Research Questions: How to minimize bias in causal effect estimation with observational studies.
Objectives: to examine and quantify biases in methods used to estimate causal effects using observational data, and to demonstrate the performance of methods developed to reduce these biases.
Scientific Rationale: The gold standard for estimating causal effects is a well-designed randomised controlled trial. However, these are not always available, and so for many questions, observational data are used – with one of the most-often used studies being UK Biobank. Analysis of such data is known to be prone to bias including unmeasured confounding, measurement bias, and selection bias. Methods developed to overcome/minimise/quantify such biases include Mendelian Randomization, multiple imputation, inverse probability weighting, g-formula, g-estimation, marginal structural models and other ways to adjust for confounders. Increasing emphasis is placed on examining robustness of conclusions to biases, including via quantitative bias analyses and sensitivity analyses. All these methods require simulation data for their development, but also testing, refining and comparing using large, realistically complex real-data analyses.
Many research publications usually provide a qualitative discussion of the potential biases from their causal effect estimation, only a few provide a quantitative assessment or adjustment of their impact. The lack of implementation of bias adjustments is potentially due to the scarcity of analyst-friendly methods and software. Therefore, there is need for developing more flexible methods, guidelines and software that non-technical analysts (such as medical researchers and epidemiologists) can easily use. Furthermore, even when such biases are identified methods that can provide estimates that are robust to the particular source of bias are often unavailable or not well understood.