Interval Estimation: Pivotal Quantities & Exact Confidence Sets

Let $(\mathcal{X}, \mathcal{A}, \mathcal{P})$ be a statistical experiment where $\mathbf{X} = (X_1, \dots, X_n) \sim P_\theta$ with $\theta \in \Theta \subseteq \mathbb{R}$. An interval estimator for a scalar parameter $\tau(\theta)$ is a pair of observable statistics $L(\mathbf{X})$ and $U(\mathbf{X})$ satisfying:

The realization $[L(\mathbf{x}), U(\mathbf{x})]$ computed from observed empirical data $\mathbf{x}$ is termed a confidence interval.

Coverage Probability and Confidence Coefficient

Prior to data collection, the interval bounds $L(\mathbf{X})$ and $U(\mathbf{X})$ are random variables; the parameter $\tau(\theta)$ is a fixed, non-random constant. The coverage probability function of the interval estimator is:

The confidence coefficient (or nominal confidence level), denoted $1 - \alpha$ with $\alpha \in (0, 1)$, is the infimum of the coverage probability over the entire parameter space:

If $C(\theta) = 1 - \alpha$ identically for all $\theta \in \Theta$, the interval is said to be an exact $(1 - \alpha)$ confidence interval. If $C(\theta) \ge 1 - \alpha$ for all $\theta$, the interval is conservative. If $C(\theta) < 1 - \alpha$ for some parameter values, the interval is anticonservative or liberal, exhibiting coverage failure.

Frequentist Interpretation of Confidence

A notorious epistemological pitfall in applied statistics is the misinterpretation of realized confidence intervals. Once experimental data $\mathbf{x}$ are collected and specific numbers are evaluated—such as $[12.4, 15.8]$—it is mathematically meaningless in frequentist probability to assert that "there is a $95\%$ probability that the true parameter $\theta$ lies between $12.4$ and $15.8$." The parameter $\theta$ is not a random variable; it is an objective, deterministic state of nature. The realized interval $[12.4, 15.8]$ either contains $\theta$ or it does not: the conditional probability is either $1$ or $0$. The correct frequentist interpretation is: If the identical experiment is repeated indefinitely under identical conditions, generating an infinite sequence of independent intervals $[L(\mathbf{X}^{(k)}), U(\mathbf{X}^{(k)})]$, exactly $100(1 - \alpha)\%$ of the resulting random intervals will successfully contain the fixed true parameter $\theta$.

Neyman's Confidence Belt

Jerzy Neyman visualized interval estimation through a confidence belt in the $(\mathbf{x}, \theta)$ plane. For each fixed parameter value $\theta_0$, define an acceptance interval in the sample space:

such that $\mathbb{P}_{\theta_0}(T(\mathbf{X}) \in A(\theta_0)) = 1 - \alpha$. Plotting the curves $c_1(\theta)$ and $c_2(\theta)$ against $\theta$ traces out a two-dimensional confidence belt. Once a specific numerical statistic $t = T(\mathbf{x})$ is observed on the horizontal axis, projecting vertically intersects the boundaries of the belt at two points: $\theta_L(t)$ and $\theta_U(t)$. Inverting the confidence belt produces the confidence set:

The most powerful and elegant analytical tool for constructing exact confidence intervals in parametric models is the Pivotal Quantity Method, formulated by R. A. Fisher (1930).

Definition of a Pivotal Quantity

A function of the sample data $\mathbf{X}$ and the unknown parameter $\theta$, denoted $Q(\mathbf{X}, \theta)$, is called a pivotal quantity (or simply a pivot) if its probability distribution is completely independent of $\theta$ (and any nuisance parameters).

Crucially, a pivot must satisfy two distinct conditions:

1. $Q(\mathbf{x}, \theta)$ must depend explicitly on both the data $\mathbf{x}$ and the unknown parameter $\theta$.
2. The probability law governing the random variable $Q(\mathbf{X}, \theta)$ under $P_\theta$ must be a fixed, known distribution containing zero unknown parameters (e.g., standard normal $\mathcal{N}(0, 1)$, Student's $t_\nu$, Chi-square $\chi_\nu^2$, or Snedecor's $F_{\nu_1, \nu_2}$).

