Industry Issues: Dataset Preparation for Time to Event Analysis

Download Report

Transcript Industry Issues: Dataset Preparation for Time to Event Analysis

Industry Issues:
Dataset Preparation for Time to
Event Analysis
Davis Gates
Schering Plough Research Institute
Not all Clinical Trials are
Designed for Time to Event
Analysis
However, time to events are common
secondary analyses in many clinical trials,
later requiring PRIMARY attention. These
comments address those types of studies.
Obstacles to optimal Time-to-event
Analysis
 Secondary Parameter: study not powered for this
analysis
 Low Incidence of Events: resulting in poor or
impossible estimate of time-to-event by treatment
 Determination of Event: During vs. after the
treatment period
 Censoring Issues: Rate very high or poorly defined
 Bias due to Per Patient Duration of Treatment:
another term meaning early drop outs
Secondary Parameter
 Study may be powered for a defined delta of
a continuous outcome variable.
 Time-to-event among a long list of
secondary analyses.
 Cannot rely on statistical inference for final
conclusions (PROC LIFETEST), therefore, a
clinically meaningful difference should be
assessed in lieu of a p-value.
Low Incidence of Events
 Time to event analysis may not be feasible
for a population of patients who experience
a low proportion of events.
 How many times have we seen non
estimable MEDIAN time to events?
 May have to break down to a categorical
analysis (EVENT=YES or NO), losing the
time to event element. (We should have
enrolled more sickly subjects)
Determination of Event
 Even though the event itself may be clearly
defined, the probability of event vs. relation to
treatment may not be so clear. (are all events
treated equally?)
 The evaluation of event can differ in two
contrasting treatment methods such as daily
topical vs. monthly intravenous.
 Generally analyze ALL data, regardless of
temporal relationship with treatment, but why not
define the drug relation interval, and use this to
enforce follow-up?
Simple Case – Daily Dosing
Regimen
Daily/periodic Inhaled/Topical Drug with minimal
/no Blood Level Impact:
Treatment Period:
Last Day of Dose + Delta
Delta = period of time no shorter than
treatment duration claim, but can include a
wash-out period
Treatment Period for Drugs with Long
Half-Life/Infrequent Dosing Regimen
Long-acting Drug such as a monthly IV or single dose with long Blood
level half life…
Treatment Period can last through:
Last Dose Date +N*t1/2
N = number of half lives (t1/2), each of which can be of several days or
weeks in length (N=5 generally determines drug wash-out)
This criteria can require a considerable amount of follow-up.
Censoring Issues
General Time-to-event methodology, even though censors
are handled, were not optimally designed for data with
heavy censoring.
Large scale clinical trials can have a high proportion of
censored subjects (20% or more).
This places an emphasis in improvements in follow-up for
all subjects, even those discontinued from the trial.
This is related to the issue of low proportion of events,
which leave the potential for high proportion of censors.
Patient Duration of Treatment
 Is time-to-event (or survival) analysis the
ultimate missing data analysis? Certainly
not, subjects drop out for many reasons,
related or not related to treatment or
covariates.
 Calculation of event rates can be biased by
early drop-outs.
Duration of Treatment (Cont…)
 Early Drop Outs: it is in our interest to follow
up subjects, give them a chance to report an
event – even if discontinued from a trial, to
reduce this bias.
 In many cases the probability of a subject
reporting an event is reduced with an early
drop out, producing by treatment event rates
biased downward.
The Perfect Dataset
 First Day of Treatment Documented
 Each Event Date Recorded with pre or post dose
information for that day
 Last Day of Treatment Documented
 Last Visit Day Documented
 All Subjects receive follow-up regardless of treatment or
discontinuation status (Insure open communication
opportunity – at least remotely).
 Last contact date Documented
 Last Day of follow-up clearly documented (last contact date
is optimal, but may have to settle for last visit date)
 Censoring and Events to be determined by availability of
above data and a pre-specified algorithm.