The authors compute the scaled varlogram (the variances of kth differences scaled by the variance of first differences) of the log of annual per capita real aggregate output (GDP or GNP), as measured by (1) the long series for the United States and United Kingdom; (2) Angus Maddison's (1982) long series for twelve countries; and (3) the postwar IFS data for thirty-two countries. Simulations show that the scaled varlogram of real output is nearly always more consistent with the data being generated by parsimonious difference stationary than trend stationary univariate processes. In fact, the data reveal some "excess nonstationarity" relative to parsimonious ARIMA models. The power of the scaled varlogram to discriminate between trend stationary and difference stationary processes is somewhat greater than that of a Dickey-Fuller type F test. Copyright 1990 by Ohio State University Press.