GMDS Biostatistics Competition 2026

The Biostatistics section of the German Association for Medical Informatics, Biometry and Epidemiology (GMDS) is hosting the GMDS Biostatistics Competition, a new format aiming to foster collaboration between biostatisticians and data scientists from academia and industry. In this year, the competition is about "Reliable Subgroup Identification and Analysis", a very relevant topic proposed by colleagues at Merck Healthcare KGaA. Besides strengthening the collaboration of current and emergent experts in the field, the competition aims to facilitate scientific advances in a constructive and competitive environment.  

A central methodological challenge is identifying which patients will benefit most from a given regimen. Whereas current methods often rely on single biomarkers from simple subgroup analyses, machine learning (ML) approaches may be able to integrate multiple biomarkers and learn more nuanced decision rules to better match patients to treatments. Several novel methods for subgroup analysis in clinical trials have been developed in recent years to leverage ML [1]. However, these methods often rely on black-box algorithms for patient selection, which provide limited practical guidance and lack explainability. Highlighting the continued interest in the field, a workflow for assessing treatment-effect heterogeneity was recently proposed by Sechidis et al. [2]. 

For the Biostatistics Competition 30 simulated datasets akin to clinical trial data should be analysed by participants. Each dataset represents a randomised phase II or phase III trial with a binary or normally distributed endpoint. True patient assignment to a subgroup or its complement is defined using a subset of the available covariates. The data‑generating mechanism and the true subgroup assignment will not be disclosed to participants to ensure a blinded analysis that reflects real‑world practice. Whilst similar challenges have been published [3,4, 5], our aim is to tap into the diverse backgrounds and expertise of participants that take part in this competition. 

One or more of the following concrete tasks may be addressed by the participants: 

• Task 1: "Assessing treatment effect heterogeneity"  
  Indicating whether a subgroup and a complement exist, whose treatment effects exhibit heterogeneity beyond what can be attributed to randomness. 
• Task 2: "Performing variable selection" 
  Identifying prognostic and predictive baseline variables.
• Task 3: "Assigning individuals" 
  Assigning patients to the subgroup or complement, with a clear definition of the decision criteria used. 
• Task 4: "Estimating subgroup proportion" 
  Defining the proportion of the total population that may be assigned to the subgroup with enhanced benefit. 
• Task 5: "Estimating treatment effect" 
  Estimating the treatment effect in the identified subgroup and complement.

References 

[1] Lipkovich I, et al. Statistics in Medicine. 2024; 43(22): 4388-4436. doi: 10.1002/sim.10167 

[2] Sechidis K, et al. Pharmaceutical Statistics. 2025; 24: e2463. doi: 10.1002/pst.2463 

[3] Bornkamp B, et al. Pharmaceutical Statistics. 2024; 23(4): 495-510. doi: 10.1002/pst.2368 

[4] Ruberg SJ. Pharmaceutical Statistics. 2021; 20(5): 939-944. doi: 10.1002/pst.2110 

[5] Ruberg SJ, et al. Biometrical Journal. 2023; 66(1): 2200164. doi: 10.1002/bimj.202200164

Registered participants will receive four files: 

1. A sheet with additional guidance and explanations beyond information provided on this web page. 
   
   
2. A sheet to indicate which of the tasks their approach(es) is (are) designed to address. 
   
   
3. A zip-file with 30 data sets in csv format:
   
   Figure: Example of a provided dataset where Y is the outcome, x1, ..., xn are baseline covariates, and W is the treatment group.
   
   
4. A results file that allows to collect dataset-level information, including subgroup existence, predictive and prognostic biomarkers, subgroup size, treatment effect estimates, and other relevant metrics: 
   
   Figure: Example of a returned results file where each row represents a dataset ID, along with corresponding results depending on the selected task. 

Participants will be asked to submit the following documents / files: 

1. Sheet indicating which of the following tasks their approach(es) is (are) designed to address. 
2. If subgroup/complement assigning was addressed, please return a ZIP file containing the 30 datasets with patient-level subgroup assignments added as an extra column, using a variable 'S' where 1 indicates subgroup and 0 indicates complement. 
3. Results file that allows for the collection of dataset-level information if one of the other tasks is selected. 
4. Provide a short summary of the methods/algorithms applied for each task that was addressed. If multiple approaches were used for a task, indicate how the final decision was derived. Additionally, if the approach includes hyperparameters, please include those as well, specifying whether they were fixed or determined via techniques such as cross-validation. For each task, please limit your submission to no more than 2 pages in Word.

Those receiving positive feedback on their submission, i.e. being pre-selected for presentation at GMDS, are asked to submit the corresponding code (which will be made publicly available after the conference) or share a GitHub link, along with their poster/presentation before the conference. 

Please note that, akin to real-world drug development practices, even if multiple algorithms used yield inconsistent results, participants are expected to reach a final decision for each task on each dataset. 

The primary evaluation outcome will be an overall rating that takes all selected tasks into account. Additionally, there will be task-specific ratings. These ratings will be derived by comparing participants' outputs with those from the true data-generating model.   

• Task 1: "Assessing treatment effect heterogeneity"  
  Agreement rate over all data sets 
• Task 2: "Performing variable selection" 
  Average score with score per data set based on proportion of correctly classified variables into neither/predictive/prognostic/both plus a bonus of 0.25 if the exact correct set of predictive variables is selected. Note, declaring predictive as both or vice versa is considered as correct decision for the bonus. 
• Task 3: "Assigning individuals" 
  Average over all agreement rates per data set 
• Task 4: "Estimating subgroup proportion" 
  Average difference between estimated proportion and true rate per data set 
• Task 5: "Estimating treatment effect" 
  Average root mean square error (RMSE)

To be included in the rating for each task, at least 80% of the provided data sets must have been analysed. 

It is anticipated that the results of the Biostatistics Competition will be presented and discussed at the ISCB/GMDS conference in Freiburg (ISCB GMDS 2026 Conference). The team(s) with the best solutions will be invited to present their approach. One representative of the winning team will receive a free conference ticket.  It will also be possible to remotely participate in the final workshop. 

In addition, the most interesting approaches will be invited to be published in a joint scientific publication.  

Q: I missed the announcement event on February 18. Can I still get access to a recording of the announcement event?
A: Yes, please contact competition@gmds.de for that matter.

Q: What happens after I register for the competition?
A: You will receive an automatic email confirmation immediately after registration. The data and additional information needed to actively participate is sent out usually within 1-2 business days.

Q: Can I register as a member of multiple groups?
A: No, you can't be a member of multiple groups. However, multiple independent teams can participate from the same department. 

Q: Can team members still be changed after the initial registration?
A: The final list of team members can deviate from the initial list and needs to be included in the submission of solutions. At least the team leader (the person who performed the initial registration) needs to stay the same.

Q: Are there intermediate evaluations or a leaderboard?
A: No, we expect only one submission per team. You may update your submission before the deadline (June 30). Only the last submitted solution per team will be evaluated.

Q: Will the code (for data generation) be made public after the competition?
A: Yes, all relevant code for data generation will be made publicly available after the competition has ended. Winning teams and those participating in the joint publication effort are also expected to share relevant code used to tackle the competition problems.  

We are looking for voluntary members to join our evaluation committee at the end of the competition (July and August 2026). If you are interested, please contact competition@gmds.de. 

• Max Westphal (GMDS)
• Anika Großhennig (GMDS)
• Heiko Götte (Merck)
• Aslihan Gerhold-Ay (Merck)
• Clara Eléonore Pavillet (Merck)

If you have questions or suggestions, feel free to contact us at competition@gmds.de.