Economic Behavior and the Partisan Perceptual Screen
Mary C. McGrath∗
July 17, 2015
Abstract
Partisans report different perceptions from the same set of facts. According to
the “perceptual screen” hypothesis, this difference arises because partisans perceive
different realities. An alternative hypothesis is that partisans take even fact-based
questions as an opportunity to voice support for their team. In 2009, Gerber and
Huber conducted the first behavioral test of the perceptual screen hypothesis outside
of the lab. I re-analyze Gerber and Huber’s original data and collect new data from
two additional U.S. elections and four other countries. Gerber and Huber’s finding
of a relationship between partisanship and economic behavior does not hold when
observations from a single state-year (Texas in 1996) are excluded from their analysis.
Out-of-sample replication based on the two U.S. presidential elections since the original
study similarly shows no evidence of an effect. Given these results, the balance of
evidence tips toward the conclusion that economic perceptions are not filtered through
partisanship.
∗
I would like to thank Alan Gerber and Greg Huber for making their data publicly
available for replication, and for their support of this work. Don Green provided guidance
throughout the research process. Peter Aronow, John Bullock, David Mayhew, and Pia
Raffler gave valuable feedback.
1
1
Introduction
Students of political behavior have long asserted that identification with a party alters a
voter’s perception of the world. In 1954, Berelson et al. wrote, “The mirror that the mind
holds up to nature is often distorted in accordance with the subject’s predispositions.” Campbell et al. (1960) argue that partisanship is stable not because it acts as a reliable representation of political belief, but because it shapes belief. Partisanship generates a “perceptual
screen” through which people filter information, causing adherents of opposing parties to
perceive different information from the same set of facts. Voter belief and behavior vary
by party and hold steady through life because partisans of different stripes move through
different versions of reality.
Until recently, the primary evidence in support of a perceptual screen came from what
voters said in interviews and surveys. Bartels (2002) demonstrated that Democratic and
Republican respondents to the 1988 American National Election Studies (ANES) survey reported different answers to questions of fact related to the state of the economy. Subsequent
survey-based and lab studies have found partisan differences in reasoning and interpretation of facts (e.g., Taber and Lodge 2006, Gaines et al. 2007, Kernell and Mullinix 2013)
and reported spending (Enns and Anderson 2009). Stanig (2009) documents a relationship
between party identification and assessments of the economy outside of the United States,
using survey data from 33 countries.1 And Gerber and Huber (2010) used a panel survey,
interviewing respondents immediately before and immediately after the 2006 U.S. midterm
elections to show that divergent partisan reports are not the result of changes in respondent partisanship, different criteria by which partisans evaluate the economy, or selective
perception or exposure.
1
Stanig’s sample includes the four countries from which I collected data on economic
behavior: Australia, France, Germany, and Ireland. Full analysis of the data from these
countries is presented in the Appendix.
2
But distinguishing true belief from mere talk is difficult in the self-reported data of surveys. Divergent partisan survey responses regarding questions of fact could simply be a
form of political cheerleading. Did the Democratic respondents to the 1988 ANES truly look
back over the eight years of the Reagan administration and perceive a crumbling economic
landscape, while Republican respondents saw a flourishing economy? Partisan respondents
may give different reports not because they perceive different realities, but because people responding to surveys take even fact-based questions as an opportunity to express an
opinion.2 Experiments encouraging accurate responses to fact-based questions suggest the
presence of cheerleading in survey responses: partisan differences can be reduced either by
offering monetary incentives for correct answers or by asking respondents to answer as accurately as possible for the sake of scientific validity (Bullock et al. 2013; Prior, Sood, and
Khanna 2015).
Economic behavior is a more candid expression of belief than survey response. In 2009,
Gerber and Huber drew upon economic data to conduct the first large-scale study examining partisan behavior for evidence of a perceptual screen.3 Using data on taxable sales
2
Gerber and Huber (2010) make an effort to distinguish partisan cheerleading from a
perceptual screen by looking for partisan divergence on measures of self-reported happiness
and projected vacation/holiday spending in the 2006 Cooperative Congressional Election
Survey. The authors argue that these “alternative measures appear unrelated to politics and
therefore unlikely to impel the respondents to tailor their response to their general partisan
disposition.” This argument rests on a narrow construal of partisan cheering, and of which
elements in a political survey respondents will take as related to politics. Shortly after
an election in which a respondent’s favored party suffers an unexpected blow, the partisan
may indicate displeasure with the outcome of the election in her responses to a political
survey by expressing a general “things are bad” sentiment, reporting low mood and belttightening intentions along with a shabby national economy. More to the point, given that
the outcome of an election can be expected to affect the mood of self-reported partisans
for reasons unrelated to a perceptual screen, and considering the complicated relationship
between mood and spending intentions (see, e.g., Cai, Tang, and Jia 2009; Lerner, Small, and
Loewenstein 2004; Mano 1999; Saliagas and Kellaris 1986), these measures seem unsuited to
isolate the effect of a perceptual screen.
3
Trigger (2006) conducted a similar test on a smaller scale. Gerber and Huber note that
the relationship Trigger identifies seems overly sensitive to influential data points: “[Trigger’s]
3
from roughly 1,450 counties, the authors looked at post-presidential-election change in local
consumption as a function of local partisanship. If partisans truly perceive different economic realities, the economic activity in areas with different partisan leanings should reflect
this difference in belief. For example, say that a Democratic candidate wins an election. If
partisans adjust their economic expectations and behavior according to whether their party
won, then post-election sales should increase relative to pre-election sales in more Democratic
counties, and decrease relative to pre-election sales in more Republican counties.
Gerber and Huber’s analysis of taxable sales data over four U.S. presidential elections
(1992-2004) showed just such a pattern—a positive relationship between an area’s partisanship and consumption after a party win. Partisans appeared to increase their spending when
their candidate won. This pattern looked like a behavioral manifestation of the partisan perceptual screen: partisans acting on different beliefs about the economy under either party’s
administration. Gerber and Huber’s analysis thus seemed to provide powerful new evidence
that partisans not only report different versions of reality, but make economic decisions as
if living in different versions of reality.
This paper reassesses the evidence for a perceptual screen in partisan economic behavior.
I re-analyze the original dataset and collect new data from the 2008 and 2012 U.S. presidential
elections. This new data expands the original sample from 19 states to 23, drawing an
additional 356 counties into the analysis. My analysis shows that there is no systematic
evidence of a perceptual screen in partisans’ economic behavior. Examining each of the
elections in the original dataset separately reveals a positive relationship in only one year,
1996. A positive result only in this year is difficult to reconcile with the theoretical model,
which indicates that, all else equal, 1996 should be the U.S. election least likely to exhibit an
effect. Disaggregating the data reveals that the 1996 result is driven by a strong relationship
results appear to be driven by a small number of outlier counties...” (Gerber and Huber 2009,
p. 408, fn. 1).
4
in a single state, Texas. When observations from this state-year—Texas 1996—are excluded
from analysis of the original sample, the estimated effect shrinks to one third of the result
reported in Gerber and Huber 2009. Incorporating the new data collected from the 2008
and 2012 U.S. presidential elections, the overall effect is a precisely estimated zero.
I also collected data on economic behavior in four countries outside of the United States:
Australia, France, Germany, and Ireland. Similar to the results within the United States, of
the fourteen elections observed across these four countries, a statistically significant relationship appears in only one election—Ireland 2007. Again, this finding is not easily reconciled
with the theoretical model: of all five countries in the sample, Ireland, with its weak party
identification and relatively non-ideological parties, would seem the least likely to exhibit
such a perceptual screen.4 However, the data from these countries are limited, and so the
results are more ambiguous than in the United States.
This paper is organized as follows. In Sections 2 and 3, I review key elements of Gerber
and Huber’s theory and analysis. First I outline Gerber and Huber’s theoretical model, a
version of the life-cycle/permanent-income hypothesis which indicates that, all else equal,
an effect of partisanship on economic behavior should be most pronounced in elections characterized by high uncertainty or in areas where partisans expect especially large economic
differences under each party. Next, I review Gerber and Huber’s primary specification, which
I take as the starting point for my analysis. In Section 4 I re-analyze the original dataset
and analyze new data I collected from the 2008 and 2012 U.S. presidential elections. The
Appendix contains a full analysis of the data from Australia, France, Germany, and Ireland.
