Crunch-Time Manual: EPAID 7120
Epidemiologic Methods 2 Final
Exam Prep.
A Comprehensive Exam Study Guide
for a Guaranteed Top Score with
Grade A+
What are two common models of causation? - ANSmodel of sufficient and component causes
and the counterfactual
DAGs - ANSvisual presentation of causal assumptions
Bradford Hill Criteria - ANSstrength, consistency, temporality, biological gradient,
coherence, experiment, analogy, specificity, plausibility; not a checklist
strength - ANSA small association does not mean that there is not a causal effect, though the
larger the association, the more likely that it is causal.
Consistency - ANSConsistent findings observed by different people in different places with
different samples strengthens the likelihood of an effect.
Temporality - ANSThe effect has to occur after the cause
Biological gradient - ANSGreater exposure should generally lead to greater incidence of the
effect. In other cases, greater exposure leads to lower incidence.
Coherence - ANSCoherence between epidemiological and laboratory findings increases the
likelihood of an effect but "... lack of such [laboratory] evidence cannot nullify the
epidemiological effect on associations".
Experiment - ANS"Occasionally it is possible to appeal to experimental evidence".
Analogy - ANSThe effect of similar factors may be considered.
Specificity - ANSCausation is likely if there is a very specific population, at a specific site,
with a specific disease with no other likely explanation.
Plausibility - ANSA plausible mechanism between cause and effect is helpful
types of causal relationships - ANSdirect & indirect
cause - ANSAn event, condition, or characteristic that preceded the outcome, and had the
event, condition, or characteristic been different in some way, the outcome would not have
occurred at all or not for some time later.
sufficient cause - ANSA minimal set of conditions and events that are sufficient for the
outcome to occur. (e.g. exposure by itself leads to an outcome or set of exposures together
leads to the outcome)
necessary cause - ANSis a particular type of component cause that is required for the
outcome to occur. (remove exposure and do not get the outcome)
counterfactual - ANSwhat you do not actually observe (e.g. Does Lipitor (cholesterol-
reducing medicine) change the risk of heart disease? Iheart disease that would have been
observed if these persons had not taken Lipitor = counterfactual (not actually observed).
Directed Acyclic Graphs (DAGs) - ANSLike conceptual frameworks, DAGs are visual
representations of "the world"
These visual representations use a common set of "rules" to facilitate communication and
collaboration among investigators
One approach to translating causal pathways into diagram, that be can be a useful tool for
epidemiologic studies of causal associations
,Crunch-Time Manual: EPAID 7120
Epidemiologic Methods 2 Final
Exam Prep.
A Comprehensive Exam Study Guide
for a Guaranteed Top Score with
Grade A+
are a set of arrows drawn along a timeline, characterizing [theorized] causal and temporal
relationships between variables. There can never be a cycle (hence acyclic) because we can
never go back in time."
What are the rules of causal diagrams? - ANS1) no cycles
2) nodes (measured and unmeasured variable)
3) path (association)
4) directed path (theorized causation)
For diagram to be "causal" (i.e. allow determination of bias) must include ALL common
causes of every variable on the diagram
open pathway - ANSA -> B -> C (obesity to diabetes to cardiovascular disease)
blocked pathway - ANSA->B<-C (obesity ->diabetes<-congenital hyperinsulinism)
collider - ANSa node with two arrow heads pointing at it
decendant - ANSE (diabetes) is a descendant of B (glycated hemoglobin A1c)
mediators - ANSaffected by exposure AND on causal pathway between exposure & disease
AND cause of disease; translate part of the effect of exposure on disease; we do not want to
adjust for mediators
confounders - ANSassociated with the exposure, AND associated with disease, AND not on
the causal pathway; often want to adjust for confounders
direct causal pathway - ANSe.g. obesity causes MI
indirect causal paths - ANSconnect exposure and outcome of interest through other
variable(s) (e.g. obesity causes MI through High BP)
confounding - ANSdistortion of the estimated effect of an exposure on an outcome due to the
presence of of a common cause of the exposure and the outcome (e.g. taking aspirin and risk
of a heart attack (more elderly people take aspirin and more elderly people have a heart
attack).
What are the major problems for causal inference? - ANS1. Were there confounders?
2. Were these confounders measured accurately?
3. Was the analysis done appropriately?
4. sampling and measurement processes adequate
residual confounding - ANSconfounding that remains after inadequate control, e.g. because
of measurement error in measuring the confounder, improper categorization or analysis
Definitions for confounding - ANS1. "mixing of effects" leading to a change in estimate
2. "classical" definition
3. "counterfactual" model
4. "collapsibility" definition
mixing of effects - ANSthe effect of the exposure of interest is mixed together with the effect
of another variable, leading to an incorrect effect estimate ("change in estimate")
,Crunch-Time Manual: EPAID 7120
Epidemiologic Methods 2 Final
Exam Prep.
