Professional Documents
Culture Documents
Aggregation and The Estimated Effects of Economic Conditions On Health Jason M. Lindo
Aggregation and The Estimated Effects of Economic Conditions On Health Jason M. Lindo
Aggregation and The Estimated Effects of Economic Conditions On Health Jason M. Lindo
November 2014
Abstract
This paper considers the relationship between economic conditions and health with a
focus on different approaches to geographic aggregation. After reviewing the tradeoffs as-
sociated with more- and less-disaggregated analyses, I update earlier state-level analyses
of mortality and infant health and then consider how the estimated effects vary when the
analysis is conducted at differing levels of geographic aggregation. This analysis reveals
that the results are sensitive to the level of geographic aggregation with more-disaggregated
analysesparticularly county-level analysesroutinely producing estimates that are smaller
in magnitude. Further analyses suggest this is due to spillover effects of economic conditions
on health outcomes across counties.
I am extremely grateful to Coco Padilla-Romo and Eric Steinmann for research assistance. I also thank
Alan Barreca, Ben Hansen, Mark Hoekstra, Doug Miller, Adriana Lleras-Muney, Marianne Page, Chris Ruhm,
Jessamyn Schaller, Ann Huff Stevens, Glen Waddell, Wes Wilson, and Xiaohan Zhang along with seminar par-
ticipants at the University of Maryland, the University of Oregon, and the 2012 Meetings of the Society of Labor
Economists for their valuable comments. Contact at Department of Economics, Texas A&M University, College
Station, TX 77843; jlindo@econmail.tamu.edu.
1 Introduction
Although Harvey Brenners pioneering research suggested that health deteriorates during
recessions (1973, 1975, 1979), follow-up work has revealed that estimates based on aggregate
time-series data are quite fragile (Forbes and McGregor 1984; McAvinchey 1988; Joyce and
Mocan 1993; Laporte 2004; Gerdtham and Johannesson 2005). As Ruhm (2000) points out,
this fragility is not surprising since any lengthy time-series is likely to suffer from substantial
omitted variables bias.1 Out of concern for such biases, researchers have largely stopped
considering nationwide changes in favor of an area approach that considers how the health
of individuals living in an area changes over and above changes occurring across all areas when
its economic conditions change over and above changes occurring across all areas. The intent,
however, has remained the same: to estimate the degree to which health outcomes respond to
changes in macroeconomic conditions. These studies have repeatedly concluded that recessions
are good for health in developed countries, though this interpretation relies on the assumption
that health is similarly influenced by macroeconomic conditions (broadly defined) and more local
area-specific economic conditions.2 This study is motivated by the idea that this assumption
cannot be tested directly since credible estimates of the effects of macroeconomic conditions
(broadly defined) would appear to be out of our reach, but that it can be tested indirectly by
investigating the extent to which more local and less local economic conditions have different
estimated effects on health. More generally, this paper is concerned with the way in which
geographic aggregation (or disaggregation) influences the conclusions we draw from our analyses
There are several issues to consider when choosing the level of geographic aggregation and
interpreting the results of the analysis. From an economic perspective, we must recognize that
there are many mechanisms through which economic conditions can affect health and, from a
statistical perspective, we must recognize that the level of geographic aggregation influences the
degree to which the estimates capture these different mechanisms. For example, we would expect
1
For example, it is problematic for this approach that penicillin became increasingly available as the United
States began to recover from the Great Depression.
2
See Ruhm (2000, 2003, 2005, 2007, 2013), Dehejia and Lleras-Muney (2004), Johansson (2004), Neumayer
(2004), Tapia Granados (2005), Gerdtham and Ruhm (2006), Lin (2009), Miller et al. (2009), Stevens et al.
(2011).
3
With few exceptions nearly all area studies using U.S. data define areas as states. To my knowledge, the
only exceptions are Currie and Tekin (2011) who consider foreclosures at the zip-code level in four states and
Dehejia and Lleras-Muney (2004) who consider unemployment rates and supplement their state-level analysis
with a county-level analysis of California.
1
the effects of individuals job losses to be captured by changes in economic conditions in the
local area where individuals work. However, economic conditions both near and far may affect
about economic conditions, traffic congestion, levels of pollution, the quality of medical care,
government policies, and through effects on the members of ones social network. In terms of
identification, it is important to keep in mind that the estimated effects of an areas economic
conditions are fully inclusive of spillover effects across subareas within the area whereas the
estimated effects of subarea economic conditions are not. For example, estimated effects of
state economic conditions are fully inclusive of spillover effects across counties within a state
At the same time, more-disaggregated analyses can offer more precise estimates because they
use variation in economic conditions idiosyncratic to the area in addition to variation driven by
variation in economic conditions that are masked in more-aggregate measures. For example, a
contraction in one part of a state that is offset by an expansion in another part of the state
would contribute to county-level estimates but not to state-level estimates. Similarly, a county-
level analysis would exploit variation in the severity of contractions (and expansions) across
different parts of the state whereas a state-level analysis would not. Thus, more-disaggregated
analyses may be able to detect statistically significant effects of economic conditions where less-
aggregated conditions cannot, even if they do not capture some of the spillover effects captured
in more-aggregated analyses.4
error, which is especially problematic for fixed-effects estimators (Griliches 1977; Griliches and
Hausman 1986; Hausman 2001). The most common measure of economic conditions used in
this literature is the unemployment rate, produced by the Bureau of Labor Statistics (BLS).
