t -d h support for the republican party · graduate from high school by their early 20s, with...
TRANSCRIPT
THE ANTI-DEMOCRAT DIPLOMA:HOW HIGH SCHOOL EDUCATION INCREASES
SUPPORT FOR THE REPUBLICAN PARTY∗
JOHN MARSHALL†
DECEMBER 2016
Attending high school can alter student life trajectories by affecting labor marketprospects and exposure to ideas and networks. However, schooling’s influence com-petes with powerful early socialization forces, and may be confounded by well-establishedselection biases. Consequently, little is known about whether or how high school edu-cation shapes downstream political preferences and voting behavior. Using a difference-in-differences design exploiting variation in U.S. state dropout laws across cohorts, Ifind that raising the school dropout age substantially increases Republican partisanidentification and voting later in life. Instrumental variables estimates, which reflectdropout laws principally impacting minorities, show that each completed grade of highschool increases Republican support by around 10 percentage points. High school’seffects operate by increasing income, which increases support for conservative eco-nomic policies and ultimately the Republican party. These effects influence futurestate legislature composition, suggesting that recent Democrat attempts to raise statedropout ages are strategically misguided.
∗I thank Jim Alt, Charlotte Cavaille, Dan Carpenter, Jon Fiva, Anthony Fowler, Andy Hall, Shigeo Hirano, TorbenIversen, Larry Katz, Horacio Larreguy, Rakeen Mabud, Jonathan Phillips, Jim Snyder, and Harvard University work-shop participants for many insightful comments on earlier drafts. I wish to thank Paul Bolton, John Bullock, PhilipOreopoulos and Jim Snyder for kindly sharing data.†Department of Political Science, Columbia University. [email protected].
1
1 Introduction
Education is a fundamental component of U.S. economic and social policy, and continues to be a
salient political issue.1 However, as recently as 2010, nearly 20% of students per cohort still fail to
graduate from high school by their early 20s, with substantially lower graduation rates among mi-
nority populations (Murnane 2013). Particularly for the Democratic party, raising the high school
dropout age is a key tool helping to ensure that children from disadvantaged backgrounds share
in the economic returns to education (see Oreopoulos 2009). President Obama underscored the
importance of this issue in his 2012 State of the Union address, stating that “When students are
not allowed to drop out, they do better. So ... I am proposing that every state ... requires that all
students stay in high school until they graduate or turn 18.” Dropout ages have generally increased
over the 20th century, and many states have recently considered raising, or successful raised, their
minimum leaving age to 17 or 18.2
While high school education undoubtedly shapes labor market prospects and social interac-
tions, its capacity to impact voters’ political preferences and voting behavior is not theoretically
obvious. First, it is not clear that schooling is able to overcome competing influences on partisan
identification and voting behavior. In particular, a prominent literature argues that partisan iden-
tities and vote choices are highly durable over the course of an individual’s lifetime (e.g. Bartels
2010; Campbell et al. 1960; Converse and Markus 1979; Jennings, Stoker and Bowers 2009). Such
behaviors are believed to develop primarily at home from a young age (e.g. Jennings, Stoker and
Bowers 2009; Kroh and Selb 2009) or as a young adult (e.g. Meredith 2009; Stoker and Bass
2011). From this perspective, schooling—and especially its labor market consequences only expe-
1Across federal, state and local levels, spending on education represents 5% of GDP in 2015. It is thethird largest budget item behind pensions and health care.
2In 2013, Kentucky increased the dropout age to 18 (or receiving a high school diploma). Similarly,Rhode Island raised its leaving age to 18, effective of 2011. Maryland also legislated in 2012 to raise itsdropout age to 17 for 16-year-old students for the 2015-2016 school year, and to 18 for the 2017-2018school year. Similar bills have been introduced in Alaska, Delaware, Illinois, Massachusetts, Montana,South Carolina and Wyoming since 2011.
2
rienced later in life, once preferences are well-established—could have limited political impact in
the face of powerful habits and early socialization.
Second, even if high school education is able to influence partisan preferences, it is not clear
whether this predisposes voters toward the Democratic or Republican party. One one hand, edu-
cation increases downstream income (e.g. Acemoglu and Angrist 2000), which may cause voters
to support lower taxation (e.g. Meltzer and Richard 1981; Romer 1975). At least once the returns
to schooling are realized later in life, remaining in high school could thus increase Republican
support. Conversely, high school education may increase exposure to liberal values, either in the
classroom (e.g. Dee 2004; Hillygus 2005; Niemi and Junn 1998) or through relatively liberal so-
cial networks (e.g. Alwin, Cohen and Newcomb 1991; Green, Palmquist and Schickler 2002; Nie,
Junn and Stehlik-Barry 1996). This could instead increase support for the Democratic party. Both
sets of arguments suggest that schooling has the potential to transform students’ life trajectories,
causing students to change their policy stances and ultimately support different political parties
later in life.
The extant evidence investigating high school education’s partisan effects is correlational and
often conflicting. Classic survey studies document a positive association between greater education
and voting Republican (e.g. Campbell et al. 1960; Verba, Schlozman and Brady 1995). However,
such correlations may simply reflect the well-documented differences in the types of people that
become better educated (e.g. Kam and Palmer 2008; Sondheimer and Green 2010), or the fact that
education is itself often a cause of the variables researchers control for. Furthermore, by failing
to disentangle either the direction of this relationship or its mechanisms, scholars have struggled
to square the widely-documented correlation between income (which education increases) and
support for conservative economic policies (e.g. Erikson and Tedin 2015; Fisher 2014; Gelman
et al. 2010) with the correlation between education and socially liberal attitudes (e.g. Dee 2004;
Gerber et al. 2010).
This article identifies how raising the high school dropout age, and each additional completed
3
grade of late high school education that this imparts, affects voter policy preferences and voting
behavior in later life. To estimate these partisan effects of schooling, I employ a difference-in-
differences design leveraging cross-state differences in the dropout age across different cohorts
(see also Acemoglu and Angrist 2000; Angrist and Krueger 1991). The high school dropout age
remains an important policy instrument; despite substantial progress, in 2010 16.3% of students
per cohort failed to graduate from high school by age 20-24 (Murnane 2013). My estimates show
that raising the dropout age by an additional year significantly increases the average student’s
completed grades of schooling. In addition to estimating the effects of raising the dropout age, I
also use school leaving laws to instrument for the number of grades of schooling that an individual
completes. Two-sample instrumental variables (IV) methods address the difficulty that the number
of grades completed is rarely measured in large-scale political surveys (Angrist and Krueger 1992;
Inoue and Solon 2010).3 Given that the dropout age does not affect post-high school education,
the IV estimates identify the effect of completing an additional grade of high school for students
that only remained in school due to a higher state dropout age.
Examining large nationwide surveys conducted around the 2000, 2004 and 2008 Presidential
campaigns, I find that remaining in high school causes voters to become substantially more con-
servative in later life. I first demonstrate that raising the dropout age has a large political effect in
its own right. Increasing the minimum school leaving age by a year causes a 1-3 percentage point
shift, per birth-year cohort, toward identifying as a Republican partisan or voting for the Repub-
lican party. The decline in Democrat support is commensurate. The IV estimates further indicate
that each additional completed grade of high school increases the probability that a student will
subsequently identify as a Republican partisan and vote for a Republican Presidential candidate
by around 10 percentage points. Among compliers, who I demonstrate to be disproportionately
black and Hispanic, the dropout age thus dramatically affects downstream political behavior, caus-
3As discussed below, instrumenting for an indicator of completing high school could considerably up-wardly bias IV estimates (Marshall 2016).
4
ing voters to become considerably more conservative. Tentative estimates indicate that these shifts
have accumulated across cohorts to increase the Republican seat share in state House and Senate
chambers.
Furthermore, I show that high school’s partisan effects operate primarily via an income-based
channel. Previous evidence from the U.S. indicates that an additional year of schooling increases
wages by around 10% (e.g. Acemoglu and Angrist 2000; Angrist and Krueger 1991; Ashenfelter
and Rouse 1998; Oreopoulos 2006). Consistent with education’s economic returns inducing voters
to support low redistribution (Meltzer and Richard 1981; Romer 1975), the political effects of edu-
cation are greatest at an individual’s mid-life earnings peak, while each additional completed grade
causes voters to favor Republican positions on economic issues. Although this could simply reflect
voters becoming Republican partisans and adopting the party’s policy positions (Lenz 2012; Zaller
1992), education does not increase support for Republican positions on non-economic issues such
as abortion, gun control, health care spending and military spending. Furthermore, contrary to the
predictions of civic education (Dee 2004; Hillygus 2005) and social network (Green, Palmquist
and Schickler 2002; Mendelberg, McCabe and Thal forthcoming; Nie, Junn and Stehlik-Barry
1996) socialization theories, I find no evidence to suggest that high school education affects polit-
ical engagement or socially liberal values. Together, these results indicate that, by increasing an
individual’s subsequent income, high school education causes voters to support economic policies
that, ultimately, lead them to identify as and vote Republican.
These findings raise a “catch 22” for the Democratic party: while increasing access to education
by raising the school dropout age is an important plank of their policy agenda, it comes at the cost
of the beneficiaries of additional schooling electing Republican politicians in the future. Moreover,
given that the beneficiaries are principally black and Hispanic voters, the results further suggest
that income motivations can in part explain why some minority voters support a Republican party
that has generally opposed minority interests and immigration over the period my analysis covers.
The results also recast our understanding of the life-cycle development of political behavior.
5
Previous studies have emphasized both the “primary principle,” that what is learned first—typically
from parents (e.g. Jennings, Stoker and Bowers 2009)—is learned best (see Erikson and Tedin
2015), and particularly that early adulthood is central to political socialization (Mendelberg, Mc-
Cabe and Thal forthcoming; Stoker and Bass 2011). In contrast, although there is some evidence
that high school increases political participation (Sondheimer and Green 2010), adolescent civics
classes and peer interactions at school have produced limited effect on civic and partisan attitudes
(Erikson and Tedin 2015; Green et al. 2011). However, my findings show that high school edu-
cation has important consequences for partisan identification and vote choice later in life, which
can overcome other influences before and after high school, and are driven by downstream income
opportunities rather than socialized values. Given that education’s pro-Republican effects coincide
with voters’ earnings peak, and thus vary across a voter’s lifetime, the results imply that parti-
san preferences are significantly more responsive to changes in circumstances over a lifetime than
previous studies highlighting the stickiness of partisan identity suggest.
By identifying how education affects which party an individual supports, this article departs
from the large literature debating the relationship between education and political participation
(e.g. Berinsky and Lenz (2011); Henderson and Chatfield (2011); Kam and Palmer (2008); Sond-
heimer and Green (2010)). This article instead emphasizes high school’s partisan effects. In this
respect, it complements recent evidence that attending an affluent college with a norm of finan-
cial gain induces students to support conservative economic policies (Mendelberg, McCabe and
Thal forthcoming). This article demonstrates that a similar income-oriented mechanism applies
to the realization of income gains accruing to high school education for disadvantaged popula-
tions, such that schooling shapes voters to oppose redistribution and vote Republican party later in
life. Given that high school’s effects are most pronounced at a voter’s earnings peak, my findings
suggest that Mendelberg, McCabe and Thal’s (forthcoming) study of students at college—before
financial gains are realized—could understate the long-term consequences of conservative college
campuses. Combined, these findings provide a causal interpretation to the increasing likelihood of
6
voting Republican as a voter becomes more educated, at least until the postgraduate level where
the correlation reverses (Fisher 2014).
2 Theoretical perspectives
Education plays a central role in the lives of American voters. However, although much has been
written about how education affects income, socioeconomic outcomes and civic participation, ed-
ucation’s effect on partisan behavior has received surprisingly limited attention. To the extent that
it is able to overcome other powerful early life forces, I build upon contrasting economic and soci-
ological mechanisms to consider how schooling could affect which political party a voter identifies
with and ultimately votes for.
2.1 Education increases income, decreasing support for taxation?
Perhaps the most obvious, but potentially most politically relevant, effect of education may be to al-
ter an individual’s earnings trajectory. A large body of research now shows that education increases
an individual’s wage income later in life. Twin studies in the U.S. suggest that an additional year of
schooling increases annual wages by 11-16% (e.g. Ashenfelter and Rouse 1998), while studies us-
ing compulsory schooling laws to instrument for schooling estimate this rate of return to be 8-18%
(Acemoglu and Angrist 2000; Angrist and Krueger 1991; Oreopoulos 2006, 2009). The human
capital interpretation of this relationship claims that education imparts productive skills, which are
rewarded in competitive labor markets (e.g. Becker 1964).
Linking schooling’s effect on income to political behavior, Romer (1975) and Meltzer and
Richard (1981) (henceforth RMR) argue that individuals with higher wages will prefer lower in-
come tax rates and less income redistribution. To the extent that redistributive policy preferences
influence vote choice, preferring lower income taxation entails supporting candidates advocating
low tax rates. Particularly since Ronald Reagan’s Presidency, the Republican party has consistently
7
supported lower income taxation. Furthermore, as ideological and policy differences between the
Democratic and Republican parties have grown (McCarty, Poole and Rosenthal 2006), the differ-
ence between the Democratic and Republican parties on this issue has become easy for voters to
discern. According to American National Election Surveys conducted since 2000, 71% of voters
identify the Republican party as more right-wing than the Democrat party. Moreover, since the
1980s, survey correlations have consistently shown that higher-income voters are more likely to
identify with and vote for the Republican party (Erikson and Tedin 2015; Fisher 2014; Gelman
et al. 2010).
Together, the human capital and RMR models imply that increased schooling should make
voters more favorable toward Republican economic policies, and particularly their proposals for
lower income tax rates, by increasing their income. Furthermore, given that voters can relatively
accurately predict their future income, redistributive preferences may also reflect expected earnings
(Alesina and La Ferrara 2005). Although policy preferences do not necessarily translate into vote
choices, differences in vote choices are likely to increase with realized returns to education.
Critics of human capital theory have instead argued that education is a costly and unproductive
signal of a worker’s underlying productivity (Spence 1973). By this logic, education does not itself
impart skills that improve a worker’s productivity, but instead represents a costly signal enabling
workers that are productive along other dimensions to distinguish themselves from less produc-
tive workers, for whom the costs of becoming educated are greater. Accordingly, an increase in
education should not affect the national income distribution or a voter’s support for the economic
policies proposed by different political parties. While such a signaling model, in conjunction with
the RMR logic, could still be consistent with a correlation between education and vote choices,
there should remain no empirical relationship when using a research design that controls for the
underlying differences in productivity that drive educational choices in the signaling model.
8
2.2 Education increases exposure to liberal values?
From the perspective of political socialization, education’s effects on political behavior could be
very different. A considerable literature shows that education is correlated with socially and polit-
ically liberal attitudes. For example, more educated voters display greater support for freedom of
expression (Dee 2004; Gerber et al. 2010), ethnic diversity (e.g. Bobo and Licari 1989; Campbell
et al. 1960) and immigration (e.g. Hainmueller and Hiscox 2010). At least in the recent decades
examined in this article, these values have become strongly associated with the Democratic party
(Levendusky 2009). Some studies find that these correlations also extend to economic policy pref-
erences (e.g. Gerber et al. 2010), although others do not (e.g. Weakliem 2002). Thus, while
education may bestow economically valuable knowledge and skills, it could also instill liberal
values.
Two main mechanisms have been proposed to explain these associations: direct effects of
curricula; and changes in social network composition. First, civic education and social science
classes generally encourage political engagement and trust in the political system (Niemi and Junn
1998), and seek to instill liberal values of tolerance (Dee 2004; Hillygus 2005). Such classes
typically start in late high school, and can intensify if students choose to concentrate in such areas
at college. Furthermore, formative experiences appear to affect political beliefs decades later in
life (Jennings, Stoker and Bowers 2009; Meredith 2009). However, recent experimental evidence
finds that while enhanced high school civics classes increase political knowledge, this knowledge
quickly decays and does not affect support for civil liberties (Green et al. 2011). Given that other
empirical evidence has emphasized the importance of college education (see Galston 2001), civic
education’s importance may be confined to higher education.
Second, education could more indirectly affect a voter’s political beliefs by altering the com-
position of their social network. More educated individuals sort into more prestigious, ethnically
diverse and politically-connected networks (Green, Palmquist and Schickler 2002; Nie, Junn and
9
Stehlik-Barry 1996), or at least become exposed to such groups (Alwin, Cohen and Newcomb
1991). Such networks are often characterized by liberal social values and high levels of political
interest, knowledge and discussion, and might thus increase support for the Democratic party. At
lower levels of education, individuals may enter social networks like labor unions with a clear
left-wing political agenda (Abrams, Iversen and Soskice 2010).
