This paper presents an information criterion, referred to as the unit-root information criterion (URIC), that is developed for considering model selection between models with and without a unit root. It is based on small-sample adjustments to the penalty term of the Bayesian information criterion (BIC) and is asymptotically equivalent to BIC. The development of URIC is based on maximizing the frequency of choosing a stationary or trend stationary model when there is no unit root, for given implicit sizes (the frequency of choosing a stationary or trend stationary model when there is a unit root). The implicit sizes for URIC regardless of trend are designed to follow those of BIC when there is no trend allowed. When the sample size (after adjusting for lags and first-differencing) is 100 or more and Gaussian errors are used, URIC appears to have reasonable implicit size and power properties in comparison to hypothesis testing using some traditional unit root test. A demonstration of the use of URIC with various real-world datasets is also provided.