Unit:
Κατεύθυνση ΒιοστατιστικήLibrary of the School of Health Sciences
Supervisors info:
Παναγιώτα Τουλούμη Αναπλ. Καθηγήτρια, Φώτης Σιάννης Επικ. Καθηγητής, Νίκος Πανταζής Βιοστατιστικός
Original Title:
Ελλείπουσες τιμές λόγω περικοπής μετρήσεων σε διαχρονικά συνεχή δεδομένα: χρήση μοντέλων για πληροφοριακή περικοπή όταν η αποκοπή είναι αγνοήσιμη.
Translated title:
Missing values due to dropout in longitudinal data: applications of joint models for ignorable missing data
Summary:
The analysis of longitudinal data is more complicated when there are missing
data. It is a well known fact that in most cases the repeated measurements are
truncated due to a clinical event. When studying the CD4 count evolution during
HIV natural history, the series of this marker’s measurements can be naturally
censored due to death or lost to follow up and artificially censored due to
clinical AIDS onset or antiretroviral treatment (ART) initiation. In the
combined ART (cART) era (i.e. post 1996), the most prevalent censoring
mechanism is cART initiation. If the censoring mechanism is informative then
models that jointly model the marker’s evolution and the drop-out mechanism
such as the Joint Multivariate Random Effects (JMRE) model, will give unbiased
estimations. On the other hand, if the censoring mechanism is non-informative
then models which are based in maximum likelihood estimations lead to unbiased
estimations. The aim of the present study is to investigate the performance of
linear mixed models and JMRE models when analyzing such data under various
plausible scenarios regarding cART initiation and timing.
This study focuses on the analysis of CD4 cell count, which is an important
marker of the HIV natural history. The decision for cART initiation is based on
observed CD4 cell count and the censoring mechanism is considered as ignorable.
We used real data from Athens Multicenter AIDS Cohort Study – AMACS. We applied
a linear mixed model and a JMRE model in the real data. However, the
performance of the above methods is based on simulated data drown from the
European CASCADE study. Four distinct scenarios which illustrate the censoring
mechanism due to cART initiation have been explored.
The missingness process plays a vital role on determination of the ideal method
to be applied. The random effect model always results in unbiased estimates in
presence of completely at random (MCAR) and missing at random (MAR) data, while
in presence of non-ignorable missingness it could result in biased estimates.
In case of right informative censoring, the application of a JMRE model that
minimizes bias is highly recommended.
Keywords:
Joint models, Missing data, Missing at random (MAR), Mixed effect models, Bias
Number of references:
110
