Andrew M. Jones 569an instrument for socioeconomic conditions in infancy in order to avoid the prob-
lems of unobservable heterogeneity bias that plague cross-section comparisons.
The impact of early life conditions is analyzed nonparametrically, comparing those
born in booms and recessions, and through duration analysis of the individual
mortality data.
12.2.2.3 Educational reforms
Lleras-Muney (2005) shows how historical changes in the US educational system
can be used as a natural experiment, based on a discontinuity design, to identify
the effect of education on adult health. The estimated effect is larger than previous
studies have suggested, with the magnitude of the instrumental variables estimate
of the local average treatment effect three times larger than the OLS estimate. The
natural experiment is based on changes across states in compulsory schooling and
child labor laws between 1915 and 1939. Identification stems from variation over
states and across time in the age at which children had to enter school, the age
at which they could leave school and get a work permit, and whether those with
work permits had to continue in school part time. Estimation uses a regression
discontinuity design, which attributes any jumps associated with school leaving
age to the policy effect. This is applied using a linear probability specification for
deaths as a function of years of schooling. Synthetic cohorts are constructed from
successive US Censuses (1960, 1970 and 1980) to select those who were aged 14
between 1915 and 1939 and to follow-up subsequent mortality rates. These are
synthetic in the sense that they do not follow the same individuals and are based
on gender, birth cohort and state of birth.
Educational reforms are also used as a natural experiment in Arendt (2005). In
this case the analysis focuses on Denmark and reforms in 1958 that removed for-
mal tests before middle school, and 1975, that increased the compulsory minimum
school leaving age. Data are taken from the Danish National Work Environment
Cohort Study (WECS), with two waves in 1990 and 1995, and covers workers
aged 18–59 in 1990. The impact of years of schooling on outcomes later in life
is estimated using two-stage conditional maximum likelihood (2SCML) estimates,
allowing for a random individual effect, for self-reported health. Models are also
estimated for body mass index and for an indicator of never having smoked. The
latter is included for comparison as, for most people, it is determined while they
are still in education. The impact of education on health is amplified when instru-
ments are used, but, at the same time, the standard errors are inflated so that
exogeneity is not rejected. However, tests suggest that there may be a problem of
weak instruments, as the reforms have low explanatory power in the reduced form
equations.
12.2.2.4 Health policies and reforms
Bleakley (2007) is a good example of combining a natural experiment with a
long-term follow-up to explore the economic consequences of a public health