The Costs of Early School Leaving in Europe

The reduction of early school leaving to less than 10 percent of the relevant population by 2020 is a headline target in the Europe 2020 strategy and one of the five benchmarks of the strategic framework for European cooperation in education and training. Designing adequate policies to combat early school leaving is a difficult task that requires both the identification of causal links and the measurement of costs and benefits. In this paper, we review the issues surrounding the measurement of the costs of early school leaving to individuals and societies, and examine several implemented policies that are expected to affect early school leavers. These include broad policies – such as changes in minimum school leaving age, tracking and school resources – as well as more targeted policies. While our focus is mainly on Europe, we also consider important evidence from across the Atlantic.


Introduction
According to the definition used by Eurostat and the European Commission, early school leaving occurs when an individual aged 18 to 24 has attained at most lower secondary education and is not engaged in education and training. In 2009, 14.4% of the European population  was in this condition, with some countries close to the 10 percent mark (Austria, Germany, Belgium, The Netherlands, Sweden, Czechia, Poland and Slovenia) and other countries very far from this mark (Spain, Portugal, Italy, and -somewhat surprisingly -Norway and Iceland).
What are the individual and social costs of early school leaving? In a frictionless economy, where individuals are free to choose their optimal level of education and the effects of individual choice do not spill over into society at large, this question makes little sense: the level of education each of us chooses or has chosen is both privately and socially optimum. In practice, however, economies and societies are far from frictionless, and optimal choice may be restrained both by the presence of credit constraints -both short and long term -and by other market frictions. Moreover, since individual choice has effects that are typically ignored when choosing education, there are externalities driving a wedge between the social and private costs and returns to education.
Many view early school leaving as a serious economic and social phenomenon that has important consequences both on individuals and on society. The recent economic literature has highlighted that the benefits of education are broad (see Oreopoulos and Salvanes, 2011). For individuals, education generates benefits not only because it improves occupational prospects, wages 1 and job satisfaction, but also because it leads to more informed decisions affecting health, marriage, parenting and retirement. Moreover, schooling affects individual non-cognitive skills and attitudes, such as risk aversion, patience and motivation, that influence economic choices.
Individual decisions to undertake further education have social consequences, and affect both state finances -by raising tax revenues and reducing welfare benefit payments -and social welfare, because of their effects on crime, attitudes toward minorities and immigrants and political participation (Moretti, 2007).
The broad perception that early school living affects in a significant way both society and individuals, who could benefit from increased educational attainment, has induced policy makers to design policies that try to address the problem. Reducing early school leaving to less than 10 percent of the relevant population by 2020 is a headline target in the Europe 2020 strategy and one of the five benchmarks of the strategic framework for European cooperation in education and training (European Commission, 2010). Designing adequate policies to combat early school leaving is, however, a difficult task, that requires the identification of causal links and the evaluation of the expected costs and benefits. It is unfortunate that in the European policy debate these requirements are often overlooked or receive limited attention.
In this paper, we review and discuss the issues surrounding the measurement of the costs of early school leaving to individuals and societies. We also examine a set of implemented policies that are expected to affect the percentage of early school leavers, by distinguishing between broad policies -such as changes in minimum school leaving age, tracking and school resources -and policies explicitly targeted to the sub-population of early school leavers. While our focus is mainly on Europe, we also consider important evidence from across the Atlantic.

The Costs of Early School Leaving
The costs of early school leaving are private, fiscal and social (see Belfield, 2008, andPsacharopoulos, 2007). Table 1 from Belfield is a good summary of the expected benefits from additional education. Private benefits include the expected gains in earnings and wealth, improved health and life expectancy and higher lifetime satisfaction. From these benefits one should subtract the additional private costs, which include foregone earnings. Fiscal benefits include increased tax payments, lower reliance on government transfers and reduced expenditures on criminal justice. If the government subsidizes education, the public costs of additional schooling should be subtracted from expected benefits. Social benefits include productivity externalities, the social value of better health and the gains from reduced crime.
Evaluating the costs of early school leaving and the expected benefits of additional education -typically high school graduation -is a complex task, based on the comparison of outcomes for a treatment and a control group. It is customary to identify the treatment group with early school leavers and the control group with high school graduates. Ideally, individuals in these two groups should differ only in their educational attainment. In practice, however, they differ also in a number of observable and un-observable characteristics. While the analyst can control for the former using standard multivariate regression analysis, less can be done about the latter. To illustrate the implications of failure to control for un-observables, consider the possibility that high school graduates have better (unobserved) talent than early school leavers. If labour market outcomes depend both on formal education and on talent, assigning additional education to the treatment group will not produce for this group the labour market returns attained by high school graduates. If this is the case, the average observed difference in earnings between the 4 control and the treatment group over-estimates the expected private benefits of additional education.

Private costs of early school leaving: earnings and employment
Early school leaving typically reduces lifetime earnings and leads to higher unemployment where F is the annual expected net income flow, i the (real) rate of interest, s the age when school ends and labour market entry occurs and T is end of life. The expected income flow is defined as where K=HS,ESL, D is a dummy equal to 1 if the individual is active and to zero if she is retired, u is the unemployment rate, w the net wage, b net welfare benefits, p are net pension benefits and indices a and t are for age and time respectively, with a=s+t. We also assume that pensions are a function π of the average wage, and that social security contributions are used to fund pension outlays.
The expected cost of early school leaving -relative to attaining high school education -is given by The first component on the right hand side of [4] is the discounted sum of the (expected) difference in earnings (and pensions) over a common length of working and non-working life; the second component captures the foregone earnings associated to increasing schooling from ESL to HS; the third and final component captures the difference in life expectancy between ESL and HS graduates.
The computation of [4] for the average early school leaver requires information on lifetime earnings profiles, that typically last between 40 and 50 years, and on future pension benefits.
Longitudinal data covering these long life spans are rare outside of Scandinavia, and often available for specific birth cohorts. They can be used in [4] with the additional assumption that the earnings profiles do not vary by cohort. In the absence of long longitudinal data, the empirical literature has used cross sectional data. For this to make sense, the following assumption is required This assumption states that the expected income flow accruing to individuals aged a+x at time t+x is equal to the income flow accruing to individuals aged a+x at time t, augmented by average productivity growth  . To illustrate, the earnings flow of an individual aged 50 in year 2030typically not observable -is equal to the earnings flow of an individual aged 50 in 2013observable in a cross section of individuals of different age -inflated by estimated productivity growth.

