Causal Inference or Truth in the Universe

1
Causal InferenceorTruth in the Universe

• Importance of clinical trials
• Major pitfalls in clinical trials
• Low power
• Not randomized
• Unblinded
• Incomplete follow-up

2
Framework

• Untruth - spurious associations
• chance (small sample size)
• bias (selection bias and other biases)
• Truth - real associations, not always causal
• effect - cause
• effect - effect (confounding)
• cause - effect (truth in the universe)

3
Estrogen and CHD in Women

• RQ Does postmenopausal estrogen therapy reduce
  CHD risk in women?
• Design Cross-sectional
• Subjects 20 postmenopausal women - entire
  population of my Tuesday clinic
• Measurements estrogen therapy (ever/never)
  self-report CHD (yes/no) chart review

4
Estrogen and CHD in WomenCross-Sectional Study

• CHD No CHD
• E
• No E

1
4
5
9
15
6
20
13
7
RR 0.5
5
Estrogen and CHD in Women

• RQ Does estrogen therapy reduce CHD risk?
• Design Case-control
• Cases 1000 women admitted to SFGH over a 5-year
  period with discharge diagnosis of CHD (ICD-9
  codes)
• Controls 1000 women identified by random digit
  dialing in SF who report no CHD
• Measurements CHD based on discharge diagnosis
  estrogen therapy based on self-report

6
Estrogen and CHD in WomenCase-Control Study

• CHD No CHD
• E
• No E

200
300
500
700
1500
800
2000
1000
1000
OR .6 p .01
7
Estrogen and CHD in Women

• RQ Does estrogen therapy reduce CHD risk?
• Design Case-control
• Cases 1000 women admitted to Kaiser over a
  5-year period with discharge diagnosis of CHD
• Controls 1000 women admitted to Kaiser over the
  same 5-year period with no discharge diagnosis of
  CHD
• Measurements CHD based on discharge diagnosis
  estrogen therapy based on computerized pharmacy
  records

8
Estrogen and CHD in WomenCase-Control Study

• CHD No CHD
• E
• No E

9
Confounding
All
CHD
No CHD
130
370
500
E
870
630
1500
No E
1000
1000
2000
OR .25 p .001
Age 50-64
Age 65-79
CHD
No CHD
CHD
No CHD
360
400
E
90
10
100
E
40
60
540
600
No E
810
90
900
No E
900
100
1000
100
900
1000
OR 1.0 p .9
OR 1.0 p .9
10
Controlling Confounding

• Design stage
• Matching
• Specification
• Randomization
• Analysis stage
• Stratification
• Multivariate modeling

11
Estrogen and CHD in Women

• RQ Does estrogen therapy reduce CHD risk?
• Design Prospective cohort
• Subjects 59,337 PM nurses followed for 16 years
• Measurements Self-reported estrogen use
  self-reported CHD events validated by chart
  review
• Analysis Multivariate logistic regression - age,
  ethnicity, education, blood pressure, diabetes,
  smoking, alcohol, family history of CHD and
  hypercholesterolemia

12
Nurses Health Study

• Hormones N PYAR CHD RR P-value
• Never 20,034 324,748 452 1.0 referent
• Past 12,503 150,238 195 0.8 0.06
• Current 14,000 166,371 98 0.6 0.01

Grodstein, NEJM, 1996
13




RISK FOR CORONARY HEART DISEASE IN ESTROGEN USERS
VS. NONUSERS
Cohort Studies
Grodstein, 2000

Falkeborn, 1992

Wolf, 1991

Henderson, 1991

Sullivan, 1990

Avila, 1990

Criqui, 1988

Petitti, 1987

Bush, 1987

Wilson, 1985

Angiographic Studies

McFarland, 1989

Sullivan, 1988

Gruchow, 1988


Case-Control Studies
Mann, 1994

Rosenberg, 1993

Croft, 1989

Beard, 1989

Szklo, 1984

Ross, 1981

Bain, 1981

Adam, 1981

Rosenberg, 1980

Pfeffer, 1978

Talbott, 1977

Rosenberg, 1976

RR 0.65
Summary Relative Risk

s
0.01
0.1
1
10
Relative Risk
14
Potential Mechanisms

• ESTROGEN
• Improves lipoproteins
• Reduces LDL 10-15
• Increases HDL 10-15
• Retards atherosclerosis
• Prevents coronary vasoconstriction

