LEARNING TO BE CONSERVATIVE: HOW BRITAIN AND THE US STAYING IN HIGH SCHOOL CHANGES POLITICAL PREFERENCES IN GREAT JOHN MARSHALL

Transcription

1 LEARNING TO BE CONSERVATIVE: HOW STAYING IN HIGH SCHOOL CHANGES POLITICAL PREFERENCES IN GREAT BRITAIN AND THE US JOHN MARSHALL THIS DRAFT: AUGUST 2013 FIRST DRAFT: DECEMBER 2012 Abstract Education may increase political participation, but does high school affect partisan identification and voting? Political economy theories imply that schooling makes citizens more conservative by increasing current or expected income, or by imparting general skills that reduce demand for social insurance. Sociological theories suggest schooling more immediately affects political preferences by increasing political engagement, post-materialism and network externalities. This paper utilizes compulsory schooling laws (CSLs) in the US and Great Britain to identify late high school s effect on political preferences. Raising CSLs by a year induces a five percentage point partisan swing toward the Republican or Conservative party. Instrumental variable estimates show an additional year of late high school increases the probability that CSL-compliers support right-wing parties by more than ten percentage points. Assessing the theoretical mechanisms, the income-based political economy channel receives clear support. These results pose serious strategic problems for left-wing parties traditionally supporting further public education. PhD candidate, Department of Government, Harvard University. jlmarsh@fas.harvard.edu. I wish to thank Paul Bolton, Philip Oreopoulos and Jim Snyder for kindly providing very useful data. I thank James Alt, Charlotte Cavaille, Anthony Fowler, Larry Katz, Jonathan Phillips, Stephen Fisher, Torben Iversen, Rakeen Mabud, James Robinson, Jim Snyder, Tom Woodin, and participants at the Harvard University comparative politics and graduate student political economy workshops for comments on earlier versions. 1

2 1 Introduction Given every individual goes through the high school system in industrialized countries, understanding its role in shaping political behavior is key. Surveys and election exit polls generally show more highly educated voters are more likely to turn out, but are also more conservative at least up until university. 1 While schooling s participation bias is relatively well-established, the causal link from schooling to partisanship and voting is not. This paper examines whether high school has such a partisan bias, before differentiating theoretical explanations by examining the mechanisms. Given its centrality in young people s lives, there are various avenues through which education could affect political preferences. This paper focuses on two prominent approaches. Political economists claim that richer individuals prefer lower tax rates and so vote for conservative parties (Romer 1975; Meltzer and Richard 1981). If human capital theory is correct and schooling is rewarded financially in the labor market, high school education makes voters more conservative once the income benefits are in sight. Others suggest the acquisition of general skills also reduces demand for social insurance (Iversen and Soskice 2001). Alternatively, political sociologists have argued that schooling changes an individual s political values (Inglehart 1981; Marshall 1950), and imparts skills that affect political engagement (Nie, Junn and Stehlik-Barry 1996; Rosenstone and Hansen 1993) and the constitution of their social network (Nie, Junn and Stehlik-Barry 1996). These hypotheses differ in their partisanship predictions, when their effects should occur, and their theoretical mechanisms. Researchers have long been interested in political socialization, including the impact of education on individual-level political behavior (e.g. Almond and Verba 1963; Brody 1978; Nie, Junn and Stehlik-Barry 1996; Rosenstone and Hansen 1993; Verba, Schlozman and Brady 1995), but until recently struggled to address issues of causality (Sondheimer and Green 2010). Berinsky and 1 E.g. this comment by Andrew Gelman. 2

3 Lenz (2011) note that political scientists are increasingly addressing the education as a cause versus education as a proxy debate (Kam and Palmer 2008). Recent contributions have used innovative research designs to seek causal estimates of different levels of education on voter participation (Berinsky and Lenz 2011; Dee 2004; Henderson and Chatfield 2011; Kam and Palmer 2008; Milligan, Moretti and Oreopoulos 2004; Sondheimer and Green 2010; Tenn 2007). This literature has generally found high school and university education increase participation. This research agenda has so far focused on explaining measures of citizenship, information and participation predominantly in the US, ignoring the important issue of education s partisan bias. The important question of vote choice has not yet been analyzed by researchers emphasizing causal identification. To estimate the partisan effects of an additional year of high school, this paper uses compulsory schools laws (CSLs) to generate variation in the numbers of years of schooling a student acquires in the US and Great Britain. Because CSLs do not induce university attendance, the results speak only to high school education. This policy natural experiment has been used widely among labor economists (e.g. Acemoglu and Angrist 2000; Angrist and Krueger 1991; Oreopoulos 2006), and applied to voter turnout (Milligan, Moretti and Oreopoulos 2004) and measures of citizenship (Dee 2004). CSLs make good instruments because they effectively induce students to stay in school, but are unlikely to be correlated with family or individual characteristics, or affect political behavior through channels other than continued schooling. Furthermore, the assignment of CSLs is not endogenous to government partisanship or partisan trends, and are unaffected by controlling for a variety of policy-level confounders. With US states continuing to change their laws and the UK raising the education leaving age to 17 in 2013 and 18 in 2015, CSLs are themselves both policy-relevant and politically salient. Estimates in the US and Britain show that CSLs have a large causal effect in their own right: increasing the minimum school leaving age by a year induces a five percentage point swing toward conservative parties per cohort. Instrumental variable estimates show that an additional year 3

