📊 statistics

The Complementary Bell-Kumaraswamy-G Family of Distributions: A Flexible Class for Modeling Lifetime and Reliability Data

This paper introduces the novel Complementary Bell-Kumaraswamy-G (CBell-Kw-G) family of distributions, a flexible statistical framework for modeling complex lifetime and reliability data that is rigorously analyzed through theoretical derivation, maximum likelihood estimation, and validation on real cancer survival datasets.

Original authors: Seyed Jamal Khorashadizadeh, Fatemeh Yousefzadeh

Published 2026-08-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Seyed Jamal Khorashadizadeh, Fatemeh Yousefzadeh

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of data, where scientists try to understand how long things last before they fail, the standard tools often fall short. Imagine trying to describe the lifespan of a machine, a patient, or a lightbulb using a single, rigid shape. Real life is rarely so simple. Sometimes, things fail quickly and then the risk drops; other times, the danger starts low, climbs to a peak, and then fades away. These complex patterns, known as hazard rates, are the heartbeat of reliability engineering and medical survival analysis. For decades, statisticians have tried to build flexible mathematical models that can bend and twist to fit these messy, real-world curves. The goal is always the same: to find a formula that captures the true story of how long something survives, whether that is a cancer patient in remission or a daily count of deaths during a pandemic.

A team of researchers from the University of Birjand in Iran has taken a significant step forward in this quest by introducing a new, highly flexible family of statistical models. They call it the Complementary Bell-Kumaraswamy-G family. To understand what they did, one must first appreciate the building blocks they combined. They started with a well-known method called the Kumaraswamy generator, which is like a versatile mold that can stretch or squeeze a basic distribution to fit skewed data. To this, they added a mechanism rooted in the Complementary Bell distribution, a structure that helps account for situations where the specific cause of failure might be hidden or where multiple risks compete against one another. By merging these two approaches, the researchers created a new mathematical framework that is not just a slight adjustment of existing models, but a robust tool capable of producing a wide variety of shapes. It can mimic increasing risks, decreasing risks, or the classic bathtub curve where danger is high at the start, low in the middle, and high again at the end.

The researchers did not stop at theory; they built a comprehensive mathematical map of this new family. They derived formulas to calculate the average lifespan, the spread of data, and the amount of uncertainty or "disorder" within the system, a concept known as entropy. They also developed methods to estimate the model's parameters using real data, specifically testing two different approaches: a standard method and a "penalized" version that adds a small mathematical guardrail to prevent the model from overreacting to small, noisy datasets. To see if their new tool actually worked, they ran thousands of computer simulations. These tests showed that while the standard method sometimes struggled with small amounts of data, producing shaky results, the penalized version remained steady and accurate. This suggests that for smaller studies, the new, guarded approach is the safer choice for getting reliable answers.

To prove the model's worth in the real world, the team applied it to two very different, high-stakes datasets. The first consisted of 89 daily records of confirmed COVID-19 deaths, a dataset known for its volatility and unpredictable spikes. The second involved the remission times of 128 bladder cancer patients, tracking how long they stayed free of the disease after treatment. In both cases, the researchers compared their new model against several established competitors, including the standard Weibull distribution and other modern variations. The results were telling. While the new model fit the data very well, capturing the complex patterns of survival and failure, it did not always win the prize for the absolute best fit. In fact, for both the pandemic data and the cancer data, a slightly simpler version of their own model, which lacked one layer of complexity, actually provided a marginally better match to the numbers.

This finding is a crucial part of the story. It suggests that while the new, highly complex model is incredibly flexible and capable of describing almost any pattern, it does not always need that extra complexity to do the job. The simpler version often achieved the same level of accuracy with fewer moving parts. This is a vital lesson in statistical modeling: more parameters do not always mean a better result. The researchers also used their model to look at risk and inequality. They calculated how much uncertainty existed in the survival times and measured the risk of extreme events, such as a patient surviving much longer than expected or a death toll spiking unexpectedly. They found that their model could estimate these extreme risks reliably, avoiding the mathematical errors that caused other models to produce impossible results, such as negative variance.

