Bias Assessment in Progression-Free Survival Analysis Using Interval-Censored Methods - Interval-Censored-Time-to-Event Data: Methods and Applications - page 285

In general, the results show that point estimates based on Finkelstein's
method are almost unbiased under all scenarios considered. When assessments
are strictly conducted on schedule (scenario I), the point estimates based on
right-point imputation (and mid-point imputation) are always slightly nega-
tively biased (toward null) and thus underestimate the treatment effect. Fur-
thermore, the bias becomes more pronounced when a longer assessment inter-
val is scheduled. Similar observations are seen when random deviations from
scheduled assessments (scenario II) or missed scheduled assessments (scenario
III) exist. The mid-point imputation shows less negatively biased point esti-
mates than right-point imputation. The biases under scenario II and III are
comparable to those under scenario I, which suggest that such random proto-
col noncompliance may not introduce additional bias in this setting. However,
in scenario IV, when early determination of progression exists in the experi-
mental arm, we find significant negative bias (up to 30%) that underestimates
the true treatment effect using conventional approaches. This negative bias
in conventional methods is expected, as shorter \true" progression time in the
T arm is observed than it should be compared with those in the C arm, and
correspondingly underestimates the treatment effect (negative bias). We ob-
serve analogous findings when early determination of progression occurs in the
control arm (scenario V), where observing a shorter \true" progression time
in the C arm than in the T arm leads to overestimating the treatment effect
(positive bias up to 24%). It is noted that in both scenarios IV and V, the
right-point and mid-point imputations also gives the same results based on
the Cox regression model because their ranks are retained, and therefore only
right-point imputation results are reported.
In all scenarios, the standard deviations from the conventional method are
comparable to those from the Cox model with exact times, while those from
Finkelstein's method have slightly larger standard deviations. This nding
is reasonable because conventional methods, by assuming exact event times,
overstate the information retained in the observed data. For Finkelstein's