Research Design, Causality, and Inference
From Questions to Identification
Causal inference
Causal inference asks not only whether two variables are associated, but whether changing one would change the other under a specified counterfactual framework. Advanced research design therefore focuses on identification, which is the logic by which causal effects are separated from confounding.
Key threats to validity
Common causal designs
| Design | Strength | Typical limitation |
|---|---|---|
| Randomized controlled trial | High internal validity | Ethical and practical constraints |
| Difference-in-differences | Useful with policy shocks | Parallel trends assumption |
| Instrumental variables | Can address endogeneity | Hard-to-justify exclusion restriction |
| Regression discontinuity | Strong local causal claims | Limited external validity |
Identification first
A sophisticated empirical project begins with a causal estimand, then asks what assumptions are needed to identify it from the available data.
In causal inference, what is the main purpose of an instrumental variable?
An instrument provides variation in the treatment that is plausibly unrelated to unobserved confounders, enabling causal estimation under strong assumptions.
Correct answer: To isolate exogenous variation in the treatment
What does the parallel trends assumption mean in difference-in-differences?
This assumption supports the interpretation that post-treatment divergence is attributable to the intervention.
Correct answer: Absent treatment, the treated and control groups would have followed similar outcome trends over time.
Counterfactual logic
The counterfactual framework formalizes causality by comparing what happened with what would have happened under a different exposure state. Because the counterfactual is unobserved, credible designs approximate it using randomization, quasi-experimental variation, or model-based assumptions.
Which design estimates a local treatment effect around a cutoff?
Regression discontinuity exploits a threshold assignment rule to compare units just above and below a cutoff.
Correct answer: Regression discontinuity
Why can high internal validity still coexist with limited external validity?
Precise identification in a narrow setting does not automatically justify broad generalization.
Correct answer: A design may identify a causal effect very well in one context, but that effect may not generalize to other populations or settings.