Multinomial logistic regression models are popular in multicategory classification analysis, but existing models suffer several intrinsic drawbacks. In particular, the parameters cannot be determined uniquely because of the over-specification. Although additional constraints have been imposed to refine the model, such modifications can be inefficient and complicated. In this paper, we propose a novel and efficient simplex-based multinomial logistic regression technique, seamlessly connecting binomial and multinomial cases under a unified framework. Compared with existing models, our model has fewer parameters, is free of any constraints, and can be solved efficiently using the Fisher scoring algorithm. In addition, the proposed model enjoys several theoretical advantages, including Fisher consistency and sharp comparison inequality. Under mild conditions, we establish the asymptotical normality and convergence for the new model, even when the numbers of categories and covariates increase with the sample size. The proposed framework is illustrated by means of extensive simulations and real applications.

Problem definition: Warranty reserves are funds used to fulfill future warranty obligations for a product. In this paper, we investigate the warranty reserve planning problem faced by a manufacturing firm who manages warranties for multiple products. Academic/practical relevance: It is nontrivial to determine a proper amount of reserves to hold, because warranty expenditures are random in nature and reserving either excess or insufficient cash would incur losses. How can warranty reserve levels be optimized and promptly adjusted is a focal issue, especially for firms selling multiple products. Methodology: Inspired by the general pattern of empirical warranty claims data, we first develop an aggregate warranty cost (AWC) forecasting model for a single product by coupling stochastic product sales and failure processes, which can be used to plan for warranty reserves periodically. The reserve levels are then optimized via a distributionally robust approach, because the exact distribution of AWC is generally unknown. To reduce the losses generated from managing the funds, we further investigate two potential loss-reduction approaches: demand learning and funds pooling. Results: For the demand learning algorithm, we prove that, as the sales period grows, the optimal learning parameter asymptotically converges to a constant in a fairly fast rate; our simulation experiments show that the performance of demand learning is promising and robust under general warranty claim patterns. Moreover, we find that the benefits of funds pooling change over different stages of the warranty life cycle; in particular, the relative pooling benefit in terms of reserve losses is nonincreasing over time. Managerial implications: This study offers guidelines on how manufacturers should adaptively forecast and dynamically plan warranty reserves over the warranty life cycle.

Field data provide important information about product quality and reliability. Many large organizations have developed ambitious reliability databases to trace field failure data of a variety of components on the systems they operate and maintain. Due to the exponential distribution assumption for the component lifetimes, the data in these databases are often aggregated. Specifically, individual lifetimes of the components are not available. Instead, each recorded data point is the cumulative operating time of one component position from system installation to the last component replacement, and the number of replacements in between. In the literature, the gamma distribution and the inverse Gaussian (IG) distribution have been used to fit the aggregate data, while the operating environment of different systems is often assumed the same. In order to capture possible heterogeneities among the systems, this study proposes the gamma random effects model and the IG random effects model. The expectation-maximization algorithm is used for point estimation of the parameters and an algorithm based on the generalized fiducial inference method is proposed for interval estimation. Simulation studies are conducted to assess the performance of the proposed inference methods. A real aggregate dataset is used for illustration.