ISCB 2021

Du 18/07/2021 au 22/07/2021

Domaine de Rockefeller - 8 Avenue Rockefeller, 69008 Lyon, FRANCE


STRATOS Mini-Symposium

Organizer: Willi Sauerbrei (Freiburg, Germany) and Els Goetghebeur (Ghent, Belgium)

Providing accessible and evidence-based guidance for key topics in the design and analysis of observational studies is the overarching goal of the STRATOS (STRengthening Analytical Thinking for Observational Studies) initiative. Guidance is intended for applied statisticians and other data analysts with varying levels of statistical education, experience and interests. In this mini-symposium we will have six talks illustrating the broad range of topics handled in STRATOS and the importance to cooperate with researchers from other areas to develop interdisciplinary approaches to cope with complex analytical challenges.

Talks are planned for a duration of 22 minutes, followed by a discussion of 3 minutes.
Talks with two speakers will be 5 minutes longer.

1. Session 1:00 PM – 2:30 PM, Central European Summer Time

Understanding and accounting for measurement error when prediction equations are used in observational studies
Paul Gustafson; University of British Columbia, Canada for TG4

When an important variable in an observational study is hard to measure, an appealing strategy is to predict its values from other variables. In essence, this induces Berkson measurement error, the implications of which may not be widely understood. We discuss three scenarios. In the first, the marginal distribution of the variable being predicted is of inferential interest. In the second and third, this variable is respectively an exposure variable and an outcome variable, with the exposure-outcome association being of interest. We show that both the implications of the measurement error, and appropriate mitigation strategies, can vary across these scenarios. As part of this discussion, we necessarily grapple with an assumption of non-differential measurement error. We consider both the general plausibility of this assumption, and how it facilitates statistical adjustment for the measurement error. The ideas presented will be illustrated via an example from nutritional epidemiology, with data from the Hispanic Community Health Study / Study of Latinos (HCHS/SOL).

Guidance for performance assessment in prediction models for survival outcomes
David J McLernon1, Terry Therneau2, Daniele Giardiello3,4, Ben Van Calster4,5, Laure Wynants6, Maarten van Smeden7, Ewout W Steyerberg4, on behalf of STRATOS TG6 and TG8

Speakers: David J McLernon and Terry Therneau

1Institute of Applied Health Sciences, University of Aberdeen, Aberdeen, UK; 2Division of Biomedical Statistics and Informatics, Mayo Clinic, Rochester MN, USA; 3Netherlands Cancer Institute, Amsterdam, the Netherlands; 4Department of Biomedical Data Sciences, Leiden University Medical Center, Leiden, the Netherlands; 5Department of Development and Regeneration, KU Leuven; 6School for Public Health and Primary Care, Maastricht University; 7Julius Center for Health Sciences and Primary Care, University Medical Center Utrecht, Utrecht, the Netherlands.

Summary: Risk prediction models need careful validation to understand their performance. Validation of models for survival outcomes poses challenges due to the censoring of observations and the varying time horizon at which predictions can be made. We aim to give a hands-on description of methods to evaluate predictions and decisions from survival models based on Cox proportional hazards regression. We discuss statistical measures, that evaluate model performance in terms of discrimination, calibration, and overall performance, and decision analytic measures. As a motivating case study, we consider the prediction of event free survival in breast cancer patients following surgery.

Guiding the path from Patient Reported Outcomes to treatment registration based on randomised and single arm studies: STRATOS engaged in the European IMI-SISAQOL project. (https://qol.eortc.org/projectqol/sisaqol/)
Saskia le Cessie (Leiden, the Netherlands; Ghent, Belgium) and Els Goetghebeur (Ghent, Belgium)

Whether cancer presents itself as a late stage terminal disease or in almost chronic form, its impact on quality of life and other patient reported outcomes is typically overwhelming. This justifies a far greater role for patient reported outcome measures (PROMs) besides survival in primary and secondary parameters targeted for inference in cancer clinical trials. In addition, recent years saw many new cancer treatments entering the market based on evidence drawn from non-randomized – often single arm - studies. The challenge of reaching evidence on treatment effects without concurrent randomized control is daunting. In IMI-SISAQOL statisticians and clinicians from academia and industry team up with regulators and patient advocates, to reveal current practice in published and unpublished supporting material. We critically examine data, methods and results uncovering opportunities and threats to reliable, relevant and actionable evidence. Our aim is to give guidance on (un)acceptable methodological approaches in specific contexts. We discuss estimands that are or should be targeted with due attention to reference values, minimal acceptable difference and sensitivity analysis starting from available background when concurrent controls are missing while repeated measures inadvertently vary in timing.

