Transcript Science, epidemiology and progress

Science, epidemiology and
progress
Simon Thornley
True or False
Is it possible to prove a hypothesis in
science?
Objectives
• Also to pick up on a couple of important
epidemiological concepts we have missed so
far.
• To understand the philosophy of science and
how it relates to epidemiological study design.
Epidemiology in a nutshell
Aim
• does exposure cause
disease?
• does drug treat disease?
Is change in exposure distribution
temporally related with change in
disease distribution?
Statistical power calculation (type-1, type-2
error, prevalence of disease in unexposed,
minimum detectable effect)
Design study
Can I randomise?
• Ethical?
• Clinical equipoise?
Yes
Randomised study
Report (RR)
No?
Observational
study
Rare disease?
One outcome?
Case-control
(report OR)
Rare
Exposure? Many outcomes?
Cohort
(report RR)
Define case and
exposure status
Table 1
Check missing data, duplicates, ranges, bivariate scatterplots and lowess
curves, principal components.
Are there systematic
differences between
exposure and
unexposed groups
(confounding)
Population divided by
exposure status?
What population is
the study sample
drawn from?
Yes (shouldn’t be in
RCT!)
Are they adjusted for
in the analysis if
confounders?
Is it representative of
underlying population
or is there likely
selection bias?
Results: Analysis
Check data distributions
Transform?
Outcome variable?
Continuous
Categorical
Report crude
measures of
association
(OR/RR/HR)
t-test
Chi-square or
Fisher exact test if cell
counts <5
Confounders?
• “Shared common
cause of exposure and
disease”
Multiple linear
regression
If difference between crude
and adjusted >10%, then
Statistical evidence of
confounding
Logistic regression
and or stratification
Report adjusted measures of
association (OR/RR/HR)
Interpret study results
Estimate OR/RR/HR and 95% C.
I.
Is there an association
between exposure
and outcome?
Is P <0.05 or 95% CI
for measure of
association contain
null value (1)?
Yes
Exposure associated with disease
Is there another
explanation?
Bias
Confounding
Information
(recall)
Selection
Shared common
cause of exposure and
disease?
(survivor; loss to follow up, hosp.
controls)
Could study design be
improved?
Regression or
stratified analysis
No
Hypothesis
likely false
Consider type-2
error;
confounding,
bias, other
studies
Type-1 error
(consider strength of
association)
How does my study
compare with others?
Discussion
Is the association I have
detected causal?
Bradford Hill criteria
Temporality: (cohort study? Not cross sectional or case-control which do not
separate exposure and disease)
Strength of association: (odds ratio or relative risk, does it indicate >50% increase)
Dose response: is there increasing association with increased exposure?
Biological plausibility: (are there any laboratory studies to support your assertions?)
Consistency: (do other studies using different methods, with different groups come
up with similar findings?)
Experimental evidence: (Any randomised studies?)
Analogy: (Any similar findings from related fields of science?)
Specificity: Is exposure to the cause reliably followed by disease?
Also: are there any other competing explanations? Are there any studies which shed
light on these? If not then…
Yes (on balance)
Exposure causes disease
Calculate Risk difference, NNT and
PPAR.
Epidemiological tidy up...
• Risk difference
– Chapter 6: Pocket Guide to Epidemiology,
Kleinbaum
– “Excess risk of exposure on risk of outcome”
• Population attributable risk
• Chapter 6: Pocket Guide to Epidemiology, Kleinbaum
Risk difference
• Cohort study of smokers after MI – 5 year CI of
death – exposure (Quit or continue to smoke)
Outcome
Death
Survival
Total
Exposure
Smoke
Quit
27
14
48
67
75
81
Risk (Smokers)
Risk (Quit)
Overall risk
RR
RD
NNT
PPAR
Total
41
115
156
0.36
0.17
0.26
2.08
0.19
5
34
Risk difference
Risk of
outcome
1
0
PPAR
• What is effect of exposure on disease in
population?
• Rare exposure with high RR may be more
hazardous than common exposure with low
RR
• PPAR= Overall risk – risk among unexposed
--------------------------------------------Overall risk
PPAR
Population attributable risk
Population attributable Risk
60
50
40
30
20
10
0
0.00
0.20
0.40
0.60
Prevalence of exposure
0.80
1.00
What two factors determine the
PPAR?
A) Absolute risk difference and confounding
B) Prevalence of exposure and effect
estimate
C) Sample size and statistical power
D) Type-1 error rate and Type-2 error rate
E) The number of cases and number of
controls
Causation
Positivism and Epidemiology
• Roots in Western philosophy
• Idea to “enlighten the masses”
• Human reason as a vehicle to combat ignorance,
superstition, and tyranny.
