Not too long ago, I posted
Dani Rodrik's response to my
challenge to tell us what was wrong with estimates of the gains from trade, and I used a 2006 paper by Scott Bradford, Paul Grieco, and Gary
Hufbauer, "The Payoff to America from Globalisation," cited in a recent Ben Bernanke
speech as an example. Dani claimed that the gains the authors estimate overstate the true gains due to some of the modeling choices the authors make.
In the paper, the authors estimate the gains from trade raise incomes anywhere from $4,000 to $12,000 per household, but Rodrik raised several points challenging these estimates as too high.
At the end of the post, I promised to follow-up soon and as part of that I sent an email to the authors of the study offering them a chance to respond. Paul Grieco has has taken me up on the offer and his remarks are below. As he notes, an additional follow-up is expected from the other authors and I will post that when it becomes available:
Paul Grieco: First, let me say that I find Dani Rodrik's work interesting and compelling and I thank him for taking the time to comment on our paper. I also want to be very clear that these are my remarks alone, they do not necessarily represent the views of my coauthors or the IIE (now the Peterson Institute). They will be giving a more complete response on the Peterson website in a few days.
Rodrik makes the very good point that the results that come out of any empirical research are heavily dependent on the assumptions that go in. In producing the numbers in the papers we make lots of assumptions. We relied heavily on the work of others in pulling together estimates from a variety of sources. We tried to document our sources and our assumptions as carefully as possible. Many of the points which Rodrik brings up are actually explicitly addressed in the text. We believe the assumptions that we made were reasonable, although clearly others can disagree, I hope your readers will look at the paper and see for themselves.
I think Rodrik's critique fails to consider his own assumptions. Namely, he argues our numbers are too large, but fails to mention that his alternatives assume (rather than show) that a lot of the effects we try to capture don't exist (such as improved product variety, technological spillovers, and import competition, see box 2.1, page 66-67 of our paper). For example let's look at the back of the envelope number. He calculates the gains to be 0.25% of GDP compared to our lowest estimate of 3.4% of GDP.
Rodrik presents this as a ballpark estimate, but what does it assume? Among other things, it assumes that foreign and domestic goods are perfect substitutes, and that the US charges uniform tariffs on all imports. We all know that the US tariff schedule is far from uniform, instead some goods are have prohibitive tariffs, while others are tariff-free. This in and of itself can generate distortions that are much worse than what an "equivalent" uniform average tariff would produce. Perfect substitutability assumes away any gains from variety (among other things) and removes the possibility of meaningful import competition (since all goods are the same from the start). Neither of these assumptions seem very likely to me, and I feel fairly confident in saying they bias Rodrik's estimate downward. Given this, Rodrik's ballpark number is best understood as a lower bound on the gains from liberalization. As such, I'm not sure how it can serve as a reasonable reality check. In our paper, we explicitly say we are trying to account for gains which this number rules out by assumption...so it's not much of a surprise that we get a bigger number.
Next there is the issue of CGE models. In our paper we do cite several CGE models (including one based on the GTAP model from Purdue), not just the one from the University of Michigan. Of course all of these models are built on a large number of assumptions. Having read the work, it is my belief that all of these researches take a lot of care in creating and documenting their models. The real learning comes from understanding how estimates change with changing assumptions. Rodrick seems to imply that a constant returns to scale model is a more realistic assumption, but I'm not sure why, since the world is full of examples of increasing returns to scale. For the record, Rodrick's preferred model omits services liberalization. Services are a huge part of the US economy today, so ignoring this sector must cause a downward bias.
My last comment is really just a technical issue. Rodrick states that, "Rose says 'belonging to a regional trade agreement raises bilateral trade by (exp(1.17)-1»)222%,' whereas Bradford et al. use 118%, without explanation."
Our paper is online here: http://www.petersoninstitute.org/publications/chapters_preview/3802/2iie3802.pdf, the sentence Rodrick is referring to is footnoted (footnote 60, page93):
The dependent variable was the natural log of bilateral trade, the regional FTA coefficient controlling for fixed country-pair effects is 0.78, 100*(exp(0.78)-1)= 118 percent. This is the smallest regional FTA coefficient of the three reported benchmark regressions.
I assume that Rodrick was looking at a different version of the paper where this footnote does not appear for some reason. Nonetheless I wanted to set the record straight.
If you look at table 1 of Rose (2003, http://faculty.haas.berkeley.edu/arose/Comparer.pdf), the .78 estimate is right in the table, just next to the 1.17 coefficient. The reasons we use .78 over 1.17 are pretty simple. It seems reasonable to us that country-pairings that naturally trade a lot will be more likely to form an FTA, and this would imply that the OLS and random effects estimators will be biased upwards, while the fixed effects estimator won't have this problem. Therefore, it's best to use the fixed effects estimator. It seems a little odd that here Rodrick, who says we are trying to inflate numbers, is now criticizing us for not using one we believe is biased upwards.
Update: Dani Rodrik responds (briefly).