Advanced Causal Inference quiz

Question 1

What is the Potential Outcomes Framework (Rubin Causal Model)?

Accepted Answer

The Potential Outcomes Framework is also called the Rubin Causal Model

Answer

The fundamental problem of causal inference is that all potential outcomes are observed

Answer

ATE is the average effect among the treated population only

Answer

Randomization prevents estimation of ATE

Question 2

What is the backdoor criterion for causal identification?

Accepted Answer

Backdoor criterion blocks all paths entering X from behind to remove confounding

Answer

The backdoor criterion requires conditioning on all descendants of X

Answer

Backdoor criterion blocks all paths going forward from X to Y

Answer

Backdoor criterion is only applicable in randomized experiments

Question 3

What is Instrumental Variables (IV) estimation?

Accepted Answer

IV requires the instrument to correlate with treatment (relevance)

Answer

IV estimates the Average Treatment Effect for the full population

Answer

The exclusion restriction in IV is testable from observational data

Answer

IV is only valid in randomized controlled trials

Question 4

What is Difference-in-Differences (DiD) and what assumptions does it require?

Accepted Answer

The parallel trends assumption requires treated and control groups to have the same trend without treatment

Answer

DiD requires that treated and control groups have identical pre-treatment levels

Answer

DiD is only valid when treatment is randomly assigned

Answer

The DiD estimator requires at least 3 pre-treatment periods

Question 5

What is a Synthetic Control Method?

Accepted Answer

Synthetic control constructs a counterfactual from a weighted combination of control units

Answer

Synthetic control requires at least 50 treated units for validity

Answer

Synthetic control requires parallel trends assumption like DiD

Answer

Synthetic control inference uses bootstrap standard errors

Question 6

What is the doubly robust estimator (AIPW)?

Accepted Answer

AIPW is consistent if either the propensity or outcome model is correct

Answer

AIPW requires both propensity and outcome models to be correctly specified

Answer

AIPW requires randomized data and cannot be used observationally

Answer

AIPW always produces the same estimate as simple propensity score matching

Question 7

What is the Local Average Treatment Effect (LATE) and for whom does IV identify it?

Accepted Answer

LATE identifies the average effect for compliers — units whose treatment changes with the instrument

Answer

LATE is the average treatment effect for the entire population

Answer

LATE identifies the causal effect for always-takers

Answer

LATE equals ATE when the instrument is binary

Question 8

What is the Conditional Average Treatment Effect (CATE) and how is it estimated?

Accepted Answer

CATE captures treatment effect heterogeneity across covariate values

Answer

CATE is the same as ATE and does not vary with covariates

Answer

CATE estimation requires an IV instrument

Answer

CATE can only be estimated in randomized experiments

Question 9

What is the exclusion restriction in IV and how can it be partially tested?

Accepted Answer

The Sargan-Hansen J test tests whether multiple instruments are mutually consistent

Answer

The exclusion restriction is directly testable from observational data

Answer

A strong first-stage F-statistic confirms the exclusion restriction

Answer

A weak instrument has a first-stage F > 10

Question 10

What is the Average Treatment Effect on the Treated (ATT) and when is it the right estimand?

Accepted Answer

ATT is the average causal effect among those who actually received treatment

Answer

ATT is always larger than ATE

Answer

ATT is the average treatment effect for the control group

Answer

ATT and LATE are identical when the instrument is perfectly strong

Question 11

What is collider bias (Berkson's paradox) and how does it arise?

Accepted Answer

A collider is caused by both X and Y; conditioning on it opens a spurious path

Answer

Conditioning on a collider removes confounding between X and Y

Answer

Colliders should always be included in regression models as controls

Answer

Collider bias can only occur in randomized experiments

Question 12

What is the Rubin-Holland causal model's counterfactual consistency assumption?

Accepted Answer

Consistency requires that observed outcome equals the potential outcome under received treatment

Answer

Consistency allows treatment to have different versions with different effects

Answer

Consistency is the same as the parallel trends assumption in DiD

Answer

Consistency violation means the exclusion restriction is violated

Question 13

What is a confound vs a confounder vs a mediator?

Accepted Answer

A confounder causes both treatment and outcome; a mediator lies on the causal path

Answer

A mediator should be included as a control in regression to remove its confounding

Answer

A confounder is always caused by the outcome only

Answer

Conditioning on a mediator always increases precision of ATE estimates

Question 14

What is regression adjustment (ANCOVA) and how does it improve experiment precision?

Accepted Answer

ANCOVA reduces variance by using pre-experiment covariates as controls

Answer

Including covariates in RCT regression biases the ATE estimate

Answer

ANCOVA is equivalent to CUPED only when sample sizes are very large

Answer

Regression adjustment requires propensity score estimation

Question 15

What is the fundamental problem of causal inference?

Accepted Answer

The fundamental problem is that only one potential outcome per unit is observable

Answer

The fundamental problem is that treatment effects are always zero on average

Answer

The fundamental problem is that confounders cannot be measured

Answer

The fundamental problem only applies to observational studies, not experiments

Question 16

What is the difference between internal and external validity in causal inference?

Accepted Answer

Internal validity measures freedom from confounding within the study; external validity measures generalizability

Answer

Internal validity ensures results generalize to other populations

Answer

External validity is higher in studies with strict randomization than observational studies

Answer

IV estimates have higher external validity than OLS estimates

Question 17

What is the Goodman-Bacon decomposition for TWFE DiD?

Accepted Answer

Goodman-Bacon decomposition shows TWFE is a weighted average of 2x2 DiD comparisons

Answer

TWFE estimator is unbiased under staggered adoption with heterogeneous effects

Answer

Staggered adoption always produces positive TWFE estimates

Answer

Goodman-Bacon decomposition applies only to continuous treatment variables

Question 18

What is the Angrist-Imbens-Rubin (1996) framework for IV with heterogeneous effects?

Accepted Answer

The AIR framework shows IV identifies LATE for compliers under monotonicity

Answer

Monotonicity allows both compliers and defiers in the IV framework

Answer

LATE equals ATE when treatment effects are constant across all individuals

Answer

The exclusion restriction is testable from observed data under the AIR framework

Question 19

What is the Frisch-Waugh-Lovell theorem and its importance in causal regression?

Accepted Answer

FWL theorem: the OLS coefficient on X1 equals the coefficient from regressing residualized Y on residualized X1

Answer

FWL theorem states OLS coefficient on X1 changes when X2 is added to the model

Answer

FWL theorem requires X1 and X2 to be independent

Answer

Double ML violates the FWL theorem by using ML for residualization

Test yourself