Do per capita health care expenditures converge among OECD countries? Evidence from unit root tests with level and trend-shifts

Abstract

This study examines the stochastic conditional convergence of per capita health care expenditures (PCHCE) among 19 OECD countries over the period 1972–2008. Specifically, newly developed LM and RALS-LM unit root tests with allowance for two endogenously determined structural breaks are employed. The results indicate support for PCHCE convergence among most OECD countries. The results are stronger in more general tests that control for two breaks and nonnormal errors. Panel unit root tests provide additional support for the stochastic convergence of PCHCE.

Keywords:

• per capita health care expenditures
• OECD
• unit roots
• trend-shifts

JEL Classification:

• I15
• C22

Notes

¹ See Hitiris and Posnett (1992); Murthy and Ukpolo (1994); Hansen and King (1996); Hitiris (1997); DiMatteo and DiMatteo (1998); Okunade and Karakus (2001); Dreger and Reimers (2005); Chou (2007); Baltagi and Moscone (2010); Potrafke (2010); Panopoulou and Pantelidis (2012), among others.

² See Murthy and Ukpolo (1994); Blomqvist and Carter (1997); Hansen and King (1996, 1998); McCoskey and Selden (1998); Gerdtham and Löthgren (2000, 2002)); Okunade and Karakus (2001); Freeman (2003); Jewell et al. (2003); Carrion-i-Silvestre (2005); Dreger and Reimers (2005); Narayan (2006); Chou (2007); Wang and Rettenmaier (2007); Baltagi and Moscone (2010); Moscone and Tosetti (2010); Wang (2011), among others.

³ Sala-i-Martin (1996) distinguishes between absolute and conditional -convergence. Absolute -convergence occurs when all countries in a sample converge to the same steady state whereas conditional -convergence occurs when each country converges to its own steady state due to explanatory variables reflecting differences across countries. However, several important studies have reported shortcomings in testing for -convergence and advocate time series approaches for stochastic conditional convergence based on unit root tests; see Evans (1996), Evans and Karras (1996) and Quah (1996), among others. Furthermore, given the different endowments of each country, the Solow (1956) model implies conditional convergence rather than absolute.

⁴ Given that the optimal AR lag is used, the F-test for the significance of breaks works reasonably well regardless of whether or not the null of a unit root is supported; see Lee et al. (2012) for the simulation results on this point. Also see Kim and Perron (2009) and Perron and Yabu (2009) who examine a related issue.

⁵ One relevant question is the reverse case where nonnormal errors are not present. In such cases, one may expect a loss of power for the RALS-based tests since the additional terms are included in the testing procedure. Im et al. (2012) report a mild loss of power and it is not significant.

⁶ Exogenous critical values are used here because the test can identify the break with 100% accuracy when the break is relatively large. Refer to Lee et al. (2012) for a detailed discussion of this issue.

⁷ Here, we consider only the significance of the trend breaks.

⁸ Results from two-break using TR-LM and TR-RALS-LM are quite consistent for most of the countries except for Canada, Iceland, New Zealand and Sweden. In the cases in which the RALS-LM unit root tests are significant but the LM unit root test insignificant (Canada and Sweden), the difference could be attributed to the LM unit root test is not recognizing the nonnormal errors whereas the RALS-LM unit root test captures the nonnormal errors. On the other hand, in the cases in which the RALS-LM unit root tests are insignificant with the LM unit root test significant (Iceland and New Zealand), the difference could be attributed to the situation in which nonnormal errors are not present. However, results from one-break using TR-LM and TR-RALS-LM are less consistent than results from two breaks when they ignore apparent additional breaks.

⁹ Note that the optimal lags given in the last column in Table 1 are used to construct the panel statistic in Equation 13. Also note that the standardized panel statistic follows a standard normal distribution (see Im et al., 2012 for details).

¹⁰ The panel statistic is based on the univariate LM tests using the optimal number of breaks.