Statistical Modelling 14 (6) (2014), 549–566

Unobserved confounder effects in models for clustered dental failure time data.

Frank Eriksson
Department of Biostatistics,
University of Copenhagen,
Denmark
e-mail: ank.eriksson@sund.ku.dk


Thomas Alexander Gerds
Department of Biostatistics,
University of Copenhagen,
Denmark


Emmanuel Lesaffre
Department of Biostatistics,
Erasmus MC,
Rotterdam, the Netherlands and
L-Biostat School of Public Health,
University of Leuven,
Belgium


Abstract:

The marginal approach and the conditional approach are two different ways to model clustered dental failure time data. We compare the two approaches in the context of a Cox regression analysis, where the aim is to estimate the effect of a covariate (e.g., dental treatment) on the risk of failure. Specifically, we treat within-cluster correlation as if it was introduced by unobserved cluster level covariates, and study the small sample behaviour of the marginal and the conditional approach. We show that in a non-randomized setting where an unobserved cluster variable is correlated with the variable of interest, both the marginal and the conditional approaches can give misleading results. We argue that this is an important message, since most often it is assumed that the frailty term and the covariates of interest are independent.

Keywords:

clustered data; marginal model; frailty; attenuation bias; variable selection;
back