Start With the Question, Not the Number of Columns
A data set can contain one sample, two groups, or several measurements on the same people. The layout alone does not determine the right kind of inference. The first question is what population quantity is the research question asking about? Is it a feature of one population compared with a fixed reference value, or a difference between two populations?
The distinction matters because these questions target different parameters. A survey might ask whether the average commute time in one city differs from a stated 30-minute benchmark. Another might ask whether vaccine uptake differs between two populations. A treatment study might compare outcomes under a treatment and a control condition. Each situation calls for a procedure suited to its target, outcome, and design—not merely one that fits the appearance of the data table.
Identify the Target Parameter
A population parameter describes a population, while a sample statistic summarizes observed data. For a quantitative outcome, \(\mu\) denotes a population mean and \(\bar{x}\) a sample mean. For a binary outcome, \(p\) denotes a population proportion and \(\hat{p}\) a sample proportion. The relevant parameter depends on the outcome and the question; the notation does not decide the research question for us.
In a one-sample comparison, the benchmark is a specified number, not a second population estimated from new observations. For example, asking whether a population mean differs from 30 minutes concerns \(\mu\) relative to 30. In a two-group comparison, the target is a contrast, such as \(\mu_1-\mu_2\) or \(p_1-p_2\). Both quantities describe how the populations differ, with the order of subtraction determining the sign.
The response type also matters. A quantitative response leads to a question about a mean; a binary response leads to a question about a proportion. These are broad parameter choices, not procedures by themselves. Choosing between one and two populations comes first; the particular inferential procedure requires further attention to the outcome and design.
| Question structure | Target | What the data contribute |
|---|---|---|
| One population versus a fixed value | One population parameter, such as \(\mu\) or \(p\) | A statistic from one sample, such as \(\bar{x}\) or \(\hat{p}\) |
| Two populations or conditions | A contrast, such as \(\mu_1-\mu_2\) or \(p_1-p_2\) | A statistic from each group, combined to estimate the contrast |
| Same individuals measured in two conditions | A population parameter for within-person differences | One difference per individual; this is a paired design |
The last row is important. Two columns of measurements do not necessarily represent two independent groups. If each person supplies both measurements, the observations are linked. The earlier material on paired designs treats the within-person differences as the data for a one-sample analysis. Do not discard that pairing and analyze the columns as though they came from unrelated people.
Why a Difference of Sample Summaries Targets a Difference of Parameters
For two independent groups with quantitative outcomes, the sample difference \(\bar{x}_1-\bar{x}_2\) is a natural estimate of the population difference \(\mu_1-\mu_2\). The reason follows from the definition of an unbiased sample mean. If each group is sampled so that its sample mean has expected value equal to its population mean, then:
The first equality uses the linearity of expectation: the expected value of a difference is the difference of the expected values. Thus, when the sample means target their corresponding population means, their difference targets the difference in population means. This justifies matching a two-group mean question to a contrast of means; it does not by itself establish that a particular inferential procedure's conditions are satisfied.
The same logic applies to a difference in proportions: each sample proportion estimates its population proportion, and their difference estimates the population difference. In all cases, the study design and sampling process still matter. A well-matched parameter does not repair a biased sample or turn an observational comparison into a randomized experiment.
Worked Examples
Worked Example: A Survey Compared With a Fixed Benchmark
A hypothetical city survey randomly selects 80 adult commuters. Their reported commute times have a sample mean of 31.4 minutes. The city asks whether the population mean commute time differs from its planning benchmark of 30.0 minutes. Which broad type of inference matches the question?
Identify the outcome and population. Commute time is quantitative. The population is the adult commuters in the city to whom the survey is intended to apply, and the target parameter is their mean commute time, \(\mu\).
Identify the comparison. The value 30.0 minutes is a fixed benchmark. It is not a sample mean from a second population. The question asks whether one population mean differs from a specified number, so it calls for one-sample inference about a mean.
Compare the observed statistic with the benchmark. The sample's descriptive difference is:
The sample mean is 1.4 minutes above the benchmark. That arithmetic describes the sample; it does not alone determine whether the population mean differs from the benchmark. A valid analysis also depends on how the 80 commuters were sampled and on the conditions for the specific procedure. The random selection supports representing the target population, provided the sampling process was carried out as intended and responses were obtained appropriately.
Worked Example: A Survey Comparing Two Populations
In a hypothetical survey, 96 of 160 insured adults report receiving a seasonal vaccine, while 73 of 150 uninsured adults report receiving it. The question is whether the population uptake proportions differ between insured and uninsured adults.
