reveal curvature or other features that a single linear summary does not describe. Conversely, when a scatterplot looks approximately linear, \(r\) can provide a concise description of the direction and tightness of that pattern.
Common Mistakes
- Calling \(r\) a percentage. The reasoning “\(r=0.80\), so 80% of observations follow the pattern” is incorrect. Correlation is not a percentage of people, observations, or outcomes. Say that the sample has a positive linear association, and describe its magnitude in context.
- Reading \(r\) as a slope. The claim “\(r=0.80\) means the outcome rises by 0.80 units for each unit increase in the explanatory variable” confuses two different summaries. Correlation has no measurement units and is not a change-per-unit rate.
- Treating a small correlation as proof of no relationship. A value near zero indicates little linear association in the sample. A curved pattern can still be clear, so inspect the scatterplot before making a broader claim.
- Reporting direction without naming the variables. “The correlation is negative” gives only part of the interpretation. State which variables were measured and what larger values of one tend to accompany in the other.
- Assigning a universal strength label. There is no single set of numerical cutoffs that makes every correlation “weak,” “moderate,” or “strong” in every field. Use the scatterplot and context to support any qualitative label.
Check Your Understanding
Try these questions without looking back.
- A hypothetical sample has \(r=0.67\) for weekly exercise hours and a fitness-test score. State the direction and explain what the sign means in context.
- Which indicates the stronger linear pattern in magnitude: \(r=-0.71\) or \(r=0.54\)? Show the absolute-value comparison.
- A report says, “The correlation between study time and score is 0.75, so each extra hour of study adds 0.75 points.” Identify the error.
- What does \(r=0\) tell you about linear association, and what does it not tell you about every possible relationship?
- Why should a scatterplot be examined even when a correlation coefficient has already been reported?