I will be giving a seminar within the Faculty of Health and Social Care Sciences on 11 October 2012, 13:00 to 14:00, introducing for the first time some methodological work that has taken up a lot of energy over the last two years. All are welcome; it would be helpful if you e-mailed N.Greenwood@sgul.kingston.ac.uk so we can get the numbers right. I will be submitting the paper soon so it could be a while before this appears again in the public eye! I am not one to exaggerate my own work but I think this is an important new method for epidemiology and observational studies generally.
Title: A new method for dealing with residual confounding: a practical introduction for researchers
In this seminar I will outline recent work I have carried out to develop a new statistical method, in practical and non-mathematical terms. Confounding is an almost universal problem in observational (non-randomised) studies, where the predictor of interest is correlated with other factors, causing one to over- or under-estimate the effect it truly has on an outcome. There are many tools for separating the predictor of interest from the confounder, but these fail if the confounder has been imperfectly measured, for example recording smoking simply as current / ex / never. This is the situation called residual confounding, and the received wisdom is that nothing can be done about it.
It is however possible to adapt modern methods for missing data (multiple imputation) and use this to correct the imperfections and remove residual confounding. Some assumptions have to be made and these are safer in some situations than others. I will explain what is required in terms of information and expertise, and show some examples.