Confidence Intervals: Concept and Construction

Suppose a factory samples 50 light bulbs and finds a mean lifetime of 1020 hours. What can honestly be said about the true average lifetime of every bulb the factory will ever produce? A single number, like 1020, pretends to a precision that no finite sample can deliver. Confidence intervals answer the deeper question. They replace a point guess with a range of plausible values. This range comes with an explicit, provable guarantee about how often it will be right. This matters because point estimates alone are silent about their own reliability. Two estimates of 1020 hours, one from 5 bulbs and one from 5000, deserve very different levels of trust. A confidence interval encodes this difference directly into the width of the reported range. The entire discipline of statistical inference rests on being able to quantify uncertainty this way, rather than merely asserting it. The big picture is that a confidence interval is not a probability statement about the unknown parameter. The parameter is a fixed, non-random number. Instead, it is a probability statement about the procedure used to generate the interval. This procedure is evaluated before the data are observed. This distinction, subtle on first encounter, is the entire content of the idea of coverage probability. Getting it right is what separates rigorous inference from folklore. This article gives the precise definition of a confidence interval. It proves the general pivotal quantity method that manufactures such intervals from a sampling distribution. It works through the construction in full using the sample mean and sample variance. It closes by setting up the natural next steps. These include the exact small-sample pivots built from the tt, χ2\chi^2, and FF distributions. It also links to the Hypothesis Testing Framework.

Key takeaways
  • A confidence interval is a random interval, computed from data, that traps the fixed unknown parameter with a guaranteed long-run frequency.
  • The guarantee — the coverage probability — is a property of the procedure, evaluated before data are seen, not a probability statement about the parameter itself.
  • The pivotal quantity method builds intervals systematically: find a function of data and parameter whose distribution is known and parameter-free, then invert an inequality on it.
  • Interval width shrinks like 1/n1/\sqrt{n} for the sample mean; more data means a tighter, still-valid guarantee.
  • Confidence level (e.g. 95%) is a design choice, traded off against interval width — higher confidence costs a wider interval.
  • Key idea: invert a pivot's known distribution to convert a probability statement about an interval into a probability statement about coverage.

Where this lives today

Confidence intervals are the default unit of reported uncertainty across empirical science: clinical trials report confidence intervals for treatment effects, political polling reports margins of error (half-widths of confidence intervals for a proportion), and physical measurements report intervals for constants like the gravitational constant or particle masses. In machine learning, bootstrap and conformal prediction methods generalize the pivotal idea to build prediction intervals without assuming a parametric family, directly inheriting Neyman's coverage criterion. In engineering reliability analysis, confidence intervals for failure rates (often built from Exponential or Weibull pivots) drive maintenance schedules and warranty design. In econometrics, confidence intervals for regression coefficients — built from tt-pivots exactly analogous to Example 2 — underlie almost all reported empirical results. In cryptography and algorithm analysis, concentration inequalities (e.g. Chernoff-type bounds) are used to build confidence intervals around estimated success probabilities of randomized algorithms. The pivotal method itself reappears throughout statistics as the standard technical tool for deriving exact intervals whenever an appropriate distributional fact — such as those catalogued in the The t, Chi-Squared, and F Distributions — is available, and it dovetails precisely with the logic of the Hypothesis Testing Framework, where the same pivot inverted differently yields a test rather than an interval.

:::quiz[A researcher reports a 95% CI for a population mean as [40.2, 44.8]. Which statement is correct?]

  • There is a 95% probability that the true mean lies in [40.2, 44.8].
  • The interval [40.2, 44.8] was constructed with 95% probability.
  • The procedure used produces intervals containing the true mean in 95% of repeated samples ✓
  • 95% of the population's data values lie in [40.2, 44.8].

Ask the tutor: How does inverting the pivotal quantity's inequality relate to inverting a hypothesis test to build a confidence interval?

