Working Paper: NBER ID: w16566
Authors: Joshua Angrist; Ivan Fernandez-Val
Abstract: This paper develops a covariate-based approach to the external validity of instrumental variables (IV) estimates. Assuming that differences in observed complier characteristics are what make IV estimates differ from one another and from parameters like the effect of treatment on the treated, we show how to construct estimates for new subpopulations from a given set of covariate-specific LATEs. We also develop a reweighting procedure that uses the traditional overidentification test statistic to define a population for which a given pair of IV estimates has external validity. These ideas are illustrated through a comparison of twins and sex-composition IV estimates of the effects childbearing on labor supply.
Keywords: Instrumental Variables; Local Average Treatment Effects; External Validity; Overidentification Tests
JEL Codes: C01; C13; C31; C53
Edges that are evidenced by causal inference methods are in orange, and the rest are in light blue.
| Cause | Effect |
|---|---|
| LATEs (J46) | causal effect of an instrument-induced shift in treatment (C26) |
| differences in IV estimates between instruments (C26) | differences in characteristics of the compliers (D10) |
| twins and sex-composition instruments (C36) | different estimates of labor supply consequences of childbearing (J22) |
| Abadie (2003) weighting theorem (C46) | reconcile differences in estimates (C13) |
| CEI assumption (C20) | rule out selection bias from unobserved gains (C24) |
| treatment effect heterogeneity is attributable to observable characteristics (C21) | differences in IV estimates (C26) |