I made some comments on Sunday about a recent critique by Thomas Herndon, Michael Ash, and Robert Pollin of an influential 2010 paper by Carmen Reinhart and Kenneth Rogoff. Yesterday Econbrowser hosted a reply from Pollin and Ash to my remarks. Here I would like to add a few further thoughts on this discussion.

Pollin and Ash first produce a graph that supports the statement I made that “At the moment, the interest rate on U.S. government debt is extremely low, so that despite our high debt load, the government’s net interest cost is currently quite reasonable.” However, they dismiss the possibility that interest rates may rise from their current low levels as merely the projections of “deficit hawks,” stating that

this pattern of low interest payments will almost certainly continue at least until the Fed alters its current monetary policy stance. Chair Bernanke has stated repeatedly that the Fed will continue with its current policy course at least until unemployment falls below 6.5 percent.

Pollin and Ash apparently did not follow the link I provided in support of my claim that many objective forecasters expect interest rates to rise. That link points to the discussion of long-term interest rates given by Ben Bernanke on March 1, in which the Fed chairman noted that predictions of rising interest rates have been made by (1) Blue Chip consensus forecast, (2) Survey of Professional Forecasters, (3) Congressional Budget Office, and (4) the Fed’s own interest rate model.

It is true that all four of these forecasts could be wrong– interest rates in 4 years could easily be higher or lower than those portrayed in the figure above. But if for example they follow the trajectory indicated by the green line, which is the baseline scenario recently constructed by theCongressional Budget Office, then net interest cost will be as big as the entire nondefense discretionary budget by 2018 and the entire defense budget by 2019.

Pollin and Ash then repeat the arguments in their paper about the desirability of treating all country-years equally without responding to the particular critique of their argument that I originally provided. The issue I raised has nothing to do with serial correlation. The issue instead is whether the expected GDP growth rate should be regarded as if it is the same number across different countries. A well-known econometric method for dealing with this is referred to as “country fixed effects.” In this method, one uses the average for the Greek observations as an estimate of the Greek growth rate and the average of the U.S. observations as an estimate of the U.S. growth rate. This is a widely used procedure. By contrast, the weighting proposed by Herndon, Ash, and Pollin assumes that the expected growth rate is the same across different countries, an approach that is less widely chosen for panel data sets and in my opinion less to be recommended. Given that the ultimate goal in this case is to infer an average effect across different countries, I personally feel that a random-effects approach would be superior to fixed-effects estimation, particularly given the unbalanced nature of the panel (that is, given the fact that we have many more observations on the 90% debt state for some countries than for others). As I noted in my original piece, this would yield an estimate that would be in between those or RR and HAP. But to suggest that there is some deep flaw in the method used by RR or obvious advantage to the alternative favored by HAP is in my opinion quite unjustified.

Finally, Pollin and Ash repeat the claim that the HAP estimates are substantially different from those in RR. Again, in doing so they have failed to respond to my original point, which was summarized in the table found here. To repeat, in Reinhart and Rogoff’s original (2010) paper there were separate analyses of three different data sets, and each data set was summarized using two different methods, the first based on means and the second based on medians. HAP had quarrels with only one of these three datasets (the postwar panel of advanced economies) and only one of the two methods (the sample means). F.F. Wiley has a helpful graph comparing the original RR summary of the postwar advanced data set using the median (blue in the graph below) with HAP’s summary of that same data set using HAP’s preferred reweighted means (in red).

And for those Econbrowser readers who have been trying to point to this 2011 commentary by Reinhart and Rogoff as evidence of some claims that the HAP reanalysis somehow overturns, please note that the numbers Reinhart and Rogoff bring up in that article refer to the sample medians (blue lines) in the graph above.

*This piece is cross-posted from Econbrowser with permission.*

But what no one seems to realize is that a significantly lower mean or median growth rate for the right-most of the four data bins (90% to infinity) is no proof whatsoever for the dominant interpretation of RR's work: that there is a nonlinear "cliff" at 90% indebtedness.

Using just four bins to analyze this data is incredibly crude and does not exploit the full information. The 90% upper bin threshold is purely arbitrary (or is it? – that was the US debt rate around 2010!). If RR wanted to make a serious case for a debt cliff, they would have had to test for a breakpoint, or use the very pretty nonparametric estimator HAP use in their figures 3 and 4, fully exploiting the information without imposing any arbitrary bin boundaries.

See my "The Reinhart-Rogoff Dragon Revisited": http://silverberg-on-meltdown-economics.blogspot….