Constructing Confidence Intervals via Inversion of Pivots

Let $Q(\mathbf{X}, \theta)$ be a pivotal quantity with known continuous cumulative distribution function $F_Q(q)$. For any chosen significance level $\alpha \in (0, 1)$, we can select two critical values $q_1 < q_2$ from the known distribution of $Q$ such that:

Typically, one selects symmetric tail probabilities: $F_Q(q_1) = \alpha / 2$ and $F_Q(q_2) = 1 - \alpha / 2$, so $q_1 = q_{\alpha/2}$ and $q_2 = q_{1 - \alpha/2}$. Now, assume that for every sample realization $\mathbf{x}$, the function $Q(\mathbf{x}, \theta)$ is a strictly monotonic function of $\theta$. Then the algebraic inequality:

can be inverted to isolate $\theta$:

• If $Q(\mathbf{x}, \theta)$ is strictly decreasing in $\theta$:

• If $Q(\mathbf{x}, \theta)$ is strictly increasing in $\theta$:

Setting $L(\mathbf{X}) = Q^{-1}(\mathbf{X}, q_2)$ and $U(\mathbf{X}) = Q^{-1}(\mathbf{X}, q_1)$ yields an exact $(1 - \alpha)$ confidence interval:

Consider an i.i.d. random sample $X_1, X_2, \dots, X_n$ from a Gaussian distribution $\mathcal{N}(\mu, \sigma_0^2)$ where the population variance $\sigma_0^2$ is known. The sample mean is $\bar{X} = \frac{1}{n} \sum_{i=1}^n X_i$.

The Standard Normal Pivot

By linear combination properties of Gaussian variables:

Standardizing $\bar{X}$ produces the pivotal quantity:

Notice that $Z$ depends on both the data (via $\bar{X}$) and the unknown parameter $\mu$. Its sampling distribution is strictly standard normal:

which contains zero unknown parameters. Thus $Z$ is an exact pivot!

Inversion and Interval Construction

Let $z_{1 - \alpha/2}$ denote the $(1 - \alpha/2)$-quantile of the standard normal distribution, such that $\Phi(z_{1 - \alpha/2}) = 1 - \alpha/2$. By symmetry of the Gaussian density about zero:

Multiply across the compound inequality by the standard error $\frac{\sigma_0}{\sqrt{n}} > 0$:

Subtract $\bar{X}$ from all three parts:

Multiplying by $-1$ reverses the inequalities:

The exact $100(1 - \alpha)\%$ confidence interval for the population mean $\mu$ is:

Geometric Properties

1. Symmetry: The interval is centered precisely at the sample mean $\bar{X}$.

2. Deterministic Width: The length of the interval is:

Notice that the interval width $W$ is a deterministic constant: it does not depend on the observed data values!

3. Sample Size Determination: If an experimenter demands that the margin of error not exceed a bound $E$, setting $z_{1 - \alpha/2} \frac{\sigma_0}{\sqrt{n}} \le E$ yields the exact required sample size:

In real scientific experiments, the population variance $\sigma^2$ is almost never known with certainty. If we attempt to use $Z = \frac{\bar{X} - \mu}{\sigma / \sqrt{n}}$, the quantity involves the unknown nuisance parameter $\sigma$, rendering it unusable as a pivot. If we substitute the sample standard deviation $S = \sqrt{\frac{1}{n-1} \sum (X_i - \bar{X})^2}$, what is the exact distribution of the resulting ratio?

Derivation of Student's $t$ Pivot (Gosset, 1908)

Consider the Studentized statistic:

We decompose $T$ algebraically as follows:

where:

1. $Z = \frac{\bar{X} - \mu}{\sigma / \sqrt{n}} \sim \mathcal{N}(0, 1)$ is standard normal.
2. $V = \frac{(n-1)S^2}{\sigma^2} \sim \chi_{n-1}^2$ is Chi-square with $\nu = n - 1$ degrees of freedom.
3. By Basu's Theorem (or Cochran's Theorem), the sample mean $\bar{X}$ and sample variance $S^2$ are mutually independent! Therefore, the numerator $Z$ and denominator $V$ are strictly independent random variables.

