In 2014, the R&D Department of Unicancer, a national academic network of comprehensive cancer centers in France, launched the Epidemiological Strategy and Medical Economics (ESME) Research program to centralize real-world patient data in oncology. The program's first project concerns the construction of a comprehensive database on patients with metastatic breast cancer (MBC), called ESME MBC Database, using data obtained from the 18 French Comprehensive Cancer Centers. Among the studied issues, ESME MBC Database (NCT03275311) [NCT03275311?titles=ESME%26rank=2] has served to describe the Overall Survival (OS) in patients with HER2-negative MBC treated with first-line paclitaxel chemotherapy, with or without bevacizumab [1Delaloge S., et al. Ann Oncol 2016 ;  27 : 1725-1732 [cross-ref]

Cliquez ici pour aller à la section Références]. Nevertheless, results have been criticized due to the lack of randomized trials, and the eventual presence of bias due to unmeasured confounders. In this work, we investigated two methods to address this issue : instrumental variables (IV) [2Greenland S. Int J Epidemiol 2000 ;  29 : 1102

Cliquez ici pour aller à la section Références] and quasi-trials analysis [3De Gramond A., et al. JCO 2007 ;  25 : 3224-3229

Cliquez ici pour aller à la section Références].

The IV should satisfy two conditions : IV should be correlated to the endogenous variable treatment ; IV should not be correlated to the output (OS) and affect the OS only indirectly through its effect on the treatment. We used “Activity level of FCCC”, which represents the number of cases in the whole database per contributing FCCC as IV with five strata. Analysis was carried out with the two-stage residual inclusion method (2SRI). For the quasi-trial analysis, the idea is to use a “superficial” randomized trial using variables, for example centers, reflecting a definite preferences for certain treatment options. Indeed, comparisons of results between centers can be informative and it is likely such differential preferences exist in real-life practice. In this study, we have considered the percentage of patients with bevacizumab+paclitaxel in each center, regardless of treatment arm.

In the first analysis published by Delaloge and al., the OS adjusted for major prognostic factors was significantly longer in the paclitaxel and bevacizumab group compared with paclitaxel [hazard ratio (HR) 0.672, 95 % confidence interval (CI) 0.601–0.752] [1Delaloge S., et al. Ann Oncol 2016 ;  27 : 1725-1732 [cross-ref]

Cliquez ici pour aller à la section Références]. Using IV method to support these results seems difficult due to the lack of candidate variables. The level of activity of FCCC, with five strata, leads to different conclusions compared with those obtained in Delaloge's paper.

For the quasi-trial analysis, the percentage of patients with bevacizumab+paclitaxel had an independent and significant impact on OS [HR=0.180, CI=0.043–0.748]. Centers in which more than 50 % of the patients were reintroduced had an adjusted HR for OS of [HR=0.735, CI=0.645–0.837] compared with centers in which 50 % or less of patients were treated with bevacizumab+paclitaxel. These results consolidate those obtained with a conventional multivariate Cox model.