Persistence in Turkish Real Exchange Rates: Panel Approaches

Testing whether real exchange rates are stationary and, thereby, obtaining evidence of whether the absolute version of the purchasing power parity (PPP) hypothesis holds, have, initially, be done by using the ADF statistic to test for a unit root. Subsequently, to mitigate the low power of the ADF test, several alternatives have been used for the same purpose. Panel unit root testing is one of these alternatives. In Erlat (2003), I had previously considered two other alternatives; namely, introducing multiple structural shifts in the deterministic terms and fractional integration, in the context of the two primary bilateral Turkish real exchange rates; the $US and the German DM based rates. This investigation did indicate that these two rates may, in fact, be taken to be stationary with significant long-memory components. In the present paper, I utilise panel procedures to see if they, also, give corroborating evidence. I used monthly data for the period 1984.01-2001.06 and constructed a panel of 17 bilateral CPI-based real exchange rates corresponding to Turkey's main trading partners for which complete data were available. I implemented seven panel procedures. The first two, Levin, Lin and Chu (LLC) (2002) and Im, Pesaran and Shin (IPS) (2003) are the most commonly used procedures. LLC assumes a common coefficient for the lagged dependent variable in the autoregressions while IPS recognises the full heterogeneity of the coefficients. The third procedure utilised, Hadri (2000), also assumes full heterogeneity but has stationarity as its null hypothesis. These three procedures take account of the dependence between the series that make up the panel by subtracting the means obtained for each time period across cross sections, from the observations. On the other hand, the remaining four procedures, due to Taylor and Sarno (TS) (1998), Breuer, McNown and Wallace (BMW) (2001), Pesaran (P) (2007)and Bai and Ng (BN) (2004a) handle the problem of dependence in a somewhat more elaborate manner. TS and BMW do this by considering the autoregressions corresponding to each series as set of seemingly unrelated regressions. TS consider a joint test of a unit root while BMW consider individual tests, thereby complementing each other. P and BN, on the other hand, assume that there is a common factor in the panel of series. P adds this common factor, proxied by the time-wise mean, as a regressor to the autoregressions and performs the ADF test while BN decompose the series into this common factor and the idiosyncratic components and test for a unit root in both components, thereby enabling us to determine the source of the persistence if it exists. Of these seven procedures, LLC and IPS lead to the rejection of the null hypothesis of a unit root, while Hadri, TS and BMW do not. The LLC result has the, rather sharp, implication that all 17 series are stationary which, obviously, is not realistic. The IPS result, on the other hand, implies that, at least one series is stationary. This is corroborated by individual ADF tests for, say, the UK, Italy, France, the Netherlands and Belgium based series. The same corroboration is, however, lacking from the other panel approaches, implying that the evidence about the stationarity of the Turkish real exchange rate is mixed and not very strong if panel procedures are used alone as an alternative to univariate ADF tests. Structural shifts in the deterministic terms may need to be introduced into these procedures to obtain stronger evidence of stationarity but this is the subject of further research.