Conditions Are the Reason an Inference Is Trustworthy
A calculator can produce a t interval from almost any list of numbers. That does not mean the interval is justified. In the previous tutorial, “Writing the Full Four-Step t Interval Solution,” the Plan step checked how data were collected, whether observations could be treated as independent, and whether the data shape supported a t procedure. This tutorial explains why those checks matter.
A one-sample t procedure uses the sample mean \(\bar{x}\) to estimate a population mean \(\mu\), and uses the sample standard deviation \(s\) to estimate the population standard deviation \(\sigma\). Its interval has the form \(\bar{x}\pm t^*(s/\sqrt{n})\). The formula depends on more than arithmetic: the sampling process and the distribution of the data must make the t-based uncertainty calculation meaningful.
Three ideas work together. Randomness concerns how the sample or experimental units were selected. Independence concerns whether the observations provide separate information. Normality and shape concern whether the t procedure’s reference model is a reasonable approximation for the sample mean. A failure in one area is not repaired by success in another.
Why Randomness Matters
Random selection gives members of the target population a chance to be represented according to a known chance process. That is the basis for using sample data to make claims about the population. As covered in “Sampling Distribution of a Sample Mean Defined,” the sampling distribution describes the values of \(\bar{x}\) across possible samples selected by a specified method. If the actual method systematically favors some kinds of individuals, that distribution need not describe the estimates produced by the method.
For instance, people who volunteer for a survey may differ from people who ignore it. A very large volunteer sample can still overrepresent people with strong opinions. Increasing \(n\) generally reduces random sampling variability when the design supports that conclusion, but it does not automatically remove bias caused by a nonrandom method; “Does a Larger Sample Reduce Bias in the Mean” established this distinction.
Random sampling and random assignment have different purposes. A random sample can support generalizing to the population from which it was selected. Random assignment in an experiment can support a cause-and-effect conclusion for the experimental units. One does not automatically provide the other: randomly assigning volunteers to treatments does not, by itself, make those volunteers representative of everyone.
Why Independence Matters
The usual estimated standard error \(s/\sqrt{n}\) reflects the variability expected when observations contribute independent information. Earlier, in “The 10% Condition for Sample Means,” we used the 10% condition for a simple random sample without replacement from a finite population: \(n<0.10N\). That condition helps justify treating observations as independent in that setting. Other designs require attention to how the observations were obtained.
Dependence can arise when the same person is measured repeatedly, when sampled individuals come in related groups, or when nearby measurements over time tend to resemble one another. If observations move together, the sample does not contain as much independent information as \(n\) separate observations would. The usual standard error may then misstate the uncertainty—often making it too small when observations are positively related. The usual t reference distribution may also fail to describe the statistic’s behavior.
The issue is not that dependent data are useless. It is that the one-sample t method described here may not account for their structure. A calculator’s standard one-sample t procedure treats the entries as independent observations; it cannot detect or adjust for a sampling design or relationship that was not entered into the calculation.
Why the Data Shape Matters
If the observations are independent and come from a Normal population, the sampling distribution of \(\bar{x}\) is Normal for any sample size, and the one-sample t statistic has its stated t distribution. That exact result is one reason Normality matters. When the population is not exactly Normal, t procedures can still work well in many situations, but their accuracy depends on the sample size and how far the data depart from a reasonable shape.
For small samples, a graph of the sample data should not show strong skewness or extreme outliers. With such a small amount of data, a few unusual observations can have a large effect on \(\bar{x}\) and \(s\), and the t-based model may not approximate the behavior of the statistic well. When the sample is larger, the Central Limit Theorem, introduced in “Central Limit Theorem Explained,” helps explain why the sampling distribution of \(\bar{x}\) may be approximately Normal even when the population itself is not. But there is no single sample-size cutoff that makes every shape problem disappear. “Is n = 30 Large Enough for the CLT” emphasized that the population’s skewness and outliers matter.