By definition, the ratio of a standard normal variable to the square root of an independent Chi-square variable divided by its degrees of freedom follows Student's $t$-distribution with $\nu = n - 1$ degrees of freedom:

The distribution $t_{n-1}$ depends only on the known sample size $n$, and is completely independent of both unknown parameters $\mu$ and $\sigma^2$. Thus $T$ is an exact pivotal quantity!

Exact Interval Construction

Let $t_{n-1, 1 - \alpha/2}$ denote the $(1 - \alpha/2)$-critical value from the Student's $t$-distribution with $n - 1$ degrees of freedom:

Inverting the inequality yields the exact $100(1 - \alpha)\%$ confidence interval for $\mu$:

Comparison with Known Variance Case

Because Student's $t$-distribution incorporates the additional sampling variability of estimating $\sigma$ by $S$, its probability density has heavier tails than the standard normal distribution. Consequently:

The confidence interval is wider, formalizing the penalty of variance uncertainty. Furthermore, the length of the interval $W = 2 t_{n-1, 1-\alpha/2} \frac{S}{\sqrt{n}}$ is a random variable that varies from sample to sample according to the realized sample variance $S^2$. As $n \to \infty$, by Slutsky's Theorem, $S \xrightarrow{P} \sigma$ and $t_{n-1} \xrightarrow{d} \mathcal{N}(0, 1)$, recovering the Gaussian interval.

Paired-Sample Inferences and Matched-Pairs Design

In many experimental settings, observations in two groups are not independent, but intrinsically linked. For example, in a medical study of an antihypertensive compound, systolic blood pressure is recorded on the same $n$ patients before treatment ($X_i$) and after treatment ($Y_i$). Because observations $(X_i, Y_i)$ are coupled on the same subject, they exhibit positive correlation $\text{Cov}(X_i, Y_i) = \rho \sigma_X \sigma_Y > 0$. Applying a naive two-sample independent pooled $t$-interval would be catastrophic: the variance of the difference under independence is $\frac{\sigma_X^2 + \sigma_Y^2}{n}$, whereas the true variance of the pairwise difference is:

To exploit this variance reduction, we define the univariate difference series:

Assuming $D_i \stackrel{\text{i.i.d.}}{\sim} \mathcal{N}(\mu_D, \sigma_D^2)$ where $\mu_D = \mu_X - \mu_Y$, the problem reduces to a single-sample Student's $t$ inference on the differences:

• Sample mean difference: $\bar{D} = \frac{1}{n} \sum_{i=1}^n D_i = \bar{X} - \bar{Y}$
• Sample standard deviation of differences: $S_D = \sqrt{\frac{1}{n-1} \sum_{i=1}^n (D_i - \bar{D})^2}$

The exact $100(1 - \alpha)\%$ paired $t$ confidence interval for $\mu_D$ is:

By subtracting out subject-specific baseline variability, the matched-pairs design dramatically narrows the confidence interval, yielding immense statistical power.

Now suppose the primary parameter of scientific interest is the population variance $\sigma^2$ (or population standard deviation $\sigma$, or precision $\tau = 1/\sigma^2$), while the mean $\mu$ is an unknown nuisance parameter.

The Chi-Square Pivot

Recall from normal sample distribution theory that:

follows a Chi-square distribution with $\nu = n - 1$ degrees of freedom:

Because the distribution $\chi_{n-1}^2$ depends solely on the sample size $n$ and contains neither $\mu$ nor $\sigma^2$, $V(\mathbf{X}, \sigma^2)$ is an exact pivotal quantity for $\sigma^2$.

