Progressive Interval Estimation

Updated 4 July 2026

Progressive interval estimation is a suite of methods that produces interval-valued conclusions via sequential sampling, censoring, or online refinement.
These techniques are applied in reliability analysis, survival studies, and nonlinear system state estimation, offering both deterministic and stochastic guarantees.
The approach combines methods like bisection coverage tuning, pivotal and bootstrap inference, and adaptive censoring to ensure accurate uncertainty quantification.

Searching arXiv for relevant papers on progressive interval estimation and closely related usages of the term. arxiv_search(query="\"progressive interval estimation\" OR \"multistage estimation\" OR \"progressive type-I interval censoring\" interval estimation", max_results=10) Progressive interval estimation denotes a family of inferential procedures in which interval-valued conclusions are produced under a progressive acquisition, censoring, or refinement mechanism. In the arXiv literature, the expression spans several technically distinct settings: multistage sampling that stops when a sequential random interval attains prescribed coverage; confidence or credible interval construction for lifetime models observed under progressive censoring or interval censoring; nonparametric estimation of conditional future-entry distributions from current-status progressive multistate data; and interval state estimation in nonlinear dynamical systems by progressively tightening bounded-error enclosures (0809.1241, Mondal et al., 21 Jan 2025, Anyaso-Samuel et al., 2024, Xu et al., 19 Sep 2025).

1. Terminological scope and conceptual distinctions

A central distinction is that the word interval does not always refer to the same object. In Chen’s multistage framework, the interval is the inferential output: a sequential random interval $(L(\hat\theta,n),U(\hat\theta,n))$ with prescribed coverage probability, where the sample size $n$ is itself random because it is determined by a stopping rule (0809.1241). In reliability and survival settings, by contrast, interval may refer to the observation mechanism, as in progressive type-I interval censoring, where failures are only localized to inspection intervals rather than observed exactly (Teimouri, 2018).

The word progressive is equally overloaded. It may denote stagewise sampling, as in multistage estimation; progressive removal of units during a life test, as in progressive type-I, progressive type-I hybrid, improved adaptive Type-II progressive censoring, or block adaptive progressive Type-II censoring; or online refinement of a state enclosure, as in zonotope-based interval observers for bounded Jacobian nonlinear systems (Dutta et al., 2021, Zhang, 2023, Xu et al., 19 Sep 2025).

This distinction matters because the technical guarantees differ. Some frameworks target exact finite-sample coverage, some rely on asymptotic normality, some use generalized pivotal quantities or bootstrap calibration, and some produce deterministic set inclusions rather than probabilistic confidence statements. A common source of confusion is therefore the conflation of interval censoring with interval estimation: the former is a data-collection scheme, whereas the latter is an inferential objective. The literature surveyed here contains both.

2. Sequential random intervals in multistage estimation

The most explicit formalization of progressive interval estimation appears in Chen’s “A New Framework of Multistage Estimation” (0809.1241). There, sampling is carried out in stages. At stage $\ell$ , the available sample size is $n_\ell$ , the decision variable is $D_\ell\in\{0,1\}$ , and the procedure stops at the first stage for which $D_\ell=1$ . The output is a sequential random interval satisfying

$\Pr\{L(\hat\theta,n)<\theta<U(\hat\theta,n)\mid \theta\}\ge 1-\delta,\qquad \forall \theta\in\Theta.$

This framework is explicitly presented as broad enough to subsume fixed-sample-size interval estimation, point estimation with error control, bounded-width confidence intervals, confidence sequences, interval estimation following hypothesis testing, and multistage estimation under absolute, relative, or mixed error criteria. The inferential core is the construction of a controlling confidence sequence $[L_\ell,U_\ell]$ together with an inclusion principle: sampling stops when the desired interval at stage $\ell$ contains the controlling confidence sequence. Theorem 3 formalizes the resulting coverage transfer, while Theorem 2 gives an abstract stopping-rule guarantee based on stagewise tail controls (0809.1241).

Two technical devices organize the theory. The first is the unimodal-likelihood estimator (ULE), which is more general than the MLE and is used to prove monotonicity of one-sided tail probabilities. The second is coverage tuning, based on a parameter $\zeta>0$ . By making $n$ 0 sufficiently small, the minimum coverage can be forced above the nominal level; by bisection coverage tuning, $n$ 1 is then made as large as possible while retaining the guarantee. Complementary coverage probabilities

$n$ 2

are checked using monotonicity and reduction-to-extrema arguments rather than exhaustive search (0809.1241).