Ultimately, the work confirms that the new Complementary Bell-Kumaraswamy-G family is a powerful addition to the statistician's toolkit. It offers a way to model the messy, non-linear reality of survival and reliability with high precision. The study demonstrates that by combining different mathematical generators, one can create a distribution that is both analytically sound and practically useful. However, the research also serves as a reminder that flexibility has a cost. The most complex model is not always the best one; sometimes, the simpler structure is sufficient to tell the truth about the data. For scientists analyzing everything from disease outbreaks to mechanical failures, this new family provides a robust option, but with the wisdom to choose the right level of complexity for the specific problem at hand.

Technical Summary: The Complementary Bell-Kumaraswamy-G Family of Distributions

Problem Statement
Statistical modeling in reliability engineering, survival analysis, and actuarial science frequently encounters datasets characterized by complex behaviors, including heavy tails, varying degrees of skewness, and non-monotonic hazard rates (e.g., bathtub or unimodal shapes). Classical probability distributions often lack the structural flexibility required to capture these nuances. While numerous generalized families (e.g., Exponentiated-G, Marshall-Olkin, Kumaraswamy-G) have been proposed to address these limitations, there remains a need for models that can simultaneously offer analytical tractability and the capacity to model diverse hazard rate shapes through a robust compounding mechanism.

Methodology
The authors propose a novel class of distributions termed the Complementary Bell-Kumaraswamy-G (CBell-Kw-G) family. This family is constructed by compounding the Complementary Bell generator with the Kumaraswamy-G (Kw-G) mechanism.

Model Construction:
- Let $G(x; \theta)$ be a baseline cumulative distribution function (CDF) with probability density function (PDF) $g(x; \theta)$ .
- The Kw-G CDF is defined as $G_{KwG}(x) = 1 - [1 - G(x; \theta)^a]^b$ , where $a, b > 0$ are shape parameters.
- The Complementary Bell-G (CBell-G) generator introduces a parameter $\lambda > 0$ and is defined as $F(x) = \frac{\exp(e^{\lambda G(x)} - 1) - 1}{\exp(e^\lambda - 1) - 1}$ .
- The proposed CBell-Kw-G family is formed by substituting the Kw-G CDF into the CBell-G generator, resulting in a four-parameter family (plus baseline parameters) with CDF:
  $F(x; \lambda, a, b, \theta) = \frac{\exp(e^{\lambda G_{KwG}(x)} - 1) - 1}{\exp(e^\lambda - 1) - 1}$
Theoretical Development:
- Linear Representation: The authors derive an infinite linear representation of the PDF and CDF in terms of exponentiated-G densities. This allows complex properties of the new family to be expressed as weighted sums of simpler baseline properties.
- Statistical Properties: Explicit forms are derived for moments, incomplete moments, the moment generating function (MGF), and quantile functions.
- Entropy Measures: The paper provides closed-form expressions for Shannon entropy and Rényi entropy, utilizing the linear representation to handle the complex density structure.
- Inequality Curves: Theoretical derivations for Lorenz, Bonferroni, and Zenga curves are established to analyze inequality and concentration within the distribution.
- Estimation: Maximum Likelihood Estimation (MLE) is employed for parameter inference. The score functions for all parameters are derived, and numerical optimization (Newton-Raphson, BFGS) is recommended for solving the non-linear likelihood equations.
Special Case:
- The Complementary Bell-Kumaraswamy-Weibull (CBell-Kw-W) distribution is developed as a specific sub-model by selecting the Weibull distribution as the baseline. This model is analyzed in depth for its asymptotic behavior, hazard rate shapes, and specific entropy formulas.
Simulation Study:
- A Monte Carlo simulation study ( $N=10,000$ replications) was conducted to evaluate the performance of MLEs and Penalized Maximum Likelihood Estimators (PMLEs) across various sample sizes ( $n=30, 50, 100, 200$ ) and parameter configurations.
- Performance metrics included empirical bias and Mean Squared Error (MSE).
Empirical Application:
- The model was applied to two real-world datasets: daily confirmed COVID-19 death cases ( $n=89$ ) and bladder cancer remission times ( $n=128$ ).
- The CBell-Kw-W model was compared against five competing models: CBell-Weibull (CBell-WD), Alpha-Power Weibull (APWD), standard Weibull (WD), Exponentiated Weibull (ExpWD), and CBell-Exponentiated Weibull (CBell-EXWD).
- Goodness-of-fit was assessed using Anderson-Darling (AD), Cramér-von Mises (CVM), and Kolmogorov-Smirnov (KS) tests, alongside information criteria (AIC, CAIC, BIC, HQIC).
- Additional analyses included entropy calculations, risk measures (VaR, TVaR, TV, TVP), and inequality curve plotting.