DefinitionConfidence Interval and Coverage Probability

Let X1,…,XnX_1, \dots, X_n be a random sample from a distribution FθF_\theta depending on an unknown parameter θ∈Θ\theta \in \Theta. Let L(X1,…,Xn)L(X_1,\dots,X_n) and U(X1,…,Xn)U(X_1,\dots,X_n) be two statistics (functions of the data only, not of θ\theta) satisfying L≤UL \le U with probability one. The random interval [L,U][L, U] is a 100(1−α)%100(1-\alpha)\% confidence interval for θ\theta if

Pθ(L(X1,…,Xn)≤θ≤U(X1,…,Xn))  =  1−αfor every θ∈Θ.\mathbb{P}_\theta\big(L(X_1,\dots,X_n) \le \theta \le U(X_1,\dots,X_n)\big) \;=\; 1-\alpha \qquad \text{for every } \theta \in \Theta.

The quantity 1−α1-\alpha is the confidence level, and the probability itself — computed with θ\theta held fixed and the randomness coming entirely from the sample — is the coverage probability of the procedure. If the equality above holds only in the limit as n→∞n \to \infty, the interval is said to have asymptotic confidence level 1−α1-\alpha. When exact equality cannot be achieved (e.g. for discrete data), an interval satisfying Pθ(L≤θ≤U)≥1−α\mathbb{P}_\theta(L \le \theta \le U) \ge 1-\alpha for all θ\theta is called conservative.

Intuition

The crucial mental shift is this: before the data are collected, LL and UU are random variables, and it is meaningful to ask how often the random interval [L,U][L,U] will contain θ\theta. Once the data are observed and L,UL,U become numbers, θ\theta is either inside the realized interval or it is not — there is no more randomness left to talk about. The 95% refers to the manufacturing process of the interval: if you repeated the entire experiment — resample, recompute LL and UU — a very large number of times, approximately 95% of the resulting intervals would contain the true θ\theta. A useful physical picture is a "confidence machine": you feed it a sample, and it stamps out a numbered interval. The guarantee is about the machine's long-run success rate across all its stampings, not about any single stamped interval you happen to be holding. The figure below shows twenty simulated 90% confidence intervals for a fixed true mean μ=5\mu = 5; on average two out of twenty (10%) miss the target, which is exactly what the guarantee promises — no more, no less, and never a promise about which ones miss.

Figure
Twenty simulated 90% confidence intervals for a fixed true mean μ = 5. About 10% are expected to miss.
μ = 5 (true value) 18/20 contain μ (green); 2/20 miss (red) ≈ 90% coverage

Building the machine: the pivotal quantity method

TheoremConstruction of Confidence Intervals via a Pivotal Quantity

Let X1,…,XnX_1,\dots,X_n be a random sample from a distribution depending on an unknown parameter θ∈Θ⊆R\theta \in \Theta \subseteq \mathbb{R}. Suppose there exists a function

Q=Q(X1,…,Xn,θ)Q = Q(X_1,\dots,X_n,\theta)

called a pivotal quantity (or pivot), satisfying: 1. (Computability) QQ is a function of the data X1,…,XnX_1,\dots,X_n and the parameter θ\theta only (no other unknown quantities enter). 2. (Pivotality) The distribution of QQ does not depend on θ\theta (or on any other unknown parameter); write this common distribution as GG. 3. (Monotonicity) For each fixed value of X1,…,XnX_1,\dots,X_n, the map θ↦Q(X1,…,Xn,θ)\theta \mapsto Q(X_1,\dots,X_n,\theta) is continuous and strictly monotonic on Θ\Theta. Fix α∈(0,1)\alpha \in (0,1) and choose constants qα/2<q1−α/2q_{\alpha/2} < q_{1-\alpha/2} from the distribution GG such that

P(qα/2≤Q≤q1−α/2)=1−α.\mathbb{P}\big(q_{\alpha/2} \le Q \le q_{1-\alpha/2}\big) = 1-\alpha.

Then, solving the inequality qα/2≤Q(X1,…,Xn,θ)≤q1−α/2q_{\alpha/2} \le Q(X_1,\dots,X_n,\theta) \le q_{1-\alpha/2} for θ\theta produces statistics L(X1,…,Xn)L(X_1,\dots,X_n) and U(X1,…,Xn)U(X_1,\dots,X_n), with L≤UL \le U, such that

Pθ(L≤θ≤U)=1−αfor every θ∈Θ.\mathbb{P}_\theta\big(L \le \theta \le U\big) = 1-\alpha \qquad \text{for every } \theta \in \Theta.

