Why the Unit of Observation Matters
In “Lurking Variables and Confounding,” you considered how a third variable can contribute to an observed association. There is another important question to ask when reading a correlation: what does each point represent? A point might represent one person, or it might represent an average for a whole group.
Those are different data sets. A correlation calculated from individual measurements describes the linear association among individuals. A correlation calculated from group averages describes the linear association among groups. Often, the group averages follow a clearer, stronger linear pattern. Averaging can smooth out individual differences that make a pattern harder to see. But this is a tendency, not a rule, and the group-level pattern does not necessarily describe the individuals in those groups.
The distinction is not just a change in how the same points are summarized. A group average replaces many individual observations with a single value, and the groups—not the people—become the cases in the correlation. If the question is about individuals, a correlation among group averages alone cannot answer it.
How Averaging Can Strengthen a Pattern
Imagine that different groups have different typical values of \(x\) and \(y\), and that groups with larger average \(x\) also tend to have larger average \(y\). Within any one group, however, individual values may vary for many reasons. One person may be far above the group average for \(x\), while another is far below it; their \(y\)-values may also differ from their group’s average.
When measurements are averaged within each group, some of those individual ups and downs no longer appear as separate points. The group mean represents the group’s typical value instead. If the remaining differences between groups line up in a roughly straight pattern, the group averages can have a stronger correlation than the individuals did.
This explanation depends on the setting. Averaging does not guarantee a stronger correlation. It changes which variation is represented: variation among individual people is set aside, while variation among group means becomes the basis for the correlation. The groups might show a weak pattern, no pattern, or a pattern in the opposite direction from the individuals.
Worked Example: A Moderate Individual Correlation, a Stronger Group Correlation
Consider a fictional set of 12 students in three study groups. For each student, \(x\) is minutes spent on a practice activity and \(y\) is a score on a follow-up task. The values are invented for illustration. Each group has four students.
| Study group | Students’ \(x\)-values | Students’ \(y\)-values | Group mean \(\bar{x}\) | Group mean \(\bar{y}\) |
|---|---|---|---|---|
| A | 7, 9, 11, 13 | 19, 23, 17, 21 | 10 | 20 |
| B | 9, 11, 13, 15 | 21, 25, 19, 23 | 12 | 22 |
| C | 11, 13, 15, 17 | 23, 27, 21, 25 | 14 | 24 |
Correlation among group averages: The three points are \((10,20)\), \((12,22)\), and \((14,24)\). They fall exactly on the line \(\bar{y}=\bar{x}+10\), so the correlation among the three group averages is \(r=1\). This is a perfect positive linear association among these three group averages.
Correlation among individuals: The overall means are \(\bar{x}=12\) and \(\bar{y}=22\). To calculate the individual correlation, use the sums of cross-products and squared deviations. The between-group contribution to each sum of squares is \(4(4+0+4)=32\), because each group has four students and the group means differ from the overall means by \(-2,0,2\). Within each group, the \(x\)-deviations from that group’s mean are \(-3,-1,1,3\), with squared deviations summing to \(20\). The \(y\)-deviations from its group mean are \(-1,3,-3,1\), also with squared deviations summing to \(20\). Their within-group cross-products sum to \(0\). Across three groups, this gives sums of squares \(32+3(20)=92\) for each variable and a sum of cross-products of \(32+3(0)=32\).
Interpret in context: Among these 12 students, minutes spent on the practice activity and follow-up task score have a moderate positive linear association. Among the three group averages, the association is perfectly positive and linear. The group-level correlation is stronger here, but it does not mean every student who spends more time on the activity will have a higher score.
A Strong Group Pattern Can Hide an Opposite Individual Pattern
Aggregation can do more than make a pattern stronger. It can change the direction of the pattern. Differences between groups can point one way, while the relationship among individuals within groups points another way. This is one reason not to use a correlation among group averages as a substitute for an individual-level correlation.
The example below uses the same group mean locations as the previous one, but different individual \(y\)-values within each group. The example is deliberately constructed to show that the group-level direction does not determine the individual-level direction.
Worked Example: Group Averages Rise While Individual Values Fall
In a fictional program, each of three workshops includes four participants. For each participant, \(x\) is a practice score and \(y\) is a follow-up score. The table gives the paired individual measurements and each workshop’s averages.
| Workshop | Participants’ \(x\)-values | Participants’ \(y\)-values | Mean \(x\) | Mean \(y\) |
|---|---|---|---|---|
| A | 7, 9, 11, 13 | 23, 21, 19, 17 | 10 | 20 |
| B | 9, 11, 13, 15 | 25, 23, 21, 19 | 12 | 22 |
| C | 11, 13, 15, 17 | 27, 25, 23, 21 | 14 | 24 |
Correlation among workshop averages: The group-mean points are again \((10,20)\), \((12,22)\), and \((14,24)\). They fall on a straight line with positive slope, so their correlation is \(r=1\).
Correlation among individuals: The overall means are \(12\) and \(22\). The between-workshop sum of cross-products is \(32\), as before. Within each workshop, the \(x\)-deviations are \(-3,-1,1,3\), while the \(y\)-deviations are \(3,1,-1,-3\). The within-workshop cross-products sum to \((-3)(3)+(-1)(1)+(1)(-1)+(3)(-3)=-20\). The total sum of cross-products is therefore \(32+3(-20)=-28\). The sum of squared deviations for each variable is \(32+3(20)=92\).