6
Suppose that we plot age -earnings profiles by educational attainment using cross sectional data (see for instance Rouse, 2007). The assumption above says that -conditional on age -the ratio of income flows by education is constant over time In other words, the ratio of the expected income of high school graduates to the income of early school leavers at time t is equal to the ratio at time t +x.
While this assumption is necessary in practice to estimate lifetime earnings, it is important to be aware of its limits. Relative wages by education are the outcomes of the interaction between demand and supply. Consider first demand. There is a well -documented tendency of modern labour markets to become increasingly polarized, with an expansion of jobs in the upper and lower part of the wage distribution and a reduction of "middle class" routinized jobs (see Autor, Levy andMurnane, 2002, for the US andGoos, Manning andSalomons, 2010, for Europe).
Polarization implies that the relative demand for high school graduates, who have typically filled middle class jobs, is bound to decline relative to the expanding demand for professionals, who have a college degree or more, and for the less educated, who fill elementary occupations. Figures   1 and 2, taken from Goos, Manning and Salomons, illustrate these patterns for Europe. With increasing polarization and given relative supply, we expect that the income premium from completing high school -relative to early school leaving -might fall in the future. If this is the case, estimates that rely on current earnings profiles run the risk of over-estimating the costs of early school leaving.
Next consider relative supply. Policies that encourage individuals to complete high school, if successful, will increase the relative supply of high school graduates. Given demand, this increase will trigger a decline in the relative wage paid to high school graduates compared to early school leavers. If this is the case, using current wage ratios to estimate future ratios may lead again to over-estimating the costs of early school leaving. 2 A key ingredient of the first component of equation [4] is the difference of earnings by education. This difference should be measured for individuals who differ only in their educational attainment. In practice, most analysts use econometric methods to control for an array of observable characteristics that differ across education levels, such as parental background and labour market experience. Inevitably, important un-observables are omitted. In the presence of unobserved differences in ability, and of measurement error in observed education, standard estimates of the relationship between earnings and education fail to uncover causal effects.
A large literature has addressed this issue, in an attempt to identify sources of exogenous variation in educational attainment that could be used to pin down these effects. There is growing consensus in this literature that a genuine source of exogenous variation originates from changes in compulsory school reforms or minimum school leaving laws (see for instance Oreopoulos, 2006). 3 When these reforms are used as instruments for education, average estimated returns are typically much higher than those uncovered using standard ordinary least squares techniques. In the presence of heterogeneity in the returns to schooling, this result is often interpreted as showing that the individuals affected by these reforms (compliers) have higher returns to schooling than those who are not affected. 4 In the rest of this section, we review five studies that have tried to estimate the cost of early school leaving relative to high school or college graduation. For each study, we briefly discuss the key working assumptions and the main results.

The Rouse study
The study by Rouse, 2007, is part of the broader project aimed at evaluating the cost of early school leaving or of inadequate education in the United States, which was undertaken by a group of scholars coordinated by Henry Levin of the Teachers College, Columbia University. Rouse uses CPS (Current Population Survey) data for the years 2003 and 2004 to obtain average ageearnings profiles by education (high school or more, high school diploma or GED and high school dropouts). She disregards pension income and makes the following assumptions: a) the expected income flow accruing to individuals aged a+x at time t+x is equal to the income flow accruing to individuals aged a+x at time t augmented by average productivity growth  ; b) annual productivity growth lies between 0 and 2 percent; c) the discount rate is either 3.5 or 6 percent; d) the relevant time horizon starts at 18 and lasts about 45 years; e) no differences in pension income; 8 f) no differences in employment rates by education; g) ordinary least squared estimates of the returns to education, on the ground that omitted ability bias and measurement error bias more or less cancel out.
Rouse estimates that the per capita lifetime earnings difference between dropouts and those with a high school diploma ranges from 121,000 to 294,000 dollars (at 2004 prices), depending on the assumed rate of productivity growth and on the discount rate.
This interesting study has three drawbacks. First, assumption g) is not supported by recent research using instrumental variables techniques, which consistently finds that the estimated local average treatment effect is substantially higher than the ordinary least squares estimate. Second, by considering the age range between 18 and 67, Rouse downplays the foregone earnings of early school leavers, who may start working well before age 18. Third, by ignoring differences in employment and unemployment by education, Rouse under-estimates costs, because early school leavers are less likely than high school graduates to be continuously employed.

The Oreopoulos study
The study by Oreopoulos, 2006, shares most of the assumptions made by Rouse, with the important exception of assumption g). He estimates the causal effect of education on earnings separately for the United States, Canada and the United Kingdom, using the variation across cohorts and states in minimum school leaving age. He finds that the earnings gain induced by an additional year of schooling assigned to a dropout who left school at age 15 is 15.4 percent return in the United States, 12.9 percent in Canada and 13.7 percent in the United Kingdom. These returns are substantially higher than those estimated by standard ordinary least squares techniques. 5 Using these estimates, he derives age -earnings profiles from age 15 to age 64. He assumes a 3% discount rate and zero productivity growth and estimates that the average present value gain from staying an additional year in school is 103,593 dollars (2000 US dollars) in the United States, 82,572 dollars in Canada and 120,354 dollars in the United Kingdom. The estimated amount for the United States exceeds foregone earnings by a factor of 13.8.
Since computed gains from an additional year in high school exceed foregone earnings many times over, Oreopoulos asks why individuals drop out from school and offers several alternative explanations: a) dropouts strongly dislike schools; b) they and their families are financially constrained; c) students would like to stay on but parents are not willing to bear the costs, given the uncertainty about future returns; d) investing in further schooling is too risky; e) students do not act rationally and / or fail to correctly predict future higher earnings.

The study of school failure in Estonia
Anspal and co-authors, 2011, have carried out a very detailed study of the cost of school failure in Estonia. This study shares several of the assumptions made by the Rouse study, and stands out mainly because it removes assumption f), which disregards unemployment risks. These risks vary both with age and with educational attainment, and are much higher for those who have not completed upper secondary education. Table 2 presents the main results of this study, separately by gender and by educational level (tertiary relative to upper secondary, upper secondary relative to lower secondary). When only those continuously employed are considered, the expected gain from having an upper secondary degree is estimated at 4 percent for males and at 10 percent for females. However, when the wage gains are corrected for the probability of employment, the expected gain associated to having an upper secondary education raises to 35 percent for males and 46 percent for females, a substantial increase. In terms of lifetime earnings, this corresponds to 52,000 euro (in 2009) for males and to 41,000 euro for females.
The Estonian study clarifies that overlooking the substantial difference in unemployment risks experienced by individuals with different education can lead to grossly under-estimating the monetary costs of early school leaving.

Conditions (EFILWC)
This study examines the economic costs borne by individuals who are not in employment, education or training (NEET), using a European -wide dataset, the European Survey on Income and Living Conditions (EUSILC). This dataset has several advantages, and should be considered as a key resource in any investigation of the costs of early school leaving in Europe. First, it is based on a common questionnaire and harmonized across more than 20 European states; second, it includes crucial information for the estimation of the costs of early school leaving: gross and net earnings, welfare payments -including unemployment benefits -and labour market status. Third, it covers the period from 2005 to 2011.
Compared to the studies illustrated so far, this study differs also in the empirical methodology, because it replaces the more standard regression analysis with a matching approach. This is a statistical tool designed to match "treated" individuals (early school leavers) and "control" individuals (high school graduates) using a large set of observables. The underlying idea is to find for each early school leaver a high school graduate with the same set of observable characteristics. Unfortunately, since matching cannot be done on un-observable characteristics, this method does not allow us to identify a causal relationship between education and earnings.

