Abstract
Reliability engineering often relies on classifying distributions into aging categories such as Increasing Failure Rate Average (IFRA) and New Better than Used (NBU). Traditional statistical tests typically focus on detecting deviations from an exponential distribution rather than confirming membership within these classes. This paper introduces a novel approach to testing the null hypothesis that a distribution belongs to the IFRA or NBU class. By leveraging the geometric properties of the cumulative hazard function, specifically the Greatest Star-shaped Minorant, the authors develop a weighted Kolmogorov-Smirnov distance metric to test the null hypothesis of class membership. The tests are shown to be consistent and demonstrate high power in simulations, providing a practical tool for reliability engineering and economic analysis.
Introduction
In fields ranging from reliability engineering to economics, understanding how a system or population ages is critical. Properties like IFRA and NBU allow researchers to establish bounds on failure probabilities and optimize maintenance schedules. However, a significant gap exists in the statistical toolkit: most existing tests are "goodness-of-fit" tests against the exponential distribution, which represents a state of no aging. This paper addresses the need for "membership tests," allowing engineers to formally justify the use of specific reliability models by testing the null hypothesis that a distribution belongs to the IFRA or NBU class.
Methodology
The proposed tests utilize the geometric characterization of the cumulative hazard function, $\Lambda(t)$. A distribution is classified as IFRA if $\Lambda(t)$ is "star-shaped," meaning the average failure rate $\Lambda(t)/t$ is non-decreasing. For the NBU class, $\Lambda(t)$ must be "super-additive," satisfying the condition $\Lambda(t_1 + t_2) \geq \Lambda(t_1) + \Lambda(t_2)$.
The core of the testing procedure involves three main steps:
- Empirical Estimation: Calculating the empirical cumulative hazard function from observed data.
- Geometric Approximation: Finding the "best fit" within the target class. For IFRA, this is the Greatest Star-shaped Minorant (GSM)—the largest star-shaped function that remains below the empirical data.
- Distance Metric: Measuring the gap between the raw empirical data and the GSM using a weighted Kolmogorov-Smirnov (K-S) distance.
A key theoretical finding is that the exponential distribution serves as the "least favorable" case. If the test is valid for the exponential distribution (the boundary of the class), it remains conservative for any other distribution within the class. This property facilitates the calculation of p-values via Monte Carlo simulations.
Results and Discussion
The effectiveness of the tests was evaluated using both synthetic and real-world datasets. Monte Carlo simulations using Weibull distributions with varying shape parameters ($a$) were conducted to assess the power and significance levels of the tests.
The results indicate that the tests are statistically consistent, with accuracy improving as sample size increases. Notably, the choice of weighting significantly impacts the test's power. Using inverse hazard weights ($w^{(1)}$) yielded substantially higher power in identifying non-IFRA distributions compared to other weighted versions. For instance, at a sample size of $n=100$ and $a=0.6$, the weighted test correctly rejected the null hypothesis 99.7% of the time.
The tests were also applied to household expenditure data from the National Sample Survey Organization (NSSO) of India. The analysis strongly rejected the Decreasing Failure Rate Average (DFRA) pattern, highlighting the test's utility in characterizing economic inequality and expenditure distributions.
Conclusion
By shifting the traditional hypothesis testing framework to focus on class membership, this research provides a valuable tool for decision-making in reliability and economics. While the current reliance on Monte Carlo simulations for p-values presents some computational challenges, the extension of these methods to censored data ensures their relevance in practical industrial and medical contexts. Future work will look toward extending these tests to randomly right-censored data to broaden their application in survival analysis.
Summary updated