We were fortunate to welcome Björn Bornkamp of Novartis to the EUGM 2016 presenting work he has developed jointly with Marius Thomas (1) on methods of adjusting treatment effect estimates in subgroup analyses with a focus on early phase trials.
As Bornkamp notes at the outset of the presentation, The question of subgroup analysis is currently considered a hot topic with multiple reviews having been published in literature in the last few months. The EMA released a draft guidance in 2014 with a finalization of the guidance expected this year, and the FDA also has a working group on subgroup analysis.
The challenges of conducting subgroup analyses are well documented. In early phase trials, these challenges are exacerbated due to low sample sizes and limited prior knowledge. As Bornkamp notes , a challenge with standard practice is that when searching for a group with an increased treatment effect, one may find a subgroup by chance or over fitting . In this way, the treatment effect estimate in the subgroup will be upwards biased. Inevitably, this problem increases the more subgroups that are investigated. The presentation explores three methods of deriving adjusted treatment effect estimates including Model Averaging, Resampling and Penalized Multivariate Regression ( LASSO).