Key Results

Theoretical Flexibility: The CBell-Kw-G family admits a linear representation that simplifies the derivation of moments and generating functions. The hazard rate function of the CBell-Kw-W sub-model is shown to accommodate increasing, decreasing, bathtub, and unimodal shapes.
Estimation Performance: The simulation study indicates that while MLEs are consistent (bias and MSE decrease as $n$ increases), they can exhibit instability and high bias in small samples ( $n < 100$ ). The PMLE method demonstrated superior stability, lower bias, and reduced MSE, particularly for shape parameters in small-sample regimes.
Empirical Fit:
- For both the COVID-19 and bladder cancer datasets, the CBell-Kw-W model provided an acceptable and statistically adequate fit, with non-significant p-values across all goodness-of-fit tests.
- However, the CBell-WD (a simpler, three-parameter model) consistently yielded slightly better goodness-of-fit statistics and lower information criterion values (AIC, BIC) than the five-parameter CBell-Kw-W model in both applications. This indicates that while the additional parameters of the CBell-Kw-W enhance theoretical flexibility, the simpler CBell-WD was sufficient to capture the main structural features of these specific datasets, and the proposed model did not strictly outperform the simpler alternative in these empirical tests.
- Standard models like the Weibull (WD) and Alpha-Power Weibull (APWD) failed to capture the data patterns, evidenced by poor fit statistics and, in the case of APWD, numerical anomalies (negative tail variance).
Risk and Entropy Analysis: The CBell-Kw-W model produced stable, physically meaningful risk estimates (VaR, TVaR) and effectively captured the heterogeneity of the data, as evidenced by the Lorenz and Zenga curves. The entropy analysis confirmed that the COVID-19 dataset exhibited higher uncertainty and dispersion compared to the bladder cancer data.

Significance and Claims
The paper claims that the CBell-Kw-G family offers a robust and flexible alternative for statistical modeling in reliability and survival analysis. Its primary significance lies in:

Structural Versatility: The ability to model diverse hazard rate shapes and heavy-tailed data through the compounding of the Complementary Bell and Kumaraswamy mechanisms.
Analytical Tractability: The provision of a linear representation that facilitates the computation of complex statistical properties (moments, entropy, order statistics) without resorting to purely numerical integration.
Practical Utility: The model's capacity to provide competitive fits to real-world survival data, often capturing tail behavior and heterogeneity effectively, though the specific CBell-Kw-W sub-model did not surpass the simpler CBell-WD in the empirical applications tested.
Estimation Stability: The demonstration that penalized estimation methods (PMLE) can effectively mitigate the convergence issues and bias often associated with high-parameter generalized distributions in small samples.

The authors conclude that while the specific CBell-Kw-W model did not strictly outperform the simpler CBell-WD in these specific empirical applications, the broader CBell-Kw-G family provides a valuable framework for future modeling of complex lifetime data, with potential extensions into bivariate/multivariate settings and regression formulations.

More like this