Interpret in context: The workshop averages have a perfect positive linear association, but the 12 individual practice and follow-up scores have a weak negative linear association. The group-level result cannot be used to claim that participants with higher practice scores tend to have higher follow-up scores. In this constructed example, the individual data show the opposite direction.
A common danger in this situation is making an ecological inference: using a pattern among groups to make a claim about the individuals within those groups. A group with a higher average \(x\) and a higher average \(y\) does not tell us how \(x\) and \(y\) are associated for the people inside that group. To answer an individual-level question, examine individual paired data.
Calculate and Describe a Correlation of Group Averages
When the question is specifically about groups, calculate \(r\) from the paired group means just as you would calculate it from paired individual measurements. Each group supplies one pair: its mean \(x\) and its mean \(y\). Then interpret the resulting correlation at the group level, naming the variables and the groups in context.
For example, suppose a fictional community survey reports average recreational screen time and average nightly sleep for four neighborhoods. The values below are neighborhood averages, not measurements for individual residents.
Worked Example: Screen Time and Sleep Across Neighborhoods
The table lists average recreational screen time per resident and average nightly sleep per resident in four fictional neighborhoods. Treat the four neighborhoods as the cases.
| Neighborhood | Average screen time, \(x\) (hours) | Average sleep, \(y\) (hours) |
|---|---|---|
| A | 1 | 8 |
| B | 2 | 7 |
| C | 3 | 6 |
| D | 4 | 4 |
Identify the cases: Each point represents a neighborhood. The quantitative variables are average screen time in hours and average sleep in hours.
Calculate the correlation: The means are \(\bar{x}=2.5\) hours and \(\bar{y}=6.25\) hours. The sum of cross-products is \((-1.5)(1.75)+(-0.5)(0.75)+(0.5)(-0.25)+(1.5)(-2.25)=-6.5\). The sums of squared deviations are \(5\) for \(x\) and \(8.75\) for \(y\). Therefore:
Interpret in context: Among these four neighborhoods, average recreational screen time and average nightly sleep have a very strong negative linear association. Neighborhoods with higher average screen time tend to have lower average sleep. This describes the neighborhood averages; it does not establish that individual residents with more screen time sleep less, and it does not show that screen time causes less sleep.
What to Check Before Interpreting the Result
As in “What the Correlation Coefficient Measures,” \(r\) describes the direction and strength of a linear association. The calculation alone does not tell you whether the data are individual measurements or group summaries. Before interpreting \(r\), check the cases represented by the points and the units or definitions behind each group average.
- Identify the observational unit. Is one point a person, a class, a school, a neighborhood, or another group?
- State what each average summarizes. An average for a group may conceal a wide range of individual values.
- Match the conclusion to the cases. A correlation among schools supports a description of the schools in the data, not automatically a claim about students.
- Inspect the pattern and context. As in “Judging Strength of an Association,” the strength of \(r\) concerns how closely points follow a linear pattern. A high group-level \(r\) does not reveal the individual-level scatter.
- Avoid causal claims. As discussed in “Why Correlation Does Not Imply Causation,” an association by itself does not establish cause and effect. Averaging does not remove that limitation.
The number of groups also matters for how much information the group-level correlation contains. In the first two examples, there were only three group averages. Their perfect group-level correlation came from three points lying on a line; it did not show that a broad population of groups would follow that pattern. Describe the data you have, and avoid claiming more than they support.
Common Mistakes and AP Exam Tips
- Treating group averages as individual measurements. A point for a school represents that school’s average, not a typical student’s paired values. Full-credit wording identifies the groups as the cases.
- Assuming averaging must make \(r\) stronger. It often can smooth individual variation, but the examples show that aggregation can also leave a pattern weaker or reverse its direction. Say “can” or “often,” not “always.”
- Using a group correlation to make an individual claim. A strong association among neighborhood averages does not prove the same association among residents. State the level of the data in your interpretation.
- Confusing a high correlation with a causal effect. A strong group-level association does not prove that changing one group average would change the other. Describe the observed association and keep causal language out unless the study design supports it.
- Leaving out the observational unit. “There is a strong negative correlation” is incomplete when the question concerns group averages. Name the groups and both variables, and specify that the correlation is among their averages.
A strong AP response is explicit about the level of analysis: “Among the neighborhoods in the data, average screen time and average sleep have a very strong negative linear association.” It does not substitute “people who use screens more sleep less,” because that is an individual-level claim the group averages alone cannot establish.
Check Your Understanding
For each question, distinguish the group-level pattern from any claim about individuals.
- A researcher finds a strong positive correlation between average weekly exercise and average fitness scores across six schools. What does this correlation describe, and what individual-level claim would it not establish?
- Why can averaging measurements within each group make a linear pattern among group averages look stronger than the pattern among individuals?
- In a data set, the group-mean points have \(r=-0.90\), while individual observations have \(r=0.15\). State one careful interpretation of each correlation.
- Can a correlation among group averages be stronger than the individual correlation in every setting? Explain why or why not.
- What is an ecological inference, and how could a student avoid making one when interpreting a correlation of group averages?