a primer to methodologic issues in...
TRANSCRIPT
A primer to methodologic issues
in Pharmacoepidemiology
Reimar W. Thomsen, overlæge, lektor, ph.d.
Klinisk Epidemiologisk Afdeling, Aarhus Universitetshospital
”Farmakoepidemiologi er epidemiologiens
Saudiarabien” [Jørn Olsen]
Pharmacoepi in Denmark…
A primer to methodologic issues in pharmacoepi -
OUTLINE
1. Study designs & methods -
how to measure drug exposure in real life
2. Common types of epidemiologic bias in analytic pharmacoepi
– illustrated by some examples
– including a little bit on ”newer” analytic approaches in
pharmacoepi – propensity scores and instrumental variables
(For in-depth coverage, attend a 1-week-course,
e.g. AU course by Til Stürmer et al.!)
1. Study designs and how to measure drug exposure
Pharmacoepidemiology
• The study of distribution and determinants of drug-related events in
populations, and application of this study to efficacious drug treatment
– Last: A Dictionary of Epidemiology. Oxford University Press. New
York 1988
• The study of the use and effects of drugs in large numbers of people
– Strom: Textbook of Pharmacoepidemiology, 2006
Pharmacoepidemiology
• Descriptive methods
– Patterns of- / Determinants of- drug utilization
• Analytic methods
– Intended effects (benefit)
– Unintended effects (not only harm!)
– Comparative Effectiveness Research
Why do we need observational designs in
pharmacoepidemiology?
5 Shortcomings of RCTs (T. Stürmer)
• Too Small
– to detect rare outcomes
• Too Simple
– to detect interactions
• Too Selected
– to be generalizable to all users and all indications
• Too Specific
– to assess all relevant outcomes
• Too Short
– to detect long-term effects
11
Examples for Incidence of Adverse Drug Reaction
Drug Event Incidence
Chinidine Syncopy 1 / 100
Clozapine Agranulocytosis 1 / 1,250
Enalapril Angioedema 1 / 3,000
Lovastatin Rhabdomyolysis 1 / 3,000
Dextrane Anaphylactoid reaction 1 / 4,000
Clopidogrel Agranulocytosis 1 / 5,000
Halothane Liver cell necrosis 1 / 30,000
Choramphenicole Aplastic anemia 1 / 40,000
Cyclosporine A Malignancy ? 12
Nonexperimental Studies of Drug Effects (T. Stürmer)
• Not restricted by 5 S of RCTs
– Large enough to study rare outcomes
– Include people with co-morbidity
– Include people with co-medication
– Include elderly, children, pregnant women
– Include wider indication (e.g., less severe disease), off-
label use
– Wide variety of outcomes
– Lagged and long term effects
13
ADR Burden
...”Most of what we learn,
and will continue to learn,
about adverse drug
effects is from
observational studies”
Walker & Stampfer
Lancet 1996;348:489
Need for nonexperimental studies of drug effects
• Drugs rarely studied in RCTs:
– Inexpensive off-patent drugs
– Within-class comparisons, e.g. NSAIDs:
• Comparative effectiveness (pain)
• GI toxicity
• Cardiovascular morbidity/mortality
Effectiveness more important than efficacy in public health!
15
How to ascertain drug exposure in practice…
October 2008
Main data contents of the Danish prescription registries
The personal identifier (CPR number)
The type of drug (Anatomic Therapeutic Chemical (ATC)
classification codes)
The date of dispensing of the drug
The quantity of drug (number of packages, number of pills in the
package, quantity expressed by the defined daily dose (DDD))
NOT complete data on dosing information for each prescription
NOT complete data on the medical indication for prescribing the
drug
(A. Brookhart)
Anatomical Therapeutic Chemical (ATC) Classification
• http://www.whocc.no/atc_ddd_index/
• Tool for drug utilization research
• To improve quality of drug use
• Combined with Defined Daily Dosage (DDD)
• 1996: WHO International Working Group for Drug Statistics
Methodology, Oslo, Norway
• Strong reluctance to make changes
ATC Structure
A Alimentary tract and metabolism
B Blood and blood forming organs
C Cardiovascular system
D Dermatologicals
G Genito-urinary system and sex hormones
H Systemic hormonal preparations, excluding sex hormones and insulins
J Antiinfectives for systemic use
L Antineoplastic and immunomodulating agents
M Musculo-skeletal system
N Nervous system
P Antiparasitic products, insecticides and repellents
R Respiratory system
S Sensory organs
V Various
Level 1: Anatomical main group - consists of one letter.
There are 14 main groups: hierarchical
Hierarchical coding ideal for PE!