2. Session 3:30 PM – 5:00 PM, Central European Summer Time)

Statistical analysis of high-dimensional biomedical data: Analytical goals, common approaches and challenges
Jörg Rahnenführer, Technische Universität Dortmund, Germany for TG9

In high-dimensional data (HDD) settings, the number of variables associated with each observed individual or experimental unit is very large. In biomedical research, prominent examples are omics data and electronic health records data. Advances in statistical methodology and machine learning methods offer new opportunities for innovative analyses of HDD, but at the same time require a deeper understanding of some fundamental statistical concepts. We discuss key aspects of HDD analysis to provide a gentle introduction both for non-statisticians and for classically trained statisticians with little experience specific to HDD analysis. Main analytical goals are outlined, and situations are identified where traditional statistical methods cannot, or should not, be used in the HDD setting, or where adequate analytic tools are still lacking.

Current education and practical guidance in statistical (non-linear) modeling for researchers with limited statistical background (level-1)
Christine Wallisch (Vienna, Austria) and Geraldine Rauch (Berlin, Germany) on behalf of TG2

In a recent systematic review about issues related to selection of variables and functional forms, we investigated the transferred knowledge of statistical modeling to medical researchers through series of statistical notes and tutorials in medical journals. We found that some areas of statistical modeling were underrepresented and code reproducing the presented results was mostly missing. In particular, Poisson regression, variable selection methods and methods to deal with non-linear relations in multivariable models were rarely found. The latter topic was addressed by us in an educational shiny app, with which one interactively learns how fractional polynomials, b-splines and natural splines handle non-linear relations between an outcome and an independent variable.

In defense of correct use of statistical significance
Michal Abrahamowicz 1, Marie-Eve Beauchamp 1, James Carpenter 2,3 and Victor Kipnis 4 Speakers: Michal Abrahamowicz and Victor Kipnis
1 Department of Epidemiology & Biostatistics, McGill University, Montreal, Canada
2 Department of Medical Statistics, London School of Hygiene & Tropical Medicine, UK
3 MRC clinical trials unit at UCL, Holborn, London, UK
4 Biometry Research group, National Cancer Institute, USA

Recently, Amrhein et al, in a highly cited Comment in Nature (2019), recommended banning statistical significance, based on a pre-specified dichotomization of p-values, as a criterion for assessing the strength of the evidence provided by empirical studies. About 850 researchers, representing a wide range of empirical sciences, including several statisticians, endorsed this position, by signing the Amrhein et al’s comment, which got > 100 citations in the 6 first months after its publication. However, other experienced statisticians raised serious concerns about the potential impact of such a black-and-white recommendation.

To stimulate further discussions about this controversial yet very impactful issue, we present some empirical evidence in ‘defense’ of an accurate use of statistical significance. We start by taking a 2nd look at the very flagship empirical example used by Amrhein et al to illustrate the absurd consequences of incorrect interpretation of the results of significance testing, applied separately to two independent studies and then informally compared. We demonstrate that a proper use of a simple statistical test of the significance of the difference between the results of the two studies eliminates the risk of incorrect inference and illogical conclusions. On the other hand, to illustrate the potentially risky implications of the research paradigm advocated by Amrhein et al, we briefly review some recent empirical studies that were cited to back up their decision to report ‘effects’ regardless of the lack of their statistical significance. Finally, we will clarify an important conceptual and formal difference between (i) p-values as used by Fisher in significance testing versus (ii) the role of an a priori chosen significance level (α) in hypothesis testing as suggested by Neyman and Pearson. Based on this empirical evidence and relevant statistical arguments, we conclude that the statistical research community should focus its efforts on better educating our collaborators to improve the understanding of the pros and cons of statistical significance testing and its accurate use in applications rather than on “political” interventions aimed at banning it from the applied research. We believe that the STRATOS initiative can play an important role on this front.