4
The main Irish parties differ little in their economic ideology, with partisan division instead based on each party’s historical position towards the 1921 Anglo-Irish Treaty (Hutcheson 2011)—partisans in Australia, Germany, or France, where parties are defined largely
by their economic ideology, should have more reason to expect economic differences under
opposing parties than should partisans in Ireland. What is more, party identification is
weaker in Ireland than in Australia, Germany, or the United States (Marsh 2006; France
was not included in Marsh’s comparison of 13 countries. The relative importance of party
identification vs. ideological identification in France is addressed in the Appendix.)
5
Overall, the data analyzed here show scant evidence that partisans make economic decisions from behind a perceptual screen. Though partisan reports about the state of the
economy depend on which party is in power, partisan economic behavior does not reflect
these reports. My findings at the population level are consistent with the individual-level
experimental work mentioned above (Bullock et al. 2013; Prior, Sood, and Khanna 2015).
At both the aggregate level and the individual level, the evidence now suggests that it is
what partisans say, not what they see, that is distorted in accordance with predispositions.
2
Electoral Uncertainty and Implications for 1996
In this section I briefly sketch the main elements of Gerber and Huber’s theoretical model to
highlight three points regarding the role of electoral uncertainty in the model’s predictions.
The main point is that, at any given level (or strength) of the perceptual screen, electoral
uncertainty determines a partisan’s change in spending. When a partisan is more uncertain
about the outcome of an election, her spending will change more upon learning the election
result. If that partisan were very certain about the outcome of an election, her spending
would change very little upon learning the result. A second point to note is that even in
areas where the partisan composition would most amplifly the aggregate economic effect of
a perceptual screen (i.e., in highly partisan areas), the effect is small when the electoral
outcome is fairly certain. The final point considers the importance of uncertainty if the
strength of the perceptual screen varies from election to election. Observing an aggregate
effect of the perceptual screen that is larger in a fairly certain election than in an uncertain
election would require the perceptual screen in the certain election to be five times the
strength of that in the uncertain election.
A positive result only in 1996 is difficult to reconcile with the theoretical model. Hibbs’
“Bread and Peace” model predicted a Democratic win in the 1996 U.S. presidential election
with a probability between 94-98% (Hibbs 1996). In the leadup to the election, Bill Clinton
6
consistently led by a wide margin in public opinion polls (see, e.g., Gallup 1996). This
contrasts with the other elections covered in the sample, which all featured at least one
reversal of the lead and were characterized by much closer margins in general (Gallup 2008;
Gallup 2012). Of the six U.S. elections in the sample, partisans would be expected to have
the least uncertainty about the outcome of the election in 1996. The theoretical model would
indicate that, barring dramatic growth of the perceptual screen during 1996 as compared
to the other election years, 1996 should exhibit the least evidence of a partisan perceptual
screen.
2.1
Gerber and Huber’s Theoretical Model
Gerber and Huber base their model of the relationship between partisanship and economic
behavior on the life-cycle/permanent-income hypothesis.5 Following Hall (1978), the model
assumes that only new information can affect expectations of lifetime income. Gerber and
Huber invoke the certainty-equivalence solution, assuming quadratic utility (see Gerber and
Huber 2009, Appendix), so that the smoothing solution holds under uncertainty (see Zeldes
1989). The model indicates that, for any fixed difference in economic outlook partisans
expect under each party (i.e., for a given level of perceptual screen), an effect of partisanship
on economic behavior should be most pronounced in elections characterized by the greatest
uncertainty.
In Gerber and Huber’s model, the perceptual screen takes the form of a differential effect
of partisan control on partisans’ expectations of lifetime income. For any partisan J of party
V , the difference in expected lifetime income under a win (W = V ) and under a loss (W 6= V )
is YJ=V |W =V −YJ=V |W 6=V = DV > 0, where W indicates the winning party. Before an election,
uncertainty about the winning party causes uncertainty about Y , so that the expected income
for a Democratic partisan prior to an election is E(YD ) = pYD|D + (1 − p)YD|R , where p is the
5
For a detailed exposition, see Gerber and Huber 2009.
7
common prior probability of a Democratic win. Similarly, a Republican partisan’s expected
income prior to the election is given as E(YR ) = pYR|D + (1 − p)YR|R .
Upon learning the outcome of the election, uncertainty is resolved. In light of a Democratic win, Republican partisans—who before the election had been accounting for the possibility of YR|R , the higher lifetime income they foresaw should a Republican win—respond
to the resolution of uncertainty by adjusting their consumption downward; Democratic
partisans—who had been accounting for the eventuality of the lower lifetime income YD|R
they associated with a Republican win—correct for their uncertainty-driven downweighting
by adjusting consumption upwards.
More formally, the life-cycle consists of three periods of consumption, C1 + C2 + C3 =
E(YJ |W ), with electoral uncertainty p during period 1 resolved immediately prior to period
2. A Democrat’s optimal consumption when a Democratic candidate wins with probability
p is:
C1 = E(YD )/3,
C2 (D) = C3 (D) = C1 + (1 − p)DD /2,
(1)
and a Republican’s optimal consumption under the same Democratic win probability p is:
C1 = E(YR )/3,
C2 (D) = C3 (D) = C1 − (1 − p)DR /2.
(2)
Gerber and Huber’s model predicts that an area’s change in consumption after an election
is a function of that area’s partisan composition and the outcome of the election. Recall
from equations (1) and (2) that (1 − p)DD /2 represents a Democrat’s post-win change in
consumption, and −(1 − p)DR /2 represents a Republican’s post-loss change in consumption.
Defining the proportion of winning partisans in an area (A) as 0 ≤ αA ≤ 1, the area’s change
8
in consumption following a Democratic victory is αA [(1 − p)DD /2] − (1 − αA )[(1 − p)DR /2].
More generally, the model predicts that change in an area’s consumption after an election
occurs as a positive linear function of the area’s proportion of winning-party partisans.
∆C1→2,A = αA [(1 − p)DW =V /2] − (1 − αA )[(1 − p)DW 6=V /2]
(3)
Areas with a greater proportion of Democratic partisans should increase consumption more
after a Democratic win and decrease consumption more after a Republican win. The greater
the proportion of Republicans in an area, the more that area’s consumption should increase
after a Republican win.
2.2
The Role of Electoral Uncertainty in the Theoretical Model
The differential effect of partisan control on partisans’ perceived lifetime income—the partisan perceptual screen, DV > 0—drives divergent partisan economic behavior in response
to the election. The magnitude of the change is a function of the common prior probability of a win—that is, the magnitude of the electoral uncertainty. If the probability of a
Democratic win is 90%, a Democrat will downwardly adjust consumption by 10% of the income she expects under a Republican win, spread over each period of the life-cycle, in order
to smooth income given the possibility of a negative shock. If the Democratic candidate
wins, the Democratic partisan will then increase her spending by 10% of the difference in
income she expected under a Republican win, in order to make up for her pre-election downward adjustment. For any given level of perceptual screen, as uncertainty over the outcome
increases—that is, as the 90% and 10% probabilities both move towards 50%—the difference
between pre-election and post-election spending becomes more pronounced. All else equal,
an effect of partisanship on economic behavior should be most apparent in elections characterized by the greatest uncertainty, least apparent in elections characterized by the least
9
uncertainty.
For a sense of the extent to which uncertainty matters when there is a mix of Democrats
and Republicans in a given area, consider that the aggregate change in consumption in an
area is given by [(1 − p)/2] ∗ [αDD − (1 − α)DR ], where p is the probability of a Democratic
win, α is the proportion of Democrats in the area, and Di is the differential effect of party
win on expected income for partisans of party i. All else equal, the effect of a Democratic win
will be more pronounced in more highly Democratic areas. Taking an area where partisan
composition would allow for a large effect, say, a very Democratic county with α = .80, the
aggregate change in consumption after an election with p = .9 would be .04 ∗ DD − .01 ∗ DR .