• Example: Insulin glargine
• A Alimentary tract & metabolism
• A 10 Drugs used in diabetes
• A 10 A Insulin and analogues
• A 10 A E Insulins and analogues for injection, long-acting
• A 10 A E 04 Insulin glargine
All these difficult terms in pharmacoepi…
• Drug exposure
• DDDs
• Current, former, or never drug use
• New use
• Wash-out period
• Adherence
• Medication Possession Ratio
• Persistence
• Drug Stop
• Augmenting / Intensification
• Switching
• Gap / Drug holiday
• Grace period
• (…)
Drug utilization: Incidence and prevalence
Example: Pottegård A et al: Use of exenatide and liraglutide in
Denmark: a drug utilization study. Eur J Clin Pharmacol 2013.
How is ”drug use” defined?
(from: Pottegaard et al, Eur J Clin Pharmacol 2013)
• “For the first day in each quarter, the number of
persons currently treated (point prevalence) was
estimated by finding the number of unique persons that
had redeemed a prescription that covered this day”
• “As the prescribed daily dose is not recorded in our
data, we defined the duration of the single prescription
as the redeemed quantity divided by the minimum
recommended daily dose (1.2mg for liraglutide and 10
μg for exenatide) and adding 20 % to account for
noncompliance and irregular prescription renewal”
Average daily dose
Pottegaard et al, Eur J Clin Pharmacol 2013
• “The amount of drug used per day in a period
between two dispensings was calculated as the
amount of active drug substance redeemed at the first
prescription divided by the number of days between
the two prescriptions”
• “The ‘“current dose used” was then calculated as a
moving average of the drug used per day in the last
three periods, weighed by the length of each period”
Adherence
2. Common types of epidemiologic bias in analytic
pharmacoepidemiology
• Illustrated by some examples
No Yes
No
Likely
Yes
Unlikely
Cause
Bias in selection or
measurement
Chance
Confounding
Cause
Explanation Finding
Important biases in analytic pharmacoepidemiology
• Misclassification of drug intake
– Recall bias (e.g., pregnant women with/without malformed child)
– Prescriptions: independent data, but prescription received or even
redeemed is ≠ actual drug use
• Usually bias towards the null, but CAVE if comparing drugs
• Confounding by indication
– “Drugs are usually given for a reason”
-> indication is complex and multifactorial, and often associated with the
outcome (can be considered as selection bias)
• Healthy user effects
– Healthy initiator, healthy adherer, sick stopper - CAVE preventive drugs
• Reverse causality / protopathic bias
– E.g. cancer and NSAIDs, pancreatitis and diabetes drugs
• Example 1: Glucose-lowering drugs and risk of acute
pancreatitis
• Example 2: Statin use and pneumonia prognosis
Nationwide data linkage in Denmark
Personal
Identification
Number
Civil Registry
System 1968-
(address, death)
Prescription
Data 1991-, 1996-
(2004-)
Primary care
health services
late 1990s- (vaccines etc.)
Laboratory
databases
late 1990s-
Hospital
Discharge
Registry 1977-
Clinical database
(e.g. DD2 2010-)
GP data / DAMD
2010s-
Example 1: Glucose-lowering drugs and risk of
acute pancreatitis
Modified from: ADA 74th Scientific Sessions, San
Fransisco 2014
Diabetes Care, in press
Incretin Therapy and Risk of Acute Pancreatitis:
A Nationwide Population-based Case-control Study
Background • Incretin-based therapies (the ‘incretins’ =
glucagon-like peptide 1 (GLP-1) receptor
agonists and dipeptidyl peptidase 4 (DPP4)
inhibitors) increasingly used in type 2 DM 1
• Concerns that incretins may cause
pancreatitis through stimulation of GLP-1
receptors in the pancreas 2
• However, type 2 diabetes per se has been
associated with a 1.5- to 2-fold increased risk
of acute pancreatitis 3
• A number of observational studies found
pancreatitis RRs close to 1.0 in users of
DPP4 inhibitors or GLP-1 receptor agonists,
vs. other glucose-lowering drug (GLD) use 4
• Some found increased RRs with incretins 5