The t distribution accounts for estimating \(\sigma\) with \(s\), as explained in “Why We Use t Instead of z for Means.” It does not fix biased sampling, dependence, or severe departures from the shape assumptions. A heavier-tailed critical value is not a universal correction for flawed data collection.
Worked Examples
Worked Example: Filter Efficiency in a Random Sample
A fictional testing team selects a simple random sample of 16 air filters from a production run of 500 filters. Under the same test conditions, the sample mean efficiency is 42.0 percentage points and the sample standard deviation is 8.0 percentage points. A graph shows a roughly symmetric pattern with no outliers. Construct and interpret a 90% confidence interval for the mean efficiency of filters in the run, and explain why the conditions support the method.
State. Let \(\mu\) be the true mean efficiency, in percentage points, of all filters in this production run. We will estimate \(\mu\) with a 90% confidence interval.
Plan. Use a one-sample t interval because the goal is to estimate one population mean and \(\sigma\) is unknown. The filters were selected by simple random sampling, supporting inference to the production run. Because \(16<0.10(500)=50\), the 10% condition is met, so treating the observations as independent is reasonable. The sample graph is roughly symmetric with no outliers, supporting a t procedure for this sample size.
Do. Here \(n=16\), so \(df=n-1=15\). The 90% critical value is \(t^*\approx1.7531\). The estimated standard error and margin of error are:
Therefore, the interval is:
Conclude. We are 90% confident that the true mean efficiency of filters in this production run is between approximately 38.494 and 45.506 percentage points. The conditions make the t interval a reasonable method here; they do not mean that 90% of individual filters have efficiencies in this interval.
Worked Example: Volunteer Survey at a Community Center
A fictional community center wants to estimate the mean number of minutes residents spend traveling to the center. Staff post a survey link on the center’s social media account, and 240 residents choose to respond. The mean reported travel time is 22 minutes. Can a one-sample t interval based on these responses be justified as an estimate of the mean travel time for all residents in the area?
State. The target parameter is \(\mu\), the true mean travel time, in minutes, for all residents in the area. The question is whether the survey responses justify a t interval for this population mean.
Plan. A one-sample t interval would require a suitable sampling process, reasonable independence, and data shape that supports the t model. The main concern is the sampling process: respondents chose to click the link, so this is a voluntary-response sample rather than a random sample of area residents. People who use the center or follow its social media account may be more likely to respond than other residents.
Do. The reported mean of 22 minutes is a summary of the 240 respondents. A calculator could also compute a standard error and interval if the standard deviation were supplied, but neither a larger response count nor a t critical value corrects the self-selection in the survey. Since the sample method does not provide a sound basis for generalizing to all residents, a conventional one-sample t interval is not justified for that target population.
Conclude. The response data describe the people who chose to answer, but they do not, by themselves, provide convincing evidence about the mean travel time for all residents in the area. A probability sample or another defensible design would be needed to support that population-level inference.
Worked Example: Consecutive River-Level Readings
A fictional environmental station records a river’s water level once every hour for 25 consecutive hours. The sample mean is 3.8 meters and the sample standard deviation is 0.5 meter. Water levels at nearby hours tend to be similar. A student proposes using a one-sample t interval with standard error \(0.5/\sqrt{25}\) meters to estimate the mean level over a season. What is wrong with this plan?
State. The intended parameter is \(\mu\), the true mean water level, in meters, over the season. The data consist of 25 hourly readings during one continuous period.
Plan. A t interval would require the observations to provide independent information and the time period to represent the seasonal target. The readings are consecutive, and nearby water levels tend to resemble one another. This pattern is evidence of dependence, so the usual one-sample t calculation is not automatically appropriate. Also, a single 25-hour stretch may not represent the range of conditions across the season.