An election with p = .5 (i.e., complete electoral uncertainty) in that same county would
result in a post-election change in consumption of .20 ∗ DD − .05 ∗ DR . That is, under near
certainty, the predicted change is only 4% of the Democratic differential minus 1% of the
Republican differential, while under complete uncertainty the predicted change is 20% of
the Democratic differential minus 5% of the Republican differential. In the case of a certain
electoral outcome (say, if a candidate were running unopposed), the model predicts no effect
of partisanship on consumption even when a partisan perceptual screen exists (DV > 0).
In other words, even if partisans are making economic decisions from behind a perceptual
screen, no difference in pre- and post-election spending would be discernible in a case of
electoral certainty.
With a given level of perceptual screen, partisans who expect a better economy when
their party wins will expect it to be better by roughly the same amount from one election to
the next. But the strength of partisans’ perceptual screen might vary across elections. To
have a sense of how much the strength of the perceptual screen would need to differ from
one election to the next in order to offset the effect of electoral uncertainty, consider again a
fairly certain election (p = .9) and a completely uncertain election (p = .5). Assuming that
the perceptual screen affects Democrats and Republicans to the same extent, but varies from
10
election to election (i.e., DD = DR = Dq , where q identifies the election), the perceptual
screen in the fairly certain election would have to be more than five times stronger than the
perceptual screen in the uncertain election in order to produce a larger effect. For example,
in the highly Democratic county described above:
.04 ∗ DD − .01 ∗ DR > .20 ∗ DD − .05 ∗ DR ⇒
.04 ∗ Dc − .01 ∗ Dc > .20 ∗ Du − .05 ∗ Du ⇒
Dc > 5 ∗ Du ,
(4)
where q = c indicates the fairly certain election, and q = u indicates the completely uncertain
election. To observe an effect of the perceptual screen in a predictable election but not in
an uncertain election, something must cause the perceptual screen to increase by a factor of
five in the predictable election relative to the toss-up.
3
Estimation
Gerber and Huber apply their model to data from the four U.S. presidential elections between 1992 and 2004 to test whether partisan economic behavior after an election suggests
evidence of a partisan perceptual screen. This section describes Gerber and Huber’s primary
specification and the measurement of each variable. The same variable definitions apply
to the new U.S. data I collect, and I take Gerber and Huber’s primary specification as the
starting point for my analysis in Section 4.
The model indicates an area’s change in consumption as the dependent variable, and
the area’s proportion of partisans as the independent variable. The dataset consists of
quarterly records of county-level taxable sales as a measure of consumption and countylevel Democratic vote share as a measure of partisanship. Gerber and Huber’s primary
11
specification is:
ln(
salesi,t+1,1stQtr.
) = B0 + B1 ∗ avg.Dem.votesharei ∗ winningpartyt
salesi,t,3rdQtr.
D.V.i,t−1 + D.V.i,t−2 + D.V.i,t−3
+ B2 ∗
+ B 0 State-year + B 0 County + e,
3
(5)
for area i in election year t.
The dependent variable, change in partisans’ consumption from C1 to C2 , is represented
by each county’s log change in taxable sales from Quarter 3 of the election year (the quarter
immediately preceding that in which the election takes place) to Quarter 1 of the following
year (the quarter immediately following that in which the election takes place). Gerber and
Huber collected taxable sales data from 26 states, a census of states with recoverable sales
tax data.6 They report that no geographic or attrition bias is apparent in which states make
data available. The primary specification includes two sample restrictions that exclude some
states and counties: (1) analysis is limited to states for which data are available during at
least one Democratic win (1992 or 1996) and one Republican win (2000 or 2004); and (2)
any county with a recognized Native American reservation is excluded from analysis due to
concerns about casino sites and on-reservation sales reporting. These restrictions reduce the
functional sample in the original analysis to the 19 states shown in Table 2 below.
To account for a county’s recent consumption history, Gerber and Huber include a threeyear lagged average of the dependent variable on the right-hand side. Quarter-3-to-nextyear-Quarter-1 log change in sales is calculated for each of the three years since the previous
election, and the average of those three inter-election values is included as a covariate for
each election-year observation.
County-level partisanship score, the independent variable, is obtained by averaging the
6
Data are obtained from the following states: AL, AR, AZ, CA, CO, FL, GA, IA, ID, IL,
KS, ME, MN, MO, NC, NE, OK, SC, SD, TN, TX, UT, VA, VT, WA, WI.
12
county’s Democratic share of the two-party presidential vote over the four elections from
1992 through 2004, and multiplying this county-level average by −1 in the two Republicanwon years (2000 and 2004). This partisanship score thus takes on just two values for each
county: for the years in which the Democratic presidential candidate wins, the partisanship
score is the county’s average Democratic proportion of the two-party presidential vote over
the elections analyzed; for the years in which the Democratic presidential candidate loses,
partisanship score is the negative of that value. In the first two periods studied, every
county is characterized by some positive fraction Democrat partisanship; in the next two
periods studied, every county is characterized by the negative of that fraction. Partisanship
measured in this manner can be thought of as a dose-response type treatment variable,
wherein each county receives a dose of partisanship of a magnitude two times its average
Democratic voteshare. Because each county’s average Democratic voteshare is multiplied
by −1 in Republican-won years, the more Republican a county is, the more positive its
2000/2004 partisanship score relative to less Republican counties. For example, compare a
county with an average Democratic voteshare of .7 (largely Democratic) to a county with an
average Democratic voteshare of .2 (largely Republican). In 2000/2004, the more-Republican
county’s partisanship score would be −.2, while the more-Democratic county’s score would
be −.7. In each election, an increase in spending by partisan winners after the election would
be marked by a positive coefficient.
Fixed effects are included for state-year and for county. State-year indicators differentiate
each state in each election year, so that, e.g., Texas-1992 observations are distinguished from
Texas-1996 observations. Inclusion of state-year indicators accounts for the effect of partisanship on consumption change that is particular, within a single election, to the collection
of counties that make up a state. Inclusion of county indicators accounts for cross-election
differences in effect particular to each county. The relationship between partisanship and
consumption change is then estimated on the basis of variation within a county over different
13
elections.7 Standard errors are clustered at the county level.
The change variable on the left-hand side paired with the dose-response-type independent variable produces a difference-in-difference-in-difference form of analysis. Taking into
account the fixed effects for state-year and county and the lagged dependent variable, this
specification asks the following question: considering a county’s 3rd-to-1st quarter change in
taxable sales relative to the average change in its state in that year and to the county’s average change over the elections observed (call this the county’s “relative-sales-change value”),
and controlling for the average value of the county’s 3rd-to-1st quarter sales change in the
three years since the last election, how does the magnitude of the county’s partisanship dose
relate to the change that occurs in its relative-sales-change value when moving from favorable
(party wins) to unfavorable (party loses) conditions?
4
Analysis
In this section, I begin by replicating Gerber and Huber’s original analysis. I also present
results from each election year of the original sample separately. I then disaggregate the
data to show an anomalous result in Texas, 1996, and present results from the original data
when the TX-1996 observations are excluded from analysis. Following my re-analysis of the
original data, I present results from the updated dataset, incorporating data from the 2008
and 2012 U.S. presidential elections. An appendix presents the results graphically, mapping
the cross-election estimates for each county, as well as analysis of the data from Australia,
France, Germany, and Ireland.
7
Referring back to the theoretical model, county (i.e., area, A) indicators stipulate that
the effect of the independent variable, partisanship (α), on the dependent variable, consumption (C2 : C1 ), is estimated within an area over time rather than across areas within
an election.
14
4.1
Replication and Re-analysis of Original Data
Table 1 replicates the original analysis in the first column, and presents estimates from each
election of the original sample separately in the next four columns. In the original analysis,
fixed effects are included for county and state-year, so that the pooled estimate is based
on variation within each county across Democratic wins (1992 and 1996) versus Democratic
losses (2000 and 2004). In the columns for an individual election year, fixed effects are
included for each state, but not for county or state-year, since only one observation is present
for each county in a given year. This means that in the individual-election specifications,
estimates are based on variation across counties within each state-year, pooling states at
each election year. When each election is estimated separately, only the 1996 election shows
a positive relationship between county partisanship and post-election change in spending.
