Causal inference with observational data: the need for triangulation of evidence

Abstract

The goal of much observational research is to identify risk factors that have a causal effect on health and social outcomes. However, observational data are subject to biases from confounding, selection and measurement, which can result in an underestimate or overestimate of the effect of interest. Various advanced statistical approaches exist that offer certain advantages in terms of addressing these potential biases. However, although these statistical approaches have different underlying statistical assumptions, in practice they cannot always completely remove key sources of bias; therefore, using design-based approaches to improve causal inference is also important. Here it is the design of the study that addresses the problem of potential bias – either by ensuring it is not present (under certain assumptions) or by comparing results across methods with different sources and direction of potential bias. The distinction between statistical and design-based approaches is not an absolute one, but it provides a framework for triangulation – the thoughtful application of multiple approaches (e.g. statistical and design based), each with their own strengths and weaknesses, and in particular sources and directions of bias. It is unlikely that any single method can provide a definite answer to a causal question, but the triangulation of evidence provided by different approaches can provide a stronger basis for causal inference. Triangulation can be considered part of wider efforts to improve the transparency and robustness of scientific research, and the wider scientific infrastructure and system of incentives.

Effects of glycemic traits on left ventricular structure and function: a mendelian randomization study

“…In addition, the intercept obtained from MR-Egger regression can be used as a measure of unbalanced pleiotropy ( P < 0.05 indicated significance) [ 15 ]. Although the weighted median and MR-Egger methods are not as precise as the IVW method due to lower statistical power [ 33 ], the estimates will be more reliable if there are consistent results in the same direction across all those three methods [ 34 ]. MR-PRESSO method was used to detect and correct for any SNP outliers that reflected potentially pleiotropic biases.…”

Assessment of long-term outcomes post living liver donation highlights the importance of scientific integrity when presenting transplant registry data

“…Even within the context of appropriate study design, it is important to note that analysis of observational registry data can lead to errors related to causal inference, the concept that a specific exposure results in a causal effect on an outcome being analyzed. 5 Studies using observational data can only compare the risk of the outcome in those who were exposed to the risk of the outcome in the unexposed population. In the context of the studies compared in this Viewpoint, the survival of the "healthy" group does not represent the counterfactual outcome of the living liver group had they not proceeded with donation.…”

Post-acute COVID-19 cognitive impairment and decline uniquely associate with kynurenine pathway activation: a longitudinal observational study

“…In conclusion, KP activation was found to be the first biomarker for post-acute COVID-19 cognitive impairment/decline, highlighting the KP as a monitoring tool and therapeutic target. 41 Severe acute disease: Patients who had been hospitalized (see 52 for more details) Medical comorbidities which were investigated/included: chronic cardiac disease (n=10); hypertension (treated) (n=16); obesity (n=7); chronic lung disease (n=1); asthma (treated) (n=11); diabetes (treated) (n=8); chronic kidney disease (n=3); liver disease (n=0); cancer (treated) (n=7); HIV infection (treated) (n=1); transplant recipient (n=1); rheumatologic or autoimmune disorder (n=9); obstructive sleep apnoea (treated but 1), n=3); other comorbidities (n=2: 1 early Parkinson's disease treated) Note: 4 women were pregnant but in the gestation period); 8.7% (n=11) has a history of a psychiatric diagnosis (anxiety disorder and major depressive disorder on treatment) Only the √KYN is illustrated as it also presents the individual values in a more realistic fashion that when using the dummy code. Note however that the dummy coded outcome model (0=within normal range; 1=KYN >3uM), showed the best fit (AIc=574 versus 897), but same direction of effect.…”