Why the Data Collection Process Matters
A one-sample t interval uses a sample mean and sample standard deviation to estimate a population mean. But the arithmetic alone cannot establish that the interval is appropriate. We also need to ask how the observations were selected and whether one observation gives information about another. These questions matter because the interval’s uncertainty calculation relies on a sample that provides suitable information about the population and on observations that behave independently.
For example, a carefully calculated interval from volunteers who chose to answer an online poll may not represent the population of interest. And a sample of 40 measurements is not necessarily equivalent to 40 independent observations if several measurements came from the same person or group. Checking the data collection process is therefore part of assessing an interval, not an optional detail after the calculation.
For independent observations drawn from the same population, the standard one-sample t interval is
where \(t^*\) is the critical value for the chosen confidence level and \(df=n-1\). This tutorial focuses on the data collection conditions behind that calculation: random sampling and independence. Even if these conditions are supported, a t interval also requires the usual suitability of the sample or population distribution for the t procedure.
Random Sampling and the Population of Interest
A random sampling process gives members of a defined population a chance, determined by the sampling design, to be selected. A simple random sample is one in which every possible sample of the specified size has the same chance of selection. Random selection helps connect the observed sample to the population named in the study question. It does not guarantee that the sample will look exactly like the population: chance variation can produce an unbalanced sample, especially when \(n\) is small.
The population must be stated clearly. If a researcher randomly selects students from one university, the sampling process may support inference about that university’s students, but it does not automatically support inference about all university students. A random method applied to the wrong sampling frame does not fix a mismatch between the sampled group and the intended population.
Random sampling is different from random assignment. Sampling concerns how individuals enter a study and can support generalization from the sample to a population. Assignment concerns how study participants are allocated to treatments. Random assignment is important for other kinds of conclusions, but it does not make a convenience sample representative of a wider population.
Worked Example: A Random Sample of Study Times
A university has a complete list of its 2,000 enrolled students. For a hypothetical survey, a computer selects five students at random, and each reports the number of hours studied last week. The reported values are 6, 7, 8, 9, and 10 hours. Assess the sampling process and calculate a 95% one-sample t interval, assuming the five observations are independent draws from the same population distribution and the distribution is suitable for a t interval.
Sampling assessment. The computer selected students by chance from the stated university population. This is a random sampling process, so the study has a defensible basis for estimating the mean study time among students on that list. Selecting five distinct students from a list of 2,000 is sampling without replacement, so the selections are technically dependent. Because five is a tiny fraction of 2,000, the dependence is negligible, and treating the five observations as independent (as the problem states) is a reasonable approximation. The small sample still makes distributional suitability important; it is assumed here rather than inferred from the sampling method.
The sample mean is
The squared deviations from the mean sum to \((6-8)^2+(7-8)^2+(8-8)^2+(9-8)^2+(10-8)^2=10\). Therefore,
For a 95% interval with \(df=5-1=4\), use \(t^*\approx 2.776\). The margin of error is \(2.776(0.7071)\approx1.963\) hours, giving
Under the stated assumptions, the interval is an estimate of the mean weekly study time for students in the defined university population. The random selection supports that target; it does not make the five values a perfect miniature of all 2,000 students.
Independence Is About How Observations Relate
Independence is a property of the way observations are generated, not a visual feature that can always be checked by looking at a short list of values. If observations are independent, learning one observation does not change the probability distribution for another. In a random sample with replacement from a population, draws can be independent. In sampling without replacement, the draws are dependent: after one member is selected, the composition of the remaining population changes. Whether this dependence is small enough to ignore is assessed using a separate sampling-fraction condition.
Dependence also arises when observations share a person, setting, or group. Repeated measurements on one participant are often more alike than measurements on different participants. Patients treated at the same clinic may share clinicians or local conditions. Students in the same classroom may share a teacher. In each case, treating every recorded value as a separate independent observation can make the apparent amount of information too large.
The reason can be seen from the variance of a mean. For \(n\) observations with the same variance \(\sigma^2\), the variance of their average includes not only individual variances but also covariances between pairs of observations:
If observations are independent, each covariance is zero, and the expression reduces to
This is the variance pattern behind the standard error of the sample mean. If observations have positive covariance, the pairwise covariance terms add to the variance; the mean can be less precise than the usual calculation suggests. A formula using \(s/\sqrt{n}\) does not account for such dependence simply because it has been computed correctly.