As noted in Section 2, 1996 would seem the least likely of the U.S. elections to show evidence
of a partisan perceptual screen, as it was characterized by the least uncertainty of the six
U.S. elections in the sample.
Table 2 shows each state in the sample considered separately. Due to the restrictions
described in Section 3 above, the functional sample for the original analysis consists of 19
states comprising roughly 1,450 counties. The number of observations (counties) on which
each estimate is based is shown in the column marked “N”. In the left panel, fixed effects
are included for county and for election year, so state-level estimates are based on variation
within each county across Democratic wins versus losses. This specification follows that used
by Gerber and Huber, except pooled only to the state level rather than to the national level.
The right panel isolates the 1996 election. Here, estimates are based on variation across
counties within each state at the 1996 election, and no pooling takes place.
The left panel shows that, estimating across all four elections in the original sample, a
roughly equal number of states show positive coefficients (9 states) and negative coefficients
(10 states). In the right panel, which shows only the data from 1996, the coefficient for
15
Table 1: Regression Estimate of the Effect of Partisanship & Election Outcomes on Local
Consumption
Dependent Variable: Pre- to Post-Election Change in Taxable Sales
Pooled 1992 1996 2000
2004
Partisanship Score .030* -.002 .200* -.033
-.033
[.015] [.027] [.078] [.025]
[.047]
3-yr Avg. LDV
.144
.877
.594
.764
.813
[.042] [.040] [.182] [.050]
[.048]
Constant
.035
-.177 -.190 .001
.093
[.023] [.011] [.076] [.014]
[.024]
Observations
5426
1087 1453 1422
1464
R-squared
.811
.733
.549
.644
.655
County FE
Yes
No
No
No
No
State FE
No
Yes
Yes
Yes
Yes
State-Year FE
Yes
No
No
No
No
Robust standard errors clustered by county in brackets.
Significance marked only for Partisanship Score: *p < .05
Texas stands out dramatically, ten percentage points larger than the next largest estimate.
Without Texas, the distribution of the other states’ coefficients has a mean of .031 and a
standard deviation of .106, indicating that the Texas estimate is more than three standard
deviations higher than the mean.
The original finding is not reproduced when observations from this single state-year—
Texas 1996—are excluded from analysis. Table 3 shows estimates based on the original data,
pooling both counties and state-years as in the original analysis. The left column shows the
original estimate. The middle column shows the coefficient estimate when the Texas counties’
1996 observations are removed from the analysis, and the right column shows the estimate
when all 1996 observations are excluded. County and state-year fixed effects are included in
all three columns.
The middle column shows that the coefficient estimate based on all of the original data
except Texas 1996 is one-third the size of the original estimate, and less than its standard
16
Table 2: Estimated Effect of Partisanship & Election Outcomes in Sampled States
Coefficient Estimates from Regression of
Pre- to Post-Election Change in Taxable Sales on Partisanship Score
Across Elections: 1992-2004
1996 Election
State Coef. Std. Err.
N
Elections
State Coef. Std.Err
N
AZ
-.001
.092
12
4
AZ
n/a
3
CA
.061
.061
152
4
CA
-.059
.077
38
CO
.071
.061
240
4
CO
.031
.312
60
†
FL
.101
.051
247
4
FL
.087
.104
61
IA
-.050
.059
392
4
IA
.062
.121
98
ID
.079
.084
117
3
ID
.053
.155
39
IL
-.066
.060
306
4
IL
-.028
.073
102
KS
-.019
.051
412
4
KS
-.076
.082
103
MO
-.004
.042
345
3
MO
-.024
.066
115
NC
-.008
.056
294
3
NC
.014
.064
98
NE
-.026
.079
338
4
NE
.265
.186
84
OK
-.075*
.037
292
4
OK
-.054
.057
73
SC
.034
.041
180
4
SC
-.004
.064
45
TN
.049
.057
301
4
TN
-.033
.199
69
TX
.083*
.034
1003
4
TX .366**
.086
251
UT
-.150
.145
80
4
UT
-.123
.192
20
VA
.002
.047
535
4
VA
.014
.092
134
WA
-.073
.124
69
3
WA
.220
.224
23
WI
.078
.232
111
3
WI
.184
.337
37
†
p < .10, *p < .05, **p < .01
Robust standard errors in both panels.
The left panel includes fixed effects for county and year, std. err. clustered on county.
17
Table 3: Pooled Estimate with and without Texas, 1996
Dependent Variable: Pre- to Post-Election Change in Taxable Sales
Original Sample Excluding TX-1996 Excluding 1996
Partisanship Score
.030*
.010
.007
[.015]
[.015]
[.020]
3-yr Avg. LDV
.144
.129
.252
[.042]
[.043]
[.091]
Constant
.035
.043
-.113
[.023]
[.022]
[.047]
Observations
5426
5175
3973
R-squared
.811
.826
.833
County FE
Yes
Yes
Yes
State FE
No
No
No
State-Year FE
Yes
Yes
Yes
Robust standard errors in brackets
Significance marked only for Partisanship Score: *p < .05
error. Disaggregating the original data reveals that the original finding appears to be driven
by a strong positive relationship between partisanship and post-election spending in a single
state during a single election year, 1996. The estimate based solely on the data from this
year (.200 [.078], shown in Table 1) falls well outside the 95% confidence interval estimated
using data from the other three elections (95% CI: -.031, .046), as can be seen from the right
column of Table 3.
4.2
Analysis of Updated Dataset
Section 4.1 highlights the strong influence of data from a single state-year in Gerber and
Huber’s original estimate. But influential points do not invalidate an estimate: just because
a model is sensitive does not mean it is wrong.8 New data must be collected to learn
whether out-of-sample tests support the weight given to influential points, or reinforce the
interpretation of these points as outliers.
8
See e.g., Stigler’s critique of Kramer (1971).
18
I updated Gerber and Huber’s original dataset by collecting county-level taxable sales
records and partisanship data through the 2012 U.S. presidential election, thus expanding
the dataset from four elections to six. Adding data from the 2008 and 2012 elections also
serves to draw four new states into the analysis—Alabama, Arkansas, Georgia, and South
Dakota—which together contribute 356 counties, 25% of the original sample N . These states
were excluded from the original analysis due to the restriction that a state’s data must include
at least one Republican and one Democratic win. The original data for these states covered a
range of years with only Republican-won presidential elections. Adding the Democratic-won
elections of 2008 and 2012 allows the inclusion of these states in the updated analysis. The
new data comprises the 19 states listed in Table 2 as well as AL, AR, GA, and SD.9
For the analyses based on this updated data, I recalculate each county’s partisanship
score to cover the elections included in the updated sample. Instead of averaging the county’s
Democratic share of the two-party presidential vote over the four elections between 1992 and
2004, the updated partisanship score averages over the six elections between 1992 and 2012.
I also corrected minor data errors present in the original dataset. These changes do not
alter results from the original sample in any meaningful way. The original finding can be
reproduced from the updated dataset by reverting to the 1992-2004 partisanship score and
restricting analysis to the observations available in the original analysis.
Table 4 shows an estimate based on all six elections in the updated dataset taken together,
along with each of the six U.S. elections estimated separately—an updated version of Table
1. Estimates and sample sizes of the four original elections differ slightly from the original
estimates due to the minor data error corrections. The coefficient estimate based on the
pooled data from all six of the elections in the updated dataset is almost exactly 0, as shown
in the first column of Table 4. The 95% confidence interval for this updated estimate now
nearly excludes the original estimate of .030 shown in the first column of Table 1 (95% CI:
9
Data will be available at isps.yale.edu/research on publication.
19
Table 4: Updated Sample
Pre- to Post-Election Change in Taxable Sales
Pooled 1992 1996 2000 2004 2008
Partisanship Score
.001
.014 .201* -.016 -.002 -.008
[.017] [.031] [.077] [.025] [.038] [.070]
3-yr Avg. LDV
.135
.904
.593
.772
.831
.794
[.042] [.034] [.182] [.047] [.042] [.064]
Constant
.084
-.189 -.188 -.041 .109 -.358
[.030] [.015] [.077] [.004] [.019] [.036]
Observations
9390
1080 1446 1555 1773 1730
R-squared
.624
.748
.548
.640
.649
447
County FE
Yes
No
No
No
No
No
State FE
No
Yes
Yes
Yes
Yes
Yes
State-Year FE
Yes
No
No
No
No
No
Robust standard errors in brackets.