Hence [L,U][L,U] is an exact 100(1−α)%100(1-\alpha)\% confidence interval for θ\theta.

Proof

Technique: inversion of a pivotal quantity's known distribution. By hypothesis (2), QQ has a fixed distribution GG that does not involve θ\theta, so the event {qα/2≤Q≤q1−α/2}\{q_{\alpha/2} \le Q \le q_{1-\alpha/2}\} has a probability that can be computed from GG alone, and by construction this probability equals 1−α1-\alpha for every true value of θ\theta:

Pθ(qα/2≤Q(X1,…,Xn,θ)≤q1−α/2)=1−α,∀ θ∈Θ.(∗)\mathbb{P}_\theta\big(q_{\alpha/2} \le Q(X_1,\dots,X_n,\theta) \le q_{1-\alpha/2}\big) = 1-\alpha, \qquad \forall\, \theta \in \Theta. \tag{$\ast$}

This is the key step enabled by pivotality: although the event in (∗)(\ast) is defined using the true (unknown) θ\theta, its probability is a known number, 1−α1-\alpha, independent of θ\theta. This is precisely what makes the subsequent algebra produce a usable interval rather than an unverifiable statement. Now fix the realized data X1,…,XnX_1,\dots,X_n and regard Q(X1,…,Xn,θ)Q(X_1,\dots,X_n,\theta) as a function of θ\theta alone. By hypothesis (3) this function is continuous and strictly monotonic in θ\theta, hence injective, so it has a continuous inverse on its range. There are two cases: Case A: QQ is strictly increasing in θ\theta. Then for fixed data, the map θ↦Q(⋅,θ)\theta \mapsto Q(\cdot,\theta) is order-preserving, so the double inequality

qα/2≤Q(X1,…,Xn,θ)≤q1−α/2q_{\alpha/2} \le Q(X_1,\dots,X_n,\theta) \le q_{1-\alpha/2}

holds if and only if

Q−1(qα/2)≤θ≤Q−1(q1−α/2),Q^{-1}(q_{\alpha/2}) \le \theta \le Q^{-1}(q_{1-\alpha/2}),

where Q−1(⋅)Q^{-1}(\cdot) denotes the inverse in the θ\theta-argument, holding the data fixed. Set L:=Q−1(qα/2)L := Q^{-1}(q_{\alpha/2}) and U:=Q−1(q1−α/2)U := Q^{-1}(q_{1-\alpha/2}); these are statistics, since they are computed from the data alone (the constants qα/2,q1−α/2q_{\alpha/2}, q_{1-\alpha/2} are fixed numbers from GG, not unknowns). Case B: QQ is strictly decreasing in θ\theta. Then the map reverses order, so the same double inequality on QQ holds if and only if

Q−1(q1−α/2)≤θ≤Q−1(qα/2),Q^{-1}(q_{1-\alpha/2}) \le \theta \le Q^{-1}(q_{\alpha/2}),

and we instead set L:=Q−1(q1−α/2)L := Q^{-1}(q_{1-\alpha/2}), U:=Q−1(qα/2)U := Q^{-1}(q_{\alpha/2}), again both statistics. In either case, we have shown the pointwise logical equivalence of events (for every fixed data value and every θ\theta):

{qα/2≤Q(X1,…,Xn,θ)≤q1−α/2}  =  {L(X1,…,Xn)≤θ≤U(X1,…,Xn)}\big\{ q_{\alpha/2} \le Q(X_1,\dots,X_n,\theta) \le q_{1-\alpha/2} \big\} \;=\; \big\{ L(X_1,\dots,X_n) \le \theta \le U(X_1,\dots,X_n) \big\}

as events in the sample space (for the fixed true θ\theta, these are literally the same set of outcomes (X1,…,Xn)(X_1,\ldots,X_n)). Since these are the same event, they have the same probability under Pθ\mathbb{P}_\theta. Combining this with (∗)(\ast):

