Data

To measure the health effects of the reform, we use data from the Constances study, a large prospective community-based cohort study started in 2012.23 Participants were randomly selected to take part in the study and invited to undergo a 1 day clinical examination at 1 of 22 Health Screening Centres run by the National Health Insurance Agency, which covers 85% of the French population. At baseline, a range of comprehensive health assessments were carried out by health professionals, and participants were asked to complete questionnaires about their health, behaviours and socioeconomic circumstances. Our original sample comprised 33 762 individuals aged 45 years and older. We follow previous research19 to choose the bandwidth (number of cohorts born before and after the reform) and focus on respondents born 48 months before or after the cut-off for eligibility to the reform (n=18 928).

To document the impact of the reform on school leaving age we use the French Labor Force Survey (LFS), a household survey representative of the French population living in private households. We use data on the age at which respondents left full-time education measured in the 2012 wave. We include respondents born within 48 months of the reform, which includes 72 133 individuals.

Outcomes

In Constances, mental health was measured using the 20-item version of the Centre for Epidemiologic Studies Depression scale (CES-D),24 a validated depressive symptoms scale. Higher scores indicate higher levels of depressive symptoms and a cut-off of 16 for men and 20 for women was used to define elevated levels of depressive symptoms.25

Cognitive function was assessed by trained neuropsychologists using five tests.26 The Mini-Mental State Examination (MMSE) is a general test of cognitive function consisting of 30 questions assessing five areas of cognition: orientation, registration, attention and calculation, recall and language. The Digit-Symbol Substitution Task of the Wechsler adult intelligence scale IV assesses psychomotor speed. The respondents were presented with nine digit-symbol pairs, followed by a list of numerical digits. They were asked to write the corresponding symbol under each digit as fast as possible. We retain the number of symbols correctly associated in 90 s.27 The Free and Cued Selective Reminding Test (FCSRT) measures verbal episodic memory. Respondents were first asked to read and memorise 16 words, then asked to recall as many words as possible in 2 min. The neuropsychologist provided a cue for each word that has not been retrieved. For this study, we include the FCSRT free immediate recall score. Finally, two verbal fluency tasks were included to assess language abilities: number of words related to the ‘animals’ category named in 1 min for the semantic fluency task and number of words starting with the letter ‘R’ named in 1 min for the phonemic fluency task. As the scale of each individual cognitive test differs, each score is converted to a z-score by subtracting the mean and dividing by the SD. A global cognitive score is obtained by summing the standardised tests. High cognitive performance is defined as the top quartile of the global score’s distribution.

Physical functioning is assessed based on three objective health measures. Hand grip strength was assessed using a handheld dynamometer. Participants were asked to sit and squeeze the dynamometer for a few seconds with their dominant hand. The exercise was repeated three times and grip strength was determined as the average of the three results. Finger tapping, a measure of motor performance, was assessed by the number of taps respondents could do with the index finger of their dominant hand for 15 s. Walking speed was measured as the time taken by respondents to walk 3 m at a rapid pace. For each physical health outcome, high levels of physical functioning are defined based on the top quartile of the distribution.

Years of schooling and completed education

Schooling is measured as the age at which respondents left full-time schooling (in the French LFS) and the highest level of education completed (in the Constances data). Educational attainment in Constances is reclassified into four categories: primary education or less, lower secondary education, upper secondary education and tertiary education.

Parental socioeconomic position

Parental SEP during the respondent’s adolescence is based on the French classification of professions and socioprofessional categories. It is recategorised into high (managerial, professional occupations), intermediate (trade and services related occupations) and low parental SEP (manual occupations).

Analysis design

We implement a Regression Discontinuity Design (RDD), a quasi-experimental approach used to isolate the causal effect of an exposure using observational data.28 It aims to address unmeasured confounding and reverse causality by exploiting the discontinuity in schooling duration generated by the reform across otherwise similar cohorts. The reform raised the minimum school leaving age for all children born after 1 January 1953 from 14 to 16 years. Our approach classifies respondents born after this cut-off date as the eligible group, while respondents born prior to the cut-off were classified as the ineligible comparison group. RDD compares outcomes between these groups, appropriately accounting for secular trends using local linear regression.

Because some respondents would have stayed at school beyond the age of 14 in the absence of the reform, we implement a fuzzy RDD and estimate an intent-to-treat effect. Our estimates do not capture the impact of additional schooling per se, but the impact of eligibility to a policy that requires students to stay until age 16 at school. Our main model specification is as follows:

(1)

where Health _ict is a health outcome measure for individual i from birth cohort c at time t; D _ic is a binary variable taking the value of 1 if an individual was born after the cut-off date, and of 0 if the individual was born prior to this date; R _ic is an individual’s birth cohort relative to 1 January 1953 measured in months, which aims to capture secular trends and X _ict is a vector of individual characteristics: age, age squared, gender and month of birth. The coefficient γ₁ denotes the health effect of being eligible to the reform. Using local linear regression we model birth cohort effects as a piecewise linear function, with different slopes on each side of the cut-off date.

