Block 1: X-Centered Research- Causal Arguments and Experiments (Lesson 2) Flashcards
Definition of causality
Remember: This is the goal of most x-centered research!
To say that a factor, X, is the cause of an outcome, Y
= a change in X generates a change in Y,
through a mechanism (M)
relative to what Y would otherwise be (counterfactual),
given certain ceteris paribus assumptions
and scope conditions
The Causal Diagram
[Image]
General Criteria of a good causal argument (NOT descriptive)
- MSMCII*
- A manipulable separation mechanism with a clear, independent impact*
Manipulability: Is X manipulable?
Separation: How separable is X relative to Y?
Mechanism: How does X generate Y? (M)
Clarity: Can X and Y be operationalized?
Independence: Is X independent of other causes of Y?
Impact: How much of the variation in Y can X explain? (significance)
Moving from argument to analysis
Going from claims to verifying
What is the standard theory of causation called in the social sciences? What does it comprise of?
Neyman-Rubin-Holland Theory of Causation
Comprises of a:
– Factual
– Counterfactual
– Antecedent (event)
– Causal effect
Formal (mathematical) way of writing a treatment effect on variable “A”
EA = YA(1) – YA(0)
What is experimental data?
Where you control the data generation process
What is observational data?
Where you do NOT control the data generation process (real-world data) Difficult to control for all confounders
Issue with experiments?
Internal vs. external validity Might be highly valid internally, but have no external validity Also, experimental situations are often artificial or context-dependent
Describe the problem of causation (causal inference) for observational data
We cannot control for common-cause confounders without knowing the data generation process, which can be difficult in real-world scenarios Might create problems of endogeneity
Definition of a confounder
Any factor that might interfere with finding causality from covariational evidence => Creates a wrong or spurious X-Y relationship)
What is the selectivity problem?
Unless you can control for the selection process, chosen criteria for putting cases into treatment and control groups may influence the outcome in unwanted ways
Explain conditional independence
Use control variables to control for the selection process, so the remaining variation in X/Y is the result of the treatment effect
Experimental studies and types of these
Study where the researcher controls the data generation process – Laboratory experiments – Field experiments
Non-experimental studies and types of these
Researchers do NOT control the data-generating process – Natural experiments – Quasi-experiments – Classic observational studies
The Three Hallmarks of an Experiment
– Response of subjects to a ‘treatment’ is compared to the response of other subjects in a ‘control’ group (often no treatment).
– The assignment of subjects to treatment and control groups is random (identical sampling probability).
– Application of the treatment is under the control of the researcher (control for confounders)
The strengths of experiments
Explanation:
– Experiments help determine causality by creating a good counterfactual
– Randomization is used to rule out any other explanatory variable
– Experimental control enables researchers to create the difference between treatment and control groups
– With perfect experiments, the control group is therefore extremely comparable to the counterfactual for the treatment group.
Effects:
– High internal validity
Why do we not see more experiments in the social sciences?
– Issues pertaining to external validity (whether results are can broadly be replicated and generalized)
– Experiments are best suited to analyze individual behavior
– Some outcomes cannot be replicated through experiments
– Experiments are expensive
– Ethical issues
Definition of field experiments
Field experiments try to simulate the conditions under which a causal process occurs, to enhance external validity
Difference between laboratory and field experiments
Laboraty experiments:
– Have better control over the experiment
– Create environments that do not exist in reality
– Greater range of variation can be induced
– BUT: Always suffer under questions about external validity
Definition of natural experiments
- Natural experiments are observational (non-experimental) studies 2. Assignment of the subjects to treatment and control groups is “as if” random (no self-selection!) Important: If this assumption doesn’t hold, we cannot speak of a natural experiment!
What kind of treatment effect does causal analysis focus on?
Focuses on mean causal effects, i.e. the Average Treatment Effect (ATE), at the population level
What is ITE?
Individual Treatment Effect is the impact of the treatment on a single unit compare to its control group
What is ATE?
Average Treatment Effect (population level) ATE is the mean impact of a change in X on Y across the population It equals the average of ITE
Typology of Treatment Effects
[Image]
