The tables below show which of the vars that we use in the main analyses still have some missing data. I also show which specific countries we are missing social registry data for because that var is causing us to drop a few countries.
The first regression uses the same specification as the one that you sent me.
The second regression adds in a few variables (log deaths as of end of 2020, ATM penetration, and social registry coverage) which seemed useful in the last set of regressions.
The third regression adds in COVID-19 spending but instruments for this variable using log deaths and change in real GDP per capita.
Overall, it seems like log deaths might be a useful addition to the regression but that the other variables (and the IV spec) don’t add much predictive value.
| Dependent variable: | |||
| Coverage | |||
| OLS | instrumental | ||
| variable | |||
| (1) | (2) | (3) | |
Pre-CovCov
|
0.643*** | 0.621*** | 0.333 |
| (0.131) | (0.136) | (0.314) | |
| rGDPg2020 | -0.008* | -0.008* | |
| (0.004) | (0.004) | ||
| digIDcov | 0.069 | 0.021 | -0.059 |
| (0.073) | (0.080) | (0.105) | |
Digital registration
|
0.235*** | 0.213*** | 0.122 |
| (0.052) | (0.056) | (0.102) | |
| FSI | -0.002 | 0.005 | -0.022 |
| (0.015) | (0.016) | (0.021) | |
EAP dummy
|
0.177*** | 0.280*** | -0.001 |
| (0.063) | (0.073) | (0.120) | |
| log_deaths | 0.038** | ||
| (0.015) | |||
| soc_reg_coverage | -0.001 | -0.0004 | |
| (0.001) | (0.001) | ||
| i_ATMs_pop | -0.00001 | 0.0001 | |
| (0.001) | (0.001) | ||
Cov spend
|
28.744 | ||
| (18.621) | |||
| Constant | 0.017 | -0.127 | 0.223 |
| (0.146) | (0.165) | (0.196) | |
| Observations | 82 | 76 | 67 |
| R2 | 0.693 | 0.739 | 0.679 |
| Adjusted R2 | 0.669 | 0.703 | 0.635 |
| Residual Std. Error | 0.182 (df = 75) | 0.171 (df = 66) | 0.189 (df = 58) |
| F Statistic | 28.227*** (df = 6; 75) | 20.762*** (df = 9; 66) | |
| Note: | p<0.1; p<0.05; p<0.01 | ||
The first figure below shows predicted versus actual coverage. I wasn’t able to see any pattern in the outliers but I could be missing something. The next few graphs are partial adjustment graphs which are sometimes useful for checking if there are any important non-linearities that we should be taking into account. (There don’t seem to be any.)