Download Effectiveness Evaluation - DEP

Design and analysis strategies to
assess the effectiveness of
long-term drug therapies
Mirko Di Martino
Rome, October 15th - 16th, 2012
The randomized controlled trial design
The randomized controlled trial (RCT) design is essential
to evaluate the efficacy of medications and to obtain
regulatory approval for their use in clinical practice.
1) Yet, it rarely provides information on their pragmatic
benefit in routine care.
2) In some cases, very important RCTs have not allowed
to fill certain gray areas.
The TORCH study is an emblematic example. This RCT was one of the
largest and longest trials of pharmacotherapy in patients with chronic
obstructive pulmonary disease (COPD). However, no definitive
conclusions can yet be drawn on whether inhaled corticosteroids (ICS)
has an effect on mortality.
The explosion in observational studies
A large number of observational studies have
been conducted in an attempt to fill this gap by
assessing long-term effects of medications on
infrequent outcomes. However, observational
studies can be misleading.
Are observational studies reliable?
In 2008 Samy Suissa identified 20 observational
studies on the effects of commonly prescribed drugs
that were subject to immortal time bias.
At the end of the nineties, findings from non
randomized studies influenced millions of women to
use hormone replacement therapy for the presumed
cardiac benefits. Healthy drug user effect and
chronology bias (prevalent users) have greatly
contributed to this illusion.
Objectives
Our goal is to critically explore the potential and limits of
non-experimental research, comparing different design
and analysis strategies.
The methods will be developed within the framework of
the OUTPUL study to assess the benefits of inhaled
corticosteroids in reducing mortality among Chronic
Obstructive Pulmonary Disease (COPD) patients.
The study population
We enrolled more than 21,000 patients, aged 45 or
more, resident in two Italian regions (Lazio and
Emilia-Romagna; data of Lombardia will be available
soon), discharged from hospital with diagnosis of
COPD between January 1st 2006 and December 31st
2009 and prescribed with respiratory drugs no later
than six months after discharge.
The study design: the “two step” enrollment
Six months after discharge
Time
First COPD hospital discharge
during the enrollment period
Index prescription for respiratory drugs.
The beginning of follow-up.
We considered an equal follow-up time per subject of one year.
Drug exposure and outcome
We dynamically evaluated prescriptions of long-acting beta-agonists
(LABA), tiotropium, inhaled corticosteroids (ICS) and other
respiratory drugs. Prescription patterns were classified as follows.
• ICS monotherapy
• Long acting bronchodilators without ICS (LABA and/or tiotropium)
• Long acting bronchodilators plus ICS
• Other respiratory drugs (No ICS, no LABA, no tiotropium)
The outcome was mortality from all causes.
Patient characteristics
Patients were characterized according to sociodemographic factors (age, gender, educational level and
area of residence), COPD severity (hospitalizations for
COPD, diagnosis of respiratory failure, invasive respiratory
procedures, staying in intensive care unit during a COPD
hospitalization, emergency visits for COPD and use of
oxygen), concomitant respiratory diseases (asthma, chronic
respiratory disease other than COPD, pulmonary infections
and acute pulmonary symptoms), previous use of
respiratory drugs, previous use of oral corticosteroids and
antibacterials, previous use of non-respiratory drugs and
comorbidities (more than 20 conditions such as diabetes,
hypertension or heart failure).
Estimation of the treatment effectiveness
Patients may use drugs very irregularly, medication patterns
may contain a plethora of different drugs, doses and switching
between treatments.
Several solutions are under evaluation:
•
•
•
•
•
•
cohort study using the censoring at switching approach;
longitudinal study using time-varying determinants;
longitudinal study using marginal structural models;
incidence density nested case-control design;
self-controlled designs;
new-user designs.
The censoring at switching approach
We censored patients at switching or discontinuation and kept the
person-time follow-up until the time of censoring.
Prescription Coverage
ICS
ICS
Grace
Time (7 days)
Prescription Coverage
ICS
ICS
LABA
CENSORING
AT
DISCONTINUATION
Grace
Time (7 days)
CENSORING
AT
SWITCHING
Prescription coverages were estimated on the basis of the
Defined Daily Doses.
ICS
The trade-off between “power” and validity
Intention to treat
analysis
Censoring at
switching
Censoring at
switching and new
user design*
Number of Patients
21,276
21,276
4,284
Number of Events
2,603
411
65
Person-Years
19,932
4,226
803
* The new user design requires the exclusion of patients with any filled
prescription of respiratory drugs in the six months prior to the index prescription.
The use of respiratory drugs in clinical practice is characterized by
frequent changes in therapy and discontinuation. Therefore, the
censoring at switching approach largely reduces the amount of events
and person-years. The new user design further reduces the study
population.
Censoring at switching and new user design:
preliminary results (1)
Number of
Patients
Number of
Events
Personyears
ICS monotherapy
831
9
113
Long acting bronchodilators
(LABA and/or tiotropium, without ICS)
668
12
150
Long acting bronchodilators plus ICS
1621
29
367
Other respiratory drugs
(No ICS, no LABA, no tiotropium)
1164
15
173
Censoring at switching and new user design:
preliminary results (2)
Crude Mortality
Rate (per 1,000 p-y)
Censoring Rate
(per 100 p-y)
ICS monotherapy
79.89
711.12
Long acting bronchodilators
(LABA and/or tiotropium, without ICS)
79.89
402.82
Long acting bronchodilators plus ICS
78.89
408.34
Other respiratory drugs
(No ICS, no LABA, no tiotropium)
86.44
640.85
Censoring at switching: key points
If preceded by an appropriate washout period, the censoring at
switching design is free from carryover effect, and should provide
an “independent” estimate of the relationship between treatment
and outcome.
However, it is necessary a careful evaluation of the balance
between the “validity” of the study and the loss in follow-up time.
The length of the grace time plays a central role.
Moreover, there is a concern whether an excessive amount of
these censorings carries information for patients' prognosis
(informative censoring).
Multiple propensity score could improve the control of confounding.
Critical aspects of the other methods
Longitudinal study using time-varying determinants. Such estimates may be
biased in the presence of time-dependent confounders which are
themselves affected by prior exposure.
Longitudinal study using Marginal Structural Models. This method can
adjust for time-varying confounders affected by prior exposure. However,
will be necessary to verify a) the plausibility of the assumption of no
unmeasured confounding and b) the applicability of such a complex model
in patients whose medication patterns are characterized by such a high
switching rate.
Self-controlled design. These designs are generally used to study short
transient exposures and acute effects. Therefore, there are concerns
whether self-matched designs may be applied to study long-term effects of
chronic treatments.
Future prospects
We think that a critical comparison between these
methods will contribute to identify optimal strategies to
assess the effectiveness of long-term therapies and will
provide important elements on the real potential of
observational studies in clinical research.
Be careful about reading health books.
You may die of a misprint.
Mark Twain
Mirko Di Martino
Department of Epidemiology, Lazio
Regional Health Service.
Via di Santa Costanza, 53 00198
Roma, Italy.
E-mail: [email protected]