Inversion and Interval Construction

Unlike the Gaussian and Student's $t$ distributions, the Chi-square distribution is strictly non-negative and asymmetric (skewed to the right). To construct an exact $(1 - \alpha)$ confidence interval, select two critical cutoffs $\chi_{n-1, \alpha/2}^2$ and $\chi_{n-1, 1 - \alpha/2}^2$ from the lower and upper tails:

where $\mathbb{P}(\chi_{n-1}^2 \le \chi_{n-1, \alpha/2}^2) = \alpha / 2$ and $\mathbb{P}(\chi_{n-1}^2 \le \chi_{n-1, 1 - \alpha/2}^2) = 1 - \alpha / 2$. Invert the compound inequality by taking reciprocals (which reverses the inequalities):

Multiplying through by $(n - 1)S^2 > 0$ yields the exact $100(1 - \alpha)\%$ confidence interval for $\sigma^2$:

Confidence Intervals for Standard Deviation and Precision

By the monotonicity of the square root function, the exact $100(1 - \alpha)\%$ confidence interval for the population standard deviation $\sigma$ is obtained by taking the square roots of the endpoints:

Similarly, inverting the variance endpoints gives the exact confidence interval for the precision parameter $\tau = \frac{1}{\sigma^2}$:

Notice that because the Chi-square density is asymmetric, equal-tailed cutoffs do not minimize the expected interval length $\mathbb{E}[U - L]$. Tate and Klett (1959) tabulated shortest-length critical values $(c_1, c_2)$ that satisfy the optimality condition $f_{\chi^2}(c_1) c_1^2 = f_{\chi^2}(c_2) c_2^2$, producing strictly shorter confidence intervals for $\sigma^2$.

In comparative scientific experiments (such as clinical drug trials or agricultural A/B treatments), researchers compare two independent groups:

• Sample 1: $X_1, \dots, X_{n_1} \stackrel{\text{i.i.d.}}{\sim} \mathcal{N}(\mu_1, \sigma^2)$
• Sample 2: $Y_1, \dots, Y_{n_2} \stackrel{\text{i.i.d.}}{\sim} \mathcal{N}(\mu_2, \sigma^2)$

where both populations share a common, unknown variance $\sigma^2$. The target parameter is the difference in population means $\tau = \mu_1 - \mu_2$.

The Pooled Sample Variance

The point estimator of $\mu_1 - \mu_2$ is $\bar{X} - \bar{Y}$. Its variance is:

To estimate the common variance $\sigma^2$, we pool the sum of squared deviations from both samples:

Because $\frac{(n_1 - 1)S_1^2}{\sigma^2} \sim \chi_{n_1 - 1}^2$ and $\frac{(n_2 - 1)S_2^2}{\sigma^2} \sim \chi_{n_2 - 1}^2$ are independent, their sum is Chi-square:

The Two-Sample Pooled $t$ Pivot

Forming the Studentized ratio:

Inverting this exact pivot yields the exact $100(1 - \alpha)\%$ confidence interval for $\mu_1 - \mu_2$:

What if the two populations possess unequal, unknown variances $\sigma_1^2 \ne \sigma_2^2$? This is the celebrated Behrens–Fisher Problem, which baffled Ronald Fisher, Jerzy Neyman, and Harold Jeffreys. When $\sigma_1^2 \ne \sigma_2^2$, the true variance of $\bar{X} - \bar{Y}$ is:

The natural plug-in standard error is:

Consider the ratio:

Remarkably, Linnik (1966) proved that no exact pivotal quantity exists for $\mu_1 - \mu_2$ based on sufficient statistics! The distribution of $T^*$ depends on the unknown variance ratio $\theta = \sigma_1^2 / \sigma_2^2$.

The Welch–Satterthwaite Solution

B. L. Welch (1938) and F. E. Satterthwaite (1946) proposed approximating the distribution of the denominator variance $V = \frac{S_1^2}{n_1} + \frac{S_2^2}{n_2}$ by a scaled Chi-square distribution $\frac{c \chi_\nu^2}{\nu}$ by matching the first two moments:

Matching $\text{Var}(V) = \frac{2(\mathbb{E}[V])^2}{\nu}$ yields the Welch–Satterthwaite effective degrees of freedom:

The approximate $100(1 - \alpha)\%$ Welch confidence interval is:

Notice that $\min(n_1 - 1, n_2 - 1) \le \nu^* \le n_1 + n_2 - 2$. In modern applied statistics, Welch's interval is almost universally preferred over the pooled $t$-interval because it maintains robust, near-exact nominal coverage even when variances differ by orders of magnitude.