While both mechanisms could induce socially liberal attitudes, the observed correlations may
nevertheless be spurious. Family background, values, cognitive abilities and life experiences are
likely to be correlated with both education and socially liberal attitudes (Jencks et al. 1972; Kam
and Palmer 2008). Excepting Green et al.’s (2011) randomized control trial, which provides little
evidence of durable value change, this concern is particularly salient because these studies do not
exploit plausibly random variation in education, and often control for variables like income that
are themselves downstream consequences of education.
3 Empirical design
The difficulty of identifying the partisan effects of education is widely acknowledged. As noted
above, there are many confounding explanations for any correlation between schooling and policy
preferences and voting behavior (see e.g. Kam and Palmer 2008; Sondheimer and Green 2010).
Beyond the omitted variable concern that has besieged the extant literature, another concern is in-
terpreting potential heterogeneity in schooling’s political effects. First, the effect of schooling is
likely to be heterogeneous across types of individual (see Card 1999). Second, schooling’s effects
may depend upon the grade in question, e.g. because middle school, high school and college cover
different content. As suggested above, different levels of schooling may thus differentially stimu-
late income-based and socialization-based mechanisms. Accordingly, a failure to separate levels of
schooling would return a misleading null finding if, for example, high school’s conservative effects
were counterbalanced by liberal effects of attending college.
10
To address these concerns, I exploit cross-state variation in state minimum dropout ages—both
as an instrument for schooling, but also as a topical policy variable of interest in its own right—
across cohorts using a difference-in-differences design. Since I show that dropout ages do not
induce students to attend college, my estimates identify the effect of completing an additional grade
of high school for individuals that would otherwise have dropped out. This study thus offers a rare
opportunity to cleanly identify the causal effects of a specific level of schooling on partisanship
and voting behavior.
3.1 State school dropout age laws
Compulsory schooling laws are legislated by state governments and define the minimum age at
which a student must enter and may drop out of schooling. The first such law was adopted in
Massachusetts in 1852 and 41 states had adopted similar laws by 1910, principally to meet de-
mand for educated workers and promote assimilation (Goldin and Katz 2008; Kotin and Aikman
1980). The laws have since predominantly focused on the dropout age, which gradually increased
throughout the twentieth century. The minimum leaving age is now 18 for almost half the states
in the country. While the enforcement of dropout laws is relatively weak, states can and do punish
habitual truancy (see Oreopoulos 2009). Labor economists have thus frequently used compulsory
schooling laws as instruments to estimate the economic effects of schooling in the U.S. (e.g. Ace-
moglu and Angrist 2000; Angrist and Krueger 1991; Lochner and Moretti 2004; Milligan, Moretti
and Oreopoulos 2004).
Figure 1 plots the minimum age at which a student is permitted to drop out of school across
the 48 contiguous states and Washington, DC between 1920 and 2004.4 Unsurprisingly, there has
4Alaska and Hawaii are omitted throughout due to lack of data. Although more complex measures ofcompulsory schooling have previously been adopted (e.g. Acemoglu and Angrist 2000; Angrist and Krueger1991), the minimum dropout age captures the relevant legislation for the empirical analyses conducted here.In particular, I omit child labor laws because they offer limited variation and are mostly irrelevant for theperiod I study, and thus produce a weak first stage. Their inclusion as additional instruments does not affectthe results.
11
1214
1618
1214
1618
1214
1618
1214
1618
1214
1618
1214
1618
1214
1618
1920 1940 1960 1980 2000 1920 1940 1960 1980 2000 1920 1940 1960 1980 2000 1920 1940 1960 1980 2000 1920 1940 1960 1980 2000 1920 1940 1960 1980 2000 1920 1940 1960 1980 2000
AL AR AZ CA CO CT DC
DE FL GA IA ID IL IN
KS KY LA MA MD ME MI
MN MO MS MT NC ND NE
NH NJ NM NV NY OH OK
OR PA RI SC SD TN TX
UT VA VT WA WI WV WY
Min
imum
sch
ool l
eavi
ng a
ge
Year
Figure 1: U.S. state minimum school leaving age laws, 1920-2004
Note: Dropout law data are from Oreopoulos (2009), and based on the National Center for EducationStatistic’s Education Digest.
been a general upward trend in minimum leaving ages. However, there remain numerous instances
of reversal as in Maine, Mississippi, and Oregon. Given that high school dropout rates remain non-
trivial—consistently registered at around 20% for all students since the 1970s, with considerably
higher rates among ethnic minorities (Murnane 2013)—many states are actively seeking to raise
the legal dropout age. Since 1914, the majority (74.6%) of state-years specified a dropout age of
16; 8.8% are below 16; 7.8% use 17; and 8.7% use 18. Since the leaving age reached in 16 in all
states in 1993, when Mississippi increased its dropout age to 16, today’s policy debate centers on
raising the leaving age to 17 or 18.
12
I use two indicators—1(dropout age = 16) and 1(dropout age ≥ 17)—to examine the effect
of these leaving ages on a given cohort, relative to the baseline of a dropout age below 16.5 The
former indicator primarily captures changes in dropout ages in the early and mid twentieth century,
while the latter captures more recent reforms that affected today’s younger generations. The results
below—which use recent surveys—separate between these thresholds, and show similar effects
when focusing only on dropout ages of 17 or above.
3.2 Identification strategy
To identify the effects of schooling on which party voters identify with and vote for later in life,
I leverage state-level reforms in the dropout age using a difference-in-differences design. I first
identify the political effects of raising the dropout age in its own right, before further applying an
IV strategy to identify the effect of completing an additional grade of schooling for individuals that
only remained in school because of changes in the dropout age affecting their cohort.
3.2.1 Reduced form: identifying the effect of increasing the school dropout age
To identify the political effects of state dropout laws, I use cohorts from states that did not change
their dropout age as controls for the same cohorts in states where a reform of the minimum leav-
ing age occurred. This difference-in-differences design thus separates out common trends across
cohorts in political support from the impact of dropout laws on cohorts in states where a reform
occurred. The design entails estimating equations of the form:
Yicst = δ11(dropout agecs = 16)+ δ21(dropout agecs ≥ 17)+
αc +θs +φsbirth yearc +ηt +Witγ + εicst , (1)
5The leaving ages of 17 and 18 are combined because forcing students only months from completinghigh school to remain in school until 18 has little impact on schooling outcomes (see also Lochner andMoretti 2004). Using 1(dropout age = 18) as an additional instrument, or using indicators for each mini-mum leaving age, provide similar results.
13
where Yicst is an outcome (e.g. Republican identifier/voter) at survey year t for an individual i from
cohort c in state s. To exploit only variation across different cohorts within a state, the difference-
in-differences strategy includes fixed effects αc for birth-year (cohort) and fixed effects θs for the
state in which a respondent state grew up. Cohort fixed effects control for all differences across
cohorts, including generation effects reflecting political events that may have occurred at more im-
pressionable ages—such as early adulthood—for some cohorts (see Stoker and Bass 2011). State
fixed effects capture all time-invariant characteristics of states, such as varying levels of school-
ing and political support. In addition, state-specific linear cohort trends, φsbirth yearc, allows for
underlying trends in Yicst across cohorts to vary by state.6 Pre-treatment individual characteris-
tics Wit—gender and race indicators, and quartic age polynomial terms—are included to enhance
estimation efficiency, while survey year fixed effects ηt capture common shocks across electoral
campaign periods. Standard errors are clustered by state-cohort.
This estimation strategy thus compares changes in support for the Republican party across
cohorts within states that changed their dropout age to changes across cohorts within states that did
not. The key identifying assumption is that in the absence of changes in dropout laws, individuals
from treated states would experience parallel trends in Republican support to individuals from
control states. There are two main types of challenge to this assumption: individual selection into
particular dropout laws; and school dropout age reforms reflecting unobserved trends which also
affect Republican support.
One concern is that individuals likely to be impacted by the reform may move in response to
dropout age reforms. However, among poor women of childbearing age—whose children are more
likely to be remain in school only because they are required to—cross-state migration is especially
low (Molloy, Smith and Wozniak 2011), and is concentrated well before their children typically
reach high school age (Gelbach 2004). Furthermore, the principal reasons to move across states—
for college, marriage, family reasons, and natural disaster (Molloy, Smith and Wozniak 2011)—are
6Table A5 in the Appendix shows similar results excluding state-specific cohort trends.
14
unlikely to be linked to dropout age reforms. Moreover, it is likely that if a parent is willing to
move to improve their child’s education, they would also persuade their child to remain in school
past the dropout age anyway. Robustness checks in the Appendix show that the selective migration
required to nullify the results is not plausible.
Since legislation is not randomly enacted, another concern is the context of dropout age re-
forms. Perhaps the most plausible violation of the exogeneity of minimum school leave ages with
respect to political outcomes is if reforms correlate with state-level political changes. For example,
state legislatures dominated by Republicans may disproportionately raise the minimum leaving
age. However, Table A2 in the Appendix shows that changes in the school leaving age are not
predicted by partisan political forces. Specifically, leaving ages are uncorrelated with Republican
control of upper, lower and both state houses, Republican state legislative seat shares, and Repub-
lican governors. The lack of such correlations suggest that dropout age reforms were unlikely to
have been associated with waves in local political sentiment.
Although political trends are uncorrelated with changes in state school leaving ages, reforms
could still reflect socioeconomic changes. However, Lochner and Moretti (2004) show that state
high school graduation trends are uncorrelated with school age reforms. More specific potential
concerns are addressed below by controlling for state-level political, educational, economic and
demographic conditions at the time when a voter attended high school. Most importantly, the state-
specific cohort trends in the baseline specification provide a demanding general check against the
possibility that leaving age reforms could reflect long-run processes that may differ across states,
such as changes in school quality (Stephens and Yang 2014), and that could also affect political
behavior.
3.2.2 Instrumental variables: identifying the effect of an additional grade of late high school
To estimate the effect of an additional year of schooling—as opposed to increasing the state’s
school dropout age—on identifying as and voting Republican, I build upon the reduced form
15
specification by using the dropout age to instrument for schooling. I use the same difference-
in-differences design to estimate the following structural equation:
Yicst = βSicst +αc +θs +φsbirth yearc +ηt +Witγ + εicst , (2)
where Sicst is a measure of schooling received by individual i. I instrument for schooling by
estimating the following first stage:
Sicst = π11(dropout agecs = 16)+π21(dropout agecs ≥ 17)+
αc +θs +φsbirth yearc +ηt +Witγ +υicst . (3)
Using the predicted values from equation (3), IV estimates of equation (2) identify the local av-
erage causal response for dropout law compliers (Angrist and Imbens 1995), i.e. the effect of an
additional year of schooling on voters that only completed an additional year of schooling due to a
change in their state’s dropout age.7
In addition to the parallel trends assumption needed to estimate the effect of the minimum
school leaving age, identification further requires a strong first stage, that the instruments satisfy
monotonicity, and an exclusion restriction requiring that the dropout age only affects political
outcomes through increased schooling. First, the results in Table 3 below confirm that increasing
the dropout age significantly increases the level of schooling obtained. Second, it is unlikely that a
higher leaving age would cause an individual to choose less schooling, although Figure A3 in the
Appendix more formally supports the monotonicity assumption by showing that the cumulative
distribution of years of schooling for higher dropout ages lies strictly to the right of the distribution
for lower dropout ages. Third, additional education is temporally proximate to the point at which
the dropout age binds. Consequently, most downstream behavior is likely to be a function of a
7This estimand weights the reduced form effects of completing each grade of schooling by the proportioninduced by the respective instruments to achieve that grade.
16
respondent’s level of schooling. There is thus limited scope for changes in the minimum school
leaving age to affect political preferences through channels other than additional education, and
thereby violate the exclusion restriction. I reinforce this claim empirically below, showing that
there is little evidence to support plausible violations.
3.3 Data
I use data from the National Annenberg Election Survey (NAES) and the American Community
Survey (ACS). The NAES collates rolling surveys conducted throughout the 2000, 2004 and 2008
Presidential election campaigns. More than 50,000 randomly-sampled adults were interviewed by
telephone over the period of each campaign, and together these yield a maximum pooled sample of
176,796 respondents.8 In addition to its large sample size, a key advantage of the NAES over other
widely used political surveys is the wide-ranging set of questions.9 The diverse range of politi-
cal outcomes is essential for assessing the mechanisms underpinning education’s political effects.
To demonstrate the robustness of the results, I also examine the American National Election Sur-
vey (ANES); this serves as an important out-of-sample robustness check because it uses different
survey protocols and covers all Congressional elections since 1952.
The ACS, which has conducted interviews with U.S. citizens at monthly intervals since 2000,
supplements the NAES’s political questions by measuring the number of completed grades of
schooling. To match the years when NAES respondents were surveyed, I use ACS data from the
publicly-available microdata from the 2000, 2001, 2003, 2004, 2007 and 2008 samples (Ruggles
et al. 2010).
I now briefly describe how the main variables are measured. Summary statistics are provided
in Table 1. The Appendix provides detailed variable definitions.8Interviews were conducted for around a year over the following months: 12/1999-1/2001, 10/2003-
11/2004, 12/2007-11/2008.9The General Social Survey collects the number of years of schooling, but only registers a respondent’s
Census division at age 16 (there are only 9 divisions). Furthermore, questions regarding the political mech-anisms are asked irregularly or not at all.
17
3.3.1 Dependent variable: Republican support
I use three measures of support for the Republican party in the NAES. The first, which does not
rely on voter turnout, is partisan self-identification. In the sample, 31% of respondents identify as
Republicans, while 35% identify as Democrats; the residual are independents or “don’t know.”10
Throughout, I use an indicator to capture Republican partisanship. Secondly, I measure vote in-
tention at the forthcoming Presidential election, coding an indicator for intending to vote for the
Republican Presidential candidate (either George W. Bush or John McCain). Finally, I code an
indicator for whether a respondent voted Republican at the last Presidential election (for Bob Dole
or George W. Bush).11 In the sample, the Republican candidates received 47% of intended Pres-
idential votes while 45% of respondents actually voted for the Republican Presidential candidate
at the previous election.12 I show that the results are mirrored if Democrat indicators are used
instead.
3.3.2 Assigning state dropout ages
Survey respondents are mapped to the dropout age affecting their cohort based on their year of
birth and the state in which they resided when the minimum dropout age binds. Birth year cohort
is inferred from a respondent’s age when surveyed by the NAES, and given as year of birth in the
ACS.13 Survey respondents are then mapped to the dropout age in their state at the age at which
the law binds their decision to leave school. Since state of residence as a teenager is not available,
this is approximated by current state of residence in the NAES and state of birth in the ACS.14
10Considering only strong partisans and removing “don’t know” responses provides similar results.11Abstention is coded as 0, so result do not reflect turnout shifts.12This corresponds to 53% of voters that turned out for the Republican candidate. The Republican Presi-
dential candidate in the 1996, 2000, 2004 and 2008 elections ultimately received 40.7%, 47.9%, 50.7% and45.7% of the votes cast. Weighting by the NAES sample sizes, the vote share corresponding to the intendedvote measure is 48.7%, while 47.9% of voters reported voting Republican in the previous election.
13This approach is common and yields similar first stage results to studies using month of birth (Ace-moglu and Angrist 2000).
14Assigning dropout ages by current state of residence yields a similar first stage.
18
Tabl
e1:
Sum
mar
yst
atis
tics
Nat
iona
lAnn
enbe
rgE
lect
oral
Stud
yA
mer
ican
Com
mun
ities
Surv
eyO
bs.
Mea
nSt
d.de
v.M
in.
Max
.O
bs.
Mea
nSt
d.de
v.M
in.
Max
.
Dep
ende
ntva
riab
les
Rep
ublic
anpa
rtis
an17
6,79
60.
300.
460
1In
tend
tovo
teR
epub
lican
forP
resi
dent
144,
742
0.46
0.50
01
Vote
dR
epub
lican
forP
resi
dent
atla
stel
ectio
n12
5,19
60.
450.
500
1
Edu
catio
n(e
ndog
enou
sva
riab
les)
Com
plet
edgr
ades
443,
539
11.6
41.
330
12B
eyon
d12
thgr
ade
443,
539
0.52
0.50
01
Exc
lude
din
stru
men
tsD
ropo
utag
e=16
177,
653
0.72
0.45
01
443,
539
0.73
0.44
01
Dro
pout
age≥
1717
7,65
30.
260.
440
144
3,53
90.
250.
430
1
Pre
-tre
atm
entc
ontr
olva
riab
les
Age
177,
653
49.2
816
.67
1897
443,
539
49.3
016
.51
1895
Mal
e17
7,65
30.