Heterogeneous returns to compulsory education
Most estimates of the cost of early school leaving focus on the cost born by the average school leaver, thereby ignoring that early school leavers are heterogeneous, for instance because they differ in their early life conditions. Brunello, Weber and Weiss, 2012, use comparable European data from the survey SHARE (Survey on the Health, Age and Retirement in Europe) and retrospective information to estimate the lifetime earnings of Europeans born between 1920 and 1956 in 9 European countries. Compared to the Rouse study, this investigation allows for unemployment spells, includes estimated pension income in some specifications and identifies the causal effects of education on earnings using the cross country and cross cohort variation in compulsory school years.
Brunello, Weber and Weiss find that an additional year of schooling increases the lifetime earnings of compliers -who typically have less than upper secondary education -by 9 percent.
They then estimate the returns to education separately for individuals who have had no books in the house at age 10 and for individuals with had books and find that an additional year of schooling increases earnings by 5.2 percent in the former group and by a very high 21.1 percent in the latter group. The availability of books in the house at age ten captures important features of the early environment, which contribute to shaping individual ability.
These estimates suggest that the costs of early school leaving in terms of lifetime earnings vary substantially among school leavers, and could be particularly severe for those with relatively high individual and parental ability and high marginal costs of education. They also point out that policies that increase minimum school leaving age pay off rather handsomely for individuals with good environmental conditions (books in the house at age 10) but are unlikely to deliver high returns for those forced into higher education from a less privileged background.

Fiscal Costs of Early School Leaving: Tax Revenues and Welfare Benefits
In his study of the costs and benefits of early school leaving, Psacharopoulous, 2007, classifies foregone tax revenues as fiscal costs. Since early school leavers earn on average lower lifetime earnings, they also contribute less to society in term of income taxes. When data on gross and net earnings are available -as in the EUSILC dataset -foregone tax revenues can be computed as the difference between lifetime gross and net earnings.
Since early school leavers are more likely to be unemployed or out of the labour force, they typically receive higher transfers from the government, in the form of unemployment benefits, disability benefits, and other welfare payments. In its study of NEET, the European Foundation for the Improvement of Living and Working Conditions (EFILWC), 2012, considers these transfers as public finance costs, defined as the benefits received by NEETs from public schemes in excess of those received by individuals with at least upper secondary education. These costs include: unemployment benefits, survivors benefits, sickness benefits, disability benefits and education-related allowances. When pension schemes are not fully funded by beneficiaries, one should also include pension benefits funded by public resources.
Using the matching approach discussed in sub-section 1.1.4, the EFILWC report estimates the average annual public finance costs of NEET in the 26 European countries at 742 euro per head (2008 prices), with a minimum of 3 euro in Bulgaria and a maximum of 5204 euro in Denmark. 6 Table 1 shows that the effects of additional education on health include private benefits (improved health status and better life expectancy), fiscal benefits (lower reliance on government health problems) and social benefits (higher social value of health). There are many possible channels through which education may improve health, which include: stress reduction, better decision making and information gathering, higher likelihood of having health insurance, healthier employment, better neighbourhoods and peers, and health behaviours, such as smoking, drinking, eating calorie-intensive food and refraining from exercising (see Lochner, 2011).

Costs of Early School Leaving: Health
The relationship between education and health -the "health-education gradient" -is widely studied. While there is abundant evidence that a gradient exists (Cutler and Lleras Muney, 2010), more controversy exists on whether the gradient reflects a causal effect. The empirical literature that has addressed this issue typically focuses on single countries and identifies the effect of education on health using the exogenous variation generated by mandatory schooling laws. Most of these studies consider self-reported health and mortality as the key outcome. Some find that education improves health and reduces mortality -see for instance Mazumder, 2008  across-cohorts variations in compulsory education to estimate the causal effect of education on self-reported poor health. Their OLS estimates of the gradient are -2.5 percentage points for females and -1.7 percentage points for males. The magnitude of the gradient increases sensibly when they use the exogenous variation provided by compulsory schooling. They find that one additional year of schooling decreases the probability of poor health by 4 to 8.5 percentage points for females and by 5 to 6.4 percentage points for males. This study, and several others, suggests that estimating the health costs of low education using non causal estimates such as ordinary least squares is likely to grossly under-estimate these costs.
Some of the health benefits associated to higher education accrue to the individual and some accrue to society. In modern society, several health costs are borne by governments, which are often involved in the supply of health services. By improving health, better and more education has the potential of reducing public health costs. As argued by Muennig, 2007, however, two countervailing effects may limit the savings attributable to higher education. First, better educated individuals use public health services more intensively, for instance because they screen more frequently. Second, they live longer, and therefore are in need of more services later in life. The presence of these countervailing factors imply that it is difficult to pin down the public health savings associated to higher education.
Measuring the health costs of early school leaving typically involves two steps. In the first step, the analyst estimates the effect of additional education on individual health capital, defined as the utility from the stock of current and future quality-adjusted life years. In the second step, the savings in (mainly public) health expenditures are measured. In this section, we review two studies that have addressed these issues.

The study of school failure in Estonia
This study defines health capital HC as where V is the monetary value of a year of life, H is the health flow, i is the rate of interest, a is age, K is education and E is the expectations operator. In its simplest version, the health flow is equal to 1 in the event of life and to 0 in the event of death. Conditional on being alive at age a, 14 the expected health flow is the probability of being alive at age a+k, k a P  , and Eq. (7) can be rewritten as Each year of life is characterized by different health, with the quality of health declining with age. Rather than simply adding up -with a discount -future probabilities of survival, a more encompassing measure of health should weight each year with its quality. Define this weight as QALY (quality-adjusted life year), an index that ranges from 0 (no quality) to 1 (perfect health).
One QALY is a year of life lived in perfect health. Using this weight in (8) The effect of graduating from high school on health capital is  In their study on Estonia, Anspal and co-authors, 2012, assume that QALY does not vary with age, and estimates education specific QALY using the approach suggested by Groot and van den Brink, 2007. In this approach, the authors employ as a measure of quality self-reported health 7 , which takes the following values: very good, good, fair, poor, very poor health, regress it on education and additional covariates, which include parental background, and use a normalization of the coefficient associated to education to obtain education specific QALY. As discussed above, their estimates are likely to be lower bounds, because they do not use exogenous variation to identify the effect of education on health. Anspal et al, 2011, estimate that males failing to attain upper secondary education have a 4.3% lower QALY than males who attain higher education. They combine the information on QALY with the effects of education on life expectancy -which they measure at age 25 -use a 3 percent discount rate and estimate that the difference in discounted quality -adjusted life years between individuals with upper secondary and lower secondary education is close to approximately 3 years for males and to 1 year for females. Using as value of a life year approximately 40,770 euro in 2010 prices, they estimate the value of health capital by gender and education for Estonia. Results are reported in Table 3. If we compare the estimates in Tables 2 and 3, we see that the estimated costs of early school leaving in terms of lower health capital are higher than the costs in terms of lower lifetime earnings. We also notice that the estimates in Table 3 do not include the higher public health expenditures associated to a lower health capital.

