Statistics › Section H · Confidence intervals › Topic H.3
Interval for a mean with the t distribution
The previous topic's formula needed one thing real studies rarely have: a known population spread. When the sample's s must stand in for σ, the interval grows slightly wider and the normal's z* gives way to the t distribution's critical values - honesty, priced in degrees of freedom.
The realistic case: σ unknown
The known-\(\sigma\) interval was clean but rare. In most studies the population spread must be estimated from the very sample being used, with \(s\) standing in for \(\sigma\). That substitution adds a second layer of uncertainty - the mean is estimated AND the spread is estimated - and pretending the normal's \(z^*\) still covers both would produce intervals that miss too often. The fix has a name and a story: the \(t\) distribution, invented for exactly this situation.
The t distribution and its story
In 1908, William Sealy Gosset - a chemist at the Guinness brewery who needed trustworthy intervals from small batches of barley - derived the corrected distribution and published it under the pen name "Student". The \(t\) distribution looks like the standard normal but with heavier tails, and its shape depends on one parameter: the degrees of freedom, \(\text{df} = n - 1\) for a one-sample interval. Small samples give heavy tails and tall critical values; as df grows, \(t\) collapses onto the normal, and by df in the dozens the two are nearly the same curve.
The one-sample t interval
The formula swaps two letters into the known-\(\sigma\) version:
with \(t^*\) the critical value at \(n - 1\) degrees of freedom. Because \(t^*\) exceeds \(z^*\) for small samples, the interval is wider than its normal cousin - the honest price of estimating \(\sigma\) from the same data. At \(\text{df} = 24\), the 95% critical value is \(t^* \approx 2.064\), against \(z^* = 1.96\): a widening that fades as samples grow.
Conditions for t procedures
Three conditions license the \(t\) interval. Random sample, and independent observations - as always. Shape: the sampling distribution must be roughly normal. For small samples (under about 15), that means the data themselves should look roughly symmetric without outliers; for moderate samples, mild skew is tolerable; for \(n \ge 30\), the central limit theorem usually settles the question on its own. The conditions are really one condition - trust the bell for \(\bar{x}\) - graded by how much the CLT has had to work with.
Worked example: lightbulb lifetimes
A sample of \(n = 25\) lightbulbs gives \(\bar{x} = 1200\) hours and \(s = 100\) hours. The 95% interval: 1. Degrees of freedom: \(25 - 1 = 24\), so \(t^* \approx 2.064\). 2. Standard error: \(100/\sqrt{25} = 20\) hours. 3. Margin: \(2.064 \times 20 \approx 41.3\) hours. 4. Interval: \((1200 - 41.3, 1200 + 41.3) = (1158.7, 1241.3)\) hours. 5. Interpretation: we are 95% confident that the mean lifetime of this brand's bulbs lies between about 1159 and 1241 hours.
Worked example: the price of estimating σ
What if \(\sigma = 100\) hours had been known all along? The known-\(\sigma\) formula would give margin \(1.96 \times 20 = 39.2\), interval \((1160.8, 1239.2)\). The \(t\) interval's margin is \(41.3\) - about 2.1 hours wider per side, the measurable price of estimating the spread instead of being handed it. Check: \(2.064 - 1.96 = 0.104\), and \(0.104 \times 20 = 2.08\) - the extra width exactly, so both intervals are consistent with their formulas. As \(n\) grows, \(t^*\) shrinks toward \(z^*\) and the toll fades.
Confidence interval or prediction interval?
The interval above estimates the mean lifetime - the centre of all bulbs - not the lifetime of any single bulb. A prediction interval for one new observation must also cover individual variation, and so is much wider: individuals scatter by \(\sigma\) (about \(s\)), while means scatter only by \(s/\sqrt{n}\). Engineers designing for the worst case - a bracket that must hold under one extreme load, a device that must survive one unlucky unit's use - need prediction intervals or tolerance intervals, not confidence intervals for the mean. The two answers sound similar; the questions are entirely different.
Robustness and alternatives
The \(t\) procedure is fairly robust: moderate departures from normality move the stated confidence level only a little, especially at larger \(n\). Its genuine weakness is outliers in small samples - one wild value can drag \(\bar{x}\) and inflate \(s\) at once, wrecking both ends of the interval. For skewed data or data with outliers, honest alternatives exist: transform the data (the log trick from the distribution-choice topic), use a median-based method, or bootstrap - resampling the data many times to estimate the sampling variability directly, no bell assumption required.
Common mistakes to avoid
First, using \(z^*\) when \(\sigma\) is unknown - the classic small-sample error, and the one the \(t\) distribution was built to prevent. Second, forgetting the degrees of freedom: \(\text{df} = n - 1\), not \(n\). Third, interpreting the interval as covering individual values - it estimates the mean. Fourth, treating matched or paired data as two independent groups; the differences want a one-sample interval. Fifth, ignoring outliers in a small sample, where a single point can hold the interval hostage. Sixth, quoting the interval without a full interpretation - level, parameter, interval, units.
Key takeaways
With \(\sigma\) unknown - the ordinary case - the interval for \(\mu\) is \(\bar{x} \pm t^*\frac{s}{\sqrt{n}}\) at \(n - 1\) degrees of freedom, \(t^*\) exceeding \(z^*\) and fading toward it as \(n\) grows. Conditions: random, independent, roughly normal sampling distribution - graded from symmetric-data checks at small \(n\) to CLT confidence at \(n \ge 30\). The bulbs: \(t^* = 2.064\), margin 41.3, interval (1158.7, 1241.3) hours - about 2.1 hours wider per side than the known-\(\sigma\) version would be. The interval estimates the mean, not individuals; outliers in small samples are its weak point, with transformation, median-based and bootstrap methods as the honest backups.
Flashcards
Flashcard
The one-sample t confidence interval
Test your understanding
Question 1 of 5
A sample of 25 lightbulbs has x̄ = 1200 and s = 100 hours. With t* ≈ 2.064 at 24 df, what is the 95% margin of error?
Question 2 of 5
When σ is unknown, the honest interval for a mean uses:
Question 3 of 5
For a one-sample t interval with n = 25, how many degrees of freedom?
Question 4 of 5
Compared with the known-σ interval on the same data, the t interval is:
Question 5 of 5
A 95% t-interval for mean bulb lifetime is (1158.7, 1241.3) hours. Which statement is correct?
