Check Conditions for Mean Procedures

Why Conditions Matter

A t procedure uses sample data to make an inference about a population mean or a difference between population means. Before calculating an interval or test, check whether the data could reasonably produce the sampling behavior that the procedure assumes. A large sample may help with skewed data, for example, but it cannot make dependent observations independent.

The conditions depend on the study design. With one sample, assess the observations used to estimate one mean. With two independent groups, assess each group and whether the groups are independent. With paired data, assess the pairwise differences: those differences are the observations used to estimate the mean change.

Definition: Conditions for mean inference are features of the study design and data that support using a t procedure. They include appropriate independence, no influential outliers, and a distribution shape and sample size that make the procedure's sampling distribution reasonably well approximated by a t distribution.

The Conditions to Check

Begin with the design, not a graph. A t calculation cannot repair a sample that was selected in a way that fails to represent the population of interest, or observations that are dependent in a way the procedure does not account for. Random sampling or random assignment can support independence, but the details matter: repeated measurements on the same person are paired, and observations from people in the same household or group may be related.

Independence. The observations treated as separate data points should be independent, or the design should explicitly account for their relationship. When sampling without replacement from a finite population, a common check is that the sample size is no more than 10% of the population size. In a two-sample procedure, the groups must also be independent of one another. In a paired procedure, pairs should be independent of other pairs; the two values within a pair are not independent observations.
Outliers. Look for observations far from the rest of the data, using a graph such as a dotplot, boxplot, or histogram. Outliers can strongly affect a mean and standard deviation, and therefore the t statistic and interval. A single extreme value is especially concerning when the sample is small.
Distribution shape and sample size. For small samples, the relevant data should be roughly symmetric and without strong skewness or outliers. For larger samples, the sampling distribution of a mean is often approximately normal even if the individual observations are not, provided there is no severe skewness or influential outlier. A sample size around 30 is often used as a rough guide, not as a guarantee or a universal cutoff.

The exact normality condition applies to the population distribution for a one-mean procedure and to each population for a two-sample procedure. For a paired procedure, it applies to the population distribution of the pairwise differences. In practice, the population shape is not directly visible, so sample graphs and the sample size are used to judge whether the condition is plausible.

Check the Data That Match the Procedure

The first decision is whether the data are one sample, two independent samples, or paired. This determines what should be checked for shape and outliers.

Procedure	Data used as observations	Shape and outlier check
One-mean t	Values from one sample	Check the sample values
Two-sample t	Values in each of two independent groups	Check each group separately
Paired t	One difference for each pair	Check the differences

For paired data, the individual measurements within a pair may vary widely, yet the differences can be consistent. The reverse can also happen: each set of measurements may look regular, but the differences may contain an outlier. Since paired t inference is based on the mean difference, its conditions concern the differences.

Key distinction: Independence concerns how observations are related by the design; shape and outliers concern the values used in the chosen procedure. For paired inference, the values used are the within-pair differences, not the two original columns considered separately.

Worked Examples

Worked Example: Checking a One-Mean Procedure

Suppose a hypothetical analyst wants to estimate the mean time, in minutes, that students at a particular college spend commuting to campus. A random sample of 12 students is drawn from a college population of 500. Their commute times are 18, 19, 20, 20, 21, 21, 22, 22, 23, 23, 24, and 25 minutes. Is a one-mean t procedure reasonable?

Independence. The students were randomly sampled. The 10% check is \(12 \leq 0.10(500)=50\), so the sample is no more than 10% of the population. Provided no students were sampled as a dependent cluster, independence is plausible.

Outliers and shape. The values range from 18 to 25 minutes, with observations spread fairly evenly around the center. There is no isolated extreme value or pronounced skew in this small sample. A useful numerical check is the sample mean: the values sum to 258 minutes, so \(\bar{x}=258/12=21.5\) minutes. The data are arranged around this center rather than being dominated by one unusually large or small observation.

Decision. Because \(n=12\) is small, sample size alone does not justify relying on a large-sample approximation. The roughly balanced shape and lack of an apparent outlier support using a one-mean t procedure, conditional on the stated sampling design. If a graph showed strong skewness or a clear outlier, the small sample would make that a serious concern.

Worked Example: Checking Two Independent Groups

Imagine a hypothetical comparison of commute times for students who use a campus shuttle and students who cycle. Two independent random samples are taken, one of 42 shuttle users and one of 38 cyclists. The sample mean and sample standard deviation are 72.4 and 2.4 minutes for shuttle users, and 75.1 and 3.0 minutes for cyclists. Suppose a review of separate group plots shows only mild skew and no isolated extreme values. Are two-sample t methods suitable?

Independence. The groups consist of different students, and the samples were drawn independently. If each group is sampled without replacement from a population of 1,000 or more students, then 42 and 38 are each below 10% of their respective populations. The design supports treating observations within and between groups as independent.

Shape, outliers, and sample size. Check each group separately, because the two-sample procedure uses information from both distributions. The stated plots show no severe skew or extreme values. Both sample sizes exceed 30, which makes the t approximation more robust to mild departures from normality, though it does not excuse serious outliers or severe skew.

