Are Mortality Rates Random Walk? Panel Unit-
root Tests with the Evidences from Micro Data*
University of Notre Dame
April 19, 2013
Abstract
This study investigates the relationship between economic conditions and health. Previous
studies have estimated that they have a significantly negative correlation. However, there is a
concern that mortality rates are non-stationary series, with which a chance of spurious
correlation exhibits. By taking first difference on regressors and conducting unit-root tests for
panel data, it shows some evidences that the unit-root concern is reasonable and the
association between health and economic shocks is exaggerated. An accompanying analysis of
micro data indicates that the influence of risky behaviors on mortality rate in response to
economic conditions cannot match the extent to which the basic fixed-effect model predicted.
1
* I am extremely grateful to Professor William Evans for his constant guidance and support. All errors are mine. ѱ Department of Economics, University of Notre Dame, IN 46556. Email: [email protected]
1. Introduction
A recent research shows a procyclical fluctuation of total mortality (Ruhm 2000) by Fixed-effect
(FE) models of panel data regression with health and economic condition proxied by mortality
rates and unemployment rates respectively. By using the same method with the undated data
for the 1984-2010 periods, I draw a similar conclusion with even large coefficient (absolute
value) on the unemployment rate. The negative correlation of total mortality and
unemployment rates can be partially explained by an evident increase in risky behavior,
including smoking and drinking; and a reduction of physical activities and medical care during
the economic expansion era. However, is that statistically enough to say a strengthening
economy is bad for people’s health?
In light of the concern that two independent random walks may have a statistically
significant correlation, I am curious about whether these two series are non-stationary. Many
literatures have already shown that we do not have to worry about a random walk for some
important economic indicators, including unemployment rates. However, few researches focus
on whether the mortality rates are random walk.
This study uses unit-root tests to investigate the reliability of the negative relationship
between people’s health and economic condition. Differences over one period are taken for
both dependent and independent variable to make the potential un-stationary variables
stationary. A Fisher-Type unit root test for panel data is put into use to explore which variable is
randomly dependent from time.
The analysis provides strong evidence that the negative correlation between health and
economic condition is exaggerated. When taking a one-period difference for both state
mortality and unemployment rate respectively, the regression result shows a more than 80%
shrinking of the coefficient, though it is still significant. Consistent with this result, the Fisher-
Type test shows that the state mortality rate fails to pass the unit-root test, despite that the
unemployment rate is not a random walk.
In addition, micro data from the Behavioral Risk Factor Surveillance System (BRFSS) are
used to provide some evidences for potential problem of unit root by showing the negative
coefficient between mortality and unemployment rates cannot be explained by some related
risky behaviors. Ruhm (2000) indicates that the mortality rates are procyclical by decomposing
the variation in mortality into several components, including tobacco use, drinking, physical
activity, diet and medical care. Furthermore, Sparks, Cooper, Fried, and Shirom (1997) argue
that a rising workload has a negative effect on people’s health status by using meta-analysis to
examine the relationship between the length of the working week and health symptoms. White
and Beswick (2003), and Siegrist and Rodel (2006) draw the same conclusion. Nonetheless, I
show none of these can lead to such a sharp increase in mortality rate in the expansion era with
data analysis and intuitive induction.
Fixed-effect models are estimated using panel data for the smoking rate for 1996-2010
period and alcohol-related mortality for 1984-1992 period. The results reveal that both smoking
and drinking related deaths are significantly increasing during the expansion era. However, the
coefficients are far smaller relative to the total mortality rate. Aguiar, Hurst and Karabarbounis
(2011) explore how foregone market work hours are allocated to other activities over the
business cycle of 2003-2010, with special focus on the era of the great recession. They show
that a 4% increase of the unemployment rate during the 2008-2010 leads to 7% reduction in
working hours for average workers. The result is significant but still too small to alter people’s
health status. In addition, a fairly small amount of the reduced working hours are allocated to
activities that improve people’s health (physical practices, medical care and sleep). As a matter
of fact, a very large fraction of time is used to non-market production and housework, which
consume energy as well.
The paper is structured as follows. In section 2 I use the Fixed-effect model to examine the
relationship between health and economic shock with updated data. In section 3 I use some
methods to examine if the result I got in section 2 can pass the unit-root tests. In section 4 I use
micro data to explore to what extent the risk behaviors which is responsive to the economic
shock can alter people’s health status. Finally in section 5 I make conclusions.
2. Fixed-effect model with state data
2.1 Model
I use the subscripts j and t to index the state and year. The basic regression equation is
(1)
for H the natural log of the mortality rate, E the proxies for economic conditions, X a vector of
supplementary regressors, and ε the error term. The fixed-effect controls for time-invariant
state characteristics, accounts for nationwide time effects and γ captures the impact of
within-state deviations in economic conditions. The effect of national business cycle is also
obtained by estimating
(2)
where indicates national economic conditions and the time effects are excluded.
2.2 Descriptive statistics
Summary statistics, weighted by the total resident population in each state by year, are
displayed in Table I and are largely self-explanatory. The death rate refers to all-caused deaths
per 10000 persons. The state and unemployment rates are the weighted average of period
1984-2010 by state. The per capital personal income is the weighted average of nominal dollar
of the same period.
