Gerald Kramer, Steven Brams, Peter Ordeshook, Samuel Popkin, and I were MIT undergraduates interested in political science in the late 1950s. Kramer, after taking a job at Rochester under the new chairman, William Riker, spent 1966–1967 at the Cowles Foundation at Yale University studying econometrics under Marc Nerlove. He began investigating the question of whether macroeconomic fluctuations had any effect on voting behavior. After discovering, somewhat to his (and Nerlove's) surprise, that there did seem to be some effect, he{} began exploring the question in depth, developing a more formal theoretical framework and also using several different econometric approaches, which jointly led to this article.

The question had been studied by several earlier researchers, but in rather primitive fashion and with inconclusive results. Kramer wrote an extensive, thoughtful literature review. His analysis represented a giant step in methodology with respect to almost everything published previously. One innovation was that he did not just write down a regression equation with an ad hoc collection of plausible-sounding variables, but derived the equation from a formal model, based on “bounded rationality” assumptions (rationalized more fully in Kramer and Lepper 1972). The model started by recognizing that voting on the basis of the incumbent's past performance is a shortcut for voters who cannot possibly be well informed. The dependent variable was the aggregate vote share for the Republicans. The vote was represented as a sum of three parts: the party's “normal” vote, incumbency, and the error term. The incumbency part combined the effect of raw incumbency advantage and the effect of past economic performance.

A central modeling issue was how to measure performance, which Kramer saw as the difference between an observed short-term growth rate and an expected rate. He pointed out that the expected rate was unobservable; thus, the estimated effects of incumbency and performance will tend to be underestimated, the underestimate being an increasing function of expectations. He chose to focus on the House vote in the aggregate as the best measure of the overall popularity of the incumbent's party “team,” seeing it as less subject to individual candidate effects than the presidential vote.

Another important methodological innovation concerned coattails effects. Despite his focus on House elections, Kramer added an equation for presidential elections, and specified a variance-components structure for the error terms (variance-components models were an active interest of Nerlove's at the time), with an additional parameter that captured a “coattails” effect. The two-equation model contained two very strong assumptions—that the effects of the independent variables would be the same in presidential and congressional elections, and that the noncoattails part of the error term was the same for presidential elections, congressional on-year elections, and congressional midterm elections—which would be relaxed by later researchers, but which did permit maximum likelihood estimation of the coattail parameter. Kramer's basic finding was that what really mattered, accounting for about half the variance in the vote itself, was the change in real personal income during the election year. Inflation, unemployment, and coattails did not matter. The addition of a trend term for the normal vote led the model to account for 62% of the variance in the congressional vote. There was no net raw incumbency advantage.

An important implicit finding, not explicitly discussed in the article, was that only economic fluctuations over the year preceding the election had any effect; voters paid no attention to earlier years. Kramer had experimented extensively with different lags—both longer and shorter—and lag structures, using both annual and quarterly data, in an unpublished 1967 manuscript, but because none of these specifications yielded better results than the annual data with a 1-year lag, he simply reported those results, intending to publish the lag results in a separate article. But he was obliged to abandon the project after discovering that that his erstwhile Cowles colleague Ray Fair (who, according to Kramer, had been shown the unpublished 1967 manuscript), had essentially replicated his findings in a 1975 paper, based on a slightly different sample period and data set, which had been accepted for publication in an economics journal (Fair 1978).

As it turned out, a data error in the annual series used in the Review article was discovered by George Stigler (1973). On learning of the error, Kramer reran his models on the corrected data, and was relieved to find that most of his original findings held up, except that inflation now had a modest independent effect. That bobble aside, and aside from its substantive explanatory findings, the model also seemed to have some modest forecasting ability, according to Atesoglu and Congelton (1982), who used it to forecast postsample congressional elections from 1966 to 1980, and found that its predictive power “compare[d] favorably with the post-sample forecasts of various [macro]economic models,” when judged by the usual criteria. The use of such models for forecasting elections, especially presidential elections, has since become something of a mini-industry, though in recent elections the various macro forecasting models, particularly the ones that amount to single-equation variants of Kramer's, seem to have lost their predictive ability (Bartels and Zaller 2001).

Kramer's interest was primarily theoretical, in trying to model the dynamics of interactions between the economy and the party system, and he mentioned the possibility of forecasting only as an aside. But even for explanatory modeling, the decision to focus on the congressional vote as the best measure of the overall popularity of the parties, although defensible for the sample period he used, is no longer tenable, as individual congressional races have become increasingly decoupled from national and presidential trends because of incumbent perquisites such as the franking privilege and earmarks, independent and differential incumbent-challenger access to campaign finance, gerrymandering to protect incumbents, and the like. One consequence of this decoupling is that divided government has become much more common. Kramer measured incumbency by the presidency, but recognized that divided government posed a problem. His sample period covered the 68 years ending in 1964, during which only 8 of the 35 elections occurred under divided government; the next 42 years would produce 22 years of divided government. Another possible implication of this decoupling is that there may be no “normal vote”; the lagged House vote turns out to be significant (Alesina, Londregan, and Rosenthal 1995).

Kramer believes that other important structural changes since 1964 as well, in both the U.S. economy and the American political process, invalidate his model. For most of the twentieth century, the pains and gains of economic fluctuations were widely distributed: “a rising tide floats all boats,” and conversely. But this has ceased to hold in recent years, as the incomes of lower class and middle-class earners/voters have remained virtually stagnant, even during periods of respectable aggregate economic growth. Equally important is the recent emergence of “base politics” as the norm, in which the parties contest elections not by courting moderates and independents, or the representative “average” (or “median”) voter, but by concentrating on mobilizing their respective electoral bases. He thinks that a more realistic approach to modeling the dynamics of electoral–economic interactions in the present era would probably require a more disaggregated approach, perhaps as sketched out in section 3 of Kramer and Lepper 1972.

Kramer's results were questioned by survey researchers who had been unable to replicate their findings with cross-sectional survey data. Kramer (1983) addressed the question of why this was so, and whether it was possible to reconcile the aggregate- and individual-level findings. He made the important point that the cross-sectional variation that the survey researchers were working with was related to the underlying structural parameters of interest in a quite different and more tenuous way than that of the time-series variations that the macro modelers were using; and that under plausible data and parameter assumptions, the cross-sectional survey estimates would be downwardly biased, and could even have the wrong sign. Kramer (1971, 1983) thus made two seminal contributions to why we say, “It's the economy, stupid.”