Theory Building and Causal Inference in the Social World
From Concepts to Causal Claims
Theory as a Testable Structure
A theory in social science is a structured set of concepts, assumptions, and mechanisms that implies observable implications. Strong theory does more than correlate variables: it predicts when and where effects should change, what mediates them, and which boundary conditions limit generalization.
A Useful Rule
If a theory cannot distinguish between competing observations, it is too vague to guide research design or cumulative science.
Core Causal Concepts
Identification and Estimation
Modern causal inference distinguishes identification from estimation. Identification asks whether a causal effect is recoverable from assumptions and data; estimation asks how to compute it with finite samples. This distinction is central to regression discontinuity, difference-in-differences, instrumental variables, matching, synthetic control, and randomized experiments.
Common Causal Designs
Randomized controlled trial
- Strong internal validity under compliance
- May have limited external validity
- Often expensive or ethically constrained
Observational design
- Broader real-world scope
- Requires stronger modeling assumptions
- Highly vulnerable to confounding and selection
What does identification primarily concern?
Identification asks whether causal effects can be derived from a design and assumptions, not merely whether a dataset is large.
Correct answer: Whether a causal effect is recoverable under explicit assumptions
Why is 'endogeneity' a problem for causal inference?
Endogeneity often arises from omitted variables, reverse causality, measurement error, or selection.
Correct answer: Because the explanatory variable is correlated with unobserved determinants of the outcome, biasing causal estimates.
Frontier Issues
Current research increasingly focuses on treatment effect heterogeneity, causal inference with interference and spillovers, transportability across populations, and the limits of standard assumptions in networked or dynamic settings. A major open problem is how to learn credible causal mechanisms when the social environment adapts to the intervention itself.
Which design is especially associated with a discontinuity in assignment near a threshold?
Regression discontinuity leverages a cutoff rule to estimate local causal effects around the threshold.
Correct answer: Regression discontinuity design