Wrong propensity weight for the ATT estimator
Table 2 weights control inventors by 1/(1−p), but the ATT estimand requires p/(1−p). With low base rates the current weight evaluates to ≈ 1, so it barely upweights the cities most similar to Rochester.
Document
Feedback
Chat
The Effect of High-Tech Clusters on the Productivity of Top Inventors
Enrico Moretti · UC Berkeley · American Economic Review 111(10)
I.A. Empirical approach
For the Rochester event study I complement the baseline with a propensity-score reweighting strategy that re-weights non-Rochester cities to match the observable characteristics of Rochester before the Kodak collapse, following Abadie (2005).
Note 12. The weight for inventors who are not in Rochester is 1/(1−p), where p is estimated using a logit. Data and estimation details are reported in Moretti (2021a).
The baseline specification relates the log number of patents to the log size of the local cluster, excluding the inventor's own contribution.
What changes because of this weighting error?
What changes because of this weighting error?
This error means that one robustness check in Table 2 did not function as intended. Column 5 therefore does not actually provide the intended propensity-score balanced specification.