Classical regression estimates of the determinants of the OECD health expenditures are useful for policy formulation and evaluation. However, if the underlying timeseries data are not collectively stationary in levels, the estimated parameters are faulty and can misguide health policy. Until very recently, the crucial stationarity tests were ignored in a large number of studies on international comparisons. Stationarity (ADF, Phillips-Perron, IPS heterogeneous panel) and cointegration (Engle-Granger bivariate, Johansen's multivariate) tests are conducted here using 1960-1997 health expenditures data (1998 CD ROM) of 19 OECD countries. It is found that extending the time series data length affects the order of integration and number of cointegrating vectors. However, it is arguable whether the order of integration decreases or increases as more observations are added for testing. The failure of the Johansen and Engle-Granger cointegration tests for most of the OECD countries cautions policy makers against reliance on earlier research findings that were based on unstable relationships among variables in the regression models. (This is not the case for the UK, Greece and Ireland; policy implications have been derived for the UK.) Consequently, data calibrated in growth rates may be more appropriate for investigating the long run relationships collectively in a panel of OECD health expenditure model specifications.