Knot selection strategies for semiparametric varying coefficient models applied to longitudinal cohorts with multiple dropout reasons

Abstract

Dropout is a common source of missing data in longitudinal studies, and often occurs for reasons that may be related to an outcome of interest. When dropout depends on unobserved outcomes, even after conditioning on observable data, data are potentially missing not at random and dropout is therefore not ignorable. When the dropout mechanism is unspecified, semiparametric varying coefficient models can be used to account for non-ignorable dropout. This method is more robust than the parametric conditional linear approach. However, fitting these varying coefficient models requires the specification of the number and location of spline knots, which may influence model fit. We present simulation results comparing the natural cubic B spline varying coefficient method with a knot location selection algorithm to the same method with knots evenly placed at the quantiles of the dropout distribution, as well as to the classic conditional linear model. In addition, semiparametric varying coefficient models accounting for dropout time are extended to account for dropout reason as well. These methods are applied to data from the Acute Infection and Early Disease Research Program (AIEDRP) to determine the effect of injection drug use on the longitudinal trajectory of CD4+ T cell count in untreated HIV seropositive subjects.