The framework is notable for coupling theory to computation. The paper introduces bisection coverage tuning, the Adaptive Maximum Checking Algorithm (AMCA), Adapted Branch and Bound (ABB), recursive computation, domain truncation, consecutive-decision-variable bounding, triangular partition, and interval splitting. Concrete schemes are then developed for binomial parameters, finite population proportions, Poisson means, normal means and variances, exponential parameters, gamma scale, multinomial and hypergeometric proportions, regression, and quantiles (0809.1241).

3. Progressive censoring schemes as the observational basis for interval inference

In lifetime analysis, progressive interval estimation is often inseparable from the censoring design. The main schemes appearing in the arXiv literature surveyed here differ by when failures are inspected, when units are withdrawn, and whether stopping is governed by a failure count, a time threshold, or both.

Design	Observation structure	Interval-related procedure
Progressive type-I interval censoring	Fixed inspection times $n$ 3; counts $n$ 4 in $n$ 5; random removals $n$ 6 at inspections	Numerical ML and corrected EM; bias and MSE reported, but no coverage probabilities or confidence intervals (Teimouri, 2018)
Progressive type-I hybrid censoring	Progressive removals until either the $n$ 7-th failure or time $n$ 8	Normal-approximation intervals, log-transformed MLE intervals, and HPD credible intervals (Dutta et al., 2021)
Improved adaptive Type-II progressive censoring	Warning time $n$ 9, maximum time $\ell$ 0; after $\ell$ 1 the test is accelerated and no units are censored	Approximate confidence intervals from asymptotic normality and observed Fisher information (Zhang, 2023)
Block adaptive progressive Type-II censoring	Sample split into facilities/blocks; each block uses adaptive progressive Type-II censoring	Asymptotic confidence intervals and generalized confidence intervals for parameters and reliability functionals (Mondal et al., 21 Jan 2025)

For generalized exponential lifetimes, the progressive type-I interval censoring scheme is represented by $\ell$ 2, with $\ell$ 3. The likelihood under a general cdf $\ell$ 4 is written as

$\ell$ 5

and, for the generalized exponential model with parameterization $\ell$ 6 of Mudholkar and Srivastava (1993), must be maximized numerically (Teimouri, 2018).

For progressive type-I hybrid censoring of the logistic exponential distribution, the experiment starts with $\ell$ 7 units, a prefixed failure count $\ell$ 8, and a prefixed time $\ell$ 9. If $n_\ell$ 0, the test ends at the $n_\ell$ 1-th failure; otherwise it ends at time $n_\ell$ 2, with all remaining units withdrawn then. The unified likelihood uses $n_\ell$ 3 in the first case and $n_\ell$ 4 in the second, where $n_\ell$ 5 is the number of failures observed before $n_\ell$ 6 (Dutta et al., 2021).

Improved adaptive Type-II progressive censoring for the Chen distribution introduces two thresholds, $n_\ell$ 7 and $n_\ell$ 8, following Yan et al. (2021). Before $n_\ell$ 9, the planned censoring scheme is followed; after $D_\ell\in\{0,1\}$ 0, the test is accelerated and no units are censored; the experiment is guaranteed to stop no later than $D_\ell\in\{0,1\}$ 1. This design is presented as addressing a drawback of adaptive Type-II progressive censoring, namely that the observed number of failures can become too small (Zhang, 2023).

Block adaptive progressive Type-II censoring extends adaptive progressive censoring to multiple facilities. For facility $D_\ell\in\{0,1\}$ 2, one observes failure times $D_\ell\in\{0,1\}$ 3, block-specific removals $D_\ell\in\{0,1\}$ 4, a threshold time $D_\ell\in\{0,1\}$ 5, and the number $D_\ell\in\{0,1\}$ 6 of failures occurring before that threshold. The full likelihood factorizes across blocks, with facility-specific $D_\ell\in\{0,1\}$ 7 and a common $D_\ell\in\{0,1\}$ 8 in the inverted exponentiated Pareto model (Mondal et al., 21 Jan 2025).

4. Parametric interval procedures under progressive censoring

The interval procedures built on these censoring schemes are heterogeneous. Some are fully asymptotic Wald-type constructions, some use transformation to preserve parameter constraints, and some replace asymptotic approximations by generalized pivots or posterior simulation.

