The State of Quantitative Political Methodology
Larry M. Bartels and Henry E. Brady
In the first edition of this volume, Achen ( 1983a 69) complained that "political methodology has so far failed to make serious theoretical progress on any of the major issues facing it." In the intervening decade, thanks in no small part to Achen's prodding, the field of political methodology has progressed significantly on several fronts. While political methodologists have still "done nothing remotely comparable" to the invention of factor analysis by psychometricians or structural equation methods by econometricians ( Achen 1983a, 69), they have invented, adopted, or further developed an impressive variety of useful techniques for dealing with event counts ( King 1989d), dimensional models ( Poole and Rosenthal 1991; Brady 1985b, 1990; Enelow and Hinich 1984), pseudo-panels ( Franklin 1990), model misspecification ( Bartels 1991b), parameter variation ( Rivers 1988; Jackson 1992a), aggregated data (Achen and Shively n.d.), selection bias ( Achen 1986b), non- random measurement error ( Brady 1985a; Palmquist and Green 1992), missing data (Mebane n.d.), and time series data ( Freeman, Williams, and Lin 1989; Beck 1990).
Moreover, while the volume and sophistication of basic methodological research in political science have increased significantly in the decade since Achen surveyed the field, the impact of methodological research on empirical work throughout the discipline has probably increased even more significantly. In order to document that fact, in addition to describing recent methodological advances in a variety of areas, we point to innovative and important applications of quantitative methods in every part of political science. To that end, in addition to reviewing every issue of Political Methodology ( 1974- 1985) and the four volumes of Political Analysis dated 1989-1992 (but published in the next year), we reviewed and classified according to methodology over 2,000 articles in other political science journals, including every article published in The American Political Science Review from 1981 to 1991 and every article published in The American Journal of Political Science, Comparative Political Studies, Comparative Politics, Journal of Conflict Resolution, and International Studies Quarterly from 1984 to 1991. 1
Despite having cast so wide a net, we must emphasize that, due to inevitable limitations of space and expertise, our discussion of problems and techniques remains selective. Models and analyses based on computational, rather than statistical, logic are gaining a foothold ( Alker 1988; Schrodt 1991). Data graphics, inspired by the ideas of Tufte ( 1983) and Cleveland ( 1985) and by the power and convenience of computer graphics packages, are becoming increasingly sophisticated. These strands of research and others have added significantly to the vitality and utility of political methodology in the last decade. We omit them here not because we consider them unimportant or uninteresting, but because we prefer to focus our finite attention upon what we consider the mainstream of contemporary political methodology: the armamentarium of techniques developed to relate statistical models to quantitative data of various sorts.
There is no one way to organize the many topics touched upon in our survey of the field. We have chosen to follow a rough logical progression from data collection through modeling to estimation. We begin in Section 1 with data collection, where there has been a notable resurgence in the use of experiments, innovations in survey design, and a proliferation of events data in a variety of research areas.
Sections 2 and 3 are organized around distinctive units of analysis. Time-series research has profited greatly from the development of Box-Tiao methods, vector autoregression, cointegration and error correction models, and the Kalman filter; these developments are described in Section 2. In Section 3 we survey innovations in the use of pooled time-series cross- sectional data, panel data, and auxiliary data sets, as well as the current state of aggregate data analysis.
Another familiar distinction is among nominal, ordinal, interval, and ratio data. Special methods are necessary for dependent variables that are not interval or ratio measures; in Section 4 we discuss methods for analyzing polychotomous "categorized" data such as standard Likert opinion questions, polytomous or multiple choice data such as choice of party or candidate, and