cohort study design and estimating drug exposures
TRANSCRIPT
11/10/2021
1
Cohort Study Design and Estimating Drug Exposures
Presenter: Dr Andrea Schaffer, Centre for Big Data Research
in Health, University of New South Wales, Australia
13th Asian Conference on Pharmacoepidemiology
October 13-15, 2021
2
Disclosures
I am supported by the following:
• National Health and Medical Research Council (NHMRC) Centre of
Research Excellence in Medicines Intelligence (#1196900)
• NHMRC Early Career Fellowship (#1158763)
The Centre for Big Data Research in Health, UNSW received funding
from AbbVie Australia in 2020 to conduct research unrelated to this
current work
3
Acknowledgements
Special thanks to:
• Dr Sallie-Anne Pearson, UNSW Sydney
• Dr Vin Re Lo, University of Pennsylvania
• Dr Wei Zhou, MSD & Co
• Drs Kristian Filion and Laurent Azoulay, McGill University
• Dr Tobias Gerhard, Rutgers University
• Dr Almut G Winterstein, University of Florida
1
2
3
11/10/2021
2
4
Educational objectives
Understand the selection of drug exposure and
comparators within pharmacoepidemiology databases
Understand epidemiologic study designs
Understand the strengths and limitations of cohort
studies for pharmacoepidemiology research2
3
1
5
Epidemiology study designs
Adapted from Centre for Evidence Based Medicine: http://www.cebm.net
Treatment/control groups similar except treatment
assignment
How many people are taking drug X?
Is drug X associated with Y?
Does drug X cause Z?
6
What is a cohort?
A well-defined group of people (who, where, when)
• Classified based on presence and absence of an exposure
Individuals in the cohort are followed forward through time and
observed for occurrence of outcome(s) of interest
What question does a cohort study answer?
Among the two subgroups defined by exposure
(exposed/unexposed), how does the incidence of a particular
outcome differ?
4
5
6
11/10/2021
3
7
Cohort studies
8
Cohort
Follow-up
Exposure
Outcome
Kim et al. Pharmacoepi Drug Saf 2019;28:507
9
Checklist for simple cohort study
1. Define population(s) of interest and comparator
2. Define inclusion/exclusion criteria
3. Define how people will be sampled from the underlying source population
4. Define explicitly (!!) date of cohort entry and exit
5. Classify person time by exposure
6. Sum person-time of the different exposure categories
7. Assign number of events to these categories
8. Calculate incidence and incident rates
7
8
9
11/10/2021
4
10
Checklist for simple cohort study
1. Define population(s) of interest and comparator
2. Define inclusion/exclusion criteria
3. Define how people will be sampled from the underlying source population
4. Define explicitly (!!) date of cohort entry and exit
5. Classify person time by exposure
6. Sum person-time of the different exposure categories
7. Assign number of events to these categories
8. Calculate incidence and incident rates
11
Incident versus prevalent users
Time
12
New user design
Identify and selects only drug initiators (incident users)
Follow-up begins at initiation of therapy (time zero = t0)
Restrict to people with minimum period of non-use prior to t0 (washout
period)
Baseline characteristics obtained in specified period before initiation (t0)
This accounts for depletion of susceptibles:
• Risk of outcome is often highest shortly after initiation—can lead to discontinuation
• Prevalent users are more likely to be people who did not experience the outcome
• Can make harmful exposures appear to be protective
10
11
12
11/10/2021
5
13 Renoux et al. Pharmacoepi Drug Saf 2017;26:554
14
Selecting appropriate comparator
Consider the study question—needs to be fit-for-purpose and clinically
meaningful
What about “non-users” as comparator?
• Non-users often differ substantially to users (confounding by indication)
• Difficult to define date of cohort entry
Inactive comparator (different indication)
• Ensures all patients are interacting with health care system
• Easier to define date of cohort entry
Active comparator (same indication)
• Patients are more similar (indication for use, comorbid disease)
• Reduces confounding
15Setoguchi et al. Circulation 2007;115:27. Dong et al. Br J Clin Phamacol 2018;84:1045..
Inactive
comparator
Active
comparator
13
14
15
11/10/2021
6
16
Checklist for simple cohort study
1. Define population(s) of interest and comparator
2. Define inclusion/exclusion criteria
3. Define how people will be sampled from the underlying source population
4. Define explicitly (!!) date of cohort entry and exit
5. Classify person time by exposure
6. Sum person-time of the different exposure categories
7. Assign number of events to these categories
8. Calculate incidence and incident rates
17
Cohort entry/exit
When to start follow-up (t0)
• Calendar day
• Time (x days after inclusion in registry)
• Event (date of diabetes diagnosis, first dispensing of drug)
When to stop follow-up (censoring)
