R ecent work in the time-series literature has stressed the importance of testing for unit roots as well as the existence of long-run relationshipsor cointegration-between variables. 1 Since the presence or absence of each of these characteristics ultimately determines the appropriate model, failure to perform such pretesting makes spurious inferences more likely. Even with existing tools designed to identify unit roots and test for cointegration, short series, the weak power of statistical tests, and the dangers of overfitting make pretesting time-series data particularly problematic. Although recent articles have helped to identify these issues (Grant and Lebo 2016;Keele, Linn, and Webb 2016), users have been left without a straightforward solution about how to deal with such problems. 2 I propose using the autoregressive distributed lag model and associated bounds testing procedure (ARDLbounds) developed by Pesaran, Shin, and Smith (2001) as a comprehensive approach to model specification and Andrew Q. Philips is assistant professor, Department of Political Science, University of Colorado at Boulder, UCB 333, Boulder, CO 80309-0333 (andrew.philips@colorado.edu).I would like to thank Lorena Barberia, Allyson Benton, Harold Clarke, Peter Enns, Nathan Favero, Eric Guntermann, Mark Pickup, Joe Ura, B. Dan Wood, and participants of the Texas A&M methodology brownbag lunches. Special thanks go to Soren Jordan, Paul Kellstedt, and Guy D. Whitten. Despite this helpful advice, any errors and omissions remain my own.1 Covariance stationary series exhibit constant mean, variance, and covariance. A linear combination of two or more first-order nonstationary series that yields a stationary series is said to be cointegrating.2 Grant and Lebo (2016) provide two solutions, including the one discussed herein. However, their discussion is brief. cointegration testing. Depending on the results of the cointegration test, this strategy absolves users from having to distinguish between stationary (henceforth I(0)) and first-order nonstationary (I(1)) regressors. This is an advantage since unit root testing is difficult in short series and introduces "a further degree of uncertainty into the analysis" (Pesaran, Shin, and Smith 2001, 289). The ARDL-bounds procedure involves the following:1. Ensuring the dependent variable is I(1). 2. Ensuring the independent variables are not explosive or higher orders of integration than I(1). 3. Estimating the ARDL model in error correction form, and ensuring there is no autocorrelation. 4. Performing the bounds test for cointegration.