Such collider bias could occur through initial selection, or selective drop-out, or both–for example, a study could be representative of its intended population initially, but become less representative as those of poorer health drop out due to death. The main difference between these two scenarios–initial selection and selection through attrition–is in the amount of information available on the missing individuals. Where some data are available for all participants (e.g. in the case of drop-out), then multiple imputation or inverse probability weighting can be used28 under some assumptions which are untestable given the observed data, to recover unbiased estimates of associations. However, where there is no information on missing individuals (e.g. we have no data on individuals who did not volunteer for participation into a study), then such methods cannot be used. External information (such as the expected proportion of males and females in the general population) could be used to investigate likely factors related to participation and to derive bias-adjusted estimates.
