Working Paper: NBER ID: w17816
Authors: T. Kirk White; Jerome P. Reiter; Amil Petrin
Abstract: Within-industry differences in measured plant-level productivity are large. A large literature has been devoted to explaining the causes and consequences of these differences. In the U.S. Census Bureau's manufacturing data, the Bureau imputes for missing values using methods known to result in underestimation of variability and potential bias in multivariate inferences. We present an alternative strategy for handling the missing data based on multiple imputation via sequences of classification and regression trees. We use our imputations and the Bureau's imputations to estimate within-industry productivity dispersions. The results suggest that there is more within-industry productivity dispersion than previous research has indicated. We also estimate relationships between productivity and market structure and between output prices, capital, and the probability of plant exit (controlling for productivity) based on the improved imputations. For some estimands, we find substantially different results than those based on the Census Bureau's imputations.
Keywords: productivity; imputation; manufacturing; Census Bureau
JEL Codes: C80; L11; L60
Edges that are evidenced by causal inference methods are in orange, and the rest are in light blue.
| Cause | Effect |
|---|---|
| Census Bureau's imputation methods (C80) | underestimation of within-industry productivity dispersion (L69) |
| CART-based imputations (C59) | greater productivity dispersion (D29) |
| imputed data (Y10) | estimated relationships between productivity and market structure (L11) |
| Census Bureau's imputations (C80) | negative association between plant-level prices and exit probabilities (L11) |
| CART-imputed data (Y10) | statistically insignificant relationship between plant-level prices and exit probabilities (L11) |
| imputation method (C36) | alters coefficients in regressions analyzing demand density's effect on productivity (C29) |