Identify the outcome and groups. The outcome is binary: each respondent either reports receiving the vaccine or does not. The two populations are insured adults and uninsured adults. They are distinct groups of people, not two measurements on each person.
Identify the target. Let \(p_1\) be the population proportion of insured adults who received the vaccine and \(p_2\) the corresponding proportion of uninsured adults. The target is \(p_1-p_2\), so this is a two-group question about a difference in population proportions.
Calculate the sample summaries and contrast. The two sample proportions are:
Their difference, using insured minus uninsured, is:
In these hypothetical data, reported uptake is about 11.33 percentage points higher among insured respondents. This is a sample contrast, not a conclusion about the populations by itself. The survey would need to represent the populations of interest, and respondents should not contribute to both groups. Because insurance status was observed rather than assigned, a difference in uptake would not establish that insurance caused the difference; other population characteristics could also be associated with uptake.
Worked Example: A Randomized Treatment Comparison
In a hypothetical randomized study, adults are assigned to a new treatment or a control condition. At eight weeks, the mean systolic blood pressure is 128.4 mm Hg for the 45 people assigned to treatment and 132.7 mm Hg for the 45 assigned to control. The research question is whether the treatment and control conditions have different population mean outcomes.
Identify the outcome and groups. The outcome is quantitative, measured in mm Hg. There are two groups of different people: those assigned to treatment and those assigned to control. This is not a one-population comparison with a fixed clinical benchmark.
Identify the target and calculate the observed contrast. Let \(\mu_T\) and \(\mu_C\) be the population mean eight-week blood pressures under the treatment and control conditions, respectively. The target is \(\mu_T-\mu_C\), estimated by the difference in sample means:
The treatment group's sample mean is 4.3 mm Hg lower than the control group's sample mean. The two-group structure is clear because the research question compares outcomes under two conditions. Random assignment, if properly conducted, supports a causal interpretation of the treatment contrast; the mere presence of two groups would not. Appropriate interpretation also depends on the trial's implementation, including whether outcomes were measured as planned and whether missing data or other problems distorted the comparison.
A Practical Decision Process
Name the population or populations, the outcome, and the comparison being requested. Avoid starting from the number of columns in a data file.
Ask whether the question concerns one population parameter compared with a fixed value, or a contrast between two population parameters.
Determine whether groups contain different individuals, or whether the same individuals supplied linked observations. A paired design calls for analysis of within-person differences, not an independent two-group comparison.
Classify the outcome as quantitative or categorical, then check how observations were selected or assigned. These features guide the specific procedure and what conclusions are justified.
Why the Distinction Matters
Choosing the wrong comparison changes the question being answered. If the scientific question concerns a single population relative to a policy standard, treating the standard as if it were a second sampled population invents data that do not exist. Conversely, comparing two populations by testing each one separately against a benchmark does not directly answer whether the populations differ from each other. The target must match the question.
Study design also affects what a result can mean. A survey can estimate or compare population characteristics when its sampling supports that goal, but a comparison between naturally occurring groups may reflect confounding as well as the characteristic being studied. Randomized treatment assignment can support a causal comparison under suitable conditions. A procedure choice cannot compensate for a design that does not support the desired conclusion.
Common Mistakes
- Wrong reasoning: “The table has two columns of numbers, so this is a two-group analysis.” Correction: Find out whether the columns represent different groups or repeated measurements on the same individuals. Paired measurements are linked.
- Wrong reasoning: “A benchmark is the second group.” Correction: A fixed value such as 30 minutes is not a second population sample. Comparing one population parameter with it is a one-sample question.
- Wrong reasoning: “If each group differs from the benchmark, the groups must differ from each other.” Correction: The direct two-group target is the difference between the population parameters. Separate comparisons to a benchmark do not replace that contrast.
- Wrong reasoning: “A two-group difference proves that group membership caused the outcome.” Correction: Causal interpretation depends on the design, especially whether treatment was randomly assigned; observational group differences alone establish no such cause.
Check Your Understanding
For each situation, identify the target comparison and explain which broad type of inference fits.
- A random sample of clinic patients is used to compare the population mean waiting time with a stated 20-minute service target. Is the target one population parameter or a contrast between two populations?
- A survey compares the proportion of renters and homeowners who report having a regular health-care provider. What is the population contrast?
- A clinic measures each patient's blood pressure before and after an intervention. Why should the two columns not automatically be treated as independent groups?
- A randomized study compares the mean symptom score of people assigned to a new treatment with that of people assigned to control. What feature makes this a two-group question, and what does random assignment contribute?
- A sample mean is 2.7 units above a fixed benchmark. Does that sample difference alone establish that the population mean differs from the benchmark? Explain.