ICE FALCON - Inference on Causation from Elimination of FAmiliaL
by Professor John Hopper
Abstract: Inference on Causation from Elimination of FAmiliaL CONfounding (ICE FALCON) is an approach to the analysis of data for pairs of traits measured for pairs of related individuals that tests the hypothesis that at least part of the association between two traits is due to a causal relationship.
It is well known that finding an exposure to be associated with a disease or condition does not of itself prove that the exposure is a cause; the exposure could be associated with the real cause. For example, people who smoke tend also to be drinkers, and vice versa. Therefore, if smoking causes lung cancer there will also be an association between alcohol consumption and lung cancer that is not causal. This is called "confounding". We make inference about causation by trying to eliminate it. We do this by considering smoking and alcohol consumption together as predictors of lung cancer; if smoking is causal we should find that there is no association between drinking and lung cancer once smoking has been taken into account.
The same principle applies in Ice Falcon, only this time we study the association between a person's outcome (e.g. disease) and both their own exposure and the exposure for their twin. If an exposure is causal, and correlated in pairs of relatives, then a person's risk of the disease will depend on their relative's exposure. For example, if the disease is breast cancer and the exposure is having a high-risk mutation in a breast cancer susceptibility gene like BRCA1 or BRCA2, then a woman who has a sister who carries a mutation is at increased risk (because she has a probability 1/2 of being a carrier as well). However, if one knows the exposure of the woman herself, in terms of her breast cancer risk it becomes irrelevant what her sister's status is for that exposure. That is, we make inference about causation, this time by trying to eliminate familial confounding.
We have developed a regression approach that fits a person's outcome
(Yself) against one or both of their own exposure (Xself) and/or their relatives's exposure (Xrelative). The method will only work if Xself and Xrelative are correlated - so if this method is applicable a twin design is the most appropriate. It is straightforward to allow for Yself and Yrelative to be correlated; e.g. if Y is continuously distributed, by using a bivariate normal model (and perhaps a suitable transformation) and fitting the correlation. The method can also be applied if Y is a binary or ordinal variable by using a suitable generalised linear modelling.
We will demonstrate application of this approach to outcome and exposure data for twin and sibling pairs, and also to repeated/longitudinal measures for the same individual. It can also be applied when X is a measured genetic variant, either rare as in the example above or common, and has the potential to help resolve issues about the clinical significance of so-called 'unclassified variants'. This method might open up twin research to a new group of researchers.
For More Information: Contact Quoqi Qian (firstname.lastname@example.org) or Paul Norbury (email@example.com)