Selection biases can distort causal estimates and undermine policy evaluations, yet detecting and quantifying selection remains empirically challenging. This paper introduces a novel method for detecting and quantifying selective survival using genetic data. Specifically, Polygenic Indices (PGIs) are predetermined at birth and should be equally distributed between treatment and control groups in the absence of selection. Any systematic differences in PGIs after an intervention suggest the presence of selection or survival effects.
Our primary case study focuses on the sharp introduction of the National Health Service (NHS) in the UK in July 1948, which provided free and universal healthcare. Following the NHS implementation, hospital births increased, leading to improved neonatal health. First, we use a regression discontinuity design (RDD) and analyze administrative data from the Registrar General’s Statistical Review to find that the NHS reduced infant mortality rates (IMR) by 5.48 per 1,000 births, a 16.1% decline relative to the mean.
This sharp decline left a genetic imprint on affected cohorts. NHS exposure increased PGI values predictive of detrimental medical conditions, such as chronic obstructive pulmonary disease (COPD) and depression, while reducing those linked to beneficial traits, such as life satisfaction and self-rated health. These shifts are substantial, with effect sizes reaching 10% of a standard deviation in the UK Biobank (UKB) and up to 40% in the English Longitudinal Study of Ageing (ELSA). These results remain robust when the UKB is restricted to sibling pairs born before and after the NHS implementation, and when family fixed effects are included.
Moving on to the estimation of long-term effects, we leverage the availability of genetic data to partially control for selection effects and bound the estimated treatment effect. In the nationally representative ELSA dataset, NHS exposure is associated with improved self-reported health and lower body mass index (BMI) among individuals aged 45 to 80. Controlling for PGIs increases estimated effect sizes by up to 43%, suggesting that selection bias attenuated previous estimates. However, these effects are not replicated in UKB, where no significant long-term health impacts are observed.
Our findings highlight the importance of accounting for survival bias in policy evaluations and demonstrate how genetic data can improve causal inference by detecting and adjusting for selection effects.
Presented by:
Dr Nicolau Joaquim Martin Bassols (University of Bologna)
Date & time:
May 28, 2025 12:30 pm - May 28, 2025 1:30 pm
Venue:
2N2.4.16
External seminars home