Significance marked only for Partisanship Score: *p < .05
2012
.017
[.089]
.635
[.162]
-.013
[.039]
1806
.474
No
Yes
No
−.032, .037). The 1996 estimate remains more than ten times greater in magnitude than any
other election-year coefficient, and as in the original sample (see Table 3) falls well outside
the 95% confidence interval estimated using data from the other five elections (95% CI:
−.049, .028).
5
Discussion
The intersection of politics and economics represented by the perceptual screen hypothesis
has profound theoretical and substantive importance. Bartels (2002) states that his analysis
of survey evidence “suggests that partisan loyalties have pervasive effects on perceptions of
the political world” (p.138). But the evidence Bartels adduces in support of the perceptual
screen hypothesis implies a stronger claim than he articulates: partisan bias in perception of
objective economic conditions would indicate that the influence of partisan loyalties stretches
beyond the explicitly political realm. Gerber and Huber (2009) presented evidence support-
20
ing this strong form of the perceptual screen hypothesis: partisanship altered economic
perception, shaping individual consumption decisions. Gerber and Huber’s behavioral data
seemed to reveal the economic manifestation of a partisan perceptual screen. The magnitude
of the estimate from their analysis implied a dramatic effect of partisanship on aggregate
consumption.
My analysis shows that Gerber and Huber’s finding of a relationship between partisanship
and economic behavior is driven by a single anomalous state-year. But identification of the
model’s sensitivity to this outlier is post-hoc analysis. More telling is my finding that out-ofsample replications based on the 2008 and 2012 U.S. presidential elections, data unavailable
to Gerber and Huber at the time of their analysis, show no evidence of an effect. Considering
all six U.S. elections together, the data show the relationship between partisanship and
economic behavior to be a precisely-estimated zero. Outside of the United States, the results
are more ambiguous. As elections occur and data accumulate, replications in other countries
will shed additional light on this question.
Data from twenty elections across five countries provide scant evidence that economic
behavior reflects the partisan divergence in economic assessments documented in surveys.
There are two notable implications of this finding. First, the results presented here reverse
the balance of evidence on the perceptual screen hypothesis. Whether or not partisanship
changes the way citizens see the world is central to debate over the nature of partisanship—
Gerber and Huber remark that “perceptual bias is a (perhaps the) key channel through which
partisanship affects political behavior.” (2009, p. 423). The data provide no indication that
partisanship causes perceptual bias. Second, the results speak to the methodological question
of whether surveys provide a valid instrument for studying behavior. In their conclusion,
Gerber and Huber write: “Had we found no partisan differences in consumption following
elections, this would have cast doubt on the importance of survey reports of different partisan
expectations regarding the economy, as well as on the whole collection of partisan-tinged
21
responses to economic assessment items.” The results presented here suggest conclusions
about the link between partisanship and economic behavior that differ diametrically from
those of the original analysis.
22
A
Appendix
A.1
Map of county-level cross-election coefficient estimates
Figure A1 maps the population-weighted cross-election coefficient estimate for each county
in which at least four elections are observed. This map provides a geographic rendering
of the data along the original dimension of analysis, in which a pooled estimate is based
on variation within each county across Democratic wins (1992, 1996, 2008, 2012) versus
Democratic losses (2000 and 2004).
Orange indicates a positive relationship between partisanship and pre- to post-election
change in sales. A county in which partisans increase spending after a win would be colored
orange. Green indicates a negative relationship: a county in which partisans tend to increase
spending after a loss would be colored green. The certainty of the cross-election estimate
is indicated by fill-pattern. In this map, then, behavior indicative of a partisan perceptual
screen would appear as a more-orange, and more solidly orange, United States.
[Figure A1]
23
A.2
Evidence Outside the United States
To investigate the evidence for a partisan perceptual screen beyond the United States, I
gathered new data from four additional countries: Australia, France, Germany, and Ireland.
As a measure of economic behavior in each country, I collected monthly records of new
passenger car sales. New passenger car sales provide three primary advantages over local
taxable sales as a measure of individual-level economic behavior. New passenger car sales
(1) represent a significant and largely discretionary purchasing decision, likely to correspond
with a consumer’s personal economic expectations, (2) have records that more closely track
the timing of the consumer’s behavior, and (3) tend to reflect the purchasing decisions of
individual consumers, as distinct from business spending. In addition, new car sales records
are often available on a monthly basis, rather than quarterly, which provides a practical
advantage in cases where an election does not fall neatly between quarters, as it does in the
United States.
I expanded the sample to Australia, France, Germany, and Ireland because these countries
met the following three criteria. Each country (1) presented a national setting in which
identification with a party plays a role in political behavior,10 (2) had accessible information
on partisanship at a sub-national level, and (3) maintains records of new passenger car sales
on a monthly basis, at a geographic level that can be mapped to electoral districts, and over
a period of time during which at least two national elections took place. This new dataset
adds fourteen elections to the sample, from 1996 to 2013.
I follow Gerber and Huber (2009) in constructing the dependent variable as the log change
in economic behavior from before to after the election. Because this dependent variable is a
ratio—whether based on records of new car sales, as here, or on records of taxable sales, as
in Gerber and Huber—there is no need for a scale conversion to make the two sets of results
10
See, e.g., Evans 2004 for the French context; Kroh and Selb 2009 for the German; Marsh
2006 for the Irish; and Jackman 2003 for the Australian.
24
comparable. In the analyses that follow, I use log change from the three months leading up
to the election month to the three months following the election month, which allows for a
narrower window around the election than using quarters.11
Of the fourteen elections represented in this new dataset, a strong and highly significant
relationship exists in one: Ireland, 2007. The other thirteen elections provide no evidence of
an effect. This result is similar to that in the United States. As with the U.S. 1996 result, the
finding of an effect in Ireland 2007 is not easily reconciled with the theoretical model. The
model predicts that, all else equal, an effect of partisanship on economic behavior should be
most pronounced in elections characterized by high uncertainty, or in areas where partisans
expect especially large economic differences under each party. Among the five countries
in the sample, Ireland has the smallest proportion of self-identified partisans12 and is the
only country in the sample whose leading parties do not distinguish themselves on economic
ideology.13 From a theoretical standpoint, Ireland should be the least likely of these countries
11
A narrower window around the election is more likely to capture any short-lived effects
and less likely to capture the influence of events other than the election on the dependent
variable. Since U.S. presidential elections are always held in November, Gerber and Huber’s
use of quarterly data means there is a one-month buffer on either side of the election month:
change is measured from Q3 leading up to the election (July, August, September) to Q1
following the election (January, February, March). The results in each country differ when
using a one-month buffer on either side of the election month—for example, the pooled
Ireland estimates are no longer statistically significant at p = .05 (.341 [.271] with year
fixed effects, .239 [.275] with year and unit fixed effects)—but do not change any substantive
conclusions.
12
Australia, with a stable 2-party system and compulsory voting, has the highest rates
of party identification of any country in the sample. In 2007, the most recent CSES data
available on Australia, 89% of Australian CSES respondents reported being either close to a
political party or closer to one party than other parties. While parties are relatively volatile
in France, left-right ideological identification—easily mapped to the main parties—is very
strong and stable (see, e.g., Haegel 1993, Bélanger et al. 2006). Identification is somewhat
weaker in Germany, where in 2002 74% of CSES respondents reported being at least closer
to one party than others, declining to 62% in 2009. Ireland consistently reports the lowest
levels of identification among these countries, with only 54% reporting being closer to one
party than others in 2002, 55% in 2007, and 35% in 2011.
13
While the main Australian, French, and German parties are defined primarily by eco25
to show evidence of partisan perceptual screen in post-election economic behavior.
