Normality Requirements For Exact t Intervals

Why the Assumptions Matter

A one-sample Student t interval is often used when a population standard deviation is unknown. Its familiar form is easy to calculate, but its exact coverage is not automatic. Exactness comes from a particular probability model: the observations must be independent and normally distributed with a common mean and variance. If that model is doubtful, the same calculation may still be useful, but a nominal 95% interval is not thereby guaranteed to cover the population mean 95% of the time.

The issue is especially important with small samples. A few unusual observations can substantially affect both the sample mean and sample standard deviation, while a strongly skewed population may produce samples that do not resemble the normal model. This tutorial is aligned with the SOA VEE Mathematical Statistics syllabus.

Definition: Let $X_1,\ldots,X_n$ be independent observations from the same population $N(\mu,\sigma^2)$, where $\mu$ and $\sigma^2$ are unknown and $n\geq 2$. Define the sample mean and sample variance by $$ \bar X=\frac{1}{n}\sum_{i=1}^n X_i,\qquad S^2=\frac{1}{n-1}\sum_{i=1}^n(X_i-\bar X)^2. $$ For a confidence level $1-\alpha$, the usual two-sided one-sample t interval for $\mu$ is $$ \bar X\ \pm\ t_{1-\alpha/2,n-1}\frac{S}{\sqrt n}, $$ where $t_{1-\alpha/2,n-1}$ is the $1-\alpha/2$ quantile of the Student t distribution with $n-1$ degrees of freedom. Under the stated model, its repeated-sampling coverage is exactly $1-\alpha$.

Why the Interval Is Exact Under Normality

The formula is not simply a normal interval with an estimated standard deviation substituted into it. Its exact coverage follows from the distribution of a specific pivot—the standardized quantity whose distribution does not depend on the unknown parameters.

Standardize the sample mean.
Under the model $X_i\sim N(\mu,\sigma^2)$ independently, $(\bar X-\mu)/(\sigma/\sqrt n)$ has a standard normal distribution.

Describe the sample variance.
For a normal sample, $(n-1)S^2/\sigma^2$ has a chi-square distribution with $n-1$ degrees of freedom. It is independent of $\bar X$.

Form the ratio.
The Student t distribution is the distribution of a standard normal variable divided by the square root of an independent chi-square variable divided by its degrees of freedom.

Combining those facts gives the pivot

T=\frac{\bar X-\mu}{S/\sqrt n} =\frac{(\bar X-\mu)/(\sigma/\sqrt n)} {\sqrt{\{(n-1)S^2/\sigma^2\}/(n-1)}} \sim t_{n-1}.

Therefore, under the model,

P\left(-t_{1-\alpha/2,n-1}\leq \frac{\bar X-\mu}{S/\sqrt n} \leq t_{1-\alpha/2,n-1}\right)=1-\alpha.

Rearranging the inequalities to isolate $\mu$ gives the stated interval. This is why independence and normality matter: they give the numerator and variance estimate the particular distributions, and independence, needed for the t pivot. If these conditions fail, the pivot need not have a Student t distribution, so the displayed probability need not equal $1-\alpha$.

Worked Example: Computing an Exact Interval Under the Model

A hypothetical laboratory measures the concentration of a substance in five independently selected samples. Assume explicitly that the measurements are independent and identically distributed as $N(\mu,\sigma^2)$, with both parameters unknown. The observed concentrations, in units per litre, are $8,9,10,11,12$. Find the exact 95% t interval for $\mu$.

Method. Use the one-sample t interval with $n-1=4$ degrees of freedom. The normal, independence, and common-variance assumptions are given, so the interval has exact 95% coverage under this model.

The sample mean is

\bar x=\frac{8+9+10+11+12}{5}=\frac{50}{5}=10.

The deviations from 10 are $-2,-1,0,1,2$, so their squared sum is $4+1+0+1+4=10$. Thus

s^2=\frac{10}{5-1}=2.5,\qquad s=\sqrt{2.5}=1.5811,\qquad \frac{s}{\sqrt 5}=\sqrt{0.5}=0.7071.

The 97.5th percentile of $t_4$ is approximately $2.776$. The margin of error is $2.776(0.7071)=1.963$. The interval is $10\pm1.963$, or $(8.037,11.963)$, rounded to three decimal places.

Check. The interval is centered at the sample mean 10, and its half-width is $1.963$. The method avoids the mistake of using a standard normal critical value when $\sigma$ is unknown and the specified exact procedure uses $S$.

What Changes When the Population Is Not Normal?

