W1.1
ARTICLE BY FOSTER (2010) – CAUSAL INFERENCE AND DEVELOPMENTAL
PSYCHOLOGY
Fundamental problem of causal inference = Moving from association to causation
↓
It is not that causal relationships can never be established outside of random assignment, but
that they cannot be inferred from associations alone – Additional assumptions are required
Current practices:
1. Authors hold causal inference as unattainable – One is left to wonder about the
usefulness of such information
2. Authors embrace causality, but apply tools that have limitations for performing causal
inference
3. Authors stray into causal interpretations of what are associations – Usage of other
terms to describe the relationships that are identified, such as ‘predictive’
Causal inference is essential to accomplishing the goals of developmental psychology:
1. A major goal of psychology is to improve the lives of humanity
2. Causal thinking is unavoidable – Sloman: Causality is one of the fundamentally
invariant relationships that humans use to make sense of the world
3. Even if researchers can distinguish associations from causal relationships, lay readers,
journalists, policymakers and other researchers generally cannot – Bad causal
inference can do real harm
Two conceptual tools moving from associations to causal relationships:
1. Directed acyclic graph (DAG): A graphical representation used in statistics and causal
inference to model the relationships between variables and to illustrate the possible
causal pathways among them
- Directed edges: Arrows in a DAG represent the direction of causal influence
- Acyclic: There are no loops or cycles in the graph – One cannot start at a node
and follow the arrows in a closed path to return to the same node
- The DAG cannot contain bidirectional arrows implying simultaneity
- Fewer linkages are preferred to the more complex
2. Potential Outcomes Framework: Defines the causal effect as the difference between
the outcomes that would be observed versus without the intervention under
consideration
- Counterfactual: Something that did not actually happen but is used to assess
the causal impact of an intervention
Three different variables:
1. Confounder: A variable that is associated with both the independent variable and the
dependent variable in a study – Not controlling for the confounder can create a
spurious or misleading association between the independent and dependent variables
and distort the true causal relationship
, 2. Collider: A variable that is influenced by two or more variables – Controlling for a
collider can create a spurious or misleading association between the independent and
dependent variables and distort the true causal relationship
3. Mediator: A variable that lies on the causal pathway between the independent variable
and the dependent variable – Provides insight into the underlying causal mechanisms –
Controlling for a mediator prevents information to flow from X to Y
CHAPTER BY PEARL (2018) – CONFOUNDING AND DECONFOUNDING: OR,
SLAYING THE LURKING VARIABLE
Confounding bias: Occurs when a variable influences both who is selected for the treatment
and the outcome of the experiment
‘Adjusting for x’/‘controlling for x’: Taking x into account or adjusting for its influence when
examining the relationship between other variables
Two issues:
1. Overrating the importance of adjusting for possible confounders = Controlling for
many more variables than needed; even for variables that should not be controlled for
2. Underrating the importance of adjusting for possible confounders
Deconfounders: The methods that are employed to address confounding and isolate the true
causal relationships between variables
Latin square design: A type of experimental design used in statistical research and
experimental studies that is particularly useful when there are two sources of variation, and
researchers want to control for both of them efficiently – The sources of variation are often
referred to as ‘rows’ and ‘columns’
Randomisation brings two benefits:
1. Eliminates confounder bias – Randomisation is a way of simulating the world we
would like to know – Severs every incoming link to the randomised variable,
including the ones we do not know about cannot measure
2. Enables the researcher to quantify his uncertainty – The uncertainty comes from the
randomisation procedure, which is known
Do-operator: A symbolic notation used in the field of causal inference and the study of
causality – Represents an intervention or an action that sets a particular variable to a specified
value in a causal model P(Y | do(X = x)) – Used to express, in precise mathematical language,
what counterfactual interventions would look like – Provides scientifically sound ways of
determining causal effects from non-experimental studies
, Surrogate definitions of confounding → Both wrong!
