Advertisement

Understanding Confounding in Observational Studies

Open ArchivePublished:March 08, 2018DOI:https://doi.org/10.1016/j.ejvs.2018.02.028
      Even under ideal conditions, clinical research is threatened by typical flaws that may adversely affect data interpretation and the generalisability (i.e. validity) of findings. Critical appraisal of scientific reports is crucial and requires a clear understanding of such flaws. Whereas the first part of this edutorial dealt with various forms of bias,
      • Meuli L.
      • Dick F.
      Major sources of bias.
      this part aims at illustrating the concept of confounding.
      Confounding is a typical hazard of observational clinical research (as opposed to randomised experiments). Unfortunately, it may easily pass unrecognised even though its recognition is essential for meaningful interpretation of causal relationships (e.g. when assessing treatment effects).
      Confounding occurs when an apparently causal relationship between an exposure (e.g. a treatment) and an outcome is, in reality, distorted by the effect of a third variable (the confounder). By definition, confounding factors must fulfil three criteria
      • Porta M.
      A dictionary of epidemiology.
      : (1) they must be related to both exposure (i.e. risk factor, intervention, or treatment) and outcome; (2) they must be distributed unequally between study groups; and (3) they must not be an intermediary step in a causal pathway between exposure and outcome.
      A classic example of confounding refers to the observed correlation between birth order and the risk of Down syndrome, where the incidence of Down syndrome may seem to be linked directly to the position in birth order (i.e. the more older siblings, the higher the risk of Down syndrome).
      • Stark C.R.
      • Mantel N.
      Effects of maternal age and birth order on the risk of mongolism and leukemia.
      However, women giving birth to their second or third child are usually older than women giving birth to their first child. Thus position in birth order is linked to maternal age. If maternal age, in turn, was also linked to Down syndrome (precondition 1), the correlation between birth order and Down syndrome could potentially be explained (i.e. be confounded) by maternal age. Indeed, stratification by both maternal age and birth order (see below) shows clearly that the apparent predictor “position in birth order” does, in reality, not correlate independently with the risk of Down syndrome, but only via confounding by “maternal age”.
      • Stark C.R.
      • Mantel N.
      Effects of maternal age and birth order on the risk of mongolism and leukemia.
      Like bias, the risk of confounding is influenced by the study design. Randomised experiments for instance avoid precondition 2 above (unequal distribution of potential confounders) and are thus “immune” to confounding, whereas observational evaluations of clinical routine are highly vulnerable for reasons explained further below. In contrast to bias, however, the risk of confounding can (and should!) always be addressed by statistical measures, that is statistical adjustments for all potential confounding factors. Of note, the effect of such adjustments is naturally limited to known (and measured) confounding factors.
      The simplest statistical method to adjust for confounding is stratification, that is grouping the sample into different layers (“strata”) according to potential predictors. The objective is to fix one predictor to assess the “independent effect” of others, that is potential confounding factors. In the above example, comparing groups of similar maternal age or equal birth order regarding prevalence of Down syndrome showed that within any given maternal age group, Down syndrome prevalence did not change with increasing birth order. But for any given position in birth order, it did correlate with increasing maternal age. Thereby, assessment of potential predictors became “independent” of the other. The Cochran-Mantel-Haenszel test is a well known statistical stratification model to adjust for single variables. But there are many (and more advanced) statistical models to adjust for any number of variables (i.e. multivariate models including linear, logistic, Cox, and Poisson regression models). Of note, each of these models is based on specific premises, which need to be observed.
      • Rothman K.J.
      Epidemiology an introduction.
      In vascular surgery, clinical decision making typically considers patient related factors including age, gender, frailty, cardiovascular risk profile, and many others. These factors may not only affect treatment decisions and choice of interventions, but they typically also influence outcomes (e.g. patient survival or complication rates). Therefore, studies observing clinical routine to evaluate treatment efficacy (“outcome studies”) are at particular risk of confounding. Crude (i.e. unadjusted) comparisons may just reflect (unrecognised) confounding by above contextual factors rather than actual treatment effects. As a consequence, observational studies should always be scrutinised for whether outcome assessment was adjusted for potential and plausible confounding factors. Keeping in mind that any statistical adjustment is limited to known and measured factors, a risk of residual confounding will usually persist in outcome studies.
      In conclusion, acceptance of reported outcome differences between treatment groups (i.e. of treatment effects) must always be pondered against the possibility of unaccounted confounding by a third factor, which might have caused the effect in reality (and not the ‘treatment’). Thereby, methods and results sections must be scrutinised for whether potential confounding factors were adjusted for and whether identification of such factors was plausible and comprehensive. More in depth information may be found in the specialised literature.
      • Rothman K.J.
      Epidemiology an introduction.

      W.W. Lamorte and L. Sullivan, Confounding and effect measure modification, Weblearn program, Boston University School of Public Health, [Accessed 2 March 2018], http://sphweb.bumc.bu.edu/otlt/mph-modules/bs/bs704-ep713_confounding-em/index.html.

      References

        • Meuli L.
        • Dick F.
        Major sources of bias.
        Eur J Vasc Endovasc Surg. 2018; 55: 736
        • Porta M.
        A dictionary of epidemiology.
        6th ed. Oxford University Press, Oxford2014
        • Stark C.R.
        • Mantel N.
        Effects of maternal age and birth order on the risk of mongolism and leukemia.
        J Natl Cancer Inst. 1966; 37: 687-698
        • Rothman K.J.
        Epidemiology an introduction.
        2nd ed. Oxford University Press, 2012
      1. W.W. Lamorte and L. Sullivan, Confounding and effect measure modification, Weblearn program, Boston University School of Public Health, [Accessed 2 March 2018], http://sphweb.bumc.bu.edu/otlt/mph-modules/bs/bs704-ep713_confounding-em/index.html.

      Comments

      Commenting Guidelines

      To submit a comment for a journal article, please use the space above and note the following:

      • We will review submitted comments as soon as possible, striving for within two business days.
      • This forum is intended for constructive dialogue. Comments that are commercial or promotional in nature, pertain to specific medical cases, are not relevant to the article for which they have been submitted, or are otherwise inappropriate will not be posted.
      • We require that commenters identify themselves with names and affiliations.
      • Comments must be in compliance with our Terms & Conditions.
      • Comments are not peer-reviewed.