CFE 2019: Start Registration
View Submission - CMStatistics
Title: Three causal lessons from a simulation learner: On SUTVA, instrumental variables and causal estimands Authors:  Els Goetghebeur - Ghent University (Belgium) [presenting]
Saskia le Cessie - Leiden University (Netherlands)
Abstract: To support Causal reasoning from DAGS and a choice among available estimators, the STRATOS causal inference group developed a `simulation learner'. This engine generates per subject alongside observed exposure(s) and outcome a range of alternative exposures with their potential outcome. As in the Promotion of Breastfeeding Intervention Trial, we `randomize mother-infant pairs to standard of care or a breastfeeding encouragement (BFE) intervention. Main outcome is weight at 3 months. The path from randomization to outcome meets the intervention uptake (education program), followed by the start and a specific duration of breastfeeding. Simulated parallel worlds then enable visualization of various potential outcomes and causal estimands in specific populations. The necessary intermediate steps highlight that SUTVA must be context specific. We see randomisation act as an instrument for one exposure (e.g receiving an offer for the BFE programme or actually following the BFE programme), but not others (e.g. actually starting breastfeeding). We recognize that averaging causal effects over an observed (experimental) instrument may be irrelevant unless one conditions on the instrumental variable. We thus explore distinct estimation methods and compare results with the simulated population parameters. R code is available on, where SAS and Stata code for analysis is also provided This is work on behalf of STRATOS TG 7 Causal Inference