440.
500
144
3,53
90.
440.
500
1W
hite
177,
653
0.87
0.34
01
443,
539
0.87
0.34
01
Bla
ck17
7,65
30.
080.
280
144
3,53
90.
080.
270
1A
sian
177,
653
0.01
0.08
01
443,
539
0.01
0.08
01
Coh
ort(
year
aged
14)
177,
653
1968
.59
16.4
019
2020
0444
3,53
919
68.6
916
.44
1920
2004
Surv
eyye
ar17
7,65
320
03.8
73.
0720
0020
0844
3,53
920
03.9
93.
0420
0020
08
Mec
hani
sms
Red
uce
tax
136,
870
0.35
0.48
01
Ban
abor
tion
131,
690
0.21
0.41
01
Low
gun
cont
rols
81,6
330.
380.
490
1L
owhe
alth
spen
ding
69,5
500.
310.
460
1hi
ghm
ilita
rysp
endi
ng78
,741
0.48
0.50
01
Rep
ublic
anno
n-ec
onom
icis
sues
163,
365
0.00
1.00
-1.1
52.
42Po
litic
alin
tere
stsc
ale
177,
624
0.00
1.00
-2.2
32.
33Po
litic
alkn
owle
dge
scal
e12
3,43
60.
001.
00-1
.19
1.36
Dis
cuss
polit
ics
168,
961
0.00
1.00
-1.1
81.
56Tu
rnou
t(pr
es.e
lect
ion)
150,
306
0.83
0.28
01
Trus
tfed
eral
gove
rnm
ent
54,6
310.
250.
430
1Pr
otec
tenv
iron
men
tmor
e12
6,52
20.
140.
350
1B
anga
ym
arri
age
72,3
350.
320.
470
1
19
The absence of state of residence at high school age could induce bias if compliers supporting
particular political parties systematically migrate to states with different dropout ages. However,
there is little evidence to justify this concern. First, relatively few respondents move across states:
61% of ACS respondents live in their state of birth, while 72% of ANES respondents live in
the state they lived in at age 14. Moreover, I show below that increasing the dropout age does
not increase the likelihood of attending college—a key source of cross-state migration. Third,
a bounding exercise in the Appendix demonstrates that, even under generous assumptions, the
selective migration required to account for the results is not plausible. Finally, the Appendix shows
that the ANES dataset—which measures state of residence at age 14—yields similar estimates.
3.3.3 Completed grades of high school
To translate the reduced form effect into an IV estimate for completing an additional grade of
schooling, a measure of the number of grades of completed schooling is required. The NAES
only measures completing high school, but does not distinguish between finer levels of partial
high school completion that different dropout ages could impact. While previous studies have
used an indicator for completing high school (e.g. Lochner and Moretti 2004; Milligan, Moretti
and Oreopoulos 2004), Marshall (2016) shows that this is likely to significantly upwardly bias
the magnitude of IV estimates. Intuitively, this is because coarsening years of schooling into a
dichotomous variable for completing high school causes the first stage to only capture the effect of
the instrument on completing high school, neglecting the effect of the dropout age on increasing
levels of schooling without inducing the completion of high school. Consequently, the exclusion
restriction is violated if any other level of schooling (induced by changes in the dropout age) affects
identifying as or voting Republican. For example, Marshall (2016) shows in the United Kingdom
that the effects of education are not solely a function of graduating from high school, but rather are
monotonically increasing in the number of years of schooling, and thus upwardly bias estimates of
20
completing high school by three times.15
Fortunately, an interval measure of schooling, like the number of completed grades of school-
ing, allows for consistent estimation of the local average causal response, provided—as in this
case—that multiple grades are affected by the instrument (Marshall 2016). This causal quantity
weights the local average treatment effect for each additional year by the extent to which leaving
age laws contribute to the first stage for each additional year of schooling.
Since years of schooling is not measured in the NAES (or the ANES), I combine the NAES data
with data from the ACS using two-sample IV methods (see below). The ACS provides fine-grained
data on each citizen’s education, and permits schooling to be measured by the number of grades
completed. To avoid complications with coding different types of post-high school education, the
variable is top-coded at completing 12th grade, and thus ranges from 0 to 12. This coding is
inconsequential because, as shown in Table 3 below, dropout ages do not affect schooling beyond
12th grade.
3.4 Two-sample IV estimation
For the IV estimates, I use two-sample 2SLS (TS2SLS) to estimate equation (2). This entails
estimating the first stage in equation (3) using the ACS dataset and the reduced form in equation
(1) using the NAES dataset, before efficiently combining the two as a consistent two-step estimator
of the effect of each additional grade of schooling (Inoue and Solon 2010).16 This approach was
first proposed by Angrist and Krueger (1992) and Franklin (1989), and has since been employed to
answer a variety of empirical questions (see Angrist and Pischke 2008). I derive the cluster-robust
covariance matrix for the TS2SLS estimator in the Appendix. The reduced form estimates require
only the NAES data, and thus do not rely on two-sample methods.
Beyond the standard IV assumptions discussed above, TS2SLS requires that the reduced form
15Table A7 in the Appendix provides IV estimates instrumenting for completing high school, and reportsimplausibly large estimates consistent with coarsening bias.
16TS2SLS is implemented in R; the program is available in the replication code.
21
and first stage datasets independently sample from the same population (Inoue and Solon 2010).
This in turn implies that key sample moments—variable means, variances and cross-products—are
identical in expectation, and are thus “exchangeable.” Both the ACS and NAES sample voting age
citizens, once ineligible voters from the NAES sample and those aged below 18 and born outside
the U.S. are removed from the ACS sample. To address possible differences in response rates, I
guarantee that the sample moments match by stratifying by year of birth, male, race and survey
year to randomly draw ACS observations to replicate the distribution of NAES respondents over
these pre-treatment covariates.17 Full details of this procedure are provided in the Appendix. The
summary statistics in Table 1 demonstrate that this approach almost exactly replicates the first
two moments for each variable in the samples. Encouragingly, dropout ages—which were not
matched—are also produce almost identical sample moments across the two datasets. Table A4 in
the Appendix shows similar first stage estimates without using this matching procedure.
4 Effects of high school dropout age and completed grades of
high school on Republican support
I now present the downstream political effects of raising the high school dropout age. The main
finding is that both higher dropout ages and each additional grade of high school substantially
increase the probability that an individual will identify with, or vote for, the Republican party later
in life.
4.1 Reduced form estimates: raising the dropout age
I first estimate the effect of increasing the school dropout age on the propensity of an individual
to identify as or vote Republican. Columns (1)-(3) in panel A of Table 2 report the reduced form
17By correcting for any differences between samples, TS2SLS is also consistent when sampling ratesdiffer with any control variable (Inoue and Solon 2010).
22
difference-in-differences estimates, where a positive coefficients signifies an increase in support
for the Republican party. Respectively, columns (1), (2) and (3) report the effects on identifying as
a Republican partisan, intending to vote for the Republican Presidential candidate, and voting for
the Republican Presidential candidate at the previous Presidential election. Columns (4)-(6) report
analogous measures for the Democratic party.
Table 2: The effect of school dropout age on identifying as and voting Republican
Republican support Democrat supportPartisan Intend Vote Partisan Intend Vote
(1) (2) (3) (4) (5) (6)
Dropout age=16 0.013 0.035*** 0.026* -0.017 -0.026** -0.010(0.011) (0.013) (0.014) (0.011) (0.012) (0.014)
Dropout age≥17 0.030** 0.050*** 0.043*** -0.029** -0.037*** -0.028*(0.012) (0.014) (0.016) (0.012) (0.013) (0.015)
Observations 176,796 144,742 125,196 176,796 144,742 125,196Outcome mean 0.30 0.46 0.45 0.33 0.42 0.36Test: Dropout age=16 = 0.002 0.010 0.006 0.019 0.050 0.004
Dropout age≥17 (p value)
Notes: Outcome variables are: identifying as a Republican/Democrat partisan (“partisan”), intendingto vote Republican/Democrat for President (“intend”), and voting Republican/Democrat in previousPresidential election (“vote”). All specifications include male, white, black and Asian dummies, quartic(demeaned) age polynomials, state-specific cohort trends, and state grew up, cohort and survey yearfixed effects, and are estimated using OLS. Differences in the number of observations reflect differencesin the NAES modules in which each outcome was included. The omitted dropout age category is dropoutage≤15. Standard errors clustered by state-cohort in parentheses. * p < 0.1, ** p < 0.05, *** p < 0.01.
The results show that raising the dropout age significantly increases Republican support. The
first coefficient estimate in each column indicates that, compared to cohorts in states requiring
students to remain in school until at most age 15, raising the dropout age to 16 increases the
proportion of Republican partisans by 1.3 percentage points, those intending to vote Republican
by 3.5 percentage points and those actually voting Republican by 2.6 percentage points. Relative
to the sample mean, these estimates imply around a 5% increase in Republican support among
cohorts affected by the reform, and thus indicate that changing the dropout age substantially alters
23
partisan identification and voting behavior.
To examine the more topical question relating to the impact of further increasing the dropout
age to 17 or above, the first coefficient in each column can be subtracted from the second. These
comparisons shows that, relative to cohorts in states requiring students to remain in school until age
16, raising the dropout age to 17 or higher increases the proportion of Republican partisans by 1.7
percentage points, those intending to vote Republican by 1.5 percentage points and those actually
voting Republican by 1.7 percentage points. Coefficient tests at the foot of Table 2 demonstrate that
each of these differences is statistically significant. The results again imply substantial increases
in support, of around 3-5%, for the Republican party. The slightly larger effect at 16 most likely
reflects the comparison with a more varied omitted category containing respondents facing laws
mandating that students only remain in school until between 12 and 15. Most students in this group
face a leaving age of 14.
Columns (4)-(6) show broadly commensurate decreases in Democrat support, suggesting that
increasing the dropout age causes would-be Democrat supporters to become Republican support-
ers. For both dropout age indicators, comparing the increase in Republican support to the decrease
in Democrat support indicates that relatively few voters become non-partisans or independents,
vote for third-party candidates or abstain. At least among those affected by the reform, this sug-
gests that staying in high school is able to reverse the powerful early life forces that otherwise drive
voters to become Democrats.
4.2 Instrumental variables estimates: completing an additional grade of late
high school
The IV estimates identify the effect of completing an additional grade of high school for individuals
that only remained in school because of an increase in the dropout age affecting their state-cohort.
In contrast with the reduced form estimates, which aggregate across entire state-cohorts, these
24
estimates apply only to those individuals induced to remain in education by dropout law reforms.
In Table 3, the first stage estimates from the ACS sample is presented in column (1), column (2)
examines the effect on post-high school education, and the IV estimates are presented in columns
(3)-(5).18
Table 3: The effect of an additional completed grade of high school on identifying as and votingRepublican
Completed Beyond Republican supportgrades 12th grade Partisan Intend VoteOLS OLS TS2SLS TS2SLS TS2SLS(1) (2) (3) (4) (5)
Dropout age=16 0.248*** -0.013*(0.037) (0.007)
Dropout age≥17 0.284*** -0.007(0.040) (0.008)
Completed grades 0.089* 0.165*** 0.140**(0.046) (0.055) (0.058)
First stage (ACS) observations 443,539 443,539 443,539 436,333 443,539Reduced form (NAES) observations 176,796 144,742 125,196Outcome mean 11.64 0.51 0.33 0.42 0.36First stage F statistic 25.4 2.8 25.4 25.4 25.4Test: Dropout age=16 = 0.004 0.107
Dropout age≥17 (p value)
Notes: See Table 2. First stage observations decline for vote intention because the ACS sample isreduced to match the variables in the NAES sample.
The first stage estimates in column (1) show that raising the dropout age significantly increases
schooling. Relative to state-cohorts facing a dropout age below 15, raising the leaving age to 16
increases the average number of grades of schooling completed by 0.25 grades. Further raising the
dropout age to at least 17 keeps students in school for an additional 0.04 grades. These positive
effects are broadly similar in magnitude to previous estimates pertaining to cohorts born somewhat
earlier in the twentieth century (Acemoglu and Angrist 2000; Dee 2004; Stephens and Yang 2014).
18The results for Democrat support are similar to Table 2, and henceforth omitted to save space.
25
Moreover, the F statistic at the foot of column (1) indicates that the dropout age instrument is
sufficiently strong to avoid weak instrument biases.
To understand which levels of schooling the IV estimates pertain to, it is important to examine
which levels of education are affected by the instruments. While column (1) demonstrated that
raising the dropout age increases in the number of completed grades of schooling up to high school
level, it is possible that the minimum school leaving age also influences post-secondary education.
However, column (2) shows that, on average, increasing the dropout age does not increase the
probability that an individual completes schooling beyond 12th grade (including any amount of
college).19 Importantly, the estimates in this article thus only apply to completing additional years
of high school, without also capturing potentially countervailing effects of attending college.
Reinforcing the reduced form findings, the TS2SLS estimates in panel C show that high school
has a large pro-Republican effect among those induced to remain in school by a higher dropout
age. Columns (3)-(5) show that an additional grade of late high school substantially increases
the probability that an individual identifies as a Republican or supports the Republican Presiden-
tial candidate. This effect is particularly large for vote choices, where an additional year of late
high school increases the probability of voting Republican by approximately 15 percentage points.
Nevertheless, the 9 percentage point increases in Republican partisan identification is also notable
and, combined with the reduced form estimates, suggests that high school education has major
downstream conservative effects on voters. These estimates likely reflect particularly large effects
of late high school among a group of compliers from relatively disadvantaged backgrounds where
education is uncommon, and thus the marginal effects of schooling on life opportunities may be
greatest.
19Acemoglu and Angrist (2000) and Lochner and Moretti (2004) also find that university attendance isunaffected in different samples.
26
Table 4: Heterogeneity in the effect of school dropout age on schooling and identifying as andvoting Republican, by black and Hispanic voters
Completed Republican supportgrades Partisan Intend Vote
(1) (2) (3) (4)
Dropout age=16 0.047 0.004 0.020 0.019(0.042) (0.013) (0.014) (0.015)
Dropout age=16 × Black or Hispanic 0.917*** 0.042** 0.063*** 0.058**(0.139) (0.018) (0.020) (0.027)
Dropout age≥17 0.067 0.025* 0.037** 0.032*(0.045) (0.014) (0.016) (0.016)
Dropout age≥17× Black or Hispanic 1.003*** 0.027 0.059*** 0.086***(0.140) (0.019) (0.021) (0.028)
Observations 443,539 176,796 144,742 125,196
Note: See Table 2.
4.3 Which voters do dropout age reforms induce to become Republicans?
It is essential to recognize which types of voters fail to complete high school in order to understand
which voters were induced, by school leaving age reforms, to support the Republican party. As
Murnane (2013) shows, blacks and Hispanics are significantly less likely than whites to graduate
from high school by age 20-24. Around 80% of each cohort of white voters born since World War
II have consistently completed high school. In contrast, only 60% of black and Hispanic students
born just after World War II completed high school, although the difference relative to whites
dropped below 10% for those born in the 1980s.
Given that rates of high school completion among whites are relatively high, changes in the
dropout age are likely to disproportionately influence black and Hispanic voters.20 The heteroge-
neous effects in column (1) of Table 4 show that increases in the dropout age indeed principally
impact respondents that identify as either black or Hispanic. In fact, most of the increase in com-
pleted school grades is driven by such students: the results show that for each dropout age category,
20Higher female high school graduation rates are also reflected in unreported results.
27
the increase in completed grades of school is primarily driven by black and Hispanic students—
especially in the case of raising the dropout age to 16.21
Consistent with the types of students induced to remain in school by higher school leaving
age laws, columns (2)-(4) confirm that increases in Republican support are also primarily driven
by blacks and Hispanics.22 In particular, the effect of raising the leaving age to 16 is entirely
concentrated among black or Hispanic students, while the effect of raising the leaving age to 17
or above is 2-4 times greater among such minorities. These results indicate that, because of their
lower baseline probability of remaining in high school, raising the school leaving age has induced
significant numbers of black and Hispanic voters to vote Republican.23 After demonstrating the
robustness of these estimates, I investigate the mechanisms underlying this effect.
4.4 Robustness checks
The main findings rest on two key identifying assumptions. The parallel trends assumption is
required for both the reduced form and IV estimates. The IV estimates additionally require an ex-
clusion restriction that the dropout age only affects political behavior through additional schooling.
I now show that the results are robust to potential violations of these assumptions.
