Andrew M. Jones 565(PSDP) on hookworm infection rates and on school attendance. The program
included drug therapy and public health education on avoiding hookworm infec-
tion, with the assignment of treatment randomly phased. Randomization was
done at the level of schools rather than individuals: one group of schools received
treatment in 1998 and 1999, another group only in 1999 and a third group only in
- Data was collected in 1998 and 1999 so, in the Miguel and Kremer study, the
first group are the treated and the second and third groups make up the controls.
Miguel and Kremer argue that randomization at the levels of schools is crucial in
this context as it avoids biases created by spillover effects of the deworming pro-
gram in reducing infection rates. They argue that an ideal prospective study would
randomize treatments across pupils within schools, across schools within clusters
and across these clusters. This multilevel variation in the assignment of treatments
could then be used to estimate different levels of the treatment effect in the case
where spillovers are important.
12.2.2 Natural experiments
12.2.2.1 Health shocks
Almond (2006) makes inventive use of the 1918 influenza pandemic as a natu-
ral experiment to provide evidence in favour of the “fetal origins hypothesis.”
Cohorts that were in utero during the pandemic, between the fall of 1918 and
January 1919, are shown to have poorer outcomes: lower educational attainment,
more disability, lower income, lower socioeconomic status and higher transfer pay-
ments. The pandemic has the potential to be used as a natural experiment: it was
unanticipated, the period of exposure was short and the impact varied systemat-
ically across states. The study uses discontinuity across birth cohorts to identify
the long-term effects, drawing on data from the 1960, 1970 and 1980 US Census
microdata (which identify quarter of birth). Geographic variation is also exploited,
based on the “laggard” states where the epidemic had less pronounced long-term
effects, although this does reduce the sample size available. This is a paper where
simple graphical analysis tells the main story, although it is backed up by thorough
statistical modeling.
Doyle (2005) makes innovative use of data on severe traffic accidents to measure
variation in unanticipated health shocks and finds that, in the United States, the
uninsured receive 20% less treatment and have a substantially higher mortality
rate. The Crash Outcome Data Evaluation System (CODES) links police accident
reports to hospital discharge data. This study uses data for Wisconsin covering
1992–97, with a sample of 28,236 individuals, 10% of whom were uninsured.
Severe traffic accidents are assumed to be unanticipated at the time that insurance
is taken out and the consequent use of health care is non-discretionary. Descriptive
evidence suggests that the uninsured are riskier drivers and have worse health prob-
lems, creating a problem of selection bias. To deal with selection a control group
is selected from those with medical insurance, but without car insurance. Within-
hospital variation and time effects are controlled for. The robustness of the findings
is checked by using the sub-sample where both insured and uninsured individuals