Health and Education
Eating disorders are correlated with but not caused by stress at school
Johanna Sophie Quis, Leibniz Universität Hannover
Eating disorders are among the most common chronical diseases during childhood and adolescence. In Germany, one in five adolescents shows symptoms of eating disorders such as anorexia, bulimia, binge eating or obesity. These disorders link to stress, in particular at school. However, it is not clear ex ante , whether the relationship is causal and if so whether eating disorders cause stress at school (e.g. via a lacking ability to concentrate when underfed or overly full) or vice versa (stress at school causes bad feelings which then lead to loss of appetite or overindulgence of foods). In order to shed more light on the potential mechanism, we test whether there is a causal effect of more stress at school on eating disorders.
Our data consists of KiGGS wave 1, collected in the years 2009 – 2012. We use school grades and subjective well-being at school as proxies for stress at school. In a first step, we use a linear regression model to show the relation between our stress proxies and the probability of eating disorders. As a next step, we introduce a LASSO to identify covariates that can explain part of this relation. Additionally, we instrument our stress proxies with relative age and use this exogenous variation in stress levels to analyze the causal effect of stress on eating disorders.
Our preliminary results show a negative association between stress at school and symptoms of eating disorders. Part of this association comes from child specific covariates identified by LASSO. When we instrument stress at school, the negative association disappears.
Stress at school is correlated with eating disorders. We find that part of this relation is explained by child specific covariates. IV results suggest that there is no causal path from stress at school to eating disorders. Our results imply that preventing eating disorders is one intervention to reduce stress at school and increase child wellbeing.
Simon Reif, ZEW
Johanna Sophie Quis, Leibniz Universität Hannover
The effect of education on life expectancy: Evidence from death records
Jan Marcus, Universität Hamburg
A broad range of empirical studies from different disciplines document positive correlations between education and longevity. However, these correlations do not imply that education *causes* longer lives. Reverse causality and confounding third factors can be alternative explanations for the observed correlation patterns. For example, healthier individuals may just obtain more education, or better-educated individuals may just be more patient and, therefore, restrain from unhealthy behaviors. This project studies the causal effect of education on longevity exploiting arguably exogenous variation in individuals’ education stemming from school entry cut-off rules. We leverage the robust finding that individuals who are born just after the school entry cut-off date are relatively older in their cohort, and have a higher probability to attend a higher secondary school track compared to individuals born before the cut-off.
We use two large data sets from Germany, both including the exact date of birth for several million individuals. The first is the West German Census of 1970 and the second is the official Cause of Death Statistics for the period 1992-2017, which includes the universe of deceased individuals in Germany. We map information on state-specific school entry regulations into both data sets for the birth cohorts 1945-1959 and compare individuals born just before and just after the school entry cutoff in a regression discontinuity framework.
First, we provide evidence that individuals born just after the school entry cut-off (who are relatively older at school entry) are more likely to enter the middle secondary school track (Realschule) or the academic track (Gymnasium) by about 3 percentage points. Second, we find no differences in the life expectancy of individuals born just after the cut-off, despite their higher educational degrees. Third, we run simulations with our actual data and show that our large sample size and methodological approach would have sufficient statistical power to detect even small increases in life expectancy.
While higher educated individuals live longer, on average, our study cannot find evidence for a causal effect of education on longevity, despite very large administrative data. We conclude by discussing several explanations for this finding.
Mathias Huebener, DIW Berlin
Shushanik Margaryan, Universität Hamburg
Performance of Parallel Pre-Trend Tests in Difference-in-Difference Applications in Health Economics – A Monte-Carlo Study
Clara Pott, Universität Hamburg (HCHE)
The difference-in-difference (DID) method is a quasi-experimental research design often used by health economists to study causal relationships when randomized control trials are infeasible. An identifying, however, non-testable assumption is that in absence of an intervention, the trends in outcomes of treatment and control group would have evolved in parallel. The plausibility of this assumption is often examined by testing whether differences in pre-intervention trends are present. To date, there is no consensus about which methods to use and an evaluation of test performances is needed.
First, to investigate which parallel pre-trend tests are employed in health economic literature, we conducted a literature survey of 276 DID studies published in the 11 most cited health economics journals. Second, we evaluated performances of the three most common tests applied in the literature (Event studies, Placebo tests, Group-specific linear testing) using Monte Carlo simulation for varying sample sizes (10-2000 individual observations per group) and numbers of pre-periods (2-20), when no, small (10%, 25%), medium (75%) or large (100%) linear pre-trend deviations occur.
We find great diversity in the number and types of tests conducted. 90 of the 276 (32.6%) identified DID studies did not report any testing. Those assessing pre-trends mostly employed graphical inspection (89/32.2%), Event studies (36/13.0%), Placebo regressions (24/8.7%) and/or Group-specific linear trend testing (9/3.3%). Examining the power and type I error probability, we find substantial variations among these statistical tests. Event studies and Group-specific linear trends perform very well with regard to the type I error probability (0-0.03% and 0-0.12%), while Placebo testing exhibits quite a high type I error probability (0-28.8%). All tests perform equally well, when many observations are available and the trend deviation is large (100% avoidance of type II error). However, their power substantially differs for small and medium deviations of the linear trend. Group-specific linear trends and Event studies are insensitive and do not detect small trend deviations for sample sizes up to 200 (power: 0-20.7% and 0-32.1% for deviations up to 25%) while Placebo testing exhibits a power of 0-75.9%. Compared to the other tests, it performs especially well, if many pre-period data are available.
DID is a widely used tool for health policy analysis and there is no consensus about how to validate its vital assumption. We find that the most prominent tests for parallel pre-trends vary substantially in their performance and the choice of tests should rely on the study setting.
Clara Pott, Universität Hamburg (HCHE)
Jan Marcus, Universität Hamburg (HCHE)
Tom Stargardt, Universität Hamburg (HCHE)
Pandemics and human capital - Evidence on the long-run impact of school closures during the Spanish Flu
Daniel Kühnle, University Duisburg-Essen
A substantial body of evidence examines the effects of pandemics on mortality and the economy in the short and medium run. However, little evidence exists with respect to specific policies that aim to limit the spread of the disease, and in particular their long-run effects. Our paper addresses this gap and examines the long-run effects of school closures during the Spanish Flu of 1918 on human capital development. We use Swedish register data on the universe of individuals born between 1900 and 1914, and we observe their human capital outcomes at the 1960 and 1970 census. To examine the mortaliy effects of the pandemic, we use the universe of all deaths occuring between 1914 and 1921. We merge self-collected data on school closures to the data. We will exploit the staggered introduction of school closures within an event-study design to estimate the effect of school closures on individual's long-run human capital development. As we are still collecting data on school closures, we only have preliminary results so far. We are very confident that we will have the full paper ready in time for the conference.