We can also use a number of additional specifications to explore whether selection into the sample is a source of bias in the primary results, focusing on the results for exports and output; these results are reported in table 2. In panel A, we restrict the sample to county-years that report export data. In panel B, we include for each variable only the subset of counties that reports at least eight observations for that variable, to avoid bias due to the entry and exit of counties from the sample. In panel C, we characterize each county and each variable as to whether the number of observations is above or below the median number of observations for that variable, and interact the dummy variables for a high number of observations with year fixed effects. In all three cases, the results are generally robust, though the coefficients on secondary and tertiary output are in some cases noisily estimated. The consistency across a range of specifications suggests that selection into the sample is not a significant source of bias.

Table 2.

Main Specifications Controlling for Selection into the Sample

ExportsPrimarySecondaryTertiaryGDPPer Capita
(1)(2)(3)(4)(5)(6)
A: Sample reporting export data
Post $×$ NTR gap .185 −.033 .032 .052 .067 .059
(.083)** (.014)** (.023) (.017)*** (.017)*** (.016)***
Observations 5,158 2,981 3,005 3,005 5,152 5,004
B: Sample excluding prefectures with sparse observations
Post $×$ NTR gap .196 .004 .032 .020 .038 .005
(.081)** (.018) (.014)** (.013) (.013)*** (.016)
Observations 5,117 13,479 14,388 14,112 22,989 21,187
C: Including observation number quartile - year fixed effects
Post $×$ NTR gap .188 .001 .030 .022 .030 .036
(.083)** (.016) (.015)** (.014) (.012)** (.016)**
Observations 5,158 14,722 15,688 15,375 29,782 26,333
ExportsPrimarySecondaryTertiaryGDPPer Capita
(1)(2)(3)(4)(5)(6)
A: Sample reporting export data
Post $×$ NTR gap .185 −.033 .032 .052 .067 .059
(.083)** (.014)** (.023) (.017)*** (.017)*** (.016)***
Observations 5,158 2,981 3,005 3,005 5,152 5,004
B: Sample excluding prefectures with sparse observations
Post $×$ NTR gap .196 .004 .032 .020 .038 .005
(.081)** (.018) (.014)** (.013) (.013)*** (.016)
Observations 5,117 13,479 14,388 14,112 22,989 21,187
C: Including observation number quartile - year fixed effects
Post $×$ NTR gap .188 .001 .030 .022 .030 .036
(.083)** (.016) (.015)** (.014) (.012)** (.016)**
Observations 5,158 14,722 15,688 15,375 29,782 26,333

The base specification and the dependent variables are identical to those described in table 1. In panel A, the sample is restricted to county-year observations reporting export data. In panel B, the sample for each variable is restricted to the subset of counties that report at least eight observations for this variable. In panel C, we characterize each county by the number of observations reported for each variable and generate a dummy variable for whether the number of observations is above the median; the specification then interact between this dummy variable and year fixed effects. Significant at ***1%, **5%, and *10%.

Close Modal