A.2.1
Australia
The Australian dataset contains observations for each of the areas represented in Australia’s
parliament over the seven federal elections between 1996 and 2013. The Australian constitution mandates that general elections must be held at least every three years, though the
Prime Minister may call for an early election. The Prime Minister, the Australian head of
government, is traditionally the leader of the party that wins a majority in the House of
Representatives, the lower chamber in Australia’s bicameral parliament.
Due in large part to the electoral system, national elections in Australia function as twoparty contests with high levels of turnout and of party identification. Australian elections
employ compulsory voting, and for the House of Representatives, a preferential vote system
that tends to promote partisan voting. Voters must assign every candidate on the House
ballot a unique rank in order to cast a valid ballot. First-preference votes (i.e. first-place
rankings) are tallied, and any votes for lagging candidates are redistributed according to the
voter’s ordering, until only two candidates remain. This final distribution of votes between
a winner and a runner-up is called the two-party preferred vote. In House elections, voters
commonly consult How to Vote cards handed out at the polls by partisan campaigners.14 In
2004, over 75% of voters reported identifying with either the Australian Labor Party or the
Liberal-National Coalition15 (Farrell and McAllister 2006).
I use two-party preferred vote in House elections to measure partisanship in Australia.16
nomic ideology, differences between the main Irish parties are historical divides that bear no
relation to economic ideology. Instead, partisan division in Ireland is based on each party’s
historical position towards the 1921 Anglo-Irish Treaty (Hutcheson 2011).
14
Senate races also promote partisan voting: above-the-line voting on Senate ballots allows
voters to check a single box to adopt a party’s ranking of all Senate candidates.
15
Though in fact distinct parties, the Liberal, National, and Western Australia’s Country
Liberal parties have been in long-standing coalition.
16
State and territorial returns are available at The University of Western Australia’s Aus26
Table 5: Australia
Pre- to Post-Election Change in New
Pooled
1996 1998 2001
Partisanship Score .046
.086
.576
.069 1.202
[.188] [.214] [.495] [.316] [.552]
3-yr Avg. LDV
.918
.859
.997
.626 1.225
[.087] [.098] [.169] [.239] [.515]
Constant
.053
.010 -.332 -.120 -.506
[.096] [.105] [.251] [.159] [.257]
Observations
56
56
8
8
8
R-squared
.775
.796
.876
.593
.559
Year FE
Yes
Yes
No
No
No
State FE
No
Yes
No
No
No
Car Sales
2004 2007
-1.116 .453
[.506] [.602]
.932
-.199
[.389] [.722]
.561
.213
[.254] [.283]
8
8
.721
.762
No
No
No
No
2010
.369
[.644]
.843
[.215]
.187
[.321]
8
.766
No
No
2013
-.381
[.300]
1.016
[.096]
.176
[.150]
8
.962
No
No
In both the Pooled and the 1996 regressions, an abbreviated 2-year avg. LDV is used for 1996
because data was not available prior to January 1994. The abbreviated 2-year avg. LDV consists
of the average of the one-year LDV and a two month (as opposed to three-month) version of the
two-year LDV.
The independent variable is the seven-election average Liberal-National proportion of the 2party vote, multiplied by −1 in 2007 and 2010, years in which Labor won. For the dependent
variable, I obtained monthly new passenger car sales by state/territory, available in a time
series back to January 1994 from the Australian Bureau of Statistics. Because data was not
available prior to January 1994, an abbreviated 2-year average lagged dependent variable is
used for 1996.17 Coefficients on the partisanship score are not significant at p = .05.
A.2.2
France
The French dataset contains observations for each of the country’s ninety-six departments
for the 2007 and 2012 presidential elections. As of 2002, French presidential elections are
tralian Politics and Elections Database.
17
A one-year lag is calculated as usual, based on the three months prior to and the three
months after the election month, and an abbreviated two-year lag is calculated based on the
two months prior to and the two months after the election month. The average of the oneyear and the abbreviated two-year lags is used in place of the standard three-year average
for observations in 1996.
27
held every five years. French presidential elections follow a two-round system. The first
round is a multi-party contest, which featured twelve parties in 2007 and ten in 2012. If no
absolute majority is won by a candidate in the first round, a second-round run-off is held
between the two leading candidates. A second round was held in both the 2007 and 2012
elections.
Political parties in France are less stable than in the other countries analyzed here, and the
weight of partisan identification as opposed to ideological identification has been a matter
of debate in studies of French voters. Evans (2004) writes that, “contemporary French
voters identify with parties much the same way as their traditional American counterparts,”
though other scholars argue that ideological identification is more meaningful than party
identification (Bélanger et al. 2006). The major French parties are easily classified on a
classic left-right divide (Bélanger and Lewis-Beck 2010). The Socialist party (PS) has been
the dominant party of the left for over thirty years. In 2002, Union for a Popular Movement
(UMP) emerged on the right as a merger between the previously dominant Rally for the
Republic and factions of other right and center-right parties. Since PS and UMP have
remained the primary left and right parties over the period analyzed, ideological identification
would function the same way as partisan identification.
I use the second-round Presidential vote to measure partisanship in France.18 The independent variable is the two-election average PS proportion of 2-party voteshare, multiplied
by −1 in 2007, when the UMP candidate won. For the dependent variable, I obtained
monthly new private car registrations by department, available in monthly time series back
to January of 2003. Coefficients on the partisanship score are not significant at p = .05.
18
Departmental returns are available from Sciences Po, Centre de Données Sociopolitiques.
28
Table 6: France
Pre- to Post-Election Change in New Car Registrations
Pooled
2007
2012
Partisanship Score .119
.079
.003
.237
[.086] [.075] [.101]
[.148]
3-yr Avg. LDV
.773
.257
.787
.784
[.038] [.156] [.046]
[.063]
Constant
-.042 -.005 .065
-.100
[.043] [.056] [.051]
[.073]
Observations
192
192
96
96
R-squared
.739
.905
.767
.634
Year FE
Yes
Yes
No
No
Département FE
No
Yes
No
No
A.2.3
Germany
The German dataset contains observations for each of the country’s sixteen Länder for the
2009 and 2013 federal elections. Elections for the Bundestag, Germany’s parliament, are
contested using mixed-member proportional representation. This system combines firstpast-the-post for election of representatives with proportional representation to establish
party-allocation of seats in the Bundestag. Each voter casts two votes. The first vote
is for a particular candidate, and determines which candidate is elected to represent the
constituency. The second vote is for a party list. Each state is allocated a number of seats
based on population, and the second vote determines what proportion of a state’s seats go
to each party. After a federal election, the party or a coalition of parties that make up the
majority of the Bundestag forms a government. Traditionally, the party holding the most
seats in the Bundestag elects their party’s leader as Chancellor of Germany, the head of
government.
Two parties, the Christian Democrats (CDU/CSU19 ) and the Social Democrats (SPD)
19
The Christian Social Union (CSU) in Bavaria is the sister party of the Christian Democratic Union, which operates in the other 15 states. The CDU and CSU are in long-standing
29
have been the dominant parties in the German political landscape for more than fifty years. A
third party, the Free Democratic Party (FDP) has been in government longer than any other
party, frequently acting as a junior coalition partner necessary to obtain a majority. The
FDP took part in the governing coalition with CDU/CSU in 2009. In 2013, the FDP failed
to meet the minimum 5% threshold necessary for party representation in the Bundestag,
and was ousted for the first time since the party’s founding in 1948. The CDU/CSU formed
a grand coalition with the SPD. Because the CDU/CSU remained in government for both
of the elections observed, I use the 2-election average FDP proportion of the two-party
FDP-SPD vote to construct the Partisanship Score variable.20
I measure partisanship in Germany using the second (party-list) vote.21 The independent
variable is the 2-election average FDP proportion of the two-party FDP-SPD vote, multiplied
by −1 in 2013 when SPD replaced FDP in the governing coalition. Since the FDP is a minor
party, I also calculate the results using the winning-coalition share of the three-party vote
in each election as the independent variable: CDU/CSU-FDP for 2009, CDU/CSU-SPD for
coalition and function essentially as one party within the Bundestag.
