📊 statistics

Weibull-MBUR {Y} A New Family of Generalized Unit Distributions with Correlated Parameters: Is the Correlation, Distribution or Data Driven or Interaction Between Them

This paper introduces the Weibull-MBUR {Y} family of generalized unit distributions by applying the T-X{Y} framework to link a baseline Median Based Unit Rayleigh distribution with a Weibull generator through various link functions, deriving their statistical properties and analyzing their parameter correlations using real-world COVID-19 recovery data to determine if such correlations are driven by the data, the distribution, or their interaction.

Original authors: iman attia

Published 2026-10-06✓ Author reviewed ⓘ

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: iman attia

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 by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of statistics, scientists often need to describe how things are distributed, from the height of people to the time it takes for a machine to fail. When the data falls between zero and one—like percentages, recovery rates, or proportions—standard tools often struggle to capture the full picture. To solve this, researchers have spent decades building "families" of distributions. Think of these as flexible templates that can be stretched, squeezed, or twisted to fit the unique shape of real-world data. One popular method involves taking a simple, well-understood curve and using a mathematical engine to transform it, adding new knobs and dials that allow the curve to bend in more complex ways. This flexibility is crucial because real life is rarely simple; data often skews to one side or has heavy tails, and a rigid model cannot tell the whole story.

A researcher at Cairo University has recently expanded this toolkit by creating a new family of these flexible curves, specifically designed for data that lives between zero and one. The work focuses on a specific base curve called the Median Based Unit Rayleigh distribution. While this base curve is useful, it is somewhat limited in how it can adapt to different shapes. To overcome this, the author combined it with a powerful generator known as the Weibull distribution, which is famous for its ability to model time-to-failure data. By linking the base curve to the generator through five different mathematical bridges—using distributions named Lomax, Dagum, exponential, exponentiated exponential, and Log-Logistic—the researcher created five new, highly adaptable distributions. The goal was not just to create more curves, but to see if these new models could better explain a specific real-world phenomenon: the recovery rates of patients from COVID-19 in Spain.

The study began with a dataset of 66 observations representing the recovery rates of COVID-19 patients in Spain. These numbers ranged from a low of 0.4286 to a high of 0.8628, with an average recovery rate of 0.7240. The data showed a slight skew, meaning the recovery rates were not perfectly balanced around the average. The researcher first tried to fit this data using older, established models like the Beta and Kumaraswamy distributions, as well as the original, unmodified base curve. The results were telling: the original base curve failed to capture the data's shape, and while the older models performed adequately, they did not offer the best possible fit.

When the five new generalized distributions were applied to the same Spanish recovery data, the results improved significantly. The new models, particularly the one built using the Lomax bridge, provided a much closer match to the actual observations. Statistical measures confirmed that these new curves described the data more accurately than the older alternatives. The researcher calculated the "log-likelihood," a score that measures how well a model explains the data, and found that the new Weibull-MBUR-Lomax model achieved the highest score. Other metrics, which penalize models for being too complicated, also favored the new distributions, suggesting they were not just over-fitting the noise but capturing the true underlying pattern.

However, the study uncovered a fascinating and somewhat puzzling side effect of this increased flexibility. As the new models fit the data better, the mathematical relationship between their internal parameters became extremely tight. In the best-performing model, the parameters were so closely linked that they moved almost in perfect unison. The researcher described this as a very high correlation, where changing one number to improve the fit forced the others to change in a predictable, almost locked-step manner. This created a situation where the statistical confidence intervals for these parameters became very wide, making it difficult to pinpoint the exact value of any single parameter with high precision.

The central question the paper poses is the source of this intense connection between the numbers. Is this high correlation a feature of the data itself, meaning the Spanish recovery rates naturally demand such a tight relationship between these variables? Or is it a feature of the mathematical construction, where the way the new distributions were built simply inherits these strong links from the components used to create them? The author suggests that the latter is a strong possibility, noting that the components used to build these models are known to have their own internal relationships. Yet, the paper stops short of declaring a final verdict. It proposes that the correlation might be a mix of both the data's nature and the mathematical architecture used to model it.

To resolve this, the researcher concludes that more work is needed, specifically through computer simulations. By generating fake data with known properties and testing these new models against it, future studies could determine if the high correlations appear even when the data is perfectly random or follows a simple rule. Until then, the study stands as a demonstration of how adding more mathematical flexibility can dramatically improve the fit to real-world data, even if it introduces new complexities in understanding the individual parts of the model. The new distributions successfully captured the nuances of the COVID-19 recovery data that older models missed, proving their utility, while simultaneously highlighting a deep statistical mystery about how these complex models behave.

Technical Summary: Weibull-MBUR {Y} A New Family of Generalized Unit Distributions with Correlated Parameters

Problem Statement
The paper addresses the need for flexible statistical models capable of accommodating diverse shapes (skewed, bathtub, J-shaped, etc.) for data defined on the unit interval $(0, 1)$ , such as proportions, rates, and percentages. While existing generalizations like the Beta-generated (BG) and Kumaraswamy-generated (Kw-G) families exist, the author identifies a gap in the literature regarding the specific combination of the Median Based Unit Rayleigh (MBUR) baseline distribution with the Weibull generator and various link distributions. Furthermore, the paper investigates a critical, often overlooked issue in generalized distributions: the nature of correlation between estimated parameters. The author questions whether observed high correlations in parameter estimates are driven by the specific dataset, the mathematical structure of the distribution family, or an interaction between the two.