The study by Muennig
The study on Estonia produces estimates of health capital by education but is silent about the effects of education on public health expenditures. An antecedent study by Muennig, 2007, uses US data to estimate both the private and the public health costs associated to less than upper secondary education. As in the Estonian study, the estimate of the private costs relies on two basic ingredients, the effects of education on life expectancy (the number of years an individual is expected to live at a given age) and the effects on quality -adjusted life years, or QALY.
Muennig estimates that attaining a high school degree increases perfect health over the lifetime by 1.7 years. Valuing one additional year of perfect health at 110,000 dollars (at constant 2005 prices), he estimates the benefits of high school graduation at approximately 183,000 dollars.
Since high school graduates in the US are less likely to be enrolled in public health programs such as Medicare and Medicaid, a reduction in the share of early school leavers translates into large differences in annual public costs. While the average school dropout consumes 2,700 dollars (at 2005 prices) in public health insurance costs per year, the average high school graduate consumes 1000 dollars. Muenning (2007)

Fiscal and Social Costs of Early School Leaving: Crime
Belfield, 2008, classifies the effects of education on crime into fiscal benefits -reduced expenditures on criminal justice -and social benefits -the gains to society from reduced crime.
There are several reasons why education may have an effect on crime. Machin, Marie and Vucic, 2011, consider the following three channels: income effects, time availability, and patience or risk aversion. On the one hand, education increases the returns to legitimate work, raising the opportunity costs of illegal behaviour. On the other hand, punishment for criminal offences may lead to imprisonment. By raising wage rates, schooling makes any time spent out of the labour market more costly. These arguments suggest that those who can earn more are less likely to engage in crime. But it is also possible that schooling raises the returns to crime. For example, certain white collar crimes require higher levels of education (see Moretti, 2007).
Time spent in education reduces the time available for participating in criminal activity. Whilst youngsters are at school, they stay off the streets. Education may also influence crime by affecting patience and risk aversion (see Lochner and Moretti, 2004). More patient or more risk averse individuals are likely to place a greater weight on the possibility of future punishment and future income losses from imprisonment. In a similar vein, additional schooling may affect the taste for crime by increasing the psychological costs of breaking the law.
As in the case of earnings and health, the fact that education and crime are negatively correlated does not automatically guarantee that education has a causal effect on crime. In order to identify a causal relationship, Lochner and Moretti, 2004, for the US and Machin, Marie and Vucic, 2011, for the UK, use the exogenous variation provided by changes in compulsory school laws, which affect different cohorts of individuals going through their schooling experience.
Lochner and Moretti found that US states that raised high school graduation rates by increasing minimum school leaving age experienced significant declines in incarceration rates. Machin and co-authors found that the exogenous increase in education induced by the 1947 school reform in the UK produced a significant reduction in property crimes.
Using the estimates by Lochner andMoretti, 2004, Moretti, et al. 2007, computes the expected social benefits from reduced crime associated to a one percent increase in high school graduation rates. He multiplies the estimated total cost per crime -which consists of incarceration costs, property losses and victim costs -by the estimated reduction in the arrest and crime rate associated to higher education. Admittedly, this estimate is a lower bound to the cost of crime, because it ignores important cost components, which include law enforcement and judicial costs, private security measures and the steps people take to avoid being victimized. Even so, expected savings are estimated at 1.6 billion dollars, a substantial amount. The positive externality in crime reduction induced by an extra male high school graduate is between 14 and 26 percent of the private return to high school graduation.
Machin and co-authors, 2011, estimate the social benefits from crime reduction that would follow from a 1 percent reduction in the percentage of individuals with no educational qualifications. They find that the estimated crime reduction ranges between 43,921 and 88,469 offences, and that the associated costs range from 54 to 109 million pounds. Assuming that an extra year of school at age 16 is sufficient to obtaining an educational qualification, they also estimate that the cost of achieving a one percent reduction in the pool of individuals without qualification is 22 million euro, way below the expected social benefits from crime reduction.

Evaluating the general equilibrium effects of changes in the share of early school leavers
Policies that aim at reducing the share of early school leavers need to compare expected benefits and costs. Since relative earnings are affected by changes in relative supply and demand, a potential difficulty with the estimate of expected benefits is that these benefits may vary when the percentage of early school leavers is altered. Policies can affect not only wages by education but also the shape of age -earnings profiles, a crucial ingredient in the computation of expected lifetime gains.
The availability of estimates that link changes in relative supply with changes in relative wages may allow the analyst to correct estimated benefits by adjusting wages by age and education. To illustrate, consider a setup where employers hire four different types of labour, the young with less than secondary education, and suggests that a reduction in the relative share of early school leavers increase their relative wage. Similarly, we have that A reduction in the share of early school leavers over youth employment reduces their wage relative to that of older early school leavers.
Following Katz and Murphy, 1992, equations [14] and [15] can be estimated using repeated cross sections from the EUSILC dataset. These estimates can then be used to adjust wages by age and education and the estimate of the monetary costs of early school leaving.

Estimating the Cost of Early School Leaving in Europe. Guidelines for Empirical Studies
The review of the existing empirical literature on the costs of early school leaving suggests the following guidelines for an empirical study: 1. identify the causal effect of education on earnings and health, using credible exogenous variation. A source of exogenous variation that has become increasingly popular in the 4. evaluate general equilibrium effects. If the proposed change in the share of early school leavers is small, because of proximity to the European benchmark of 10 percent, these effects can probably be disregarded. If instead the change is large, as one would foresee for countries such as Portugal, Spain, Italy or Romania, adjust age earnings profiles to take into account the effects of changes of quantities on relative prices; 5. include the risk of unemployment as a key ingredient of the cost of early school leaving.
Differences in the probability of employment may be more serious than differences in earnings.

