Publication bias
Causality
Andrea Dalecká
Hynek Pikhart, Martin Bobák
Epidemiologie E0350
1
Publication bias
High-impact journals prefer clear, positive results!
Bias in systematic reviews
Form of selection bias arising if null studies are not published
If not included the overall estimate is biased upwards
Minimised by searching grey literature, trial registers and
conference proceedings to include null/negative results
e.g. the ‘drug effectiveness cycle’ (β-blocker-mortality example in
session 7), selective serotonin reuptake inhibitors in treating
depression
2
Publication bias
Failure to publish
 a negative or inconclusive trial result
 a small trial may be abandoned
Duplicate publication
 a large treatment effect
 need for research output
Eg. nine trials of ondansetron (antiemetic) in 23 publications (Tramer et al
BMJ 1997)
3
How to avoid publication bias
 To make sure studies are not double counted.
 To search for unpublished studies (e.g. contact researchers directly).
 To use non-English language publications.
 Statistical checking (funnel plots: smaller studies report more extreme
results).
 Registration of studies and to make sure all results are in public domain (not
yet fully achieved).
 Trial registration: assigns unique trial identification numbers, and to record
other basic information about the trial so that essential details are made
publicly available.
 From 2004 International Committee of Medical Journal Editors (ICMJE)
would consider trials for publication only if they had been registered before
the enrolment of the first participant.
4
Funnel plot:
5
Asymmetrical plot in the presence of bias: some smaller studies (open
circles) are of lower methodological quality and therefore produce
exaggerated effect estimates
Funnel plot:
6
Beta-blockers and total mortality after MI: meta-analysis
Egger & Davey Smith 1997 7
Causality
8
1. We find an association between exposure and outcome
2. We need to ask whether the association is causal = does
the exposure cause the outcome?
What is a cause?
Rothman (1986): Modern epidemiology:
An event, condition, or characteristic that plays an essential
role in producing an occurrence of the disease.
- Something that has an effect
- Alters disease frequency or health status
9
Association versus Causation
10
• Epidemiological research aims to discover aetiology of
disease.
• Epidemiology is the study of the association between a
potential cause (risk factor/determinant) and a specific disease
(outcome).
• Presence of a valid statistical association does not imply
causality.
• Association is not the same as causation!
• Causation goes beyond association.
• When is there evidence in support of a causal association?
Criteria for causal inference
(inferring causality)
11
Koch’s postulates (1890)
1. The agent must be found in every case of disease, but should not
be found in healthy organisms
2. The agent must be isolated from diseased host and grown in
culture
3. The cultured agent should cause same disease when introduced
into healthy host
4. The agent must be recovered from the new host and shown to be
the same as the original agent
Potential issues:- (1) asymptomatic hosts, (2) not all grown in culture.
Satisfying Koch’s postulates is sufficient, but not necessary to establish
causation.
12
Sir Austin Bradford Hill (1897-1991)
Exposure and Disease: Association or Causation?
1. Strength
2. Consistency
3. Specificity
4. Temporality
5. Dose-response
6. Biological plausibility
7. Coherence
8. Reversibility
The Bradford-Hill criteria of causation (J Royal Soc Med 1965; 58: 295-300)
13
1. Bradford Hill Criteria
Strength
 The stronger the association between exposure and outcome, the
more likely it is to be a causal relationship (less likely to be due
entirely to bias and confounding)
 A weak association does not rule out the possibility of causality
For example
• Effect measure for heavy smoking & lung cancer supported by pvalue
and narrow 95% CI
14
2. Bradford Hill Criteria
Consistency
 Observing a relationship repeatedly improves the case for causation.
 Less likely that same biases present in all of them.
 A lack of consistency in one study does not rule out causality.
 Evidence of causation by replication of results
 different samples, circumstances, study designs
For example
• Smoking and lung cancer – shown
• Oral contraceptives and breast cancer – not shown
• Meta-analysis - supportive
15
 Outcome is best predicted by one primary factor.
 One factor predicts one primary outcome.
 If demonstrated, adds evidence of causality.
 Lack of specificity does not negate causal association.
For example
 asbestos and mesothelioma – shown
 HIV and AIDS – shown
 Low lead exposure and IQ – not clear. IQ is not a definable brain
condition so there is the potential for confounding e.g. SES
16
3. Bradford Hill Criteria
Specificity
4. Bradford Hill Criteria
Temporality
 Exposure must precede outcome. This is essential
What study design is appropriate to meet this criteria?
 Optimal study design = randomised intervention study or prospective
cohort study
 Weak design for temporality = cross-sectional, case-control study
 Diseases with long latent periods can make it difficult to establish
temporality
17
5. Bradford Hill Criteria
Biological gradient (dose-response)
 Gradient or dose-response relationship between exposure &
outcome.
 Lack of a biological gradient does not rule out causality.
