Confidence Intervals for a Mean

A confidence interval (CI) for a population mean μ is a range of values, calculated from sample data, constructed so that if the sampling procedure were repeated many times, a specified proportion of the resulting intervals — conventionally 95% — would contain the true value of μ. The sample mean, x̄, is a point estimate: a single best guess. The confidence interval is an interval estimate: it acknowledges that x̄ varies from sample to sample and expresses that variability as a range around x̄. The width of the interval is governed by the standard error of the mean (SEM = s/√n), which measures how much sample means are expected to fluctuate around μ due to random sampling. A small SEM (large sample, low variability) gives a narrow, precise interval. A large SEM (small sample, high variability) gives a wide, imprecise interval. The interval is centered on x̄ and extends outward by a margin of error: a critical value multiplied by the SEM. The critical value comes from the t-distribution rather than the standard normal (z) distribution whenever the population standard deviation σ is unknown and must be estimated from the sample as s — which is essentially always the case in real biological work. The t-distribution has heavier tails than the normal distribution, reflecting the extra uncertainty introduced by estimating σ from a finite sample. Its exact shape depends on the degrees of freedom, df = n − 1: with few data points, t has notably fatter tails than z; as n grows large, t converges to z, because s becomes a very precise estimate of σ.

Step 1: Understand why the sample mean has a sampling distribution at all. Imagine repeating the mouse-cortisol study over and over, drawing a fresh sample of 18 mice each time from the same wild population. Each sample would give a slightly different mean, purely from the luck of which mice got caught. Plotted across many hypothetical repetitions, those sample means form the sampling distribution of the mean. Its standard deviation is the SEM. By the Central Limit Theorem, this sampling distribution is approximately normal (or t-shaped, once σ is estimated) even if individual mouse cortisol values are not perfectly normal, provided n is not tiny.

Step 2: Estimate the standard error from one real sample. You only get one actual sample, so σ (population SD) is unknown. You estimate it with the sample standard deviation s, and then estimate the SEM as: SEM = s / √n This is a plug-in estimate — a source of extra uncertainty on top of the sampling variability itself, which is exactly why t (not z) is needed next.

Step 3: Pick the confidence level and find the critical t-value. For a 95% CI, you need the value t<strong> such that 95% of the t-distribution with df = n − 1 lies between −t and +*t**, leaving 2.5% in each tail. This value is read from a t-table or computed with software, and it always exceeds the corresponding z-value (1.96) for finite n — the t-distribution "pays a penalty" for the added uncertainty of estimating σ.

Step 4: Compute the margin of error and build the interval. Margin of error (ME) = *t** × SEM. The 95% CI is: x̄ − ME to x̄ + ME

Step 5: Interpret correctly. The interval is a statement about the method's long-run reliability: if you repeated the whole sampling-and-computing procedure many times, 95% of the resulting intervals would contain the true μ. Any single computed interval either does or does not contain μ — there is no probability left to talk about once the numbers are in hand.

Confidence intervals rest on two empirically and mathematically established pillars: the sampling distribution behavior described by the Central Limit Theorem, and the correction for small-sample uncertainty worked out by William Sealy Gosset.

The Central Limit Theorem in action. The claim that sample means follow an approximately normal (or t) distribution — even when individual measurements do not — is not an assumption taken on faith; it is repeatedly demonstrable by simulation and was rigorously proved mathematically over the 18th–20th centuries (Laplace, and later refined proofs by Lyapunov and Lindeberg). A simple, reproducible demonstration: take a wildly non-normal population — say, a skewed distribution of parasite loads per host, most hosts near zero with a long tail of heavily infected individuals — and repeatedly draw samples of size 30, computing the mean each time. Histogram those thousands of sample means: the result is strikingly bell-shaped and centered on the true population mean, with spread shrinking as √n grows. This is routinely demonstrated in statistics courses and simulation studies, and it is the empirical justification for treating the sampling distribution of x̄ as approximately normal in large samples.

Gosset's small-sample problem, and the birth of "Student's t." The historically decisive case is William Sealy Gosset's work at the Guinness brewery in Dublin in the early 1900s. Gosset needed to make inferences about the quality of barley and yeast strains from very small samples — sometimes just 3 to 4 measurements — where the normal-distribution-based (z) methods of the day gave overconfident, too-narrow intervals. He derived, and empirically tested against distributions generated from actual height and finger-measurement data (and later through explicit random-sampling experiments drawing cards to simulate samples), the exact distribution that the quantity t = (x̄ − μ) / (s/√n) follows when the population is normal and σ is estimated from a small sample. He published this in 1908 under the pseudonym "Student" (Guinness restricted publication of trade-relevant work by employees), giving tables of the distribution now called the t-distribution. His central empirical/mathematical finding: this ratio is not normally distributed for small n — it has fatter tails — and the excess spread depends predictably on degrees of freedom. This meant that using z-critical values (like 1.96) for small samples systematically understated uncertainty, producing intervals that were too narrow and "significant" results that were too easy to obtain. Gosset's correction — using the wider, df-dependent t critical values — was later placed on rigorous theoretical footing by Ronald Fisher in the 1920s, who derived the exact sampling distribution of s² and formalized degrees of freedom. Fisher's theoretical derivation matched Gosset's empirically-motivated tables, cementing the t-distribution as the correct tool whenever σ is estimated rather than known — which describes nearly all real biological data.

