Why Larger Samples Change the Picture
A population’s measurements do not have to be normally distributed for every procedure about its mean to be useful. With a sufficiently large sample, the sampling distribution of the sample mean can be approximately normal even when individual observations are skewed. This is one reason one-sample t procedures are often reasonably robust to moderate departures from population normality in larger samples.
That robustness has limits. A large sample does not repair dependence between observations, make an infinite-variance population behave normally, or guarantee that an extreme observation is harmless. The relevant question is not simply whether \(n\) is “large,” but whether the sampling conditions and the observed data make the approximation plausible.
How the Sampling Distribution Supports the Procedure
The central limit theorem gives the key large-sample result. Suppose \(X_1,\ldots,X_n\) are independent observations from the same population, with finite mean \(\mu\) and finite, nonzero variance \(\sigma^2\). As \(n\) grows, the standardized sample mean approaches a standard normal distribution:
The notation \(\xrightarrow{d}\) means that the distribution of the quantity on the left approaches the distribution on the right. The result concerns the sample mean across repeated samples; it does not say that the individual observations become normal. In particular, a histogram of the raw data can remain skewed even when the sampling distribution of \(\bar{x}\) is approximately normal.
In practice, \(\sigma\) is usually unknown, so a t procedure replaces it with \(s\). Under the same independent, identically distributed model with finite variance, \(s\) estimates \(\sigma\) consistently: as the sample size grows, \(s\) tends to be close to \(\sigma\). Replacing \(\sigma\) by \(s\) therefore gives the large-sample justification for the t statistic:
The t distribution with \(n-1\) degrees of freedom also approaches the standard normal distribution as \(n\) grows. Thus using a \(t_{n-1}\) reference distribution is asymptotically reasonable under these conditions. This is an approximation for nonnormal populations, not a claim that the statistic has an exact t distribution for every large sample.
For an actual t procedure, other assumptions still matter. Observations should come from a suitable random sample or randomized process, and they should be independent, or sufficiently close to independent for the procedure being used. The observations should measure the same population quantity and have a finite variance. If the data are dependent—for example, measurements are clustered within clinics or repeatedly taken from the same person—the ordinary one-sample t calculation does not account for that dependence. A large \(n\) alone does not fix it.
Worked Example: A Larger Sample With Mild Skewness
Worked Example: A Larger Sample With Mild Skewness
Suppose, in a hypothetical quality-control exercise, 80 independently selected packets are weighed. The measurements are somewhat right-skewed, but a plot shows no severe tail and no isolated observation far from the rest. The sample mean is 102.4 g and the sample standard deviation is 12.8 g. Consider testing \(H_0:\mu=100\) g using the one-sample t statistic.
Check the justification. Assume the packets were selected by a random process, their measurements are independent, and they represent one population with finite variance. The mild skewness and absence of an influential extreme value do not establish these assumptions, but they give no immediate indication of a serious shape problem. With \(n=80\), the sampling distribution of the mean may be approximately normal under the stated model.
Calculate the estimated standard error and statistic. The standard error is \(SE=s/\sqrt{n}\):
The statistic is about 1.677, and the reference t distribution has \(df=80-1=79\). This calculation illustrates how a t statistic can be used when the raw measurements are not perfectly normal: its justification is the approximate sampling distribution of the mean, supported by the model and the sample conditions. This example does not establish a test conclusion or p-value.
When a Large Sample Is Not Enough
The central limit theorem is an eventual result: it describes what happens as sample size increases, but it does not give a universal cutoff that guarantees a good approximation. How quickly the sampling distribution becomes close to normal depends on the population. Mild skewness may be compatible with useful t inference at a moderate sample size; pronounced skewness or very heavy tails can require much more information.
A particularly important limitation is the finite-variance condition in the usual central limit theorem. If the population has extremely heavy tails or infinite variance, the standard justification above does not apply. Even when the population variance is finite, a few very large values can make the sample mean and standard deviation unstable in a particular sample. Therefore, diagnostics remain useful for larger samples, though raw-data normality is not a requirement in the same way it can be for small samples.
Worked Example: A Strongly Right-Skewed Sample
In a hypothetical sample of 36 independent repair times, the mean is 42 minutes and the standard deviation is 55 minutes. Most repairs take between 10 and 50 minutes, but the distribution has a pronounced right tail, including one repair lasting 310 minutes. Consider a one-sample t procedure for a population mean under the model that these are independent observations from one population with finite variance.
