Enhancing Causal Inference

1
Facts lost Facts are never what they seem to
be Nothing there! No information of any
kind. Facts are simple and facts are
straight Facts are lazy and facts are late Facts
all come with points of view Facts dont do what
I want them to Facts just twist the truth
around Facts are living turned inside out Facts
are getting the best of them Facts are nothing on
the face of things Facts continue to change their
shape Cross-eyed and Painless The Talking
Heads
2
Enhancing Causal Inference

• Bradley Evanoff, MD, MPH

3
Summary of How Research Works
infer infer Truth in the
Truth in the Findings in
Universe Study the study
  Research Random
Study Plan Random Actual
Study Question systematic
systematic error
error  
4
Criteria Used To Infer Causes Of Diseases In
Humans

• Data from well-designed randomized studies show
  an association

49
5
Criteria Used To Infer Causes Of Diseases In
Humans

• Data from well-designed randomized studies show
  an association
• Can we get all the answers we need from RCTs?
  (expensive, differences between centers,
  contamination of intervention, placebo effect,
  long-term follow-up, etc.)

49
6

• Science is the process of discovering truth, and
  truth is sampled each time we do a study. The
  results from all of our studies will be
  distributed around the truth, and different study
  designs give different amounts and different
  qualities of sampled material. Truth is
  ascertained only when sufficient numbers of
  appropriate studies are conducted, and no one
  study or one study design has a monopoly on
  truth.
• Trudy Bush, Beyond HERS Some (not so) random
  thoughts on randomized clinical trials 2001

7
Criteria Used To Infer Causes Of Diseases (Hill
criteria)

• Data from nonrandomized studies show an
  association and
• Suspected cause precedes disease
• Association is strong
• Association makes biological sense (coherence)
• Magnitude of association is strongest when its
  predicted to be so (dose-response)
• No more plausible explanation
• Specificity (exposure leads to one or few
  effects)
• Consistency of results across different studies

49
8
(No Transcript)
9
Summary of How Research Works
infer infer Truth in the
Truth in the Findings in
Universe Study the study
  Research Random
Study Plan Random Actual
Study Question systematic
systematic error
error   Target Intended
Actual Population design sample
implement subjects Phenomena Intended
Actual Of
interest variables
measurements  
External Internal Validity Validity

10
The five explanations when an association between
coffee drinking and myocardial infarction (MI) is
observed in a sample
11
Type of Design Phase Analysis
Phase Spurious (How to prevent the
(How to evaluate Association rival explanation)
the rival explanation)
Chance Increase sample size Interpret p value
in (due to random error) and other
strategies context of prior evidence Bias C
arefully consider the Obtain additional data to
(due to systematic error) potential of each
see if potential biases difference between
the have actually occurred research
question and the study plan Check
consistency with Subjects other studies
(especially Predictor
those using different Outcome
methods)
12
Strengthening the inference that an association
has a cause-effect basis ruling out other real
associations
Type of Association Design Phase Analysis
Phase
Effect-Cause Do a longitudinal
study Consider Biologic (outcome is
actually plausibility cause of the
predictor) Obtain data on the historic seque
nce of the variables Effect-Effect (Confounding
variable is cause of both the predictor and
the outcome)
13
Design phase strategies for coping with
confounders
14
Design phase strategies for coping with
confounders
Strategy Advantages Disadvantages
Matching Can eliminate influence May be time
consuming and of strong constitutional expensiv
e, less efficient than confounders like
age increasing the number of and
sex subjects (e.g., the number of
controls per case) Can eliminate
influence Decision to match must be of
confounders that are made at outset of
study difficult to measure and can have
irreversible adverse effect on analysis
and conclusions
15
Design phase strategies for coping with
confounders
Strategy Advantages
Disadvantages Matching Can increase precision
Requires early decision
(Continued) (power) by balancing about which
variables the number of cases and are
predictors and controls in each stratum which
confounders Removes option of
studying matched variables as
predictors or as intervening
variables Requires matched
analysis Creates the danger of
overmatching ( i.e., matching on a
factor that is not a confounder, thereby
reducing power)
16
Analysis phase strategies for coping with
confounders
Strategy Advantages
Disadvantages
Stratification Easily
understood Number of strata
Flexible and reversible Limited by sample
size can choose which needed for each
stratum variables to stratify
Few strata per co-variable leads to
less complete control of confounding
Relative co-variables must have been
measured
17
Analysis phase strategies for coping with
confounders
Strategy Advantages Disadvantages
Statistical Multiple confounders Model
May not fit can be controlled
simultaneously Incomplete
control of confounding (if
Information in model does not
fit continuous variables
confounder-outcome can be fully used
relationship) As flexible and
Inaccurate estimates reversible as
of strength of
effect stratification (if model
does not fit predictor-outcome
relationship)
Results are hard to
understand Relevant
co-variables must have been
measured
18
(No Transcript)
19
Case-control study of screening sigmoidoscopy and
mortality from colon cancer (Selby et. al. 1992)

• 261 cases (died from cancer of rectum or distal
  colon) vs. 868 controls matched for age and sex
• Examined use of rigid sigmoidoscopy during 10
  years prior to cancer diagnosis
• 8.8 of cases vs. 24.2 of controls had
  undergone screening crude OR 0.30 (0.19 - 0.48),
  adjusted OR 0.41 (0.25 - 0.69)

20
yoursigmoidoscopy.com
21
Table 2. History of Screening Tests during the
10-Year Period before the Diagnosis of Fatal
Cancer within Reach of the Rigid Sigmoidoscope in
the Case Subjects

• Although the data are presented as unmatched
  analyses, the P values were obtained from matched
  analyses with conditional logistic regression.
  For sigmoidoscopes the P value is for a model
  that contained two indicator variables for the
  three levels of exposure. For the other tests,
  the P values are for parameters associated with
  the variables treated continuously.
•  

22
Additional study

• Examined 268 cases of fatal cancer of more
  proximal colon (gt 20 cm from anus), 268 controls
• out of reach of rigid sigmoidoscope
• OR 0.96 (0.61 - 1.50)

23

• The specificity of the negative association
  within the reach of the sigmoidoscope is
  consistent with a true efficacy of screening
  rather than a confounding by unmeasured selection
  factors.

24
Finagle's Laws of Information ? 1. The
information you have is not what you want. ?2.
The information you want is not what you need.
?3. The information you need is not what you can
obtain. ?4. The information you can obtain costs
more than you want to pay.