When I saw these two graphs from Jason Smith, I was immediately a little wary:

It seems that the slope of the Phillips curve in Japan has decline significantly over the last few decades.

But I thought I would do some of my own calculations, so I took the same data I was using before, except now expressed as a change from twelve months ago, and found the slope and p-values over the prior five years for each year between 1992 and 2015. The results were a little different than Jason's (which is understandable given the shorter time horizon):

Interestingly, if you ignore the periods with really high p-values (1992-1997, 2003-2007, and 2015), the Phillips curve seems to have a pretty consistent slope over the previous five years.

I'm not sure what to make of the previously mentioned periods with really high p-values except that the slope of the Phillips curve was effectively zero at those times. Basically the 2003-2007 period saw little change in inflation corresponding with consistent, high employment growth. Theoretically, I'd say this probably reflects non-cyclical changes in employment, which would mean that the Phillips curve wouldn't be seen in the data, but this explanation may not be satisfactory.

I don't know how to explain '92-'97 (which also includes the prior five years: 1987-1991), although I would note that Jason seems to have gotten a relatively high slope for this period and that the period I was most concerned with was the period directly following the great recession, in which the Phillips curve seems to have reasserted itself -- which does indeed make sense, considering the recession of 2008 was almost definitely cyclical.

My hypothesis is that Jason's slopes for the 2008-2015 period are held down by his inclusion of the Koizumi boom.

Note the high p-value regions are also regions where the estimated slope is close to zero. This is no accident. The p-value hypothesis test is for a slope that is non-zero. You both found a close to zero slope and the p-value test says you can't reject the null that it is zero.

ReplyDeleteI think part of the problem is that inflation is extremely noisy, so perhaps filtering is in order.

DeleteInteresting John. My slope for Dec 2003 to Dec 2013 is 0.77 with a p-value of 1.1e-5. 0.77 seems roughly consistent with averaging your 2014 and 2009 slopes together (although that might not be very valid to do!). Same goes for the log of my p-value actually.

ReplyDelete