Robust Does Not Mean “Works No Matter What”
A t procedure can often give trustworthy results even when the population is not exactly Normal. This ability to remain reasonably reliable under some departures from its assumptions is called robustness. It is especially useful because real data rarely follow a perfectly Normal pattern.
Robustness has limits. A t procedure is generally more tolerant of modest skewness when the sample is larger than it is of strong skewness or outliers in a small sample. It is not a remedy for biased data collection or dependent observations. As in “Checking the Normal/Large Sample Condition” and “What to Check With a Small Sample of 12 Observations,” use sample size and graphs to assess shape. Then consider randomness and independence separately.
For mean inference, the Normal/Large Sample condition is supported when the population is approximately Normal, or when the sample is large enough for the sampling distribution of \(\bar{x}\) to be approximately Normal. In AP Statistics, \(n\geq30\) is the large-sample route. For a smaller sample, inspect the data for strong skewness and pronounced outliers, as practiced in the preceding tutorials.
The large-sample route makes t procedures more tolerant of moderate departures from Normality; it does not mean every shape concern disappears at \(n=30\). A few mild irregularities in an otherwise balanced distribution are different from a severe tail or an extreme observation. With a small sample, a t procedure is most defensible when the data show no strong skewness or pronounced outlier. These are judgments based on evidence, not a rule that every departure automatically makes a procedure unusable.
Which Departures Matter Most?
The main shape concerns are strong skewness and outliers. Both can have a substantial effect on a sample mean and sample standard deviation. With a small sample, one extreme observation can pull \(\bar{x}\) toward itself and inflate \(s\), while also making the sample’s shape a poor match for the Normal model underlying t inference. A t procedure does not automatically correct for that effect.
Moderate skewness without an extreme observation is more likely to be tolerable when the sample is large. The sample mean’s sampling distribution tends to be more nearly Normal as the sample size grows, provided observations are independent and the population has a finite mean and standard deviation. Still, a sample size is not a substitute for looking at the data. If the graph shows a particularly extreme value or severe skewness, describe that concern rather than asserting that sample size settles the matter.
The other conditions are different. Random selection supports generalizing a result to the population from which the sample was selected. Independence supports the usual calculation of uncertainty; when sampling without replacement from a finite population, the 10% condition is used to assess independence. A large \(n\) does not turn a volunteer sample into a random sample or remove dependence among observations.
Random assignment has a separate role. It supports a cause-and-effect conclusion in an experiment; it does not by itself make the participants representative of a broader population. If a study uses both random selection and random assignment, they support different parts of the inference: selection helps generalization, while assignment helps causal interpretation. Neither type of randomization should be credited with the other’s role.
A Practical Way to Judge Robustness
When deciding whether a t procedure is reasonable, do not ask only, “Is the population exactly Normal?” Instead, identify what evidence is available and how serious each concern is. Keep the following checks distinct.
Determine whether there was a random sample or an appropriate randomized process. Random selection supports a population claim; random assignment supports a cause-and-effect conclusion.
Consider whether observations are independent. For a sample without replacement from a finite population, apply the 10% condition, as explained in “Checking the 10% Condition for Independence.”
Use sample size and graphs together. For a small sample, look for strong skewness or pronounced outliers. With a larger sample, moderate non-Normality is generally less concerning, but still describe notable features.
Say whether the evidence supports using the t procedure and name any limitation. Do not claim that a graph proves Normality or that a large sample fixes every condition.
This approach is a judgment, not a mechanical pass-or-fail checklist. When the sample is small and the graphs show clear skewness or an outlier, the shape evidence does not support relying on a t procedure. If the pattern is only mildly irregular, especially in a larger sample without influential outliers, a t procedure may still be reasonable. Be specific about the evidence rather than making an unqualified claim that the procedure is “robust.”
Worked Examples
Worked Example: Moderate Skewness in a Larger Random Sample
A parks department randomly selects 36 trail segments from a large network and measures the number of minutes needed to clear a standard section after a light snowfall. The sample has \(\bar{x}=52.4\) minutes and \(s=6.0\) minutes. A histogram shows a modest right tail but no pronounced outlier. Suppose the department wants a 95% confidence interval for the mean clearing time across the trail segments represented by the sample.
State. Let \(\mu\) be the mean clearing time, in minutes, for the population of trail segments from which the random sample was selected. We will use a one-sample t interval to estimate \(\mu\).
Plan. The sample was randomly selected, supporting generalization to that population. The network has more than 360 segments, so 36 is no more than 10% of the population; the 10% condition supports treating observations as independent. The sample size is \(n=36\), which meets the AP large-sample route \(n\geq30\). The histogram shows only modest skewness and no pronounced outlier, so the shape evidence does not raise a strong objection to the t procedure.
Do. The degrees of freedom are \(df=n-1=35\). For a 95% confidence interval, \(t^*\) is approximately 2.030. The estimated standard error is:
The interval is:
Thus, the 95% confidence interval is approximately \((50.37,\ 54.43)\) minutes, rounded to two decimal places.
Conclude. We are 95% confident that the interval from 50.37 to 54.43 minutes contains the true mean clearing time for the population of trail segments represented by the random sample. The modest right skew does not by itself prevent using the t interval here: the sample is large, and the histogram shows no pronounced outlier. This conclusion depends on the stated sampling and independence conditions as well as the shape evidence.
