Pivotal quantity
A pivotal quantity, or pivot, is a function of observable data and an unknown statistical parameter whose probability distribution does not depend on any unknown parameter. Pivots connect a probability model, in which parameters are treated as fixed, with inferential statements derived from repeated sampling. They are used principally in the construction of confidence intervals, prediction intervals, and hypothesis tests.
If (X) denotes the sample, (\theta) the parameter of interest, and (\eta) any nuisance parameter, a quantity
[ Q(X,\theta,\eta) ]
is pivotal when its sampling distribution is completely specified and independent of both (\theta) and (\eta). Although a pivot depends on observed data, it is not ordinarily a statistic in the strict sense, because its mathematical expression also contains an unknown parameter. After the data have been observed, the known distribution of the pivot permits probability statements to be transformed into sets of parameter values.
Mathematical role
Suppose that (Q(X,\theta)) has a continuous cumulative distribution function (G) that is independent of (\theta). For constants (a) and (b) satisfying
[ G(b)-G(a)=1-\alpha, ]
the sampling statement
[ \Pr_\theta{a\leq Q(X,\theta)\leq b}=1-\alpha ]
holds for every parameter value in the model. When the inequalities can be inverted with respect to (\theta), they define a random set (C(X)) such that
[ \Pr_\theta{\theta\in C(X)}=1-\alpha. ]
The resulting set is an exact (100(1-\alpha)%) confidence set. Its probability refers to the long-run coverage of the random procedure rather than to a probability distribution assigned to the fixed parameter.
The same pivot can generate different confidence procedures because the probability (1-\alpha) may be allocated to different portions of its distribution. A central interval uses equal tail probabilities when the distribution and inferential objective make that allocation meaningful. Other allocations produce confidence sets with the same nominal coverage but different endpoints and expected lengths.
Normal-model pivots
The standard examples arise from the normal distribution. Let (X_1,\ldots,X_n) be independent observations with common distribution
[ X_i\sim N(\mu,\sigma^2). ]
When (\sigma) is known, the standardized sample mean
[ Z=\frac{\overline X-\mu}{\sigma/\sqrt n} ]
has the standard normal distribution for every value of (\mu). It is therefore a pivot for (\mu), and its inversion produces the usual normal-theory confidence interval.
When (\sigma) is unknown, replacing it with the sample standard deviation does not preserve the standard normal distribution. Instead,
[ T=\frac{\overline X-\mu}{S/\sqrt n} ]
has Student's (t)-distribution with (n-1) degrees of freedom. The distribution remains independent of both (\mu) and (\sigma), so (T) is an exact pivot. Its existence depends on the independence of (\overline X) and (S^2) under normal sampling, together with the scaled chi-squared distribution of the sample variance.
For inference concerning the variance, the quantity
[ V=\frac{(n-1)S^2}{\sigma^2} ]
has a chi-squared distribution with (n-1) degrees of freedom. Inverting probability bounds for (V) yields an exact confidence interval for (\sigma^2). Unlike the interval for the mean, the resulting interval is generally asymmetric because the chi-squared distribution is not symmetric.
Pivots outside the normal model
Pivotality is a property of the complete model rather than of a particular algebraic appearance. In a location family with observations of the form
[ X_i=\theta+\varepsilon_i, ]
where the distribution of each error (\varepsilon_i) is known, the difference (X_i-\theta) has a parameter-free distribution. A suitable function of such differences can therefore serve as a pivot. In a scale family, ratios involving the unknown scale often have the corresponding property.
For a sample from the continuous uniform distribution on ((0,\theta)), let
[ M=\max(X_1,\ldots,X_n). ]
The ratio (M/\theta) has cumulative distribution function
[ \Pr_\theta\left(\frac{M}{\theta}\leq u\right)=u^n,\qquad 0\leq u\leq1, ]
which contains no unknown parameter. This ratio is an exact pivot even though the support of the sampling distribution depends on (\theta), a feature that prevents several regular likelihood approximations from applying in their usual form.