However, as one considers smaller areas, one needs to be more and more concerned about
measurement error in unemployment rates since they are based in part on household surveys
(Bartik 1996; Hoynes 2000).5 For this reason, employment-to-population ratios would seem
4
On a related note, more-disaggregated analyses can have improved power because they allow for a richer set
of control variables in a manner that reduces the amount of unexplained variation in outcomes. This will not
necessarily be the case, however, because smaller areas may have more more variation in outcomes overall. The
richer set of control variables, of course, may also help to mitigate concerns about omitted variable bias.
5
Angrist and Krueger (1999) provide intuition: errors tend to average out in aggregate data. Another
problematic aspect of unemployment rate data is that the BLSs substate estimates prior to 1990 are no longer
considered official BLS data because they have not been revised to be consistent with the BLSs current
2
preferable because they are based solely on administrative data. Still, one may have concerns
It is also important to consider the fact that migration is influenced by economic condi-
tions (Blanchard and Katz 1992; Saks and Wozniak 2011), particularly among highly educated
(Bound and Holzer 2000; Glaeser and Gyourko 2005; Wozniak 2010; Notowidigdo 2011) and
healthy individuals (Halliday 2007). This heterogeneity implies that the estimated improve-
ments in health associated with recessions will understate the true improvements in health
if standard demographic controls do not fully capture these sorts of compositional changes.
That said, it is unknown whether education and other characteristics associated with health
considers different types of moves and different measures of economic conditions. As such, it
is unclear whether this sort of composition bias is likely to be of greater or lesser concern for
more-disaggregated analyses.6
Given the complexity of these issues, I take as my starting point that it is not at all clear
what level of geographic aggregation is preferred but that the tradeoffs deserve consideration
and that much can be learned by comparing the results of alternative approaches. As such, after
describing the different ways that I define areas throughout the subsequent sections in Section
2, I then replicate and update earlier state-level estimates of the relationship between economic
conditions and mortality (Ruhm 2000 and Stevens et al. 2011) and the relationship between
economic conditions at the time of conception and infant health (Dehejia and Lleras-Muney
2004) in Section 3. I then consider how and why estimated effects vary when the analysis is
conducted at differing levels of geographic aggregation for mortality and infant health in sections
1. The estimated links between economic conditions and health outcomes are sensitive to the
2. Analyses that simultaneously consider the economic conditions of a county and the eco-
estimation procedure.
6
The systematic outmigration that occurs when an areas unemployment rises also highlights the impor-
tance of well measured population denominators in calculating mortality ratespopulation measures that do not
account for the systematic outmigration caused by economic downturns will lead to mechanical reductions in
mortality rates.
3
nomic conditions of surrounding areas reveal significant spillover effects on health out-
comes. For example, the economic conditions outside a county in the same state have
explanation for why the estimated effects of state (or economic area or region) economic
conditions are larger than the estimated effects of county economic conditions when each
spillover effects across counties within a state, which are quite important.
3. Because they have more power, more-disaggregated analyses have the potential to reveal
statistically significant links between economic conditions and health outcomes even when
the estimated effects are smaller in magnitude. For example, while state-level estimates
using recent years of data suggest that the link between mortality and economic condi-
tions may no longer exist (Ruhm 2013), more-disaggregated analyses indicate that the
4. Two-way clustering (Cameron et al. 2011) on areas and years tends to yield larger
standard-error estimates than the typical approach that clusters only on areasparticularly
within areas over time and to be correlated across all areas within year in estimating the
2 Defining Areas
Throughout my analyses, I separately identify the effects of economic conditions using data
at five different levels of geographic aggregation, from counties to regions. Except where repli-
cating earlier research, all of my analyses are based on data that are available at the county
level to ensure that all estimates are based on the same underlying population. In particular,
I do not use data from: Alaska or Hawaii where county population data (described in greater
detail below) are unavailable prior to 1990; counties without at least one person of each age
(085) in 1990 so that age-adjusted mortality rates can be calculated at the county level; and
Virginia where there have been substantial changes to county definitions over time.7 Section
7
In an additional additional effort to achieve consistency in the data over time, for all years I combine the
counties of La Paz and Yuma (Arizona) which split in 1983; Washabaugh and Jackson (South Dakota) which
merged in 1983; and Adams, Boulder, Jefferson, and Weld (Colorado) from which Broomfield split off in 2001.