4 of high school specifically in the final years of high school makes citizens that comply with CSLs upwards of ten percentage points more likely to identify as or vote for the Republicans or Conservatives. These results raise a catch 22 for left-of-center parties: while increasing access to education for the least advantaged is an important plank in the left-wing agenda, it comes at the cost of these future voters becoming more conservative. The large effects for CSL-compliers reflect late high school acting as a critical career juncture, and the greater benefits of schooling for the disadvantaged population of compliers. This interpretation fits with the empirical description of compliers. Furthermore, the analysis of the mechanisms underpinning schooling s effect shows that high school increases income, its effect is most pronounced at the mid-life earnings peak before tapering off into retirement, and high school reduces support for taxation and welfare. With the proposed mechanisms of sociological explanations receiving no support and high school not pushing students toward less skill-specific vocations, the results strongly support the Romer-Meltzer-Richard political economy explanation. This paper is organized as follows. Section 2 generates hypotheses linking schooling and partisanship. Section 3 describes the US and British data. Section 4 explains the research design. Section 5 provides the results and characterizes CSL-compliers. Section 6 considers the potential mechanisms through which high school makes individuals more conservative. Section 7 concludes. 2 Theory and hypotheses Although much has been written about how education increases civic participation, education s effect on partisan preferences has received limited attention. This section considers economic and sociological mechanisms through which high school could affect partisan preferences. It emphasizes different predictions and observable implications that rigorous empirical tests can differentiate. 4

5 2.1 Political economy arguments Much research suggests that an individual s education increases their wage income (e.g. Angrist and Krueger 1991; Oreopoulos 2006). The human capital interpretation suggests that education imparts productive skills, which are rewarded in competitive labor markets. Romer (1975) and Meltzer and Richard (1981) (RMR) argue that richer individuals will support less redistribution, which typically entails supporting right-wing political parties at least where taxation is a salient political issue. While the national-level implications of the RMR model receive mixed support (Karabarbounis 2011), surveys consistently show higher-income individuals are more conservative (e.g. Gelman 2009) while class voting remains prevalent (e.g. Thomassen 2005). Together, the human capital and RMR models imply that increased schooling should make voters more politically conservative. But does this link depend upon the remuneration benefits of schooling having been realized? If so, the effects of schooling should become most pronounced at an individual s earnings peak in their mid-40s (e.g. Heckman 1976), when the education premium is greatest. However, if current government policy affects future policy, voters expecting high future incomes could pre-emptively become more conservative. Alesina and La Ferrara (2005) find support for redistribution declines with one- and five-year expected future income. However, if the returns to schooling are uncertain, risk-averse voters should become more conservative once the benefits actualize. In retirement, the education ceases to provide monetary rewards, so schooling s effect should subside. Schooling could equally affect partisan preferences through demand for social insurance. Iversen and Soskice (2001) argue schooling provides students with general skills. Since general skills are rewarded in all industries, employees with such transferable skills expect shorter spells of unemployment and smaller wage reductions if forced to shift industries. As industry-specific skills are accumulated over a career, the importance of general skills declines. This theory is consistent with the human capital/rmr predictions, but distinctively implies schooling reduces demand for social 5

6 insurance against the possibility of future job loss, and most pertinently increases the relative contribution of general skills to wages. However, this phenomenon could still be explained by greater income facilitating the purchase of private insurance. Critics of human capital theory have argued that education merely serves as a costly signal of productive underlying characteristics without adding much productive value itself (Spence 1973), or a reflection of early life characteristics, values, cognitive abilities and experiences (Kam and Palmer 2008) or social class (Nie, Junn and Stehlik-Barry 1996). While both arguments imply that education is correlated with conservative voting, if selection problems can be resolved there should be no causal relationship or reason to expect any life-cycle differences. 2.2 Political sociology arguments Socialization theories offer a variety of plausible mechanisms pointing in different directions. First, the experience of education itself may directly affect political preferences. Continuing education could instill greater respect for government and the state (Marshall 1950), leaving voters favorable toward big government policies. Darden and Grzymala-Busse (2006) suggest curricula in pre- Communist nations transmitted nationalistic values that persisted across generations into post- Communist voting patterns. Less directly, schooling especially by late high school where citizenship and politics represent a greater focus in curricula may increase political information and interest, or at least provide skills reducing acquisition costs (Nie, Junn and Stehlik-Barry 1996; Rosenstone and Hansen 1993). Although greater knowledge and interest may not be intrinsically politically biased, they could create more balanced voting shortcuts (Lupia 1994) or encourage reading higher-quality newspapers (Heath et al. 1985). If the voting cues used by low-educated voters bias them toward left-wing parties, perhaps because union and social networks exert relatively larger pulls, further schooling could expose voters to right-wing arguments and immediately induce a rightward shift. In Britain, there remains debate about whether educational expansion has engendered partisan dealignment 6

