With the availability of large public databases such as NSQIP, anesthesiology researchers can quite easily assess the relationship between an intervention and an outcome for a wide range of research questions. In such non-randomized studies we try to remove selection bias by statistically adjusting for any available baseline variables that might confound the association between intervention and outcome. We basically adjust for all baseline variables, ‘the kitchen sink’, either through propensity matching (followed by comparing matched groups on outcome) or else as part of a single multivariable model assessing the relationship of interest. The adjustments nicely ‘balance’ the two intervention groups on the potentially confounding baseline variables and allow more unbiased inference on the relationship of interest.
However, the study by Turan et al
assessing the effect of smoking on postoperative outcomes in this month’s edition presents an interesting and important exception. When the exposure of interest is not an intervention applied in the hospital, but rather a chronic habit (e.g., smoking) or disease (e.g., diabetes) or condition (e.g., BMI), adjusting for all available baseline variables would often be expected to wash away some of the true treatment effect. For example, smokers may be expected to do worse than never-smokers because smoking causes hypertension, coronary artery disease, COPD or other conditions, and adjusting for those medical history variables might ‘explain away’ much of the smoking effect. Adjusting for all such variables might be expected to result in no effect whatsoever!
So what to do?
To address the issue, Turan et al a priori named baseline conditions plausibly resulting from smoking, i.e., which may represent the main reasons why smokers would be expected to do worse. Those ‘mediator’ variables were not included in the set of confounding variables that the smokers and non-smokers were matched on. This approach allowed estimation of the overall smoking effect as the aggregate effect of those ‘mediator’ variables (and other unmeasured variables) on the composite morbidity outcome, resulting in an odds ratio of 1.4 for current versus never-smokers. The so-called ‘direct effect’ of smoking was estimated by then forcing those mediator variables into the model, giving an odds ratio of 1.27. Finally, by comparing those 2 models they were able to estimate just how much of the overall smoking effect was due to the specified mediator variables (about 30%). Focus on the epidemiology of the smoking exposure in the planning stages was needed to conduct a meaningful analysis.
I expect this sort of mediation analysis will be gaining steam in anesthesiology and perioperative medicine for database studies in which the exposure is a diseases or condition. But clinical trials can benefit too. For example, one might hypothesize that an intervention such as glucose control affects postoperative major morbidity outcomes via some mechanism such as lowering of inflammation, measured by change in cytokine levels. Assessing how much of the effect on the clinical morbidity outcome is mediated through changes in cytokines would give valuable information to the research conclusions.Your comments
about the statistics are welcome.
Posted by Edward Mascha, Ph.D.