A 2004 Cato Policy Report by William A. Niskanen claiming there is evidence that cutting taxes increases government spending has reemerged and made a bit of a splash lately since it runs contrary to the "Starve the Beast" hypothesis.
I hate to do this because the result that "starving the beast" does not work is useful in political battles, but after reading the paper I feel obligated to follow up whether it runs contrary to particular politics or not. The bottom line is that this paper did not deserve the attention it received.
Here is a 2004 update to the 2002 paper Niskanen mentions in his Cato report (the 2004 paper wasn't easy to find, I finally got it from the WayBack Machine). For background, Daniel Drezner provides a link to "a Jonathan Rauch column in the Atlantic Monthly (non-subscribers can click here to read the whole thing)" summarizing Niskanen's results. He also notes:
I see that this paper made the blog rounds a few years ago -- but it does not appear to have been published. Furthermore, the link to the original conference paper is not working.
The 2002 paper is “Comments,” in Jeffrey A. Frankel and Peter R. Orszag, American Economic Policy in the 1990s, Cambridge, Massachusetts: MIT Press 2002, but as Niskanen notes below the 2004 paper supersedes the 2002 paper. Here's the 2004 paper followed by comments on the methodology used to obtain the regression results:
Some Intriguing Findings About Federal Spending, by William A. Niskanen and Peter Van Doren, The Cato Institute, presented at the annual meeting of The Public Choice Society, Baltimore, Maryland, March 11-14, 2004: In two prior exploratory articles, I found ... that the federal spending share of GDP from 1981 through 2000 was a negative function of the federal revenue share... Our ... joint paper presents ... new estimates...
The “Starve the Beast” Hypothesis
For nearly three decades, many conservatives and libertarians have argued that reducing federal tax rates, in addition to increasing long-term economic growth, would reduce the growth of federal spending by “starving the beast.” This position has recently been endorsed, for example, by Nobel laureates Milton Friedman1 and Gary Becker2 in separate Wall Street Journal columns in 2003.
The problem with this hypothesis is that it is not consistent with the evidence, at least beginning in 1981. In my article published in 2002, I presented evidence that the relative level of federal spending over the period from 1981 through 2000 was coincident with the relative level of the federal tax burden in the opposite direction; in other words, there was a strong negative relation between the relative level of federal spending and tax revenues.3 Controlling for the unemployment rate, federal spending during this period increased by about one-half percent of GDP for each one percentage point decline in the relative level of federal tax revenues. What is going on? The most direct interpretation of this relation is that it represents a demand curve – that the demand for federal spending by current voters declines with the amount of this spending financed by current taxes. Future voters will bear the burden of any resulting deficit but are apparently not effectively represented by those making the current fiscal choices.
One implication of this relation is that a tax increase may be the most effective policy to reduce the relative level of federal spending. On this issue, I would be pleased to be proven wrong...
Updating the Test of the “Starve the Beast” Hypothesis
The following equation is used to test the “Starve the Beast” hypothesis:
(S/Y) = a + b(R/Y) + cU + dAR(1) + u,
S = current federal expenditures
R = current federal revenues,
Y = GDP,
U = civilian unemployment rate, and
AR(1) = a first-order autoregressive term.
Our new estimates of this equation revise and expand my prior estimate in two ways: The data for federal expenditures, federal revenues, and GDP are from the major new revision of the national income accounts.
A separate estimate of this equation is made for the sample 1949 through 1980, in order to test the stability of this relation over a longer sample. These tests would be consistent with the “Starve the Beast” hypothesis only if the estimated coefficient on (R/Y) is positive and significant.
The results from these tests are summarized by Table 1 below. Standard errors of the estimated coefficients are in parentheses.
Table 1 Tests of the “Starve the Beast” Hypothesis
Sample 1949-1980 1981-2000
Constant 18.069 30.651 (6.763) (2.001) (R/Y) - .038 - .624 (.304) (.111) U .321 .443 (.237) (.044) AR(1) .877 .429 (.049) (.135) Adjusted R-sq. .790 .943 S.E.R. .791 .275 D-W 1.629 1.503
The first conclusion from these tests is that there was no significant effect on the relative level of federal spending of either the relative level of federal revenues or the unemployment rate in the sample 1949-1980; (all of the explanatory power of this equation is in the autoregressive term.) This finding is not consistent with the “Starve the Beast” hypothesis in that the coefficient on (R/Y) is not positive and significant, but it does not suggest that a tax cut would increase spending.
The more important conclusion, consistent with my prior estimate of this equation, is that there was a very strong negative relation between the relative level of federal spending and the relative level of federal revenues in the sample 1981-2000. Controlling for the unemployment rate, federal spending during this period appeared to increase by about .6 of one percent of GDP for each one percentage point decline in the relative level of federal revenues. During the same period that conservatives and libertarians promoted the “Starve the Beast” hypothesis, the developing evidence was strongly contrary to that hypothesis.
Finally, we are left with a puzzle. Why was the relation between the relative level of federal spending and federal receipts strongly negative in the period 1981-2000 but insignificant in the period 1949-1980? What happened in federal fiscal politics that might explain the substantial difference in the estimates from these two samples? We do not know, but we suspect that the growing influence of the “supply siders,” who have a strong case that high marginal tax rates significantly reduce economic growth, undermined the influence of the traditional fiscal conservatives’ commitment to a balanced budget...
I don't want to get overly technical about the econometrics, but this regression is not convincing as it stands. For example, to name one problem, there is a simultaneity issue. The regression presumes that unemployment causes government spending but the reverse may be true as well, government spending may cause unemployment and the methodology does not account for this which can bias the results.
Also, briefly, the serial correlation properties of the model need more investigation. The paper doesn't say if the data are quarterly or annual, so I can't evaluate the DW statistic, but I think the data are annual (Durbin's h-statistic should be used anyway, or better yet, an LM test, etc.). What happens to the results if a second lagged endogenous variable is added to the regression as it appears may be needed? As for the sample, if the data are annual then the sample size is relatively small, just 20 observations.
Throwing out a few more questions, what if R/Y is lagged to allow time for political pressure to build before spending is changed? Do the results still hold? What if the data are lagged? What is logs are taken? Are the results robust to the use of other scale measures for aggregate activity besides the unemployment rate? What if spending is broken into different components to isolate components that might be under more legislative discretion? Should other variables such as the interest rate be included to capture interest payments on the debt? Why is the sample broken at 1981? What are the results for the full sample?
There are many more questions and until there is a much more thorough and complete investigation of this issue, I don't think these results should be taken very seriously.