Body Weight, Eating Patterns, and Physical Activity: The Role of Education

The items listed include foods such as bread, rice, pasta, salt-cured meats, poultry, beef, pork, milk, cheeses, eggs, fish, green vegetables, tomatoes and other vegetables, fruit, green salad, legumes, potatoes, salty snacks, sweets and desserts, olive oil, seed oils, butter, and lard; and beverages, such as water, soda, beer, wine, aperitifs, and liquor.

The nutritional charts are available online (http://​www.​inran.​it/​646/​tabelle\_​di\_​composizione\_​degli\_​alimenti.​html).

Following the charts provided by Harvard Medical School (http://​www.​health.​harvard.​edu/​newsweek/​Calories-burned-in-30-minutes-of-leisure-and-routine-activities.​htm), we compute the averages of kilocalories burned performing an hour of light, medium, or heavy physical activity in the professional or household environment.

We use the varimax rotation. Varimax relies on orthogonal rotation and maximizes the variance of the squared loading for each factor.

We also perform the factor analysis by gender. Using the alternative set of results in the estimation, however, provides identical results.

Compliance with the reform was not immediate. Although it obliged all students to graduate from the unique track, the change did not bring instant improvements owing to persisting selective mechanisms within school structures and among teaching staff. As a result, full compliance with the reform occurred only for the cohorts born during the 1970s.

According to Oreopoulos (2006), such individuals represent less than 10 % of the population exposed to the instrument.

An alternative approach entails a simple inclusion of interaction terms between our relevant covariates and the gender dummy variable in the estimation of the model for the pooled sample. However, by splitting the sample, we provide a more flexible specification and, therefore, more precise estimates. The estimates with interaction terms are available upon request.

This parsimonious model limiting the analysis to the sole exogenous controls excludes variables related to SES, which are potentially endogenous.

The implementation of the estimator was offered by Abadie et al. (2002), following the original work of Angrist and Imbens (1994). The estimation procedure in Stata follows Frolich and Melly (2010). We thus estimate the effect of education (E) on each Y, as instrumented by the schooling reform R (the treatment). We define Y ¹ as the Y value for individuals with lower secondary school; Y ⁰ is the Y value for other individuals. Moreover, E ¹ is the education status for individuals subject to the reform (R = 1), and E ⁰ is that for individuals born before the reform implementation (R = 0). The identification strategy is based on assumptions that Y ⁰, Y ¹, E ⁰, and E ¹ are jointly independent of R for covariates X. Furthermore, we assume “no defiers” (Pr(E ¹ ≥ E ⁰|X) = 1, nontrivial assignment (0 < Pr(R = 1|X) < 1), and first-stage relevance E[E ¹|X] ≠ E[E ⁰|X]. This set of assumptions ensures that the estimation is again confined to the treatment effect for compliers, who would not have graduated from lower secondary school if the reform had not been implemented. It does not capture the always-takers and never-takers, who make the educational choices independent of the reform regulations; nor does it include defiers, who are excluded from the analysis by assumption. Chernozhukov and Hansen (2005) propose an alternative approach to the estimation of quantile treatment effects that delivers a global identification strategy. It is, however, impossible to implement the strategy here because of its reliance on rank invariance or rank similarity, which is unlikely to hold in our setting.

To check the validity of the results, we estimate analogous specifications with a “placebo” IV, in which we artificially place the reform in different periods. However, we cannot reject the hypothesis that the coefficients on the placebo policy placed randomly in the seven years after the actual reform is null. Although this outcome may weaken our inference, it most certainly results from the gradual implementation of the reform that has spread its effect over time.

For all our IV estimates, the control group not affected by the reform consists of cohorts born in the proximity of the war era. These individuals, usually referred to in the literature as “survivors,” have better average education and health status, which may point to underestimation of the effect of education in our case.

To explore in more detail the nature of the educational gradient for BMI, CI, and CE, we investigate whether the estimated impact of education might be explained by income. Because individual income is very likely to be endogenous in our setting, we stratify our subsamples according to geographical areas of residence, which determine strong income differences in Italy. We thus divide the individuals (in two, three, and four groups) according to the average regional disposable income of families as registered in the 1960s and reestimate our OLS and 2SLS specifications. However, this additional exercise offers almost identical inferences in terms of statistical significance, magnitude, and gender heterogeneity, independent of the income level. In case of BMI, the reform seems to have been marginally stronger in terms of the compliers’ subpopulation for the low-income regions, which is also reflected in slightly stronger education coefficient estimates in 2SLS results. However, in both cases, we cannot reject the null hypothesis of equality of education estimates across income-specific subsamples. These results are available upon request.

For each outcome variable, we test whether quantile regression coefficients on education are statistically significant across conditional quantiles. In particular, we test two null hypothesis, the first one of the equality of coefficients across all quantiles ([q10]edu = [q25]edu = [q50]edu = [q75]edu = [q90]edu), and the second one of the equality of coefficient estimates between the 25th and the 75th quantile ([q25]edu = [q75]edu). The null hypothesis of equality is rejected for all subsamples and specifications. The only exception is for quantile estimates on calorie intake for females; however, single coefficients are not statistically significant and do not provide any inference for our study.

Toward this end, we ran an alternative specification, based on stratified samples according to macro-area income levels. Nevertheless, we did not obtain any additional inferences from the analysis, where conditional on the average level of disposable income of the macro-area residents, the gender heterogeneity in the effect of education and CI and CE remains unaltered. The macro aggregation of income measures is not likely to capture sufficiently the relevant variation explaining this particular educational gradient. The results are available upon request.