4.4.1 Parallel trends
The state-specific cohort trends included in the baseline specifications provide a powerful check
against general parallel trend concerns. Another common test includes lags and leads of dropout
age reforms (see Angrist and Pischke 2008). If the estimates indeed pick up the effects of the
reform, rather than differential pre-trends across states that did and did not make reforms, leads
21The difference between the two lower-order dropout age coefficients is not quite statistically significant,but suggests that some white voters also remained in school longer in states raising the dropout age to 17 ormore.
22The IV estimates are omitted due to the weak first stage for non-black/Hispanic students.23Given that schooling and Republican support increase broadly commensurately, it is unlikely that dif-
ferences in how black and Hispanic voters respond to high school drive these heterogeneous effects.
28
-.05
0.0
5.1
Mar
gina
l effe
ct
-5 -4 -3 -2 -1 0 1 2 3 4Cohort relative to the reform
Effect of dropout age=16 on Republican partisanship
-.02
0.0
2.0
4M
argi
nal e
ffect
-5 -4 -3 -2 -1 0 1 2 3 4Cohort relative to the reform
Differential effect of dropout age>=17 on Republican partisanship-.0
50
.05
.1.1
5M
argi
nal e
ffect
-5 -4 -3 -2 -1 0 1 2 3 4Cohort relative to the reform
Effect of dropout age=16 on Republican vote intention
-.02
0.0
2.0
4.0
6M
argi
nal e
ffect
-5 -4 -3 -2 -1 0 1 2 3 4Cohort relative to the reform
Differential effect of dropout age>=17 on Republican vote intention
-.1-.0
50.0
5.1
.15
Mar
gina
l effe
ct
-5 -4 -3 -2 -1 0 1 2 3 4Cohort relative to the reform
Effect of dropout age=16 on Republican vote
-.02
0.0
2.0
4.0
6M
argi
nal e
ffect
-5 -4 -3 -2 -1 0 1 2 3 4Cohort relative to the reform
Differential effect of dropout age>=17 on Republican vote
Figure 2: Effect of raising the dropout age on cohorts either side of the reform date
Notes: All estimates from specifications including five pre- and post-reform variables. Bars represent95% confidence intervals.
should not affect Republican support. Figure 2 examines this expectation, comparing the five co-
horts before and after a reform. Consistent with the parallel trends assumption, there is limited
evidence of pre-trends and a relatively durable increase in Republican support among cohorts af-
fected by an increase in the minimum school leaving age. The increase in Republican support
among cohorts facing an increased dropout age of 16 is somewhat more volatile, while the addi-
tional effect of increasing the leaving age to 17 or above is precise and durable.
Nevertheless, these general checks—linear state-specific cohort trends, and lags and leads—
could still fail to capture certain more particular concerns. Accordingly, I also simultaneously
29
control for various potential confounds that vary across states and cohorts around the time when
respondents attended high school. First, I include the Presidential vote share by state at the most
recent election to control for changes in state political context when respondents were in high
school.24 Second, statewide labor market opportunities are proxied for by state personal income
per capita at ages 16 and 18. Third, I include the state’s pupil-teacher ratio to control for changes
in the quality of education. Finally, the black proportion of the state population is included to
control for the possibility that white voters start to support the Republican party as the share of
black voters rises. Columns (1)-(3) of Table 5 show that controlling for these state-level variables
at age 14, 16 and 18 does not meaningfully alter the results.
It is also possible that the types of states that changed their dropout laws were also starting
to follow a different, and potentially non-linear, trajectory that differentially influenced different
cohorts. As Figure 1 shows, many of the reforms were concentrated in the south and midwest—
areas that have experienced substantial economic and demographic changes over the past 50 years
in which this sample attended school. To ensure that other experiences which could differ across
cohorts facing different dropout ages within these states, column (4) of Table 5 includes cohort-
region fixed effects.25 The results show that, even when only exploiting changes in the minimum
schooling leaving age across cohorts within similar parts of the U.S., the dropout age and com-
pleted grades of schooling continue to substantially increase Republican support.
A final issue is that the electoral context of the survey differentially impacts certain types of
voter that may also have faced higher school leaving ages. I address this by including cohort-
survey year and state-survey year fixed effects, and thus leveraging the difference-in-differences
variation only from within each year of a Presidential election campaign. The results in column
(5) reinforce the robustness of the findings.
24At the cost of sample size, similar results are obtained when controlling for state House and state SenateRepublican seat shares.
25Census regions are: midwest, northeast, south and west.
30
Table 5: Further parallel trends checks
Cohort-Cohort- and state-region year
State-level controls at... fixed fixed...age 14 ...age 16 ...age 18 effects effects
(1) (2) (3) (4) (5)
Panel A: Reduced form effect on Republican partisanship
Dropout age=16 0.011 0.010 0.008 -0.000 0.013(0.012) (0.013) (0.012) (0.012) (0.011)
Dropout age≥16 0.028** 0.027* 0.026* 0.018 0.029**(0.013) (0.014) (0.013) (0.014) (0.013)
Observations 173,605 171,202 171,146 176,714 176,674Test: Dropout age=16 = Dropout age≥17 (p value) 0.003 0.003 0.001 0.001 0.003
Panel B: Reduced form effect on Republican vote intention
Dropout age=16 0.032** 0.026* 0.025* 0.024* 0.034***(0.014) (0.014) (0.014) (0.014) (0.013)
Dropout age≥16 0.045*** 0.040** 0.041*** 0.037** 0.049***(0.015) (0.016) (0.015) (0.015) (0.014)
Observations 142,265 140,365 140,235 144,677 144,742Test: Dropout age=16 = Dropout age≥17 (p value) 0.037 0.025 0.010 0.030 0.012
Panel C: Reduced form effect on Republican vote
Dropout age=16 0.018 0.016 0.021 0.017 0.025*(0.015) (0.016) (0.015) (0.015) (0.014)
Dropout age≥16 0.034** 0.031* 0.036** 0.032* 0.043***(0.017) (0.017) (0.017) (0.016) (0.015)
Observations 122,942 120,914 120,247 125,163 125,196Test: Dropout age=16 = Dropout age≥17 (p value) 0.027 0.032 0.042 0.022 0.008
Panel D: IV effect of schooling on Republican partisanship
Completed grades 0.167** 0.235*** 0.191** 0.244** 0.076*(0.073) (0.086) (0.089) (0.121) (0.048)
First stage F statistic 17.0 16.9 17.0 10.0 24.3
Panel E: IV effect of schooling on Republican vote intention
Completed grades 0.199** 0.243** 0.208** 0.297** 0.149***(0.085) (0.098) (0.101) (0.131) (0.053)
First stage F statistic 14.1 14.0 14.1 10.4 24.3
Panel F: IV effect of schooling on Republican vote
Completed grades 0.202** 0.268** 0.235** 0.304** 0.135**(0.088) (0.105) (0.109) (0.146) (0.059)
First stage F statistic 12.3 12.3 12.3 10.0 24.3
Notes: See Table 2. State-level controls include: the pupil-teacher ratio, the previous Republican Presidential vote share, state income per
capita, and the black proportion of the population. Differences in the number of observations in columns (1)-(3) reflects missingness on state-
level controls. Column (4) drops those born before 1909 due to the lack of observations for estimating cohort-region fixed effects. Column (5)
drops observations from cohort-survey year and state-survey year cells with 1 or fewer observations. The number of reduced form (NAES)
observations for IV specifications is identical to the corresponding reduced form outcome panel. Panels A-C are estimated using OLS and
panels D-F are estimated using TS2SLS.
31
4.4.2 Exclusion restriction
The preceding robustness checks provide strong support for the parallel trends assumption under-
pinning the reduced form estimates, but also suggest that the main estimates are robust to con-
trolling for plausible exclusion restriction violations. Nevertheless, it still remains possible that
the dropout age could have directly increased Republican support through other channels; this
would upwardly bias the IV estimates, although the reduced form estimates would remain unbi-
ased. Given the temporal proximity of a binding dropout age to the decision to remain in school,
the most plausible exclusion restriction violations reflect short-term or concurrent changes induced
by school leaving age reforms.
First, by simply keeping a student out of the labor force (rather than through the direct effects
of an additional year of schooling), a greater minimum leaving age could affect early life choices
like marriage and fertility. If, for example, schooling reduced early marriage and fertility, then
students may be more likely to avoid interaction with the welfare state and ultimately support
the Republican party. However, Table A6 in the Appendix shows that the dropout age does not
systematically decrease the likelihood that a respondent has ever been married or lower the age of
either a respondent’s oldest or youngest child living at home.
Second, raising (or lowering) the minimum leaving age could create cross-cohort spillovers,
causing older (younger) cohorts to behave more like treated cohorts. However, this would re-
duce between-cohort differences within states, and thus downwardly bias estimates of schooling’s
effects.
Finally, I conduct a sensitivity analysis to calculate the extent of exclusion restriction viola-
tion required to nullify the results. Using Conley, Hansen and Rossi’s (2012) most conservative
sensitivity test—their union of confidence intervals method—I show in the Appendix that almost
one half of the reduced form effect of the dropout age on the Republican vote outcomes must
operate through avenues other than schooling for the TS2SLS estimates to become statistically
32
insignificant at the 90% level. Given the preceding checks, such large violations are unlikely.
5 Mechanisms
Having shown that staying in high school increases support for the Republican party, I now ex-
amine the channels through which this substantial effect operates. I use the same difference-in-
differences and IV methods to test whether the dropout age and additional completed grades of
high school affect the mechanisms underpinning the income and socialization-based theories con-
sidered above.26 The results indicate that, by increasing income, schooling affects economic policy
preferences, which in turn increase the likelihood that such voters ultimately support the Republi-
can party.
5.1 Education increases income and opposition to taxation
I provide four important pieces of evidence consistent with the income-based RMR channel. First,
although potentially common to some alternative explanations, it is well-established that education
increases wages. As noted above, studies using school dropout ages as instruments consistently
show that each additional year of schooling increases wages at peak working age in the U.S. by
around 10% (see Acemoglu and Angrist 2000; Angrist and Krueger 1991; Oreopoulos 2006).
Second, although it is reasonable to expect voters to support the party that they expect will
benefit them in the future (Alesina and La Ferrara 2005), if education is driving opposition to taxa-
tion by significantly increasing an individual’s income, then this effect should be most pronounced
when the return to education is greatest—at an individual’s earnings peak. To examine this im-
plication, I compare the effects of the dropout age on the 40% of respondents aged between 45
and 65 at the date of the survey to all other respondents. Based on the ACS sample, wages are
26Although establishing the effects of potential mediators requires strong assumptions, examining a rangeof potential mediators and placebo-like tests can support some mechanisms and eliminate others.
33
highest between these ages.27 The results in Table 6 show that the effect of the dropout age on
support for the Republican party is indeed substantially higher among 45-65 year-olds: the inter-
actions indicate that, at an individual’s earnings peak, the effect of raising both dropout age levels
on identifying as and voting Republican are around 15 percentage points higher, while intention to
vote Republican is also higher but is not statistically significant. This indicates that realizing the
returns to education is a key component of education’s political effect, although the estimates still
suggest a non-zero effect outside the peak earnings period. These results further imply that even
though voter policy preferences are likely to be sticky from early adulthood (Erikson and Tedin
2015; Stoker and Bass 2011), they can nevertheless respond to changing economic incentives over
their lifecycle.
The lack of differential effect on years of schooling in column (4) shows that these earnings
peak findings do not reflect the possibility that higher minimum schooling leaving ages produced
larger effects on education for a particular generation. Moreover, columns (5)-(8) show that the
results are robust to simultaneously controlling for the interaction of the earnings peak indicator
with all other fixed effects and covariates. Most importantly, such interactive controls ensure that
the results reflect variation across voters within cohorts around their earnings peak, and therefore
are not driven by certain cohorts generally behaving differently around their earnings peak.
Third, I use survey attitudes to examine whether additional grades of schooling increase op-
position to taxation—the key driver of the RMR model. The IV estimate in column (1) of Table
7 shows that educated respondents regard themselves as more conservative when asked to place
themselves on a (standardized) 100-point liberal-to-conservative scale; Table A8 in the Appendix
shows similar reduced form estimates. An additional year of late high school thus equates to almost
a quarter of a standard deviation increase in conservative preferences. Although economic policy
preferences are an important element determining how voters place themselves on such scales (e.g.
27I regressed (log) wage on dummies for age, controlling for state and cohort fixed effects. The coeffi-cients for ages 45-65 systematically produced the largest coefficients (ignoring the tiny fraction of workingrespondents aged 80 or higher).
34
Table 6: Heterogeneous effects of dropout ages on Republican support by earnings peak
Partisan Intend Vote Schooling(1) (2) (3) (4)
Panel A: Baseline estimatesDropout age=16 0.012 0.034*** 0.018 0.216***
(0.011) (0.013) (0.015) (0.037)Dropout age=16 × Earnings peak 0.143*** 0.059 0.169*** -0.118
(0.034) (0.066) (0.042) (0.108)Dropout age≥17 0.029** 0.048*** 0.034** 0.238***
(0.012) (0.014) (0.017) (0.038)Dropout age≥17 × Earnings peak 0.142*** 0.063 0.168*** -0.141
(0.035) (0.066) (0.043) (0.108)
Observations 176,796 144,742 106,879 440,892
Panel B: Including earnings peak interactive controlsDropout age=16 0.001 0.018 0.008 0.126***
(0.012) (0.014) (0.017) (0.039)Dropout age=16 × Earnings peak 0.183*** 0.101 0.190*** -0.112
(0.047) (0.084) (0.055) (0.134)Dropout age≥17 0.012 0.027* 0.015 0.152***
(0.013) (0.016) (0.019) (0.040)Dropout age≥17 × Earnings peak 0.189*** 0.126 0.198*** -0.164
(0.049) (0.086) (0.057) (0.138)
Observations 176,796 144,742 106,879 440,892
Notes: See Table 2. Earnings peak is an indicator for respondents aged between 45 and 65. Thespecifications in panel B interact all covariates with the earnings peak indicator. Differences in samplesize reflect data availability.
Conover and Feldman 1981), this evidence does not necessarily pinpoint economic policies as the
source of the change. To better capture economic policy preferences, column (2) shows that an
additional grade of high school increases the belief that taxes should be reduced by 14 percentage
points. Furthermore, this belief is strongly correlated with identifying with or voting for the Re-
publican party,28 and thus consistent with the argument that high school causes voters to become
28While this link from potential mediator to outcome cannot be well-identified, specifications akin toequation (1) show that voters believing that taxes should be reduced are more than 10 percentage pointsmore likely to be Republican partisans or vote for the Republican Presidential candidate.
35
more fiscally conservative, and ultimately support the Republican party.
Fourth, if voters adopt the policy positions of the political party or candidate they identify
with (e.g. Lenz 2012; Zaller 1992), changes in economic policy preferences could instead re-
flect changes in partisanship. If voters reflect their party’s positions, they should also shift toward
prominent Republican stances on other issues. Columns (3)-(6) of Table 7 provide no evidence
for such changes: an additional completed grade of high school does not significantly increase
the likelihood that respondents oppose abortion, gun control, school spending, health spending,
or cutting military spending.29 The effects of schooling on these issues are consistently smaller
in magnitude than schooling’s effect on economic policy preferences, while the coefficients for
military spending, gun controls and banning abortion point in the opposite direction to what the
correlation with Republican support would suggest. Combining all the non-economic policy indi-
cators as a scale, column (7) shows that schooling does not shift respondents toward Republican
positions on non-economic policy issues. These results are in line with Gerber, Huber and Wash-
ington (2010), who find that experimentally inducing a shift in the party a voter identifies with
does not affect their policy positions. In sum, these tests strongly suggest that it is economic policy
preferences that drive the political choices of those receiving more high school education.
5.2 No evidence of political socialization
The socialization theories reviewed above argue that schooling cultivates political engagement and
induces voters to adopt socially liberal values. Although such changes are typically associated with
supporting the Democratic Party, it remains possible that socialization mechanisms operate along-
side income channels or instead drive support for the Republicans. However, I find no evidence
for the basic mechanisms underpinning these theories.
The civic education and social network hypotheses emphasize the importance of changes in
political engagement in changing political attitudes. However, consistent with existing research
29These issues are all strongly correlated with identifying as and voting Republican in the NAES surveys.
