The problem of unit roots has been the centre of attention of many works on theoretical and applied statistics and econometrics. This subject has its own intrinsic interest because of the various implications on the properties of the tests and estimators. For the applied worker, establishing the presence of unit roots remains a keystone in the analysis of possible cointegrated systems: firstly, as a preliminary tool to test the integration of the variables of interest and secondly as a cointegration test on its own. On this work, we start with the model yt =ρ yt-1 +et, t = 1,2, …,n, where yo=0, the variables e1,…,en are independent and identically distributed (iid) N(0,σ2), ¦ρ¦≤1 and we will study approximations to the distribution of the autocorrelation coefficient as an estimator of ρ. Firstly we develop the general case for all ρ provided that ¦ρ¦≤1 and then we focus our attention on the non stationary case under the presence of unit roots, that is when ρ = 1. The approximations are obtained using the expansions developed by Francis Ysidro Edgeworth in 1904 to approximate the distributions of estimators. We present an analysis of the results using comparisons between Monte Carlo almulations and our approximations. Different authors established that Edgeworth approximations are not accurate enough on practical situations. As a matter of fact, Edgeworth approximations to the density function can give negative values on the tails of the distributions. But the advantage of this kind of approximations is that they produce analytical results (in particular, formulas) that are easy to handle and interpret and very suitable when we want to compare them with other methods. Our conclusions are similar to these.