Inference on the Reliability Function with Type-I Censored Weibull Data
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: I'm Tom.
Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.
Tom: We're looking at the paper: "Inference on the Reliability Function with Type-I Censored Weibull Data".
Jane: Here's the gist of the paper. Reliability inference based on parametric distributions is an important problem in electrical and mechanical engineering because existing methods often rely on approximations or bootstrap procedures that may not perform satisfactorily when data are censored or sample sizes are small.
Tom: This article introduces a new approach for deriving exact parametric tests and confidence intervals for distributions such as the lognormal, loglogistic, and Weibull.
Jane: Problem Motivation Reliability inference based on parametric distributions is an important problem in electrical and mechanical engineering applications.
Tom: When expressed in alternative forms, the problem also arises in clinical research survival models.
Paper summary: Tom: So the paper, "Inference on the Reliability Function with Type-I Censored Weibull Data", summarizes that reliability inference is really hard when you have censored data and small sample sizes because current methods are often just approximations or bootstrapping >
Jane: They focus on Weibull distributions, which are super common in electrical and mechanical engineering for modeling component lifetimes, but they need exact solutions beyond just the point estimation >
Lu: The authors introduce a new method that uses the Generalized Pivotal Quantity approach to derive exact parametric tests and confidence intervals for these distributions >
Meng: They specifically address the issues with previous exact solutions, like Xiang et al. two thousand fifteen by fixing a glitch they found in how it constructed the solution > <ref:2604.12011#pg1>
Lalam: The method involves transforming the data into Gumbel data because that distribution is better behaved and fits well with the location-scale family structure of Weibull >
Tom: They use regression on the logarithm of the ordered observations to estimate parameters for this transformed distribution, using least squares estimators based on censored or uncensored data >
Jane: They then construct generalized confidence intervals using quantiles from these generalized pivotal quantities, which gives them those exact probabilities they were looking for >
Lu: It’s a clever way to handle the Type-I censoring issue by relating it back to known results for the Gumbel distribution which is free of unknown parameters >
Meng: So what they claim is that this approach works even when data is censored, which is a big deal for practical engineering applications where you often don't have perfect data >
Lalam: The simulation study shows that their method performs much better than the existing method by Xiang et al. two thousand fifteen and it gives coverage probabilities closer to the nominal level of ninety-five percent > <ref:2604.12011#pg1>
Tom: It also showed that while other methods like bootstrapping give under-coverage, this Gumbel-based approach stays closer to the target confidence level >
Jane: The numbers they present in their simulations suggest that if you use this method, you get narrower confidence intervals compared to some of the existing ones, even though it maintains good accuracy >
Conclusion: Tom: So looking at the paper, "Inference on the Reliability Function with Type-I Censored Weibull Data", the main thing is that they solved a glitch in an older exact method by using a Gumbel transformation >
Jane: They used generalized pivotal quantities based on estimating parameters for Gumbel data to create confidence intervals for the reliability function of the Weibull distribution >
Lu: It’s important because it gives us an exact way to do this inference, especially when dealing with Type-I censored data and small sample sizes where approximations fail >
Meng: For someone shipping models, this means you have a more reliable metric for component lifespan that isn't just based on a guess from bootstrapping >
Lalam: Essentially, they propose a new tool to get accurate estimates for reliability quantities without relying on the flawed MLE-based construction mentioned in the paper >
Tom: It balances getting coverage probabilities close to ninety-five percent with having confidence intervals that aren't unnecessarily huge, which is what they showed in their real data example >
Jane: The implication is that for reliability analysis, especially in fields like mechanical engineering or even clinical research survival models, this Gumbel transformation method offers a more robust and accurate way to infer the true reliability function >
Bowen Liu, Samaradasa Weerahandi, Malwane M. A. Ananda
University of Maryland, Kansas City University of Kansas, New Jersey
stat.ME, stat.AP
Submitted: 2026-04-13
Updated: 2026-11-03
Comments: 57 pages, 2 figures. Submitted to Reliability Engineering & System Safety
License: http://creativecommons.org/licenses/by-nc-nd/4.0/
Importance score: 90/100
The gist: Reliability inference based on parametric distributions is an important problem in electrical and mechanical engineering because existing methods often rely on approximations or bootstrap procedures…
Key concepts
- Generalized Pivotal Quantity (GPQ)
- A GPQ is a random quantity designed to simplify inference. It must satisfy two conditions: when applied to observed data, it should reduce directly to the parameter being estimated, and its distribution must not depend on any unknown parameters. This makes it ideal for constructing exact confidence intervals.
