Working Paper: NBER ID: w21248
Authors: Melvin Stephens Jr; Takashi Unayama
Abstract: Survey non-response has risen in recent years which has increased the share of imputed and underreported values found on commonly used datasets. While this trend has been well-documented for earnings, the growth in non-response to government transfers questions has received far less attention. We demonstrate analytically that the underreporting and imputation of transfer benefits can lead to program impact estimates that are substantially overstated when using instrumental variables methods to correct for endogeneity and/or measurement error in benefit amounts. We document the importance of failing to account for these issues using two empirical examples.
Keywords: Instrumental Variables; Underreporting; Imputed Data; Program Benefits; Causal Inference
JEL Codes: C80; E21; H53; H55
Edges that are evidenced by causal inference methods are in orange, and the rest are in light blue.
| Cause | Effect |
|---|---|
| underreported and imputed data (C80) | inflated estimates of causal effects (C51) |
| underreported benefit amounts (J32) | overstated estimates of causal effect on household outcomes (H31) |
| imputed benefit amounts (J32) | overstated estimates of causal effect on household outcomes (H31) |
| lack of correlation between imputed values and instruments (C26) | inconsistency in first stage of IV estimation (C26) |
| nonreporting is random (C83) | consistent estimates when restricting to nonimputed subsamples (C20) |
| nonreporting is not random (C52) | potential methods to account for selection bias (C34) |
| imputation process fails to account for age of recipients (H55) | inflated estimates of impact on living arrangements (H31) |
| low reporting of eligible households (C83) | upward bias in IV estimates of consumption response (D11) |