1. Declarative → Classical epidemiological definition of confounding: ‘A confounder is
any variable that is correlated with both X and Y’
- Association with X: A confounder is a variable that is associated with the
exposure or risk factor under investigation
- Association with Y: A confounder is also associated with the outcome or health
effect being studied
- Not on the causal pathway: A confounder is not an intermediate variable or on
the causal pathway between the exposure and outcome – It is not a variable
that is affected by the exposure and, in turn, influences the outcome, but
operates independently of the exposure-outcome relationship
2. Procedural → Non-collapsibility: ‘If you suspect a confounder, try adjusting for it and
try not adjusting for it; if there is a difference, it is a confounder, and you should trust
the adjusted value; if there is no difference, you are off the hook’ (Hernberg)
Proxy variable: A variable that is used as a substitute when it is difficult or impossible to
directly measure a certain concept of interest
Greenland & Robins → Exchangeability: No confounding exists in a study when the
treatment group is considered, one imagines what would have happened to its constituents if
they had not gotten treatment, and then judges whether the outcome would be the same as for
those who actually did not receive treatment – Counterfactual framework
Backdoor path: A path between an exposure variable and an outcome variable that is not a
direct path
M-bias: A type of bias that can occur in observational studies when there is an uncontrolled or
inadequately controlled mediator on the causal pathway between the exposure and the
outcome
CHAPTER BY SHADISH (2008) – CRITICAL THINKING IN QUASI-
EXPERIMENTATION
Locke: ‘A cause is that which makes any other thing begin to be, and an effect is that which
had its beginning from some other thing’
INUS-condition: An insufficient but non-redundant part of an unnecessary but sufficient
condition (Mackie) (E.g.: a fire started by lighting a match)
1. Insufficient: By itself not enough (in the example, other factors such as oxygen and
inflammable materials are needed)
2. Non-redundant: It is an important factor (in the example, lighting the match is
important because without it, without it, the rest of the conditions are not sufficient
for the fire
3. Unnecessary: There are other causes possible (in the example, there are other factors
that could cause the fire)
4. Sufficient: Part of a set of required factors (in the example, lighting the match could
start a fire with the other conditions present)
Hume: ‘An effect is the difference between what happened and what would have happened’
→ The discrepancy between reality and the counterfactual → Random assignment is the best
approximation to the counterfactual that we can obtain
ARTICLE BY FOSTER (2010) – CAUSAL INFERENCE AND DEVELOPMENTAL
PSYCHOLOGY
Fundamental problem of causal inference = Moving from association to causation
↓
It is not that causal relationships can never be established outside of random assignment, but
that they cannot be inferred from associations alone – Additional assumptions are required
Current practices:
1. Authors hold causal inference as unattainable – One is left to wonder about the
usefulness of such information
2. Authors embrace causality, but apply tools that have limitations for performing causal
inference
3. Authors stray into causal interpretations of what are associations – Usage of other
terms to describe the relationships that are identified, such as ‘predictive’
Causal inference is essential to accomplishing the goals of developmental psychology:
1. A major goal of psychology is to improve the lives of humanity
2. Causal thinking is unavoidable – Sloman: Causality is one of the fundamentally
invariant relationships that humans use to make sense of the world
3. Even if researchers can distinguish associations from causal relationships, lay readers,
journalists, policymakers and other researchers generally cannot – Bad causal
inference can do real harm
Two conceptual tools moving from associations to causal relationships:
1. Directed acyclic graph (DAG): A graphical representation used in statistics and causal
inference to model the relationships between variables and to illustrate the possible
causal pathways among them
- Directed edges: Arrows in a DAG represent the direction of causal influence
- Acyclic: There are no loops or cycles in the graph – One cannot start at a node
and follow the arrows in a closed path to return to the same node
- The DAG cannot contain bidirectional arrows implying simultaneity
- Fewer linkages are preferred to the more complex
2. Potential Outcomes Framework: Defines the causal effect as the difference between
the outcomes that would be observed versus without the intervention under
consideration
- Counterfactual: Something that did not actually happen but is used to assess
the causal impact of an intervention
Three different variables:
1. Confounder: A variable that is associated with both the independent variable and the
dependent variable in a study – Not controlling for the confounder can create a
spurious or misleading association between the independent and dependent variables