7 (Albright 2009). However, examining year-olds, Tenn (2007) finds that while being a university student increases voter turnout and vote registration, years of education has no effect. Showing that the effects of education on participation quickly decay, this also casts doubt on the plausibility of immediate partisan effects. Second, Inglehart s (1981) scarcity hypothesis proposes that citizens become more socially liberal (in Mill s sense) and democratic, expressing post-materialist values, as education increases financial and cognitive resources. Although mapping these values to party choice is not obvious, US post-materialists identify as Democrats (Layman and Carmines 1997), while Labour and especially the Liberals in Britain have been associated with pro-environmental, anti-nuclear and political inclusion policies (Heath et al. 1985). Heath et al. (1985) suggest socially liberal values develop at universities rather than high schools in Britain, and translate into voting for the Liberal party, although such values are generally a weak determinant of voting relative to class. Inglehart s socialization hypothesis suggests schooling s effects may only apply once income has increased, but differs from political economy arguments in predicting increased post-materialist values and political engagement. However, looking cross-nationally, Weakliem (2002) suggests such effects should be ascribed to individualism. Finally, education opens doors to opportunities beyond the labor market. Better-educated voters have greater access to politically-engaged networks (Nie, Junn and Stehlik-Barry 1996), whose greater affluence, prestige or ideological diversity could expose them to more conservative perspectives. Such peer-group logic applies to voter turnout (Abrams, Iversen and Soskice 2010). However, extended high school could instead increase the probability of (left-leaning) union membership. 7

8 Table 1: Schooling s effect theoretical predictions Theory Causal direction Effect timing Mechanisms Human capital/rmr More conservative Maximized at earnings peaks, Higher income, lower preference for declines into retirement redistribution and social welfare Unproductive signal No effect No difference Correlation with income, correlation with being conservative Social insurance More conservative Rises in early life, declines in Relatively more general skill vocation, late career lower preference for social insurance School ideology Less conservative Immediate, age-invariant Greater trust in government Political engagement Probably more conservative Immediate, age-invariant Better politically informed, interested and active Post-materialism Uncertain Increasing with age Higher income, higher job prestige, post-materialist preferences, attend university Social networks Uncertain Probably increasing with Higher income, discuss politics age more, union membership 8

9 2.3 Moving forward Table 1 summarizes the central predictions of the different theories. Although the plurality hypothesize that high school makes voters more conservative, they differ in their timing and mechanisms. Given the difficulty of establishing causal relationships, we currently lack convincing evidence identifying either the direction of this relationship or illuminating which mechanisms apply. [Table 1 about here] Another estimation issue is that particular levels of education affect different types of individuals differently. Previous analyses have typically correlated educational qualification indicators or the number of years of schooling with partisanship, implicitly assuming constant effects across individuals. However, the marginal effect of schooling is likely to vary substantially across individuals. For example, where schooling has the biggest effect on income or alleviating scarcity most likely among low-ses citizens the marginal effect is probably largest. This paper addresses both problems. I first test whether additional years of late high school affect political preferences. Finding a strong conservative effect, I then probe the mechanisms by examining the observable implications of different theories, which are fully-consistent only with the human capital/rmr explanation. 3 Data In order to examine schooling and political behavior in the US and Great Britain I use data from the American National Election Survey (ANES), extracts from the US Census, and the British Election Survey (BES). This section describes the main variables. See Appendix for variable definitions, sources and summary statistics. 9

10 3.1 United States The ANES collects many political variables which can be pooled across presidential and mid-term elections and The sample size varies across surveys but typically surveys interview several thousand randomly-sampled US households, producing a maximum pooled sample of 35,873 respondents. 3 Since the number of years of schooling is not measured in the ANES, this paper complements the ANES with extracts from the decennial Censuses. The Appendix explains how Census extracts were reduced to a pooled sample of 972,860 individuals comparable to the ANES sample. The two datasets are combined using two-sample methods new to political science Dependent variables This paper examines four partisan outcomes. I first consider partisan self-identification: 35% of respondents identify as Republicans while 53% identify as Democrats; the residual are independent, while non-responses were excluded. This appears a little Democrat-skewed, although for much of the sample the Democrats controlled Congress, turnout is lower among low-income voters (Pontusson and Rueda 2010), and Democrats identifiers have been more likely to switch (Miller 1991). I define a Republican partisan indicator. Secondly, I focus at the cost of losing non-voters and observations from off-presidential and off-senate elections on self-reported vote choices. I code indicators for Vote Republican for President, Vote Republican for House and Vote Republican for Senate. The Republicans score higher by these metrics, receiving 48%, 43% and 45% of votes respectively. This likely reflects Republican turnout bias, perhaps correcting response biases in the partisanship measure. Using both partisanship and voting measures serves as a consistency check surveys did not ask where respondents grew up. 3 Observations suffering missingness were deleted. 10

