Κατεύθυνση ΒιοστατιστικήLibrary of the School of Health Sciences
Γιαννούτσος Κωνσταντίνος (Καθηγητής), Τουλούμη Παναγιώτα (Αν. Καθηγήτρια), Πανταζής Νικόλαος (ΕΔΙΠ, Βιοστατιστικός PhD)
Inverse probability of censoring weights under missing not at random with application to patient adherence in HIV-Positive patients in Kenya.
Στάθμιση μέσω αντίστροφης πιθανότητας λογοκρισίας μη τυχαίως ελλιπών δεδομένων με εφαρμογή στην συμμόρφωση στη θεραπεία οροθετικών ασθενών με HIV στην Κένυα.
Routine estimation of the rates of drug adherence in a population are based on
patients who continue to be engaged in care (i.e., continue to come to clinic
visits). It is well understood that lower levels of adherence in a population
is associated with differential rates of dropout (disengagement from care) as
well as adverse clinical outcomes. This is particularly the case among
patients infected with HIV. This project will use a technique called inverse
probability of censoring weights (IPCW) to account for the fact that patients
who are less adherent and, consequently, may be sicker than those who are more
adherent, will experience higher rates of disengagement from care. By weighing
patients who remain in care by the inverse probability of remain uncensored,
patients who have lower probability of remaining in care (uncensored) are
weighted higher and thus represent more patients who dropout. Conversely,
patients who have higher probability of remaining in care represent less
patients who dropout. We hope that this approach will provide a (downward)
adjustment of the usually optimistic estimates of patient adherence estimated
by using data on only those patients who remain in care.
Extensions: The entire approach presented earlier assumes that patients
remaining in care (i.e., those who remain uncensored) are, conditional on their
covariate and treatment history, representative of those who dropout. This is
called the missing at random (MAR) assumption. If this is not the case, and the
patients who dropout are at higher risk for adverse outcome (and non-adherence)
at rates which cannot be predicted by those still in care, we have what is
called missing not at random (MNAR). Usually this situation is untestable
since there are no data on these patients after disengagement. Fortunately,
data are available on a (hopefully random) sample of patients disengaging from
care. We propose to use these data on the sub-sample of the dropouts (vital
status, treatment access in other facilities) to further refine the adjustments
to the estimates of adherence. This work, involving a binary variable (i.e.,
perfect versus non-perfect adherence) will follow recent work by Dr.
Yiannoutsos and his colleagues at Indiana and Harvard Universities involving
continuous longitudinal markers such as CD4 count.
Human immunodeficiency virus (HIV), Causal inference, Inverse probability of censoring weights, Μissing not at random (MNAR), Adherence to treatment
File access is restricted.