BACKGROUND Companion diagnostics (CD) testing aims to stratify the patient population. It conditions the choice of best available therapeutic options, limiting targeted therapy (TT) to subgroups most likely to benefit and triggers potential cost savings. OBJECTIVES To provide a general framework and a list of key methodological points to be addressed while conducting a model-based economic evaluation of CD-TT pairings, especially in oncology. METHODS Based on a health economic literature review and a clinical expert panel with examples drawn from cases of CD testing selection biomarker predictive of the response level towards an anti-cancer TT. RESULTS As CD and TT have embedded values, it is important to assess them concomittantly within a shared modeling framework. We propose a decision tree to model the patient population stratification and to incorporate impacts of analytical and clinical validity. The former refers here to the inability of the CD to accurately and reliably inform the biomarker resulting in true (false) positive/negative cases, whereas the latter relates to the penetrance, i.e. the strength of association between the biomarker and clinical phenotypes (treatment effect). Such parameters are crucial especially in cases multiple distinct lab-tests (commercial vs. home-brew, technics, amount of informations provided regarding the biomarker). Each patient sub-group outcomes are required to be modeled (costs and health effects). A Markov state-transition model either based on treatment pathway and/or disease staging represents both adequate approaches to simulate the clinical outcomes, incorporating specific efficacy parameters per sub group depending on their biomarker expression levels. In instances, time spent until the CD result delivery exceeds a clinical significant threshold, testing delay shall be modelled such as all parameters driving loss of opporunity. CONCLUSIONS Beyond reasonable simple binary-type of selection biomarker, more complex types of biomarkers and CD technologies (full sequences) has risen additional complexity and poses new methodological challenges.