Bernard Berelson wrote a memoir about Human Behavior: An Inventory of Scientific Findings, a book that brought together 1,003 findings about human behavior with some reasonable supporting evidence. Berelson offered a threefold grand summary of all social science research: (1) Some do, some don't. (2) The differences aren't very great. (3) It's more complicated than that.

A type 3 study, my 1975 Review article suggested that U.S. congressional midterm elections were more complicated than the then-prevailing view: a pendulum swing against the incumbent president party's caused by interacting turnout and preference effects unique to midterm elections. My idea was that the magnitude of the swing in the national congressional vote against the White House party could be explained historically—and also predicted—by national macroeconomic conditions prior to the election and by the incumbent president's approval rating. The model fit well enough that it was possible to back-calculate, say, William Howard Taft's presidential approval rating, absent the existence of Gallup Polls then, for October 1910 on the basis of the 1910 midterm election vote and 1910 economic data. I decided not to mention such antic possibilities in the Review.

My work politely fit into referendum models of elections and in particular Gerald Kramer's fine study of economic conditions and presidential elections. The midterms research grew out of my earlier 1973 Review article on seats and votes, which described how nationally aggregated preferences of voters translated into congressional seats. That translation had decayed so that by the 1970s, changes in the national party vote for congressional candidates produced small and biased changes in the party composition of the House of Representatives. This grotesquely undemocratic arrangement has worsened since, delinking changes in voter preferences from changes in party composition of the House for most congressional elections.

The midterm model was extended in my 1978 book Political Control of the Economy to on-year congressional and presidential elections. That book dealt largely with how politics affected macroeconomic conditions via the electoral–economic cycle and partisan preferences in economic policies. Put together, the two Review articles, Political Control of the Economy, and some models in my Data Analysis for Politics and Policy (1974) measured how economic and political factors translated into economic policies, the votes of citizens, and who got elected.

In addition to providing a quantitative description of some causal factors in elections, this work encouraged validation by honest prediction (now a thriving industry) rather than by after-the-fact statistical tests.

As the literature on political economy grew in the late 1970s, I developed doubts about multiple regressions on historical data. For all 17 studies of elections and economic conditions published by 1978, I constructed distributions of published statistics (Figure 1), such as published values of R² and t-statistics (as opposed to hypothetical distributions under ridiculous null hypotheses). Because competing researchers on a topic seek to publish decisive results, the meta-analytic result is a peculiar distribution of published test statistics in the literature on a particular topic: heapings that strongly favor particular points of view and few results in the dreaded Zone of Boredom, Ambiguity, and Unpublishability (ZBAU). For t-tests on regression coefficients, for example, ZBAU values fall between 1.6 and 2.0. The 17 then-published multiple regression studies of economic conditions and elections contested whether the national economy affected election outcomes (readers of a certain age will recall the divergent decisiveness of Kramer versus Stigler). Sure enough, the distribution of published t-values reveals a thinness in the ZBAU region, a result replicated for studies of economic conditions and presidential approval ratings, and for the distribution of published values of R².

Values of Published t-Statistics for 248 Regression Coefficients in All 17 Then-Published Studies of Economic Conditions and Election Outcomes.

As it turned out, at age 35, I had completed my work on political economy. I had expected to study the microeconomic side (tentative book title: Economic Control of Politics). Instead, returning to my quantitative roots and seeking to produce forever knowledge, I came to write four books on visual and statistical evidence: The Visual Display of Quantitative Information (1983, 2001), Envisioning Information (1990), Visual Explanations (1997), and Beautiful Evidence (2006).

As a Stanford sophomore in 1962, under the wonderful influence of Kenneth Prewitt and Raymond Wolfinger, I first subscribed to the Review. Perhaps even then I expected to publish here. But I never imagined that I would write something for the Review in the nostalgic first-person singular. I am grateful for the opportunity.