LIESEL - A Software Framework for Bayesian Semiparametric Distributional Regression

LIESEL是研究贝叶斯半参数分布回归的一个新的软件框架，它特别关注模块化、可扩展性和可靠性。分布回归最近在统计学界受到了相当大的关注，因为它使研究人员能够探索超出平均值的解释变量和响应变量之间的复杂关系，例如对分布特征的影响，如规模、偏度、依赖性等。LIESEL促进了统计软件在这一领域的发展，它弥合了贝叶斯推理的专用实现和通用软件包之间的差距。它既足够灵活，可以作为实现复杂模型扩展的基础，又足够方便，可以对统计模型和统计推断算法进行快速原型化和评估。为此，我们依靠最先进的工具来处理所有计算方面以及文档、社区建设和质量保证。LIESEL是作为一个带有R接口的Python包实现的，并为分布式回归模型的规范和估计提供了一个可访问且一致的API。它基于贝叶斯统计模型的有向无环图表示，这导致了模块化和可理解的代码库。大量的模型组件，包括非线性、空间和随机协变量效应，不同的马尔可夫链蒙特卡罗算法，以及其他有用的工具，例如用于模型可视化和模型检查，都在LIESEL中预实现，可用于模型扩展的开发。通过LIESEL，我们为开发可靠的统计软件、可重复的统计研究和统计方法的进步做出了贡献，为实现模型扩展和新的推理算法提供了一个经过良好测试和可靠的基础。LIESEL的原型已在RTG 2300“欧洲山毛榉森林与针叶树的丰富”的研究项目“位置、规模和形状的半参数回归模型”和子项目“时空分布回归模型”中开发。在拟议的项目中，我们将通过(i)系统地重新评估和改进核心软件组件，以确保软件的长期可持续性和重用，（ii）在与研究伙伴合作进行的几个应用案例中展示LIESEL的潜力，以及（iii）广泛投资于传播和推广活动，使LIESEL引起应用研究人员的注意，从而实现我们软件框架的全部潜力。为了实现我们的目标，我们将利用我们在统计软件开发方面的丰富经验，以及我们将使用该软件的科学家网络。

LIESEL is a new software framework for research on Bayesian semiparametric distributional regression with a particular focus on modularity, extensibility, and reliability. Distributional regression has recently received considerable attention in the statistical community since it enables researchers to explore complex relations between explanatory and response variables beyond the mean, e.g. effects on distributional features such as scale, skewness, dependence, etc. LIESEL facilitates the development of statistical software in this area by bridging the gap between specialized implementations on the one and general purpose software packages for Bayesian inference on the other hand. It is both flexible enough to serve as a basis for the implementation of complex model extensions and convenient enough for quick prototyping and evaluation of statistical models and statistical inference algorithms. For this, we rely on state-of-the-art tools for all computational aspects as well as for documentation, community building, and quality assurance.LIESEL is implemented as a Python package with an R interface and provides an accessible and consistent API for the specification and estimation of distributional regression models. It is based on directed acyclic graph representations of Bayesian statistical models, which leads to a modular and comprehensible codebase. A large number of model components including non-linear, spatial, and random covariate effects, different Markov chain Monte Carlo algorithms, and other helpful tools, e.g. for model visualization and model checking, are pre-implemented in LIESEL and can be used for the development of model extensions. With LIESEL, we are contributing to the development of trustworthy statistical software, reproducible statistical research, and advances in statistical methodology by providing a well-tested and reliable basis for the implementation of model extensions and new inference algorithms.A prototype of LIESEL has been developed in the research project "Semiparametric Regression Models for Location, Scale, and Shape" and subproject "Spatio-Temporal Distributional Regression Modeling" of RTG 2300 "Enrichment of European Beech Forests with Conifers". In the proposed project, we will realize the full potential of our software framework by (i) systematically reevaluating and improving the core software components with the goal of ensuring long-term sustainability and reuse of the software, (ii) demonstrating the potential of LIESEL in several application cases conducted in collaboration with research partners, and (iii) extensively investing in dissemination and outreach activities to bring LIESEL to the attention of applied researchers. To achieve our goals, we are going to build on our considerable experience with the development of statistical software and our network of scientists that will work with the software.