Finite sample effects of pure seasonal mean shifts on Dickey-Fuller tests: a simulation study

In this paper, it is demonstrated by simulation that, contrary to a widely held belief, pure seasonal mean shifts — i.e., seasonal structural breaks which affect only the seasonal cycle — really do matter for DickeyFuller long-run unit root tests. Both size and power properties are affected by such breaks but using the t-sig method for lag selection induces a stabilizing effect. Although most results are reassuring when the tsig method is used, some concern with this type of breaks cannot be disregarded. Further additional evidence on the poor performance of the t-sig method for quarterly time series in standard (no break) cases is also presented.