Fieller's Theorem for the Ratio of Two Normal Means

A notoriously difficult problem in bioassay, pharmacokinetics, and econometrics is estimating the ratio of two population means:

for example, the relative potency of two pharmaceutical formulations or the price elasticity of demand. Let $(\bar{X}_1, \bar{X}_2)$ be sample means from bivariate normal data with known or estimated covariance matrix $\text{Var}(\bar{X}_1) = v_{11}$, $\text{Var}(\bar{X}_2) = v_{22}$, and $\text{Cov}(\bar{X}_1, \bar{X}_2) = v_{12}$. A naive plug-in estimator $\hat{\theta} = \bar{X}_1 / \bar{X}_2$ has a Cauchy-like heavy-tailed distribution with infinite mean and infinite variance! E. C. Fieller (1940, 1954) resolved this problem by constructing an exact linear pivot. Consider the auxiliary quantity:

Notice that $\mathbb{E}[Y(\theta)] = \mu_1 - \theta \mu_2 = \mu_1 - \left( \frac{\mu_1}{\mu_2} \right) \mu_2 = 0$. The variance of $Y(\theta)$ is:

The standardized statistic follows an exact Student's $t$ (or standard normal) distribution:

Inverting $T(\theta)^2 \le t_{\nu, 1-\alpha/2}^2 = t^2$ produces a quadratic inequality in $\theta$:

Let $A = \bar{X}_2^2 - t^2 v_{22}$, $B = \bar{X}_1 \bar{X}_2 - t^2 v_{12}$, and $C = \bar{X}_1^2 - t^2 v_{11}$. Depending on whether $A > 0$ or $A \le 0$, Fieller's theorem generates three distinct geometric topologies:

1. Bounded Interval ($A > 0$): If the denominator $\bar{X}_2$ is significantly different from zero at level $\alpha$ ($t_{\text{denom}}^2 = \bar{X}_2^2 / v_{22} > t^2$), the parabola opens upward, yielding a standard finite confidence interval:

2. Complement of an Interval ($A < 0, B^2 - AC > 0$): If the denominator is not significantly different from zero, the confidence set is the union of two unbounded rays $(-\infty, \theta_1] \cup [\theta_2, \infty)$.

3. The Entire Real Line ($A < 0, B^2 - AC < 0$): If neither the numerator nor denominator is distinguishable from zero, the confidence set is $(-\infty, \infty)$.

Fieller's theorem demonstrates that demanding guaranteed frequentist coverage for non-linear ratios inherently requires accepting unbounded confidence sets when data carry insufficient information.

To compare the precision or dispersion of two independent normal populations, the target parameter is the variance ratio:

The $F$-Distribution Pivot

Recall that $V_1 = \frac{(n_1 - 1)S_1^2}{\sigma_1^2} \sim \chi_{n_1 - 1}^2$ and $V_2 = \frac{(n_2 - 1)S_2^2}{\sigma_2^2} \sim \chi_{n_2 - 1}^2$ are independent. The ratio of two independent Chi-square variables divided by their degrees of freedom follows Snedecor's $F$-distribution:

Because the distribution $F_{n_1 - 1, n_2 - 1}$ depends solely on sample sizes $n_1$ and $n_2$, this ratio is an exact pivotal quantity for $\lambda = \frac{\sigma_1^2}{\sigma_2^2}$.

Inversion and Interval Construction

Select critical values $F_{n_1 - 1, n_2 - 1, \alpha/2}$ and $F_{n_1 - 1, n_2 - 1, 1 - \alpha/2}$:

Inverting yields the exact $100(1 - \alpha)\%$ confidence interval for $\frac{\sigma_1^2}{\sigma_2^2}$:

Using the reciprocal property of $F$-quantiles, $F_{\nu_1, \nu_2, \alpha/2} = \frac{1}{F_{\nu_2, \nu_1, 1 - \alpha/2}}$, the interval can be rewritten equivalently as:

Equivariant Interval Estimation in Location-Scale Families

In a location-scale family $f(x; \mu, \sigma) = \frac{1}{\sigma} f_0\left( \frac{x - \mu}{\sigma} \right)$, physical measurements are subject to changes in measurement origin (translation $X \to X + c$) and scale units (dilation $X \to a X$ for $a > 0$). A confidence interval $[L(\mathbf{X}), U(\mathbf{X})]$ is said to be equivariant under the affine group if:

