Conference Agenda

Rima Izem¹, Paola Rebora², Nicholas Bakewell³, Mitchell Gail⁴, Suzanne Cadarette⁵ for TG5

¹Statistical Methodology, Novartis Pharma AG, Basel, Switzerland
²School and Medicine and Surgery, University of Milano-Bicocca, Italy
³Health Services Research, University of Toronto, Canada
⁴Biostatistics Branch, National Institutes of Health, Rockville, Maryland, USA
⁵Leslie Dan Faculty of Pharmacy, University of Toronto, Canada

The STRATOS Study Design Topic Group (TG5) aims to offer guidance on planning and designing observational studies. Proper planning, informed by subject-matter expertise, ensures that research objectives are clearly defined, clinically relevant, and that the chosen study design is appropriate and valid. Despite its apparent simplicity, flaws in study design are frequently reported, highlighting the need for robust guidance from this subteam.

One TG5 topic of interest includes the review of main challenges in planning clinical trials or observational studies to answer causal questions about the long-term risks and benefits of treatments for chronic conditions. In chronic care, extended exposure to treatments raises questions about long-term safety or effectiveness, necessitating further studies.

The current practice often involves designing studies to compare the initiation of a new treatment with standard care on long-term outcomes. However, the treatment landscape is dynamic. Patients may experience multiple intercurrent events after initiating a treatment, such as dose escalation, switching treatments, therapy gaps, or concurrent treatments, which can influence outcomes. Ignoring these intercurrent events or censoring follow-up at these events allows estimation but muddies the causal inference. Therefore, estimands often focus on quantifying the effect of treatment initiation at the expense of complex exposure patterns.

A potential alternative that TG5 is exploring is to ask cumulative exposure questions at fixed follow-up periods informed by drug utilization patterns in real-world settings.

Carsten. O. Schmidt¹, Marianne Huebner², Lara Lusa³

¹Institute for Community Medicine, University Medicine of Greifswald, Greifswald, Germany
²Department of Statistics and Probability, Michigan State University, East Lansing, MI, USA
³Department of Mathematics, Faculty of Mathematics, Natural Sciences and Information Technology, University of Primorska, Koper, Slovenia

The key principle of Initial Data Analysis (IDA) is to provide reliable knowledge about the data underlying the main statistical analyses (MDA). The STRATOS topic group TG3 “Initial data analysis” aims to improve awareness of IDA as an important part of the research process and to provide guidance on conducting IDA in a systematic and reproducible manner in pursue of transparent and reproducible science. IDA focuses on the workflow from metadata setup, data cleaning, data screening, data quality assessments, reporting prior to conducting the MDA. This talk will provide an overview on these steps and introduces an international effort to develop a statistical analysis plan template in cooperation with all STRATOS topic groups for observational studies that incorporates a systematic IDA plan.

Sabine Hoffmann¹ for the Open Science panel

¹Institute for Medical Information Processing, Biometry, and Epidemiology, Ludwig-Maximilians-Universität Munich, Germany

The scientific community, publishers and funders are increasingly encouraging open science practices with the idea that “scientific knowledge of all kinds, where appropriate, should be openly accessible, transparent, rigorous, reproducible, replicable, accumulative and inclusive” [1]. The STRATOS Open Science panel was funded to promote open science practices by providing guidance on ways to achieve this idea. This talk with give a general overview of the importance of open science practices in the design and analysis of observational studies in biomedical research and then focus on two ongoing projects concerning guidance on data sharing through synthetic data generation and a project that illustrates how to deal with analytical choices (“researcher degrees of freedom”) in the analysis of observational studies.

1. Parsons S, Flavio Azevedo F, Elsherif MM, Guay S, Shahim ON,Govaart GH, Norris E, Aoife O’Mahony A, Parker AJ, Todorovic A, et al. A community-sourced glossary of open scholarship terms. Nature Human Behaviour, 6(3):312–318, 2022

Aris Perperoglou¹, Mohammed Sedki², Anne Thiébaut³, Michal Abrahamowicz⁴, Paul Gustafson⁵, Victor Kipnis⁶, Laurence Freedman⁷ on behalf of the STRATOS TG2 & TG4 collaborative groups

¹ GSK, London, UK
² Université Paris-Saclay, France
³ INSERM National Institute of Health and Medical Research, Villejuif, France
⁴ McGill University, Montreal, Canada
⁵ Department of Statistics, The University of British Columbia, Vancouver, British Columbia, Canada
⁶ Biometry Research Group, Division of Cancer Prevention, National Cancer Institute, Bethesda, Maryland, USA
⁷ Biostatistics Unit, Gertner Institute for Epidemiology and Health Policy Research, Sheba Medical Center, Tel Hashomer, Israel

Covariates in medical research are often measured with error, biasing estimates of exposure-outcome relationships, especially when these relationships are non-linear. This study compares methods for measurement error correction in such non-linear settings.

This blinded, multi-stage simulation project, a collaboration within the STRATOS initiative (Topic Groups 2 and 4), involved a Data Generation and Evaluation team and three Methods teams. These teams applied Bayesian methods, Imputation/Regression Calibration (MI/RC), and Simulation Extrapolation (SIMEX), combined with flexible modelling techniques (B-splines (BS), P-splines (PS), Fractional Polynomials (FP2), and Natural Splines (NS)). Datasets featured a binary outcome, a continuous covariate with classical error (X*), and a replicate substudy. The true non-linear functional form, covariate distribution, error variance, and error distribution were initially withheld. Phase 1 used 5 pilot datasets; Phase 2 expanded to 155 unique datasets by varying sample sizes, measurement error (ME) variance, error distribution (Normal, shifted-Gamma), and true functional forms. Performance was assessed by log Mean Absolute Error (logMAE).

SIMEX methods consistently demonstrated the highest accuracy. P-splines, FPs, and NS generally outperformed BS, especially with SIMEX or Bayesian approaches. Following SIMEX, Bayesian methods (excluding BS) performed best, then RC (excluding BS), and MI. Bayesian BS combinations typically performed poorest, particularly with smaller samples. Accuracy generally improved with larger sample sizes and smaller ME. Linear relationships were estimated most accurately; J-shaped forms were most challenging. A shifted-Gamma ME distribution yielded slightly better accuracy for most methods. Notably, SIMEX was less sensitive to increased ME magnitude and, unlike MI and Bayes, showed no substantial accuracy improvement with larger replication substudy sizes.