If the observations are independent with a common mean and finite variance but are not normally distributed, the usual t interval generally does not have exact finite-sample coverage. The sample mean and sample variance still exist when the relevant moments exist, but the exact normal-sample pivot derivation no longer applies. A t interval may have useful approximate coverage in some settings, especially with a sufficiently large sample and a population that is not severely skewed or heavy-tailed. That is an approximation, not an exact result, and sample size alone does not guarantee its quality.

Three features deserve particular attention:

Small sample size: With few observations, the sample may not adequately reflect the population’s shape. In particular, a small sample from a skewed population can remain quite asymmetric.
Skewness: A long tail can make the sample mean’s sampling distribution asymmetric, especially for small $n$. A symmetric t-based interval may then have unequal tail errors or coverage below its nominal level.
Outliers and heavy tails: An extreme value can move $\bar X$ and inflate $S$. The resulting interval can be much wider, but that widening does not restore the exact t distribution or guarantee nominal coverage.

A large sample often makes inference for a mean more stable under suitable conditions, but “large” is not a universal cutoff. Strong dependence, infinite variance, very heavy tails, or a small number of extreme observations can still cause trouble. The t formula should not be described as exact merely because a sample size seems reasonably large.

Worked Example: Recognizing the Effect of an Outlier

A hypothetical assessor records six independent claim amounts, in thousands of dollars: $1,1,2,2,3,15$. Calculate the conventional 95% t interval for the mean, then explain what the calculation does and does not establish. No normal population model is supplied.

Method. Calculate the sample summary and apply the t formula as a descriptive interval calculation. Because no normal model is given, do not call its coverage exact.

The sum is $24$, so $\bar x=24/6=4$. The deviations from 4 are $-3,-3,-2,-2,-1,11$, and their squared sum is

$$ 9+9+4+4+1+121=148. $$

Therefore,

s^2=\frac{148}{6-1}=29.6,\qquad s=\sqrt{29.6}=5.4406,\qquad \frac{s}{\sqrt6}=\sqrt{4.9333}=2.2211.

For $5$ degrees of freedom, $t_{0.975,5}\approx2.571$. The margin is $2.571(2.2211)\approx5.710$, giving $(4-5.710,4+5.710)=(-1.710,9.710)$ thousand dollars, rounded to three decimals.

Interpretation and check. The observation 15 is far from the other five and contributes $121$ of the total squared deviation $148$. It strongly affects the estimated spread and hence the interval. The calculation is arithmetically correct, but the data alone do not establish exact 95% coverage. Nor is deleting the value a valid correction without an independent, defensible reason to treat it as an error or as outside the target population. The common mistake avoided here is equating “used a t critical value” with “has exact t coverage.”

A Small-Sample Example with Skewed Data

A discrete population makes the finite-sample risk especially clear. Consider a population in which an observation is 4 with probability $0.2$ and 0 with probability $0.8$. Its mean is $0.2(4)+0.8(0)=0.8$. This population is strongly right-skewed. The following calculation shows that a nominal t interval can miss its target very frequently when the sample is small.

Worked Example: Exact Coverage Fails for a Skewed Population

Take a sample of $n=4$ independent observations from the population just described, and form the conventional two-sided 95% one-sample t interval. Find its actual coverage for $\mu=0.8$. Use $t_{0.975,3}=3.182$.

Method. Let $K$ be the number of observations equal to 4. Then $K\sim\operatorname{Binomial}(4,0.2)$. Assess whether the calculated interval contains 0.8 for each possible value of $K$.

If $K=0$, all four observations are 0, so $\bar x=0$ and $s=0$. The formula gives the degenerate interval $[0,0]$, which misses 0.8. If $K=4$, the interval is $[4,4]$, which also misses 0.8.

For $K=1$, $\bar x=1$. The squared deviations sum to $12$, so $s^2=12/3=4$, $s/\sqrt4=1$, and the interval is $1\pm3.182=(-2.182,4.182)$, which contains 0.8.

For $K=2$, $\bar x=2$. The squared deviations sum to $16$, so $s^2=16/3$, and $s/\sqrt4=\sqrt{4/3}\approx1.155$. The interval is approximately $2\pm3.674=(-1.674,5.674)$, which contains 0.8.

For $K=3$, $\bar x=3$, the squared deviations sum to $12$, and the interval is $3\pm3.182=(-0.182,6.182)$, which contains 0.8. Thus coverage occurs exactly when $K=1,2,$ or $3$. The coverage probability is

P(1\leq K\leq3) =1-P(K=0)-P(K=4) =1-(0.8)^4-(0.2)^4 =1-0.4096-0.0016 =0.5888.