A Comprehensive Exam Study Guide
for a Guaranteed Top Score with
Grade A+
anticonservative confounding - ANSoverestimation of effect (away from the null effect)
conservative confounding - ANSunderestimation of effect (towards null effect)
qualitative confounding - ANSinterference changes qualitatively (reaches the null or crosses
the null effect)
classical definition of confounding - ANSmust meet these criteria:
1. a confounder must be associated with the exposure
2. must be a "risk factor" for the outcome
3. must not be an intermediate step in the causal pathway between the exposure and the
outcome (must not be affected by the exposure)
What are potential problems with the classical definition of confounding? - ANS-need to
have subject knowledge
-it is a univariate view of the world because you are going through each variable one by one;
you will not be able to know if you have multiple variables and one may cancel the other out.
counterfactual definition of an effect - ANSUnder the counterfactual, an effect is the
difference in the outcome that is caused by different exposure states in one study population
during one time period.
exchangeability assumption - ANSExchangeability of comparison groups regarding the
outcome distributions under homogenous exposure (treatment assignment); key assumption
for valid causal inference
Confounding according to counterfactual definition - ANSConfounding is present if the
assumption of exchangeability of response distributions under homogenous exposure
(treatment assignment) is not correct, i.e. if the exposed and unexposed groups are not
exchangeable.
collapsibility definition of confounding - ANS1. The effect is homogeneous (the same) across
the strata defined by the confounder
2. The crude and common stratum-specific (adjusted) effect estimates are unequal/not the
same(i.e., non-collapsible)
If crude ≠ adjusted measure of associationi.e. "change in estimate"(e.g. by ≥10% or p-value)
= estimates are non-collapsible = confounding is present
collapsibilty criterion - ANSStatistical criterion
Assumes that you already have data to calculate crude and adjusted measures of association
Assumes that the "truth" is in your data
Problematic to decide on confounding using statistics only
If p-values are used, remember what they represent
Pros and cons of mixing of effects - ANSIntuitive
Easy to communicate
Not very helpful for study design
Might help to understand the statistical control for confounding
Pros and cons of classical definition - ANSProvides rules
, Crunch-Time Manual: EPAID 7120
Epidemiologic Methods 2 Final
Exam Prep.
A Comprehensive Exam Study Guide
for a Guaranteed Top Score with
Grade A+
Good to check individual variables formally (for example at the analysis stage)
Can be displayed using DAGs
Helpful for study design and measurement Substantial subject/literature knowledge required
Pros and cons of counterfactual definition - ANS'Modern' way to think about confounding
Very useful for study design
Gave rise to powerful statistical methods (g-estimation of structural models, IPW of marginal
structural models)
Analysis of time-dependent confounding in longitudinal studies
Pros and cons of Collapsibility definition - ANSGives us ways to statistically test
Does not require prior knowledge
Limited value for identifying confounders
Helpful to examine how previously identified confounders "behave" in your data
unmeasured confounding - ANSif a confounder has not been measured we can't easily adjust
for it
How are confounders identified? - ANSThe identification of confounders requires expert or
substantive knowledge about the causal network of which exposure and outcome are part.
Attempts to select confounders solely based on observed statistical associations may lead to
bias.
knowledge-based approaches to identify potential confounders - ANSpublic health or clinical
subject knowledge; literature; conceptualization (e.g. DAGs)
data-driven approaches - ANSassociations within own study and adjusted versus crude
estimates (collapsibility)
conditioning - ANSadjusting for a variable; In an open pathway, if you condition on a non-
collider, you block the path. In a closed pathway, if you condition on a collider, you open the
path. In a closed pathway, if you condition on the descendant of a collider, you open the path.
data driven approaches within own study - ANScomparison at baseline (poor guide alone to
decide exchangeability) and association with outcome
What are the limitations of significance testing comparisons at baseline? - ANSA mix of
effect and sample size Does not "prove" comparabilityCharacteristics may not cause the
outcomeBivariate viewNot very informative with respect to outcome expectations of the
groups
What are the limitations of significance testing association of potential confounders with
outcome? - ANSDoes not prove causality Does not show that a factor is a confounderDoes
not consider the definition of confounding irrespective of being univariate or multivariateA
mix of effect and sample size
What are the pros and cons of knowledge-based approaches? - ANSConsider multi-
dimensional nature of confoundingDecision separate from own dataBased on body of
evidenceMinimizes risk of over- and under-adjustmentRequire judgment (subjective)Not
easy to report/replicate
Epidemiologic Methods 2 Final
Exam Prep.
