On Tue, 12 Oct 2010, Marcin B�^Bażejowski wrote: [ script that compares adf results with and without the --gls option ] Fixed in CVS. The problem was related to something Sven mentioned lately, namely the "reach-back" to pre-sample values. This was definitely too liberal with adf --gls, for the regression that estimates the GLS parameters. One thing to note is that your commands to generate an AR(1) process were not optimal: series y = 10 series u = normal(0,1) genr y = 0.1*y(-1) + u This is rather slow, and requires a "lead in" period before the steady state is attained (which is what tripped up adf). You're better using the filter() function: series u = normal(0,1) series y = filter(u, 1, 0.1, 10) Allin