Biomedical Engineering Reference
In-Depth Information
terms of evaluating the appropriateness of asymptotic results in the current
investigation, due to parallels with respect to expected number of events
and intra-subject event time correlations. For this data conguration, no
evidence of inappropriateness is revealed for the large sample approxima-
tions employed.
A previous analysis of this birth cohort
21
treated hospitalization for
asthma during age 0-4 as a binary variable and employed logistic regres-
sion. Under the assumption that most children who experience asthma very
early in life tend to suer it the rest of their lives, not incorporating infor-
mation on the number of episodes per child or their timing was deemed
appropriate at the time given the goals of the study. Our objective in the
current investigation was to assess the eect of birth characteristics on the
number of asthma-attributable hospitalizations and days hospitalized. It is
reasonable to assume that both quantities reect disease severity as well as
health care costs. Hence, while the original study served strictly an etiologic
purpose, the objectives of the current analysis were also pertinent from a
public health perspective.
Often in biomedical studies when patients are followed either prospec-
tively or retrospectively over time, multiple occurrences of the event of
interest are possible. Patients may leave the study before its conclusion, or
may complete the study without experiencing the event of interest. Meth-
ods are well-established to analyze the potentially censored time until rst
event by survival analysis. A more informative analysis would incorporate
all event times, requiring a multivariate survival analysis, for which meth-
ods are now fairly well known (Prentice et. al.
20
; Andersen and Gill
2
; Wei
et. al.
25
; Lin et. al.
15
). Most multivariate failure time methods in-
volve modelling the hazard function. However, in the context of recurrent
event data, the mean or rate is often of direct interest, and is a more inter-
pretable quantity. Compared to the marginal hazards model, an advantage
of the marginal means/rates model is that estimates of the mean number of
events are directly obtainable from model parameters. Both models could
be used to generally assess the eect of covariates on the event process of
interest; both have appeal from an etiologic perspective. However, among
clinicians, public policy ocials and health administrators, direct interest
often lies in the mean number of events and, in such cases, the marginal
means/rates model would be preferred. A natural area of application for the
marginal means/rates model is health economics. For example, if cost data
are available for the event of interest, the mean increase in costs associated
with various attributes could be estimated and compared with the costs
Search WWH ::
Custom Search