11 3.1.2 Years of schooling The central independent variable is years of completed schooling. This is preferred to binary high school completion because it is more responsive to CSLs mandating minimum leaving ages; misspecified binary treatments can seriously accentuate estimates (Angrist and Imbens 1995). While the ANES uses a seven-category educational qualifications question, the Census permits calculation of completed years of schooling. Schooling is coded according to the number of grades completed; college and postgraduate education are top-coded as completing 12th grade as this study focuses on high school education. 4 In the sample, 66% of students attained at least 12th grade. [Figure 1 about here] Since the Census does not collect political variables, the ANES data provide a cursory graphical analysis. Figure 1 shows that there are large differences over 20 percentage points dividing some categories between educational groups in partisan identification and voting, with the better educated generally more Republican. 3.2 Great Britain In Britain, the BES randomly samples several thousand citizens from the British (excluding Northern Irish) electoral register after general elections for in-person interviews. 5 Although surveys have been run since 1964, only data from seven elections 1974(Feb.)-1997 contained comparable variables. 6 The maximum sample used for the analysis is 20, Results are robust to alternative definitions. 5 The 1997 exception randomly sampled households. Pre-election and non-interview surveys were excluded and do not contain relevant education questions; 1964 was unavailable via the UK Data Archive. 7 Observations suffering missingness were deleted. 11

12 Figure 1: Republican identification by education Grade school (12.5%) Some high school (14.5%) Finish high school (26.5%) Additional non-academic (8.3%) Some college (20.5%) Finish college (12.2%) Advanced degree (4.9%) % Republican Partisan Pres. vote House vote Senate vote 12

13 3.2.1 Dependent variables Like the US, I construct measures of partisan self-identification and self-reported voting. In the sample, 35% of respondents identify (following the initial prompt) as Conservative, while 36% identify as Labour, and 12% identify as Liberal; the residual are for small parties, non-partisans, or uncertain. I define three indicator variables: Conservative partisan, Labour partisan and Liberal partisan. General elections, which are not concurrent with local elections, in Britain entail casting a single ballot. With three large parties I code three vote indicator variables Conservative vote, Labour vote and Liberal vote for self-reported vote choice. Pooling across surveys, the Conservatives, Labour and the Liberals received 38%, 37% and 19% respectively Years of schooling In Britain state-funded schooling continues until 18, at which point students decide whether to pursue further education. Before 1998 university tuition was also free. The BES contains richer data on a student s education than the ANES, more recently asking what age a student left continuous full-time education and, up to 1983, what age a student left school. Unfortunately, coding is not consistent across surveys: in the 2000s, the lower bound was 15 and upper bound was 19, while preceding surveys had been unbounded. To ensure comparability across surveys and since the analysis will focus on inducing additional years of schooling for 14 and 15-year-olds, this paper focuses on high school education: Schooling combines the age left education and school variables, top-coding at >18 years of schooling. The elections were omitted because their bottom coding misses critical variation. [Figure 2 about here] Figure 2 shows the distribution of Conservative partisan identifiers and voters by age left schooling. Given we are pooling across elections with larger swings and using smaller education 13

14 Figure 2: Conservative partisans by schooling, <15 (28.7%) 15 (26.1%) 16 (22.5%) 17 (7.9%) 18 (7.6%) > 18 (7.2%) % Conservative Partisan Vote 14

15 intervals than the US, a slightly weaker positive relationship between schooling and partisanship is unsurprising. Note also that most of the data is clustered around leaving school aged 14-16, which will prove useful for estimation. 4 Research design 4.1 What we would like to estimate To identify the effect of schooling on partisanship and voting, the ideal linear probability model would be: Pr(Y it = 1) = βs i + X it γ + ε it, (1) where Y it is the binary observed outcome (conservative identifier/voter) at survey period t; S i {1,...,S} is the discrete treatment intensity of the number of completed years of school; X it is a vector of pre-treatment covariates; and ε it is the error term. If β is constant across individuals and interval intensities, OLS identifies the average treatment effect on the treated provided there is no interference across individuals and, conditional on X it, S i is randomly assigned and thus uncorrelated with potential outcomes. Estimating equation (1) in the hope of identifying a population causal effect of an additional year of schooling is problematic for three main reasons. First, which individuals receive longer schooling is very unlikely to be (conditionally) random (Kam and Palmer 2008). Since the assignment of S i is complicated, it is hard to be confident that including many covariates avoids selection bias (Henderson and Chatfield 2011). Second, as noted above, the effect of schooling likely differs across types of individual. For a variety of reasons, such as the policy of interest only affecting certain citizens, it is important to be clear which salient types are affected. Third, the causal response is likely to be non-linear since the effect of an additional year of schooling depends upon the year 15

16 in question: a year of early schooling or university is not equivalent to a year of high school Causal identification assumptions and interpretation This paper uses variation in compulsory schooling laws C i as a natural experiment to generate plausibly random variation in individual schooling S i. While exogeneity identifies the causal effect of CSLs themselves, an instrumental variable (IV) strategy estimates schooling s average causal response for individuals that comply with the laws. In a constant-effects IV framework, the causal effect β is identified by instrumenting for S i with excluded instruments C i Z it. To consistently estimate β the instruments Z it = (C i,x it ) must be conditionally randomly assigned with respect to S i and Y it, satisfy non-interference, affect S i, and satisfy the exclusion restriction that C i only affects political behavior Y it through schooling S i. Random assignment permits a causal interpretation of the effect of C i on S i and Y it, where the latter association is the reduced form. This requires that in the first-stage, S i = C i Π + X it Γ + η it = Z it δ + η it, (2) Π is non-zero, which is tested with an F test. However, differing effects across individuals are very likely. The heterogeneous potential outcomes framework (Angrist, Imbens and Rubin 1996; Imbens and Angrist 1994) reinterprets IV models to estimate the local average treatment effect (LATE) for compliers individuals that only received the treatment because of the instrument. Within this category, the LATE averages across complier types. Identification requires an additional monotonicity assumption, which here implies that there exist no defiers who exited school earlier after a CSL increase. Angrist and Imbens (1995) have shown that this approach generalizes to discrete treatment 8 Although students occasionally repeat or skip school years, the effects almost certainly still pertain to high school. 16