A Comprehensive Exam Study Guide
for a Guaranteed Top Score with
Grade A+
What are two common models of causation? - ANSmodel of sufficient and component causes
and the counterfactual
DAGs - ANSvisual presentation of causal assumptions
Bradford Hill Criteria - ANSstrength, consistency, temporality, biological gradient,
coherence, experiment, analogy, specificity, plausibility; not a checklist
strength - ANSA small association does not mean that there is not a causal effect, though the
larger the association, the more likely that it is causal.
Consistency - ANSConsistent findings observed by different people in different places with
different samples strengthens the likelihood of an effect.
Temporality - ANSThe effect has to occur after the cause
Biological gradient - ANSGreater exposure should generally lead to greater incidence of the
effect. In other cases, greater exposure leads to lower incidence.
Coherence - ANSCoherence between epidemiological and laboratory findings increases the
likelihood of an effect but "... lack of such [laboratory] evidence cannot nullify the
epidemiological effect on associations".
Experiment - ANS"Occasionally it is possible to appeal to experimental evidence".
Analogy - ANSThe effect of similar factors may be considered.
Specificity - ANSCausation is likely if there is a very specific population, at a specific site,
with a specific disease with no other likely explanation.
Plausibility - ANSA plausible mechanism between cause and effect is helpful
types of causal relationships - ANSdirect & indirect
cause - ANSAn event, condition, or characteristic that preceded the outcome, and had the
event, condition, or characteristic been different in some way, the outcome would not have
occurred at all or not for some time later.
sufficient cause - ANSA minimal set of conditions and events that are sufficient for the
outcome to occur. (e.g. exposure by itself leads to an outcome or set of exposures together
leads to the outcome)
necessary cause - ANSis a particular type of component cause that is required for the
outcome to occur. (remove exposure and do not get the outcome)
counterfactual - ANSwhat you do not actually observe (e.g. Does Lipitor (cholesterol-
reducing medicine) change the risk of heart disease? Iheart disease that would have been
observed if these persons had not taken Lipitor = counterfactual (not actually observed).
Directed Acyclic Graphs (DAGs) - ANSLike conceptual frameworks, DAGs are visual
representations of "the world"
These visual representations use a common set of "rules" to facilitate communication and
collaboration among investigators
One approach to translating causal pathways into diagram, that be can be a useful tool for
epidemiologic studies of causal associations
,Crunch-Time Manual: EPAID 7120
Epidemiologic Methods 2 Final
Exam Prep.
A Comprehensive Exam Study Guide
for a Guaranteed Top Score with
Grade A+
are a set of arrows drawn along a timeline, characterizing [theorized] causal and temporal
relationships between variables. There can never be a cycle (hence acyclic) because we can
never go back in time."
What are the rules of causal diagrams? - ANS1) no cycles
2) nodes (measured and unmeasured variable)
3) path (association)
4) directed path (theorized causation)
For diagram to be "causal" (i.e. allow determination of bias) must include ALL common
causes of every variable on the diagram
open pathway - ANSA -> B -> C (obesity to diabetes to cardiovascular disease)
blocked pathway - ANSA->B<-C (obesity ->diabetes<-congenital hyperinsulinism)
collider - ANSa node with two arrow heads pointing at it
decendant - ANSE (diabetes) is a descendant of B (glycated hemoglobin A1c)
mediators - ANSaffected by exposure AND on causal pathway between exposure & disease
AND cause of disease; translate part of the effect of exposure on disease; we do not want to
adjust for mediators
confounders - ANSassociated with the exposure, AND associated with disease, AND not on
the causal pathway; often want to adjust for confounders
direct causal pathway - ANSe.g. obesity causes MI
indirect causal paths - ANSconnect exposure and outcome of interest through other
variable(s) (e.g. obesity causes MI through High BP)
confounding - ANSdistortion of the estimated effect of an exposure on an outcome due to the
presence of of a common cause of the exposure and the outcome (e.g. taking aspirin and risk
of a heart attack (more elderly people take aspirin and more elderly people have a heart
attack).
What are the major problems for causal inference? - ANS1. Were there confounders?
2. Were these confounders measured accurately?
3. Was the analysis done appropriately?
4. sampling and measurement processes adequate
residual confounding - ANSconfounding that remains after inadequate control, e.g. because
of measurement error in measuring the confounder, improper categorization or analysis
Definitions for confounding - ANS1. "mixing of effects" leading to a change in estimate
2. "classical" definition
3. "counterfactual" model
4. "collapsibility" definition
mixing of effects - ANSthe effect of the exposure of interest is mixed together with the effect
of another variable, leading to an incorrect effect estimate ("change in estimate")
,Crunch-Time Manual: EPAID 7120
Epidemiologic Methods 2 Final
Exam Prep.
