Non-Independence in Competing Risk Models 

 A central assumption in competing risk analysis is the conditional independence of the risks under analysis.  Suppose we are interested in cause-specific mortality due to causes A, B, and C.  If we assume that the process leading to death from A is independent (conditional on covariates) from the process leading to death from B, then the likelihood factors nicely, and estimation via a series of standard 0/1 hazard models is straightforward.  For example, it may be reasonable to assume that death from lung cancer (cause A) is independent of death from being struck by a meteorite (cause B).  But it is much less reasonable to assume that death from lung cancer (A) is independent of the risk of dying from emphysema (C), unless we are lucky enough to have, say, appropriate covariate information on smoking history.    
 


 The problem is partly rhetorical.  The independence assumption in competing risk analysis is the exact same as the assumption of independent censoring in standard hazard models.  Few applied papers even mention the latter (unfortunately). In competing risk analysis, however, the assumption becomes quite a bit more visible, and thus harder to hide… 

 There are a small number of strategies, none particularly popular, to cope with dependence.  Sanford C Gordon recently contributed a new strategy in “Stochastic Dependence on Competing Risks? AJPS 46(1), 2002, which builds on an earlier idea of drawing random effects.  Rather than drawing individual specific random effects, as has been suggested before by Clayton 1978, Gordon draws risk and individual specific random effects.  Thus, a K-risk model on a sample of N individuals may contain up to KxN separate random effects, one for each risk and individual.   

 The advantage of this strategy is that it allows for the estimation of the direction of dependence (previous work had to assume a specific direction).  The disadvantage is that estimation via conditional logit models is very expensive, to the order of several days for moderate size samples of a few thousand cases.