4
3 demonstrates the trivial impact these sample restrictions have on the estimated effects of
My most-aggregated analyses use region-level data based on the eight regions defined by the
Bureau of Economic Analysis (BEA), which were developed in the 1950s with the intention of
grouping states with similar economic economic and social factors.8 My next most-aggregated
analyses use state-level data, followed by analyses that use data at the BEA-Economic-Area level
and analyses that use data at the BEA-Component-Economic-Areas level.9 Last, as mentioned
above, I conduct analyses based on data at the county level. Naturally, the sample restrictions
described above reduce the number of states, Economic Areas, Component Economic Areas,
Before presenting estimates of the link between economic conditions and health outcomes
using different definitions of areas in the next section, in this section I present the estimated
effects from prior state-level studies, adapt the estimation strategies in a way that allows for
more-disaggregated analyses, consider how these changes alter the estimates, and then extend
the analyses to include additional years of data. I begin by focusing on mortality outcomes and
3.1 Mortality
In his highly influential study, Ruhm (2000) considers the effects of macroeconomic condi-
tions on health using mortality rates from Vital Statistics of the United States publications and
estimates of the unemployment rate from the BLS. In order to focus on within-area variation,
8
In contrast, the five regions used in the United States census were developed in the 1880s with similar
intentions. The BEA region groupings are as follows. New England: Connecticut, Maine, Massachusetts, New
Hampshire, Rhode Island, Vermont. Mideast: Delaware, District of Columbia, Maryland, New Jersey, New York,
Pennsylvania. Great Lakes: Illinois, Indiana, Michigan, Ohio, Wisconsin. Plains: Iowa, Kansas, Minnesota,
Missouri, Nebraska, North Dakota, South Dakota. Southeast: Alabama, Arkansas, Florida, Georgia, Kentucky,
Louisiana, Mississippi, North Carolina, South Carolina, Tennessee, Virginia, West Virginia. Southwest: Arizona,
New Mexico, Oklahoma, Texas. Rocky Mountain: Colorado, Idaho, Montana, Utah, Wyoming. Far West:
Alaska, California, Hawaii, Nevada, Oregon, Washington.
9
BEA Economic Areas define the relevant regional markets surrounding metropolitan or micropolitan sta-
tistical areas and are constructed such that all counties are a part of some BEA Economic Area. Component
Economic Areas are much the same as Economic Areas except they can be based on fewer counties and fewer
residents. For example, a Component Economic Area would be considered too small to form an Economic Area
if it consisted of fewer than three counties and fewer than 500,000 employed residents. Also note that 167 of
344 Component Economic Areas are sufficiently large that they also serve as independent Economic Areas; the
remaining 177 combine to create the remaining 12 Economic Areas.
5
Ruhms estimates are based on the regression equation
where Hat is a measure of health for individuals living in area a in year t, Eat is a measure of
economic conditions, Xat is a vector of time-varying controls, t are year fixed effects, a are area
fixed effects, and a t are area-specific time trends. To operationalize this identification strategy,
Ruhm defines a as a state, Hat as the natural log of the mortality rate (deaths per 100,000),
Eat as the unemployment rate, and Xat includes the fraction of the state population that: is
less than five years old, is 564 years old, is greater than 64 years old, is black, is hispanic,
is a high school dropout, has some college, is a college graduate.10 In Column 1 of Table 1,
I display the estimate based on these data and definitions while clustering the standard-error
estimate on the state.11 The point estimate indicates that a one-percentage-point increase in
As the starting point for my analysis, I begin with the approach used in Stevens et al.
(2011) who extend Ruhm (2000) in several ways. First, they create mortality rates using death
counts from Vital Statistics micro-record multiple-cause-of-death files and population data from
SEER (described above). Second, they pool monthly Current Population Survey data in order
to construct overall unemployment rates and unemployment rates for separate demographic
groups, which requires them to focus on data from 19782006. Third, they include a richer set
of controls for the fraction of a states population in various age brackets (less than 5, 517,
1830, greater than 64) using the aforementioned population data, construct the fraction black
with the aforementioned population data, and use CPS data to measure the fraction Hispanic
and the fraction in each education group.12 Last, they replace the natural log of the mortality
rate with the natural log of the age-adjusted mortality rate to account for uneven changes in
the age distribution across states. The age-adjusted mortality rate is calculated by taking the
weighted average of the mortality rate for each age, using as weights the fraction of individuals
in each age category in the US in 1990. After making these changes, their estimate (shown in
6
is associated with a 0.33-percent decrease in overall mortality.
Use state-level unemployment rates published by the BLS for convenience. This change
causes the estimated semi-elasticity to change from -0.0033 to -0.0039 (Column 3).
Omit controls for the fraction Hispanic and the fraction in each education category, which
cannot be reliably measured for small counties. This change has a trivial impact on the
Add data from 19761977 and 20072010 to the analysis. Incorporating the earlier years
of data causes the semi-elasticity estimate to change from -0.0040 to -0.0049 (Column 5).
