Many research questions in epidemiology can only be answered using observational data because randomized controlled trials would be practically or ethically infeasible, e.g. when investigating long-term effects or vulnerable subpopulations. However, analyses of observational data can yield highly misleading results if conducted in a way that violates basic principles of study design, as illustrated by the so-called HRT story (Hernán et al 2008 Epidemiology 19:766).
To overcome these issues and sensitize to these avoidable biases, the powerful general principle of target trial emulation (TTE) has been developed (Hernán & Robins 2016 Am J Epidemiol 183:758). It ensures the practical interpretability and meaningfulness of any analysis, and achieves a high protection against avoidable biases. The principle of target trial emulation is increasingly being adopted in the field of causal analyses of electronic health records, claims data or other observational data (Caniglia et al 2020 Neurology 95:e1322).
The overall goal of the GeTTCausal working group is to apply and further develop the TTE approach to advance our knowledge on effects and side-effects of medical treatments or interventions. We will also explore and develop methods for tailored sensitivity analyses to address unavoidable biases, e.g. by using negative controls, sequential eligibility restrictions, or quantitative bias modelling.
The working group GeTTCausal capitalises on and combines BIPS' expertise in causal analysis, clinical epidemiology and claims data analysis. With BIPS' access to the German pharmacoepidemiological research database (GePaRD) we are uniquely placed to carry out cutting-edge robust analyses that will be of direct relevance to patients, doctors and public health decision makers.
Using GePaRD for TTE provides the following opportunities:
- the database covers ~20% of the German population (~16 million per year, total number of persons in the database: 25 million); all regions and age groups in Germany are represented.
- it covers a time span of over 15 years (data beginning in 2004)
- there is no volunteer nor recall bias
- vulnerable subgroups such as pregnant women are well-represented and identifiable
- all levels of care are covered (in- and outpatient care, including specialists).
Completed, ongoing or planned projects address, for example,
- the effectiveness of mammography screening in reducing breast cancer mortality in Germany
- the effectiveness of screening colonoscopy in reducing incidence of colorectal cancer in the distal vs. the proximal colon
- the effect of antidiabetic drugs on pancreatic cancer incidence (Börnhorst et al. 2021 Clin Epi 13:1027)
- a head-to-head comparison of different direct oral anticoagulants in atrial fibrillation (effectiveness and safety)
- the effect of combined use of tamoxifen and antidepressants on breast cancer recurrence.
Vice versa, using TTE on GePaRD provides the following opportunities:
- the analysis is guided by an explicit research question so as to inform decision makers (Didelez 2016 IJE 45:2049)
- various self-inflicted (usually time-related) sources of bias are avoided, e.g. immortal time or prevalent user bias
- potential sources of bias, e.g. protopathic bias, are more easily detectable and can explicitly be addressed or assessed in sensitivity analyses
- TTE provides a framework that can be combined with a wide variety of statistical methods.
Regarding statistical approaches, the working group aims at developing a systematic procedure for choosing suitable methods e.g. to adjust for confounding and for sensitivity / bias analyses. Hence, we will compare existing, and develop novel methods for TTE with GePaRD, such as double-machine learning to adjust for high-dimensional cofounding or separable effects approaches in competing risks settings (Didelez 2019 LIDA 25:593; Stensrud et al 2020 JASA 117:175). Moreover, we will exploit opportunities to empirically illustrate the superiority of TTE over common but biased approaches, thus contributing to improved research with observational data (Pigeot et al 2021 Gesundheitswesen 2021 83:S69).