36
Tabl
e7:
The
effe
ctof
anad
ditio
nalc
ompl
eted
grad
eof
high
scho
olon
pote
ntia
linc
ome
and
soci
aliz
atio
nm
echa
nism
s
Con
serv
ativ
eR
educ
eB
anL
owL
owH
igh
Rep
ublic
ansc
ale
taxe
s†ab
ortio
n†gu
nhe
alth
mili
tary
non-
econ
omic
cont
rols
†sp
endi
ng†
spen
ding
†is
sues
(1)
(2)
(3)
(4)
(5)
(6)
(7)
Com
plet
edgr
ades
0.21
9**
0.13
7***
-0.0
38-0
.078
0.06
4-0
.028
-0.1
23(0
.092
)(0
.053
)(0
.048
)(0
.065
)(0
.072
)(0
.066
)(0
.093
)
Red
uced
form
(NA
ES)
obse
rvat
ions
172,
400
136,
830
131,
662
81,0
2269
,513
78,7
0316
3,31
4Fi
rsts
tage
(AC
S)ob
serv
atio
ns44
3,53
944
1,12
044
1,12
030
6,63
630
9,41
730
9,41
744
1,12
0Fi
rsts
tage
Fst
atis
tic25
.425
.725
.716
.717
.817
.825
.7
(8)
(9)
(10)
(11)
(12)
(13)
(14)
Polit
ical
Polit
ical
Dis
cuss
Turn
out
Trus
tPr
otec
tB
ankn
owle
dge
inte
rest
polit
ics
(pre
s.fe
dera
len
viro
.ga
ysc
ale
scal
eel
ectio
n)go
vt.†
mor
e†m
arri
age†
Com
plet
edgr
ades
-0.0
93-0
.048
-0.1
100.
037
0.01
3-0
.063
0.19
3*(0
.104
)(0
.100
)(0
.083
)(0
.037
)(0
.072
)(0
.040
)(0
.111
)
Red
uced
form
obse
rvat
ions
123,
411
177,
624
168,
961
150,
266
54,6
3112
6,52
272
,313
Firs
tsta
ge(A
CS)
obse
rvat
ions
414,
146
443,
539
443,
539
441,
120
438,
752
311,
836
185,
840
Firs
tsta
geF
stat
istic
24.7
25.4
25.4
25.7
25.1
17.8
8.3
Not
es:
See
Tabl
e2,
and
the
App
endi
xfo
rde
taile
dva
riab
lede
finiti
ons.
All
outc
ome
vari
able
sar
est
anda
rdiz
ed,e
xcep
tbin
ary
outc
omes
deno
ted
by†.
All
spec
ifica
tions
are
estim
ated
usin
gT
S2SL
S.D
iffer
ence
sin
sam
ple
size
refle
ctda
taav
aila
bilit
y.
37
(Galston 2001), columns (8) and (9) of Table 7 show that an additional completed grade of high
school does not significantly affect, and if anything reduces, multi-item scales of political knowl-
edge or political interest. Furthermore, columns (10)-(12) identify no link between schooling and
greater discussion of politics with friends or family, self-reported turnout in the last Presidential
election, or trust in the federal government. These findings suggest that voters induced to remain in
high school by higher dropout ages are not induced to engage and participate in politics more, and
thus that changes in partisanship and vote choice reflect party switching rather than mobilization.30
Moreover, policy and issue preferences do not change in accordance with socialization expla-
nations. As noted above, there is no evidence to suggest that greater schooling is associated with
socially liberal positions on gun control or abortion. Columns (13) and (14) further show that
completing an additional grade of high school is not significantly correlated with support for envi-
ronmental protection, and if anything increases support for constitutionally banning gay marriage.
In combination with not voting for the Democrats, I thus find no evidence to suggest that voters
become more socially liberal on civil liberties issues.
6 Implications of raising the dropout age for electoral outcomes
The preceding analysis has identified large political effects of implementing policies to increase
the dropout age. The increases in Republican support are particularly notable, given that many
students in any given cohort would remain in school regardless of the dropout age. This suggests
that, at least in politically competitive areas, the changes in voting behavior could have significantly
altered state-level electoral outcomes, at least once the number of affected cohorts accumulated to
reach a critical mass.31
Quantifying such effects on electoral outcomes is challenging for at least two reasons. First,
30Recent studies highlight college’s positive correlation with political engagement (e.g. Henderson andChatfield 2011; Mendelberg, McCabe and Thal forthcoming). The lack of change in engagement may thusreflect dropout laws not impacting college attendance.
31The NAES does not consistently ask about state elections.
38
it is difficult to estimate the proportion of the electorate affected by dropout age reforms at any
given election. Second, as shown above, the effects of schooling become most pronounced in
later life when the returns to education are greatest. Both challenges imply that the effect of
raising the dropout age on electoral outcomes may only become noticeable many decades after the
reform, once those affected constitute a large proportion of the electorate and are most likely to
vote according to their higher midlife income. To tentatively estimate the electoral effects of raising
the minimum school leaving age, I use the following difference-in-differences specification:
Yst = δ11(dropout agest−45 = 16)+ δ21(dropout agest−45 ≥ 17)+θs +ηt +ϕsyeart + εst , (4)
where Yst is a measure of state-level Republican political representation in state s in election year
t between 1959 and 2010. The dropout age is lagged by 45 years—approximately the number of
years after which those first affected by the reform reach the end of their earnings peak—to allow
the number of affected cohorts to accumulate over time and reach the peak effect of high school’s
influence on political behavior. State-specific trends, ϕsyeart , are again included to address the
possibility that trends in Republican representation in those states which implemented a leaving
age reform 45 years earlier are not parallel to those that did not.
The results in Table 8 suggest that prior dropout age reforms have significantly influenced state
electoral outcomes. Columns (1) and (3) show that increasing the leaving age to 16 increased the
Republican seat share in the state House and Senate 45 years later by 3.8 and 5.1 points respec-
tively. These estimates are in line with the individual-level estimates in Table 2. Although raising
the leaving age to 17 or above similarly increased Republican representation in state chambers,
relative to dropout ages below 16, it does not increase the seat share beyond raising the leaving
age to 16. This could reflect the fact that such reforms became more frequent later in the twentieth
century, and thus the effects have yet to fully materialize. Columns (2) and (4) instead use a lin-
ear measure of the dropout age and produce similar results. However, it is essential to emphasize
39
Table 8: Effect of dropout age reform on downstream Republican state-level seat shares
Republican state Republican SenateHouse seat share Senate seat share
(1) (2) (3) (4)
Dropout age (45 year lag)=16 0.038*** 0.051***(0.013) (0.016)
Dropout age (45 year lag)≥17 0.032 0.039(0.020) (0.025)
Dropout age (45 year lag) 0.014** 0.012(0.006) (0.009)
Obervations 1,121 1,121 1,043 1,043Outcome mean 0.45 0.45 0.45 0.45Test: Dropout age (45 year lag)=16 = 0.729 0.607
Dropout age (45 year lag)≥17 (p value)
Notes: See Table 2. All specifications include state and election fixed effects and state-specific trends,and are estimated using OLS. Differences in the number of observations reflect missing data.
that such estimates are tentative, given both the challenges noted above and the large time gap be-
tween treatment and outcome. Nevertheless, they suggest that the dropout age may have produced
important electoral consequences that demand further investigation.
7 Conclusion
Despite education’s key role in determining students’ life trajectories, this is the first study to
identify whether and how high school education impacts political preferences and voting behavior
later in life. While previous studies have linked education to civic and political participation,
this article gives schooling a clear partisan tint by showing that it causes voters to support more
conservative on redistributive issues, and vote on the basis of such policy preferences. The findings
sharply contrast with previous studies highlighting the stickiness of partisan preferences: I find
that raising the dropout age by a year increases Republican partisan identification and voting by
1-3 percentage points per cohort, while completing an additional grade of high school increases
40
the probability that a voter will identify with or vote for the Republican party later in life by around
10 percentage points.
The large magnitude of high school’s political effects highlight the importance of school dropout
laws. Beyond the individual-level analysis, my tentative analysis suggests that the Republican state
seat share is significantly greater 45 years after the state dropout age was raised. The results thus
imply that increasing the dropout age may have significantly altered electoral outcomes, which
are likely to have in turn influenced policy outcomes and the welfare of the electorate (e.g. Lee,
Moretti and Butler 2004). In the context of significant efforts, particularly from Democrats, to
increase dropout ages, my findings pose an important strategic problem for Democrats aiming to
simultaneously help poorly-educated voters while winning elections. This problem is magnified
by the fact that it is the disadvantaged voters, such as minorities, which the Democrats seek to help
that ultimately turn to the Republican party after realizing the returns to schooling. Conversely,
the results suggest that Republican opposition—e.g. in South Carolina in early 2016—may be
misguided from a long-run strategic perspective.
Both income and socialization theories could explain high school education’s conservative ef-
fects. However, an explanation combining human capital theory (including evidence that there is a
significant economic return to schooling) with the RMR model of income-based policy preferences
receives significant empirical support. Consistent with this explanation, I show that additional high
school education causes voters to support low tax rates, without adopting Republican positions on
non-economic issues. Furthermore, the largest effects of education are registered around voters’
mid-life earnings peak. Conversely, there is no systematic evidence suggesting that additional
schooling affects political engagement or socially liberal values.
However, school dropout ages only significantly increase levels of high school education. Con-
sequently, the results in this article speak only to high school education. While high school remains
an important policy issue, given that nearly 20% of students—particularly black and Hispanic
students—still fail to graduate from high school (Murnane 2013), another key question is whether
41
the effects of college education differ. There remains considerable debate over this issue. On
one hand, many scholars argue that college education is particularly effective at instilling socially
liberal values, whether through curricula or social interactions and networks. On the other, re-
cent research suggests that colleges with a culture emphasizing financial gain may cause students
to become economically conservative (Mendelberg, McCabe and Thal forthcoming). Although
Mendelberg, McCabe and Thal (forthcoming) focus on student attitudes, rather than downstream
voting behavior, their results nevertheless suggest that college can also produce powerful conser-
vative effects via a mechanism also linked to aspirations for financial gain. The extent to which
such findings generalize to all types of college or postgraduate education represent an important
question for future research extending my findings for high school.
42
References
Abrams, Samuel, Torben Iversen and David Soskice. 2010. “Informal Social Networks and Ratio-
nal Voting.” British Journal of Political Science 41:229–257.
Acemoglu, Daron and Joshua D. Angrist. 2000. “How Large Are Human Capital Externalities?
Evidence from Compulsory Schooling Laws.” NBER Macroeconomics Annual 15:9–59.
Alesina, Alberto and Eliana La Ferrara. 2005. “Preferences for Redistribution in the Land of
Opportunities.” Journal of Public Economics 89(5):897–931.
Alwin, Duane Francis, Ronald Lee Cohen and Theodore Mead Newcomb. 1991. Political attitudes
over the life span: The Bennington women after fifty years. Madison: University of Wisconsin
Press.
Angrist, Joshua D. and Alan B. Krueger. 1991. “Does Compulsory School Attendance Affect
Schooling and Earnings?” Quarterly Journal of Economics 106(4):979–1014.
Angrist, Joshua D. and Alan B. Krueger. 1992. “The Effect of Age at School Entry on Educational
Attainment: An Application of Instrumental Variables with Moments from Two Samples.” Jour-
nal of the American Statistical Association 87(418):328–336.
Angrist, Joshua D. and Guido W. Imbens. 1995. “Two-Stage Least Squares Estimation of Average
Causal Effects in Models With Variable Treatment Intensity.” Journal of the American Statistical
Association 90(430):431–442.
Angrist, Joshua D. and Jorn-Steffan Pischke. 2008. Mostly Harmless Econometrics: An Empiri-
cist’s Companion. Princeton, NJ: Princeton University Press.
Ashenfelter, Orley and Cecilia Rouse. 1998. “Income, schooling, and ability: Evidence from a
new sample of identical twins.” Quarterly Journal of Economics 113(1):253–284.
43
Bartels, Larry M. 2010. The Study of Electoral Behavior. In The Oxford Handbook of American
Elections and Political Behavior, ed. Jan E. Leighley. Oxford University Press pp. 239–261.
Becker, Gary S. 1964. Human Capital: A Theoretical and Empirical Analysis, with Special Refer-
ence to Education. University of Chicago Press.
Berinsky, A. J. and Gabriel S. Lenz. 2011. “Education and Political Participation: Exploring the
Causal Link.” Political Behavior 33(3):357–373.
Bobo, Lawrence and Frederick C. Licari. 1989. “Education and political tolerance testing the
effects of cognitive sophistication and target group affect.” Public Opinion Quarterly 53(3):285–
308.
Campbell, Angus, Philip E. Converse, Warren E. Miller and Donald E. Stokes. 1960. The American
Voter. New York: Wiley.
Card, David. 1999. The causal effect of education on earnings. In Handbook of Labor Economics
Volume 3, ed. Orley Ashenfelter and David Card. Elsevier pp. 1801–1863.
Conley, Timothy G., Christian B. Hansen and Peter E. Rossi. 2012. “Plausibly Exogenous.” Review
of Economics and Statistics 94(1):260–272.
Conover, Pamela Johnston and Stanley Feldman. 1981. “The origins and meaning of lib-
eral/conservative self-identifications.” American Journal of Political Science 25(4):617–645.
Converse, Philip E. and Gregory B. Markus. 1979. “Plus ca change: The new CPS election study
panel.” American Political Science Review 73(1):32–49.
Dee, Thomas S. 2004. “Are there civic returns to education?” Journal of Public Economics
88(9):1697–1720.
44
Erikson, Robert S. and Kent L. Tedin. 2015. American Public Opinion (9th Edition). New York:
Pearson.
Fisher, Patrick. 2014. Demographic Gaps in American Political Behavior. Westview Press.
Franklin, Charles H. 1989. “Estimation across data sets: two-stage auxiliary instrumental variables
estimation (2SAIV).” Political Analysis 1(1):1–23.
Galston, William A. 2001. “Political knowledge, political engagement, and civic education.” An-
nual Review of Political Science 4(1):217–234.
Gelbach, Jonah B. 2004. “Migration, the life cycle, and state benefits: How low is the bottom?”
Journal of Political Economy 112(5):1091–1130.
Gelman, Andrew, Park, Boris Shor, Joseph Bafumi and Jeronimo Cortina. 2010. Red State, Blue
State, Rich State, Poor State: Why Americans Vote the Way They Do. Princeton, NJ: Princeton
University Press.
Gerber, Alan S., Gregory A. Huber, David Doherty, Conor M. Dowling and Shang E. Ha. 2010.
“Personality and Political Attitudes: Relationships Across Issue Domains and Political Con-
texts.” American Political Science Review 104(1):111–133.
Gerber, Alan S., Gregory A. Huber and Ebonya Washington. 2010. “Party affiliation, partisanship,
and political beliefs: A field experiment.” American Political Science Review 104(4):720–744.
Goldin, Claudia D. and Lawrence F. Katz. 2008. The Race Between Education and Technology.
Cambridge, MA: Harvard University Press.
Green, Donald P., Bradley Palmquist and Eric Schickler. 2002. Partisan Hearts and Minds: Polit-
ical Parties and the Social Identities of Voters. Yale University Press.
45
Green, Donald P., Peter M. Aronow, Daniel E. Bergan, Pamela Greene, Celia Paris and Beth I.
Weinberger. 2011. “Does knowledge of constitutional principles increase support for civil liber-
ties? Results from a randomized field experiment.” Journal of Politics 73(2):463–476.
Hainmueller, Jens and Michael J. Hiscox. 2010. “Attitudes toward highly skilled and low-
skilled immigration: Evidence from a survey experiment.” American Political Science Review
104(1):61–84.
Henderson, John and Sara Chatfield. 2011. “Who Matches? Propensity Scores and Bias in the
Causal Effects of Education on Participation.” Journal of Politics 73(3):646–658.
Hillygus, D. Sunshine. 2005. “The missing link: Exploring the relationship between higher edu-
cation and political engagement.” Political Behavior 27(1):25–47.
Inoue, Atsushi and Gary Solon. 2010. “Two-Sample Instrumental Variables Estimators.” Review
of Economics and Statistics 92(3):557–561.
Jencks, Christopher, Marshall Smith, Henry Acland and Mary Jo Bane. 1972. Inequality: A Re-
assessment of the Effect of Family and Schooling in America. New York, NY: Basic Books.
Jennings, M. Kent, Laura Stoker and Jake Bowers. 2009. “Politics across generations: Family
transmission reexamined.” Journal of Politics 71(3):782–799.
Kam, Cindy D. and Carl L. Palmer. 2008. “Reconsidering the Effects of Education on Political
Participation.” Journal of Politics 70(3):612–631.
Kotin, Lawrence and William F. Aikman. 1980. Legal Foundations of Compulsory Schooling. Port
Washington, NY: Kennikat Press.
Kroh, Martin and Peter Selb. 2009. “Inheritance and the dynamics of party identification.” Political
Behavior 31(4):559–574.
