Start With the Study Design
In “Comparing Two Means: The Parameter of Interest,” the parameter \(\mu_1-\mu_2\) was defined as population 1’s mean minus population 2’s mean. Before using that parameter in a two-sample confidence interval, we need to know whether the study really provides two independent groups. Two sets of measurements can look like two groups in a table but be connected person by person.
The key is to trace the observational units: the people, animals, objects, or other entities on which the quantitative measurement is made. Ask whether each unit contributes to just one group, or whether observations from the two groups are linked by being measured on the same unit or deliberately matched. The design—not merely the labels “Group 1” and “Group 2”—determines the answer.
“Independent” here describes the relationship between the groups’ observations. It does not mean that the two groups must have the same sample size, the same variability, or different kinds of measurements. The variable should be quantitative in both groups, and the design must provide two unpaired sets of observations.
A Design-Based Decision Process
Use the following questions before deciding which mean comparison applies. First name the observational unit and the measurement. Then follow the unit through the study: was it measured under both conditions, matched to a particular unit in the other group, or placed in exactly one group without a matching link?
State what one observation represents and what quantitative variable is recorded, including its units when known.
Check whether the same unit is measured in both groups or whether units are deliberately matched into pairs.
If observations are linked within pairs, the paired approach concerns the mean of the pairwise differences. If there is no such link and each unit belongs to one group, the comparison is between two independent groups.
For independent groups, define \(\mu_1-\mu_2\) in a stated order. For paired data, define the difference for each pair and the corresponding mean difference, as in the earlier paired-data tutorials.
The distinction is about how observations are connected, not whether the groups’ sample means happen to be close or far apart. A large difference between group means does not make the groups independent, and similar measurements do not make them paired.
Common Designs That Produce Independent Groups
A common independent-groups design takes a sample from each of two populations. For example, a researcher may select different households from two towns and record the weekly water use of each household. If no household is sampled in both towns and the households are not matched, the two groups are unpaired.
A randomized experiment can also create independent groups. If each participant is randomly assigned to exactly one of two treatments and measured once, the treatment groups consist of different participants. There is no within-person pair of treatment measurements. This design can support a comparison of treatment-group means; whether conclusions generalize to a broader population depends on how participants were recruited, and whether a difference can be attributed to treatment depends on the experiment’s assignment process.
Independent groups do not require equal group sizes. One group might contain 28 people and the other 35, for instance. Unequal counts do not create pairing, just as equal counts do not prove that pairing exists.
How Pairing Changes the Question
As established in “What Makes Data Paired” and “Matched Pairs Versus Independent Samples,” pairing occurs when each unit has two related observations or when subjects are deliberately matched. In a before-and-after study, the same person supplies both measurements. In a matched study, each subject is linked to a particular partner by a planned matching rule. Either design creates one difference per pair.
The paired parameter is \(\mu_d\), the population mean of those within-pair differences, where the subtraction order must be specified. The two-sample parameter \(\mu_1-\mu_2\), in contrast, compares the means of two populations or groups. Both parameters are differences measured in the original variable’s units, but they describe different features of the study.
A useful check is to ask: “Could I point to the specific observation in Group 2 that belongs with this observation in Group 1, based on the design?” If yes, the data are paired. If no, and the units are distinct across groups, an independent-groups comparison is appropriate. Simply placing observations in corresponding spreadsheet rows does not make them pairs.
Worked Examples: Classifying the Design
Worked Example: Water Use in Two Towns
A researcher wants to compare household water use in Town Cedar and Town Lake. The researcher selects a random sample of 32 households from Cedar and a separate random sample of 29 households from Lake. For each household, the researcher records water use in liters per day. No household can appear in both samples, and the researcher does not match households across towns.
The observational unit is a household, and the quantitative variable is daily water use in liters. Each household contributes one observation and belongs to only one town’s sample. There is no design-based link between a Cedar household and a particular Lake household.
Therefore, these are two independent groups, not paired data. The relevant population comparison can be written as \(\mu_1-\mu_2\), with population 1 defined as households in Town Cedar and population 2 as households in Town Lake. The difference is Cedar’s population mean daily use minus Lake’s population mean daily use, in liters per day. The unequal sample sizes do not affect the classification.
Worked Example: Resting Heart Rate Before and After a Program
A health educator records the resting heart rate, in beats per minute, of 18 volunteers immediately before a fitness program and again after eight weeks. Every volunteer provides both measurements. The educator displays the results in two columns labeled “Before” and “After.”
The observational unit is a volunteer. Each person contributes two measurements, and the before value for a person is specifically linked to that same person’s after value. The two columns are therefore not independent groups, even though they contain different measurement occasions.
This is a paired design. Define \(d_i=\text{after}_i-\text{before}_i\) for volunteer \(i\), following the convention introduced in “Defining the Difference Variable.” The mean change is described by \(\mu_d\), not by treating all 18 before values and all 18 after values as unrelated observations. A positive difference would mean a higher heart rate after the program; a negative difference would mean a lower heart rate after the program.
