Performance of unit root tests depends on several specification decisions prior to their application, e.g., whether or not to include a deterministic trend. Since there is no standard procedure for making such decisions; therefore, the practitioners routinely make several arbitrary specification decisions. In Monte Carlo studies, the design of data generating process supports these decisions, but for real data, such specification decisions are often unjustifiable and sometimes incompatible with data. We argue that the problems posed by choice of initial specification are quite complex and the existing voluminous literature on this issue treats only certain superficial aspects of this choice. Outcomes of unit root tests are very sensitive to both choice and sequencing of these arbitrary specifications. This means that we can obtain results of our choice from unit root tests by varying these specifications.