Historical Context & Motivation
Statistical inference rests on a deceptively simple question: given a finite sample from a larger population, how confidently can we estimate the true population parameter? By the late nineteenth century, mathematicians had established the normal distribution as a cornerstone of probability theory, and the Central Limit Theorem guaranteed that sample means would follow an approximately normal shape for sufficiently large samples. However, a critical gap remained: these results presumed that the population standard deviation σ was known, a condition rarely satisfied in practice. In most real-world experiments—especially in agriculture, medicine, and quality control—researchers work with small samples and must estimate σ from the data itself, introducing additional uncertainty that the standard normal curve fails to capture.
Gosset's original problem was strikingly practical: with only a handful of barley yield measurements, how could he estimate the true mean yield and attach a meaningful measure of precision? The z-interval, which relies on knowing σ, systematically understated the true uncertainty in small samples, producing intervals that were too narrow and overconfident. The t-interval resolved this by substituting the sample standard deviation s for σ and using a distribution with heavier tails—the t-distribution—to account for the extra variability introduced by estimating σ. This section of the course equips you to construct and interpret such intervals, one of the most frequently used tools in applied statistics.
Core Principles & Definitions
Before constructing a t-interval, it is essential to understand the conceptual machinery that makes the procedure valid. A confidence interval is not a probability statement about a single fixed parameter; rather, it describes a procedure that, when repeated across many independent samples, captures the true parameter a specified percentage of the time. The confidence level (commonly 90%, 95%, or 99%) quantifies the long-run success rate of this procedure. Five foundational ideas underpin the t-interval for a mean.
Point Estimate: x̄
Standard Error: s / √n
Degrees of Freedom: df = n − 1
Critical Value: t*
Margin of Error: E = t* · (s / √n)
Visual Explanation — The t-Distribution and Interval Construction
The visual below contrasts the standard normal distribution (z) with t-distributions at various degrees of freedom. Notice how the t-distribution has heavier tails than the normal curve, especially when df is small. This additional probability mass in the tails reflects the extra uncertainty from estimating σ with s. As degrees of freedom increase, the t-distribution gradually approaches the standard normal, which is why the z-interval and t-interval produce nearly identical results for large samples.
The practical implication is immediate: for a given confidence level, the critical value t* will always be larger than z*, producing a wider interval. With df = 10 and a 95% confidence level, t* ≈ 2.228 compared to z* = 1.960—a difference of about 14%. With df = 3, t* balloons to 3.182. This widening is not a deficiency; it is the honest acknowledgment that estimating σ from a small sample introduces genuine uncertainty that must be reflected in the interval.
Mathematical Framework
The derivation of the t-interval begins with the sampling distribution of the mean. When a random sample of size n is drawn from a population with mean μ and standard deviation σ, the quantity Z = (x̄ − μ) / (σ / √n) follows a standard normal distribution. In practice, σ is unknown and is replaced by s, the sample standard deviation. The resulting quantity follows the t-distribution with n − 1 degrees of freedom.
Inverting this relationship—solving for μ—yields the confidence interval formula. We seek the range of μ values for which the observed t-statistic would fall between −t* and +t*, where t* is the critical value corresponding to the desired confidence level C.
Conditions, Mechanics, and the t-Table
In practice, constructing a t-interval involves a systematic sequence: verify conditions, compute the standard error, look up the critical value t* from a t-table (or software), calculate the margin of error, and form the interval. The diagram below illustrates this workflow and highlights the decision points where you choose between a z-interval and a t-interval.
Excerpt from the t-Table
The table below shows selected critical values t* for common confidence levels and degrees of freedom. In practice, you may also use technology—R's qt() function, Python's scipy.stats.t.ppf(), or a TI calculator's invT() command—to obtain exact values for any df.
| df | 90% (t*) | 95% (t*) | 99% (t*) |
|---|---|---|---|
| 3 | 2.353 | 3.182 | 5.841 |
| 5 | 2.015 | 2.571 | 4.032 |
| 10 | 1.812 | 2.228 | 3.169 |
| 20 | 1.725 | 2.086 | 2.845 |
| 30 | 1.697 | 2.042 | 2.750 |
| ∞ (z*) | 1.645 | 1.960 | 2.576 |
Worked Example — Estimating Mean Study Hours
A university researcher randomly selects 16 students and records the number of hours each student studies per week. The sample yields a mean of x̄ = 14.5 hours and a sample standard deviation of s = 4.2 hours. Construct a 95% confidence interval for the true mean weekly study hours μ of the student population.