A methodological caution comes from the generalized exponential model under progressive type-I interval censoring. The central result is not an interval estimator but a correction to Chen and Lio’s 2010 EM algorithm. The corrected method treats exact failure times within inspection intervals and exact failure times of withdrawn units as latent data, forms

$D_\ell\in\{0,1\}$ 9

and only then maximizes $D_\ell=1$ 0. The paper emphasizes that Chen and Lio were incorrect because they effectively took expectations after differentiating the complete-data log-likelihood. In simulations with $D_\ell=1$ 1, $D_\ell=1$ 2, $D_\ell=1$ 3, $D_\ell=1$ 4, four censoring-removal scenarios, and 1000 Monte Carlo replications, the corrected EM and the earlier EM-Chen produced identical $D_\ell=1$ 5 but different $D_\ell=1$ 6; the corrected EM had smaller MSE for the rate parameter in all four scenarios and smaller bias in three of the four. The paper does not report coverage probabilities or confidence intervals, which underscores that progressive interval censoring and interval estimation are not synonymous (Teimouri, 2018).

For the logistic exponential distribution under progressive type-I hybrid censoring, three interval constructions are compared. The first is the ordinary normal approximation based on the observed Fisher information and the asymptotic normality of $D_\ell=1$ 7. The second applies a log transformation to avoid negative lower limits and improve coverage behavior under the positivity constraints $D_\ell=1$ 8 and $D_\ell=1$ 9. The third uses posterior simulation by importance sampling to construct 90% and 95% HPD credible intervals, following the Monte Carlo HPD construction of Chen and Shao (1999). The simulation study reports average length and empirical coverage probability; the main pattern is that the normal approximation intervals are shorter than the log-transformed intervals, but the HPD credible intervals are the shortest among the compared intervals (Dutta et al., 2021).

For the Chen distribution under improved adaptive Type-II progressive censoring, interval estimation is classical and asymptotic. MLEs are obtained from the likelihood, with a closed-form $\Pr\{L(\hat\theta,n)<\theta<U(\hat\theta,n)\mid \theta\}\ge 1-\delta,\qquad \forall \theta\in\Theta.$ 0 given $\Pr\{L(\hat\theta,n)<\theta<U(\hat\theta,n)\mid \theta\}\ge 1-\delta,\qquad \forall \theta\in\Theta.$ 1 and an iterative solution $\Pr\{L(\hat\theta,n)<\theta<U(\hat\theta,n)\mid \theta\}\ge 1-\delta,\qquad \forall \theta\in\Theta.$ 2 for $\Pr\{L(\hat\theta,n)<\theta<U(\hat\theta,n)\mid \theta\}\ge 1-\delta,\qquad \forall \theta\in\Theta.$ 3. Approximate confidence intervals are then built from the inverse observed information matrix. The reported interval-performance metrics are coverage probabilities and average lengths for 95.5% intervals. The simulation summaries state that the true values $\Pr\{L(\hat\theta,n)<\theta<U(\hat\theta,n)\mid \theta\}\ge 1-\delta,\qquad \forall \theta\in\Theta.$ 4 and $\Pr\{L(\hat\theta,n)<\theta<U(\hat\theta,n)\mid \theta\}\ge 1-\delta,\qquad \forall \theta\in\Theta.$ 5 are generally well covered, that increasing the time threshold tends to increase interval length, especially for $\Pr\{L(\hat\theta,n)<\theta<U(\hat\theta,n)\mid \theta\}\ge 1-\delta,\qquad \forall \theta\in\Theta.$ 6, and that the asymptotic interval for $\Pr\{L(\hat\theta,n)<\theta<U(\hat\theta,n)\mid \theta\}\ge 1-\delta,\qquad \forall \theta\in\Theta.$ 7 is usually more satisfactory than for $\Pr\{L(\hat\theta,n)<\theta<U(\hat\theta,n)\mid \theta\}\ge 1-\delta,\qquad \forall \theta\in\Theta.$ 8 (Zhang, 2023).

For the inverted exponentiated Pareto distribution under block adaptive progressive Type-II censoring, two interval methodologies coexist. The first uses the asymptotic multivariate normality of $\Pr\{L(\hat\theta,n)<\theta<U(\hat\theta,n)\mid \theta\}\ge 1-\delta,\qquad \forall \theta\in\Theta.$ 9, the observed Fisher information, and the delta method to construct intervals not only for the parameters but also for $[L_\ell,U_\ell]$ 0, the reliability function $[L_\ell,U_\ell]$ 1, the hazard rate $[L_\ell,U_\ell]$ 2, and the median time to failure $[L_\ell,U_\ell]$ 3. The second constructs generalized confidence intervals from pivotal quantities. The data are transformed so that $[L_\ell,U_\ell]$ 4 behaves like an adaptive progressive Type-II censored sample from a standard exponential distribution; further transformations produce $[L_\ell,U_\ell]$ 5-distributed pivots for $[L_\ell,U_\ell]$ 6 and $[L_\ell,U_\ell]$ 7. A theorem shows that the pivotal function $[L_\ell,U_\ell]$ 8 is strictly decreasing in $[L_\ell,U_\ell]$ 9, which ensures a unique solution of the pivotal equation. Monte Carlo quantiles of the resulting generalized pivotal distribution then yield generalized confidence intervals. Across the simulation setups, larger $\ell$ 0 reduces bias, variance, and interval length; pivotal estimation often has smaller absolute bias and variance than MLE-based estimation; and generalized confidence intervals are usually shorter than asymptotic confidence intervals (Mondal et al., 21 Jan 2025).

