Journal of Economic and Social Measurement

This paper introduces a special volume of the Journal of Economic and Social Measurement on the development and evaluation of computer software for econometric applications. It describes at least certain aspects of the history of the development of this software during the past approximately 50 years, beginning with the first use of the programmable electronic computer by economists in the early 1950s. It considers the various types of software developed, ranging from packages that permit the user to select from a particular set of options, to those that form essentially an econometric modeling language, to those that offer econometric programming capabilities, potentially allowing the individual specification of the characteristics and properties of the operations performed. This paper provides a relatively extensive list of references and is supplemented by a separate compendium of existing econometric software packages.

The electronic computer was first created in the 1940s, but was expensive to purchase and slow to spread. Throughout the 1950s and even into the earlier 1960s, electromechanical desk calculators were generally the only means whereby economists could compute regression parameters or solve econometric models. Some of the trials and tribulations of econometric computing by hand are described in this brief memoir of the context in which the Klein-Goldberger model was developed and used.

Economists were among the earliest users of the programmable electronic computer, when it was made available for general use in the early 1950s. The first operating programmable electronic computer was the EDSAC, developed at Cambridge University. This memoir describes the context of the development of the first econometric software package, a regression program developed by the author, working together with M.J. Farrell, as well as a number of other aspects of the early computing environment at the Department of Applied Economics at Cambridge.

This brief memoir describes the context of the original development of the well-known Holt-Winters exponentially weighted average forecast methodology, both in terms of its concepts and its first computer implementation in 1960. The original motivation for the development of the methodology by Holt was the widespread need for a feasible technique that could be applied to the forecasting of sales on a product by product basis. Once introduced, the method has come to be widely applied as a practical technique. The Holt-Winters methodology also played an important part in stimulating the work of J.F. Muth on what came to be called rational expectations. Notwithstanding this implact, the methodology's theoretical basis has not before now been described in the literature, however as a result of the development of this memoir for this special issue of the Journal of Economic and Social Measurement on the history and evaluation of econometric software, by arrangement the original theory paper will now be published in a forthcoming issue of the International Journal of Forecasting.

Over the period 1964–1970, graduate students of Rex Bergstrom at the University of Auckland in New Zealand developed econometric computer programs to conduct estimation and inference for dynamic linear models, dynamic simultaneous equation models and systems of linear stochastic differential equations. Linear and nonlinear multiple equation regression programs were written in Fortran; and simulation experiments and empirical applications were performed on two small IBM mainframe computers.

This paper describes the development of algorithms to solve large-scale macroeconometric models. The bulk of this paper was written in 1970 when the problem of finding solutions for these large-scale models was still an issue. During the 1960s, large powerful (for that time period) computers became available for use by economists who proceeded to use this powerful new tool to develop and apply econometric software to the problem of estimating large-scale macroeconometric models such as the Wharton Quarterly Model and the Brookings Model. Having succeeded in estimating these models, the next problem was to find a way to solve them either in the context of simulation experiments or forecasting exercises. During the late 1960s, the author investigated the problem of solving these large-scale models and found several acceptable methods which are presented in this paper.

This article considers certain salient aspects of the development of software specifically created in order to support the construction, maintenance and use of macroeconometric models. As a type, the individual packages are commonly classified as econometric modeling languages. This nomenclature reflects that the command structures used to operate this software are language-like in many of their characteristics, particularly in the case of those commands employed to express a model's relationships, state variable transformations, and perform other essentially mathematical or statistical operations. The focus of this article is both the evolution of language structure, as an expression of the process of creating and using econometric models, and the concomitant increased degree of functional integration that has accompanied it.

The SIMSYS simulation system was first developed at the Economic Council of Canada in 1970/71 to support its development of the large-scale CANDIDE econometric model. The size of the model raised many concerns, particularly about the "correctness" of the model implementation, equation ordering and run-time performance. This article describes the modeling language and coding scheme used in SIMSYS to address these issues.

The Cambridge Growth Project (CGP) 1960–1987 developed its own software packages for (1) single-equation then matrix-equation regression analysis and (2) matrix manipulation, then full dynamic non-linear model solution. The hardware used by the project was the successor to EDSAC, the Titan computer, succeeded by an IBM mainframe, and one of the first personal computers, built by Acorn a local Cambridge company, which was replaced by the standard personal computer running under DOS. The development of the matrix approach to modelling stemmed from the emphasis on social accounting associated with the work of Richard Stone. The software was developed around the concept of the zero-, one-, two- or three-dimensional tensor (scalar, vector, matrix, or 3D tensor) a powerful way of organizing data that are both cross-section and time-series. The programs, written in FORTRAN, that were developed over the period 1973–1978 MREG and IDIOM are still in use today, although they have gone through many versions, they are able to manage much larger data sets, and they have become much faster in operation.

