574 Panel Data Methods
variation within Californian zip code areas to identify the impact of air pollution
on infant mortality. They find a statistically significant effect even at low levels
of air pollution. The impact of this effect is quantified: it is estimated that reduc-
tion in pollution in California over the 1990s saved around 1,000 infant lives. The
study takes data from the California birth cohort files and matches it to EPA data
on air quality - specifically measures of carbon monoxide, ozone and particulate
matter (PM10) – and information on weather patterns from the National Climatic
Data Center. A linear model is used to approximate the discrete hazard function
for infant deaths and month, year and zip code fixed effects are included in the
model, this relies on variation within cells of observations defined by month, year
and locality. The study complements the natural experiment presented by Chay
and Greenstone (2003) that is described above.
12.2.4 Anti-tests
One way to assess the robustness of an identification strategy is to find an anti-
test (or placebo test). Anti-tests provide counter-evidence by applying a model
or identification strategy in a context where no effect should be detected. If an
apparent “effect” is found then the validity of the identification strategy must be
called into question.
For many years the standard empirical strategy to test for the phenomenon of
supplier-induced demand (SID) in medical care has been to include a measure of the
supply of doctors – usually the physician density, measuring the number of doctors
in a locality per head of population – in empirical models of health care utilization
or expenditure. This strategy is plagued by omitted variable bias and identification
problems. To assess the robustness of the approach, Dranove and Wehner (1994)
apply the physician density strategy using the obstetrician/population ratio and the
volume of births as the measure of utilization. The physician density test shows
evidence that the number of births (and hence pregnancies) is “supplier induced”:
casting obvious doubt on the reliability of the approach. However, failure of the
methodology does not imply rejection of SID. For example, Gruber and Owings
(1996) find evidence of increased C-section rates in response to a fall in fertility in
the US between 1970 and 1982: a shift by obstetricians to more lucrative procedures
in response to economic pressures.
In their “addiction to milk” paper, Auld and Grootendorst (2004) use non-
addictive substances, such as milk, eggs and oranges, to construct an anti-test and
demonstrate that evidence for the rational addiction hypothesis based on aggre-
gate data may be spurious. Numerous studies have applied the canonical rational
addiction equation of Beckeret al.(1994) to substances such as alcohol, cigarettes
and cocaine, and claim to have found support for rational addiction; but Auld and
Grootendorst (2004) show that these findings are mimicked when the model is
applied to Canadian aggregate data for non-addictive substances. Monte Carlo sim-
ulations show that spurious evidence is likely when the time series data exhibit high
serial correlation, when prices are poor instruments, when overidentified instru-
mental variable estimators are used, or when theoretical restrictions are imposed
by fixing the implied discount rate in the model.