- Gumbel Distribution
- The Gumbel distribution is a specific type of extreme value distribution that is mathematically 'better behaved' than the Weibull distribution when dealing with logarithmic transformations. By transforming the Weibull data into this standard form, researchers can use well-known statistical properties to derive exact pivotal quantities.
- Type-I Censored Data
- This refers to situations where some observations in a dataset are not fully observed; only part of their lifetime is known. The proposed method specifically handles this type of missing information effectively, allowing for accurate confidence interval construction even when dealing with censored reliability data.
Terminology
Summary
Reliability inference based on parametric distributions is an important problem in electrical and mechanical engineering because existing methods often rely on approximations or bootstrap procedures that may not perform satisfactorily when data are censored or sample sizes are small. p. 0 Ananda
Professor, Department of Mathematical Sciences, UNLV, Las Vegas, USA
April 15, 2026
Abstract
Reliability inference based on parametric distributions is an important problem
in electrical and mechanical engineering. Most existing methods rely on approximations or bootstrap procedures, which may not perform satisfactorily when data are
censored or sample sizes are small. Hence, there is an urgent need to develop exact
inference approaches for these situations. This article introduces a new approach for
deriving exact parametric tests and confidence intervals for distributions such as the
lognormal, loglogistic, and Weibull.Where the paper says it
Where the paper says it
Problem Motivation
Reliability inference based on parametric distributions is an important problem in electrical and mechanical engineering applications. When expressed in alternative forms, the problem also arises in clinical research survival models. p. 1 1 Introduction
Reliability inference based on parametric distributions has been, and still is, an important problem in electrical and mechanical engineering applications. When expressed in
alternative forms, the problem also arises in clinical research survival models. When expressed in
alternative forms, the problem also arises in clinical research survival models. Lifetime
distributions such as Weibull (Sarhan and Tolba, 2023; Patil et al. Lifetime
distributions such as Weibull (Sarhan and Tolba, 2023; Patil et al. , 2024; Fonseca et al. , 2024; Fonseca et al. ,
2024; Das et al. ,
2024; Das et al. , 2025; Zou and Wen, 2025; Rajput et al. , 2025; Zou and Wen, 2025; Rajput et al. , 2026), loglogistic (Felipe et al. , 2026; Das et al. ,
2025) are widely used in modeling reliability problems. ,
2025) are widely used in modeling reliability problems. Despite such immense importance, beyond the point estimation, currently available
confidence intervals tend to be approximate solutions or based on bootstrapping methods. Despite such immense importance, beyond the point estimation, currently available
confidence intervals tend to be approximate solutions or based on bootstrapping methods. When data is censored and with small sample size, as usually the case in practical problems,
the problem becomes more challenging. When data is censored and with small sample size, as usually the case in practical problems,
the problem becomes more challenging. Thus, there is a need for exact inference solutions
for such problems. Thus, there is a need for exact inference solutions
for such problems. The only available exact solution by Xiang et al. The only available exact solution by Xiang et al. (2015) suffers from an
unintended glitch, as we will point out in Section 2 and demonstrate in our simulations in
Section 4.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Proposed Methodology Outline
The purpose of this section is to outline an approach that one can take in making inferences on the Reliability parameter with widely used lifetime distributions ranging from Weibull to Gamma by taking the Generalized Pivotal Quantity (GPQ) approach, cf. Weerahandi (1993), and Weerahandi (2004). p. 2 While there are multiple articles dealing with
the Weibull distribution, only Xiang et al. (2015) provide exact inference, in the sense of
Weerahandi (1993, 1995, 2004) i. Weerahandi
(1993), and Weerahandi (2004). To describe the idea of the GPQ briefly, let X1, X2, · · · , Xn
be a random sample from a certain distribution, and x1, x2, · · · , xn be the observed sample. To describe the idea of the GPQ briefly, let X1, X2, · · · , Xn
be a random sample from a certain distribution, and x1, x2, · · · , xn be the observed sample. Then, the random quantity T(X, x, θ)) is said to be a GP Q for a particular component
θ of all parameters θ if it satisfies the following two conditions: (i) at the observed data
T(x, x, θ)) reduces to θ, (ii) the distribution of T is free of unknown parameters. Then, the random quantity T(X, x, θ)) is said to be a GP Q for a particular component
θ of all parameters θ if it satisfies the following two conditions: (i) at the observed data
T(x, x, θ)) reduces to θ, (ii) the distribution of T is free of unknown parameters. The approach we propose for dealing with lifetime distributions consists of the following