E. J. G. Pitman (1939) proved that every equivariant estimator and confidence interval can be expressed in terms of the sample mean $\bar{X}$, sample standard deviation $S$, and the configuration ancillary statistic:

The optimal equivariant confidence interval conditions on the realized configuration vector $\mathbf{C} = \mathbf{c}$, adjusting the interval width to the exact geometric shape of the realized data sample. For non-normal location-scale distributions (such as Cauchy, logistic, or extreme value distributions), Pitman's conditional equivariant intervals provide exact confidence levels that strictly dominate unconditional intervals.

Equivariant Interval Estimation in Location-Scale Families

In a location-scale family $f(x; \mu, \sigma) = \frac{1}{\sigma} f_0\left( \frac{x - \mu}{\sigma} \right)$, physical measurements are subject to changes in measurement origin (translation $X \to X + c$) and scale units (dilation $X \to a X$ for $a > 0$). A confidence interval $[L(\mathbf{X}), U(\mathbf{X})]$ is said to be equivariant under the affine group if:

E. J. G. Pitman (1939) proved that every equivariant estimator and confidence interval can be expressed in terms of the sample mean $\bar{X}$, sample standard deviation $S$, and the configuration ancillary statistic:

The optimal equivariant confidence interval conditions on the realized configuration vector $\mathbf{C} = \mathbf{c}$, adjusting the interval width to the exact geometric shape of the realized data sample. For non-normal location-scale distributions (such as Cauchy, logistic, or extreme value distributions), Pitman's conditional equivariant intervals provide exact confidence levels that strictly dominate unconditional intervals.

What makes one confidence interval better than another? Two competing mathematical criteria exist:

1. Geometric Criterion (Shortest Expected Length): Minimize the expected Lebesgue measure (length or volume) $\mathbb{E}_\theta[U(\mathbf{X}) - L(\mathbf{X})]$ subject to maintaining coverage probability $\ge 1 - \alpha$.

2. Probabilistic Criterion (Uniformly Most Accurate / UMA): A confidence set $S(\mathbf{X})$ is said to be Uniformly Most Accurate (UMA) if it minimizes the probability of covering false parameter values:

for every competing $(1 - \alpha)$ confidence set $S^*$.

Pratt's Theorem Connecting Length and False Coverage

John W. Pratt (1961) discovered a profound identity linking these two criteria: Let $S(\mathbf{X})$ be a confidence set for $\theta$. The expected length of $S(\mathbf{X})$ under true parameter $\theta_0$ is mathematically identical to the integrated probability of covering false values:

Significance:

Pratt's Theorem reveals that minimizing the probability of false coverage simultaneously minimizes the expected length of the confidence interval! Inverting Uniformly Most Powerful (UMP) hypothesis tests produces Uniformly Most Accurate confidence sets with minimal expected length.

Tolerance Intervals and Prediction Intervals vs Confidence Intervals

Practitioners frequently confuse three distinct statistical interval concepts:

1. Confidence Interval for a Parameter $\theta$: Brackets an unobserved, fixed parameter of the population (such as the mean $\mu$):

As $n \to \infty$, the length of a confidence interval shrinks to zero: $\mathcal{O}(1/\sqrt{n}) \to 0$.

2. Prediction Interval for a Future Observation $X_{n+1}$: Brackets a single future random realization drawn from the same population:

For a normal population, the prediction error is $X_{n+1} - \bar{X} \sim \mathcal{N}\left(0, \sigma^2(1 + 1/n)\right)$. The exact prediction interval is:

As $n \to \infty$, the prediction interval does not shrink to zero! Its width converges to the population spread $2 z_{1 - \alpha/2} \sigma$.

3. Statistical Tolerance Interval: Brackets a specified proportion $\gamma$ (e.g., $99\%$) of the entire population with a pre-assigned confidence level $1 - \alpha$ (e.g., $95\%$):

In industrial manufacturing, tolerance intervals ensure that at least $99\%$ of manufactured ball bearings meet dimensional specifications with $95\%$ statistical confidence.

