Replication of scholarly analysis depends on individual researchers being able to explain exactly what they have done. And being able to explain exactly what one has done requires keeping good records of it. This article describes basic good computing practices and offers advice for writing clear code that facilitates the task of replicating the research. The goals are simple. First, the researcher should be able to replicate his or her own work six hours later, six months later, and even six years later.

As evidenced by the thoughtful symposium participants, political scientists have a variety of opinions about verification, replication, and data archiving. I continue to have strong reservations regarding the proposed replication, verification, data relinquishment rules debated here. The symposium, however, has convinced me of one important thing: a still broader discussion is needed before editors or organizations change the norms governing research or publication. Walter Stone's (1995) proposal to survey political scientists to obtain their opinions on verification/replication is, therefore, a good one.

Receiving thoughtful comments in this symposium from 19 of the discipline's most innovative empirical researchers is gratifying and goes a long way toward bringing issues of replication and data accessibility into public view. The progress made in the preceding pages clearly demonstrates the virtues of a community of scholars willing to share sufficient information to benefit from each other's work.

Computer models hold a continuing fascination for political scientists. Part of their attraction derives from being working miniatures like mechanical toys. They also hold the less enjoyable but more scientific promise of allowing something akin to laboratory experimentation—denied to political scientists in most circumstances because of cost or the nature of the subject matter.

Spreadsheets are powerful tools for the construction of simple or complex models. At their core, spreadsheets are nothing more than cells arranged in columns and rows into which labels, numbers, or formulas can be placed. The numbers and formulas can be arranged in infinite variety, and in modern spreadsheets they are supplemented by such features as built-in formulas called functions, macro languages which are highly flexible programming environments, data query features, and graphics capabilities.

Edith Stokey's and Richard Zeckhauser's clearly explained difference equation models lend themselves especially well to rendition on a spreadsheet (1978, 47–73). Stokey and Zeckhauser furnish the case of public housing units in a city deteriorating at the rate of 10% per year to the point of being uninhabitable. If 800 units per year are built, will the total number of public housing units reach an equilibrium and, if so, will it be stable?

Such a problem is expressed in difference equation form as follows:

Unitsn = Units n − 1 – .1

* Units n − 1 + 800

or

Units n = .9 * Units n − 1 + 800

The number of housing units in any year n equals .9 times the number of housing units in the previous year plus 800. The beginning of a spreadsheet solution to this problem is shown in Figure 1.

In a recent article in this journal, James McGregor (1993) criticizes political scientists for practicing regression analysis without recognizing the limitations in the regression technique. McGregor argues that the technique is inappropriate for political science research in many cases and suggests that we pay close attention to both the substantive (e.g., linearity and additivity) and statistical aspects of the technique.

While acknowledging that many of us have not done enough to meet the statistical assumptions of regression analysis, George Krause (1994) argues that McGregor has underestimated the value of the technique to political science research. To support his argument, Krause directs our attention to a number of complicated regression models, including those designed for situations in which the dependent measure is a discrete variable (logit, probit, ordered probit, multinomial logit, Poisson, negative binomial and generalized event count) or the independent effects on the dependent variable are lagged distributed (polynomial distributed lag). He also mentions the recent developments related to event history regression, bootstrapping (a nonparametric approach to statistical inference), and vector autoregression (VAR, a creative use of OLS in simultaneous time series data). Krause deserves credit for his attempt to show us that the scope of the regression technique is not as narrow as McGregor might have assumed.

However, in rejecting McGregor's pessimism, Krause does not realize that McGregor's major concern with regression analysis, as he illuminates by applying OLS to certain laws of nature, is not about the type of dependent variable but about relationships.