z-Interval vs. t-Interval — Strengths and Limitations
Understanding when to use the z-interval versus the t-interval is a common source of confusion in introductory statistics courses. The following table provides a side-by-side comparison of the two procedures across several important dimensions. In virtually all practical research scenarios, the t-interval is the appropriate choice because σ is almost never known.
| Feature | z-Interval | t-Interval |
|---|---|---|
| σ known? | Yes — σ is known or given | No — σ is estimated by s |
| Distribution used | Standard normal (z) | t-distribution with df = n − 1 |
| Critical value | z* (e.g., 1.960 for 95%) | t* (depends on df; always ≥ z*) |
| Interval width | Narrower for a given n | Wider — accounts for estimating σ |
| Sensitivity to n | Width shrinks as 1/√n | Width shrinks as 1/√n, plus t* → z* as df → ∞ |
| Normality requirement | Population normal or n large (CLT) | Same, but more sensitive for small n |
| Practical use | Rare — mostly textbook exercises | Standard in applied research |
Limitations of the t-Interval
- Sensitivity to outliers: Both x̄ and s are heavily influenced by extreme values. A single outlier can dramatically inflate s, widening the interval, or shift x̄, mislocating the interval entirely.
- Normality assumption: For very small samples (n < 15), the t-interval's validity depends strongly on the population being approximately normal. Strongly skewed or heavy-tailed distributions can cause the actual coverage rate to differ substantially from the nominal confidence level.
- Single population only: The one-sample t-interval addresses a single mean. Comparing two means requires the two-sample t-interval or paired t-interval, with their own assumptions.
Connections to Advanced Theory
The one-sample t-interval is the gateway to a family of inferential procedures that permeate advanced statistics. Understanding its structure—point estimate plus or minus a critical value times a standard error—reveals a template that generalizes broadly. The table below maps the one-sample t-interval to several advanced extensions you will encounter later in your studies.
| Procedure | Parameter | Key Difference from One-Sample t |
|---|---|---|
| Paired t-interval | μ_d (mean of paired differences) | Applied to difference scores; same formula with d̄ and s_d |
| Two-sample t-interval | μ₁ − μ₂ | Uses pooled or Welch SE; df calculated via Satterthwaite approximation |
| Regression CI for β₁ | Slope coefficient β₁ | SE derived from residuals; df = n − 2 |
| Bootstrap CI | Any parameter | Nonparametric; uses resampling instead of t-distribution; no normality assumption |
| Bayesian credible interval | μ (posterior) | Incorporates prior beliefs; interpretation is a direct probability statement about μ |
The relationship between confidence intervals and hypothesis tests is particularly worth noting. A 95% confidence interval for μ contains all values of μ₀ for which a two-sided t-test at α = 0.05 would fail to reject the null hypothesis H₀: μ = μ₀. This duality means that constructing a confidence interval simultaneously answers every possible two-sided hypothesis test at the corresponding significance level—a powerful conceptual unification that deepens as you study ANOVA, regression, and general linear models.
Practice Problems
Summary — Confidence Interval for a Mean (t-interval)
The t-interval for a population mean is the standard inferential tool when the population standard deviation σ is unknown—the typical situation in practice. The formula x̄ ± t* · (s / √n) combines the point estimate x̄ with a margin of error that accounts for both sampling variability (through s / √n) and the additional uncertainty of estimating σ (through the t-distribution's heavier tails relative to the normal curve). The degrees of freedom (df = n − 1) govern the shape of the t-distribution: smaller df means wider intervals, and as df → ∞, the t-interval converges to the z-interval.
Three conditions must be verified before construction: the data must come from a random sample, observations must be independent (10% condition for sampling without replacement), and the population should be approximately normal or the sample size large enough for the CLT to apply. Interpretation is crucial: a 95% confidence level means that 95% of all intervals constructed by this method would capture the true μ—it is a statement about the procedure's long-run reliability, not a probability about any single interval. This foundational tool generalizes to paired t-intervals, two-sample t-intervals, and regression inference, making it an essential building block for the rest of the statistics curriculum.