steps.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
(1). If possible, obtain the minimum sufficient statistics to summarize the information available from the data of each population. p. 3 The approach we propose for dealing with lifetime distributions consists of the following
steps. If possible, obtain the minimum sufficient statistics to summarize the information
available from the data of each population. If possible, obtain the minimum sufficient statistics to summarize the information
available from the data of each population. If no such statistics are easily available, obtain
widely used statistics from the published literature.Where the paper says it
Where the paper says it
(2). Obtain the GPQs of the parameters of the distribution, followed by the reliability parameter, directly or via a standardized distribution. p. 3 If no such statistics are easily available, obtain
widely used statistics from the published literature. Obtain the GPQs of the parameters of the distribution, followed by the reliability parameter, directly or via a standardized
distribution. Obtain the GPQs of the parameters of the distribution, followed by the reliability parameter, directly or via a standardized
distribution. Dealing with many continuous lifetime distributions, one can transform the
distribution to a standard normal distribution as illustrated by Weerahandi et al.Where the paper says it
Where the paper says it
(3). Construct generalized confidence intervals (GCIs) based on the GPQs, using the quantiles of GPQs. p. 3 Dealing with many continuous lifetime distributions, one can transform the
distribution to a standard normal distribution as illustrated by Weerahandi et al. Construct generalized confidence intervals (GCIs) based on the GPQs, using the quantiles of GPQs.Where the paper says it
Distribution Theory for Weibull Distribution
Assume that the lifetime distribution of an electrical or mechanical component follow a Weibull distribution. Consider the problem of making inferences, especially GCIs, for the reliability parameter, with possibly Type-I censored data. p. 4 2 Distribution Theory for the Weibull Distribution
Assume that the lifetime distribution of an electrical or mechanical component follow a
Weibull distribution. Consider the problem of making inferences, especially GCIs, for
the reliability parameter, with possibly Type-I censored data. Consider the problem of making inferences, especially GCIs, for
the reliability parameter, with possibly Type-I censored data. Although there are a few
summary statistics on which one can make inferences beyond point estimation, Xiang
et al. Although there are a few
summary statistics on which one can make inferences beyond point estimation, Xiang
et al. (2015) did so using some distributional results due to Thoman et al. (2015) did so using some distributional results due to Thoman et al. (1969), which
are appropriate only when one uses MLEs rather than LSEs. (1969), which
are appropriate only when one uses MLEs rather than LSEs. To briefly describe the LSE alone based theory, let X = (X1, X2, · · · , Xn) be a random
sample from a Weibull distribution, W(α, θ) having the shape parameter α and the scale
parameter θ with the probability density function as follows:
f(X; α, θ) = α
θ
x
θ
α−1
exp
−
x
θ
α
, x > 0, (2)
where both parameters are positive. The cdf of the random variable X is
F(X; α, θ) = 1 − exp
−
x
θ
α
. (3)
If m out of n observations are Type-I censored, let us denote the order statistics of the
censored sample as X(1), X(2), .Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Derivation of Generalized Pivotal Quantities
To derive GPQs for any of the above parameters, followed by the Reliability parameter, notice that regressions are actually occurring in the log transformation of X, which is distributed as minimum Gumbel. As pointed out by Weerahandi et al. p. 6 Y = ln(X) ∼ G(ν = ln(θ), σ = 1/α). (4)
As pointed out by Weerahandi et al. (4)
As pointed out by Weerahandi et al. (2024), Gumbel is a better behaved distribution
having the location parameter, ν and the scale parameter σ. (2024), Gumbel is a better behaved distribution
having the location parameter, ν and the scale parameter σ. In particular, unlike the
Weibull distribution of X, Y = ln(X) can be easily standardized as
Z = (Y − ν)/σ ∼ G(0, 1) having the cdf FZ(z) = exp(−exp(−z)), (5)
a distribution free of unknown parameters. In particular, unlike the
Weibull distribution of X, Y = ln(X) can be easily standardized as
Z = (Y − ν)/σ ∼ G(0, 1) having the cdf FZ(z) = exp(−exp(−z)), (5)
a distribution free of unknown parameters. 3 GPQs of parameters
Since the CDF of the Gumbel distribution is FY (y) = 1 − exp(−exp(y − ν)/σ), we can still
estimate the parameters ν = ln(θ) and σ = 1/α by LSE on the linear model y = ν + σw,
where depending on whether data are censored or not, w = ln(−ln(1 − p)), where p = Fˆ
Y
obtained by KM method or a plotting position, a quantity free of unknown parameters. 3 GPQs of parameters
Since the CDF of the Gumbel distribution is FY (y) = 1 − exp(−exp(y − ν)/σ), we can still
estimate the parameters ν = ln(θ) and σ = 1/α by LSE on the linear model y = ν + σw,
where depending on whether data are censored or not, w = ln(−ln(1 − p)), where p = Fˆ
Y