46
Lee, David S., Enrico Moretti and Matthew J. Butler. 2004. “Do voters affect or elect policies?
Evidence from the US House.” Quarterly Journal of Economics 119(3):807–859.
Lenz, Gabriel S. 2012. Follow the Leader? How Voters Respond to Politicians’ Policies and
Performance. University of Chicago Press.
Levendusky, Matthew. 2009. The partisan sort: How liberals became Democrats and conserva-
tives became Republicans. University of Chicago Press.
Lochner, Lance and Enrico Moretti. 2004. “The Effect of Education on Crime: Evidence from
Prison Inmates, Arrests, and Self-Reports.” American Economic Review 94(1):155–189.
Marshall, John. 2016. “Coarsening bias: How instrumenting for coarsened treatments upwardly
biases instrumental variable estimates.” Political Analysis 24(2):157–171.
McCarty, Nolan, Keith T. Poole and Howard Rosenthal. 2006. Polarized America: The Dance of
Ideology and Unequal Riches. MIT Press.
Meltzer, Allan H. and Scott F. Richard. 1981. “A rational theory of the size of government.”
Journal of Political Economy 89(5):914–927.
Mendelberg, Tali, Katherine T. McCabe and Adam Thal. forthcoming. “College Socialization and
the Economic Views of Affluent Americans.” American Journal of Political Science .
Meredith, Marc. 2009. “Persistence in Political Participation.” Quarterly Journal of Political Sci-
ence 4(3):187–209.
Milligan, Kevin, Enrico Moretti and Philip Oreopoulos. 2004. “Does education improve citizen-
ship? Evidence from the United States and the United Kingdom.” Journal of Public Economics
88(9):1667–1695.
47
Molloy, Raven, Christopher L. Smith and Abigail Wozniak. 2011. “Internal Migration in the
United States.” Journal of Economic Perspectives 25(3):173–196.
Murnane, Richard J. 2013. “U.S. High School Graduation Rates: Patterns and Explanations.”
Journal of Economic Literature 51(2):370–422.
Murphy, Kevin M. and Robert H. Topel. 1985. “Estimation and Inference in Two-Step Econometric
Models.” Journal of Business and Economic Statistics 20(1):88–97.
Nie, Norman H., Jane Junn and Kenneth Stehlik-Barry. 1996. Education and Democratic Citizen-
ship in America. University of Chicago Press.
Niemi, Richard G. and Jane Junn. 1998. Civic Education: What Makes Students Learn. Yale
University Press.
Oreopoulos, Philip. 2006. “Estimating Average and Local Average Treatment Effects of Education
when Compulsory Schooling Laws Really Matter.” American Economic Review 96(1):152–175.
Oreopoulos, Philip. 2009. Would More Compulsory Schooling Help Disadvantaged Youth? Evi-
dence from Recent Changes to School-Leaving Laws. In The Problems of Disadvantaged Youth:
An Economic Perspective, ed. Jonathan Gruber. University of Chicago Press pp. 85–112.
Romer, Thomas. 1975. “Individual welfare, majority voting, and the properties of a linear income
tax.” Journal of Public Economics 4(2):163–185.
Ruggles, Steven, J. Trent Alexander, Katie Genadek, Ronald Goeken, Matthew B. Schroeder and
Matthew Sobek. 2010. “Integrated Public Use Microdata Series: Version 5.0.”.
Sondheimer, Rachel M. and Donald P. Green. 2010. “Using Experiments to Estimate the Effects
of Education on Voter Turnout.” American Journal of Political Science 41(1):178–189.
Spence, Michael. 1973. “Job market signaling.” Quarterly Journal of Economics 87(3):355–374.
48
Stephens, Jr, Melvin and Dou-Yan Yang. 2014. “Compulsory Education and the Benefits of
Schooling.” American Economic Review 104(6):1777–1792.
Stoker, Laura and Jackie Bass. 2011. Political Socialization: Ongoing Questions and New Di-
rections. In The Oxford Handbook of American Public Opinion and the Media, ed. George C.
Edwards, III, Lawrence R. Jacobs and Robert Y. Shapiro. Oxford University Press pp. 453–465.
Verba, Sidney, Kay Lehman Schlozman and Henry E. Brady. 1995. Voice and Equality: Civic
Voluntarism in American Politics. Cambridge, MA: Harvard University Press.
Weakliem, David L. 2002. “The effects of education on political opinions: An international study.”
International Journal of Public Opinion Research 14(2):141–157.
Zaller, John. 1992. The Nature and Origins of Mass Opinion. Cambridge University Press.
49
A Appendix
Contents
A.1 Detailed variable definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A1
A.2 ANES data for robustness checks . . . . . . . . . . . . . . . . . . . . . . . . . . . A6
A.3 No evidence of political determinants of school leaving age reforms . . . . . . . . A11
A.4 Technical details for IV estimation . . . . . . . . . . . . . . . . . . . . . . . . . . A12
A.4.1 Support for the monotonicity assumption . . . . . . . . . . . . . . . . . . A12
A.4.2 Two-sample 2SLS estimation . . . . . . . . . . . . . . . . . . . . . . . . A13
A.4.3 Procedure for matching ACS and NAES sample moments . . . . . . . . . A18
A.5 Robustness checks referenced in the main article . . . . . . . . . . . . . . . . . . A19
A.5.1 Sensitivity to cross-state migration . . . . . . . . . . . . . . . . . . . . . . A19
A.5.2 First stage without sample matching restrictions . . . . . . . . . . . . . . . A20
A.5.3 Excluding state-specific cohort trends . . . . . . . . . . . . . . . . . . . . A21
A.5.4 Exclusion restriction tests . . . . . . . . . . . . . . . . . . . . . . . . . . A21
A.5.5 Upwardly biased estimates of completing high school . . . . . . . . . . . . A24
A.5.6 Reduced form mechanism estimates . . . . . . . . . . . . . . . . . . . . . A25
A.1 Detailed variable definitions
The variables used are defined in detail below, including source variable names corresponding to
National Annenberg Election Survey (NAES) and American Communities Survey (ACS) variables.
All NAES variable codes correspond to their definition in the codebook for their respective waves,
where wave 1 is the 2000 Presidential election, wave 2 is the 2004 Presidential election, and wave
3 is the 2008 Presidential election. Table 1 in the main article provides summary statistics.
A1
• (Republican) Partisan. Indicator coded 1 for the respondent identifying as a Republican.
Residual category is independents, Democrat partisans and “don’t know”. Non-answers
were deleted. This question was asked both before and after the Presidential election. NAES
variables “cV01” in wave 1, “cMA01” in wave 2, and “MA01” in wave 3.
• Intend (to vote Republican for President). Indicator coded 1 for respondents intending to
vote for the Republican Presidential candidate in the forthcoming election. The omitted
category contains any other vote, excluding non-voters and those who do not provide an
answer. This question is not asked in post-election surveys. NAES variable “cR23” in wave
1, “cRC03”, “cRC07” and “cRC10”-“cRC14” in wave 2, and “RCa03”, “RCa05”, “RCa07”
and “RCa08” in wave 3.
• Vote (Republican for President). Indicator coded 1 for respondents reporting that they voted
for the Republican Presidential candidate at the previous election (Bob Dole in 1996, George
W. Bush in 2000 and 2004, or John McCain in 2008). The omitted category contains any
other vote or not turning out. NAES variable “cR35” in wave 1, “cRD01” in wave 2, and
“RDc01” in wave 3.
• Male. Indicator variable for respondents identifying as male. Non-responses were deleted.
NAES variable “cW01” in wave 1, “cWA01” in wave 2, and “WA01” in wave 3; ACS vari-
able “sex”.
• Race. Three indicators for respondents identifying their race as white (including Hispanic,
because this is how the ACS is coded), black, or Asian (Chinese, Japanese and Other Asian
of Pacific Islanders in the ACS). The omitted category is all other races. Coded from NAES
variable “cW03” in wave 1, “cWA04” in wave 2, and “WC03” in wave 3; ACS variable
“race”.
• Black or Hispanic. Indicator coded 1 for respondents identifying their race as black or their
A2
ethnicity as Hispanic. Coded from NAES variable “cW03” in wave 1, “cWA04” in wave 2,
and “WC03” in wave 3; ACS variables “race” and “hispan”.
• Age. Years of age at date at which survey was conducted. Quartic versions of age are
standardized. NAES variable “cW02” in wave 1, “cWA02” in wave 2, and “WA02” in wave
3; ACS variable “age”.
• Cohort (year aged 14). Calculated as the year of the survey, minus the respondent’s age at
the date of the survey, plus 14. NAES variables “cW02” and “cdate” in wave 1, “cWA02”
and “cDATE” in wave 2, and “WA01” and “DATE” in wave 3; ACS variables “age” and
“year”.
• State grew up. Respondent’s current state of residence in NAES and state of birth in ACS.
NAES variable “cST” in wave 1, “cST” in wave 2, and “WFc01” in wave 3; ACS variable
name “bpl”.
• Survey year. Set of indicators for the year—2000, 2001, 2003, 2004, 2007 or 2008—in
which the survey was conducted. Because month of survey is not available in the ACS, the
precise survey completion date must be coarsened to year. NAES variable “cdate” in wave
1, “cDATE” in wave 2, and “DATE” in wave 3; ACS variable name “year”.
• Dropout ages. See definition in the Appendix of the main article. State-year dropout age
data from Oreopoulos (2009) were kindly provided by Philip Oreopoulos, and are based on
the National Center for Education Statistics’ Education Digest.
• Schooling. Number of completed years of grade school (excluding Kindergarten) or above;
top coded at 12 years of education (i.e. completing high school). Those that did not graduate
high school with a diploma are coded as 11. ACS variable name “educd”.
• Beyond 12th grade. Indicator coded 1 for respondents registering any education beyond the
high school level. Coded from ACS variable name “educ”.
A3
• Reduce taxes. Indicator coded 1 for respondents who strongly think taxes should be reduced;
“don’t know” was excluded. Coded from NAES variable “cBB01” in wave 1 (coded 1 for
respondents who see the amount Americans pay in taxes as an “extremely serious” problem),
“cCB13” in wave 2 (respond with “strongly” favoring tax reduction), and “CBb01” (respond
with “cut taxes”) in wave 3.
• Ban abortion. Indicator coded 1 for respondents who stated that abortion should never be
permitted. Coded from NAES variable “cBF03” in wave 1 (coded 1 for respondents who
agreed that the federal government should ban all abortions), “cCE01” in wave 2 (coded
1 for respondents who “strongly favor” or “somewhat favor” banning all abortions), and
“CEa01” in wave 3 (coded 1 for respondents who state that abortion is “not permitted under
any circumstances”).
• Low gun controls. Indicator coded 1 for respondents who answered that the federal govern-
ment should do “more” to restrict the kinds of guns that people can buy. Coded from NAES
variable “cBG06” in wave 1 and “cCE31” in wave 2; this was not asked in wave 3.
• Low health care spending. Indicator coded 1 for respondents who said that the federal gov-
ernment should spend more money on providing health care to those that do not already
have it. Coded from NAES variable “cBE02” in wave 1 and “cCC02” in wave 2; this was
not asked in wave 3.
• High military spending. Indicator coded 1 for respondents who said that the federal govern-
ment should spend more on military defense. Coded from NAES variable “cBJ07” in wave
1 and “cCD03” in wave 2; this was not asked in wave 3.
• Republican non-economic issues. Standardized summative rating combining the ban abor-
tion, do not increase gun controls, do not increase health care spending, and increase military
spending. Cronbach’s alpha inter-item reliability coefficient of 0.39, pooled across samples.
A4
• Political knowledge scale. Standardized summative rating scale combining correct/incorrect
answers to up to six general political knowledge questions, including: identifying Republi-
cans as more conservative than Democrats; the Congressional majority required to overturn
a Presidential veto; which body determines whether a law is constitutional; identity of the
majority party in the House; who the Vice-President is; ability to name own Senator. Cron-
bach’s alpha inter-item reliability coefficient of 0.60 in 2003/2004 and 0.60 in 2007/2008;
underlying variable were unavailable for 2000/2001. NAES variables “MC01”-“MC04” in
wave 3, and “cMC01”, “cMC03”, “cMC05”, “cMC07”, “cMC09” and “cUA02” in wave 2.
• Political interest scale. Standardized summative rating scale combining the following five
measures of interest in politics: four-point scale measuring regularity with which the respon-
dent reports following government and public affairs (“cK01” in wave 1, and “cKA03” in
wave 2); four-point scale measuring regularity with which the respondent reports follows
the Presidential campaign (“cK02” in wave 1, “cKA01” in wave 2, and “KA01” in wave 3);
number of days in the last week that the respondent watched a public or cable news program
(“cE01” and “cE02” in wave 1, and “cEA01” and “cEA03” in wave 2); number of days in
the last week (0 to 7) that the respondent read a daily newspaper (“cE13” in wave 1, and
“cEA01” in wave 2); number of days in the last week (0 to 7) that the respondent discussed
politics with family or friends (“cK05” in wave 1, “cKB01” in wave 2, and “KB01” in wave
3). Cronbach’s alpha inter-item reliability coefficient pooled across all surveys is 0.61.
• Discuss politics. Standardized number of days in the last week (0 to 7) that the respondent
discussed politics with family or friends. NAES variable “cK05” in wave 1, “cKB01” in
wave 2, and “KB01” in wave 3.
• Turnout (pres. election). Self-reported turnout in the most recent Presidential election.
• Trust federal government. Indicator coded 1 for respondents that trust the federal government
to what is right either “always” or “most of the time”; residual category includes “some of
A5
the time”, “never” and “don’t know”. NAES variable “cM01” in wave 1, “cMB01” and
“cMB02” in wave 2, and “MB01” in wave 3.
• Protect environment more. Indicator coded 1 for respondents who believe that the federal
government should do more to protect the environment; responses of the same, less or noth-
ing were coded as 0. NAES variable “cBS01” in wave 1 and “cCF08” in wave 2; this was
not asked in wave 3.
• Ban gay marriage. Indicator coded 1 for respondents that would favor a constitutional
amendment banning gay marriage.
• Age of eldest child. Age of oldest child in the respondent’s household. ACS variable name
“eldch”.
• Age of youngest child. Age of youngest child in the respondent’s household. ACS variable
name “yngch”.
• Single. Indicator coded 1 for respondents that have never been married. ACS variable name
“marst”.
A.2 ANES data for robustness checks
The American National Election Survey (ANES) collects many political variables which are pooled
across Presidential and mid-term elections 1952-2000 and 2008.32 Surveys typically interview
several thousand randomly-sampled voting-age citizens; pooling across elections produced a max-
imum sample of 35,873 respondents.33 The ANES thus has a far smaller sample size than the
NAES. Given the large number of fixed effects and trends included in the main specifications, the
greater sample size of the NAES is preferred. Nevertheless, the ANES has several advantages over
322002-2006 surveys were omitted because they did not ask where respondents grew up, a central com-ponent of the identification strategy.
33Observations suffering missingness were deleted.
A6
Figure A1: US years of schooling by cohort (Census data)
the NAES, and thus showing similar results is an important robustness check: the ANES covers a
wider range of election extending much further back in time; by extending further back in time,
the ANES also encompasses greater variation in schooling; and by asking respondents which state
they lived in at age 14, cross-state migration concerns are minimized.
Like the NAES, the number of years of schooling is not measured in the ANES. Accordingly,
this article complements the ANES with extracts from the 1950-2000 decennial Censuses. By
using decennial Census data instead of ACS surveys from the year of NAES survey, the assumption
that respondents are drawn from the same population is a little less likely to hold. Nevertheless, I
again ensure the Census sample of 636,713 individuals is approximately a random sample from the
same population the ANES is drawn from. To achieve this, I followed exactly the same stratified
procedure as for the ACS data, as detailed in the main article. The two datasets are again combined
using two-sample methods. Figure A1 shows average years of schooling by cohort, while Figure
A2 shows that the monotonicity check is satisfied.
The variables used in this analysis are defined in detail below, including source variable names
A7
020
4060
8010
0
Cum
ulat
ive
% o
f obs
erva
tions
0 5 10 15
Years of completed schooling
CSL<16 CDF CSL=16 CDF CSL>=17 CDF
Figure A2: Monotonic first-stage (Census data)
corresponding to ANES and Census variables. All ANES variables codes correspond to the Time
Series Cumulative Data File. Table A1 provides summary statistics.
• Republican partisan. Indicator coded 1 for the respondent identifying as a Republican, in-
cluding those weak partisans who when pressed leaned toward the Republicans. Residual
category is independents and Democrat partisans; non-answers were deleted. ANES variable
VCF0303.