It is common that the discovery or development of a statistical finding precedes its implementation, but in certain cases this is not true. The development of the Theil BLUS technique is a specific example of a two-way relationship between econometric theory and software design. In particular, the fact that the sum of squares of the N-K BLUS residuals is equal to the sum of squares of the N OLS residuals was first observed experimentally and subsequently proved in theory.

The development of programming languages as a means of communicating with digital computers is surveyed with emphasis on those languages and characteristics of particular relevance to economists and statisticians. There are several different types of software: (1) True programming languages which are used for many purposes in communicating with the machine, such as FORTRAN. (2) Software with greater specificity in purpose and function such as MATHEMATICA, GAUSS, MATLAB and MAPLE. (3) "Libraries" or collections of special purpose algorithms designed to be used within programs written in one of the major languages, e.g., the LINPACK, MINPACK, EISPAC, and IMSL libraries. (4) Collections of programs, sometimes called packages, linked together and designed for special purposes. These generally must be used within some of the languages one step removed from the basics. Examples include MATLAB Toolboxes. (5) Finally, there are programs which although made generally available cannot be used stand-alone in any sense.

In the last 30 years, changes in operating systems, computer hardware, compiler technology and the needs of research in applied econometrics have all influenced econometric software development and the environment of statistical computing. The evolution of various representative software systems, including B34S® developed by the author, are used to illustrate differences in software design and the interrelation of a number of factors that influenced these choices. A list of desired econometric software features, software design goals and econometric programming language characteristics are suggested. It is stressed that there is no one "ideal" software system that will work effectively in all situations. System integration of statistical software provides a means by which capability can be leveraged.

The Wilkinson Tests, entry-level tests for assessing the numerical accuracy of statistical computations, have been applied to statistical software packages. Some software developers, having failed these tests, have corrected deficiencies in subsequent versions. Thus these tests have had a meliorative impact on the state of statistical software. These same tests are applied to several econometrics packages. Many deficiencies are noted.

This paper compares alternative econometric packages for the estimation of panel data. Concern has recently been expressed in the literature about the lack of attention paid to numerical accuracy in econometric software. We compare the numerical values of the estimates of three of the most popular econometric packages featuring built-in panel data estimation algorithms: LIMDEP, STATA and TSP. The empirical application consists in looking at the determinants of bank interest rates on deposits in Italy using a data set with more than 10,000 observations. Different specifications have been estimated using fixed- and random-effects models. Fixed- effects estimates are equal in the three packages. By contrast, small differences appear for the random-effects estimators. We discuss the reasons for which the estimators differ and compare the relative efficiency of the random-effects algorithms provided by the three packages using a set of suitably designed Monte Carlo simulations.

This paper illustrates the dangers of relying on just one software system to solve what appears to be a fairly routine probit model. It is shown that quite different results can be obtained using various well known software systems. In this example, at a superficial level of analysis, the differences were traced to the default convergence tolerances that are built into the statistical software. At a deeper level of analysis, the effect of the convergence tolerance on the solution indicates that the maximum likelihood estimate does not exist due to a formerly undetected quasi-complete separation problem. Using a Multivariate Adaptive Regression Splines (MARS) model to relax the assumption of constant coefficients, further implications are drawn from this dataset.

This paper examines the issue of modelling languages as software tools for the construction and analysis of economic models. Modelling languages are intermediate level software tools that fit between the conventional commercial programming languages (such as C++ and Java) and the higher level applications packages (such as specific econometric packages). They try to blend the advantages of both the higher level and lower level tools. This paper examines these languages for use in the construction of economic models. Specifically, it examines the issue of whether such a language is the appropriate software tool for an economic modeller. It uses the modelling language MATLAB together with a number of illustrative examples to examine the use of these software tools with dynamic economic models.

This paper addresses the need for an efficient standard data exchange and transfer format for economic and social time series data that provides not only numeric observations but also sufficient documentation for users to be able easily to understand their properties. The paper proposes that the goal in the development of such a standard should be to permit these data to be downloaded into a user's environment in such a way as to be immediately usable without the need for local editing. It is argued that to achieve this goal will require that data producers and distributors, as well as developers of econometric and statistical software packages, honor and support such a standard. A candidate standard is identified, the TSD format, originally developed in 1984 and used since by a number of organizations and econometric software packages. However, the argument is made that to support future needs in the modern Internet-based world it will be necessary to extend this format, in the process developing evolving data exchange protocols for both downloaded byte stream data, based upon XML, and file transfers, based upon some extension of the TSD format.

This compendium of existing econometric software packages is intended to be a full census of all existing packages of this type that are publicly available and used by economists today. It is intended to provide the developer's view of each package: in all but a few instances, the person providing each entry is the primary author or current principal developer of the package. These program descriptions generally provide both a development history and a statement of current capabilities. Most entries also include references to more detailed descriptions, including manuals and other publications, as well as providing a website URL.