17 intensities. By weighting the causal effect at each intensity S i = s by the proportion of people affected by the instrument at that level of schooling Pr(S i = s), we can estimate the local average causal response (LACR) for compliers using 2SLS. 9 The weighting schemes underpinning estimates of LATE and LACR highlight the variation 2SLS coefficients are identified off: citizens with similar X it but who are differentially affected by the instrument. 10 LACR thus facilitates interpretation of regression results because it can explain why IV estimates differ from OLS: IV is estimated for populations with different causal effects and only averages across levels of schooling the instruments affect. 4.3 Variation in compulsory schooling Using CSLs to generate variation in incentives to attend school was pioneered by Angrist and Krueger (1991). They interacted CSLs with quarter of birth (QOB) indicators to instrument for years of schooling when estimating the wage returns to schooling. Although QOB is not available in the ANES or BES datasets, the QOB component has been criticized for its correlation with mother s SES (Buckles and Hungerman 2011). Non-interacted CSLs, however, have been widely used by labor economists to estimate the effects of schooling in the US (Acemoglu and Angrist 2000; Dee 2004; Goldin and Katz 2008; Lochner and Moretti 2004; Lleras-Muney 2002; Milligan, Moretti and Oreopoulos 2004) and UK (Devereux and Hart 2010; Harmon and Walker 1995; Milligan, Moretti and Oreopoulos 2004; Oreopoulos 2006). Their continued use lends CSLs considerable plausibility as effectively random sources of variation in individual schooling, although the political context presents new challenges to the identifying assumptions (see below). The validity of the IV assumptions is considered after the instruments are described. 9 2SLS estimation implicitly weights the LACR over covariates X it ; Abadie (2003) addresses weighting further. 10 This is perhaps easiest to understand by considering differential latent propensities to attend school affected by the instrument (Sondheimer and Green 2010). 17

18 4.3.1 US state CSLs A CSL defines the minimum legal age or level of education at which a student may drop out of school. CSL data are from Oreopoulos (2009), and based on the National Center for Education Statistic s Education Digest. Unlike Britain, CSLs are defined by age rather than academic cohort. 11 Oreopoulos (2009) lists penalties, exceptions and exemptions to state CSLs, but also notes that enforcement can be weak. US states have considerable control over education policy, including instituting CSLs. Massachusetts first implemented a CSL in 1852 and 41 states used them by 1910, principally to meet demand for educated workers and promote assimilation (Goldin and Katz 2008). Figure 3 plots changes in CSLs across 48 US states and Washington, DC since Although there has been a general upward trend in CSLs, there remain numerous instances of reversal as in Maine, Mississippi, and Oregon. This temporal variation is critical for identification when using cohort and state grew up in fixed-effects; identification comes from within-state-cohort CSL variation. Instances of reversal reduce concerns that that there is an underlying trend causing both CSLs and political behavior. [Figure 3 about here] As Figure 3 shows, the majority (74.6%) of state-year CSLs specify a leaving age of 16: 8.8% are below 16; 7.8% use 17; and 8.7% use 18. I create two indicators CSL=16 and CSL 17 to capture the effect of these leaving ages relative to a <16 baseline. Following Lochner and Moretti (2004), leaving ages of 17 and 18 are combined because leaving at 18 is a weak further constraint. 11 Child labor laws are often used alongside CSLs, and were important before 1940 (Goldin and Katz 2008). They are omitted here because they offer limited temporal variation, provide few compliers, and predominantly vary in rural states with lower compliance. Their inclusion leaves results unaffected. 12 Following Oreopoulos (2009), Alaska and Hawaii are omitted because of different demographic and economic characteristics. 18

19 Figure 3: Variation in US state CSLs, AL AR AZ CA CO CT DC DE FL GA IA ID IL IN Minimum school leaving age 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 Year 19