• Assumptions
–
–
–
–
–
Determinism (predict outcomes from scientific laws)
Objectivity (Researcher separate from participants)
Quantification (Information derived from measurement)
Reliability (extrapolation of findings to other populations)
Generalisability (“laws” that apply to different settings)
Causation in epidemiology
• “If the subject of epidemiologic inquiry is ...
the occurrence of disease and other health
outcomes ... [therefore] the ultimate goal of
most epidemiologic research is the elaboration
of causes that can explain patterns of disease
occurrence”.
– Rothman and Greenland.
Scientific Reasoning
• “Science is built of facts the way a house is
built of bricks; but an accumulation of facts is
no more science than a pile of bricks is a
house”
– Henri Poincaré
Popper
“Falsifiability” – the
fact that a scientific
theory can be
proved false by a
single contrary
incident is the
genuine
demarcation
between science
and non-science
Popper
• Testing hypotheses against experience (data)
• Falsifiability undermined the then accepted idea
of accumulated experience.
• “The self critical spirit is the essence of science.”
• “Once we have settled on a theory, then the effort
to disconfirm [falsify] begins again.”
• “If we have two theories that account for the
same data, then we should choose the simpler
one”
Scientific Inference
• No agreement about rules
• Inference based on logic
– Hypotheses can not be proved, only disproved
– Popper: method of conjecture and refutation
– Hauck: relation of theory and data a crossword
puzzle.
Fundamental problem of causal
inference
Exposed?
Yes
(No)
No
(Yes)
Outcomes
Observable
Counterfactual
Observable
Counterfactual
Induction vs Deduction
Induction
• Observations of nature, do
they fall into a pattern?
• Assumes that what has
happened in the past will
continue to happen
• Switch controls light but
does rooster control sun?
(co-occurrence)
• Eg. vit D levels low over
winter & CVD death higher?
Causal relationship.
Deduction
• Prediction made and tested
experimentally
• Limited set of observations
• Mathematical
• Needs hypotheses from
induction
• E.g. RCT of effect of vitamin
D on CVD events.
Bradford-Hill Criteria
• Strength (OR/RR)
• Consistency (meta)
• Specificity (not
useful?)
• Temporality (cohort vs
case-control or crosssectional)
• Biologic gradient
(OR/RR)
• Experimental
evidence (RCT?)
• Analogy
• Plausibility (biology)
• Coherence
Study types
Descriptive
• Describe what disease
occurs (e.g. New Zealand
Health survey)
Analytic
• Observational
– Case control or cohort
– Explain why disease occurs
(e.g. British doctors study).
• Randomised controlled Trial
– E.g. Effect of HRT on CVD.
Toxic shock
Chemical vs Toxin
How would you test?
Refutation
• Observations support a hypothesis but do not
prove it.
• A single contrary observation can do away with
thousands of “supporting” observations
• Form a new hypothesis in light of refuted old
hypothesis
• “Refutation and conjecture”
• All knowledge is tentative
• Predictions tested based on competing
hypotheses
Generalisation (Rothman)
• Not a statistical method, but a scientific
method
• Does not involve inference from a sample to
its source population, but abstraction into a
scientific hypothesis
• Studies are stronger if subjects are more
homogeneous, rather than “representative”
Can positive hypotheses be proven?
• Most studies start with positive hypotheses
– e.g. Smoking causes lung cancer
• Impossible to prove absence; only unlikely
beyond a given strength
• Statistical analyses
– ‘Null’ hypothesis refuted (unlikely smoking does
not cause lung cancer)
• Positive hypotheses only supported, not
refuted (absence refuted).
Consensus?
• “Pure” refutationist research difficult
• Findings evaluated in light of other studies
and broader opinion.
• Evidenced-based practice
True or False
Most epidemiological research is
refutationist.
Which of the following is not part of
the Bradford-Hill causal criteria?
A) Temporality
B) Magnitude of effect
C) Dose-response
D) Scientific consensus
E) Biological plausability
Summary of Scientific process
Idea
Generate
hypothesis
(Inductive)
Experiment
Test hypothesis
(Refutation)
Interpret
Inferences from
experiment
(Inductive)
New Idea
Refine hypothesis
(Refutation)
Summary
• Principles
– Induction (observed co-occurrence → cause and effect)
– Refutation (hypothesis; then predictions made and
experimentally tested)
• Epi limitations
– Few ‘clean’ experiments
– Observations – limited experimentation
– Inductivism assesses causation
• Solutions
– Causal criteria (BH)
– Meta-analyses of RCTs ideal (consistency, experimental, strength
of association, dose response).