Methodological guidelines and guidance to employ cutting-edge quasi-experiments using pooled cohort data

Task 34 Causal Inference in pooled cohortsProper understanding of causal mechanisms is crucial to deliver a personalized treatment strategy in individuals Adoptions and further developments of methods that have originated in economics political science and psychology have recently substantially expanded the methodological toolkit for causal inference in epidemiology and biostatistics These methods include both nonexperimental innovations which lack the ability to control for unobserved confounding but improve our ability to control for observed confounders including Gcomputation marginal structure models MSM and novel variants of propensity score matching as well as quasiexperimental approaches which have the ability to control for unobserved confounding because they use quasirandom exposure assignments identified in cohort data Some quasiexperimental methods have the ability to completely control for unobserved confounding such as regression discontinuity interrupted time series and instrumental variable analysis while others can only partially control for unobserved confounding such as fixedeffects and differenceindifferences analysisIn this task we will evaluate adapt and extent existing methods for causal inference for use in pooled cohort data These methods when integrated in the development of risk prediction models can directly be used to improve the estimation of personalized treatment response see also Task 47 Based on this work we will develop guidelines and guidance documents for causal inference using pooled cohort data On the one hand we will evaluate and adopt nonexperimental methods such as Gcomputation and MSM for this purpose and we will develop approaches to include causal inference methods in socalled onestage metaanalysis models where all cohorts are simultaneously analyzed rather than in separation for causal inferenceOn the other hand we will develop guidance on identifying quasiexperimental opportunities in pooled cohort data While fixed effects opportunities are often easy to identify and generate because the unit of observation in longitudinal data collection here mostly individuals are immediately discernable in pooled cohort data differenceindifferences instrumental variable and regression discontinuity opportunities typically require integration of information external to the pooled cohort data To increase the use of quasiexperimental approaches in pooled cohort data we will establish conceptual and methodological guidance for identification of analytical opportunities and rigorous application similar to a series that we recently edited in the Journal of Clinical Epidemiology but with specific view to pooled cohort data analysis Pooled cohort analysis offer particularly promising but also challenging opportunities for quasiexperiments Instrumental variables and regression discontinuity thresholds may vary across geography and time requiring approaches that go beyond the current standard analyses to ensure poolability of analyses and results A further set of innovations will be possible through the use of prospectively planned quasiexperiments Often it is not possible to randomize an exposure of interest for legal ethical or political reasons for instance because an intervention is strongly believed to be effective However in these cases it will often be possible to induce an instrumental variable by assigning an exposure that significantly affects that intervention of interest but is highly unlikely able to affect the outcome of interest We will further improve the ability to carry out regression discontinuity analysis through prospective planning and data collection This innovation allows for costefficient regression discontinuity because data collection on specific exposure and outcome can target observations that lie close to the regression discontinuity threshold comp