Missing data is very common in observational studies, and the data may be missing for a number of reasons. Missing data will cause estimates to be less precise than they would otherwise have been. They may also result in biased estimates, depending on the mechanism causing them to be missing and the methods used to analyse the observed data.
A second common problem in observational studies of the effect of a particular treatment or exposure is confounding by indication: the decision to treat or not treat may depend on patient characteristics which affect the outcome. Therefore, the difference in outcome between treated and untreated subjects is caused partly by the treatment and partly by this difference in patient characteristics, and hence is a biased measure of the effect of treatment.
The true effect of treatment is the difference between the outcome if the patient receives treatment and the outcome if they do not receive treatment. Since only one of these outcomes can ever be observed, estimating the effect of treatment can also thought of as a missing data problem.
I will present methods which are commonly used to overcome the problems of missing data and confounding by indication, concentrating on the conditions under which they enable the effect of treatment to be estimated without bias.I will also show the parallels between methods that give reliable inference when there is missing data, and methods that give reliable inference when there is confounding by indication.
Disclosure of Interest None Declared
Statistics from Altmetric.com
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.