Pθ(L≤θ≤U)=Pθ(qα/2≤Q(X1,…,Xn,θ)≤q1−α/2)=1−α,\mathbb{P}_\theta\big(L \le \theta \le U\big) = \mathbb{P}_\theta\big(q_{\alpha/2} \le Q(X_1,\dots,X_n,\theta) \le q_{1-\alpha/2}\big) = 1-\alpha,

and since θ∈Θ\theta \in \Theta was arbitrary, this holds for every θ∈Θ\theta \in \Theta, which is exactly the definition of a 100(1−α)%100(1-\alpha)\% confidence interval for θ\theta. Finally, L≤UL \le U holds by construction in both cases (an artifact of the strict monotonicity in hypothesis (3), which guarantees Q−1Q^{-1} is well-defined and order-consistent), so [L,U][L,U] is a legitimate interval almost surely. ■\blacksquare

∎
CorollaryPivot for the Normal Mean with Known Variance

Let X1,…,XnX_1,\dots,X_n be i.i.d. N(μ,σ2)N(\mu,\sigma^2) with σ2\sigma^2 known and μ\mu unknown, and let Xˉn=1n∑iXi\bar{X}_n = \frac{1}{n}\sum_i X_i. Then

Z=Xˉn−μσ/nZ = \frac{\bar{X}_n - \mu}{\sigma/\sqrt{n}}

is a pivotal quantity: it is a function of data and μ\mu only, Z∼N(0,1)Z \sim N(0,1) regardless of the true μ\mu (by properties of the sampling distribution of the sample mean), and for fixed data it is a strictly decreasing... in fact strictly monotone — here strictly decreasing as a function of μ\mu (increasing μ\mu decreases ZZ). Applying the theorem with qα/2=−zα/2q_{\alpha/2} = -z_{\alpha/2} and q1−α/2=zα/2q_{1-\alpha/2} = z_{\alpha/2} (using the standard normal upper-α/2\alpha/2 point zα/2z_{\alpha/2}, and symmetry of N(0,1)N(0,1) about 0) yields the classical interval

[ Xˉn−zα/2σn,  Xˉn+zα/2σn ].\left[\, \bar{X}_n - z_{\alpha/2}\frac{\sigma}{\sqrt n},\ \ \bar{X}_n + z_{\alpha/2}\frac{\sigma}{\sqrt n} \,\right].
Proof

ZZ is a function of (X1,…,Xn,μ)(X_1,\dots,X_n,\mu) only, satisfying hypothesis (1). Since Xˉn∼N(μ,σ2/n)\bar X_n \sim N(\mu, \sigma^2/n) (this is the standard fact about the sampling distribution of a normal sample mean, taken as known background), standardizing gives Z∼N(0,1)Z \sim N(0,1) exactly, for every μ\mu — this is hypothesis (2), pivotality. For fixed data, Z(μ)=(Xˉn−μ)/(σ/n)Z(\mu) = (\bar X_n - \mu)/(\sigma/\sqrt n) is an affine, strictly decreasing function of μ\mu with nonzero slope −n/σ-\sqrt n/\sigma, so it is continuous and strictly monotonic — hypothesis (3). By symmetry of the standard normal density about 00, choosing qα/2=−zα/2q_{\alpha/2} = -z_{\alpha/2} and q1−α/2=zα/2q_{1-\alpha/2} = z_{\alpha/2} (where zα/2z_{\alpha/2} satisfies P(Z>zα/2)=α/2\mathbb P(Z > z_{\alpha/2}) = \alpha/2) gives

P(−zα/2≤Z≤zα/2)=1−2⋅α2=1−α.\mathbb P(-z_{\alpha/2} \le Z \le z_{\alpha/2}) = 1 - 2\cdot \tfrac{\alpha}{2} = 1-\alpha.