We examined the association between
use of incretins and other glucose-
lowering drugs (GLDs) and risk of acute
pancreatitis in a large nationwide
population-based study
1. Holst JJ, Mol Cell Endocrinol 2009 2. Elashoff M, Gastroenterology 2011 3. Gonzalez-Perez A, Diabetes Care 2010 4. Li L, BMJ 2014 5. Singh S, JAMA Intern Med 2013
Methods: Case-control study
• Cases: 12,868 patients with a first-time
hospitalization with ICD-10 diagnosis of acute
pancreatitis (K85) identified in the Danish
national patient registry 2005-2012
• Controls: 128,680 age-, gender-, index date-,
and residence-matched population controls
selected 1:10 by incidence density sampling
Exposure
• Complete individual-level data on all glucose-lowering drug use from nationwide prescription database
• Incretins: ATC codes:
– DPP4 inhibitors: Sitagliptin: A10BH01, A10BD07; Vildagliptin: A10BH02, A10BD08; Saxagliptin: A10BH03, A10BD10; Alogliptin: A10BH04, A10BD09; Linagliptin: A10BH05, A10BD11
– GLP-1 receptor analogues: Exenatide: A10BX04; Liraglutide: A10BX07
• Other glucose-lowering drugs
• Redeemed ever before
– the date of hospital admission with acute pancreatitis
– or the index date among controls
• Current use: prescription redeemed within 100 days
• Former use: prescription redeemed >100 days ago
• New use: first ever prescription redeemed within 100 days
• Intensity of use: total cumulative number of prescriptions (no prescriptions, 1–3 prescriptions, or >3 prescriptions)
Confounders
• Other risk factors for acute pancreatitis retrieved from hospital and prescription databases: – gallstone disease
– Previous biliary tract procedures / ERCP (except <=10 days)
– Alcoholism
– Obesity
– Inflammatory bowel disease
– Cancer (except <=90 days)
– Other diseases in Charlson comorbidity index
– Steroids, NSAIDs, and other pancreatitis-associated drugs
Validity of acute pancreatitis in the Danish patient registry
• Positive predictive value
(PPV) of hospital diagnoses
of acute pancreatitis is 82% 1
• Validated by clinical
presentation with acute
pancreatitis (abdominal pain)
in the hospital record
combined with:
• either a 2-fold increase in
serum amylase or positive
findings by ultrasound or CT
scan, surgery, or autopsy
1. Floyd A, Scand J Gastroenterol 2002
Statistical analysis
• Conditional logistic regression to compute acute pancreatitis ORs,
adjusted for potential confounders
Pancreatitis cases
= 12,868
+ Incretins
no GLDs
+ other GLDs
Population controls
= 128,680
+ Incretins
no GLDs
+ other GLDs
Results
Characteristic Pancreatitis cases
(n = 12,868)
Population controls
(n = 128,680)
Ever use of GLDs 8.5% 6.1%
Incretins 0.69% (n=89) 0.53% (n=684)
Other GLDs 12.8% 8.3%
Male sex 51.6% 51.6%
Age >=60 years 47.9% 47.8%
Results: pancreatitis risk factors
Characteristic Pancreatitis cases
(n = 12,868)
Population controls
(n = 128,680)
Gallstone disease 16.8% 4.0%
Obesity 7.4% 3.1%
Alcoholism-related dx 15.4% 4.4%
Inflammatory bowel disease 2.2% 0.7%
Any cancer 9.2% 7.6%
Charlson Index >=1 33.2% 21.2%
Statins 22.4% 18.2%
Oral steroids 12.7% 7.7%
Azathioprine 1.2% 0.4%
NSAIDs 66.1% 51.4%
Antiepileptics 7.6% 3.8%
Unadjusted and adjusted ORs - incretins
Exposure Pancreatitis
cases
(12,868)
Population
controls
(128,680)
Unadjusted
RR
(95% CI)
Adjusted RR*
(95% CI)
Never use GLDs 11,777 (91.5%) 120,812 (93.9%) 1.00 (ref) 1.00 (ref)
Ever use any GLD 1,091 (8.5%) 7,868 (6.1%) 1.44 (1.35-1.54) 1.05 (0.98-1.13)
Ever use incretins 89 (0.7%) 684 (0.5%) 1.36 (1.08-1.69) 0.95 (0.75-1.21)
Ever use DPP4
inhibitors 68 (0.5%) 516 (0.4%) 1.38 (1.07-1.77) 1.04 (0.80-1.37)
Ever use GLP-1
receptor analogues 30 (0.2%) 230 (0.2%) 1.35 (0.92-1.98) 0.82 (0.54-1.23)
* Adjusted for previous diagnoses of gallstone disease, alcoholism-related conditions, obesity,
inflammatory bowel disease, or any cancer; for 3 levels of the Charlson Comorbidity Index score; and
for current use of oral glucocorticoids, azathioprine, lipid-lowering drugs, antiepileptics, or NSAIDs.