A Comprehensive Exam Study Guide
for a Guaranteed Top Score with
Grade A+
anticonservative confounding - ANSoverestimation of effect (away from the null effect)
conservative confounding - ANSunderestimation of effect (towards null effect)
qualitative confounding - ANSinterference changes qualitatively (reaches the null or crosses
the null effect)
classical definition of confounding - ANSmust meet these criteria:
1. a confounder must be associated with the exposure
2. must be a "risk factor" for the outcome
3. must not be an intermediate step in the causal pathway between the exposure and the
outcome (must not be affected by the exposure)
What are potential problems with the classical definition of confounding? - ANS-need to
have subject knowledge
-it is a univariate view of the world because you are going through each variable one by one;
you will not be able to know if you have multiple variables and one may cancel the other out.
counterfactual definition of an effect - ANSUnder the counterfactual, an effect is the
difference in the outcome that is caused by different exposure states in one study population
during one time period.
exchangeability assumption - ANSExchangeability of comparison groups regarding the
outcome distributions under homogenous exposure (treatment assignment); key assumption
for valid causal inference
Confounding according to counterfactual definition - ANSConfounding is present if the
assumption of exchangeability of response distributions under homogenous exposure
(treatment assignment) is not correct, i.e. if the exposed and unexposed groups are not
exchangeable.
collapsibility definition of confounding - ANS1. The effect is homogeneous (the same) across
the strata defined by the confounder
2. The crude and common stratum-specific (adjusted) effect estimates are unequal/not the
same(i.e., non-collapsible)
If crude ≠ adjusted measure of associationi.e. "change in estimate"(e.g. by ≥10% or p-value)
= estimates are non-collapsible = confounding is present
collapsibilty criterion - ANSStatistical criterion
Assumes that you already have data to calculate crude and adjusted measures of association
Assumes that the "truth" is in your data
Problematic to decide on confounding using statistics only
If p-values are used, remember what they represent
Pros and cons of mixing of effects - ANSIntuitive
Easy to communicate
Not very helpful for study design
Might help to understand the statistical control for confounding
Pros and cons of classical definition - ANSProvides rules
, Crunch-Time Manual: EPAID 7120
Epidemiologic Methods 2 Final
Exam Prep.
A Comprehensive Exam Study Guide
for a Guaranteed Top Score with
Grade A+
Good to check individual variables formally (for example at the analysis stage)
Can be displayed using DAGs
Helpful for study design and measurement Substantial subject/literature knowledge required
Pros and cons of counterfactual definition - ANS'Modern' way to think about confounding
Very useful for study design
Gave rise to powerful statistical methods (g-estimation of structural models, IPW of marginal
structural models)
Analysis of time-dependent confounding in longitudinal studies
Pros and cons of Collapsibility definition - ANSGives us ways to statistically test
Does not require prior knowledge
Limited value for identifying confounders
Helpful to examine how previously identified confounders "behave" in your data
unmeasured confounding - ANSif a confounder has not been measured we can't easily adjust
for it
How are confounders identified? - ANSThe identification of confounders requires expert or
substantive knowledge about the causal network of which exposure and outcome are part.
Attempts to select confounders solely based on observed statistical associations may lead to
bias.
knowledge-based approaches to identify potential confounders - ANSpublic health or clinical
subject knowledge; literature; conceptualization (e.g. DAGs)
data-driven approaches - ANSassociations within own study and adjusted versus crude
estimates (collapsibility)
conditioning - ANSadjusting for a variable; In an open pathway, if you condition on a non-
collider, you block the path. In a closed pathway, if you condition on a collider, you open the
path. In a closed pathway, if you condition on the descendant of a collider, you open the path.
data driven approaches within own study - ANScomparison at baseline (poor guide alone to
decide exchangeability) and association with outcome
What are the limitations of significance testing comparisons at baseline? - ANSA mix of
effect and sample size Does not "prove" comparabilityCharacteristics may not cause the
outcomeBivariate viewNot very informative with respect to outcome expectations of the
groups
What are the limitations of significance testing association of potential confounders with
outcome? - ANSDoes not prove causality Does not show that a factor is a confounderDoes
not consider the definition of confounding irrespective of being univariate or multivariateA
mix of effect and sample size
What are the pros and cons of knowledge-based approaches? - ANSConsider multi-
dimensional nature of confoundingDecision separate from own dataBased on body of
evidenceMinimizes risk of over- and under-adjustmentRequire judgment (subjective)Not
easy to report/replicate