TABLE I VARIABLES USED IN ANALYSIS OF STATE AGGREGATE DATA
Variable Mean Standard deviation
Death rate per 10000 people 84.63 13.26 State unemployment rate in % 5.97 1.89 National unemployment rate in % 5.96 1.38 Per capita personal income (in thousands) 36.86 10.84 All variables are weighed by state populations. The data of mortality rates are from CDC WONDER. The unemployment rates are provided by Bureau of Labor Statistics, Local Area Unemployment Statistics (LAUS) program. Income data are constructed using the data from U.S. Census Bureau, Current Population Survey, Annual Social and Economic Supplements. Population statistics are provided by Population Division, U.S. Census Bureau.
Figure I displays national total mortality and unemployment rates in each year. The
variables normalized by subtracting mean of variable. The result shows some evidences of
inverse relationship between economic conditions and fatality rates. However, unlike Ruhm
(2000), the negative correlation does not hold for each period. From 1984-1992 and 2007-2010,
the inverse relationship is evident. However, during 1993-2000, and positive correlation is
observed. This problem appears if I drop the some observations of the earliest and latest years.
For instance, if the observations of 1984-1987 and 2007-2010 are dropped, the coefficient on
unemployment rate severely shrinks, though the sample is still large enough.
Figure I Total Mortality and Unemployment Rates (Normalized)
2.3 Result
This subsection examines the relationship between economic conditions and total mortality
rates. Unless otherwise stated, the dependent variable is the natural logarithm of the death
rate (per100,000 persons). Table II summarizes the results of a variety of specifications, all of
which control for state fixed-effects, and use state or national unemployment rates to proxy
macroeconomic conditions. Some specifications also include year dummy variables, and some
hold constant personal incomes.
As Table II shows, the negative relationship between unemployment rate and mortality
rate is even larger than previous predictions. One percentage increase of state unemployment
rates is associated with 1.2 percentage decrease of mortality rates. With the control of personal
income, the result holds. If the year dummy is excluded, a little bit weaker relationship is
observed. This contradicts to the result from Ruhm (2000), which states that the negative
relationship is stronger by dropping the year dummy. If the independent estimator is the
national mortality rate, the relationship is even weaker, but still very significant. One
percentage increase of national unemployment rate leads to 0.6 decline of mortality rate.
The middle and bottom panel of Table II presents separate estimates for the 1984-1997
and 1998-2010 periods, the ten highest and lowest income city based on the income in 2010.
Splitting the sample into shorter time period is likely to reduce the influence of within-state
changes in omitted factors that are correlated with unemployment rate. The rationale for
singling the states with highest and lowest income is that if the negative heath effect of
economic expansion is due to migration, a difference may be observed for those to
specifications.
Mortality is procyclical for all variations. However by splitting the period, coefficients
decline for about 30 percent. Both highest income and lowest income have larger negative
health effects. But the two specifications exhibit similar results.
TABLE II FIXED-EFFECT ESTIMATES OF THE DETERMINANTS OF TOTAL MORTALITY RATE
Full sample estimates
Basic specification
(a) (b) (c) (d) (e)
State unemployment rate -0.0120 -0.0124 -0.0083 U.S. unemployment rate
(0.0011) (0.0011) (0.0007) -0.0057 (0.0009)
-0.0066 (0.0009)
Personal income -0.0498 (0.0219)
-0.0289 (0.0046)
Year effect Yes Yes No No No
Split-sample estimates
1984-1997 1998-2010
(a) (b) (a) (b)
State unemployment rate -0.0081 -0.0085 -0.0073 -0.0043 (0.0007) (0.0009) (0.0007) (0.0007) Personal income -0.0068 -0.1262 (0.0068) (0.0112)
Split-sample estimates
10 highest income states 10 lowest income states
(a) (b) (a) (b)
State unemployment rate -0.0192 -0.0189 -0.0189 -0.0212 (0.0034) (0.0033) (0.0036) (0.0037) Personal income -0.1751 -0.1450 (0.05492) (0.0617)
The dependent variable is the natural logarithm of the total mortality rate per 10,000 populations. All specifications also include vectors of state dummy variables. Year dummy variables are also controlled for, except models (c), (d) and (e) in the top panel. Standard errors are in parentheses. The sample in the top panel includes annual observations for the 50 states and the District of Columbia covering the period 1984–2010. The ten largest states refer to rankings in 1991 and include, in descending order of size, California, New York, Texas, Florida, Pennsylvania, Illinois, Ohio, Michigan, New Jersey, and North Carolina.
3. Unit-root test
In light of concern that two independent random walks may have a statistically significant
correlation, I use two methods to check the unit-roots on both dependent and independent
variables. Many literatures (Romero-Ávila and Usabiaga, 2007; Schwert 1987) have already
shown that we do not have to worry about a random walk for some important economic
indicators, including unemployment rates. However, few researches focus on whether the
mortality rates are random walk. The rationale is that for individuals, we can separate the
causes of death to two classifications. One is the internal causes of death and the other is
external causes of death. The latter one is triggered by some risky behaviors associated with
the change of economic condition, which likely to be cyclical. However, independent from
unhealthy behaviors, the time of death is very random. The amount of fatalities possibly is
heterogeneous across years. That gives another potential cyclical of mortality rate which barely
has a correlation with economic shocks. If the internal causes are dominant, we probably can
observe a negative relationship between mortality and unemployment rate. However,
economic expansion and recession cannot be applied to explain the correlation. In turn, the
results I got in Section 2 are exaggerated.
