Confidently choosing a cut-point for a continuously valued biomarker to target patients is challenging, because there are two levels of multiplicity: the multiplicity of efficacy in the marker-positive subgroup and in the marker-negative subgroup at each cut-point, and the further multiplicity of searching through infinitely many cut-points. Currently available methods do not strongly control familywise type I error rate (FWER) across both levels of multiplicity. I will present a method that does. Taking a confidence band approach, our method in fact sets forth four principles that we believe every confident biomarker cut-point selection method should strive to adhere to.

For diseases with continuous outcome such as Type II Diabetes and Alzheimer's Disease, our method provides exact simultaneous confidence intervals for efficacy in the marker-positive and marker-negative subgroups, simultaneously for all possible cut-point values. I will demonstrate an interactive app for it.

With high-throughput drug sensitivity screens we can quickly test compounds on cancer cell lines to determine treatment efficacy. Since molecular characterisation of the cell lines by various omics data sets is frequently available, we can link molecular features to treatment efficacy. The estimation of drug synergy is important when testing multiple compounds, but in vitro cell viability measurements can be imprecise due to measurement errors, especially for drug combination experiments. To address this, we propose a Bayesian hierarchical model that uses our recently developed Bayesian model for synergy estimation with uncertainty quantification. The model accounts for synergy estimation uncertainty and selects promising biomarkers for future analysis using a horseshoe prior. Because of the typically low sample size, there is often not enough signal to decisively escape the global shrinkage in the horseshoe prior. To address this issue, we use a variable selection approach called Signal Adaptive Variable Selector to separate the posterior selection probability of a molecular feature from its conditional effect size, i.e. from the posterior mean of its coefficient conditional on it being selected. We demonstrate the model in an application on a large high-throughput dataset of melanoma cell lines. Joint work with Leiv Rønneberg, Pilar Ayuda-Durán, Sigve Nakken, Eivind Hovig, Robert Hanes, Aram Andersen, Tine Norman Alver, Jorrit Enserink and Paul Kirk.

One of the key challenges of personalized medicine is to identify which patients will respond positively to a given treatment. The area of subgroup identification focuses on this challenge, that is, identifying groups of patients that experience desirable characteristics, such as an enhanced treatment effect. A crucial first step towards the subgroup identification is to identify the baseline variables (eg, biomarkers) that influence the treatment effect, which are known as predictive variables. Many subgroup discovery algorithms return importance scores that capture the variables' predictive strength. However, a major limitation of these scores is that they do not answer the core question: “Which variables are actually predictive?” With our work we answer this question by using the knockoff framework, which is a general framework for controlling the false discovery rate when performing prognostic variable selection. In contrast, our work is the first that uses knockoffs for predictive variable selection. We introduce two novel knockoff filters: one parametric, building on variable importance scores derived from a penalized linear regression model, and one non-parametric, building on causal forest variable importance scores. We conduct extensive simulations to validate performance of the proposed methodology and we also apply the proposed methods to data from a randomized clinical trial. This talk is based on our recent paper: https://onlinelibrary.wiley.com/doi/full/10.1002/sim.9134

In the era of precision medicine, understanding treatment effect in biomarker defined subgroups in relationship with overall population is essential. For continuous outcomes, Least Square estimates include an interaction term to enable an unbiased estimation of treatment effect in the overall population (overall treatment effect) by linearly combining treatment effects of the two complementary subgroups. Such logic is carried to binary and time-to-event outcomes models in most statistical software where model parameters are linearly combined in the log scale and then exponentiated to represent overall treatment effect. Although guaranteeing logical inference in appearance, such calculations do not correspond to the true overall treatment effect which may in fact be illogical for efficacy measures such as odds ratio and hazard ratio, i.e., the overall treatment effect is outside the range of subgroups effects. To correctly derive efficacy in the overall population, a principle called Subgroup Mixable Estimation (SME) should be followed. We illustrate these common mistakes and demonstrate the application of SME using real trial data.

For many disorders a musculoskeletal problem is a potential explanation and manual therapy may be a promising treatment option. However, it is often not possible to identify the musculoskeletal problem and for many patients it remains only one of several potential explanations. It has to be expected that manual therapy is only beneficial for those patients with a musculoskeletal problem, and hence it is of interest to identify patient characteristics which may be associated with the presence of a musculoskeletal problem and hence may be predictive for the treatment benefit from manual therapy. This allows to use domain knowledge and domain expertise to identify promising treatment effect modifiers even prior to the first RCT on manual therapy for a specific disorder. As domain knowledge and domain expertise is used, it is also possible to use hypotheses about the direction of the modification to specify a priori a summary index to predict the treatment effect.
This process of identifying promising effect modifiers is exemplified using two RCTs on the effect of manual therapy on disorders potentially caused by musculoskeletal problems: colic in infants (Holm et al 2021) and recurrent headache in children (Lynge et al 2021). Both RCTs are analysed using the same strategy, combing a confirmatory part with an exploratory part. The strategy of identifying the effect modifiers is related to the success of the analysis to find indications for effect modification.
A third example focus on the identification of interactions between the treatment setting (chiropractice vs general practice) and the clinical outcome based on observational data (Hartvigsen et al 2020). In this example, interactions are not interpreted as treatment effect modifications, but as differences in the prognostic value of patient characteristics between different settings. Challenges in an interpretation of this type are discussed.
References:
Hartvigsen L, Kongsted A, Vach W, Salmi LR, Hestbaek L (2020). Baseline Characteristics May Help Indicate the Best Choice of Health Care Provider for Back Pain Patients in Primary Care: Results From a Prospective Cohort Study. J Manipulative Physiol Ther. 43(1):13-23. doi: 10.1016/j.jmpt.2019.11.001

Holm LV, Vach W, Jarbøl DE, Christensen HW, Søndergaard J, Hestbæk L (2021). Identifying potential treatment effect modifiers of the effectiveness of chiropractic care to infants with colic through prespecified secondary analyses of a randomised controlled trial. Chiropr Man Therap. 29(1):16. doi: 10.1186/s12998-021-00373-6.

Lynge S, Dissing KB, Vach W, Christensen HW, Hestbaek L (2021). Effectiveness of chiropractic manipulation versus sham manipulation for recurrent headaches in children aged 7-14 years - a randomised clinical trial. Chiropr Man Therap 29(1):1. doi: 10.1186/s12998-020-00360-3.