Worked Example: Severe Skewness and an Extreme Value in a Small Sample
A technician randomly selects 12 pumps from a warehouse shipment and records how many hours each can run before its next maintenance check. The observations, in hours, are:
\(42,\ 45,\ 47,\ 48,\ 49,\ 51,\ 53,\ 55,\ 58,\ 64,\ 82,\ 210\)
A dotplot shows most values between 42 and 64 hours, followed by two larger observations, with 210 far from the main cluster. The Normal probability plot bends at the upper end.
State. Let \(\mu\) be the mean operating time, in hours, for pumps in the shipment. We are assessing whether the sample supports a one-sample t procedure for \(\mu\).
Plan. The pumps were randomly selected, which supports generalizing to the shipment. The shipment contains more than 120 pumps, so 12 is no more than 10% of its size; the 10% condition supports independence. But \(n=12<30\), so the large-sample route is not met. We must rely on the small-sample shape evidence.
Do. The graph displays a pronounced upper tail and an observation far above the main cluster. The Normal probability plot’s upper-end bend agrees with the dotplot’s evidence of strong right-skewness. These are substantial departures, not merely minor irregularities. A calculator could still produce a t interval, but the calculation alone would not make the shape condition reasonable.
Conclude. The sample does not provide reassuring shape evidence for a t procedure for the mean operating time. The random selection and 10% condition do not overcome the severe skewness and extreme value in this small sample. The technician should verify that 210 is recorded correctly; if it is a valid observation, it should not be removed solely to make the distribution more Normal.
Worked Example: A Large Volunteer Sample Is Still Biased
A neighborhood website asks residents to volunteer their daily screen time for a survey. A total of 200 people respond from a neighborhood of 5,000 residents. The sample mean is 14.2 hours per day, and the sample standard deviation is 3.1 hours. The website proposes a one-sample t interval for the mean daily screen time of all neighborhood residents.
State. The target parameter is \(\mu\), the mean daily screen time, in hours, for all residents of the neighborhood. The proposed procedure is a one-sample t interval.
Plan. Numerically, 200 is less than 10% of 5,000, or 500, but because the respondents volunteered rather than being randomly selected, the 10% condition does not establish independence for this sample. A sample of 200 also meets the large-sample route for shape. However, the residents volunteered rather than being randomly selected. Their screen-time habits may differ from those of residents who chose not to respond. The random condition needed to support generalization is therefore not established.
Do. The reported values allow a calculator to produce a t interval, but neither the large sample nor the 10% condition addresses the volunteer-selection problem. The interval’s calculation describes uncertainty under the procedure’s assumptions; it cannot account for possible systematic differences between volunteers and the neighborhood as a whole.
Conclude. The t interval should not be used to make a reliable population claim about all neighborhood residents from this volunteer sample. The data may describe the respondents, but a large sample size does not remove possible selection bias. If residents had instead been randomly selected, that would support generalizing to the neighborhood; random assignment, which is used in experiments, would not solve this sampling problem.
Common Mistakes and AP Exam Tips
- Saying “t is robust” without naming the departure. Explain whether the concern is mild skewness, severe skewness, an outlier, biased selection, or dependence. A full-credit answer connects the specific evidence to the procedure.
- Treating \(n\geq30\) as a cure-all. The large-sample route supports an approximately Normal sampling distribution, but it does not repair biased selection or dependence. It also does not make an extreme observation disappear.
- Rejecting every small departure from Normality. A sample need not look perfectly Normal. Describe whether the evidence shows modest irregularity or a pronounced problem, especially for a small sample.
- Confusing random selection with random assignment. Random selection supports generalizing to a population. Random assignment supports cause-and-effect conclusions in an experiment. Do not claim that random assignment alone makes a sample representative.
- Assuming the calculator verifies the conditions. A calculator can compute an interval from entered values even when the data do not support using it. Check the sampling method, independence, and shape before interpreting the result.
- Deleting an inconvenient observation without justification. Check whether a suspicious value is a recording or measurement error. If it is valid, discuss its effect; do not remove it just to improve the graph.
A strong AP response uses evidence and appropriately cautious wording. For example: “The random sample and 10% condition support the sampling and independence requirements. Although the histogram shows modest right skewness, \(n=36\) meets the large-sample route and there is no pronounced outlier, so a t procedure is reasonable. This does not mean the procedure would be appropriate for a severely skewed sample with an extreme value.” When a condition is not supported, say which one and explain what that limitation prevents you from concluding.
Check Your Understanding
For each situation, decide what the evidence says about using a t procedure and explain the reason.
- A random sample of 42 observations has modest right skewness and no pronounced outlier. Which feature of the t procedure’s robustness is relevant?
- A random sample of 10 has strong left skewness and one observation far below the others. Why is the t procedure’s robustness a concern here?
- A volunteer survey has 500 responses from a population of 20,000. Does the large sample establish that a t interval generalizes to the whole population? Explain.
- In one sentence, distinguish what random selection supports from what random assignment supports.
- A calculator reports a t interval even though the sample is small and contains a valid extreme observation. Why is the reported calculation not enough to justify the interval?