Compute the estimated standard error and statistic. For illustration, take the null value to be \(\mu_0=30\) minutes:
The statistic has \(df=35\). However, the calculation does not show that the t approximation is accurate. The sample is not especially large, the observed data are strongly skewed, and one long repair may have substantial influence on both \(\bar{x}\) and \(s\). The assumption of finite population variance also cannot be verified just from these summaries. The right conclusion is caution: the usual large-sample argument is less reassuring here than it was in the first example. One should investigate whether the observation is a valid value and whether the model and sampling process are credible before relying on the t approximation.
Why Influential Observations Matter
An outlier is an observation unusually far from the rest of the data. An observation is influential for a mean procedure when it has a substantial effect on quantities such as \(\bar{x}\), \(s\), or the resulting t statistic. Because the mean and standard deviation both use every value, an extreme observation can affect both parts of the statistic. It can pull the mean toward itself, increase the estimated standard error, or do both. The net change in the t statistic is not predictable from the word “outlier” alone.
The following calculation illustrates the mechanism. It is not evidence that a particular population has any specified distribution. It shows why a large sample size cannot make an individual extreme measurement disappear from the calculations.
Worked Example: One Extreme Value in a Large Sample
In a hypothetical sample of 100 instrument readings, 98 readings are 10 units, one is 9 units, and one is 11 units. The null value is \(\mu_0=10\) units. Suppose the readings are treated as independent observations from a population with finite variance. Before any additional reading is considered, the sample mean is 10 units. The sum of squared deviations from the mean is \(2\), so:
Now suppose a further reading is 30 units. In the resulting sample, \(n=101\), the sum of all readings is \(1030\), and the sum of their squares is \(10902\). The new mean is:
Using the identity \(\sum(x_i-\bar{x})^2=\sum x_i^2-(\sum x_i)^2/n\), the sum of squared deviations is \(10902-1030^2/101\approx398.0396\). Therefore:
The single new value moved the sample mean by about 0.1980 units and increased the estimated standard deviation substantially. The statistic changed from 0 to about 0.998; the outlier did not necessarily make it enormous, because it also increased the estimated standard error. With \(df=100\), large-sample reasoning may be relevant, but it does not settle whether 30 is a valid observation or whether the population model is suitable. The data should be examined in context before relying on an automatic “large \(n\)” rule.
A Practical Assessment
For larger samples, the question is not “Are the raw data normal?” Instead, assess whether the conditions behind the sampling-distribution argument are plausible and whether the sample reveals features that could make the approximation poor. A useful assessment proceeds in this order:
Confirm that the observations represent the intended population and that repeated, paired, or clustered measurements have not been treated as independent observations without justification.
Look for severe skewness, long tails, gaps, or isolated extreme values. Moderate nonnormality is less concerning in a larger sample than severe asymmetry or a highly influential observation.
If the model and data are reasonably compatible with the assumptions, a t procedure may be useful. If dependence, severe skewness, heavy tails, or influential observations are serious concerns, the usual t reference distribution may not be reliable.
Common Mistake: Treating “Large” as a Guarantee
A common mistake is to reason, “The sample size is over 30, so the t procedure is valid regardless of the data.” This applies a fixed cutoff as if it were a theorem and ignores both the sampling design and the population’s shape. The central limit theorem supports an approximation under stated conditions; it does not promise that every sample of a particular size has a nearly normal sampling distribution.
The correction is to treat sample size as one part of the evidence. Check independence and the sampling process, consider whether the population is likely to have finite variance, and inspect the data for severe skewness or influential extremes. A larger sample often reduces the impact of moderate nonnormality, but the strength of that reassurance depends on the situation.
Check Your Understanding
Answer each question using the sampling-distribution argument and its assumptions.
- Why can a sample mean have an approximately normal sampling distribution even when the individual observations are right-skewed?
- What assumptions allow the central limit theorem argument in this tutorial, and why does dependence between observations matter?
- A large sample has one extremely high measurement. Why is it not enough to say that the t procedure must be reliable because \(n\) is large?
- In a t statistic, why can an extreme observation change both the numerator and the estimated standard error?
- Under what population condition does the usual central limit theorem justification described here fail?