• First occurrence of outcome
• Loss to follow-up or death
• Switch in drug exposure
• End of study period (specific date) / end of data capture
18 Schneeweiss et al. Ann Intern Med 2019;170(6):398.
16
17
18
11/10/2021
7
19 Ou et al. Br J Clin Pharmacol 2021
Research
question
To compare the short-term risk of cardiovascular events associated with the use
of tramadol (exposed) compared with codeine (active comparator)
CohortPeople without cancer (≥18 years) who initiated either tramadol or codeine with a
12-month washout period (new user design)
Entry (t0) Date of initiation after ≥12 months of continuous history in the CPRD
Censoring
Event (hospitalisation/death due to myocardial infarction), death, end of CPRD
registration, cancer diagnosis, end of study period (31 March 2017), maximum
follow-up (30 days)
AnalysisTime to event (Cox proportional hazards) adjusted for high dimensional
propensity score
*CPRD = Clinical Practice Research Datalink
20
Checklist for simple cohort study
1. Define population(s) of interest and comparator
2. Define inclusion/exclusion criteria
3. Define how people will be sampled from the underlying source population
4. Define explicitly (!!) date of cohort entry and exit
5. Classify person time by exposure
6. Sum person-time of the different exposure categories
7. Assign number of events to these categories
8. Calculate incidence and incident rates
21
Time-dependent classification of exposure
Exposed (A=1) Unexposed (A=0)
19
20
21
11/10/2021
8
22
Conceptual considerations for drug exposure measurement
Link exposure measurement to the study question:
• Short- vs long-term use, single vs chronic use, prevalent vs new use
• Biological mechanism (pharmacology) of the drug
Need to account for changes in exposure status
Induction period: time required for the treatment to biologically affect the
outcome
Latency period: time required between outcome onset and diagnosis
**It can be difficult to differentiate between the induction and latency periods and
they are often combined
23 Billioti de Gage et al. BMJ 2012;345:e6231.
24
Drug exposure definitions
Lee and Pickard. Chapter 4: Exposure definition and measurement. In:
Developing a Protocol for Observational Comparative Effectiveness Research: A User's Guide. 2013
1
23
22
23
24
11/10/2021
9
25
Exposed
Exposed?
Last
dispensing/
prescription
Extending drug exposure period?
26
May be appropriate if:
• A drug effect is known to continue after stopping use
• Drug supply is likely not exhausted
• Drug was stopped because of early symptoms related to the outcome
What if I get it wrong (misclassification)?
• Erroneous attribution of an event to non-exposure (if patient was
actually exposed) will increase incidence in non-users and bias the
comparison towards no association
Extending drug exposure period?
27 Gamble et al. Value in Health 2012;15:191.
25
26
27
11/10/2021
10
28
What if I am unsure?
Define exposure classification from the outset based on sound
reasoning (clinical and empirical)
Don’t forget the pharmacology in pharmacoepidemiology
Undertake sensitivity analyses
To account for
residual drug effects
To account for informative
discontinuation or
continued use
To account for
induction period and
reverse causation
Initial Current Recent Past
29 Margulis et al. Eur J Clin Pharmacol 2018;74:193.
30
Checklist for simple cohort study
1. Define population(s) of interest and comparator
2. Define inclusion/exclusion criteria
3. Define how people will be sampled from the underlying source population
4. Define explicitly (!!) date of cohort entry and exit
5. Classify person time by exposure
6. Sum person-time of the different exposure categories
7. Assign number of events to these categories
8. Calculate incidence and incident rates
28
29
30
11/10/2021
11
31
Measures of frequency and risk
1. Incidence proportion (cumulative incidence): number of
new cases in the population over a defined period of time
• Denominator = Total population at risk
2. Incidence density (incidence rate): number of new cases
per unit of time
• Denominator = Sum of the time that each individual is followed
until the event, death, or lost to follow-up (person-time)
32
Incidence measurement
PY = person-years
33
Comparing incidence – relative measures
Incidence risk ratio (“relative risk”) (RR)
RR =Incidence proportion in exposed
Incidence proportion in unexposed
Incidence rate ratio (IRR)
IRR =Incidence density in exposed
Incidence density in unexposed
Hazard ratio (HR) (for time-to-event analyses)
HR =Hazard in exposed
Hazard in unexposed
31
32
33
11/10/2021
12
34
Comparing incidence – absolute measures
Absolute risk difference: difference in two incidence proportions
or rates
Risk difference RD = Incidence in exposed − Incidence in unexposed
35
Statistical analysis
Start with descriptive/unadjusted comparisons
Need to adjust for confounding (e.g. multivariable regression,
propensity scores, IPTW)
Can use intent-to-treat analysis, or incorporate time-varying
exposure/confounding
Adjusted analysis: use regression modelling (logistic, linear,
Poisson)• For time-to-event analyses: Cox proportional hazards regression
36
No. events No. people
No. person-
years
Incidence rate
(95% CI)
Adjusted HR
(95% CI)
Codeine 646 913,966 74,544 8.7 (8.0-9.4) 1.00 (Ref)
Tramadol 106 123,390 10,051 10.5 (8.7-12.8) 1.00 (0.81-1.24)
Table. Association between tramadol vs codeine and the risk of myocardial infarction
Ou et al. Br J Clin Pharmacol 2021
34
35
36
11/10/2021
13
37
Cohort study design
Advantages Disadvantages
Timing is clear (i.e. exposure
precedes outcome)
Large sample size needed for rare
outcomes
Can calculate incidence Long follow-up may be required for
outcomes with long latency
Can study multiple outcomes Loss to follow-up can introduce bias
More intuitive Can be costly
38
Any questions?
37
38