Restrict the sample to individuals living in counties with at least one person of each
age (085) in 1990 so that age-adjusted mortality rates can be calculated at the county
level. Also, omit data from Alaska and Hawaii where county codes are unavailable in
the early years of the data and from Virginia where there have been substantial changes
to county definitions over time. So that the unemployment rates correctly correspond
to the counties contributing to the state-level estimates, use county-level BLS data to
construct unemployment rates. These changes have a negligible impact on the estimated
Use the death rate as the outcome variable rather than the log of the death rate, which
may be undefined in small counties. This changes the interpretation of the estimate, which
3.7 fewer deaths per 100,000 (Column 8). The semi-elasticity implied by this estimate,
calculated by dividing coefficient estimate by the mean of the outcome, is nearly identical
7
3.2 Infant Health
Like Ruhm (2000), Dehejia and Lleras-Muney (2004) conduct a state-level analysisalso
characterized by Equation 1but instead focus on various measures of infant health and the
effects of economic conditions at the time of conception. As one of their primary outcome
measures, they consider the fraction of children who have low birth weight (less than 2500
grams), calculated at the race-by-state-by-year level based on vital statistics birth records for
white and black mothers at least 18 years old. As their primary measure of economic conditions,
they use state unemployment rates in the year of conception, which they base on the timing
of the mothers last menstrual period. Column 1 of Table 2 displays their estimates, which
indicate that a one-percentage-point increase in the unemployment rate reduces the fraction of
analyses. In Column 2, I instead define a childs year of conception as nine months prior to
birth and include in the analysis children born to mothers for whom information on the last
menstrual cycle is missing. The subsequent columns progressively include children classified as
other race to the analysis and an indicator variable for other race to the regression model
(Column 3); include children born to mothers of all ages (Column 4); collapse the state-by-year-
by-race level data to the state-by-year level while adding controls for the fraction in each race
category (Column 5); control for the age distribution of mothers with variables corresponding
to the fraction of mothers who are less than 18 years old, 1822 years old, 2328 years old,
and 2934 years old (Column 6); update the sample to include data from 20002010 (Column
7); and then construct the sample based on data available at the county level in the same
manner described above (Column 8). These changes tend to have only minor impacts on the
estimated effects with one exception. When the data is extended from 19761999 to 19762006
effect of economic conditions on newborn health has grown over time. Column 9 indicates
that using the employment-to-population ratio as the measure of economic conditions produces
8
4 Estimates of the Unemployment-Mortality Relationship Us-
In this section, I begin by exploring how the estimated effects of economic conditions on
mortality vary with different levels of geographic aggregation using the identification strategy
described in the previous section. Though the convention in this literature is to cluster standard
errors at the area level, it may instead be appropriate to cluster on broader areas where possible,
e.g., clustering on regions for state-level analyses, because shocks to health outcomes (that are
not captured by the variables in the model) may be correlated across adjacent areas.14 That
said, the standard-error estimates based on the conventional approach tend to be the same as
or more conservative thanestimates that cluster on broader areas, except when the analysis
is conducted at the county level. For county-level estimates, clustering the standard errors at
the state level yields the largest estimates (though quite similar to the estimates that cluster on
BEA economic areas), which indicates that it is important to allow the errors to be correlated
across counties within states when conducting inference. As such, I take this approach where
applicable. Further recognizing that year-specific shocks may be correlated across areas, I have
also investigated the additional importance of allowing the errors to be correlated across areas
within a given year by estimating two-way standard estimates following Cameron et al. (2011).
I demonstrate that this can be quite importantby way of a comparison with standard-error
estimates that do not allow for two-way clusteringin the first table that follows before taking
Table 3 focuses on how the estimated effect of economic conditions on overall mortality
varies with different levels of geographic aggregation across columns 1 through 5. Using the
broadest definition of area, region, the estimate indicates that a one-percentage-point increase
in the employment-to-population ratio is associated with an additional 3.44 deaths per 100,000.
The estimate is a bit larger when the analysis is instead conducted at the state level (3.71 per
100,000) and then the estimates become smaller and smaller as we progressively consider more
narrowly defined areas. That said, these changes are relatively modest when we move from the
state level to the BEA Economic Area to the BEA Component Economic Area (3.71 to 3.03 to
2.82). In contrast, the magnitude of the estimate falls substantially to 1.75 when the analysis
9
Table A1 in the Appendix takes the same approach to separately estimate the effects on
youth mortality (age 017), working-age mortality (age 1864), and elderly mortality (age 65+).
This exercise is motivated by Miller et al. (2009) who explain that, even though unemployment
has significant effects on younger and older individuals, the additional deaths that tend to be
observed during recessions largely consist of the elderly. Again, the magnitudes of the estimates
are smaller in more-disaggregated analyses, except for working-age mortality for which there is
Table A1 also shows estimated effects on mortality due to cardiovascular problems, motor-
vehicle accidents, and suicides.15 Echoing earlier work, these estimates reveal that economic
downturns reduce deaths due to cardiovascular causes and motor-vehicle accidents while in-
creasing deaths due to suicides. Moreover, these estimates exhibit a pattern consistent with
those for overall mortalityregional and state-level analyses yield estimates largest in magni-
tude while county-level analyses yield estimates smallest in magnitude, with economic-area-level
As a whole, the results shown in Table 3 and A1 demonstrate that the estimated effects of
economic conditions on mortality depend on the level of geographic aggregation. This raises
the question: do the results vary because of the level of aggregation of the measure of economic
conditions, because of the aggregation of other variables, or something else? In addition, are
Table 4 sheds light on these questions by considering the effects of the economic conditions
of regions, states, economic areas, component economic areas, and counties in separate regres-
sions using county-level data and the set of associated county-level control variables. While
this alternative approach does bring the estimated effects of economic conditions of regions,
states, economic areas, and component economic areas closer to the estimated effects of county
economic conditions, the latter remain considerably smaller in magnitude.16 In particular, the
10
is 1.75 additional deaths per 100,000 while the estimated effect of a one-percentage-point in-
crease in state employment-to-population ratios is 3.25 additional deaths per 100,000. This
difference is statistically significant at the one-percent level. We can similarly reject hypotheses
that the effect of county economic conditions is the same as the effects of economic conditions of
regions, economic areas, and component economic areas. We generally cannot, however, reject
hypotheses that the effects of these broader measures are same.17 Table A2 in the Appendix
Table 5 takes another step towards understanding why the estimated effects of county eco-
nomic conditions are smaller than the estimated effects of broader measures by simultaneously
considering the effects of economic conditions in counties and in their surrounding areas. To
begin, Column 1 shows the results from a regression that estimates the effects of county and
state economic conditions, where statewide economic conditions are calculated for each county
using information from other counties in the state. This estimate indicates that the health
itself and economic conditions in the rest of the state. In other words, there appear to be
spillover effects across counties in the effects of economic conditions on mortality.18 This esti-
mate is notably consistent with the estimates from regressions separately considering the effects
increase in the employment-to-population ratio across all counties in a state is expected to in-
crease the mortality rate (per 100,000) by 3.69 which is quite similar to the estimate when state
employment-to-population ratios are considered on their own (3.71 and 3.25 in tables 3 and 4,
respectively).