Worked Example: Volunteers in an Online Poll
A student researcher wants to estimate the average number of hours slept by all students at a college. A link is posted on a public student forum, and 120 students choose whether to respond. The researcher calculates a one-sample t interval from the 120 reported sleep times. Is the usual interval justified for the college-wide mean by this sampling process?
Random sampling assessment. No. Students entered the data set by choosing to respond, rather than being selected at random from a list of all college students. Students who see the post or feel especially interested in sleep may be more likely to respond. The respondents may therefore differ systematically from students who did not respond.
Independence assessment. The fact that there are 120 different respondents does not, by itself, establish independence. For example, respondents may be connected through the same residence hall or student group. But even if the responses could reasonably be treated as independent, the self-selection problem remains.
Conclusion. The computed interval describes uncertainty under the assumptions of the t procedure; it cannot repair a nonrandom recruitment process. The interval is not justified as a reliable estimate of the mean for all college students on the basis of this design. It may summarize the respondents, but generalizing to the full college population requires more than a large response count.
Worked Example: Measurements Clustered Within Clinics
In a hypothetical project, researchers randomly select eight clinics and record a blood-pressure measurement for five patients at each clinic, for a total of 40 patients. They want an interval for the mean measurement among patients served by the clinics. Can the 40 measurements automatically be treated as 40 independent observations in a one-sample t interval?
Random sampling assessment. The clinics were selected randomly, which is a chance-based first stage. However, we would also need to know how the five patients were chosen within each clinic and whether the selected clinics and patients cover the intended population. Randomly selecting clinics alone does not show that all patients in the target population had an appropriate chance of entering the sample.
Independence assessment. Patients from the same clinic may share staff, procedures, or local circumstances. Their measurements could therefore be more alike than measurements from different clinics. The 40 observations are grouped into eight clusters, and the within-clinic measurements may be dependent.
Conclusion. We should not automatically use \(n=40\) as though all 40 measurements were independent draws from one population. The ordinary one-sample t interval’s standard error does not represent the extra covariance that may result from clinic clustering. The design and dependence need to be addressed before that interval can be justified.
A Practical Assessment
Before using a one-sample t interval, describe the data collection process in enough detail to answer three questions. This assessment concerns the observations themselves, not whether the final interval happens to look plausible.
Specify whose mean the interval is intended to estimate. Check that the sampling frame and recruitment process correspond to that target.
Determine whether selection used a chance-based method, or whether entry depended on convenience, volunteering, or another potentially selective process.
Ask whether measurements are repeated on the same individuals, grouped within settings, or selected in a way that makes one selection affect another. If dependence is plausible, do not assume the usual standard error handles it.
If the selection process does not support generalization to the intended population, limit the conclusion accordingly. If independence is doubtful, the ordinary one-sample t interval may not be appropriate even for a randomly selected group.
Common Mistakes
- Assuming a large sample fixes nonrandom selection. A large number of volunteers can still differ systematically from the target population. More observations reduce some forms of random uncertainty, but they do not remove selection bias.
- Treating random sampling and independence as synonyms. Random selection is about the selection design; independence concerns relationships among observations. Sampling without replacement creates dependence, and clustered data can be dependent even when clusters were selected randomly.
- Assuming random assignment makes a sample representative. Random assignment allocates participants to conditions. It does not, by itself, make the participants representative of a broader population.
- Counting every measurement as a new independent unit. Several measurements from one person or several patients from one clinic may contain less independent information than the same number of observations from unrelated units.
- Believing a correct calculation validates the design. Correct arithmetic does not establish random sampling or independence. Those assumptions must be supported by how the data were collected.
Check Your Understanding
For each question, focus on what the data collection process supports.
- A researcher randomly selects 50 names from a complete employee list, then surveys each selected employee once. What does random selection support, and what additional issue should be considered before using a one-sample t interval?
- A survey has 2,000 responses, all submitted voluntarily through a social-media link. Explain why the number of responses does not by itself establish a random sample.
- Why can measurements from several patients at the same clinic be dependent even if each measurement was recorded separately?
- Explain the difference between random sampling and random assignment, and state which one supports generalizing from a sample to a population.
- What happens to the variance of a sample mean when observations are positively correlated, compared with the independent-observation calculation?