The effectiveness of policies combating early school leaving
Early school leaving is the result of multiple factors, such as poor early school experience, lack of ability, low perceived labour market returns to education, heavy discounting of future benefits, poor social environment and peer influence. Consequently, interventions aimed at reducing this phenomenon have to address a number of different aspects. Since handling all these factors together is particularly difficult, many existing policies tend to focus only on some.
Several policies have been pursued by different countries to reduce early school leaving (see European Commission, 2010), and the debate on which policies are more effective is still very animate among both academics and politicians. While some policies are aimed at students at-risk or are targeted to disadvantaged pupils, other policies affect the entire school system with actions involving all pupils in school. 8 Unfortunately, policies in this area are rarely evaluated using approaches that identify causal effects, partly because the design and implementation of many programs disregard that a source of exogenous variation in the treatment should be available.
As reported by Meierkord and Mascherini, 2012, many policies implemented in EU countries do not even define outcome targets. Typically, the outcome target refers to the percentage of participants who are expected to achieve a qualification. Very rare are also evaluations that quantify the benefits obtained by individuals who have achieved the target qualification, relative to those accruing when the target is not reached. On top of this, we often do not know whether these benefits outweigh the program costs. Completely absent are studies that include in the costbenefit analysis the gains in terms of fiscal benefits and externalities. Unfortunately, the virtual absence of careful studies providing guidance to policy-makers can lead to an inefficient allocation of (scarce) public resources.
In this section, we review a number of evaluation studies that rely on credible identification strategies. While some of these studies only consider the impact of the policy under study on selected outcome variables, there are also studies that offer a cost-benefit analysis. We consider both interventions affecting the entire school system and interventions targeted at disadvantaged groups.

Interventions affecting the entire school system
In this section we review: a) policies that increase the number of compulsory years of schooling; b) policies that change the time of tracking; c) policies that affect the organization of teaching activities, both by changing the number of pupils per class and by changing curricula.

Changes in the length of compulsory education
Compulsory school laws establish either a minimum number of years of education that an individual should complete before leaving school or a minimum school leaving age. These laws are widespread in Europe (see Murtin and Viarengo, 2011). In several European countries, compulsory education has repeatedly increased after the end of the Second World War, from primary and lower secondary education to include at least part if not the entire cycle of upper secondary education.
Higher compulsory education addresses the dropout problem because it forces potential school leavers to stay in school longer and eventually complete upper secondary education (and even college). Several studies have estimated the effect of adding one year of compulsory education on average educational attainment for European countries. Brunello, Fort and Weber, 2009, review some of these studies for Continental (Western) Europe and show that the estimated average effect is close to 0.3 years of schooling. For the UK, Oreopoulos, 2006, estimates that the compulsory increase in the school leaving age -from 14 to 15 -in 1947 increased the average school leaving age by more than half a year.
In a recent empirical study, Cabus and De Witte, 2011, estimate the causal impact of one additional year of compulsory education on the dropout rate, using administrative data for students living in Amsterdam (for whom the information on the date of birth is available). The authors identify a treatment and a control group, depending on year of birth. While before the reform pupils were mandated to attend school until their 17 th birthday, after the reform mandatory school attendance increased to age 18 (or until a higher secondary diploma was obtained if that comes at an early age). They estimate the effect of a one-year increase in compulsory education by comparing the average dropout rate in the treatment and control groups, after conditioning for age and time effects, and find that the policy has reduced the rate in the former group relative to the latter by 2.28 to 2.52 percentage points, with the largest effect for vocational students (−3.32 percentage points) and Dutch natives (−6.12 percentage points), and no effect for students who have experienced grade retention. The estimated effect, however, is driven mainly by an increase in the dropout rate of students in the control group, probably due to the anticipated improvement in the Dutch labour market when the policy was implemented.
As argued by Oreopoulos, 2006, compulsory education can be either not binding -for those individuals who are already investing in higher education -or binding for those who would have attained less education in the absence of the law. If we believe in the investment model of education, which predicts that each individual chooses her optimal level of investment by equating perceived marginal benefits to marginal costs, and if we are prepared to assume that there are no obstacles to individual choice, then compulsory school reforms cannot be optimal, because they force some individuals to invest more than their private optimum, at a level where marginal costs are above marginal benefits. While changes in minimum school leaving age are often justified by the presence of positive externalities -for instance because education reduces crime -or because individuals are constrained by lack of financial resources or finally because they are not able to predict in a satisfactory way the expected returns to further investment, it is not clear that these changes are the most efficient way to address the dropout problem.
Importantly, the effects of compulsory school reforms on the sub-sample of compliers are far from homogeneous. As discussed above, Brunello, Weber and Weiss, 2012, have shown that the expected payoff in terms of higher lifetime earnings of adding one year of compulsory education varies substantially with the conditions that individuals face early in life. Their findings suggest that among the individuals induced by these reforms to attain higher education, only those with 22 enough books at home have been able to reap significant private economic returns. A significant group of individuals, who lived in rural areas with very few books at home, attained higher education but earned much lower private returns.

School tracking
An important issue in the organization of secondary educational systems is whether and at what age students should be allocated to academic and vocational tracks. As reported by Brunello and Checchi, 2007, countries differ in the age at which students are tracked into different types (or levels) of education. While in the US and Scandinavia secondary schools are broadly comprehensive, in Austria and Germany pupils are tracked into different types of schools as early as at age 10. Many European countries fall between these two extremes. There is also variety across countries in the number of available tracks.
Early tracking may affect early school leaving for two reasons. On the one hand, by affecting test scores they may induce some students to drop out from school too soon. On the other hand, by providing a more flexible set of curricula and a vocational track that places less emphasis on academic abilities, early tracking may be able to keep longer in schools those students who prefer practical trades to more academic education.
Some studies have tried to evaluate the effects of tracking on educational outcomes.
Hanushek and Woessman, 2006, for instance, use a differences-in-differences approach to compare the mean and the dispersion of test scores before and after tracking in two samples of countries: a treatment sample with early tracking (before age 15) and a control sample where tracking occurs after age 15 if ever. They find that early tracking increases inequality in test score performance, but has limited effects on average achievement. However, as shown by Jakubowski, 2007, their results might be affected by the fact that the international tests they consider are taken at different ages. 9 Pekkarinen et al, 2009, examine the gradual introduction of a comprehensive school system in Finland and find that the reform has a small positive effect on verbal test scores but no effect on mean performance in arithmetic or logical reasoning tests. In addition, the reform improved the test scores of students with poorly educated parents. This evidence is in contrast with that produced by Figlio and Page, 2002, who find that tracking in the US does not harm low-ability students.
In a recent study, Hall, 2012, evaluates the effects of introducing a more comprehensive school system on dropout rates in Sweden. She considers a pilot scheme (implemented only in some municipalities) that preceded an educational reform aimed at increasing the academic content and the duration of vocational curricula, thereby reducing the differences with the more academic tracks. In Sweden, the municipalities involved in the pilot scheme were decided by a central authority on the basis of observable characteristics. Hall identifies the effects of the policy on the treated municipalities by exploiting the exogenous variation provided by a set of municipality characteristics. Her estimates show that the probability that upper secondary school students drop out increased by 3.8 percentage points with the piloted policy, mainly as a result of the behaviour of low performing students. Since the piloted policy also increased workplace training, these negative effects may have been partly compensated by higher employment rates among students leaving school early.