20 Given data is unavailable for specific date of birth, Year aged 14 is calculated based on birth year. 13 I map survey respondents to CSLs using the ANES question about state of residence at age 14 and state of birth in the Census. [Figure 4 about here] Figure 4 shows schooling steadily increased in the Census data until it plateaued around 1970 with most students completing high school. 14 These numbers reflect the fact that high school drops out have only dropped from 17% in 1970 to 14% since 1990 (Oreopoulos 2009) CSLs in Britain In Britain, the minimum school leaving age is controlled by the national government. There were three landmark pieces of legislation concerning CSLs in the twentieth century (see Gillard 2011; Woodin, McCulloch and Cowan 2012). Greater historical detail is provided in the Online Appendix. First, David Lloyd George s Liberal government raised the school leaving age from 13 to 14 under the Education Act 1918, coming into effect in Although the 1918 Act had intended for further increases in the leaving age, these did not transpire for financial reasons despite repeated attempts in the 1920s and 1930s (Oreopoulos 2006). Given the small number of people in the sample who turned 13 before 1922, and because the policy s effects were limited in practice, this paper examines only two most recent acts. 15 Second, Winston Churchill s wartime coalition government passed the Education Act 1944, 13 Since school and calendar years do not always coincide, the law is only affecting those born between January and the school entry cutoff which varies across state (e.g. September-December births affected by the instrument are counted in the prior year aged 14, which if anything should downwardly bias estimates.) 14 Averages slightly drop-off each decade as some people have not yet completed twelfth grade. 15 Only 5% of the sample were 13 before Using the 1922 reform as an additional instrument makes no difference. 20

21 Figure 4: US years of schooling by cohort, Average years of completed schooling Cohort: year aged 14 21

22 which increased the leaving age from 14 to 15 in England and Wales; the Education (Scotland) Act 1945 enacted the same reform in Scotland. The new leaving age came into force 1st April The Education Act also fully funded secondary education, making education more accessible. Figure 5 shows the 1947 reform dramatically reduced the proportion of pupils in each cohort leaving school at 14. There is also a slight jump in the proportion leaving at 15, suggesting a minor knock-on effect for leaving at 16. [Figure 5 about here] Conservative Prime Minister Harold Macmillan passed the Education Act 1962 raising the school leaving age to 16, although it was Conservative Edward Heath who finalized the update to the current system under Statutory Instrument 444 (1972). The new rule was implemented in England and Wales for the academic year starting 1st September Statutory Instrument 59 (1972) also raised the leaving age in Scotland, although teacher shortages meant not all local authorities fully implemented the reform until the Education Act Figure 5 shows these reforms again raised education participation rates, but less dramatically than in Although the reforms were implemented by Conservatives, like the 1947 reform Labour had consistently pushed for the increase, 16 while education was widely seen as an economically and socially beneficial investment in the golden years (Woodin, McCulloch and Cowan 2012). I code CSLs for the leaving age in England, Wales and Scotland: CSL=15 and CSL=16 are indicators for the minimum schooling leaving age (<14 is the residual category). In the BES, age is measured in years at the date of the survey a month or so after an election typically held between April and June and mapped to cohort CSLs. Although Scottish students faced a weaker law between 1972 and 1976, they are coded identically to England/Wales as a similarly large drop in the proportion leaving occurs Labour Prime Minister Gordon Brown passed the Education and Skills Act 2008, raising the education leaving to 18 by IV results are robust to excluding Scottish students aged Discontinuity analyses 22

23 Figure 5: British proportion leaving school age 14 and 15 by cohort, Proportion leaving at 14/ Cohort: year aged 14/15 Leaving at 14 Leaving at 15 23

24 4.3.3 Assessing IV assumptions As noted above, the IV identifying assumptions are that conditional upon X it, higher CSLs not only increase the likelihood that an individual remains in high school (and never decrease it) and affect political behavior only through the measured schooling channel. Despite compliance concerns, most studies find a strong first-stage for CSLs on years of schooling in the US and Britain. As statistical tests confirm below, and Figure 5 clearly shows in the Britain, CSLs had large effects on schooling behavior. Since CSLs affect all cohorts under local control, instrument assignment is unlikely to be correlated with individual or family-level characteristics. This obviates the need to control as extensively as Kam and Palmer (2008), except as robustness checks. Furthermore, given individuals cannot manipulate their date of birth and parents are extremely unlikely to have successfully predicted reforms more than a decade in advance, selection into particular CSL regimes is unlikely. Inter-state migration is a concern in the US if certain types of CSL-compliers move more. Annual cross-state migration has declined from 5% in the 1980s to around 3% (Molloy, Smith and Wozniak 2010). Among poor women of childbearing age the rate is lower and unrelated to destination welfare benefits (Hanson and Hartman 1994). Given CSL-compliers likely have disproportionately young, childless and less educated parents, early-life migration has limited bias potential (Molloy, Smith and Wozniak 2010). Furthermore, the principal recent reasons to move across states for college, marriage, family reasons, and natural disaster (Molloy, Smith and Wozniak 2010) are unlikely to be linked to CSL changes. A greater concern is adoption context: CSLs are not random events, but are determined by legislators. The most plausible violation of conditional independence arises if CSL changes are correlated with state-level political changes that also affect partisanship. Extensive analysis in the Online Appendix finds no statistical association between CSLs and indicators of Republicans exclude Scots. 24

