Econometrics Software

Operational econometric methods require computer software. David recog­nised this early on when writing his PhD thesis, so he wrote code in Fortran for the techniques that he was developing.

David had three reasons for developing econometrics software: to facilitate his own research, seeing as many techniques were not available in other pack­ages; to ensure that other researchers did not have the excuse of unavailability; and for teaching. Early versions of GIVE demonstrated the computability of FIML for systems with high-order vector autoregressive errors and latent- variable structures. At LSE, David and his research officer Frank Srba expanded David's initial version of AUTOREG to include new techniques, especially a rapidly expanding battery of model diagnostic (misspecification) tests.

David saw diagnostic testing as a key aspect of empirical model building, functioning in much the same way that a medical doctor would run examina­tions and tests on patients to diagnose what was troubling them. Tests for predictive failure—along with numerous other diagnostics being developed at the time—were promptly implemented in AUTOREG; see Hendry and Srba (1977, 1980).

Shortly after moving to Oxford, David ported the mainframe program GIVE to a PC-based “PcGive”, a menu-driven version initially on an IBM PC 8088 using a rudimentary MS-DOS Fortran compiler; see Hendry (1986c, 1987b). That conversion took about four years, with his research officer Adrian Neale writing graphics in Assembler. One immediate benefit was a practical, graphical implementation of recursive estimation and testing procedures—a major leap forward for analysing parameter constancy.

Jurgen Doornik then translated PcGive to C++ and implemented it as a Windows-based package with a front end (GiveWin), modules for single­equation and system estimation and testing (PcGive and PcFiml), Monte Carlo simulation (PcNaive), and specialised modules for modelling volatility, discrete choice, panels, ARFIMA, and X12ARIMA; see Doornik and Hendry (2001). Jurgen subsequently converted PcGive to his Ox language, enabling further additions by anyone writing Ox packages; see Doornik (2001).

Motivated by Hoover and Perez's (1999) results on computer-automated model selection, David and Hans-Martin Krolzig designed the PcGive-based econometrics software package PcGets, expanding on Hoover and Perez's tools for model selection; see Hendry and Krolzig (2001). PcGets's simulation prop­erties confirmed many of the earlier methodological claims about general-to- specific modelling; and, through machine learning, PcGets provided significant time-savings to the researcher, especially for large problems; see Hendry and Krolzig (1999, 2005).

PcGive embodies several important features for David, and for modellers generally. First, the software is flexible and accurate, with the latter checked by standard examples and by Monte Carlo. Second, it has rapidly incorporated new tests and estimators—sometimes before they appeared in print. Examples include Sargan’s common-factor test, the system-based tests of parameter con­stancy from Hendry (1974) and Kiviet (1986) and their recursive equivalents, the Johansen (1988) reduced-rank cointegration procedure, general-to- specific model selection, and IIS and its generalisations. Notably, other com­mercially available software packages are only starting to implement IIS, in spite of its power for detecting breaks and outliers. Third, while OxMetrics is interactive, it also generates editable batch code of user sessions, helping rep­lication and collaboration—and combining the best of both batch and inter­active worlds.

Empirical modelling still requires the economist’s value added, especially through the choice of variables and the representation of the unrestricted model. The machine-learning algorithm Autometrics confirms the advantages of good economic analysis through excluding irrelevant effects and (espe­cially) through including relevant ones. Excessive pre-simplification, as might be suggested by some economic theories, can lead to a badly misspecified general specification with no good model choice from simplification. Fortunately, little power is lost from some overspecification with orthogonal regressors, and the empirical size remains close to the nominal one.

For David, automatic model selection is a new and powerful instrument for the social sciences, akin to the introduction of the microscope in the bio­logical sciences. Already, PcGets and Autometrics have demonstrated remark­able performance across different (unknown) states of nature, with Monte Carlo data generating processes being found almost as often by commencing from a general model as from the DGP itself. Retention of relevant variables is close to the theoretical maximum, and elimination of irrelevant variables occurs at the rate set by the chosen significance level. The selected estimates have the appropriate reported standard errors, and they can be bias-corrected if desired, which also down-weights adventitiously significant coefficients. These results essentially resuscitate traditional econometrics, despite data- based selection. Peter Phillips (1996) has made great strides in the automation of model selection using a related approach; see also Haldrup, Hendry and van Dijk (2003).

4