Unadjusted and adjusted ORs – other GLDs
Exposure Pancreatitis
cases
(12,868)
Population
controls
(128,680)
Unadjusted
RR
(95% CI)
Adjusted RR*
(95% CI)
Never use GLDs 11,777 (91.5%) 120,812 (93.9%) 1.00 (ref) 1.00 (ref)
Ever use any GLD 1,091 (8.5%) 7,868 (6.1%) 1.44 (1.35-1.54) 1.05 (0.98-1.13)
Ever use
metformin 732 (5.7%) 5,475 (4.3%) 1.39 (1.28-1.50) 1.01 (0.92-1.10)
Ever use
sulfonylureas 546 (4.2%) 3,748 (2.9%) 1.52 (1.38-1.66) 1.13 (1.02-1.25)
Ever use
insulin 355 (2.8%) 2,473 (1.9%) 1.49 (1.33-1.67) 0.96 (0.85-1.08)
* Adjusted for previous diagnoses of gallstone disease, alcoholism-related conditions, obesity,
inflammatory bowel disease, or any cancer; for 3 levels of the Charlson Comorbidity Index score; and
for current use of oral glucocorticoids, azathioprine, lipid-lowering drugs, antiepileptics, or NSAIDs.
Adjusted ORs by type of incretin use
Adjusted ORs by type of other GLD use
What happens with confounder adjustment?
Ever use Incretins DPP4 GLP-1 Metformin SU Insulin
Unadjusted,
age-/gender-
matched
1.36
(1.08-1.69)
1.38
(1.07-1.77)
1.35
(0.92-1.98)
1.39
(1.28-1.50)
1.52
(1.38-1.66)
1.49
(1.33-1.67)
Adjusted for:
Pancreatitis-
associated
conditions *
1.04
(0.82-1.31)
1.11
(0.85-1.45)
0.95
(0.63-1.42)
1.08
(0.99-1.18)
1.23
(1.11-1.36)
1.07
(0.95-1.21)
+ pancreatitis-
associated
drug use †
0.97
(0.77-1.23)
1.05
(0.80-1.37)
0.86
(0.57-1.30)
1.02
(0.93-1.12)
1.17
(1.06-1.29)
1.02
(0.90-1.15)
+ any
comorbidity
in Charlson
0.95
(0.75-1.21)
1.04
(0.80-1.37)
0.82
(0.54-1.23)
1.01
(0.92-1.10)
1.13
(1.02-1.25)
0.96
(0.85-1.08)
* gallstone disease, alcoholism, obesity, IBD, any cancer
†current use of oral glucocorticoids, azathioprine, lipid-lowering drugs, antiepileptics, or NSAIDs.
Strenghts
• Population-based design and setting in a comprehensive health care system with complete population coverage reduces risk of selection and referral biases
• High validity of hospital and prescription data from independent sources reduces risk of information bias
Limitations – potential biases
• Confounding by indication: Lack of exact data on diabetes duration and severity
– However, incretin therapy usually associated with more advanced diabetes and increased comorbidity
• Reverse causality / Protopathic bias: diabetes caused by pancreatic disease may lead to drug initiation
• Misclassification of exposure: prescription redemption ≠ actual drug use.
– May underestimate pancreatitis risk, bias towards the null (but CAVE if compliance differential by type of drug, when comparing drugs)
• Misclassification of acute pancreatitis – Unlikely to be related to type of drug use?
Limitations – potential biases
• Residual confounding by incompletely measured, unmeasured, or unknown confounders related to drug use
– E.g. obesity may be registered more completely in incretin and other GLD users, vs. non-users
– Limited data on socioeconomic and lifestyle factors that may be associated with use of new expensive drugs like the incretins
• Choice of confounders is critical
– Some factors may be intermediate variables (=effects of GLD use), rather than confounders (e.g. gallstones, obesity)
Interpretation
• Patients with acute pancreatitis are ~40% more likely to be users of any GLD (incretins or others), compared with other individuals
• After adjustment for available confounding factors, the use of incretin-based therapies is not associated with increased risk of acute pancreatitis
• Our null-results are consistent with two recent meta-analyses of RCTs of incretin effects, and with the vast majority of the approx. 10 observational studies on this topic conducted to date
Li L, BMJ 2014, Faillie J, BMJ 2014
Acknowledgments
• Educational material borrowed from:
• Professor Til Stürmer, UNC Gillings School of Global Public Health,
University of North Carolina, Chapel Hill, USA; and
• Professor Maurice Alan Brookhart, PhD, UNC Gillings School of Global
Public Health and UNC School of Medicine, University of North
Carolina, Chapel Hill, USA
• Jennifer Lund, PhD and Alan Kinlaw, MSPH, both UNC Gillings School
of Global Public Health
Read a book get your knowledge up…
• Pharmacoepidemiology, 5th Edition. Brian L. Strom (Editor), Stephen E
Kimmel (Editor), Sean Hennessy (Editor). ISBN: 978-0-470-65475-0.
976 pages. February 2012, Wiley-Blackwell.
Conclusions
• There are lots of possibilities, and a lot of work to do
• ”(Pharmaco-) epidemiology is not for amateurs”