Transcript Screening Pharmaceuticals for Possible Carcinogenesis: Three

Screening Pharmaceuticals for
Possible Carcinogenesis:
Three Decades of Experience
Supercourse Lecture
April 2, 2010
Gary D. Friedman, M.D., M.S.
Adjunct Investigator and Former Director
Kaiser Permanente Division of Research
Consulting Professor of Epidemiology
Department of Health Research and Policy
Stanford University School of Medicine
Pharmaceuticals and Cancer
Study:
Current team of investigators
Laurel A. Habel, Principal Investigator
Gary D. Friedman
Charles P. Quesenberry, Jr.
James Chan
Natalia Udaltsova
Ninah Achacoso
Kaiser Permanente Division of Research
Other contributors
Hans Ury
Donna Wells
Joe V. Selby
Bruce Fireman
Stephan Van Den Eeden Sheng-Fang (Sophie) Jiang
Lisa Herrinton
Nina Oestreicher
Tamirah Haselkorn
Elizabeth (Dawn) Flick
Alice Whittemore
Kristin Sainani
Stephan Woditschka
Christopher Rowan*
Christine Iodice*
*Pending as of August 2010
History
• FDA-supported drug reaction monitoring
system at Kaiser Permanente1-3
• Initial data sources: pharmacy and clinics.
• Initial (1977) and subsequent NCI grant
support for our screening-for-carcinogenesis
studies
Initial Kaiser Permanente
surveillance
• Drugs: 1969-1973 San Francisco KP
pharmacy records: cohort of 143,574
subscribers who received prescribed drugs.
• Cancers
– Pre-existing: 1968-9 manual SF hospital files
– Incident
•
•
•
•
Tumor registry: 1969Hospitalizations: 1971All manually confirmed, then only hospital
Later: total tumor registry coverage
Initial Kaiser Permanente
surveillance (cont’d)
• Users of 215 most commonly dispensed
drugs carefully followed up for cancer
compared to non-users, adjusted only for
age and sex.
– >=500 prescriptions or >=300 recipients
Initial Kaiser Permanente
surveillance (cont’d)
• Biennial analyses: each site, all combined
– Lag analyses: 1-year, 2-year (to avoid
associations due to treatment of prediagnostic cancer symptoms)
– Optional dose-response, crudely based on
number of dispensings
Initial Kaiser Permanente
surveillance (cont’d)
• Associations found are just clues
– Most due to chance or confounding
– Select interesting ones for more detailed
study, mostly chart review of exposed cases,
some comparisons with non-exposed cases,
exposed non-cases
Evolving view of screening for
associations
• “Hypothesis-seeking”
• “Data-dredging”
• “Hypothesis-free research”
Some accomplishments of our
non-drug hypothesis-seeking
• Alcohol and coronary heart disease (1974)4
• Alcohol and blood pressure (1977)5
• Leukocyte count and coronary heart
disease (1974)6
• Obesity and multiple myeloma (1994)7
Question of statistical significance
and multiple comparisons
• Thousands of comparisons
• Adjustment controversial
• What we have done:
– Switch from .05 to .01 for screening
– Constantly emphasize in publications that
most of what we see is due to chance
Examples of confounding by
prediagnostic cancer symptoms or
by indication
• Observed positive associations
– Iron and large bowel cancer (anemia treated
before the cancer, which has bled, is
diagnosed)
– Antacids or cimetidine and stomach cancer
(treatment of ulcer-like symptoms of the cancer)
– Tetracycline and lung cancer
(treatment of exacerbations of chronic
bronchitis in smokers, who are also prone to develop
lung cancer)
Iron-large bowel cancer association:
effects of lagged analysis
No lag
1-year lag
2-year lag
SMR (Obs/Exp cases)
2.4
1.8
0.7
Illustrates use of lagged analysis to exclude
some forms of confounding.
Findings in initial database
• Follow-up to 2002, up to 33 years
• 4 papers on screening results; follow-up
up to 7, 9, 15, 19 years 8-11
• Not much that can’t be attributable to
chance or confounding. Fairly reassuring.
Findings in initial database (cont’d)
• Association most studied: barbiturates and
lung cancer.
– Tumor promoter in rodent liver
– In two follow-ups can’t readily rule out
confounding by smoking—too few cases in
nonsmokers.12,13
– Negative association with bladder cancer in
smokers.14
Findings in initial database (cont’d)
• We published several findings, most of which did
not confirm drug/cancer associations, reported
mostly by others
–
–
–
–
–
–
–
–
–
Antidepressive drugs15
Cimetidine16
Clofibrate9
Digitalis17
Diphenylhydantoin (phenytoin)8,10
Iron18,19
Lindane20
Methylergonovine21
Metronidazole8,22,23
Findings in initial database (cont’d)
• Additional published findings, most of which did
not confirm drug/cancer associations, reported
mostly by others
–
–
–
–
–
–
–
Phenylbutazone8,24
Propoxyphene9,25
Rauwolfia (reserpine)8,26
Selenium sulfide11
Several drugs with colon cancer27
Spironolactone8
Tertiary amine drugs including oxytetracycline8,9
Important source of data on
humans
• IARC evaluations of carcinogenic risks to
humans.
• Six monographs concerning pharmaceuticals,
1980 through 200128-33: our data were cited on
18 drugs.
• Our data were the only source concerning
humans (other than case reports) for 9 of these.