obtained by KM method or a plotting position, a quantity free of unknown parameters. Hence, LSEs of parameters can be obtained by fitting the regression y(i) = ν + σwi
, where
y(i) = log t(i)
is the logarithm of the ordered observation and wi = log{− log(1 − pi)}.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
The GPQ for σ can be derived as Gσ = σˆ/σˆ(Z) (8), because (i) from the first expression of (8) it reduces to σ at the observed sample points, and (ii) from the second expression it is obvious that the distribution of Gσ is free of unknown parameters. p. 7 Now it is clear that
Gσ =
σσˆ
Σb
= ˆσ/σˆ(Z) (8)
is a GPQ for σ, because (i) from the first expression of (8) it reduces to σ at the observed
sample points, and (ii) from the second expression it is obvious that the distribution of Gσ
is free of unknown parameters. A GPQ for ν can be derived from its LSEs provided by (6). because (i) the distribution of Gν involves no unknown parameters, and (ii) its observed
value is ˆν − σ (¯z − w¯σˆ(z)) = ν from the starting equation of the above derivation. When the data are censored, the least-squares estimator is still based on the linear
relation
Y(i) = ν + σwi
, ν = log(θ), σ =
1
α
,
but the plotting quantities wi are no longer obtained from the fixed complete-data plotting
positions.Where the paper says it
Where the paper says it
Simulation Study Results
The coverage probabilities from WLMA and GLA, along with the coverage probabilities from bootstrapping were calculated for all scenarios. According to the simulation study, for all simulation scenarios, the WLMA method is overly conservative and the coverage probability is significantly higher than the nominal level γ = 0. p. 0 We illustrate the
method using the Weibull distribution and expect it to be applicable to other widely
used lifetime distributions such as the loglogistic distribution. Our Simulation study
show that the new approach provides better performance than existing methods when
handling complete data and type-I censored data. Two numerical examples are provided to demonstrate the application of the proposed method. The proposed method
is expected to be widely applicable in reliability engineering and survival analysis,
offering a robust alternative to existing methods, particularly in scenarios involving
censored data and small sample sizes. We illustrate the
method using the Weibull distribution and expect it to be applicable to other widely
used lifetime distributions such as the loglogistic distribution. Our Simulation study
show that the new approach provides better performance than existing methods when
handling complete data and type-I censored data.Where the paper says it
Where the paper says it
Where the paper says it
Application to Real Data
Example 1. This is a Weibull dataset that was given in Lawless (2003) and previously used by researchers for reliability inference (Xiang et al. p. 9 Specifically, the GPQ Xiang et al. (2015) used for θ is given by
G
MLE
θ = ˆθ
θ
ˆθ
! Because lifetime data are often censored
in practice, the proposed approach is designed for such settings. We illustrate the
method using the Weibull distribution and expect it to be applicable to other widely
used lifetime distributions such as the loglogistic distribution. ,
2025) are widely used in modeling reliability problems. Despite such immense importance, beyond the point estimation, currently available
confidence intervals tend to be approximate solutions or based on bootstrapping methods. ,
2025) are widely used in modeling reliability problems. Despite such immense importance, beyond the point estimation, currently available
confidence intervals tend to be approximate solutions or based on bootstrapping methods.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Conclusion
In this article, we considered the problem of making inferences about the reliability function of the Weibull distribution. To overcome a glitch in Xiang et al. p. 1 The only available exact solution by Xiang et al. (2015) suffers from an
unintended glitch, as we will point out in Section 2 and demonstrate in our simulations in
Section 4. We illustrate the
method using the Weibull distribution and expect it to be applicable to other widely
used lifetime distributions such as the loglogistic distribution. Our Simulation study
show that the new approach provides better performance than existing methods when
handling complete data and type-I censored data. (2024), Gumbel is a better behaved distribution
having the location parameter, ν and the scale parameter σ. In particular, unlike the
Weibull distribution of X, Y = ln(X) can be easily standardized as
Z = (Y − ν)/σ ∼ G(0, 1) having the cdf FZ(z) = exp(−exp(−z)), (5)
a distribution free of unknown parameters. (2024), Gumbel is a better behaved distribution
having the location parameter, ν and the scale parameter σ. In particular, unlike the
Weibull distribution of X, Y = ln(X) can be easily standardized as
Z = (Y − ν)/σ ∼ G(0, 1) having the cdf FZ(z) = exp(−exp(−z)), (5)
a distribution free of unknown parameters. Specifically, if
pi = Fˆ
Y
t(i)
denotes the estimated distribution value at the ordered observation t(i)
, then for the uncensored failure observations one defines
wi = log{− log(1 − pˆi)}. Letting yi = log t(i) and F denote the set of failure indices, the censored data least-squares