Since ZZ is decreasing in μ\mu (Case B of the theorem's proof), inverting −zα/2≤Xˉn−μσ/n≤zα/2-z_{\alpha/2} \le \frac{\bar X_n - \mu}{\sigma/\sqrt n} \le z_{\alpha/2} gives

Xˉn−zα/2σn≤μ≤Xˉn+zα/2σn,\bar X_n - z_{\alpha/2}\frac{\sigma}{\sqrt n} \le \mu \le \bar X_n + z_{\alpha/2}\frac{\sigma}{\sqrt n},

matching L,UL, U from Case B with the roles of the endpoints swapped as prescribed there. By the theorem, this interval has exact coverage 1−α1-\alpha for every μ∈R\mu \in \mathbb{R}. ■\blacksquare

∎

Worked examples

Example 1 — Known variance, asymmetric numbers. A quality engineer measures the tensile strength (in MPa) of n=36n = 36 steel cables. Past extensive testing has established that tensile strength is normally distributed with known standard deviation σ=12\sigma = 12 MPa (variability is a stable manufacturing constant, unlike the mean, which depends on a new alloy batch). The sample mean from this batch is xˉ36=812.4\bar x_{36} = 812.4 MPa. Construct a 95% confidence interval for the true mean tensile strength μ\mu of the batch. Since α=0.05\alpha = 0.05, we need zα/2=z0.025=1.96z_{\alpha/2} = z_{0.025} = 1.96 (the value with P(Z>1.96)=0.025\mathbb P(Z>1.96)=0.025 for standard normal ZZ). The standard error is

σn=1236=126=2.\frac{\sigma}{\sqrt n} = \frac{12}{\sqrt{36}} = \frac{12}{6} = 2.

Applying the interval from the corollary above:

xˉ±z0.025⋅σn=812.4±1.96(2)=812.4±3.92.\bar x \pm z_{0.025}\cdot \frac{\sigma}{\sqrt n} = 812.4 \pm 1.96(2) = 812.4 \pm 3.92.

So the 95% confidence interval is

[808.48, 816.32] MPa.[808.48,\ 816.32] \text{ MPa}.

Interpretation, done correctly: we do not say "there is a 95% probability that μ\mu lies in [808.48,816.32][808.48, 816.32]" as a statement about μ\mu's randomness — μ\mu is a fixed physical constant of this alloy batch, not random. The correct statement is that the procedure — take a sample of 36, compute Xˉ±1.96σ/n\bar X \pm 1.96\sigma/\sqrt n — produces an interval containing the true μ\mu in 95% of independent repetitions of this sampling experiment. Having observed xˉ=812.4\bar x = 812.4, the realized interval [808.48,816.32][808.48, 816.32] either contains μ\mu or it does not; the 95% describes the reliability of the method that produced it. Note also the effect of sample size: had the engineer used n=9n = 9 instead of 3636 (a quarter as many cables), the standard error would be 12/3=412/3 = 4, doubling the half-width to ±7.84\pm 7.84 — illustrating the 1/n1/\sqrt n shrinkage rate promised by the theory.

Example 2 — Unknown variance, using a Student-pivot preview. A nutrition lab measures the sodium content (mg) of n=10n = 10 servings of a snack food, believed approximately normal, with both mean μ\mu and variance σ2\sigma^2 unknown. The data give sample mean xˉ=342.7\bar x = 342.7 and sample standard deviation s=14.8s = 14.8 (computed from s2=1n−1∑(xi−xˉ)2s^2 = \frac{1}{n-1}\sum (x_i - \bar x)^2, the unbiased sample variance). Here the pivot from Example 1 cannot be used directly, because it requires the known constant σ\sigma; substituting the random ss for σ\sigma changes the distribution of the resulting quantity away from standard normal. The random variable