Worked Example: Two Randomly Assigned Study Strategies
A teacher recruits 40 students for a study of study strategies. Each student is randomly assigned to use either Strategy A or Strategy B for a week, but no student uses both. At the end of the week, each student completes the same assessment, scored in points. The teacher plans to compare the assessment averages for the two strategy groups.
The observational unit is a student, and the quantitative measurement is the assessment score in points. Each student is assigned to one strategy and supplies one score. Students in Strategy A are not matched with particular students in Strategy B, and no student contributes a score to both groups.
The study has two independent treatment groups. Define \(\mu_1\) as the true mean assessment score for students like those assigned to Strategy A under these study conditions, and \(\mu_2\) as the corresponding mean for Strategy B. The parameter \(\mu_1-\mu_2\) is the Strategy A mean score minus the Strategy B mean score, in points. Random assignment is relevant to assessing a treatment effect; it does not turn the two groups into pairs.
Worked Example: Matching Cyclists by Experience
A coach studies two training plans. Before assigning plans, the coach matches 24 cyclists into 12 pairs based on years of cycling experience and recent race times. Within each pair, a random process assigns one cyclist to Plan A and the other to Plan B. After six weeks, each cyclist’s time on a standard course is recorded in minutes.
There are two training groups, and each cyclist receives just one plan. However, the design deliberately connects each Plan A cyclist to a specific Plan B cyclist. The matching was part of the study design, so the two groups are not independent for the intended analysis.
This is a matched-pairs experiment. For each matched pair, define a difference such as \(d_i=\text{Plan A cyclist's time}-\text{Plan B cyclist's time}\). The paired parameter is the population mean of these pairwise differences, with the population of pairs defined by the study’s target. It would discard the planned matching to analyze the two sets of 12 times as if there were no links. The fact that the cyclists are different people does not eliminate pairing: deliberate matching is itself a design link.
Details That Can Mislead
Several features can make a design look independent or paired when they are not. Use the design facts rather than surface appearances to classify it.
- Two columns do not guarantee independent groups. “Before” and “After” are separate columns, but if the same people supply both, the observations are paired.
- Two columns do not guarantee paired data. Putting unrelated Group 1 and Group 2 observations on the same spreadsheet row creates no meaningful match.
- Different people can still be paired. If the study deliberately matches one person in one group to a particular person in the other group, analyze the link rather than ignoring it.
- Random assignment does not automatically mean paired. Assigning each person to one of two treatments creates independent treatment groups when people are not matched and no one receives both treatments.
- Equal sample sizes are not evidence of pairing. Independent groups can have the same number of observations by chance or by design. Pairing requires a meaningful one-to-one link, not equal counts.
- Repeated measurements are not separate independent groups. If each object is measured under both conditions, the measurements from that object are linked. The repeated-measure design calls for considering within-object differences.
Also distinguish the classification of observations from the quality of the sampling or experiment. Calling groups independent does not establish that samples were randomly selected, that participants were randomly assigned, or that the results generalize to a particular population. Those are separate design questions that matter when justifying inference and interpreting conclusions.
Common Mistakes and AP Exam Tips
- Choosing a procedure from the wording alone: “Compare two groups” does not settle whether the data are paired. Identify who or what was measured and trace any links across groups.
- Assuming equal group sizes mean pairs: The counts may match without any observation having a partner. State the actual matching rule—or state that no matching link exists.
- Ignoring deliberate matching: Matched subjects are different people, but they are still paired when the study design connects them. An AP response should name that link.
- Mixing up treatment and measurement: A participant assigned to one of two treatments and measured once contributes to one independent treatment group. A participant measured under both treatments contributes linked observations.
- Defining the wrong parameter: For unpaired groups, identify \(\mu_1\) and \(\mu_2\), their populations, and the order in \(\mu_1-\mu_2\). For paired observations, define the subtraction order for each pair and use the mean difference parameter \(\mu_d\).
- Claiming more than the design supports: Classification alone does not establish causation or generalizability. Discuss those claims using the actual random assignment and random sampling features of the study.
For full credit, give a brief design-based justification, not just the label. For example: “These are independent groups because different households were sampled for each town, each household appears in only one sample, and none were matched.” Or: “These are paired data because each participant was measured before and after, so the two measurements are linked within each person.” Such explanations show exactly why the selected comparison fits.
Check Your Understanding
For each situation, identify the observational unit and decide whether the groups are independent or paired. Give the design feature that supports your decision.
- Different randomly selected students from two schools each report their daily commute time. No students are matched across schools.
- Twenty garden plots are each measured for plant height before and after a fertilizer is applied.
- Participants are matched into pairs by age, and within each pair one person is assigned to each of two exercise plans. Each person’s endurance time is recorded once.
- A researcher assigns each of 50 volunteers to one of two phone apps and records one typing-speed score per volunteer. No matching is used.
- A data table lists two groups with 16 observations each. What additional design information would you need before deciding whether the data are paired?