Biomedical Engineering Reference
In-Depth Information
method, although the regularity conditions of its asymptotic properties are
dicult to justify, we observe reasonable correspondences between the asymp-
totic standard deviations and empirical standard errors under all scenarios,
which indicates its good asymptotic property in finite sample sizes. The 95%
coverage probability of both conventional approaches with Efron's method of
tie-handling and Finkelstein's method reasonably stay at the nominal level
when compared with benchmarks in most cases, unless the point estimation
is severely biased. The root mean square errors of Finkelstein's method esti-
mators consistently remain at similar levels across all scenarios, while their
counterparts from conventional methods are highly inflated if the estimators
are biased, as expected.
Compared with the simulation results in Table 10.2, Table 10.3 summa-
rizes the results with the same underlying distribution of progression times
and sampling scenarios, except for a lower event rate (60% versus 80%). In
general, a lower event rate would lead to less precise estimation and more un-
certain statistical inference. Not surprisingly, we find almost unbiased results
based on Finkelstein's method under all scenarios. Meanwhile, in addition to
similar findings we obtained from Table 10.2 for conventional approaches, we
observe more severe bias under the same scenarios than in their counterparts.
When comparing the results from conventional methods between Table 10.3
(60% event) and Table 10.2 (80% event), we find that the estimation biases
are comparable under per-protocol compliance (Scenario I), while the biases
under other scenarios (Scenarios II to V) are more severe. When only 60% of
events are observed, we find more pronounced bias in Scenarios II and III than
in scenario I, which can be viewed as the extra bias introduced due to devia-
tions or missing scheduled assessments. Because fewer events are observed and
less information is retained, the correspondence between asymptotic standard
deviations and empirical standard errors when 60% events are observed are
larger than their counterparts when the event rate is 80%. The 95% coverage
probability also remains at nominal levels.
Search WWH ::
Custom Search