Abstract
In observational studies on causal associations, comparison groups (e.g. groups of treated and untreated subjects) are likely to differ on prognostic characteristics. Such incomparability will bias the association under study, which is referred to as confounding. The studies presented in this thesis focused on the reporting and quantification of observed and unobserved confounding. In chapter 2 a review of methods to quantify observed and unobserved confounding is presented. Then, in chapter 3, two reviews of the reporting of confounding in observational intervention studies were presented. A systematical literature search indicated that from 1985 through 2005 in only 9% of the papers on observational intervention studies confounding was addressed in title or abstract. A detailed investigation of current reporting in leading medical and epidemiological journals showed that out of eight pre-specified essential items in the reporting of confounding, on average only four were reported in the papers that were studied. The studies described in chapter 4 focused on observed confounding. First, it was shown in a study on influenza vaccine effectiveness, that pooling of observed confounders did not result in important residual confounding. In the second part of chapter 4, a numerical example indicated that subgroup effects could be confounded in both individual trials and individual patient data meta-analyses. Chapter 5 focused on methods to quantify unobserved confounding. First, two-stage sampling was used to assess impact of potential confounders, which were initially not observed. In this study, potential confounders (such as functional health status, smoking status, alcohol consumption, and educational level) did not affect the association between influenza vaccination and mortality risk. Secondly, a reference period for which the effect of influenza vaccination was known, namely summer, was used to adjust the observed vaccine effectiveness during an influenza epidemic for unobserved confounding. Adjustment for unobserved confounding reduced influenza vaccine effectiveness from 42% to 31%. Then, several sensitivity analyses to estimate the potential impact of an unobserved confounder were reviewed and applied in a study on influenza vaccine effectiveness. Finally, the potential for instrumental variables to study influenza vaccine effectiveness was assessed. All potential instrumental variables that were studied (i.e., use of antacid medication, a history of gout, a history of orthopaedic morbidity, and GP group practice characteristics) did not meet the assumptions of instrumental variable analysis. The final chapter provides a general discussion of the methods to quantify observed and unobserved confounding. This discussion then serves as a base for recommendations on the reporting and handling of confounding in observational causal studies.
Translated title of the contribution | Quantifying confounding |
---|---|
Original language | Undefined/Unknown |
Qualification | Doctor of Philosophy |
Awarding Institution |
|
Supervisors/Advisors |
|
Award date | 19 May 2009 |
Publisher | |
Print ISBNs | 978-90-393-5038-6 |
Publication status | Published - 19 May 2009 |