Changing educational inputs: class size
The effects on student performance of the resources devoted to education are a central concern in the economic and political debate on educational policies. Starting with the controversial Coleman Report, a study commissioned by the United States Department of Health, Education, and Welfare in the mid-sixties to assess the availability of equal educational opportunities, many empirical studies have investigated the effects of class size, teaching experience and ability, peer characteristics and school resources on student achievement. However, the results obtained are far from conclusive.
The main findings of the Coleman Report, showing that inputs into schooling have negligible effects on student achievement, were confirmed by Hanushek's, 1986, influential survey of the economic research on the subject. A key ingredient of the educational production function is class size. Hanushek concludes that there does not appear to be a strong, systematic relationship between school resources, and in particular class size, and student achievement. Krueger, 2003, on the other hand, documents a positive association between resources, class size and achievement.
In traditional studies, the estimates of class size effects are plagued by serious identification problems, which produce biased estimates. When trying to assess the effect of an intervention such as a reduction in class size on student achievement, researchers are typically not able to observe all the other confounding factors that enter into the student learning process, including student effort, family support or teaching quality. The problem is that these omitted inputs are 24 likely to correlate both with the intervention itself and with the selected outcome. For instance, parents may choose schools with smaller class size for their offspring and invest more of their personal time in their children's education. The effect of this joint choice is to create an upward bias in the estimated effect of class size. Conversely, if they invest less time, a downward bias is generated. A downward bias would also emerge if school administrators assigned pupils from a disadvantaged background or with learning difficulties to smaller classes, or if better teachers were assigned to larger classes.
To handle these problems, recent empirical research investigating class size effects has relied on identification strategies that exploit exogenous variations in class size induced by controlled or natural experiments. On the one hand, Krueger, 1999, andWhitmore, 2001, have focused on the Tennessee STAR project, 10 where pupils are randomly allocated to classes of different size. On the other hand, Angrist and Lavy, 1999-and also Browning and Heinesen, 2007and Leuven, Oosterbeek and Ronning, 2008  The focus of these researches is mainly the effect of class size on test scores, which indirectly affect the decision to stay in school. More recently, a few studies have looked explicitly at the effects of class size on educational attainment. Fredriksson, Ockert and Oosterbeek, 2013, using very rich data from Sweden and the exogenous variation in class size induced by a maximum class size rule, show that reducing class size by one pupil during the final three years of primary school increases years of schooling by only 0.05 years (or two-thirds of a month). They report that smaller classes increase the probability of having a college degree by 0.8 percentage points, improve cognitive test scores and increase earnings at age 27 to 42. Using these effects, the authors carry out a costs-benefits analysis, showing that for reasonable values of the discount rate the present discounted value of benefits associated to the policy exceeds the present discounted value of costs. They also estimate an internal rate of return (the discount rate that equalizes the present values of costs and benefits) equal to 17.8 percent.
Similar studies are provided by Browning andHeinesen, 2007, andBingley, Jensen andWalker, 2005, who investigate the effects of class size in Denmark. As in Fredriksson, Ockert andOosterbeek, 2013, the Browning andHeinesen, 2007, study relies on the exogenous variation provided by a maximum class-size rule. They find that reducing class size in the 8 th grade by one pupil (a 5% decline) increases educational attainment by 0.02 years on average and the probability of completing upper secondary education by 0.4%, a small effect.
Unfortunately, this study does not include a cost-benefit analysis. This type of investigation is instead carried out by Bingley, Jensen and Walker, 2005, who examine the effects of changes in class size in Denmark, using a different estimation strategy, that combines administrative rules with between-siblings differences. The advantage of this method is that it does not rely on the identifying assumption that class size is not influenced by parents. Using data on the whole Danish population, they find that class size has a statistically significant but very small effect on educational attainment. In particular, they estimate that reducing class size in the 8 th grade by 5% increases mean schooling attainment by about 0.005 years, a tiny effect. Effects are somewhat higher for households with three or four children, and much larger when focusing on large municipalities.
Using wage returns estimated with twin data, the authors compute the ratio of discounted benefits to costs associated to a 5% reduction in class size. As shown in Table 4 below, which is based on a 5 percent rate of return to education, this ratio is very low and outweighs the costs only for a zero discount rate and for males. In the US, Chetty et al, 2011, use data from the well-known Project STAR, where pupils are randomly assigned to classes of different size, and show that being assigned to a smaller class size produces beneficial effects on educational attendance, as randomly selected students are 1.8 percentage points more likely to be enrolled in college at age 20 than students assigned to larger classes. 12 In contrast to Bingley, Jensen and Walker, their cost -benefit analysis suggests that the expected benefits from reducing class size outweighs the costs. 13 It is worthwhile to notice that both Bingley, Jensen andWalker, 2005, andChetty et al, 2011, do not estimate the effect of class size on wages. The estimates by Bingley, Jensen and Walker are based on the assumption that class size affects wages by changing years of schooling. Chetty et al, on the other hand, assume that class size modifies wages because it changes test scores. Fredriksson, Ockert and Oosterbeek (2013), instead, provide an estimation of the effect of class size directly on wages. They also show, using their data, that indirect methods tend to underestimate the true effect. 14 In sum, the reviewed studies agree on the fact that reducing class size may have a positive albeit relatively small effect on educational attainment. Less agreement exists on the estimated costs and benefits of a policy that reduces class size early on, with some studies finding that benefits are much lower than the costs and other studies finding the opposite.

The content and structure of teaching
Early school leaving is often correlated to bad school performance. As a consequence, programs that try to improve student performance can also induce individuals to stay in education longer. It is well known that one key factor in determining student performance and attainment is teacher quality. 15 In the late 1990s, two programs were introduced in England with the aim of improving the teaching of literacy and numeracy in primary schools. At the beginning, these programs were implemented only in few cities, but were subsequently extended nationally as the 'National Literacy Strategy' and the 'National Numeracy Strategy ' (in 1998 and 1999 respectively).
These programs tried to improve the quality of teaching by providing more focused instruction and a more effective classroom management. For instance, a daily literacy hour was introduced and divided into 10-15 minutes of reading or writing for the entire class, 10-15 minutes of whole-class session on word work (phonics, spelling and vocabulary) and sentence work (grammar and punctuation), 25-30 minutes of directed group activities (on aspects of writing or reading), and a plenary session at the end to revisit the objectives of the lesson. Machin, McNally and Meghir, 2004, evaluate the effects of the 'literacy hour' relying on a differences-in-differences strategy. Since the implementation of the policy was preceded by pilot dollars). They assume a 3% annual discount rate and discount all earnings streams back to the point of the intervention. They estimate the average cost per child of reducing class size by 33% for 2.14 years (the mean treatment duration for STAR students) at $9,355. According to these calculations, the benefits of reducing class size slightly outweigh the costs. 14 An early attempt is the classical study by Card and Krieger, 1992. 15 See for example Rockoff (2004) and De Paola (2009). projects, they could compare educational attainment at the end of primary school in 'treatment schools' and in schools in an appropriately defined comparison group, before and after the 'pilot' project was introduced. Results show that the literacy hour has produced a 2-3 percentage point improvement in the reading and English skills of primary school children involved in the program. The authors estimate the costs per pupil of the program at £25.52 per annum, and the present discounted value of benefits in the range of £1,918-£4,995, suggesting that the policy had a very high payoff.