More broadly, these results suggest an explanation for why the estimated effects of state (or
17
In order to estimate the standard error for the difference in parameter estimates across the different models
in a way that allows for two-way clustering on states and years, I follow Cameron, Gelbach, and Miller (2011) and
obtain (1) an estimate of the variance for the difference in the parameter estimates that allows for clustering on
states, (2) an estimate of the variance for the difference in the parameter estimates that allows for clustering on
years, and (3) an estimate of the variance for the difference in the parameter estimates that allows for clustering
on state-years. I obtain these estimates by (cluster) bootstrapping and then combine them in order to obtain
an estimate of the standard error for the difference in parameter estimates that allows for two-way clustering on
states and years. In particular, this standard-error estimate is calculated as the square root of (1) plus (2) minus
(3).
18
Though the point estimate suggests that the effect of county economic conditions is smaller than the effect
of economic conditions in the rest of the state, the difference is not statistically significant. That said, the
difference is statistically significant when focusing on youth mortality rates and motor-vehicle-accident mortality
rates, shown in Table A3 in the Appendix. This offers an interesting angle for future research as motor vehicle
accidents are the leading cause of death for youths and one might suspect that this finding is related to the
fact that over a quarter of workers commute across counties for work (McKenzie 2013) but very few of these
commuters are youths.
11
economic area or region) economic conditions are larger than the estimated effects of county
economic conditions when each is considered alonethe estimated effects of state economic
conditions are fully inclusive of spillover effects across counties within a state, which this anal-
ysis has demonstrated to be quite important. This is also evident for age- and cause-specific
Column 2 of Table 5 extends the analysis from Column 1 by also considering the effects of
the economic conditions of areas in different states of the same region. While the estimated
effect of regional economic conditions has the same sign as the estimated effects of county
and state economic conditions, it is small in magnitude and not statistically significant. This
sheds light on why the estimated effects of region and state economic conditions tend to be
similar when each is considered alonewhile the effects of regional economic conditions can
fully capture spillover effects across states within a region, these effects are not large enough to
have a meaningful influence on the estimates. The estimates for suicides, shown in Table A3 in
the Appendix along with other mortality outcomes, are something of an outlier but also serve
to demonstrate this general line of reasoning. In particular, the estimated effect of regional
with the rather large difference in the estimated effects of regional and state economic conditions
Column 3 further exploits of the county-level data to consider spillover effects by considering
the effects of the economic conditions of the following areas which are mutually exclusive by
construction: (1) the county itself, (2) counties in the same state and economic area, (3) counties
in the same state but different economic areas, (4) counties in the same economic area but
different states, and (5) counties in the same region but different states and economic areas.
The results from this analysis indicate that economic conditions outside of the state but within
the same economic area have independent effects on mortality while there is weaker evidence
that economic conditions within the state but outside of the economic area have significant
effects on mortality. While this finding could be taken as evidence regarding the mechanisms
through which economic conditions affect health, as it suggests that the effects of economic
conditions on health are unlikely to be strongly related to any state-level policy responses to
changing economic conditions, it is important to note that the estimates are not precise enough
for us to be able to reject the hypothesis that the effects of economic conditions are the same
12
Table 6 explores these issues using data from 19902010. This analysis is motivated by
Ruhm (2013), which consists of state-level analyses that suggest that the link between economic
conditions and total mortality has weakened over timeso much so that estimates using recent
15- and 20-year windows are not statistically significant, raising a question about whether there
In particular, columns 1 through 3 of Table 6 provide estimates of the effects of the economic
conditions of regions, states, and counties, respectively, on county mortality rates as in Table 4.
Consistent with Ruhm (2013), the estimated effect of state economic conditions (Column 2) is
smaller when using data beginning in 1990 than when using data beginning in 1976 (2.46 versus
3.25). Moreover, neither the estimated effect of region nor the estimated effect of state economic
conditions is statistically significant at the five-percent level.19 While the estimated effect of
county economic conditions (Column 3) is smaller in magnitude than the estimated effects of
state economic conditions, it is sufficiently precise to leave little doubt that there remains a
in a county and the economic conditions in surrounding areas, as in Table 5. Column 4, which
simultaneously considers the effects of county and state economic conditions, shows that county
economic conditions have a significant effect on overall mortality that is independent of the
effect of statewide economic conditions. While the corresponding point estimate for statewide
economic conditions is smaller in magnitude and not statistically significant, suggesting a more
limited role for spillover effects on overall mortality in recent years, we cannot reject the hy-
potheses that the independent effects of state and county economic conditions are the same.
Moreover, for the specific causes of death considered in Appendix Table A4 (cardiovascular
problems, motor vehicle, and suicide), the estimated effect of state economic conditions is statis-
ratios have positive and significant (independent) effects on mortality due to cardiovascular
causes and motor vehicle accidents while they have negative and significant effects on mortality
due to suicides, which explains why the estimated effects on overall mortality are somewhat
muted.