and distort the true causal relationship
, 2. Collider: A variable that is influenced by two or more variables – Controlling for a
collider can create a spurious or misleading association between the independent and
dependent variables and distort the true causal relationship
3. Mediator: A variable that lies on the causal pathway between the independent variable
and the dependent variable – Provides insight into the underlying causal mechanisms –
Controlling for a mediator prevents information to flow from X to Y
CHAPTER BY PEARL (2018) – CONFOUNDING AND DECONFOUNDING: OR,
SLAYING THE LURKING VARIABLE
Confounding bias: Occurs when a variable influences both who is selected for the treatment
and the outcome of the experiment
‘Adjusting for x’/‘controlling for x’: Taking x into account or adjusting for its influence when
examining the relationship between other variables
Two issues:
1. Overrating the importance of adjusting for possible confounders = Controlling for
many more variables than needed; even for variables that should not be controlled for
2. Underrating the importance of adjusting for possible confounders
Deconfounders: The methods that are employed to address confounding and isolate the true
causal relationships between variables
Latin square design: A type of experimental design used in statistical research and
experimental studies that is particularly useful when there are two sources of variation, and
researchers want to control for both of them efficiently – The sources of variation are often
referred to as ‘rows’ and ‘columns’
Randomisation brings two benefits:
1. Eliminates confounder bias – Randomisation is a way of simulating the world we
would like to know – Severs every incoming link to the randomised variable,
including the ones we do not know about cannot measure
2. Enables the researcher to quantify his uncertainty – The uncertainty comes from the
randomisation procedure, which is known
Do-operator: A symbolic notation used in the field of causal inference and the study of
causality – Represents an intervention or an action that sets a particular variable to a specified
value in a causal model P(Y | do(X = x)) – Used to express, in precise mathematical language,
what counterfactual interventions would look like – Provides scientifically sound ways of
determining causal effects from non-experimental studies
, Surrogate definitions of confounding → Both wrong!
1. Declarative → Classical epidemiological definition of confounding: ‘A confounder is
any variable that is correlated with both X and Y’
- Association with X: A confounder is a variable that is associated with the
exposure or risk factor under investigation
- Association with Y: A confounder is also associated with the outcome or health
effect being studied
- Not on the causal pathway: A confounder is not an intermediate variable or on
the causal pathway between the exposure and outcome – It is not a variable
that is affected by the exposure and, in turn, influences the outcome, but
operates independently of the exposure-outcome relationship
2. Procedural → Non-collapsibility: ‘If you suspect a confounder, try adjusting for it and
try not adjusting for it; if there is a difference, it is a confounder, and you should trust
the adjusted value; if there is no difference, you are off the hook’ (Hernberg)
Proxy variable: A variable that is used as a substitute when it is difficult or impossible to
directly measure a certain concept of interest
Greenland & Robins → Exchangeability: No confounding exists in a study when the
treatment group is considered, one imagines what would have happened to its constituents if
they had not gotten treatment, and then judges whether the outcome would be the same as for
those who actually did not receive treatment – Counterfactual framework
Backdoor path: A path between an exposure variable and an outcome variable that is not a
direct path
M-bias: A type of bias that can occur in observational studies when there is an uncontrolled or
inadequately controlled mediator on the causal pathway between the exposure and the
outcome
CHAPTER BY SHADISH (2008) – CRITICAL THINKING IN QUASI-
EXPERIMENTATION
Locke: ‘A cause is that which makes any other thing begin to be, and an effect is that which
had its beginning from some other thing’
INUS-condition: An insufficient but non-redundant part of an unnecessary but sufficient
condition (Mackie) (E.g.: a fire started by lighting a match)
1. Insufficient: By itself not enough (in the example, other factors such as oxygen and
inflammable materials are needed)
2. Non-redundant: It is an important factor (in the example, lighting the match is
important because without it, without it, the rest of the conditions are not sufficient
for the fire
3. Unnecessary: There are other causes possible (in the example, there are other factors
that could cause the fire)
4. Sufficient: Part of a set of required factors (in the example, lighting the match could
start a fire with the other conditions present)
Hume: ‘An effect is the difference between what happened and what would have happened’
→ The discrepancy between reality and the counterfactual → Random assignment is the best
approximation to the counterfactual that we can obtain