15
Estrogen and CHD in Women

• Observational findings
• Strong association
• Consistent association
• Plausible biologic mechanism

16
Reasons to be Cautious

• Observational findings susceptible to bias and
  confounding
• Estrogen has known risks
• (Was a) preventive therapy widely used among
  healthy women

17
Estrogen and CHD in Women

• RQ Does estrogen therapy reduce CHD risk?
• Design Randomized trial
• Subjects 2500 PM women with CHD
• Intervention Estrogen progestin vs. placebo
• Measurements Predictor treatment
• outcome CHD death or nonfatal MI

18
Estrogen and CHD in WomenRandomized Trial

• CHD No CHD
• HT
• No HT

19
Important Features of RCTs

• Adequate Power Rule out chance associations
• Find clinically significant associations
• Randomization Comparability at baseline
• - Bias
• - Confounding
• Blinding Comparability during follow-up
• - Placebo effect
• - Differential outcome ascertainment
• - Co-intervention
• Complete Follow-up Comparability at the end of
  the trial

20
Power of the Placebo

• Internal Mammary Artery Ligation for Angina
• In unblinded trials, IM ligation
• reduced angina 60
• In blinded trials, reduced angina 65 in

• subjects who underwent sham IM ligation

• subjects who underwent IM ligation

21
Differential Outcome Adjudication

• Canadian Cooperative MS Trial
• 165 patients with multiple sclerosis
• plasma exchange cyclo pred
• sham plasma exchange placebo meds
• Outcome structured neurologic exam by blinded
  and unblinded neurologists
• More improvement with plasma exchange by
  unblinded, but not blinded assessment

Noseworthy, Neurology, 1994
22
Co-Intervention

• Unintended effective interventions
• participants use other therapy or change behavior
• study staff, medical providers, family or friends
  treat participants differently
• Nondifferential decreases power
• Differential causes bias

23
Heart and Estrogen-progestin Replacement Study
(HERS)

• 2763 postmenopausal women lt 80 years old with
  documented CHD and a uterus
• Randomized to CEE 0.625 mg plus MPA 2.5 mg or
  identical placebo
• Followed every 4 months for 4.2 years
• Separate gynecology group managed bleeding
• Outcome nonfatal MI and CHD death

24
HERS Trial Profile
Placebo1,383
Estrogen Progestin1,380
Died - 123 Dead or completed follow-up -
91 Vital Status Known - 100
Died - 131 Dead or completed follow-up -
91 Vital Status Known - 100
25
HERS Baseline Characteristics

• HRT Placebo
• Age (years) 67 67
• White () 88 90
• Current Smoker () 13 13
• Diabetes () 19 18
• Blood pressure (mmHg) 135 135
• LDL-C (mg/dL) 145 145
• BMI gt 27 (kg/m2) 57 55
• Prior estrogen use () 24 23

26
CHD Events in HERS
R.H. 1.0 (95 CI 0.8 to 1.2)
Hulley, JAMA 1998
27
HERS Primary Outcomes

• EP Pbo
  RR p-value
• Total CHD events 172 176 1.0 0.9
• CHD death 71 58 1.2 0.2
• Non-fatal MI 116 129 0.9 0.5

28
HERS Cardiovascular Outcomes

• HRT Placebo RH p-value
• (N1380) (N1383)
• CABG 88 101 0.9 .4
• PTCA 164 175 0.9 .6
• Unstable angina 103 117 0.9 .4
• CHF 128 112 1.0 .6
• PVD 94 108 0.9 .3
• Stroke/TIA 108 96 1.1 .4

29
HERS vs. Observational Studies

• Why did the findings of HERS differ?
• HERS design different
• adverse effect of added progestin
• no benefit in women with CHD
• Observational findings wrong
• selection bias - comparison groups differ
• adherence bias

30
Benefit of Adherence with Medication

• 5 Year Mortality ()
• Adherence Clofibrate Placebo
• All 20 21

26 16
lt80 pills 22 gt80 pills 16
Coronary Drug Project, NEJM, 1980
31
Are Observational Studies Useless?

• NO
• generate important hypotheses
• provide only answer if trial not feasible
• generally produce correct answer
• But bias and confounding always worrisome
• Particularly problematic for interventions that
  require selection and adherence

32
Thank you!