20
While not one of the two main German parties, the FDP was typically considered “the
‘pivotal party’ in German political life” (Jérôme, Jérôme-Speziari, and Lewis-Beck 2013),
since its presence in government often determined the balance of power. Because the FDP
was the only classically liberal party in the Bundestag, when the possibility arose that the
party might not meet the threshold, this potential ousting was predicted to have meaningful
economic consequences (see, e.g., The Economist 2013). Using FDP and SPD votes as
the denominator rather than the two main parties (CDU/CSU and SPD) is useful for two
reasons. First, a FDP-SPD Partisanship Score measures partisanship with respect to two
parties observed under both “treatment” (in the governing coalition) and “control” (not in
the governing coalition) conditions. And conveniently, as the classically liberal free-market
party, the FDP’s economic ideology provides better contrast against the Social Democrats
than the more economically-centrist CDU/CSU would. Though a smaller proportion of the
German population identifies with the FDP than with the CDU/CSU, these FDP partisans
would seem the citizens most likely to perceive a difference in economic outlook under a
FDP-CDU/CSU governing coalition compared to a SPD-CDU/CSU governing coalition—
especially given the possibility of a complete outser of their party and the looming spector
of a Bundestag with no voice from a classically liberal party for the first time in 65 years.
21
The German Federal Statistics Office provides state-level electoral returns since the first
Bundestag in 1949.
30
Table 7: Germany
Pre- to Post-Election Change in New Car Registrations
Pooled
2009
2013
Partisanship Score -.346 -.297 -.644
.105
[.481] [.452] [.773]
[.550]
3-yr Avg. LDV
.389
.077
1.655
.033
[.423] [.561] [1.133]
[.378]
Constant
-.201 -.156 -.127
-1.312
[.128] [.151] [.202]
[.337]
Observations
32
32
16
16
R-squared
.948
.979
.183
.004
Year FE
Yes
Yes
No
No
Land FE
No
Yes
No
No
2013. These results do not substantively differ from those found in Table 7.22 For the dependent variable, I obtained monthly new passenger car registrations from Germany’s Federal
Motor Vehicle Office, accessible to January 2006. As in Australia and France, coefficients on
the partisanship score are not significant at p = .05.
A.2.4
Ireland
The Irish dataset covers the three general elections from 2002 to 2011. The data available
from Ireland present two complications in establishing units of observation. First, the electoral constituencies, the spatial unit at which vote returns are measured, do not line up with
county boundaries, the spatial unit at which the dependent variable (new car registrations)
is recorded. Second, the boundaries of the electoral constituencies were redefined between
two of the three elections that took place during the period for which monthly car registration data are available. Between the 2002 and 2007 elections, five counties that made up
22
Pooled, no Land FE: −.557[.447]; Pooled, with Land FE: −.345[.594]; 2009, FDP-CDU
’09-’13 avg.: −.568[.534]; 2013, SPD-CDU ’09-’13 avg: −1.183[1.134].
31
four 2002 constituencies were reshuffled into three constituencies for 2007.23 In each case of
overlapping or shifting boundaries, I use the larger definition of area as the unit of analysis.
Electoral returns are summed across constituency for multi-constituency counties; car registrations are summed across county for two-county constituencies; and both returns and car
registrations are summed in order to create a single unit covering the five counties reshuffled
between 2002 and 2007 (Longsford-Westmeath-Roscommon-Leitrim-Sligo). This results in
twenty units that are common across all three elections.
The Dáil Éireann is the lower house of Ireland’s parliament, from which the head of
government, the Taoiseach, is selected. Dáil elections use a proportional representation
single transferable vote system. The voter need only indicate a first preference candidate in
order to cast a valid ballot, but may also assign other candidates a ranking.
Two parties, Fianna Fail (FF) and Fine Gael (FG), have dominated Irish politics since
the 1930s. Fianna Fail held the plurality of seats with FG as the second-largest party in
every Dáil convened from 1932 until 2011 (Hutcheson 2011). The parties do not represent a
traditional left-right divide, and the two main parties have a similar center-right economic
ideology (Hutcheson 2011). Marsh (2006) writes that party identification is present in the
Irish electorate, but that it is not very strong. “At least half of the electorate does have a
bias towards one party but only one out of every two electors have anything more than a
slight inclination towards that party.” (506–507).
In 2002, a Fianna Fail-Progressive Democrat (PD) coalition returned to power, marking
the first re-election of an incumbent government since 1969 (Mitchell 2003). Polls had
23
There are 26 counties in the Republic of Ireland. In 2002, there were eight counties
that each mapped directly to a single electoral constituency, eight counties that each contained multiple electoral constituencies within the county borders, and ten counties that were
paired into five electoral constituencies each comprising two counties. Constituency changes
before the 2007 election resulted in six single-constituency counties, nine multi-constituency
counties, four two-county constituencies and one pair of constituencies (Roscommon-Leitrim
South and Sligo-Leitrim North) that covered three counties.
32
suggested that Fianna Fail was almost certain to return to power, and though it did, the
party underperformed predictions. Fine Gael suffered its worst performance ever, described
as an “electoral catastrophe”. Turnout hit a record low at 62.1%. In the 2007 election,
Fianna Fail and PD once again returned to power, this time including the Green Party in
the governing coalition, as PD had lost 6 of its 8 seats. Gallagher and Marsh (2008) write
that the 2007 election was characterized by a high degree of uncertainty over the outcome.
Although the newly elected government “was much the same as the outgoing one. . . this
was not what anyone expected during the campaign. . . The theme of the 2007 results,
then, is not just one of little change; it is one of little change in the context of widespread
expectations of major changes” (78). The following election, in 2011, returned a FG-Labour
coalition and “reshaped a political system that Fianna Fail had previously dominated for
more than half a century” (Little 2011).
Table 8: Ireland
Pre- to Post-Election Change in New Car Registrations
Pooled
2002
2007
2011
Partisanship Score .666** .797** .125 .977** .892
[.231] [.173] [.193] [.133] [.594]
3-yr Avg. LDV
.817
.673
.833
.776
.817
[.116] [.173] [.193] [.133] [.239]
Constant
.015
-.719 -.188 -.684
.739
[.035] [.230] [.220] [.141] [.373]
Observations
60
60
20
20
20
R-squared
.976
.982
.536
.766
.543
Year FE
Yes
Yes
No
No
No
Unit FE
No
Yes
No
No
No
Significance marked only for Partisanship Score: **p < .01.
In both the Pooled and the 2002 regressions, a 2-year avg. LDV is
used for 2002 because data was not available prior to January 2000.
I measure partisanship in Ireland using the first preference vote.24 The independent
24
Electoral returns are available at ElectionsIreland.org back to the 1st Dail in 1918.
33
variable is the average 2002-2011 FF share of the two-party vote, multiplied by −1 in 2011
when Fine Gael became the first party and FF was knocked out of government.25 For the
dependent variable, I obtained monthly new private car registrations in each county from
Ireland’s Central Statistics Office, available back to January 2000. The partisanship score
shows a strong positive coefficient in the 2007 Irish general election and in pooled estimates
with or without unit fixed effects, significant in all three cases at p < .01.
A.2.5
Evidence from Five Countries
Altogether, the data presented above represents a varied sample of national settings with
accessible information on sub-national partisanship and consumption behavior. The United
States and Australia are two-party systems, while France, Germany, and Ireland are multiparty systems. Australia, Germany, and Ireland are parliamentary rather than presidential,
and in these three countries, the timing of a national election is not exogenous.26 Identification with a party plays a role in political behavior in each of the five countries, making this
sample relevant for investigation of a partisan perceptual screen.
Out of twenty elections across five countries, two elections show significant relationships
between local-level partisanship and post-election change in consumption behavior. Both
of the significant relationships are in the expected direction. Both elections—the United
States in 1996, Ireland in 2007—took place in the midst of sustained economic growth and
returned sitting governments to power. One electoral setting, the United States in 1996, was
characterized by very low levels of uncertainty over the outcome, while the other, Ireland in
25
I also calculate the results using the winning-coalition share of the five-party vote in
each election as the independent variable: FF-PD for 2002, FF-PD-GP for 2007, and FG-LP
for 2011. The results do not substantively differ from those found in Table 8: Pooled, no
Unit FE: .60[.23]; Pooled, with Unit FE: .68[.26]; FF-PD, 2002: −.03[.34]; FF-PD-GP, 2007:
.85[.26]; FG-LP, 2011: 1.00[.54].
