Document resource
Background Instrumental variable (IV) analysis uses proxy variables for an exposure to provide potentially unconfounded causal effect estimates, but can be biased when IV assumptions are not met. Determining when instruments are invalid is challenging. Robust IV estimators have been developed for Mendelian randomisation (MR) (using genetic IVs) providing consistent, unbiased estimates even if a proportion of instruments are invalid. We investigate how these estimators can be used in a non-genetic context.Methods We used year-on-year changes in generosity of multiple welfare benefits to instrument for income’s effect on mental health (GHQ-12 score) and life satisfaction. Individual fixed-effects IV analyses were conducted for each instrument, and resulting IV effect estimates were analysed using meta-analysis methods from the MR field (inverse variance weighted, MR-Egger, median, and mode-based estimators). The approach was demonstrated using a simulated dataset (20 instruments affecting 10,000 individuals over 5 time points) under scenarios with 25%, 50%, 75%, and 100% valid instruments, with and without correlation between instrument strength and direct effects. We then applied the methods to UK Household Longitudinal Study (UKHLS) data covering 2010 to 2020, using predicted benefit eligibility from a static tax-benefit microsimulation (UKMOD).Results Simulations demonstrated that robust IV estimators can reduce bias when non-genetic instruments violate IV assumptions by directly affecting the outcome. The weighted median and mode estimators remained unbiased with up to 50% invalid instruments, while MR-Egger was unbiased but imprecise when direct effects were uncorrelated with instrument strength. In UKHLS data (n=59,249 individuals, 290,264 observations), six benefits were strong enough (F-statistic > 10) to include as instruments for social benefit income. The inverse variance weighted estimator suggested small effects of £100 monthly benefit income on GHQ-12 score (0.023 SD [95% CI: -0.003, 0.050]) and life satisfaction (0.009 SD [-0.018, 0.037]) in the same year; larger than the conventional fixed effects estimates of 0.000 SD (-0.002, 0.003) and -0.001 SD (-0.003, 0.002), respectively. Estimates were largely consistent across the bias-robust estimators, and robust to leave-one-out analyses, except for life satisfaction effects being primarily driven by Working Tax Credit. Results from MR-Egger were very imprecise.Conclusion Robust IV estimators previously applied in MR are feasible to use for non-genetic IVs and may be useful when multiple instruments are available. These methods may help alleviate important limitations of IV methods and expand their use in public health research. The greatest limitation is the need for a sufficient number of distinct, strong IVs.