• Vote Republican. Vote Republican for President, Vote Republican for House and Vote Re-
publican for Senate are indicators for voting Republican in any of these elections, including
at mid-terms. The omitted category contains Democrat and other party/candidate voters;
non-voters and those who do not provide an answer are deleted. ANES variables VCF0704,
VCF0707 and VCF0708.
• Gender. Indicator variable for respondents identifying as male. Non-responses were deleted.
A8
Tabl
eA
1:Su
mm
ary
stat
istic
s:A
NE
San
dC
ensu
ssa
mpl
es
Am
eric
anN
atio
nalE
lect
ion
Stud
yC
ensu
sO
bs.
Mea
nSt
d.de
v.M
in.
Max
.O
bs.
Mea
nSt
d.de
v.M
in.
Max
.
Dep
ende
ntva
riab
les
Rep
ublic
anpa
rtis
an35
,873
0.35
0.48
01
Vote
Rep
ublic
anfo
rPre
side
nt14
,712
0.48
0.5
01
Vote
Rep
ublic
anfo
rHou
se19
,081
0.43
0.5
01
Vote
Rep
ublic
anfo
rSen
ate
13,3
390.
450.
50
1
End
ogen
ous
vari
able
Scho
olin
g63
6,71
310
.75
2.62
012
Exc
lude
din
stru
men
tsD
ropo
utag
e=16
35,9
970.
750.
430
163
6,71
30.
740.
440
1D
ropo
utag
e≥17
35,9
970.
170.
370
163
6,71
30.
180.
380
1
Pre
-tre
atm
entc
ovar
iate
sA
ge35
,997
43.7
216
.04
1797
636,
713
43.0
716
.67
1893
Mal
e35
,997
0.44
0.5
01
636,
713
0.42
0.49
01
Whi
te35
,997
0.85
0.35
01
636,
713
0.86
0.35
01
Bla
ck35
,997
0.12
0.32
01
636,
713
0.12
0.33
01
Asi
an35
,997
00.
060
163
6,71
30
0.06
01
Yea
rage
d14
35,9
971,
950.
6019
.41,
914
1,99
663
6,71
31,
951.
2018
.81,
914
1,99
6C
ensu
sye
ar(d
ecad
e)35
,997
1,98
0.05
13.3
41,
950
2,00
063
6,71
31,
980.
2713
.41,
950
2,00
0
A9
ANES variable VCF0104; Census variable name “sex”.
• Race. Three indicators for respondents identifying their race as White or Hispanic, Black, or
Asian (Chinese, Japanese and Other Asian of Pacific Islanders in the Census). The omitted
category is all other races. Hispanics are grouped with whites under the Census. Coded from
ANES variable VCF0106a; Census variable “race”.
• Age. At time of survey. Quartic version of age is not standardized to ensure comparability
across ANES and Census datasets. ANES variable VCF0101; Census variable name “age”.
• Year aged 14. Calculated as the year of the survey, minus the respondent’s age at the date
of the survey, plus 14. Since no person in the Census could be 14 later than 1996, the
small number of ANES observations for respondents 14 after 1996 were dropped. Since no
person in the ANES was 14 before 1914, all such Census observations were dropped (see
above). 1914 is the omitted category when a full set of indicators is used. Coded from ANES
variables VCF0101 and VCF0104; Census variables “age” and “year”.
• State grew up. The ANES asks respondents what state they grew up in, taking the state
where more time between 6 and 18 was spent if respondents moved across states. The
Census asks which state respondents were born in, which requires a stronger no-migration
assumption when matching individuals to the state dropout ages. Foreign-born respondents
were excluded. ANES variable VCF0132 and VCF0133; Census variable “bpl”.
• Census decade. Indicator for each decade of the Census conducted, 1950-2000. For the
ANES, I count the nearest years as part of that Census decade; e.g. the ANES surveys
1986-1994 are coded for the 1990 Census indicator; the 2008 election is included with
the 2000 Census. Because TS2SLS estimation requires using exactly the same variables
across datasets, election-specific indicators (which would be possible if we were only using
the ANES) cannot be used. Coded from ANES variable VCF0004; Census variable name
A10
“year”.
• Dropout age. Same as ACS.
• Schooling. Same as ACS. Census variable name “educd”.
A.3 No evidence of political determinants of school leaving age reforms
To investigate the concern that school leaving age reforms may reflect strategic behavior by par-
ticular parties or reflect waves of political sentiment, I test whether a state’s minimum schooling
leave age is correlated with partisan political forces in state s in year t by estimating the following
difference-in-differences equation:
dropout agest = Pstλ +ϑs +ωt + ξst , (A1)
where Pst is a vector of state-level political variables, and θs and ωt are respectively state and year
fixed effects.
The results in Table A2 provide no evidence for the political endogeneity concern. Column (1)
shows no statistical association between the minimum dropout age and indicators for whether the
Republicans controlled the upper, lower or both state houses. Using the Republican seat shares
instead, column (3) again finds that political support is uncorrelated with changes in dropout ages.
Finally, column (5) finds no significant correlation with Republican state Governors. The lack
of such correlations suggest that minimum schooling leaving age reforms were unlikely to have
been implemented strategically, or have been correlated with waves in local political sentiment.34
Columns (2), (4) and (6) show that the results are also robust to including state-specific time trends.
34Unreported checks using the three leaving age indicators and lagging the independent variables tocapture delayed implementation similarly show no statistically significant associations.
A11
Table A2: Correlation between state political control and state minimum schooling leaving age
(1) (2) (3) (4) (5) (6)
Republican House majority -0.010 -0.079*(0.056) (0.041)
Republican Senate majority -0.047 -0.068(0.093) (0.061)
Unified Republican houses 0.007 0.044(0.075) (0.064)
Republican House seat share 0.194 -0.356(0.251) (0.251)
Republican Senate seat share 0.166 -0.093(0.271) (0.238)
Republican Governor 0.091* 0.031(0.054) (0.025)
Observations 3,890 3,890 3,890 3,890 3,725 3,725State-specific trends X X X
Notes: Outcome is the minimum dropout age in a given state. All specifications include state and yearfixed effects, and are estimated using OLS. Standard errors clustered by state in parentheses. * denotesp < 0.1, ** denotes p < 0.05 *** denotes p < 0.01.
A.4 Technical details for IV estimation
A.4.1 Support for the monotonicity assumption
Although monotonicity is fundamentally untestable, one implication of monotonicity with multi-
valued treatments is that the CDFs F(Si|dropout ageics = z) should not cross (Angrist and Imbens
1995). Figure A3 plots the CDFs for each dropout age instrument, and support the monotonicity
assumption as the CDF of higher dropout ages lie everywhere below the CDF of lower dropout
ages. Although important covariates such as state grew up in fixed effects are not included, these
figures provide an alternative graphical depiction of the first-stage relationship with the largest
changes in the relative CDFs occurring where the instruments apply. While this cannot prove
monotonicity holds, because the counterfactual is never observed and the controls used in the main
analysis are not used here, it is suggestive in that a possible violation receives no empirical support.
A12
020
4060
8010
0
Cum
ulat
ive
% o
f obs
erva
tions
0 5 10 15
Years of completed schooling
CSL<16 CDF CSL=16 CDF CSL>=17 CDF
Figure A3: Monotonic first-stage (ACS data)
A.4.2 Two-sample 2SLS estimation
The goal is estimate the following system of IV equations:
Yi = TiβT +Wiβ−T + ui = Xiβ + ui (A2)
Ti = ZiΠ+ εi, (A3)
where Xi includes exogenous covariates Wi and the treatment variable(s) Ti, while Zi includes Wi
and q excluded instruments. Identification requires that only p ≤ q treatment variables can be
instrumented for.
Two methods have been proposed for IV estimation with two samples. Angrist and Krueger
(1992) propose a Wald-style estimator where the reduced form estimates are divided by their first
stage counterparts, which can be generalized to the overidentified case where the number of in-
struments outnumber the number of endogenous variables. Inoue and Solon (2010) show that this
A13
estimator is less efficient than the 2SLS counterpart—first proposed by Franklin (1989)—that will
be used in the empirical application here. The advantage of this estimator is that it corrects for
finite-sample differences between the two samples.35 Furthermore, its extension to multiple in-
struments and multiple endogenous variables is straight-forward—both of which are important in
many empirical applications, including the analysis in this article.
In matrix form (stacking over i in each sample), the two-sample 2SLS (TS2SLS) estimator is:
βT S2SLS = (X ′1X1)
−1X ′1Y1, (A4)
where X1 = (T1,W1) is the matrix of predicted values in sample 1. The OLS regression coefficients
generating T1 are based on p first stage regressions estimated in sample 2:
X1 = Z1Π = Z1(Z′2Z2)−1Z′2X2. (A5)
The following assumptions are required to ensure the consistency of the TS2SLS estimator (see
Franklin 1989; Inoue and Solon 2010):
1. Random sampling from the same population: Y1i,Z1in1i=1 and T2i,Z2in2
i=1 are indepen-
dently and identically distributed draws of size n1 and n2 from the same population with
finite second moments.
2. Instrument exogeneity: E[Z′1iε1i] = E[Z′2iε2i] = 0.
3. Exclusion restriction: E[Z′1iu1i] = 0.
4. Rank conditions: (a) Z′1iZ1i and Z′2iZ2i have full rank, (b) X ′1iZ2i and X ′2iZ2i have full rank.
5. Interchangeable sample moments: (a) E[Z′1iX1i] = E[Z′2iX2i], (b) E[Z′1iZ1i] = E[Z′2iZ2i].
35Inoue and Solon (2010) show that the TS2SLS estimator remains consistent even when differences inthe sampling rates vary with some of the instrumental variables.
A14
Assumption 1 says that the samples must draw from the same population. Assumption 2 requires
that the instrument be exogenous in the first stage. Assumption 3 is implied by the exclusion re-
striction, but is written in terms of expectations. Assumption 4 is a standard rank condition required
for matrix invertibility. Assumption 5 requires that crucial samples moments can be interchanged,
thereby permitting substitution between samples. As n1 and n2 converge to the population size,
Assumption 5 necessarily holds.
Franklin (1989) proves the n1-consistency of the TS2SLS estimator. However, calculating the
TS2SLS standard errors is not obvious. Calculating the standard errors from a regression of Y1 on
X1 neglects the uncertainty in the first stage, in addition to distributional differences between the
first stage and reduced form samples.
The Murphy and Topel (1985) two-stage framework for understanding “generated regressors”—
accounting for the uncertainty introduced where a variable is estimated as a proxy to enter a
separate regression—incorporates such estimation uncertainty.36 Proposition 1 derives the ho-
moskedastic and cluster-robust variance (matrices), of which the robust variance is the particular
case of G1 = n1 and G2 = n2 clusters. (i is dropped to facilitate exposition.)
Proposition 1. The asymptotic variance of the TS2SLS estimator, V[β T S2SLS], is
[σ
2u +
n1
n2β
T S2SLS′S Ωβ
T S2SLSS
]E[X ′1X1]
−1, Ω = E[ε ′ε|X1] =
σ2
1 · · · σ1,p
... . . . ...
σp,1 . . . σ2p
(A6)
when the reduced form squared error σ2u = E[u2|X1] and the error covariances Ω of the p first
stage regressions are homoskedastic; when the reduced form and first stage errors are grouped
36Inoue and Solon (2010) acknowledge this approach but derive homoskedastic and heteroskedastic vari-ance matrices in an alternative way, but do not provide a cluster-robust variance estimate.
A15
into G1 and G2 clusters respectively, the cluster-robust variance is
E[X ′1X1]−1[
V[β T S2SLS]+n1
n2E[X ′1(β
T S2SLS′S ⊗Z1)]V(Π)E[(β T S2SLS′
S ⊗Z1)′X1]
]E[X ′1X1]
−1,(A7)
where β T S2SLSS is the vector of coefficients on p endogenous variables, the uncorrected TS2SLS
variance is given by V[β T S2SLS] = G1G1−1 ∑
G1g=1 E[X ′1gu1gu′1gX1g] and the variances from m first-
stage regressions are V(Π) = G2G2−1 Φ⊗E[Z′2Z2]−1, where
Φ =
E[Z′2Z2]−1
∑G2g=1 E[Z′2gε2g1ε ′2g1Z2g] · · · E[Z′2Z2]−1
∑G2g=1 E[Z′2gε2g1ε ′2gpZ2g]
... . . . ...
E[Z′2Z2]−1∑
G2g=1 E[Z′2gε2gpε ′2g1Z2g] . . . E[Z′2Z2]−1
∑G2g=1 E[Z′2gε2gpε ′2gpZ2g]
. (A8)
Proof : Start by separating X into its endogenous and exogenous components,
Yi1 = Xi1β−S +Ti1βS + ui = Xi1β−S + Ti1βS +[Ti1− Ti1]+ ui, (A9)
where Ti1 = Zi1Π = Zi1(Z′2Z2)−1Z′2T2 is the predicted value of the treatment using the first stage
estimates, and Ti1 is the true and unobserved treatment in sample 1. An OLS regression would
yield:
√n1
β−T −β−S
βS−βS
=
(1n1
X ′1X1
)−1 1√
n1X ′1u1 +
(1n1
X ′1X1
)−1 1√
n1X ′1[Ti1− Ti1]βS, (A10)
where subscripts i and superscripts T S2SLS are omitted to save space. Using the expansion result
A16
in Murphy and Topel (1985: 374) yields:
√n1(β −β ) ≡
√n1
β−T −β−S
βS−βS
a=
(1n1
X ′1X1
)−1 1√
n1X ′1u1
+
(1n1
X ′1X1
)−1(n1
n2
)1/2 1n1
X ′1(β′T ⊗Z1)
√n2(Π−Π),(A11)
where (β ′T ⊗Z1) is the matrix of defined in equation (12) of Murphy and Topel (1985).
Let Π be a consistent estimator of the first stage for the endogenous variables, such that√
n2(Π−Π)a∼ N(0,V(Π)). Using our consistent first stage estimate, the asymptotic variance
is therefore given by:
V(β −β ) = E[X ′1X1]−1[
V[β ]+n1
n2E[X ′1(β
′T ⊗Z1)]
−1V[Π]E[(β ′T ⊗Z1)′X1]
−1]
E[X ′1X1]−1,(A12)
where V[β ] is the variance of the naive TS2SLS estimator. (Note that E[X ′1u1] = 0, in conjunction
with a consistent first stage, implies the consistency of the estimator.)
This establishes the general asymptotic variance formula in Proposition 1. We now apply the
homoskedastic and cluster-robust error structures:
1) Homoskedastic errors. Under homoskedasticity, the naive variance from the TS2SLS re-
gression is simply σ2u (X
′1X1)−1. To correct for the first stage estimation, we have:
X ′1(β′T ⊗Z1)V(Π)(β ′T ⊗Z1)
′X1 = X ′1(β′T ⊗Z1)(Ω⊗ (Z′1Z1)
−1)(β ′T ⊗Z1)′X1 (A13)
= X ′1(β′T ΩβT ⊗Z1(Z′1Z1)
−1Z′1)X1 (A14)
= β′T ΩβT (X ′1X1), (A15)
where the first line uses the definitions of homoskedasticity given in the proposition, the sec-
ond line applies the mixed product property of Kronecker products, and the third line exploits
Z1(Z′1Z1)−1Z′1X1 = X1 (because all exogenous variables are contained in both X1 and Z1) and the
A17
fact that β ′T Ωβ ′T is a scalar. Substituting into the general variance matrix yields the homoskedastic
variance formula in Proposition 1.
2) Clustered errors. In the clustered case, we simply let V(Π) = G2G2−1 Φ⊗E[Z′2Z2]−1.
Standard errors are given by the square roots of the diagonal elements of V[β T S2SLS]/n1. Using
the analogy principle, expectations can be replaced by sample moments.
In the case of a single endogenous regressor, V(Π) is simply the standard cluster-robust vari-
ance matrix for the first stage:
E[Z′2Z2]−1
[G2
G2−1
G2
∑g=1
E[Z′2gε2gε′2gZ2g]
]E[Z′2Z2]
−1. (A16)
When there are multiple endogenous variables, the first stage estimates may be correlated across
models. This requires the more complex formulation in Proposition (1).
A.4.3 Procedure for matching ACS and NAES sample moments
As noted in the main text, I use stratified random sampling to ensure that the sampling moments of
the ACS data used to estimate the first stage match the analagous sampling moments in the NAES
dataset used to estimate the reduced form. This enables me to correct for observable differences
across these samples, which draw from the same population (once ineligible voters are removed
from the NAES and those aged below 18 and born outside the U.S. are removed from the ACS).