26
For example, in Germany in 2005 (an election not covered in the dataset), the Chancellor
arranged for early dissolution of government and concomitant election. In Ireland in 2011,
an election was called because the governing coalition crumbled.
34
2007, was characterized by a high degree of uncertainty. The voting population in the United
States exhibits strong partisan identification and the leading parties differentiate themselves
economically; Irish voters exhibit weak partisan identification, and the leading parties do
not differentiate themselves economically.
It is possible that certain admixtures of partisanship, political and economic circumstances, and electoral uncertainty present conditions in which a partisan perceptual screen
manifests itself in economic behavior. The data analyzed here do not reveal those conditions.
Outside of the United States, standard errors are large, and so the results are difficult to
interpret. Further replication and extension may shed more light.
References
Bartels, Larry M. (2002). “Beyond the Running Tally: Partisan Bias in Political Perceptions”.
Political Behavior 24(2), pp. 117–150.
Bélanger, Éric and Michael S. Lewis-Beck (2010). “French National Elections: Democratic
Disequilibrium and the 2007 Forecasts”. Canadian Joural of Political Science 43, pp. 107–
122.
Bélanger, Éric et al. (2006). “Party, Ideology, and Vote Intentions: Dynamics from the 2002
French Electoral Panel”. Political Research Quarterly 59(4), pp. 503–515.
Berelson, Bernard R., Paul F. Lazarsfeld, and William N. McPhee (1954). Voting: A Study
of Opinion Formation in a Presidential Campaign. Chicago: University of Chicago Press.
Bullock, John G. et al. (2013). “Partisan Bias in Factual Beliefs about Politics”. NBER
Working Paper 19080 (May).
Cai, Fengyan, Felix Tang, and Jian-min Jia (2009). “The Interaction Effect of Mood and
Price Level on Purchase Intention”. Advances in Consumer Research 36, pp. 963–964.
Campbell, Angus et al. (1960). The American Voter. New York: John Wiley & Sons, Inc.
35
Enns, Peter K. and Christopher J. Anderson (2009). “The American Voter Goes Shopping: Presidential Elections and the Micro-Foundations of Partisan Consumption”. SSRN
Working Paper No. 1451209. url: {http: // papers.ssrn .com /sol3 /papers. cfm?
abstract_id=1451209}.
Evans, Jocelyn A. J. (2004). “Ideology and Party Identification: A Normalisation of French
Voting Anchors?” The French Voter: Before and After the 2002 Elections. Ed. by Michael
S. Lewis-Beck. Basingstoke: Palgrave Macmillan, pp. 47–73.
Farrell, David M. and Ian McAllister (2006). The Australian Electoral System: Origins,
Variations, and Consequences. Sydney: University of New South Wales Press.
Gaines, Brian J. et al. (2007). “Same Facts, Different Interpretations: Partisan Motivation
and Opinion on Iraq”. The Journal of Politics 69(4), pp. 957–974.
Gallagher, Michael and Michael Marsh (2008). How Ireland Voted 2007: The Full Story of
Ireland’s General Election. Basingstoke: Palgrave Macmillan.
Gallup (1996). “POLL TRENDS: The 1996 Presidential Election”. url: {http://library.
law.columbia.edu/urlmirror/CLR/100CLR524/ptpreselec.html}.
— (2008). “Gallup Presidential Election Trial-Heat Trends, 1936-2008”. url: {http://www.
gallup . com / poll / 110548 / gallup - presidential - election - trialheat - trends 19362004.aspx}.
— (2012). “Election 2012 Registered Voters Trial Heat: Obama vs. Romney”. url: {http:
//www.gallup.com/poll/150743/Obama-Romney.aspx}.
Gerber, Alan S. and Gregory A. Huber (2009). “Partisanship and Economic Behavior: Do
Partisan Differences in Economic Forecasts Predict Real Economic Behavior?” American
Political Science Review 103(3), pp. 407–426.
— (2010). “Partisanship, Political Control, and Economic Assessments”. American Journal
of Political Science 54(1), pp. 153–173.
36
Hall, Robert E. (1978). “Stochastic Implications of the Life Cycle-Permanent Income Hypothesis: Theory and Evidence”. Journal of Political Economy 86(6), pp. 971–987.
Hibbs, Douglas A. (1996). “The Economy and the 1996 Presidential Election: Why Clinton Will Win with Probability Exceeding 90%”. Göteborg University Working Paper
(September).
Hutcheson, Derek S. (2011). “The February 2011 Parliamentary Election in Ireland”. IBIS
Working Paper No. 109. url: {http://www.ucd.ie/ibis/filestore/wp2011/109_
hutcheson.pdf}.
Jackman, Simon (2003). “Political Parties and Electoral Behavior”. The Cambridge Handbook of Social Sciences in Australia. Ed. by Ian McAllister, Steve Dowrick, and Riaz
Hassan. Rydon 1968. Cambridge: Cambridge University Press. Chap. 16.
Jérôme, Bruno, Véronique Jérôme-Speziari, and Michael S. Lewis-Beck (2013). “A PoliticalEconomy Forecast for the 2013 German Elections: Who to Rule with Angela Merkel?”
PS: Political Science & Politics 46(03), pp. 479–480.
Kernell, Georgia and Kevin Mullinix (2013). “The Scope of the Partisan Perceptual Screen”.
Institute for Policy Research Working Paper WP-13-15 (July). url: {http : / / www .
ipr.northwestern.edu/publications/docs/workingpapers/2013/IPR-WP-13-15A.pdf}.
Kramer, Gerald H. (1971). “Short-Term Fluctuations in U. S. Voting Behavior, 1896-1964”.
American Political Science Review 65(1), pp. 131–143.
Kroh, Martin and Peter Selb (2009). “Inheritance and the Dynamics of Party Identification”.
Political Behavior 31(4), pp. 559–574.
Lerner, Jennifer S., Deborah A. Small, and George Loewenstein (2004). “Heart strings and
purse strings: Carryover effects of emotions on economic decisions”. Psychological Science
15(5), pp. 337–341.
37
Little, Conor (2011). “The General Election of 2011 in the Republic of Ireland: All Changed
Utterly?” West European Politics 34(6), pp. 1304–1313.
Mano, Haim (1999). “The influence of pre-existing negative affect on store purchase intentions”. Journal of Retailing 75(2), pp. 149–172.
Marsh, Michael (2006). “Party identification in Ireland: An insecure anchor for a floating
party system”. Electoral Studies 25(3), pp. 489–508.
Mitchell, Paul (2003). “Fianna Fáil still dominant in the coalition era: The Irish general
election of May 2002”. West European Politics 26(2), pp. 174–183.
Prior, Markus, Gaurav Sood, and Kabir Khanna (2015). “You Cannot Be Serious: The Impact of Accuracy Incentives on Partisan Bias in Reports of Economic Perceptions”. Quarterly Journal of Political Science (forthcoming). url: {http://gsood.com/research/
papers/PartisanBias.pdf}.
Saliagas, Dena Thometz and James J. Kellaris (1986). “The Influence of Mood on Willingness
to Spend and Unplanned Purchasing”. Developments in Marketing Science: Proceedings
of the Academy of Marketing Science. Ed. by Naresh K. Malhotra. Springer International
Publishing, pp. 61–65.
Stanig, Piero (2009). “Political Bias in Economic Perceptions”. London School of Economics
Working Paper. url: {http://www.lse.ac.uk/government/research/resgroups/
PSPE/pdf/Stanig.pdf}.
Taber, Charles S. and Milton Lodge (2006). “Motivated Skepticism in the Evaluation of
Political Beliefs”. American Journal of Political Science 50(3), pp. 755–769.
The Economist (2013). “Germanys Free Democrats: The endangered queenmaker”. January
12.
Trigger, Kaitlyn (2006). “Partisan Bias in Consumer Behavior”. Yale University, Typescript.
Zeldes, Stephen P . (1989). “Optimal Consumption with Stochastic Income: Deviations from
Certainty Equivalence”. The Quarterly Journal of Economics 104(2), pp. 275–298.
38