Collider bias occurs when a model is adjusted for a collider or the descendant of a collider. This can occur when there is selection bias (e.g., because of a very low response into the study, loss to follow-up or missing data). It can also occur when analyses are restricted to subgroups of a population, such as the Nurse’s or Physician’s Health studies, other occupational based cohorts, or when studying people with a specific disease (e.g., when looking at the effect of a risk factor or treatment on prognosis/disease progression). The resulting bias in the exposure-outcome causal effect estimate can be in either direction, can mask an effect so that it falsely appears null, or can induce an effect when none exists.
In MR for testing developmental origins/intrauterine effects, adjustment for potential violation of IV assumptions via fetal genetic variants by adjusting for the fetal genetic variants can generate a spurious association between mother’s and father’s genetic variants (because maternal and paternal genetic variants collide on fetal/offspring genetic variants). If it is not possible to adjust for paternal genetic variants (this is often the case) and the paternal phenotype affects the offspring outcome, then the MR result is likely to be biased. New methods for separating maternal and fetal genetic effects and availability of genetic data on trios can be used to mitigate this. MR in people with diseases to explore its progression or results to treatments may be affected by collider bias. For example, if a hypothesised risk factor and disease progression both influence being diagnosed or selected into the study, they are colliding on study selection and this will generate a spurious association between them. The extent to which this has a major effect in different studies is a subject of active research. In two-sample MR, where summary data have been adjusted (e.g., the GWAS of waist-hip ratio adjusted for BMI) the use of these adjusted data can introduce collider bias.
- Aschard H, Vilhjalmsson BJ, Joshi AD, Price AL, Kraft P. Adjusting for heritable covariates can bias effect estimates in genome-wide association studies. Am J Hum Genet 2015;96:329-39.
- Paternoster L, Tilling K, Davey Smith G. Genetic epidemiology and Mendelian randomization for informing disease therapeutics: Conceptual and methodological challenges. PLoS Genet 2017;13:e1006944.
- Pearce N, Lawlor DA. Causal inference – so much more than statistics. International Journal of Epidemiology 2017;doi: 10.1093/ije/dyw328.
- Warrington NM, Freathy RM, Neale MC, Evans DM. Using structural equation modelling to jointly estimate maternal and fetal effects on birthweight in the UK Biobank. Int J Epidemiol 2018.
- Lawlor DA, Richmond R, Warrington N et al. Using Mendelian randomization to determine causal effects of pregnancy (intrauterine) exposures on offspring outcomes: Sources of bias and methods for assessing them. . Wellcome Open Research 2017;2:11.
Other terms in 'Sources of bias and limitations in MR':
- Assortative mating
- Dynastic effects
- Exclusion restriction assumption
- Harmonization failure (in two-sample MR)
- Homogeneity Assumption
- Horizontal Pleiotropy
- Independence assumption
- InSIDE assumption (in two-sample MR using aggregate data)
- Monotonicity assumption
- MR for testing critical or sensitive periods
- MR for testing developmental origins
- No effect modification assumption (Additional IV assumption)
- Non-linear effects
- Non-overlapping samples (in two-sample MR)
- Population stratification
- Regression dilution bias (attenuation by errors)
- Relevance assumption
- Reverse causality
- Same underlying population (in two-sample MR)
- Statistical power/efficiency
- Vertical Pleiotropy
- Weak instrument bias
- Winner's curse