T=Xˉn−μSn/nT = \frac{\bar X_n - \mu}{S_n/\sqrt n}

is still a valid pivot in the sense of the theorem — it is a function of data and μ\mu only, its distribution (Student's tt with n−1n-1 degrees of freedom, a fact established via the Student t-distribution) does not depend on μ\mu or σ2\sigma^2, and for fixed data it is strictly decreasing in μ\mu exactly as before. The only change from Example 1 is which tabulated distribution supplies the critical constant. For n−1=9n-1 = 9 degrees of freedom and 95% confidence, the correct critical value is t9,0.025=2.262t_{9,0.025} = 2.262 (larger than z0.025=1.96z_{0.025}=1.96, reflecting the extra uncertainty from estimating σ\sigma). The interval is

xˉ±t9,0.025sn=342.7±2.262⋅14.810=342.7±2.262(4.681)=342.7±10.59,\bar x \pm t_{9,0.025}\frac{s}{\sqrt n} = 342.7 \pm 2.262\cdot\frac{14.8}{\sqrt{10}} = 342.7 \pm 2.262(4.681) = 342.7 \pm 10.59,

giving

[332.11, 353.29] mg.[332.11,\ 353.29]\ \text{mg}.

This example is included specifically to show why the pivotal method's power lies in choosing the right pivot for what is actually known: swapping σ\sigma (known) for SnS_n (estimated) is not cosmetic — it changes the reference distribution from N(0,1)N(0,1) to tn−1t_{n-1}, and using z0.025=1.96z_{0.025}=1.96 here instead of t9,0.025=2.262t_{9,0.025}=2.262 would silently understate the interval's width and destroy the exact 95% coverage guarantee for finite nn. This is precisely the gap that the The t, Chi-Squared, and F Distributions machinery is built to close.

CounterexampleWhy pivotality (parameter-free distribution) cannot be dropped

Consider X1,…,XnX_1,\dots,X_n i.i.d. Uniform(0,θ)\text{Uniform}(0,\theta) with θ>0\theta > 0 unknown, and let M=max⁡(X1,…,Xn)M = \max(X_1,\dots,X_n). A tempting "pivot-like" quantity is simply

R=M−θ.R = M - \theta.

This satisfies hypothesis (1) (a function of data and θ\theta) and hypothesis (3) (for fixed data, θ↦M−θ\theta \mapsto M-\theta is continuous and strictly decreasing). But it fails hypothesis (2), pivotality: the distribution of R=M−θR = M-\theta depends on θ\theta. Indeed, MM has density fM(m)=nmn−1/θnf_M(m) = n m^{n-1}/\theta^n for 0<m<θ0 < m < \theta, so R=M−θR = M - \theta has density fR(r)=n(r+θ)n−1/θnf_R(r) = n(r+\theta)^{n-1}/\theta^n for −θ<r<0-\theta < r < 0 — a family of densities whose support and shape both depend explicitly on θ\theta. Suppose one naively tries to mimic the pivotal method anyway: fix a single numeric threshold c<0c<0 (not depending on θ\theta) and attempt to use Pθ(R≥c)=1−α\mathbb{P}_\theta(R \ge c) = 1-\alpha to build an interval by inverting M−θ≥cM - \theta \ge c, i.e. θ≤M−c\theta \le M - c. Compute

Pθ(R≥c)=Pθ(M≥θ+c)=1−(θ+cθ)n=1−(1+cθ)n,\mathbb{P}_\theta(R \ge c) = \mathbb{P}_\theta(M \ge \theta + c) = 1 - \left(\frac{\theta+c}{\theta}\right)^n = 1-\left(1+\frac{c}{\theta}\right)^n,

valid only when θ+c>0\theta + c > 0, i.e. θ>−c\theta > -c. This probability is not constant in θ\theta: as θ→∞\theta \to \infty with cc fixed, (1+c/θ)n→1(1+c/\theta)^n \to 1, so the coverage probability →0\to 0; for θ\theta close to −c-c, coverage climbs toward 11. There is no single choice of cc that delivers coverage exactly 1−α1-\alpha simultaneously for every θ\theta, which is precisely the requirement in the confidence-interval definition ("for every θ∈Θ\theta \in \Theta"). Any interval built this way would have a coverage probability that silently varies with the unknown θ\theta — undermining the entire guarantee, since a user cannot know θ\theta and hence cannot know the actual coverage they are getting. The genuine fix is to use an actually pivotal quantity for this model, such as Q=M/θQ = M/\theta, whose density fQ(q)=nqn−1f_Q(q) = nq^{n-1} on (0,1)(0,1) does not involve θ\theta at all (a direct computation via the change of variables q=m/θq = m/\theta confirms this). Inverting P(qα/2≤M/θ≤q1−α/2)=1−α\mathbb{P}(q_{\alpha/2} \le M/\theta \le q_{1-\alpha/2}) = 1-\alpha correctly produces an exact confidence interval for θ\theta, illustrating that pivotality — not mere monotonicity — is the property doing the real work in the theorem.