New drug surveillance for possible
carcinogenesis
• Pharmacy Information Management System (PIMS).
• Phased in starting in 1991, with complete coverage of
all pharmacies in the Northern California Kaiser
Permanente Medical Care Program by mid-1994.
• Over 3 million subscribers; more than 90% have at
least partial financial coverage for prescriptions filled
at Program pharmacies. Currently over 7 million
current and former subscribers in this database.
• Follow-up for cancer in the Kaiser Permanente Cancer
Registry, part of the national Surveillance,
Epidemiology, and End Results (SEER) program.
Evolving statistical methods
• SIR: Standardized incidence ratio (formerly
referred to as SMR, standardized morbidity ratio)
Observed/expected cases, expected based on
age-sex-specific incidence in non-users of the
drug. Poisson test.
• Cox model: with age as time variable, control for
calendar year to account for changes in drug
use/cancer incidence
• Nested case-control analysis.
– 10 then 50 person-time controls per case, matched
for age, sex, length of membership
– Why so many controls? (Hennessy S. et al,
AJE;149:195-7, 1999)
50 vs 10 controls/case (1)
• Example: drug not commonly used by adults:
methylphenidate (Ritalin), 3+ Rx’s
• Few exposed subjects, e.g., prostate cancer
OR(95% CI)
– Exposed cases
10
– Exposed controls (10/case) 52 1.92 (0.98-3.78)
– Exposed controls (50/case) 218 2.30 (1.22-4.33)
50 vs 10 controls/case (2)
• Methylphenidate 3+ Rx’s
• More exposed subjects, e.g., any cancer
OR(95% CI)
– Exposed cases
35
– Exposed controls (10/case) 226 1.55(1.09-2.22)
– Exposed controls (50/case) 1151 1.53 (1.09-2.14)
External adjustment
• Only confounders available for all subjects
– Age, sex, length of membership (matched)
– Some drugs, e.g., hormone use for female cancers
– Need to control for others
• External adjustment using the method of
Schneeweiss et al.; spreadsheet available for
downloading from the Internet at www.drugepi.org
• Requires estimates of:
–
–
–
–
Confounder/drug use association
Confounder/outcome association
Prevalence of confounder
Prevalence of drug use
Example of external adjustment
• Statins and smoking-related cancers.
• Drug/confounder: linked pharmacy data to
Member Health Survey
• Confounder/disease: literature
• Drug use prevalence: pharmacy, membership
• Confounder prevalence: Member Health
Survey
• Statins/lung cancer/men: OR 1.060.82
External adjustment for sensitivity
analysis
• Limitation: can only adjust for one variable
• For sensitivity analysis of breast cancer37, we
generated a range of odds ratios by assuming
– a dichotomy of overall risk among the subjects,
varying over a range of 3-fold to 1/3-fold
– prevalence of high or low risk: 10%-50%
– drug/high-or-low-risk association: 1.5, 2.0
• Example: reserpine, early data base (834 users)
– OR (95% CI)
1.07 (0.74-1.55)
– OR (sensitivity limits*) 1.07 (0.89-1.25)
– Uncertainty due to chance>uncertainty due to
uncontrolled confounding.
*Sensitivity limits are the OR’s that differed the most from the point
estimate given the above assumptions about confounding.
Some recent screening results
• Criteria for positive associations of interest
– OR for 2-year lag, 3+ dispensings >=1.50, p<0.01
– OR for 3+ dispensings > OR for 1 dispensing
(crude confirmation of dose-response)
• Recent paper 39 on 105 newly studied drugs
• There were 101 positive associations for 61 drugs.
• 66 associations were judged to have involved
substantial confounding
– e.g., smoking-related cancers, corpus uteri with
antidiabetic drugs (both related to obesity)
• 35 associations probably not due to confounding
Associations deserving further
study
• 35 associations probably not due to confounding
• 11 with some evidence that may not be due to
chance-deserve further study.
• Examples:
– sulindac and gallbladder cancer: sulindac is excreted in
bile and its metabolites found in gallstones
– hydrochlorothiazide (HCTZ} and lip cancer: HCTZ is a
photosensitizer and sun exposure is a risk factor for lip
cancer
– fluoxetine (Prozac) and paroxetine (Paxil) and testicular
cancer: both SSRI’s have been reported to cause
testicular damage in rats at high doses
Comments about screening
• A reviewer for the journal suggested that we switch
one of the screening criteria from p<0.01 to p<0.001.
• If this were done, we would have missed the
sulindac/gallbladder cancer and the
paroxetine/testicular cancer associations.
• Some reassuring findings confirming known
associations using our criteria.
– Hydrochlorothiazide and renal cancer
– Cyclophosphamide and bladder cancer and myeloid
leukemia
Not just screening
• Also studies to help evaluate other findings about
possible pharmaceutical carcinogens
– Antibiotics and breast cancer: did not confirm except
weakly for tetracyclines and macrolides. Possible
confounding by indication (acne) 35.
– Animal mammary carcinogens: weak support for
furosemide and griseofulvin 37.
– IARC: drugs with limited evidence of carcinogenicity in
humans; some supporting evidence for griseofulvin,
metronidazole, and phenytoin38.
– Methylphenidate (Ritalin) (chromosomal abnormalities):
some evidence for lymphocytic leukemia in children34.
Thanks and References
• Thank you for your interest in this lecture and
our work
• References are provided in the note below.