Biomedical Engineering Reference
In-Depth Information
when data abnormalities exist. Based on our limited Monte Carlo simulation
studies, we nd that interval-censored methods (e.g., Finkelstein's method and
Sun's generalized logrank test) outperform conventional approaches under all
scenarios considered, with almost unbiased point estimations and correct Type
I error and power rates. This provides a robust approach to evaluate interval-
censored data such as PFS, especially when various situations arise in PFS
assessment.
Our simulation studies also demonstrate the limitations of conventional
approaches. In particular, we conclude that the relative information retained
in the observed interval-censored data from exact times is the most crucial fac-
tor that influences the results of point estimates and hypothesis testing. The
information retained after various interval-censored sampling schemes (sce-
narios) primarily depends on two types of sources: (i) the overall assessment
frequency and number of events, and (ii) the sampling symmetry between
treatment arms.
The former source of information is essentially the problem of interval-
censored data, which results in a slower convergence rate than right-censored
data and many other related results in interval-censored literatures (Sun
(2006)). In our case, given fixed sample sizes, if both assessment frequency and
event proportion are low, biased point estimation and incorrect Type I error
and power rates are observed. In other words, when designing clinical trials
with interval-censored data such as PFS as the primary or secondary end-
point, the required number of PFS events based on conventional approaches
with right-censored survival data assumption should be critically reviewed,
and the assessment frequency (ratio between median PFS time and assess-
ment interval) should be jointly considered. Otherwise, the study may not
achieve the desired power at the required Type I error level. Furthermore, our
observation on balanced assessments with per-protocol compliance continues
to hold when only random deviations or missingness of scheduled assessment
occur (scenarios II and III). The randomness of data abnormalities is crucial
Search WWH ::
Custom Search