25 control of upper, lower or both state houses, or Republican seat share. In Britain, CSL reforms were popular with all political parties and based on independent reports and post-war pressure for social reforms (Woodin, McCulloch and Cowan 2012), suggesting that they were not politicallyoriented decisions. Rather than immediately follow changes in government or political sentiment, the reforms were long-advocated but required considerable preparation and financing. However, we still cannot be certain that CSL reforms are uncorrelated with waves of economic, social or political change. This concern is addressed by controlling for contemporaneous labor market conditions and educational inputs, while by looking at cohorts around the reform discontinuity it is hard to see how adjacent cohorts are differentially affected by generational or cultural shifts. Furthermore, Lochner and Moretti (2004) show state high school graduation trends are uncorrelated with CSL changes in the US, while Lleras-Muney (2002) uses placebo tests to demonstrate CSLs cannot explain prior enrollment. Another concern is that CSLs affect subsequent experiences shaping individual political behavior. However, the exclusion restriction that CSLs only affect political behavior through continued schooling is likely to hold because schooling occurs early in life, CSLs are temporally proximate to the decision to remain in school for an additional year, and many subsequent decisions stem from school experiences. Although monotonicity is fundamentally untestable since we cannot observe all potential outcomes, it is hard to see why a higher leaving age would make an individual choose less schooling (e.g. Spence 1973). Furthermore, cumulative distribution plots by CSL category in the Online Appendix are consistent with monotonicity. 5 Effects of schooling on partisan preferences This section identifies large political effects of additional years of late high school. Inducing students to attend high school using CSLs considerably increases the probability that an individual 25

26 will identify with, or vote for, the Republicans or Conservatives later in life. 5.1 Partisanship and voting in the US To estimate the effect of an additional year of high school on Republican partisanship and voting, we wish to estimate the following model for US citizens who grew up in the US: Pr(Y igct = 1) = βs i +W it γ + α g + θ c + η t + ε igct, (3) where W it includes individual-specific pre-treatment characteristics (fourth-order demeaned Age polynomials, Male dummy and Race dummies); 18 α g, θ c, and η t are state grew up in (or born in for the Census), birth-year cohort and period fixed-effects; and ε igct is the error term. Under this model, β is identified off within-state, within-cohort and within-period variation in schooling (partialling out basic demographic features). In effect, we are comparing political behavior in a particular period for voters from the same cohort in a given state with different years of schooling. The extensive fixed-effect structure eliminates a variety of threats to internal validity: for example, cohort fixed-effects eliminate concerns that different cohorts were subject to different national schooling policies, while α g removes time-invariant differences in education across states and η t allows for common period-specific shocks that could capture candidates and political events moving all voters towards one party. The control set is sparse because education is an early life event, so many characteristics explaining political behavior are post-treatment. At the cost of sample size, robustness checks include family background controls. Even if OLS estimates of equation (3) were unbiased, years of completed schooling is not measured in the ANES, so equation (3) cannot be estimated using OLS. This paper thus focuses on reduced form and IV estimates of the causal effect of CSLs and years of schooling on partisanship 18 Because age, cohort and period effects cannot be simultaneously identified even in synthetic panels, the linear age term is dropped. 26

27 and voting. As an important education policy lever, the effects of CSLs are of interest in their own right. I estimate the following reduced form equation, Pr(Y igct = 1) = δ 1 1(CSL gc = 16) + δ 2 1(CSL gc 17) +W i γ + α s + θ c + η t + ε igct, (4) using OLS in the ANES data; η t are election fixed-effects. Under the parallel trends and crosssectional composition balance assumptions, equation (4) produces repeated cross-section differencein-difference (DD) estimates leveraging differential CSLs within and across states. Parallel trends assumes that in the absence of the treatment, treated and control units would experience parallel trends in Y igct (conditional on X it ). Random sampling across surveys and cohort fixed-effects mean panel cross-sections also similar to longitudinal data. Given CSLs can be integrated into both the ANES and Census datasets, two-sample IV (Angrist and Krueger 1992) can estimate the LACR of schooling by instrumenting for S i with the indicators 1(CSL gc = 16) and 1(CSL gc 17). I use the two-sample-2sls (TS2SLS) estimator, which computes the first-stage using the Census dataset and reduced form using the ANES dataset, efficiently combining the two as a consistent two-step estimator (Inoue and Solon 2010). Unlike the reduced form models, η t are Census decade indicators. Election fixed-effects cannot be included because the Census is not conducted concurrently with the ANES. Beyond the standard LATE assumptions discussed above, the critical assumption underpinning TS2SLS is that both datasets independently sample from the same population. 19 Although not perfectly satisfied because the Census occurs every five elections, both samples randomly draw from almost exactly the same voting age citizen population once ineligible voters from the ANES sample and those below 18 from the Census are removed. 20 Stratified re-sampling from Census extracts to produce the same cohort distribution considerably enhances comparability. The On- 19 This ensures sample moments can be substituted across datasets; see Inoue and Solon (2010). 20 To enhance comparability and include cohort fixed-effects I remove Census respondents aged 27