Scope, edge cases & common traps

SituationWhat changesCorrect approach
σ2\sigma^2 known, normal dataZZ-pivot, exact for all nnUse zα/2z_{\alpha/2}, interval Xˉ±zα/2σ/n\bar X \pm z_{\alpha/2}\sigma/\sqrt n
σ2\sigma^2 unknown, normal datatt-pivot, exact for all nnUse tn−1,α/2t_{n-1,\alpha/2}, interval Xˉ±tn−1,α/2 S/n\bar X \pm t_{n-1,\alpha/2}\,S/\sqrt n (see The t, Chi-Squared, and F Distributions)
Non-normal data, large nnOnly asymptotic coverage, via CLTUse ZZ-pivot as an approximation; state "asymptotic 95% CI"
Discrete data (e.g. Binomial pp)Exact equality often impossibleReport a conservative interval, coverage ≥1−α\ge 1-\alpha
Variance parameter σ2\sigma^2 itselfPivot is (n−1)S2/σ2∼χn−12(n-1)S^2/\sigma^2 \sim \chi^2_{n-1}Asymmetric interval from chi-squared quantiles, not symmetric ±\pm form
Two-sample comparisonsPivot built from difference of meansUses FF or tt distribution depending on variance assumptions

The mathematician behind it

Origins · 1 minJerzy Neyman and how this was discovered
JN
Jerzy Neyman
1894–1981 · Warsaw, Kraków, and the University of California, Berkeley

Neyman was a Polish-born mathematician and statistician who, working with Egon Pearson in the 1930s and independently afterward, formalized the frequentist logic of interval estimation and hypothesis testing. He insisted on a strict separation between the mathematical guarantee a procedure could deliver in repeated sampling and any subjective belief about a particular unknown parameter — a distinction that remains the conceptual backbone of the confidence interval.

History
- 1894 — Jerzy Neyman born in Bendery, Russian Empire (now Moldova).
- 1928–1933 — Neyman and Egon Pearson develop the theory of hypothesis testing, laying groundwork for a fully frequentist inference theory.
- 1934 — Neyman presents the theory of confidence intervals to the Royal Statistical Society, formalizing coverage probability as the guiding criterion for interval estimators.
- 1937 — Neyman's paper "Outline of a Theory of Statistical Estimation Based on the Classical Theory of Probability" gives the rigorous general definition of confidence intervals used to this day.
- 1938 — Neyman moves to the University of California, Berkeley, founding a statistics program that becomes a center of frequentist inference research.
Is a 95% confidence interval the same as saying there's a 95% chance the parameter is in the interval?

No, and this is the single most important subtlety in the whole topic. In the classical (frequentist) framework, the parameter θ\theta is a fixed, non-random constant — it does not have a probability distribution to make such a statement about. The randomness lives entirely in the data, and hence in LL and UU, before they are observed. The correct statement is: the procedure that generated LL and UU has the property that, across repeated applications to fresh samples, it produces an interval containing the true θ\theta 95% of the time. Once you plug in actual numbers and get [40.2,44.8][40.2, 44.8], that specific interval either contains θ\theta or it doesn't — the 95% is a pre-data guarantee about the method, not a post-data probability about this instance.

What exactly is a pivotal quantity, and why does it need a parameter-free distribution?

