Multiple-Bias Modelling for Analysis of Observational Data

Abstract

Summary.Conventional analytic results do not reflect any source of uncertainty other than random error, and as a result readers must rely on informal judgments regarding the effect of possible biases. When standard errors are small these judgments often fail to capture sources of uncertainty and their interactions adequately. Multiple-bias models provide alternatives that allow one systematically to integrate major sources of uncertainty, and thus to provide better input to research planning and policy analysis. Typically, the bias parameters in the model are not identified by the analysis data and so the results depend completely on priors for those parameters. A Bayesian analysis is then natural, but several alternatives based on sensitivity analysis have appeared in the risk assessment and epidemiologic literature. Under some circumstances these methods approximate a Bayesian analysis and can be modified to do so even better. These points are illustrated with a pooled analysis of case-control studies of residential magnetic field exposure and childhood leukaemia, which highlights the diminishing value of conventional studies conducted after the early 1990s. It is argued that multiple-bias modelling should become part of the core training of anyone who will be entrusted with the analysis of observational data, and should become standard procedure when random error is not the only important source of uncertainty (as in meta-analysis and pooled analysis).

Sensitivity analysis for the effects of multiple unmeasured confounders

“…In addition to the strength of the associations of the unmeasured confounders with exposure and outcome, their mutual correlation is a key driver of the magnitude of bias resulting from unmeasured confounders. We support the recommendation by others, 6,[8][9][10][11][12] that researchers of observational studies should routinely perform sensitivity analysis of (multiple) unmeasured confounders in Based on the method described by Lin, Psaty, and Kronmal. 10 The dotted line indicates a scenario in which an unmeasured confounder increases the odds of the outcome seven times, is negatively associated with exposure (OR 0.2), and is present in 25% of the population.…”

“…[6][7][8][9][10][11][12][13][14][15] Some focus on unmeasured confounding only, [7][8][9][10][11][13][14][15] whereas in other publications an encompassing framework for unmeasured and poorly measured confounding, together with measurement error and other sources of bias sensitivity analyses is proposed. 6,12 Often sensitivity analyses of unmeasured confounding focus on a single unmeasured variable. [7][8][9][10] Consequently, when thinking about the potential for unmeasured confounding in observational studies, researchers will often have a single unmeasured variable in mind.…”

“…The possibility that there could be multiple (unmeasured) confounders for any given association has been confirmed by the finding that 6 many measured subject characteristics/confounders in cohort studies are associated with one another, much more than would be anticipated by chance. 19 Although multiple unmeasured confounders can be summarized into a single variable, 6,8 this may be hard to conceptualize and the literature provides little guidance on how to construct such a summary.…”

“…[3][4][5] Publications describing methods for sensitivity analysis of unmeasured confounding typically focus on the impact of a single confounding variable. [6][7][8][9][10][11][12][13][14][15] It is often assumed that only a variable with a strong association with both exposure and outcome can materially confound the association under study. For example, in an observational study of influenza vaccine effectiveness, sensitivity analysis of unmeasured confounding was conducted by simulating a single unmeasured confounder under a wide range of scenarios.…”

Use of evidence in economic decision models: practical issues and methodological challenges

“…For example, it is acknowledged that observational studies are often of poorer quality than RCTs as they are prone to internal biases such as selection bias, non-response bias, and/or confounding but can this be adjusted for in the analysis? Possible solutions from the statistics literature include: (i) exclusion of studies below a pre-specified quality threshold, (ii) adjustment of the weight given to a study in the analysis by quality (Tritchler, 1999), (iii) random effects modelling of bias (Spiegelhalter and Best, 2003) by choosing a parameter to describe the plausible size (and direction) of the bias for each study or study design, and (iv) full bias modelling (Greenland, 2005) that attempts to identify all sources of potential bias in the available evidence and to obtain external information on the likely form of each bias and then construct a model to correct the data analysis accordingly. The first method is very simplistic with an arbitrary threshold whereas the other methods are subjective regarding the judgement about how study quality/design affects the reliability and precision.…”