Pratt's Paradox and the Role of Conditioning in Interval Estimation

In 1961, John W. Pratt formulated an intriguing paradox regarding the blind minimization of expected interval length. Suppose an investigator wishes to estimate the mean $\mu$ of a normal population. An ancillary coin flip chooses whether to draw a sample of size $n_1 = 10$ (yielding variance $\sigma^2 / 10$) or a sample of size $n_2 = 1000$ (yielding variance $\sigma^2 / 1000$). If the experimenter minimizes the overall unconditional expected length across both outcomes:

subject only to the unconditional coverage constraint $\frac{1}{2} C_1 + \frac{1}{2} C_2 = 0.95$, the mathematical optimizer will choose $C_1 = 0$ (reporting an empty interval whenever $n_1 = 10$ occurs) and compensate by setting $C_2 = 0.999$ to achieve a tiny unconditional expected length! This strategy is scientifically absurd: when the small sample $n_1 = 10$ is realized, the experimenter reports an interval known to be empty, while claiming $95\%$ confidence! Pratt's paradox establishes that minimizing expected length unconditionally can lead to unacceptable inferential rules. Valid interval estimation must condition on ancillary statistics (such as realized sample size), enforcing conditional coverage $\mathbb{P}(L \le \theta \le U \mid \text{Ancillary}) = 1 - \alpha$ for every realized experimental condition.

Jerzy Neyman established a fundamental mathematical duality between confidence sets and hypothesis testing.

Theorem: Neyman's Test Inversion Duality

1. For each candidate parameter value $\theta_0 \in \Theta$, let $A(\theta_0) \subset \mathcal{X}$ denote the acceptance region of a level-$\alpha$ hypothesis test of $H_0: \theta = \theta_0$ versus $H_1: \theta \ne \theta_0$:

For an observed sample $\mathbf{x}$, define the set:

Then $S(\mathbf{X})$ is a $100(1 - \alpha)\%$ confidence set for $\theta$.

1. Conversely, if $S(\mathbf{X})$ is a $(1 - \alpha)$ confidence set, defining the acceptance region $A(\theta_0) = \{ \mathbf{x} : \theta_0 \in S(\mathbf{x}) \}$ yields a valid level-$\alpha$ test of $H_0: \theta = \theta_0$.

Test inversion provides a universal engine for constructing confidence intervals: to find the confidence interval, simply gather all parameter values that cannot be rejected at significance level $\alpha$ by the likelihood ratio test.

Fiducial Inference and the Neyman–Fisher Controversy

The foundations of interval estimation witnessed one of the most celebrated intellectual debates in twentieth-century science between Ronald A. Fisher and Jerzy Neyman. In 1930, Fisher introduced the concept of fiducial distributions, attempting to obtain an objective inverse probability distribution over the parameter space without postulating an a priori prior distribution.

Fisher reasoned that if $T$ is a continuous pivotal quantity whose distribution depends monotonically on $\theta$, then observing $T = t$ transfers the stochastic uncertainty directly to $\theta$:

Fisher defined the resulting measure as the fiducial probability density of $\theta$. For single-parameter location models, fiducial intervals coincide numerically with Neyman confidence intervals and Bayesian credible intervals under uniform priors.

However, in 1937, Jerzy Neyman formalized the confidence interval framework based strictly on repeated sampling behavior (frequentist coverage probability). Neyman proved that confidence intervals require no inverse probability over the parameter; instead, the interval boundaries $L(\mathbf{X})$ and $U(\mathbf{X})$ are random variables that cover the fixed, deterministic parameter $\theta$ with pre-assigned probability $1 - \alpha$.

The controversy culminated when Richard von Mises and David Blackwell demonstrated that in multiparameter problems (such as the Behrens–Fisher problem of comparing means of two normal populations with unequal variances), the fiducial approach fails to satisfy the law of total probability and can produce mutually contradictory conclusions depending on the chosen step-by-step reduction. Consequently, modern mathematical statistics embraces Neyman’s rigorous confidence framework alongside Bayesian posterior distributions, while recognizing fiducial theory as a brilliant historical attempt to synthesize objective inductive inference.