estimators are
σˆ
∗ =
P
i∈F
(wi − w¯)(y(i) − y¯)
P
i∈F
(wi − w¯)
2
, νˆ
∗ = ¯y − σˆ
∗w. The proposed method
is expected to be widely applicable in reliability engineering and survival analysis,
offering a robust alternative to existing methods, particularly in scenarios involving
censored data and small sample sizes. Keywords: Survival function, Generalized inference, Gamma distribution, Gumbel distribution. (2015) and Jia et al. (2018) provide GPQ (generalized
pivotal quantities) based solutions, but they suffer from an unintended glitch due to using
MLE-based construction for an LSE-based solution. When
X is distributed as a Weibull distribution and the available data is Type-I censored, Xiang
et al. (2015) solved the problem by taking the GPQ (generalized pivotal quantity) approach
suggested by Tsui and Weerahandi (1989). Our Simulation study
show that the new approach provides better performance than existing methods when
handling complete data and type-I censored data. Two numerical examples are provided to demonstrate the application of the proposed method. Our Simulation study
show that the new approach provides better performance than existing methods when
handling complete data and type-I censored data. Two numerical examples are provided to demonstrate the application of the proposed method. Our Simulation study
show that the new approach provides better performance than existing methods when
handling complete data and type-I censored data. Two numerical examples are provided to demonstrate the application of the proposed method. The proposed method
is expected to be widely applicable in reliability engineering and survival analysis,
offering a robust alternative to existing methods, particularly in scenarios involving
censored data and small sample sizes. Keywords: Survival function, Generalized inference, Gamma distribution, Gumbel distribution.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Precision reliability inference for components under Type-I censoring: The improved system can derive exact parametric tests and confidence intervals for distributions like lognormal, loglogistic, and Weibull by using LSE-based Generalized Pivotal Quantities (GPQ), which avoids the
unintended glitch
found in previous MLE-based solutions. -
Robust parameter estimation for mechanical components: The system can estimate the shape and scale parameters of a Weibull distribution even when data are Type-I censored, providing more accurate results than methods relying solely on Maximum Likelihood Estimation (MLE) or standard bootstrapping.
-
Improved stress-strength reliability assessment: The system can calculate the reliability parameter P(X < Y) for two independent Weibull variables with Type-I censored data using the derived GPQ for R, offering better coverage probabilities than existing WLMA methods.
-
Enhanced performance in small sample scenarios: The improved method offers better performance than existing methods when handling complete data and Type-I censored data, specifically showing better coverage probabilities compared to the bootstrapping method which shows
under-coverage.
-
Generalized inference for complex distributions: The framework can be extended to other widely used lifetime distributions such as the Gamma distribution by transforming them to a standard normal distribution, enabling LSE based generalized confidence intervals.
Abstract
Reliability inference based on parametric distributions is an important problem in electrical and mechanical engineering. Most existing methods rely on approximations or bootstrap procedures, which may not perform satisfactorily when data are censored or sample sizes are small. Hence, there is an urgent need to develop exact inference approaches for these situations. This article introduces a new approach for deriving exact parametric tests and confidence intervals for distributions such as the lognormal, loglogistic, and Weibull. We revisit several issues in classical reliability analysis based on the survival function. Because lifetime data are often censored in practice, the proposed approach is designed for such settings. We illustrate the method using the Weibull distribution and expect it to be applicable to other widely used lifetime distributions such as the loglogistic distribution. Our Simulation study show that the new approach provides better performance than existing methods when handling complete data and type-I censored data. Two numerical examples are provided to demonstrate the application of the proposed method. The proposed method is expected to be widely applicable in reliability engineering and survival analysis, offering a robust alternative to existing methods, particularly in scenarios involving censored data and small sample sizes.
Related papers
- Learning Robust Treatment Rules for Censored Data
- Neural Estimation of Summary Statistics for Simulation-Based Inference with Non-Asymptotic Guarantees
- Spectral divide-and-conquer MCMC for long stationary time series
- Improving the Efficiency of Subgroup Analysis in Randomized Controlled Trials with TMLE
- Personalized Federated Vector Autoregression with Personalization Diversity
- Doubly Robust and Efficient Calibration of Prediction Sets for Right-Censored Time-to-Event Outcomes