Applied Econometrics without Sinning

One must offer a cautious welcome to any paper (Kennedy, 2001) discussing the problems confronting the implementation of econometrics to empirical data, especially from someone who has written an elementary econometrics textbook (see Kennedy, 1985, and later editions). There is a considerable gulf between the elementary — and inappropriate — methodology inherent in many such textbooks, and what applied econometricians need to do in practice to model economic data. There is some good advice in his paper, but also some which is bad: Kennedy fails to formalize the issues, is unclear at key points, and neither draws essential distinctions nor defines central terms. Consequently, although some of his recommendations should help to narrow the theory-applications gap, others could exacerbate the difficulties. The remainder of this note amplifies these comments. First, the topic is in fact addressed in some textbooks, albeit none of which are cited by Kennedy: see, inter alia, Hendry and Doornik (2001) (and earlier editions, especially chs. 11 and 15); Hendry (1995a) provides extensive analyses of the problems of matching theory and evidence, empirical modelling, and data measurement issues, as well as de-emphasizing estimation, and highlighting the role of graphs; Spanos (1986) analyzes the difficulties inherent in the conventional paradigm, and clearly explains an alternative approach, now exposited in his introductory textbook, Spanos (1999); and Charemza and Deadman (1997) describe the practice of modelling non-stationary time series; also see Hendry and Wallis (1984; ch. 1) for an earlier general discussion. There are too many relevant papers on methodology to list additional to those Kennedy mentions, but Mizon (1995) and Hendry (1997) clarify some of his difficulties. Second, the theory and practice of model building has advanced spectacularly in the last few years (see, e.g. Hoover and Perez, 1999, 2000; Krolzig and Hendry, 2001; Hendry and Krolzig, 2001b). Many of the so-called ‘sins’ turn out to be false accusations, while others (formulating and checking the ingredients of the basic specification) remain essential if anything useful is to result from empirical modelling. Thus, considerable revision of the ‘Ten Econometric Commandments’