The estimate in Column 5, which considers spillover effects across areas in different states
19
Also consistent with Ruhm (2013), the estimated effects on mortality due to cardiovascular causes, motor-
vehicle accidents, and suicides continue to be statistically significant, as shown in Appendix Table A4.
13
in the same region, is too imprecise to reject that these effects are zero but also too imprecise
to reject that they are the same as the effects of state and county economic conditions. The
one exception is for the estimates focused on suicides (in Appendix Table A4), which indicates
that regional economic conditions are more strongly related to suicides than are county or state
economic conditions. This evidence is further supported by Column 6, which considers the
effects of economic conditions of the county itself, counties in the same state and economic
area, counties in the same state but different economic areas, counties in the same economic
area but different states, and counties in the same region but different states and economic
areas.
As a whole, the estimates across tables 3 through 6 (and the associated tables in the Ap-
pendix) demonstrate that the estimated effects of economic conditions on mortality are sensitive
to the level of aggregation of the data, that this sensitivity results primarily from the use of
aggregated measures of economic conditions, and that this sensitivity results from spillover ef-
fects across areas in the effects of economic conditions on health outcomes. In the next section,
I consider these issues in the context of the estimated effects of economic conditions on infant
health.
Panel A of Table 7 follows the identification strategy described in Section 3.2 to estimate
the effects of economic conditions at the time of conception on the fraction of children born
with low birth weight. In particular, the estimates are based on data aggregated to the area
level and models that control for area fixed effects, area-specific trends, year fixed effects, and
demographics, with different definitions of area across columns 1 through 5. Though the esti-
mates are always positive, indicating that downturns are associated with improvements in infant
health, the magnitude of the estimated effects decline as areas are defined more narrowly in a
Panel B of Table 7 begins to shed light on why the estimated effects of economic conditions
on infant health are sensitive to the level of geographic aggregation by considering the effects
of the economic conditions of regions, states, economic areas, component economic areas, and
counties in separate regressions using county-level data and the set of associated county-level
14
control variables (as in Table 4). As with the analysis of mortality outcomes, this approach
brings the estimated effects of economic conditions of regions, states, economic areas, and
component economic areas closer to the estimated effects of county economic conditions and
thus closer to one another, but the estimated effects of county economic conditions remain
Table 8 further explores why the estimated effects of county economic conditions are smaller
than the estimated effects of broader measures of economic conditions by simultaneously con-
sidering the effects of the economic conditions in counties and in their surrounding areas (as
in Table 5). Like the results for mortality, the results of this analysis indicate that health
a manner that can explain why the estimated effects tend to be larger in analyses that focus
panels indicate that infant health is significantly related to the economic conditions of other
counties in the same state. In contrast, these results also indicate that a countys own economic
In terms of the bigger picture question regarding how and why economic conditions are
related to infant health, we must consider the degree to which economic conditions affect the
number and composition of newborns. Towards this end, tables A5 and A6 in the Appendix
present analyses corresponding to tables 7 and 8 but instead focus on birth rates (defined as the
number of births per 10,000 women between the ages of 15 and 44), the fraction of newborns
with a white mother, and the fraction of newborns with a mother younger than 18 years old.
The estimates for birth rates in Table A5 are positive and usually significant, indicating that
fertility is procyclical, while the estimates in Table A6 suggest that we cannot rule out the
possibility that fertility is affected equally by local economic conditions and broader economic
conditions.20 As such, the estimated effects on infant health are difficult to interpret because
they may be influenced by non-random selection. Indeed, the estimated effects on the fraction
of children born to white mothers and to young mothers are sometimes statistically significant
and are often too imprecise to rule out economically significant effects in one direction or the
other.
20
Schaller (2012) also finds that fertility is procyclical.
15
6 Conclusion
Because time-series analyses of health outcomes are subject to such a large suite of potential
biases, approaches that make use geographically disaggregated data are a natural alternative
to understanding the effects of macroeconomic conditions. This paper shows that the level of
(dis)aggregation has important consequences for the estimated effects of economic conditions
in panel data models that exploit variation in the timing and severity of contractions and
expansions across areas using the the now-common area approach. Most significantly, the
estimated effects based on county-level variation in economic conditions are significantly smaller
in magnitude than those based more-aggregate variation. Analyses using county-level data to
simultaneously consider the effects of the economic conditions in a county and the effects of the
economic conditions in surrounding areas indicate that this occurs because there are significant
spillover effects of economic conditions on health outcomes across counties. As such, estimates
based on more aggregate analyses may be more reliable in the sense that they can fully capture
such effects. That said, analyses using more-disaggregated data can offer increased precision and
the ability to consider such spillover effects directly. Having demonstrated that such spillover
effects are empirically relevant, understanding how and why they occur would appear to be a
16
References
Angrist, J., and A. Krueger (1999): Empirical Strategies in Labor Economics, in Hand-
book of Labor Economics, ed. by O. Ashenfelter, and D. Card. North Holland.
Bartik, T. J. (1996): The Distributional Effects of Local Labor Demand and Industrial Mix:
Estimates Using Individual Panel Data, Journal of Urban Economics, 40(2), 150178.