5. Nonparametric conditional future-entry estimation with current-status multistate data

A distinct strand of progressive interval estimation arises in multistate models observed only through current status data. Each subject is seen once, at a random inspection time $\ell$ 1, together with the occupied state $\ell$ 2. The paper “Nonparametric estimation of a future entry time distribution given the knowledge of a past state occupation in a progressive multistate model with current status data” studies directed-tree progressive multistate systems without assuming Markovity and targets conditional future-state occupation probabilities of the form

$\ell$ 3

together with their time-indexed versions $\ell$ 4 and conditional entry-time distributions

$\ell$ 5

The object is explicitly not a transition probability $\ell$ 6, but a lifetime conditional “ever enter” probability (Anyaso-Samuel et al., 2024).

Two fully nonparametric estimators are proposed. The fractional at-risk set estimator (FRE) constructs synthetic pooled states and assigns subject-specific fractional weights $\ell$ 7, interpreted as the estimated probability that a subject has ever reached the conditioning state. Aalen–Johansen and competing-risks machinery is then applied to weighted at-risk sets and estimated transition counts. The ratio-of-marginal-state-occupation-probabilities estimator (PLE) exploits the directed-tree structure to express the conditional quantity as a ratio of downstream occupation probabilities: $\ell$ 8 The identification is based on unique downstream paths and remains valid without Markov assumptions, because the occupation estimator $\ell$ 9 is still consistent under the assumptions cited from Datta and Satten (2001) (Anyaso-Samuel et al., 2024).

The paper does not derive a full asymptotic theory, because isotonic regression, kernel smoothing, and recursive multistate plug-in estimation complicate analysis. It therefore emphasizes finite-sample simulation and bootstrap inference. Both FRE and PLE improve with increasing sample size; biases and mean absolute distances decrease with $\zeta>0$ 0; FRE often has slightly smaller bias, especially for later states; and PLE can propagate estimation error along the path, especially for downstream states. Confidence intervals are constructed by a smoothed bootstrap with an arcsine-square-root transformation,

$\zeta>0$ 1

because a naive bootstrap may fail to capture smoothing bias. In the breast-cancer application based on EORTC trial 10854 data reduced to current-status observations, the probability of eventually reaching distant metastasis after local recurrence given prior local recurrence is estimated as $\zeta>0$ 2 with $\zeta>0$ 3 CI $\zeta>0$ 4 under FRE and $\zeta>0$ 5 with $\zeta>0$ 6 CI $\zeta>0$ 7 under PLE. The purely marginal probability of ever reaching the same state is only about $\zeta>0$ 8, whereas the estimate from the original right-censored data gives $\zeta>0$ 9 under PLE (Anyaso-Samuel et al., 2024).

A control-theoretic usage of progressive interval estimation appears in bounded Jacobian nonlinear systems. The system is observed through a Luenberger-like observer

$n$ 00

with process disturbance, measurement noise, and initial uncertainty represented as zonotopes. Under LMIs indexed by all Jacobians $n$ 01, the estimation error dynamics are input-to-state stable and satisfy a peak-to-peak bound. This yields a coarse point-centered interval

$n$ 02

The enclosure is then tightened by propagating a zonotope for the error dynamics and converting that zonotope to interval form by rowwise $n$ 03 norms. The final comprehensive interval is the elementwise intersection of the coarse bounded-error interval and the zonotope-derived interval: $n$ 04 The method is explicitly described as a progression from a robust point-valued estimate to a feasible zonotope set of the estimation error and then to a refined interval enclosure (Xu et al., 19 Sep 2025).

Taken together, these strands suggest that progressive interval estimation is best understood as an umbrella notion rather than a single procedure. The progressive element may lie in stagewise sampling and stopping, in adaptive withdrawal and censoring, in recovery of latent event structure under coarse observation, or in online set refinement. The interval itself may be a sequential random interval with prescribed coverage, a Wald or log-Wald confidence interval, an HPD credible interval, a generalized confidence interval built from pivots, a bootstrap interval for a nonparametric functional, or a deterministic state enclosure. A plausible implication is that the unifying feature is not a shared formula, but an iterative mechanism that updates uncertainty statements as information is progressively accumulated, selectively removed, probabilistically reconstructed, or dynamically propagated.