Introduction and objectives

The goal of the High-dimensional Data (HDD) Topic Group of the STRATOS initiative (TG9) is to provide guidance amid the jungle of opportunities and pitfalls inherent in the analysis of high-dimensional biological and medical data. Methods for analysis of HDD are rapidly changing, and researchers across different fields, including biostatistics, bioinformatics, and bioengineering, contribute to their development. Advances in statistical methodology and machine learning methods have contributed to improved approaches for data mining, statistical inference, and prediction in HDD settings; however, adoption of these methods has sometimes gotten ahead of understanding of their proper application.

Methods and results

The mission of TG9 includes identification of fundamental principles for analysis of HDD, explanation of available methods, and development of broadly accessible guidance on best practices in this complex and changing landscape. In this talk we present the first published work of TG9 [1]: a comprehensive review aiming to provide a solid statistical foundation for researchers, including statisticians and non-statisticians, who are new to research with HDD or simply want to better evaluate and understand the results of HDD analyses. Following that, we will describe new research topics currently being pursued by TG9. Specifically, guidance materials specific to HDD for sample size calculation, influence and choice of tuning parameters in machine learning applications, and use of plasmode data for simulations are under development. Simulation studies are especially challenging for HDD yet they are essential tools needed to perform evaluation and comparison of different methods.

Conclusions

Proliferation of high-dimensional data in biomedical research has brought unprecedented opportunities to advance knowledge. In order to harness the power of the rapidly evolving repertoire of analysis methods to reveal useful insights from HDD, it is imperative that researchers have access to guidance on the methods available and their proper application.

References

[1] Rahnenführer, J., De Bin, R., Benner, A. et al. Statistical analysis of high-dimensional biomedical data: a gentle introduction to analytical goals, common approaches and challenges. BMC Med 21, 182 (2023). https://doi.org/10.1186/s12916-023-02858-y

Topic group 2 (TG2) of the STRATOS initiative deals with multivariable model building, in particular with issues in building suitable descriptive regression models. In particular, our TG will provide guidance on strategies for variable selection and on the specification of the functional form of nonlinear effects of continuous covariates.

Medical literature is still full of outdated statistical approaches, because there is a lack of awareness of possible pitfalls of commonly used methods even in very common scenarios. Contrary to that, a large number of new methods is proposed, but most of these methods are stuck in an early phase of development (see Heinze et al, 2023, https://doi.org/10.1002/bimj.202200222). There is insufficient evidence of them being fit for purpose. Often guidance for suitable methods is missing, and scientifically inferior procedures are used to analyse medical data, leading to questionable medical conclusions. These problems affect all STRATOS topics, but here we focus on multivariable modelling.

In our overview paper (Sauerbrei et al, 2020, https://doi.org/10.1186/s41512-020-00074-3) we identified seven areas where evidence should be created by well-designed comparison studies. These areas include: (1) investigating properties of variable selection strategies, (2) comparing spline procedures, (3) analysing variables with a spike at zero, (4) comparing multivariable procedures for variable and function selection, (5) clarifying the role of shrinkage to correct for bias induced by data-driven decisions, (6) evaluating approaches to post-selection inference, and (7) adapting model building strategies to large sample sizes.

In this talk we mention recent activities inside and outside STRATOS to address these issues. Research is ongoing in almost all of these seven areas, but it will need further neutral studies exploring the empirical properties of existing methods in a wider range of problems, and studies that are able to uncover situations where established methods may fail and clarify which assumptions of a method are crucial and which are non-critical. We encourage researchers to perform, reviewers to appreciate, and biostatistical journals to publish such carefully planned method evaluation studies that are indispensable to create the evidence for defining a state-of-the-art in multivariable modelling.

Quantitative bias analysis (QBA) permits assessment of the expected impact of various imperfections of the available data on the results and conclusions of a particular real-world study. This article extends QBA methodology to multivariable time-to-event analyses with right-censored endpoints, possibly including time-varying exposures or covariates. The proposed approach employs data-driven simulations, which preserve important features of the data at hand while offering flexibility in controlling the parameters and assumptions that may affect the results. First, the steps required to perform data-driven simulations are described, and then two examples of real-world time-to-event analyses illustrate their implementation and the insights they may offer. The first example focuses on the omission of an important time-invariant predictor of the outcome in a prognostic study of cancer mortality, and permits separating the expected impact of confounding bias from non-collapsibility. The
second example assesses how imprecise timing of an interval-censored event – ascertained only at sparse times of clinic visits – affects its estimated association with a time-varying drug exposure. The simulation results also provide a basis for comparing the performance of two alternative strategies for imputing the unknown event times in this setting.