Biomedical Engineering Reference
In-Depth Information
The simulation results when median PFS time in C increases from 12 to 24
weeks, with 80% event rate, and under same assessment intervals, are summa-
rized in Table 10.4. Results based on Finkelstein's method continue to show
a lack of bias across different scenarios. With the same assessment intervals
but longer median PFS, the conventional approaches provide almost unbi-
ased point estimation under per-protocol compliance (Scenario I) and random
protocol noncompliance (Scenarios II and III). Even under scenarios of sys-
tematically different evaluations (Scenarios IV and V), we observe much less
severely biased results when compared with those of shorter progression time.
The good statistical inference results in nite sample sizes from Finkelstein's
method remain. Similar results also hold when event rate decreases to 60%
(Table 10.5).
10.3.3
Results on Hypothesis Testing
Tables 10.6 and 10.7 summarize the empirical Type I error rate at = 5%
(two-sided) based on 12 weeks median PFS in C, and 80% or 60% event rate,
respectively. In general, interval-censored methods, for example, Finkelstein's
score test and Sun's generalized logrank test, perform reasonably well across
all scenarios when compared to the logrank tests of exact time, with slightly
inflated Type I errors when assessment frequency decreases (12 weeks assess-
ment interval) or event rate is low (60% event rate). Between Finkelstein's
score test and Sun's generalized logrank test, it seems that Sun's test con-
trols Type I error rate a little better than Finkelstein's score test, although it
is minimal. Under per-protocol compliance (scenario I), we find comparable
Type I error rates from both conventional approaches and interval-censored
methods, with slight inflation when the event rate is lower (60% event rate).
Such results from conventional approaches hold and are still around nominal
levels under Scenarios II and III. However, consistent with results in Tables
10.2 and 10.3, under scenarios IV and V, Type I errors are greatly inflated
with conventional approaches (up to 62%). In other words, when unbalanced
 
Search WWH ::




Custom Search