Targeting additional resources targeted to pupils or schools at risk
Students at risk of dropping out display easily identifiable characteristics, including among others low income, scarcely educated parents, and low attendance rates. Policies that try to change individual choices of subjects at risk are targeted either directly at pupils or at schools. In this section, we review some of the policies that were recently implemented in European countries.

Conditional cash transfers: the EMA project in UK
The use of conditional cash transfers to improve student performance is becoming a common practice in the educational systems of several countries. For example, in Colombia, the FA (Familias en Acciòn) and PACES (Programa de Ampliaciòn de Cobertura de Educaciòn Secundaria) programs assigned vouchers, respectively, to primary and secondary school students to encourage attendance and performance (Angrist, Bettinger, Bloom, King, and Kremer, 2002).
In Mexico, the PROGRESA financial incentive scheme offered cash payments to families to increase school enrolment and attendance (Schultz, 2004). In Kenya, the Girls' Scholarship Programme provided awards to female students who obtained high test scores (Kremer, Miguel, and Thornton, 2009). Similar programs have also been recently implemented in a number of developed countries. In the United States, following the "Helping Outstanding Pupils Educationally" (HOPE) program implemented in Georgia, several other states have introduced analogous scholarship programs. About 27,000 students benefitted from the financial incentives offered by a number of schools in Chicago, Dallas and New York (Fryer, 2011). As far as European Countries are concerned Leuven et al. (2010) and De Paola et al. (2012) study the effect of financial incentive respectively in the Netherlands and in Italy reaching mixed results.
Since many of these programs have been implemented following an experimental design (students were randomly assigned to a treatment and a control group), their effects have been evaluated rather extensively. It emerges that conditional cash transfers work well in increasing attendance and enrolment, especially in developing countries (Angrist, Bettinger and Kremer, 2006).
Among the programs implemented in European countries particularly relevant is the Earnings Maintenance Allowance (EMA). This project was piloted in England in 1999 before national implementation. The effects of the program have been evaluated by Dearden et al, 2009, using data from the first cohort of the EMA pilot study. 16 The respectively. 17 16 The study is based on survey data (face-to-face interviews with both the parents and the children and followup annual telephone interviews with the children). Data were collected both for eligible and ineligible individuals in pilot and control areas. Information includes: the youngster's current economic activity, the family income, parental economic activity, occupation, and education, childhood events (such as ill health and mobility), and prior school achievement. In addition the study uses administrative data on the quality of schooling in the relevant neighbourhood and other measures of neighbourhood quality measured prior to the introduction of EMA. 17 No relevant differences between males and females were detected. The effect of EMA was estimated to be largest for children with lower levels of prior educational achievement. In addition, the effect was larger for children in rented accommodation (who are more likely to be liquidity constrained) that for children living in owner-occupied housing, although the difference was not statistically significant.

29
The cost-benefit analysis conducted in this study suggests that the program benefits would outweigh the costs if students staying two additional years in school -and therefore completing upper secondary education -experienced a real increase in future earnings higher than 6.2 to 7.7 percent. These values are below the returns to education estimated by Oreopoulos, 2006, for UK -11 percent for males and 18 percent for females -but above those estimated by Devereux and Hart, 2010.

Additional resources to schools with disadvantaged children
In this section we report the results of three evaluation studies that consider programs implemented in France, the Netherlands and UK. All these programs devoted additional resources to schools with disadvantaged pupils.

The education priority zones program in France
The Education Priority Zones program (ZEP), first established in France in 1982, was targeted at both primary and junior-high schools located in disadvantaged zones. 18 As the main policy in France directed at helping students from a disadvantaged background, its main objective was to reduce the number of students who dropped out of the school system without any degree.
Schools included in the program received additional resources, which were mainly used for additional hours of instruction and to pay bonuses to teachers and other personnel. Benabou et al, 2009, use a sample of 24,455 students who entered 6 th grade in 1989 and information on school and teacher characteristics to evaluate the effects of this policy on the probability that targeted students move up either to the 8 th or to the 10 th grade of the academic track rather than switching to a vocational track after the 7 th or the 9 th grade, and on the likelihood of success at the "Baccalauréat" (the French national exam at the end of high school).
These effects are estimated using a differences-in-differences approach, which compares students in the same grade but of different cohorts, some who attended school before the program was implemented and others attending school after priority status was granted.
The results are somewhat disappointing in that the impact of ZEP on the academic achievement of students in targeted schools was never significantly different from zero, after conditioning for pre-existing differences between treated and control schools. One reason why 18 Plus a small number of high schools.

30
the policy was not successful was that it was not well focused. Furthermore, the bonuses awarded to teachers were not a sufficient incentive to attract experienced or highly qualified teachers.

Devoting additional resources to schools with disadvantaged students in the Netherlands
Devoting additional resources to schools with disadvantaged students has been a poorly performing policy also in the Netherlands. Leuven et al, 2007, use Regression Discontinuity Design (RDD) to evaluate two subsidies implemented in the Netherlands, which gave additional funds to schools with a share of disadvantaged pupils above 70 percent. While one subsidy provided extra resources to improve the working conditions of teachers, the other subsidy offered additional funding for the purchase of computers and software. By comparing schools with shares of disadvantaged students around a threshold value, the authors find that the policy has failed to increase student performance in nationwide tests, and had led to lower attainment.
These negative results could depend on the fact that the extra payment to teachers was not conditional on performance. The negative effect of additional computer resources might have occurred because instruction methods that use intensively computers are less effective than more traditional methods. These negative findings cannot be easily generalized to other contexts. The fact that these interventions focusing on schools with a large share of disadvantaged students did not succeed in increasing test scores does not imply that these policies would be ineffective if implemented in other schools.