Check. The actual coverage is $0.5888$, not $0.95$. The calculation shows why a small sample from a skewed population can make the usual t interval unreliable. The zero sample variance in the two extreme cases yields degenerate intervals; it does not create valid 95% coverage.

Independence Is Also Essential

Normal marginal distributions by themselves are not enough. The observations must also be independent for the usual t pivot. For example, suppose measurements share a random common component: $X_i=\mu+A+\varepsilon_i$, where $A$ has variance 1 and the independent errors $\varepsilon_i$ each have variance 1, with $A$ independent of all the errors. Each observation has variance 2, but two distinct observations have covariance 1 because they share $A$. Their correlation is $1/2$, not 0.

For $n=4$, the variance of the sample mean is

\operatorname{Var}(\bar X) =\operatorname{Var}(A)+\operatorname{Var}\left(\frac14\sum_{i=1}^4\varepsilon_i\right) =1+\frac14=1.25.

The usual independent-observation standard error, based on the marginal variance 2, would be $\sqrt{2/4}=\sqrt{0.5}\approx0.707$. The actual standard deviation of the mean in this model is $\sqrt{1.25}\approx1.118$. Thus treating these dependent measurements as independent understates the sampling variability. The standard one-sample t interval is not justified by normal marginals alone.

Common Exam Traps and a Practical Assessment

When deciding whether exact t reasoning applies, check the model before calculating or interpreting an interval. In particular, distinguish a statement about exact coverage from a statement that an approximation may be reasonable.

Wrong: “The population is not specified as normal, but $n$ is at least 30, so the t interval is exact.” Correction: A large-sample argument may support approximate coverage under suitable conditions; it does not establish the exact normal-sample pivot.
Wrong: “Each observation is normally distributed, so the usual t interval is exact.” Correction: Check independence and a common mean and variance as well. Dependence can change the sampling variability of $\bar X$.
Wrong: “An outlier makes the t interval conservative because it increases $S$.” Correction: A larger $S$ may widen the computed interval, but it does not prove conservative coverage. The interval’s coverage depends on the sampling distribution of the entire pivot.
Wrong: “A realized interval either has a 95% probability of containing $\mu$, or the procedure is wrong.” Correction: The 95% statement describes long-run coverage across repeated samples under the model. Once the data are observed, the interval either contains the fixed parameter or it does not.

A useful assessment sequence is to identify the observational unit, decide whether observations can reasonably be treated as independent, examine the population model or sample shape for severe skewness and outliers, and then state whether the claimed coverage is exact or approximate. This does not replace a formal model, but it prevents a formula from being mistaken for its assumptions.

Key takeaway: The one-sample t interval has exact finite-sample coverage when the observations are independent draws from a common normal population. With nonnormal data it may be a useful approximation, but skewness, outliers, small samples, and dependence can make its actual coverage differ substantially from the nominal level.

Check Your Understanding

For each problem, distinguish the interval calculation from the conditions that justify its coverage.

A sample of 9 independent observations comes from $N(\mu,\sigma^2)$, with unknown parameters. Which degrees of freedom are used, and is the usual two-sided t interval exact under this model? (Answer: 8 degrees of freedom; yes, it is exact under the stated independent normal model.)
A researcher uses 12 observations from a strongly right-skewed population with one exceptionally large value and reports a 95% t interval. Can the researcher claim exact 95% coverage from the t formula alone? (Answer: No; the formula alone does not establish exact coverage, and the skewness and extreme value warrant caution.)
Let $X$ equal 1 with probability 0.1 and 0 otherwise. For a sample of 3, what is the probability that all observations are zero, and why does that event challenge the t interval? (Answer: $0.9^3=0.729$; the sample variance is zero and the resulting degenerate interval at 0 misses the population mean 0.1.)
Four measurements on one item share a common random effect, so the measurements are dependent. Is the usual one-sample t interval justified merely because each measurement is normally distributed? (Answer: No; dependence violates a required condition for the usual exact t pivot.)
For observations $2,4,4,6$, calculate the 95% t interval using $t_{0.975,3}=3.182$. State the condition under which its coverage is exact. (Answer: $\bar x=4$, $s^2=8/3$, $s/\sqrt4=0.8165$, margin $=2.598$, so the interval is $(1.402,6.598)$ to three decimals; exact coverage requires independent observations from a common normal population.)

What You'll Learn