The standard error for the estimated difference in sample means, using the two-sample t method that does not require equal population variances, is

SE(\bar{x}_1-\bar{x}_2) =\sqrt{\frac{s_1^2}{n_1}+\frac{s_2^2}{n_2}} =\sqrt{\frac{2.4^2}{42}+\frac{3.0^2}{38}} =\sqrt{\frac{5.76}{42}+\frac{9}{38}} \approx \sqrt{0.37398496} \approx 0.61154\text{ minutes}.

The displayed approximations retain enough precision to support the final rounding to 0.612 minutes. This calculation does not establish the conditions; it illustrates the standard error once the design and data checks make the method reasonable. Given the stated independent sampling, sample sizes, and group plots, two-sample t inference is suitable.

Worked Example: Checking a Paired T Procedure

In a hypothetical study, 10 randomly sampled adults have their resting pulse rates measured before and after a short program. Define each difference as \(d=\text{after}-\text{before}\), in beats per minute. The 10 differences are \(-4,-5,-3,-6,-2,-4,-5,-3,-4,-4\). Is paired t inference reasonable for the population mean change?

Independence. Each adult supplies one pair of measurements, so the before and after values for that adult are intentionally linked. The relevant independence condition is that the 10 adults' differences are independent of one another. The random sampling supports that condition, and, if the population contains 100 or more adults, the sample of 10 also meets the 10% check.

Shape and outliers. Check the list of differences, not the before and after readings separately. The differences lie between \(-6\) and \(-2\) beats per minute and are centered without an isolated value. Their mean is

\bar{d}=\frac{-4-5-3-6-2-4-5-3-4-4}{10} =\frac{-40}{10} =-4\text{ beats per minute}.

There is no apparent extreme difference, and the values are reasonably balanced around \(-4\). Because \(n=10\) is small, that shape check matters; a strong skew or an outlying difference would weaken the case for a paired t procedure. Under the stated design and observed differences, paired t inference is reasonable for the mean change.

Why the Standard Error Uses Independent Observations

The role of independence can be seen from a short derivation. Let \(X_1,\ldots,X_n\) be independent observations with the same population mean \(\mu\) and variance \(\sigma^2\). The sample mean is \(\bar{x}=(X_1+\cdots+X_n)/n\). For independent variables, variances add, so

\operatorname{Var}(\bar{x}) =\operatorname{Var}\left(\frac{X_1+\cdots+X_n}{n}\right) =\frac{1}{n^2}\sum_{i=1}^{n}\operatorname{Var}(X_i) =\frac{n\sigma^2}{n^2} =\frac{\sigma^2}{n}.

Taking the square root gives the standard deviation of the sampling distribution of \(\bar{x}\), \(\sigma/\sqrt{n}\). Since \(\sigma\) is generally unknown, a one-mean t procedure estimates it with the sample standard deviation \(s\), giving \(SE=s/\sqrt{n}\). The same reasoning applies to paired t inference by treating each pair's difference as one observation.

This derivation also explains why dependence is not a minor technical detail. If observations are correlated, the variance of their sum includes covariance terms, so the formula based on independent observations need not be correct. A larger sample does not make those covariance terms disappear simply by being larger.

Common Mistakes and Practical Judgment

Using sample size as a substitute for every other check. “There are more than 30 observations, so the procedure is fine” ignores independence and outliers. A large sample cannot correct a dependent design, and an extreme value can still distort a mean and its standard error.
Checking only the pooled two-group data. One group could be strongly skewed or contain an outlier even if the combined graph looks ordinary. Inspect each independent group separately.
Checking paired measurements instead of differences. Paired t inference uses one difference per pair. Assess the shape and outliers of those differences.
Treating the 10% check as a rule for all designs. The check is used for sampling without replacement from a finite population. It is not a substitute for considering independence in an experiment or other design.
Calling a rough sample-size guideline an absolute cutoff. The number 30 is a convenient rule of thumb, not a theorem that guarantees valid inference. A small, nearly symmetric sample without outliers may be suitable, while a larger sample with extreme problems may not be.

A conditions check is a reasoned judgment, not a mechanical pass/fail checklist. State what the design says about independence, what the relevant graph or observations show about outliers and shape, and how sample size affects the concern. If a major condition is doubtful, do not report a routine t-based result as if the condition had been established.

Check Your Understanding

For each question, identify which condition or data feature matters and explain your reasoning.

A random sample of 24 people is drawn without replacement from a population of 180. Does it meet the 10% condition?
For a two-sample t procedure, one group has a strong right skew and a clear high outlier, while the other group is roughly symmetric. Why is it not enough to inspect only the combined data?
In a before-and-after study, what values should be checked for outliers and distribution shape before using paired t inference?
A study has 80 observations, but several observations come from each household. Why might treating all 80 as independent be inappropriate?
Explain why a sample size greater than 30 does not, by itself, guarantee that a t procedure is suitable.