Excellence in Cities (EiC) in UK
Excellence in Cities (EiC) was introduced in England in September 1999 to encourage participation in higher education. Schools in disadvantaged areas of England were given extra resources to try to improve standards. The original program was expanded over time and now covers about one-third of all secondary (state) schools. The main features of the EiC programme were: the employment of 'learning mentors', to help students overcome educational or behavioural problems; the provision of 'learning support units' to provide short-time teaching and support programs for difficult pupils; a 'gifted and talented programme', to provide extra support for 5-10 per cent of pupils in each school. Machin and McNally, 2012, evaluate the impact of the Excellence in Cities policy in secondary schools by assessing the extent to which the whole range of activities implemented by the program have led to an improvement in both learning and school attendance at age 14. The 31 analysis is based on national-level administrative records of pupil attainment and attendance before (1999) and after the policy was introduced (from 2000 to 2003). The identification strategy is based on a differences-in-differences approach and consists of comparing outcomes in schools where the EiC policy has been in place with outcomes in schools belonging to an appropriate comparison group. This strategy exploits the fact that the policy does not cover every disadvantaged or poorly performing school in England.
The authors find that EiC has contributed both to better learning, in term of higher scores in Mathematics (although not in English), and to higher pupil attendance. 19 The cost-benefit analysis involving 4,300 students in 181 classes. The intervention was randomized at the class level within each school (respectively 98 treated and 83 control classes). The principals of the secondary schools involved in the program were required to identify students at risk of dropping out and to help their families form less unrealistic aspirations and make better school choices by providing adequate information. Two meetings were held between the principal and the selected families with the purpose of illustrating the complexity of the choices available in the French educational system. These meetings had also the purpose of identifying the specific aspirations of families, evaluate whether these were realistic, provide information on alternative options, and explain the merits of vocational education.
The authors find that, as a result of the mentoring program, involved parents became less confident that their children would be able to graduate from high school. Enrolment in 19 The effects of the policy are heterogeneous, stronger the longer the policy has been in place and for medium-high ability pupils within disadvantaged schools. 20 Other studies have evaluated the effect of student involvement in extracurricular activities. These activities might help in developing skills, such as teamwork ability, self-confidence, taste for competition, that are relevant in many contexts. Eide and Ronan, 2001, analyse the impact of sport participation on educational attainment. Their identification strategy relies on an instrumental variable approach in which sport participation is instrumented with height at age 16. They do not find any effect of sport activities on high school dropout rates. The same instrument is used by Pfeifer and Cornelißen, 2010, who use data from the German Socio-Economic Panel and show that sport participation has a positive effect on educational attainment. However, the validity of the adopted instrument is questionable, since height is correlated both to education and to earnings conditional on education (see for example Persico, Postlewaite and Silverman, 2004;Case et al. 2009;Guven and Lee, 2013). Lipscomb, 2007, uses a fixed effect estimation strategy and finds that athletic participation is associated with a 2 percent increase in math and science test scores. vocational tracks increased, and both dropout rates and grade repetition fell by one third, a large effect. This effect persisted two years after the treatment, as differences between treated and control students further increased rather than to diminish as time went by. Although the authors do not provide a cost -benefit analysis, the costs of the program appear to have been small compared to the large effects produced by the policy on early school leaving.

Critical Assessment
This review of the studies evaluating policies that address directly or indirectly early school leaving in Europe leads to at least three considerations. First of all, the number of studies is limited. As demonstrated by the increasing number of evaluation studies focusing on programs implemented in other countries, such as the US and Canada, this is not due to a lack of interest in the subject. Second, serious evaluation is possible only when the policy is designed in a way that causal effects -from the policy to the outcomes -can be identified. Currently available evaluation methods rely on the definition of treatment and a control groups independently of the outcome of interest. In some cases, these groups are obtained by explicit randomization. In other cases, the evaluation uses the variations in the exposition to a certain policy that originate from an exogenous source that is not correlated to the outcome of interest (instrumental variables or natural experiment approach). Another recent technique, in the spirit of natural experiments, is the so-called regression discontinuity approach, which can be applied when the assignment to treatment is based on a specific exogenous threshold.
When public funds are allocated to policies, it is paramount to consider that rigorous policy evaluation is possible only under particular circumstances. Policies need then to be designed so as to allow sources of exogenous variation in the treatment, a necessary requirement for the identification of casual effects. 21 Only in this case it is possible to obtain reliable estimates of the policy impact, a crucial starting point for any serious cost-benefit analysis.
Third, evaluation requires accurate data. It is not surprising that most of the studies that we have reviewed in this report come from countries with a tradition of collecting and releasing to academic researchers rich individual data. Education affects individual, social and fiscal outcomes, ranging from labour market performance to health, lifelong learning participation, job and life satisfaction, crime, social cohesion, growth, tax revenues and benefits payments. A fullyfledged policy evaluation would require data on all these aspects. When such data exist, it is 21 Unfortunately, this point is often ignored by several UE member states.

33
potentially possible to measure the impact of different policies on each of these outcomes and try to assess which policy has produced the best overall impact.
Unfortunately, these data are often missing. For instance, since some of the effects of a policy intervention occur after several years, information should be collected on the life histories of treated and control groups. While this type of data is available in some North European countries, they are not available in most EU member states. In many situations, only a few outcomes are observable. In these cases, evaluations are less reliable, because policies may produce effects on variables that are important but not observed by the analyst.
As shown in our review of evaluation studies, cost-benefit analysis of European programs typically focus on wage returns, and tend to fail to calculate overall benefit-cost ratio. The computation of these ratios with the information at hand often requires strong assumptions. For instance, Levin and al, 2012, in their evaluation of US policies targeted at reducing early school leaving, rely on the very strong assumption that social benefits are a given percentage of observed private benefits.

Conclusions
In this paper, we have reviewed the issues surrounding the measurement of the costs of early school leaving to individuals and societies, and examined several policies that were implemented in Europe to reduce early school leavers. These policies include broad policies -such as changes in minimum school leaving age, tracking and school resources -as well as more targeted policies.
While our focus is mainly on Europe, we also consider important evidence from across the Atlantic.
Our assessment of the empirical literature that tries to measure the costs of early school leaving can be summarized in the following guidelines for an empirical study: 1. the causal effect of education on earnings and health should be identified, using credible exogenous variation; 2. it should be recognized that early school leavers are heterogeneous and have heterogeneous returns to education. Therefore, empirical research should go beyond estimating the average return to additional education, and should consider how these returns vary at different points of the distribution of early school leavers; 3. in the absence of long longitudinal data, cross section data can be used only at the price of strong assumptions, that have important implications on outcomes; 34 4. general equilibrium effects should be evaluated, especially if the proposed change in the share of early school leavers is large; 5. unemployment and health are key ingredients in the cost of early school leaving. In particular, differences in the probability of employment may be more serious than differences in earnings.
In the second part of the paper, we have looked at the policies implemented by EU member states to reduce early school leaving. We have highlighted that comparing different policiessome addressing all students and some others targeted at disadvantaged students -on the basis of a cost-benefit analysis is a very difficult task. Success in this exercise requires not only accurate data on outcomes and costs, but also that the same outcomes are considered, that individuals with similar characteristics are targeted and that similar evaluation methods are applied.
We have shown that polices devoting higher resources to disadvantaged students typically lead to small individual benefits. However, it is difficult to compare these low returns with those found for programs involving the whole population of students, because individuals with different characteristics are likely to differ in terms of their policy responsiveness. In addition, as the underlying assumptions that have to be met in order to obtain estimates of causal estimates differ across estimation methods, particular care must be paid when comparing evaluations using different approaches. Relative employment growth (percent) 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 year High-paid to middling Low-paid to middling Note: Employment growth averaged across countries, no imputation for countries with shorter data spans.