Bound, J., and H. J. Holzer (2000): Demand Shifts, Population Adjustments, and Labor
Market Outcomes during the 1980s, Journal of Labor Economics, 18(1), 2054.
Brenner, M. (1973): Mental Illness and the Economy. Cambridge, MA: Harvard University
Press.
(1979): Mortality and the National Economy, The Lancet, 314(8142), 568573.
Currie, J., and E. Tekin (2011): Is The Foreclosure Crisis Making Us Sick?, NBER
Working Paper No. 17310.
Dehejia, R., and A. Lleras-Muney (2004): Booms, Busts, and Babies Health, The
Quarterly Journal of Economics, 119(3), 10911130.
Forbes, J. F., and A. McGregor (1984): Unemployment and Mortality in Post-War Scot-
land, Journal of Health Economics, 3(3), 239257.
Gerdtham, U.-G., and M. Johannesson (2005): Business Cycles and Mortality: Results
from Swedish Microdata, Social Science and Medicine, 60(1), 205218.
Gerdtham, U.-G., and C. Ruhm (2006): Deaths Rise in Good Economic Times: Evidence
from the OECD, Economics and Human Biology, 4(3), 298316.
Glaeser, E. L., and J. Gyourko (2005): Urban Decline and Durable Housing, Journal of
Political Economy, 113(2), 345375.
Griliches, Z., and J. Hausman (1986): Errors in Variables in Panel Data, Journal of
Econometrics, 33(1), 93118.
Halliday, T. J. (2007): Business cycles, migration and health, Social Science & Medicine,
(64), 14201424.
Hoynes, H. W. (2000): Local Labor Markets And Welfare Spells: Do Demand Conditions
Matter?, The Review of Economics and Statistics, 82(3), 351368.
17
Johansson, E. (2004): A Note on the Impact of Hours Worked on Mortality in OECD
Countries, The European Journal of Health Economics, 5(4), pp. 335340.
Joyce, T. J., and H. N. Mocan (1993): Unemployment and Infant Health: Time-Series
Evidence from the State of Tennessee, Journal of Human Resources, 28(1), 185203.
Lin, S.-J. (2009): Economic Fluctuations and Health Outcome: A Panel Analysis of Asia-
Pacific Countries, Applied Economics, 41(4), 519530.
Miller, D. L., M. E. Page, A. H. Stevens, and M. Filipski (2009): Why Are Recessions
Good for Your Health?, American Economic Review, 99(2), 122127.
Neumayer, E. (2004): Recessions Lower (Some) Mortality Rates, Social Science and
Medicine, 58(6), 10371047.
Notowidigdo, M. J. (2011): The Incidence of Local Labor Demand Shocks, NBER Working
Paper No. 17167.
Ruhm, C. J. (2000): Are Recessions Good For Your Health?, The Quarterly Journal of
Economics, 115(2), 617650.
(2003): Good times make you sick, Journal of Health Economics, 22(4), 637658.
(2005): Healthy living in hard times, Journal of Health Economics, 24(2), 341363.
(2007): A Healthy Economy Can Break Your Heart, Demography, 44(4), 829848.
Saks, R. E., and A. Wozniak (2011): Labor Reallocation over the Business Cycle: New
Evidence from Internal Migration, Journal of Labor Economics, 29(4), 697 739.
Schaller, J. (2012): Boom, Busts, and Fertility: Testing the Becker Model Using Gender-
Specific Labor Demand, University of Arizona Working Paper.
Tapia Granados, J. (2005): Recessions and Mortality in Spain, 1980-1997, European Jour-
nal of Population, 21(4), 393422.
Wozniak, A. (2010): Are College Graduates More Responsive to Distant Labor Market
Opportunities?, Journal of Human Resources, 45(4), 944970.
18
Table 1
State-level Estimates of the Effects of Economic Conditions on Overall Mortality (Per 100,000 Residents)
Stevens Using BLS Using Controls Adding Adding Using County Not Taking Alt. Measure
Ruhm et al. Unemployment Available Earlier Later Matched Log of of Economic
(2000) (2011) Data For Counties Years Years Data Mortality Rate Conditions
Years Analyzed: 1972-1991 1978-2006 1978-2006 1978-2006 1976-2006 1976-2010 1976-2010 1976-2010 1976-2010
(1) (2) (3) (4) (5) (6) (7) (8) (9)
19
Unemployment Rate -0.0054*** -0.0033*** -0.0039*** -0.0040*** -0.0049*** -0.0044*** -0.0045*** -3.6979***
(0.0010) (0.0010) (0.0010) (0.0010) (0.0010) (0.0013) (0.0013) (0.9378)
Observations 930 1479 1479 1479 1581 1785 1680 1680 1680
Notes: See Section 3.1 for details on the data and measures used in each column. All regressions include year fixed effects, state fixed effects, and state-specific trends. In
addition, all regressions are weighted by the population counts in each cell and cluster standard-error estimates on the state.