 Beware of threshold dose – outcome associations
 J or U shaped associations
Example - Persons who have increasingly higher exposure levels have
increasingly higher risks of disease.
18
Smoking Status Lung Cancer risk
None 1.0
Ex-smoker 1.1 (0.7-1.6)
1-20 per day 2.6 (1.7-4.0)
20-40 per day 4.4 (2.8-6.9)
40+ per day 6.8 (4.3-10.7)
6. Bradford Hill Criteria
Biological Plausibility
 There is a rational biological basis for the observed association.
 The association supported by biological theory.
 No consensus on amount of evidence required.
 Not necessary, since the plausibility may yet to be discovered
(depends on current knowledge).
 Weak criterion
For example
 Cigarettes & lung cancer. Carcinogenic substance.
 Low fibre diet & colon cancer. Dietary fibre increases intestinal
motility and dilutes/absorbs fecal carcinogens.
19
7. Bradford Hill Criteria
Coherence
 Reported association does not conflict with current knowledge.
 Can lead to publication bias.
 Can discourage search for alternative associations.
 Lack of coherence should not rule out causality.
Example of coherence
 Serum cholesterol lowering effect on heart attack, regardless of
the means e.g. diet or drug
Note:- Coherence is subtly different to plausibility. Coherence –
should not conflict, while plausibility – should be in line with
knowledge.
20
8. Bradford Hill Criteria
Reversibility (Experiment)
 Removing/reducing the potential causal factor reduces the risk
of disease.
 Likelihood of the association being causal is strengthened by
showing reversibility.
 Currently perceived as the strongest type of evidence.
 May be difficult to ascertain in diseases with long lag times
between exposure and disease.
Examples
 Smoking cessation → lower risk of lung cancer?
21
Causal Inference
 Not just ticking boxes
 Weigh evidence of causal association against other explanations
 Understanding, judgement & interpretation
 Cannot prove a causal association
 Can only be inferred based on evidence
 May change in the light of new evidence
23
Evidence of
causality
Weaknesses
in the data
Reverse causality
 Refers to the possibility that the link between exposure and
outcome is a result of the disease or disease process being
studied, not the exposure.
 Reverse causality is a type of confounding in the sense that it is
‘real’ and not an artefact of study design.
24
Example of potential reverse
causality
26
Researchers are interested in the link between blood levels of
inflammatory markers and later CVD
There are 4 possible explanations:
1. Inflammation → atherosclerosis (causal association)
2. Atherosclerosis → inflammation (reverse causal association)
3. Inflammation → atherosclerosis (association is bi-directional)
4. Other processes lead both to atherosclerosis and inflammation
(confounding) e.g. diet
Causation and Public health policy
 Ideally based on ‘evidence’ - meta-analyses and systematic reviews
 Eradication of poverty for improving health?
 Reduction in social inequality for reducing health inequality?
 Removing endocrine disruptor chemical from environment reduces
immunological disorders?
Little direct causal evidence, but the weight of evidence
 Competing interests:
 Public, Industry, Government, Scientific community
 Considerations of efficiency, cost-effectiveness and harm
27
From single to multiple
causes
28
Kenneth Rothman (1986)
29
• The individual factors are called component causes.
• The complete pie, which might be considered a causal pathway,
is called a sufficient cause.
• A disease may have more than one sufficient cause, with each
sufficient cause being composed of several component causes
that may or may not overlap.
• A component that appears in every pie or pathway is called
a necessary cause, because without it, disease does not occur.
• In this figure, that component cause A is a necessary cause
because it appears in every pie.
Sufficient cause – a minimum set of factors and circumstances which if present will
produce disease
For example. In the 19th/20th century, no BCG, poor ventilation, poor nutrition,
crowding, exposure to TB ➔ TB
30
Sufficient cause
No BCG
Poor
ventilation
Poor
nutrition
Exposure
to TB
Source – Modern Epidemiology. Rothman 1986
Crowding
What is absolutely necessary
for disease to occur?
Necessary component = exposure to TB
Necessary and Sufficient
Components
• Necessary and sufficient - factor necessary for disease to
occur, and no other factors required
• Necessary but not sufficient - factor necessary for disease to
occur, but other factors also required
• Sufficient but not necessary - factor alone can cause disease,
but so can other factors in its absence
• Neither sufficient not necessary – factor cannot cause disease
on its own, nor is it a factor that is required for disease to occur
31
Multiple causes
 Rarely a single cause
 Disease results from a complex interaction of factors
 Web of causation / chain of causation
 Common in social epidemiology
32
Causal pathway model of the
association between education and
health conceptual
33
Source: https://www.jstor.org/stable/pdf/3559677.pdf Bold arrow = pathway supported by evidence
Further reading on causality
 Bhopal 2002 or 2016, Concepts of Epidemiology chapter 5.
Also available free online
http://oxfordmedicine.com/view/10.1093/med/97801987396
85.001.0001/med-9780198739685
35