28 line Appendix provides further details about TS2SLS, and summary statistics indicating sample similarity. 21 In both reduced form and IV models, standard errors are clustered by state rather than the state-cohort level where the instrument varies to most conservatively address interference across individuals. Clustering is implemented in the TS2SLS models by adapting the results from Murphy and Topel (1985); see Online Appendix Results Table 2 examines how an additional year of schooling affects the propensity for an individual to identify as or vote Republican. Across all outcomes, additional years of late high school significantly increases the probability of supporting the Republican party. [Table 2 about here] Columns 1-4 report large reduced form effects. Compared to states requiring students to remain in school until at most age 15, keeping a student in school until 16 increased their probability of identifying as Republican by 5-6 percentage points and voting Republican by a similar magnitude. Keeping a student in school until at least 17 had little additional effect especially in the smaller-sample presidential and Senate elections, which typically feature greater personality than ideological voting, and Democrat-identifiers vote Republican (Miller 1991). These effects are interpreted as the causal effect of increasing the school leaving age. The effect of completing additional years of high school is perhaps more theoretically interesting. The first-stage coefficients for the excluded instruments in model (5) show that requiring students to remain in school until 16 increases the average number of years of schooling by 0.48 years, while keeping students until at least 17 adds an additional 0.05 years. The smaller reduced form effect at 17 likely reflects the fact that not much additional schooling is required to complete 14 before 1914 because no such respondents exist in the ANES. 21 Residual sample composition differences should be captured by W i. 28

29 Table 2: Schooling s effect on Republican support (1) (2) (3) (4) (5) (6) (7) (8) (9) Partisan Pres. House Senate Schooling Partisan Pres. House Senate OLS OLS OLS OLS OLS TS2SLS TS2SLS TS2SLS TS2SLS CSL= ** *** 0.057* 0.476*** (0.020) (0.027) (0.020) (0.025) (0.087) CSL * *** (0.028) (0.034) (0.029) (0.033) (0.109) Schooling 0.128** *** (0.042) (0.057) (0.044) (0.054) Pre-treatment controls Y Y Y Y Y Y Y Y Y State and cohort fixed-effects Y Y Y Y Y Y Y Y Y Period dummies Election Election Election Election Decade Decade Decade Decade Decade Reduced form observations 35,873 14,712 19,081 13,339 35,873 14,712 19,081 13,339 First-stage observations 972, , , , ,860 First-stage F statistic Notes: standard errors clustered by state in parentheses; controls are gender, white, black, Asian and Native American dummies, and quartic demeaned age polynomials; + denotes p < 0.1, * denotes p < 0.05, ** denotes p < 0.01, *** denotes p <

30 twelfth grade. These positive effects are broadly similar to previous estimates (Acemoglu and Angrist 2000; Dee 2004; Milligan, Moretti and Oreopoulos 2004), and entail an F statistic of Unsurprisingly given the reduced form estimates, the LACR estimates show additional years of high school have a large pro-republican effect. The model predicts a greater than ten percentage point increase in the probability of supporting the Republicans across all four partisanship outcomes for each additional year near the end of high school. The schooling effect, driven primarily by keeping students in school until 16, increases Republican partisanship by 13 percentage points. The standard errors are less precise for smaller-sample presidential and Senate voting. It is important to remember the effect of additional schooling is highly localized. A LACR interpretation implies that an additional year of high school causes a large average marginal effect for voters encouraged by CSLs to attend high school longer than they would have done otherwise. The first-stage keeps students in high school but does not affect college attendance. 22 Because schooling s effect pertains only to late high school, the results remain consistent with college education producing Democrat-leaning liberals and illustrate why averaging across all levels of education can be misleading. As the British analysis more clearly shows, only in late high school is schooling s political effect large. The relatively large effects likely reflect CSL-compliers standing to benefit most from the skills imparted (see below) and engaging in positive selection such that those induced by CSLs to remain in school have latent characteristics predisposing them toward the largest benefits from schooling Robustness The DD and TS2SLS estimates are lent further credibility by their robustness to including statespecific linear time trends. This checks the plausibility of the parallel trends assumption. While the DD estimates decline somewhat, the LACR estimates are highly robust in the large samples for 22 Unreported first-stages show both instruments increase the probability of remaining in school until 14, 15, 16 and 18, but like Acemoglu and Angrist (2000) not some college. 30

31 partisanship and House voting. All robustness checks are available in the replication materials or Online Appendix. Although state CSLs reforms are unlikely to be correlated with respondent-level characteristics, they could correlate with state average characteristic shifts. At the cost of sample size, adding six father occupation indicators and five indicators for head of respondent s childhood household s education level produced very similar results, although reduced form and LACR estimates for Senate voting became insignificant and the first-stage F statistics fell below 10. Attempting to control for labor market opportunities, the results were robust to including state personal income per capita. In further specification checks, similar results are found across mid-term and presidential election samples, when attention is restricted to post-civil Rights Act 1964 elections, or when the dependent variable is switched to Democrat support. Finally, the extensive margin of completing high school using only the ANES data reiterates the positive effect on the probability of supporting the Republican party found using TS2SLS. 23 Nevertheless, because the results may be surprising Great Britain provides a valuable robustness check exploiting different identifying assumptions. 5.2 Partisanship and voting in Great Britain Turning to Britain, an additional year of late high school education similarly increases Conservative partisan identification and voting. These estimates pertain to a slightly lower schooling age than the US where high school continues for two years longer. I first examine all cohorts before focusing on the cohorts around the reforms to address cohort bias. 23 However, as Angrist and Imbens (1995) note, misspecifying a treatment intensity as binary induces multiplicative bias φ 1. Here φ is large because the first-stage of CSLs on high school completion is small, but the reduced form captures effects for CSL-compliers who do not complete high school; see Angrist and Imbens (1995). 31