* significant at 10%; ** significant at 5%; *** significant at 1%
Table 2
State-level Estimates of the Effect of Economic Conditions at the Time of Conception on Fraction With Low Birth Weight
Dehejia & Impute Including Including Collapsing Adding Adding Using County Alt. Measure
Lleras-Muney Conception All All Ages Race-Level Mother Age Recent Matched of Economic
(2004) Year Races of Mothers Data Controls Years Data Conditions
Years Analyzed: 1976-1999 1976-1999 1976-1999 1976-1999 1976-1999 1976-1999 1976-2010 1976-2010 1976-2010
(1) (2) (3) (4) (5) (6) (7) (8) (9)
Unemployment Rate -0.0180*** -0.0171* -0.0167** -0.0157* -0.0144* -0.0177** -0.0320*** -0.0325***
20
(0.0063) (0.0088) (0.0080) (0.0079) (0.0072) (0.0072) (0.0089) (0.0093)
Observations 2447 2448 3672 3672 1224 1224 1785 1680 1680
Notes: See section 3.2 for details on the data and measures used in each column. All regressions include year fixed effects, state fixed effects, and state-specific trends. In
addition, all regressions are weighted by the number of births in each cell and cluster standard-error estimates on the state.
* significant at 10%; ** significant at 5%; *** significant at 1%
Table 3
Estimated Effects of Economic Conditions on Mortality (Per 100,000 Residents)
Using Different Definitions of Areas To Aggregate Data
Notes: Data aggregated to the area level, spanning 19762010, are based on employment data produced by the BEA,
population data produced by SEER, and death counts from restricted-use Vital Statistics micro-record multiple-cause-
of-death files provided by the NCHS. All estimates control for year fixed effects, area fixed effects, area-specific trends,
and demographics, and are weighted by the population in each cell. Standard-error estimates shown in parentheses are
clustered on the area defined by the level of geographic aggregation, except for the county-level analyses where standard-
error estimates are more conservative when clustered at the state level. Two-way standard-error estimates that additionally
allow errors to be correlated across areas within a given year are shown in brackets.
* significant at 10%; ** significant at 5%; *** significant at 1%
21
Table 4
Estimated Effects of Economic Conditions on Mortality (Per 100,000 Residents)
Using County-Level Data and Aggregate Measures of Economic Conditions
Notes: Data spanning 19762010 are at the county level and are based on employment data produced by the BEA, popu-
lation data produced by SEER, and death counts from restricted-use Vital Statistics micro-record multiple-cause-of-death
files provided by the NCHS. All estimates control for year fixed effects, county fixed effects, county-specific trends, and
county demographics, and are weighted by the population in each cell. Standard-error estimates shown in parentheses
allow for two-way clustering (state and year).
22
Table 5
Estimated Effects of Economic Conditions on Mortality (Per 100,000 Residents) Using County-Level
Data to Simultaneously Consider The Effects of Local and Broader Measures of Economic Conditions
23
Table 6
Estimated Effects of Economic Conditions on Mortality (Per 100,000 Residents)
Using County-Level Data Post-1990
24
Table 7
Estimated Effects of of Economic Conditions on at the Time of Conception on Fraction of Newborns
With Low Birthweight
Notes: Data spanning 19762010 are based on employment data produced by the BEA, population data produced by
SEER, and birth information from restricted-use natality files provided by the NCHS. The estimates in Panel A are based
on data aggregated to the area level indicated by the column heading and control for year fixed effects, area fixed effects,
area-specific trends, and demographics at that level. The estimates in Panel B use county-level data and consider the
effects of aggregate measures of economic conditions (at the level indicated by the column heading) while controlling for
year fixed effects, county fixed effects, county-specific trends, and county demographics. All estimates and are weighted
by the population in each cell. In Panel A the standard-error estimates shown in parentheses allow for two-way clustering
(area and year), except for the county-level analyses where standard-error estimates are more conservative when clustered
at the state level. In Panel B the standard-error estimates shown in parentheses also allow for two-way clustering (state
and year).
* significant at 10%; ** significant at 5%; *** significant at 1%
25
Table 8
Estimated Effects of of Economic Conditions on at the Time of Conception on Fraction of Newborns
With Low Birthweight Using Data at County Level And Simultaneously Considering
The Effects of Local and Broader Measures of Economic Conditions
Notes: Data spanning 19762010 are at the county level and are based on employment data produced by the BEA,
population data produced by SEER, and birth information from restricted-use natality files provided by the NCHS. All
estimates control for year fixed effects, county fixed effects, and county-specific trends, and demographics of mothers. All
estimates are weighted by the population in each cell. Standard-error estimates shown in parentheses allow for two-way
clustering (state and year).
* significant at 10%; ** significant at 5%; *** significant at 1%
26
Appendix - Additional Tables
Table A1
Estimated Effects of Economic Conditions on Mortality (Per 100,000 Residents)
Using Different Definitions of Areas To Aggregate Data
27
Table A2
Estimated Effects of Economic Conditions on Mortality (Per 100,000 Residents)
Using County-Level Data and Aggregate Measures of Economic Conditions
28
Table A3
Estimated Effects of Economic Conditions on Mortality (Per 100,000 Residents) Using County-Level
Data to Simultaneously Consider The Effects of Local and Broader Measures of Economic Conditions
30
Table A5
Estimated Effects of of Economic Conditions on at the Time of Conception on Fraction of Newborns
With Low Birthweight
Note: This table corresponds to Table 7 for additional outcomes, omitting demographic control variables.
31
Table A6
Estimated Effects of of Economic Conditions on at the Time of Conception on Infant Outcomes
Using Data at County Level And Simultaneously Considering
The Effects of Local and Broader Measures of Economic Conditions
Note: This table corresponds to Table 8 for additional outcomes, omitting demographic control variables.
32