"In a seminal contribution, Romer and Romer (2010) (RR henceforth) estimate GDP tax multipliers of up to -3 after 3 years. These results have been criticized as implausibly large. For instance, Favero and Giavazzi (2010) (FG henceforth) argue RR's specification cannot be interpreted as a proper (truncated) moving average representation of the output process. They show that when the system is estimated in its VAR form, or its correct truncated MA representation, a unit realization of the RR shock has much smaller effects on GDP than in RR, typically about - .5 percentage points of GDP. I argue that on theoretical grounds the discretionary component of taxation should be allowed to have different effects than the automatic response of tax revenues to macroeconomic variables; existing approaches, including FG's, that do not allow for this difference, exhibit impulse responses that are biased towards 0. I show that the correct impulse responses to a RR tax shock are about half-way between the large effects estimated by RR and the much smaller effects estimated by FG: typically, a one percentage point of GDP increase in taxes leads to a decline in GDP by about 1.5 percentage points after 3 years. I also create two new datasets of tax shocks, one based on receipts and the other on liabilities; in these datasets, I distinguish between different types of taxes (personal, corporate, indirect, and social security) and their subcomponents"--National Bureau of Economic Research web site.