We explore heterogeneous effects by first stratifying our models by gender and parental SEP and then interacting the reform dummy with these variables. Ordinary least squares regressions are used to obtain coefficients and 95% CIs for continuous scores and logistic regressions for OR and 95% CIs for dichotomous outcomes. SEs are clustered at the month of birth level, as recommended when the variable which determines eligibility is discrete.29 To ensure that the functional form is correctly specified, we also estimate models with a cubic term for age. Additional models using triangular kernel weights are estimated to test whether our findings are robust to giving more weights to observations which are closer to the cut-off for eligibility. All analyses were carried out using Stata V.14.

Effects on cognition

Our findings are consistent with previous evidence of long-term positive effects of schooling reforms on memory scores,3 4 in particular among men.31 These findings suggest that increased schooling may increase cognitive reserve and thus postpone the manifestation of cognitive decline in later life.32 These effects were confined to men, which might be in part due to the relatively young age of respondents in our sample: women have lower incidence of dementia before 75 and are less affected in early old age by vascular midlife risk factors than men.33

Effects on mental health

Eligibility to the reform was associated with higher levels of depressive symptoms among women. We found a clear association between educational attainment and depressive symptoms in our sample for both genders.34 However, the sign reversed for women once we implemented an RD design exploiting the change in schooling duration. This result echoes findings from recent quasi-experimental studies in Sweden, Great Britain and Turkey15 16 35 reporting negative effects of schooling on mental health and emotional well-being. A recent review concluded that evidence of a positive effect of schooling on women’s health is overall weaker than for men.22

There are several possible explanations for these results. Education is hypothesised to influence health partly by enhancing access to health-promoting resources such as health literacy, self-efficacy, social networks and better jobs and earnings.1 Available evidence suggests that the French reform—contrary to other schooling laws in the USA and Western Europe—did not lead to substantive improvements in qualifications or earnings.36 Our analyses partly confirm this pattern: longer schooling did not translate into higher qualifications or wages for either gender but was associated with higher odds of being in employment in 2012 among men. For women, the additional years of schooling were not associated with an upgrading of their position in the labour market, potentially creating a job-education mismatch detrimental for mental health.37 The discrepancy between the expectations created by the reform and the actual long-term benefits might have been particularly harmful for women’s mental health as research suggests that women have smaller economic returns on education.38 A second hypothesis is that the health effects of education for women operated partly through the marriage market and not exclusively through employment and wages.22 39 However, the reform did not exert any effects through this pathway as eligibility was not associated with women’s marital status or number of children.

Effects on physical functioning

Consistent with our findings for physical functioning, previous studies have reported weak or inconsistent effects of schooling laws on physical health.40 There are three possible interpretations. First, we may not have enough power to detect the effect of the reform on physical health: the CIs for these estimates are wide and include substantial benefits as well as harms as plausible effects. Second, our sample was relatively young and these measures may not discriminate subtle differences in physical functioning in younger age. A third possible interpretation is that the commonly observed association between education and functioning may be partly driven by selection. A potential implication would be that increasing compulsory schooling may not be an effective policy to improve physical health. A recent theoretical framework linking human capital, schooling and health identified skills formation as a critical pathway.22 Our results indicate that degree attainment and skill formation—which were not impacted by the Berthoin reform—might be more important for health than the amount of time spent in school.

Strengths and limitations

Major strengths of our study include the use of a quasi-experimental design in a large population-based cohort, as well as an extensive battery of objective health measures. RDD is a powerful tool to estimate causal effects of policies for which eligibility is based on a clearly defined threshold or time point.28 A limitation is that our data are cross-sectional, so we were unable to examine longitudinal trajectories. Reverse causality, however, is not a concern in our study because the reform affected cohorts decades before health measures were taken. In addition, if exposure to the policy is ‘as good as’ random, using cross-sectional data should not compromise estimates of the impact of the policy on health outcomes in later life.

Taken together, our findings provide a mixed picture of the long-term health effects of a policy that increased compulsory schooling. Decades after implementation, the Berthoin reform led to better cognitive outcomes among men, but also to increased rates of depressive symptoms among women. The reform had no measurable impact on physical functioning. Our results do not suggest that education is not beneficial for health but indicate that policies focused on increasing the quantity of schooling and not its quality, for example, might not be the best tool to improve population health.