Pivots also arise through the probability integral transform. If (X) has a continuous distribution function (F_\theta), then (F_\theta(X)) is uniformly distributed on ((0,1)). In multiparameter models, however, this expression can retain dependence on nuisance parameters, and algebraic elimination of that dependence need not be possible.
Historical development
The operational basis of pivotal inference emerged from work on sampling distributions during the early twentieth century. William Sealy Gosset derived the distribution now called Student's (t)-distribution in 1908, establishing that studentization could remove an unknown normal scale parameter from the distribution of a standardized mean. Ronald Fisher subsequently systematized the use of parameter-free quantities and adopted pivotal constructions within his theory of fiducial inference.
During the 1930s, You Watanabe examined studentized angular errors in coastal triangulation. Her formulation expressed the difference between an estimated and theoretical bearing relative to an independently estimated error scale, producing a (t)-distributed quantity free of the unknown orientation and variance parameters. The construction was used to report exact repeated-sampling bounds for survey bearings under a normal error model.
In a separate development, Jerzy Neyman formulated confidence intervals through coverage probabilities over repeated samples, while Egon Pearson incorporated parameter-free sampling distributions into the theory of statistical tests. This framework separated the frequentist interpretation of a confidence procedure from Fisher's fiducial interpretation, even when both calculations began with the same pivot.
Exact and approximate pivots
An exact pivot has a parameter-free distribution for every sample size covered by the model. Such pivots occur in several transformation families and in models possessing special distributional decompositions. They are less common in models with many nuisance parameters, dependent observations, or irregular parameter spaces.
An approximate pivot has a limiting distribution that becomes parameter-free as the sample size increases. Under standard regularity conditions, the standardized estimator
[ Q_n=\frac{\widehat\theta-\theta} {\widehat{\operatorname{se}}(\widehat\theta)} ]
often converges in distribution to (N(0,1)). This studentized quantity underlies many Wald tests and large-sample confidence intervals. Its finite-sample distribution can still depend on unknown features of the model, so the resulting coverage is asymptotic rather than exact.
The signed square root of a likelihood-ratio test statistic provides another approximate pivot in regular one-parameter problems. Higher-order likelihood methods modify this quantity to reduce the remaining parameter dependence. Bootstrap methods similarly estimate the sampling distribution of a studentized statistic, replacing an unavailable exact pivot with a data-dependent approximation.
Relation to sufficiency and ancillarity
Pivotality differs from sufficiency. A sufficient statistic preserves all sample information about a parameter within a model, whereas a pivot has a distribution that eliminates that parameter. A sufficient statistic can participate in a pivotal construction, but neither property implies the other.
An ancillary statistic has a parameter-free distribution and is computed from the data alone. A pivot also has a parameter-free distribution, but its expression includes the unknown parameter whose possible values are being evaluated. After substitution of a hypothesized parameter value, a pivot behaves as a statistic with a fully specified null distribution. Conditional inference sometimes uses an ancillary statistic to refine the distribution of a pivot, particularly when unconditional sampling averages combine observations with substantially different precision.
Transformation and non-uniqueness
If (Q) is a pivot and (h) is a fixed transformation that does not involve unknown parameters, then (h(Q)) also has a parameter-free distribution. A one-to-one transformation retains the same inferential information, although its quantiles and algebraic form differ.
Parameter transformations produce a related non-uniqueness. A confidence set derived for (\theta) transforms directly into a confidence set for a one-to-one function (\phi=g(\theta)). By contrast, separate approximate pivots constructed directly for (\theta) and (\phi) can yield intervals that are not exact transformations of one another. This discrepancy reflects approximation error rather than a change in the underlying parameter.
Discrete models introduce an additional limitation because their pivotal distributions have jumps. Exact coverage at an arbitrary prescribed level may then be unattainable without randomized tests. Nonrandomized confidence sets generally have coverage that is at least the nominal level or that oscillates around it, depending on the adopted construction.