A pivotal quantity Q(X1,…,Xn,θ)Q(X_1,\dots,X_n,\theta) is a function combining the data and the unknown parameter whose probability distribution is completely known and does not depend on θ\theta (or any other unknown quantity). This parameter-free property is what lets you compute exact probabilities like P(qα/2≤Q≤q1−α/2)=1−α\mathbb{P}(q_{\alpha/2}\le Q\le q_{1-\alpha/2}) = 1-\alpha using only tables (normal, tt, χ2\chi^2, etc.), with no unknowns left in the computation. If the distribution of QQ depended on θ\theta, you could not choose fixed cutoffs qα/2,q1−α/2q_{\alpha/2}, q_{1-\alpha/2} that deliver exactly 1−α1-\alpha coverage for every possible true θ\theta, as the counterexample with R=M−θR = M-\theta demonstrates concretely.

Why is the interval for the mean with unknown variance wider than the one with known variance?

Because using the sample standard deviation SS in place of the true σ\sigma introduces additional sampling variability: SS itself fluctuates from sample to sample, and this extra noise must be accounted for to keep the coverage probability exactly 1−α1-\alpha. The resulting pivot T=(Xˉ−μ)/(S/n)T = (\bar X - \mu)/(S/\sqrt n) follows a Student's tt-distribution with n−1n-1 degrees of freedom rather than a standard normal, and the tt-distribution has heavier tails than the normal, especially for small nn. This forces the critical value tn−1,α/2t_{n-1,\alpha/2} to exceed zα/2z_{\alpha/2}, widening the interval to compensate honestly for the estimation of σ\sigma.

Does a wider confidence interval mean the estimate is worse?

Not necessarily — width trades off against confidence level and sample size, and the three cannot all be optimized simultaneously. For fixed sample size and data, increasing the confidence level (say from 90% to 99%) necessarily widens the interval, because a stronger guarantee of capturing the true parameter requires a larger safety margin. Comparing two different confidence levels on the same data, or comparing across different sample sizes, without accounting for this trade-off is a common source of confusion. The right comparison is: at a fixed confidence level, a narrower interval built from a larger, more informative sample reflects better precision, not a different (weaker) guarantee.

Can confidence intervals be constructed without a pivotal quantity?

Yes — the pivotal method is the cleanest and most common construction for exact intervals in parametric families, but it is not the only route. Alternatives include inverting a hypothesis test directly (accepting all θ0\theta_0 not rejected by a level-α\alpha test, which produces an interval by duality with the Hypothesis Testing Framework), likelihood-based intervals using the asymptotic distribution of the likelihood ratio, and resampling-based methods like the bootstrap, which approximate a pivot's distribution numerically when no closed form exists. All of these ultimately aim at the same coverage guarantee defined at the outset; the pivotal method is simply the most transparent to prove rigorously because it reduces to inverting a single known distribution.

What happens if the pivot is not monotonic in θ?

The inversion step in the theorem's proof relied critically on strict monotonicity to convert the two-sided inequality on QQ into a single interval for θ\theta. If Q(⋅,θ)Q(\cdot,\theta) is not monotonic — for instance if it is U-shaped in θ\theta — the set of θ\theta satisfying qα/2≤Q≤q1−α/2q_{\alpha/2}\le Q\le q_{1-\alpha/2} can be a union of disjoint intervals, or a more complicated region, rather than a single interval [L,U][L,U]. The resulting "confidence set" is still valid (it still has exact coverage 1−α1-\alpha, by the same probability argument), but it may not be expressible as a simple two-endpoint interval, and reporting it requires describing the full region rather than just LL and UU.

Is a conservative interval, with coverage ≥ 1−α rather than exactly 1−α, a problem?

It is a compromise forced by discreteness, not a flaw in the logic. For discrete data, such as a Binomial proportion, no continuous adjustment of LL and UU can hit the target coverage exactly for every θ\theta, because the achievable coverage probabilities jump in discrete steps as θ\theta varies. A conservative interval guarantees at least the nominal coverage for every θ\theta, erring on the side of being slightly wider than strictly necessary, which is the safer failure mode for a scientific guarantee. The trade-off is that conservative intervals are, on average, wider than the exact nominal level would technically require, sacrificing some precision for the safety of the "at least 1−α1-\alpha" guarantee.

Before this

Where this leads