To match the sample moments, I first removed respondents aged 14 before 1920 (since no
NAES respondent was 14 before 1920) and respondents who grew up in Alaska or Hawaii, were
born outside the U.S., or became naturalized citizens. I then stratified by cohort-gender to draw
a random sample of 2,000,000 ACS respondents to ensure that the subsample has the same by-
cohort-by-gender distribution as the NAES. Finally, I randomly dropped respondents to reduce
the sample to ensure that the proportion of respondents from each survey year reflects the NAES
dataset distribution, before in turn randomly dropping respondents to ensure that the ACS sample
A18
matches the NAES distribution over race (white, black and Asian dummies). This produced a
maximum pooled sample of 443,539 voting age respondents.
A.5 Robustness checks referenced in the main article
A.5.1 Sensitivity to cross-state migration
Although the main article argues that cross-state migration is unlikely to be driving the results, I
nevertheless conduct two key robustness checks to show that this is unlikely to be a concern: a
bounding exercise that shows that the type of migration required to upwardly bias the estimates is
implausible; and analysis using the ANES, where state of residence at age 14 can be used to assign
the dropout age. These tests complement the lags and leads check in Figure 2, which is hard to
reconcile with migration unless most of those driving the results moved immediately as the reform
came into effect.
The bounding exercise evaluating the extent of biased migration required to explain the re-
duced form estimates. Consider the most extreme case where voters migrate from a state with a
low dropout age (baseline category of 15 or lower) to a state with a high dropout age (17 or higher).
In the NAES sample, the average proportion of Republican partisans in these states are, respec-
tively, 32.3% and 32.1%; the sample-weighted average population in high-dropout age states is 3.3
times larger. To account for the smallest reduced form estimate (a 0.03 percentage point increase
in Republican partisanship), if 48% of migrants from the low-dropout age state are Republican
partisans (which is 50% greater than the sample average), the net migration rate from low to high-
dropout age states would have to average almost 14% per cohort. For Republican vote intention
and reported vote choice, selective migration (50% above the sample average for each measure of
voting) would respectively require net per cohort migration rates of 15% and 12% to fully explain
the reduced form estimates. Beyond the fact migration is unlikely to be so politically biased, net
migration between such states is substantially lower than the rates required to account for the main
A19
results. In the ACS data, respondents are in fact slightly more likely to migrate to a state with a
lower dropout age than the state in which they were born.37
Furthermore, I show that the results are robust to using the ANES dataset, which uses a substan-
tially smaller and less recent sample. However, the ANES has several advantages over the NAES,
and thus showing similar results is an important robustness check: the ANES covers a much wider
span of elections, and thus includes more voters from earlier cohorts when individuals left school
earlier (average schooling is 10.8 grades); and by asking respondents which state they lived in
at age 14, cross-state migration concerns are minimized because dropout ages can be mapped to
individuals by location at age 14. The second feature is particularly relevant for addressing the
migration concern because it ensures that the state dropout age is correctly assigned at the start of
high school.
The TS2SLS estimates in panel A of Table A3 again show that schooling significantly increases
Republican partisanship and self-reported Republican voting in prior House and Senate elections.
State-specific cohort trends are excluded, given the comparatively weak power of the analysis and
a weak first stage. To further address the migration concern, the results in panel B show that the
ANES estimates are robust to focusing only on respondents who live in the state where they resided
at age 14. More generally, these findings also indicate that the political effects of increasing the
dropout age do not simply reflect idiosyncratic features of the NAES sample. The ANES data and
procedures are described in detail below.
A.5.2 First stage without sample matching restrictions
Table A4 show that the first stage estimates, when no restrictions are applied to match sample
moments, are very similar to those reported in Table 3.
37Using the ACS data, I coded as 1 respondents who live in a state with a high dropout age than theirstate of birth, and and a lower-dropout age state as -1. Summing across the sample (including non-movers),net migration was 4.5 percentage points toward lower-dropout age states.
A20
Table A3: The effect of schooling on identifying as and voting Republican—ANES dataset
Panel A: Full sample Partisan Vote House Senate(1) (2) (3) (4)
Schooling 0.103*** 0.050 0.163*** 0.071*(0.022) (0.035) (0.033) (0.038)
Reduced form observations 35,873 14,712 19,081 13,339First-stage observations 636,713 636,713 636,713 636,713First-stage F statistic 196.6 196.6 196.6 196.6
Panel B: Non-migrants Partisan Vote House Senate(5) (6) (7) (8)
Schooling 0.135*** 0.091** 0.160*** 0.095**(0.027) (0.042) (0.040) (0.043)
Reduced form observations 25,823 10,505 13,700 9,504First-stage observations 389,442 389,442 388,757 388,419First-stage F statistic 141.9 141.9 141.1 142.7
Notes: Outcomes are Republican partisanship (“partisan”), and voting Republican for President(“vote”), House and Senate. All specifications are estimated using TS2SLS, and include pre-treatmentscontrols (male, white, black and Asian dummies, and quartic demeaned age polynomials), state grew upin, survey decade, and cohort fixed effects. Differences in sample size reflect data availability. Standarderrors clustered by state-cohort in parentheses. * denotes p < 0.1, ** denotes p < 0.05, *** denotesp < 0.01.
A.5.3 Excluding state-specific cohort trends
Table A5 reports the main estimates where state-specific cohort trends—a key check for parallel
trends violations—are excluded. While the effect on partisanship increases somewhat, the effect
on vote choice decreases somewhat. However, all estimates remain statistically significant.
A.5.4 Exclusion restriction tests
Table A6 reports the exclusion restriction tests cited in the main article. The results indicate that
the dropout age does not meaningful effect fertility and martial choices—the age of a respondent’s
children or their probability of being married at the time of the survey. The age of children vari-
A21
Table A4: The effect of school dropout age on completed grades of high school, without samplematching restrictions
Completedgrades
(1)
Dropout age=16 0.183***(0.021)
Dropout age≥17 0.211***(0.023)
Observations 3,084,049Outcome mean 11.62First stage F statistic 43.2Test: Dropout age=16 = Dropout age≥17 (p value) 0.001
Notes: The specification include male, white, black and Asian dummies, quartic (demeaned) age poly-nomials, state-specific cohort trends, and state grew up, cohort and survey fixed effects, and is estimatedusing OLS. The omitted dropout age category is dropout age≤15. Standard errors clustered by state-cohort in parentheses. * p < 0.1, ** p < 0.05, *** p < 0.01.
ables, in particular, suggest that adolescent life choices were not substantially affected.
I also conduct an exclusion restriction sensitivity analysis in line with the Conley, Hansen
and Rossi (2012) union of confidence intervals approach. Specifically, I replace Yicst with Yicst −
δ11(dropout agecs = 16)− δ21(dropout agecs ≥ 17) in equation (2). I then estimate this using
TS2SLS, and examine the sensitivity of the results for different δ1 and δ2 parameter specifications.
I restrict attention to violations such that the relationship between δ1 and δ2 reflects the ratio be-
tween the reduced form coefficients estimated in columns (1)-(3) of Table 2. For example, for
Republican partisanship I consider δ1 = 0.013 and δ2 = (0.030/0.013)δ . The TS2SLS estimates
cease to be statistically significant at the 90% level when δ reaches 0.002, 0.017 and 0.010 respec-
tively for the partisanship, vote intention and vote choice outcomes. These violations respectively
represent around 16%, 48% and 38% of the full reduced form estimate, and thus suggest that a
very large violation is required to nullify the vote choice outcomes.
A22
Tabl
eA
5:T
heef
fect
ofsc
hool
drop
outa
gean
dco
mpl
eted
grad
eson
iden
tifyi
ngas
and
votin
gR
epub
lican
,exc
ludi
ngst
ate-
spec
ific
coho
rttr
ends
Rep
ublic
ansu
ppor
tD
emoc
rats
uppo
rtPa
rtis
anIn
tend
Vote
Part
isan
Inte
ndVo
te(1
)(2
)(3
)(4
)(5
)(6
)
Pane
lA:R
educ
edfo
rmD
ropo
utag
e=16
0.03
6***
0.02
9***
0.01
3-0
.028
***
-0.0
20**
0.00
2(0
.010
)(0
.011
)(0
.013
)(0
.010
)(0
.010
)(0
.012
)D
ropo
utag
e≥17
0.05
1***
0.03
5***
0.02
3*-0
.038
***
-0.0
25**
-0.0
13(0
.011
)(0
.012
)(0
.013
)(0
.011
)(0
.011
)(0
.013
)
Obs
erva
tions
176,
796
144,
742
125,
196
176,
796
144,
742
125,
196
Out
com
em
ean
0.30
0.46
0.45
0.33
0.42
0.36
Test
:Dro
pout
age=
16=
Dro
pout
age≥
17(p
valu
e)0.
001
0.30
50.
063
0.02
50.
353
0.00
5
Pane
lB:I
nstr
umen
talv
aria
bles
Com
plet
edgr
ades
0.10
8***
0.07
3***
0.05
2*-0
.081
***
-0.0
54**
-0.0
33(0
.026
)(0
.027
)(0
.029
)(0
.024
)(0
.025
)(0
.027
)
Firs
tsta
ge(A
CS)
obse
rvat
ions
443,
539
436,
333
443,
539
443,
539
436,
333
443,
539
Red
uced
form
(NA
ES)
obse
rvat
ions
176,
796
144,
699
125,
196
176,
796
144,
699
125,
196
Out
com
em
ean
0.30
0.46
0.45
0.33
0.42
0.36
Firs
tsta
geF
stat
istic
51.4
51.0
51.4
51.4
51.0
51.4
Not
es:
See
abov
efo
rde
taile
dva
riab
lede
finiti
ons.
All
spec
ifica
tions
incl
ude
mal
e,w
hite
,bla
ckan
dA
sian
dum
mie
s,qu
artic
(dem
eane
d)ag
epo
lyno
mia
ls,a
ndst
ate
grew
up,c
ohor
tand
surv
eyfix
edef
fect
s,an
dar
ees
timat
edus
ing
OL
S.D
iffer
ence
sin
the
num
bero
fobs
erva
tions
refle
ctdi
ffer
ence
sin
the
NA
ES
mod
ules
inw
hich
each
outc
ome
was
incl
uded
.T
heom
itted
drop
outa
geca
tego
ryis
drop
outa
ge≤
15.
Stan
dard
erro
rscl
uste
red
byst
ate-
coho
rtin
pare
nthe
ses.
*p<
0.1,
**p<
0.05
,***
p<
0.01
.
A23
Table A6: Exclusion restriction tests
Age of Age of Singleeldest youngestchild child(1) (2) (3)
Dropout age=16 0.100 0.349* 0.007*(0.182) (0.179) (0.004)
Dropout age≥17 -0.070 0.238 0.001(0.192) (0.186) (0.005)
Observations 168,525 168,525 443,539Outcome mean 16.18 13.44 0.17
Notes: See above for detailed variable definitions. All specifications include male, white, black andAsian dummies, quartic (demeaned) age polynomials, state-specific cohort trends, and state grew up,cohort and survey fixed effects, and are estimated using OLS. The omitted dropout age category isdropout age≤15. Standard errors clustered by state-cohort in parentheses. Standard errors clustered bystate-cohort in parentheses. * p < 0.1, ** p < 0.05, *** p < 0.01.
A.5.5 Upwardly biased estimates of completing high school
Table A7 provides the estimate when dropout ages are used to instrument for completing high
school. As noted in the main text, these estimates will be biased if the instruments induce additional
schooling beyond completing high school and such years of schooling affect the political outcomes
for such compliers.
Consistent with such an exclusion restriction violations inducing considerable bias, the esti-
mates in Table A7 for both the NAES and ANES samples are extremely large. Relative the effect
sizes for each additional year of schooling, the effect magnitudes increases substantially.38 Of the
specifications, four produce a coefficient that exceeds the theoretical maximum of one. In the other
three, the coefficient is again far larger than the comparable weighted per-year estimates obtained
in the main article. These estimates are thus consistent with a large bias due to dichotomizing an
endogenous variable with multiple intensities. Furthermore, since the first stage is relatively strong
38Equations including linear state-specific cohort trends are excluded because the first stage becomesvery weak.
A24
Table A7: IV estimates of the effect of completing high school on Republican partisanship andvoting
Panel A: NAES sample Partisan Intend Vote(1) (2) (3)
Complete high school 1.372*** 0.937*** 1.307(0.360) (0.338) (0.893)
Observations 176,796 144,742 106,879First stage F statistic 12.8 10.6 2.4
Panel B: ANES sample Partisan Vote House Senate(4) (5) (6) (7)
Complete high school 0.884*** 0.587** 1.518*** 0.847**(0.235) (0.244) (0.463) (0.402)
Observations 35,669 14,638 18,980 13,262First stage F statistic 11.5 10.8 6.9 5.2
Notes: See Table A7 for variable definitions. All specifications include gender, white, black and Asian dummies,quartic (demeaned) age polynomials, and state grew up, cohort and survey fixed effects, and are estimated using2SLS. The omitted dropout age category is dropout age≤15. First stage observations decline for vote intentionbecause the ACS sample is reduced to match the variables in the NAES sample. Standard errors clustered bystate-cohort in parentheses. * p < 0.1, ** p < 0.05, *** p < 0.01.
in most specifications, the inflated coefficients cannot simply be attributed to weak instrument bias.
Nevertheless, while the bias is large in this application, it is important emphasize that the extent of
bias is application-specific (see Marshall 2016).
A.5.6 Reduced form mechanism estimates
Table A8 shows the reduced form estimates of the dropout age on the possible mediators examined
in Table 7. The results reinforce the IV estimates.
A25
Tabl
eA
8:T
heef
fect
ofdr
opou
tage
son
pote
ntia
linc
ome
and
soci
aliz
atio
nm
echa
nism
s
Con
serv
ativ
eR
educ
eB
anL
owL
owH
igh
Rep
ublic
ansc
ale
taxe
s†ab
ortio
n†gu
nhe
alth
mili
tary
non-
econ
omic
cont
rols
†sp
endi
ng†
spen
ding
†is
sues
(1)
(2)
(3)
(4)
(5)
(6)
(7)
Dro
pout
age=
160.
053*
*0.
032*
*-0
.010
0.02
50.
015
-0.0
03-0
.032
(0.0
22)
(0.0
13)
(0.0
12)
(0.0
16)
(0.0
18)
(0.0
16)
(0.0
24)
Dro
pout
age≥
170.
063*
*0.
040*
**-0
.011
0.01
40.
018
-0.0
11-0
.034
(0.0
25)
(0.0
14)
(0.0
14)
(0.0
17)
(0.0
19)
(0.0
18)
(0.0
26)
Obs
erva
tions
172,
400
136,
830
131,
662
81,0
2269
,513
78,7
0316
3,31
4
Polit
ical
Polit
ical
Dis
cuss
Turn
out
Trus
tPr
otec
tB
ankn
owle
dge
inte
rest
polit
ics
(pre
s.fe
dera
len
viro
.ga
ysc
ale
scal
eel
ectio
n)go
vt.†
mor
e†m
arri
age†
(8)
(9)
(10)
(11)
(12)
(13)
(14)
Dro
pout
age=
16-0
.026
-0.0
10-0
.026
0.00
90.
008
-0.0
160.
016
(0.0
28)
(0.0
21)
(0.0
23)
(0.0
09)
(0.0
19)
(0.0
10)
(0.0
17)
Dro
pout
age≥
17-0
.026
-0.0
14-0
.032
0.01
10.
001
-0.0
17*
0.03
4*(0
.030
)(0
.024
)(0
.025
)(0
.011
)(0
.021
)(0
.010
)(0
.019
)
Obs
erva
tions
123,
411
177,
624
168,
961
150,
266
54,6
3112
6,52
272
,313
Not
es:A
llou
tcom
eva
riab
les
are
stan
dard
ized
,exc
eptb
inar
you
tcom
esde
note
dby
†;va
riab
lede
finiti
ons
are
prov
ided
abov
e.A
llsp
ecifi
catio
nsin
clud
em
ale,
whi
te,b
lack
and
Asi
andu
mm
ies,
quar
tic(d
emea
ned)
age
poly
nom
ials
,sta
te-s
peci
ficco
hort
tren
ds,a
ndst
ate
grew
up,c
ohor
tand
surv
eyfix
edef
fect
s.A
llsp
ecifi
catio
nsar
ees
timat
edus
ing
OL
S.D
iffer
ence
sin
sam
ple
size
refle
ctda
taav
aila
bilit
y.St
anda
rder
rors
clus
tere
dby
stat
e-co
hort
inpa
rent
hese
s.*
p<
0.1,
**p<
0.05
,***
p<
0.01
.
A26