Do. The proposed arithmetic gives \(0.5/\sqrt{25}=0.1\) meter as the nominal standard error under the usual formula. That calculation does not establish that 0.1 meter is the actual uncertainty here: the formula assumes the observations can be treated as independent. Because the readings are linked in time, the stated t procedure may understate uncertainty, and the data-collection window may not support generalization to the whole season.
Conclude. The student should not report this interval as a justified t interval for the seasonal mean without addressing the time dependence and how the observations represent the season. Twenty-five recorded values are not necessarily equivalent to 25 independent observations.
Worked Example: A Small Sample With an Extreme Value
A fictional researcher randomly selects eight households and records their weekly hours of home energy monitoring. The observations, in hours, are \(4, 5, 5, 6, 6, 7, 8, 23\). The researcher wants a t interval for the mean weekly hours among similar households. Is the random sample alone enough to support the procedure?
State. Let \(\mu\) be the true mean weekly monitoring time, in hours, among the target households. The proposed estimate is based on a random sample of eight households.
Plan. Random selection supports the sampling claim, and independence would also need to be checked from the population size or sampling design. However, the sample is small and has one observation, 23 hours, far above the other seven values. That extreme value makes the data strongly right-skewed and raises concern about using a t procedure with only eight observations.
Do. The sample mean is \(64/8=8\) hours, so the value 23 has substantial influence on the center. A t interval can still be calculated from these observations, but the random sample does not remove the shape concern. With so few observations, the Central Limit Theorem does not ensure that the sampling distribution of \(\bar{x}\) is close enough to Normal in the presence of this extreme value.
Conclude. Randomness is important, but it is not sufficient on its own. The researcher should not treat a routine t interval as well supported by this small, strongly skewed sample; the interval’s usual reliability is in doubt. More data collected appropriately, or a method designed for the situation, may be needed.
Common Mistakes and AP Exam Tips
- Claiming that a large sample fixes every problem. Larger \(n\) can help with the sampling distribution’s shape, but it does not remove bias from a voluntary-response sample or make dependent observations independent. Full-credit reasoning identifies which condition is at issue and why.
- Writing “random” without describing the design. Say whether the data came from a random sample or an appropriate randomized process, and connect that fact to the population or process being studied.
- Calling observations independent because they are listed separately. Rows in a data table can still be linked by person, group, place, or time. Explain how the design supports independence, or identify the dependence concern.
- Demanding a perfectly Normal-looking sample. The relevant question is whether the t model is reasonable, not whether every point forms a perfect bell curve. For small samples, strong skewness or outliers are serious warnings; larger samples may tolerate moderate departures, but no universal cutoff guarantees success.
- Treating a calculator result as proof that conditions hold. A calculator performs the requested arithmetic. It does not assess representativeness, independence, or whether an extreme observation undermines the model.
- Confusing a condition failure with a guaranteed wrong answer. A failure means the usual justification is weakened or absent; it does not tell us exactly how far an interval misses the parameter. State what is no longer supported rather than claiming a particular outcome.
On an AP response, condition statements should use evidence from the setting. “The sample is random” is stronger when followed by the actual selection method. “Independent” should be supported by a design argument, such as the 10% condition where appropriate. A shape claim should refer to a graph, a stated Normal population, or the sample size and observed pattern. If a condition is not supported, explain the consequence for the inference instead of silently proceeding as though it were met.
Check Your Understanding
For each question, explain the reasoning in context rather than only naming a condition.
- A random sample of 12 items is drawn without replacement from a shipment of 180 items. Is the 10% condition met? What does that support?
- A random sample of 50 people is drawn from an online poll’s list of volunteers. Does the sample size make the results representative of all residents? Explain.
- A researcher measures each of 15 plants on three consecutive days and treats the 45 measurements as independent. Identify the concern and explain how it could affect the usual standard error.
- A small random sample has an approximately symmetric dotplot with no outliers. What does this evidence say about the shape condition for a t procedure?
- Why does a calculator’s successful t-interval output not establish that the interval is valid for the intended population?