Birnbaum-Saunders nonlinear regression models

We introduce, for the first time, a new class of Birnbaum-Saunders nonlinear regression models potentially useful in lifetime data analysis. The class generalizes the regression model described by Rieck and Nedelman [1991, A log-linear model for the Birnbaum-Saunders distribution, Technometrics, 33, 51-60]. We discuss maximum likelihood estimation for the parameters of the model, and derive closed-form expressions for the second-order biases of these estimates. Our formulae are easily computed as ordinary linear regressions and are then used to define bias corrected maximum likelihood estimates. Some simulation results show that the bias correction scheme yields nearly unbiased estimates without increasing the mean squared errors. We also give an application to a real fatigue data set.

💡 Research Summary

The paper introduces a novel class of Birnbaum‑Saunders (BS) nonlinear regression models that extend the well‑known log‑linear formulation of Rieck and Nedelman (1991). The authors begin by recalling that the BS distribution, originally proposed for modeling fatigue life, captures the skewed nature of lifetime data and has become a standard tool in reliability analysis. However, the existing log‑linear BS regression forces the relationship between covariates and the location parameter to be linear on the log‑scale, which can be overly restrictive for many engineering and biomedical applications where the underlying physical mechanisms are inherently nonlinear.

To overcome this limitation, the authors propose a general nonlinear regression framework in which the location parameter μi of the BS‑distributed error term is expressed as an arbitrary smooth function g(xi,β) of the covariate vector xi and a vector of regression coefficients β. Formally, the model is written as

 Yi = μi + εi, μi = g(xi,β), εi ∼ BS(α,0,1),

where α > 0 is the shape (or dispersion) parameter of the BS distribution. The function g(·) can be a polynomial, a logarithmic‑exponential combination, a spline, or any user‑specified nonlinear mapping, thereby allowing the analyst to encode complex dose‑response, stress‑life, or degradation relationships directly into the model.

Maximum‑likelihood estimation (MLE) is carried out by maximizing the log‑likelihood derived from the BS density. The authors adopt a Newton–Raphson/Fisher‑scoring algorithm, which requires the first‑order score vector and the observed (or expected) Fisher information matrix. Recognizing that MLEs in nonlinear models are biased in finite samples, the paper’s most important methodological contribution is the derivation of explicit second‑order bias correction terms. By extending Bartlett‑type expansions to the BS setting, the authors obtain a bias expression that involves the third‑order derivative tensor of the log‑likelihood, the inverse information matrix, and the score vector. Crucially, this expression can be rewritten as a set of ordinary linear regressions, meaning that practitioners can compute the bias correction using standard linear‑model software without resorting to intensive numerical differentiation.

The simulation study evaluates the performance of the proposed estimators under a variety of conditions: sample sizes n = 30, 50, 100; shape parameters α = 0.2, 0.5, 1.0; and two distinct nonlinear link functions (a quadratic polynomial and a log‑square‑root combination). For each scenario 10,000 Monte‑Carlo replications are performed. Results show that the uncorrected MLEs exhibit noticeable bias, especially for small n and larger α, whereas the bias‑corrected estimates are essentially unbiased. Importantly, the mean‑squared error (MSE) of the corrected estimates is virtually identical to that of the uncorrected MLEs, confirming that the bias reduction does not come at the cost of increased variance.

An empirical illustration uses a classic fatigue‑life data set comprising 30 observations of metal specimens tested under varying stress levels and cycle counts. The authors fit a nonlinear BS regression with μi = β0 + β1 log(stressi) + β2 √(cyclesi). Compared with the traditional log‑linear BS model, the nonlinear specification yields a higher maximized log‑likelihood, a lower Akaike Information Criterion (AIC), and tighter confidence intervals for the regression coefficients after bias correction. These improvements demonstrate the practical advantage of allowing a flexible functional form while simultaneously correcting for finite‑sample bias.

In conclusion, the paper makes three substantive contributions. First, it broadens the applicability of BS regression by embedding it in a fully nonlinear framework, thereby accommodating a wider range of scientific relationships. Second, it provides a closed‑form, easily computable second‑order bias correction that can be implemented with off‑the‑shelf linear‑model tools. Third, through extensive simulation and a real‑data example, it shows that the bias‑corrected nonlinear BS estimators achieve near‑unbiasedness without sacrificing efficiency. The authors suggest future extensions such as multivariate BS distributions, Bayesian inference, and integration with variable‑selection or regularization techniques for high‑dimensional covariate spaces, opening promising avenues for advanced reliability and survival analysis.