Block 1: X-Centered RDs: Causal strategies: Beyond X and Y Flashcards
Ross’ paper: “Is Democracy Good for the Poor”
– Measures child mortality as a proxy for welfare of the poorest
– Research wants to demonstrate causation: Democracy => Benefit to the poorest
– Based on co-variation
– Non-experimental/observational data
– Panel data research design
– Sample = population
Problem: Systematic bias with regards to missing observations (rich, autocratic states)
What is the fundamental problem of causation?
One uses a comparison of factuals to create information about counterfactuals
Refined definition of a confounder
Any factor that renders the co-variation of X and Y spurious, making evidence of causality difficult
Typology of confounders
Common incidents compound collisions with antecedent, exogenous mechanisms
Graph-based typology:
– common cause,
– incidental,
– compound treatment,
– collider,
– antecedent,
– endogenous,
– mechanism
Notation in causal graphs: C
C = Confounder that is measured and conditioned
Thus: Uncondition when found
Notation in causal graphs: [C]
[C] = Confounder that is unmeasured and unconditioned
Thus: Condition when found
Basic principles of conditioning

Typology of Confounders: Common Cause
Has a causal effect on both X and Y Solution: Condition on the common cause confounder, thereby breaking the link

Typology of Confounders: Incidental
Affects Y and is correlated with X, but not through any identifiable casual relationship

Typology of Confounders: Compound Treatment
Researcher fails to distinguish between a causal factor of theoretical interest and a confounder

Typology of Confounders: Mechanism
A conditioned factor is endogenous to X

Typology of Confounders: Collider
A conditioned factor is affected by both X and Y

Typology of Confounders: Antecedent
A conditioned factor affects Y only thorugh X Solution: Uncondition on Antecedent confounder

Typology of Confounders: Endogeneity
Situation where Y affects X

Typology of Confounders: Mechanism sub-types
1) Front-door 2) No front-door assumptions

Three general sorts of confounders
1) Pre-treatment confounders 2) Post-treatment confounders 3) Pre-/ post treatment confounders in longitudinal studies
Types of pre-treatment confounders
– assignment, –selection, – self-selection bias
List all the strategies of causal inference beyond X and Y
Instruments help condition mechanisms against rivals and alternative causal reasonings thus making them more causally heterogenuous and robus
– Conditioning on confounders
– Intrumental variables
– Mechanisms
– Alternate outcomes
– Causal heterogeneity
– Rival hypotheses
– Robustness tests
– Causal reasoning
Strategies of causal inference beyond X and Y: Conditioning on confounders
This approach conditions factors that would otherwise confound the relationship between X and Y

Strategies of causal inference beyond X and Y: Intrumental variables
A good instrument is a variable that: 1) is highly correlated with the treatment variable (X), and 2) has no effect on the outcome (Y) except through the treatment variable (X) (the exclusion restriction)

Strategies of causal inference beyond X and Y: Mechanisms and the assumptions thereof
Connection between X and Y that explains the covariational relationship
Front door approach assumptions:
– M is the only pathway between X and Y
– The components of M are isolated and measurable
– Any confounders (C) affecting X do not affect M

Strategies of causal inference beyond X and Y: Alternate outcomes
Focuses on variation across outcomes, instead of across groups or time.
- Placebo test*: Investigates alternative outcomes that a confounder should have affected, if an effect is noted the X-Y relationship is spurious
- Unconfounded outcome: T*ry and identify an alternative outcome (Y2), that correlates with Y1, but is free from confounders
- Within-unit:* Same group is divided into treatment and control sub-groups. Different outcomes, if independent, show the treatment effect

Strategies of causal inference beyond X and Y: Causal heterogeneity
When causal heterogeneity is not stochastic (random) the treatment effect can be measured through moderators (Z)
Assumption: The interaction between X and Z (X*Z) must not be influenced by confounders

Strategies of causal inference beyond X and Y: Rival hypotheses
As the name implies, strategy where instead of looking at X, the researcher looks at possible alternative causes (Z) of Y
Logic of elimination: If I cannot find any other explanation for variation in Y, there must be some truth in the X-Y link
Critique: By definition never conclusive
Strategies of causal inference beyond X and Y: Robustness tests, definition and purpose
Can be defined as any alteration of a benchmark model that test (qualitatively or quantitatively) the plausibility of key assumptions related to study findings
Purpose: Create an estimate of the range of variation possible – If the results are very robust (X=>Y), our confidence in the finding increases
Strategies of causal inference beyond X and Y: Causal reasoning
Is essentially a counterfactual though-experiment, where one thinks through the assumptions of the causal inference and especially the DGP
What does conditioning mean?
To include a factor within a statistical model or disaggregate it into its individual parts and then hold it constant when one wants to control the variable
What do robustness tests create alterations in?
Alterations in the research design:
– Opertionalization of key variables
– Sampling
– Strategies for measuring causal effects
– Estimators
– Specifications
How do you creaete a good counterfactual? (criteria)
- Clarity
- Plausibility of the antecedent
- Conditional plausibility of the consequent
- Projectability