Methodology
The study employs the T-X{Y} family approach, an extension of the T-X method introduced by Aljarrah et al. (2014). This framework constructs a new family of distributions by linking a baseline distribution ( $X$ ), a generator distribution ( $T$ ), and a link distribution ( $Y$ ) via the quantile function of $Y$ .

Components:
- Baseline ( $X$ ): Median Based Unit Rayleigh (MBUR) distribution, defined on $(0, 1)$ with shape parameter $b$ .
- Generator ( $T$ ): Weibull distribution, defined on $(0, \infty)$ with shape ( $c$ ) and scale ( $\xi$ ) parameters. The Weibull was chosen for its closed-form CDF and tractable raw moments.
- Link ( $Y$ ): Five different distributions were utilized as the link (quantile) function to generate five new sub-models:
  - Lomax
  - Dagum
  - Exponential
  - Exponentiated-Exponential
  - Log-Logistic
Derivation Process:
- The Cumulative Distribution Function (CDF) of the new family, $G_{X_g}(x)$ , is derived by substituting the CDF of the baseline MBUR into the quantile function of the link distribution $Y$ , and then substituting that result into the CDF of the Weibull generator.
- The Probability Density Function (PDF), Hazard Function, and Quantile Function are derived analytically for each of the five new distributions.
- Moment Derivation: Two approaches are used to derive the $r$ $r$ -th raw moments:
  - A series linear expansion of the PDF using generalized binomial theorems and Pochhammer falling factorials.
  - An approach isolating the variable $x$ and expressing the moments in terms of the raw moments of the Weibull generator.

Key Contributions

New Distribution Family: The paper introduces the Weibull-MBUR{Y} family, comprising five specific distributions (Lomax, Dagum, Exponential, Exponentiated-Exponential, and Log-Logistic variants). To the author's knowledge, these specific combinations have not been previously discussed in the literature.
Analytical Properties: The paper provides closed-form expressions for the PDF, CDF, quantile, and hazard functions for all five models. It also derives complex series expansions for the raw moments, facilitating computational statistics without requiring special functions in some cases.
Correlation Analysis: A significant contribution is the empirical investigation into the correlation structure of the estimated parameters. The author analyzes the variance-covariance matrices and correlation coefficients for the new models fitted to real data.

Results and Real Data Analysis
The proposed distributions were applied to a dataset of 66 observations representing the recovery rate from COVID-19 patients in Spain. The data exhibited negative skewness (-0.7214).

Model Comparison: The new Weibull-MBUR{Y} distributions were compared against the baseline MBUR, Beta, Kumaraswamy, Topp-Leone, Unit Lindley, and Generalized Odd Median Based Unit Rayleigh (GOMBUR-I and II) distributions.
Performance Metrics:
- The baseline MBUR, Topp-Leone, and Unit Lindley distributions failed to fit the data adequately (low Log-Likelihood, high KS statistics).
- Among the new models, the Weibull-MBUR{Lomax} distribution provided the best fit, evidenced by the highest Log-Likelihood (64.4163) and the lowest AIC, CAIC, BIC, and HQIC values.
- The Weibull-MBUR{Dagum} and Weibull-MBUR{Log-Logistic} also demonstrated strong performance, outperforming the standard Beta and Kumaraswamy distributions in several indices.
Parameter Correlation Findings:
- The Weibull-MBUR{Lomax} model exhibited extremely high positive correlations (near 1.0) between its parameters ( $b, c, \xi, \psi, \lambda$ ), resulting in a variance-covariance matrix determinant near zero ( $5.6073 \times 10^{-13}$ ).
- The Weibull-MBUR{Dagum} model showed perfect correlations between several parameters ( $c, \xi, a, \psi$ ) and a negative correlation between $b$ and $\gamma$ .
- The Weibull-MBUR{Exponential} and Exponentiated-Exponential models also displayed high correlations, particularly between the scale and shape parameters.
- The Weibull-MBUR{Log-Logistic} model showed moderate to high correlations, particularly between $\xi$ and $\psi$ .

Significance and Claims
The paper claims that the generalization of the MBUR distribution using the T-X{Y} technique successfully enhances the model's ability to accommodate the skewness and shape of real-world unit data, which the baseline distribution could not achieve. The addition of parameters from the generator and link distributions allows the PDF to "twist" and fit complex data shapes.

However, the author maintains a modest stance regarding the source of the observed parameter correlations. While the new distributions fit the data well, the high correlations between parameters (leading to wide confidence intervals in some cases) raise the question of whether these correlations are:

Distribution-driven: Inherited from the mathematical construction of the T-X{Y} system.
Data-driven: Specific to the characteristics of the COVID-19 recovery dataset.
An interaction: A mix of both.

The paper concludes that while the new family is effective for modeling, the issue of parameter correlation requires further investigation. The author explicitly states that simulation studies are necessary to definitively determine if these correlations are inherent to the T-X{Y} construction or specific to the data, and to evaluate how they impact confidence intervals. No definitive causal claim is made regarding the origin of the correlations in this study.

More like this