Simulation as ongoing verification. Modern confirmation of the whole CI machinery is itself an experiment repeatable on a laptop: simulate 10,000 samples of size n from a known population with a known mean, build a 95% CI from each sample using the t-formula, and count how many intervals actually contain the true mean. The tally consistently comes out close to 95% — this is the standard classroom and textbook demonstration (see Whitlock & Schluter's simulation exercises) that operationalizes what "95% confidence" really means: a property of the procedure, verified by repetition, not a probability attached to any one interval.

Example 1 — Small, asymmetric sample: enzyme activity in a rare cichlid species A researcher measures alkaline phosphatase activity (U/L) in n = 9 wild-caught individuals of a rare Lake Malawi cichlid, since larger samples are impossible to obtain ethically: Data (U/L): 142, 155, 138, 210, 149, 133, 161, 145, 158 Step 1 — sample mean: x̄ = (142+155+138+210+149+133+161+145+158)/9 = 1391/9 ≈ 154.6 U/L Step 2 — sample standard deviation: Computing deviations from 154.6 and squaring (the value 210 is a clear outlier pulling on the spread) gives s ≈ 22.9 U/L (rounded from full calculation). Step 3 — standard error: SEM = s/√n = 22.9/√9 = 22.9/3 ≈ 7.63 U/L Step 4 — degrees of freedom and critical t: df = 9 − 1 = 8. For 95% confidence, two-tailed, t* (df=8) ≈ 2.306 (notice this is much larger than z = 1.96 — the small-sample penalty is substantial here) Step 5 — margin of error and interval: ME = 2.306 × 7.63 ≈ 17.6 U/L 95% CI = 154.6 ± 17.6 = 137.0 to 172.2 U/L Interpretation: we are 95% confident, in the long-run-frequency sense, that the interval 137.0–172.2 U/L captures the true mean alkaline phosphatase activity for this cichlid population. Note how wide this is — nearly ±11% of the mean — a direct consequence of small n and one high outlier inflating s. This is a realistic and common situation in conservation physiology, where sample sizes are constrained by animal welfare and rarity.

Example 2 — Larger sample, clinical scenario: fasting glucose after a dietary intervention A clinical trial measures fasting blood glucose (mmol/L) in n = 64 patients three months after starting a low-glycemic-index diet. Reported summary statistics: x̄ = 5.62 mmol/L, s = 0.84 mmol/L. Step 1 — standard error: SEM = 0.84/√64 = 0.84/8 = 0.105 mmol/L Step 2 — degrees of freedom and critical t: df = 64 − 1 = 63. For 95% confidence, t*(df=63) ≈ 2.00 (at this sample size, t has nearly converged to z = 1.96 — the small-sample penalty has almost vanished) Step 3 — margin of error and interval: ME = 2.00 × 0.105 ≈ 0.21 mmol/L 95% CI = 5.62 ± 0.21 = 5.41 to 5.83 mmol/L Interpretation: the plausible range for true mean post-intervention fasting glucose is fairly narrow, 5.41–5.83 mmol/L, reflecting the larger sample size. A clinician comparing this to the standard "normal fasting glucose" threshold (roughly below 5.6 mmol/L for many guidelines) would note the interval straddles that boundary — illustrating that the CI itself, not just the point estimate, informs whether the intervention convincingly shifted patients into a normal range. This foreshadows the direct link between CIs and hypothesis testing: if a null hypothesized mean (say, μ₀ = 6.0 mmol/L, a pre-treatment baseline) falls outside this interval, a two-sided test at α = 0.05 would reject that null.

Clinical trials and drug approval. Regulatory bodies (FDA, EMA) require confidence intervals, not just p-values, for treatment effects — e.g., the 95% CI for the mean reduction in blood pressure from a new drug tells clinicians the plausible range of real-world benefit, which matters for weighing against side effects even when a result is "statistically significant."

Conservation biology and small population monitoring. Endangered species studies routinely work with tiny samples (a handful of tagged individuals, a few surviving nests), making the t-distribution's small-sample correction essential — using z instead of t in such studies would understate uncertainty and risk overconfident management decisions.

Agricultural field trials. Ironically returning to Gosset's original context, modern crop yield trials still report CIs for mean yield differences between varieties or fertilizer treatments, guiding real-world planting decisions across thousands of hectares.

Ecological and physiological reference ranges. Establishing "normal" ranges for physiological variables (body temperature, cortisol baselines, growth rates) in a species requires CIs around means from field or lab samples, feeding directly into veterinary diagnostics and wildlife health assessments.

Reproducibility and meta-analysis. Confidence intervals are the standard input for meta-analyses that pool results across many small biological studies — forest plots (visual summaries of CIs from multiple studies) are a direct, ubiquitous application built entirely on this machinery.

Ask the tutor: How would this confidence interval formula change if I wanted to compare two independent sample means instead of estimating just one?