3.1 Model
3.1.1 Difference on difference test
I use the subscripts j and t to index the state and year. The difference on difference regression
equation is
(3)
for is the first difference of natural log of the mortality rate, is the proxies of the first
difference of economic conditions, X a vector of supplementary regressors, and ε the error term,
accounts for nationwide time effects.
By taking the first difference between variable and one period lagged variable, I convert the
non-stationary series into stationary, by which the spurious regression can be avoided.
3.1.2 Fisher-type test
To investigate the causal relationship between mortality and unemployment rate, I further
check for the stationarity of each variable separately. Recently, there has been a heightened
development of panel-based unit root tests (Hadri, 1999; Breitung, 2000; Choi, 2001; Levin, Lin
and Chu, 2002; Im, 2003; Breitung and Das, 2005). These studies have shown that the panel
unit root tests are more powerful than tests based on times series data.
I use the subscripts j and t to index the state and year. The fisher-type ADF specification
regression equation can be written as
(4)
(5)
for is the first difference of natural log of the mortality rate, is the proxies of the
one period lagged economic conditions, X a vector of supplementary regressors, and ε the error
term.
In testing for panel-data unit roots, Fisher-type tests conduct the unit-root tests for each
panel individually and then combine the p-values from these tests to produce an overall test. In
this context, I perform a unit-root test on each of panel units j (state) separately and then I use
their combined p-values to construct a Fisher-type test to investigate whether or not the series
exhibit a unit-root. The null hypothesis is that all the panels have unit roots and the alternative
hypothesis is that at least one panel does not have unit roots or some panels do not have unit
roots. This routine provides 4 different unit-root test methods as proposed by Choi (2001).
3.2 Result
3.2.1 Difference on difference test
The way in which I interpret the Table III is analogous to the Section 2 while the numbers are
significantly different. By difference on difference test, one percentage increase of state
unemployment rate is associated to 0.26% decline of mortality rate. The coefficient is still
significant but has shrunk about 5/6 compared to the basic regression. Adding year dummy or
not, controlling personal income or not and estimating by state or national economic predictors
give similar observations.
TABLE III FIRST DIFFERENCE TEST OF DETERMINANTS OF MORTALITY RATE
Full sample estimates
Basic specification
(a) (b) (c) (d) (e)
State unemployment rate -0.0026 -0.0024 -0.0035 U.S. unemployment rate
(0.0009) (0.0011) (0.0006) -0.0037 (0.0006)
-0.0027 (0.0007)
Personal income -0.0186 (0.0090)
-0.0102 (0.0024)
Year effect Yes Yes No No No
The dependent variable is the natural logarithm of the total mortality rate per 10,000 populations. State dummy here is excluded for avoiding bias. Year dummy variables are controlled for, except models (c), (d) and (e). Standard errors are in parentheses. The sample in the top panel includes annual observations for the 50 states and the District of Columbia covering the period 1985–2010.
The test draws a conclusion that by converting the non-stationary series into stationary,
the updated results show evidences that the basic regression exaggerates the correlation
between mortality rate and unemployment rate and concerns about a spurious regression is
reasonable. Next I want to use the panel date unit-roots test to show which variable is possibly
non-stationary.
3.2.2 Fisher-type test
The Table IV shows the result of fisher-type panel data unit-root test for mortality rate and
unemployment rate for 1984-2010 period. By observing the p-value, I can reject the null
hypothesis that all panels of unemployment rate have unit root while the total mortality rate
does not pass the test.
The result is consistent with the conclusion I draw on previous subsection. Since very likely
the mortality rates have unit roots, its correlation with unemployment rates cannot be
identified by simply run a fixed-effect regression. A process of taking first difference is not long
reasonable, but also required.
TABLE IV FISHER-TYPE UNIT-ROOT TEST ON PANEL DATA WITH ONE PERIOD LAG
All mortality Unemployment Drinking mortality
Test statistic Statistic P-value Statistic P-value Statistic P-value
Inverse chi-squared P 88.75 0.82 302.36 0 134.01 0.02 Inverse normal Z 2.72 0.99 -10.55 0 -0.02 0.49 Inverse logit t L* 2.74 0.99 -11.04 0 -0.17 0.43 Modified inv. chi-squared Pm -0.92 0.82 14.03 0 2.24 0.01
3.3 Discussion
As Figure II shows, the red line represents the unemployment rate; dot line indicates the
external mortality rate, which is responsive to the economic shocks; dashed line shows the
internal mortality rate, which is determined by people’s own health status rather than risky
behavior; and cross line exhibits the aggregate mortality rate, which is an additive combination
of internal and external mortality rate.
Figure II Schematic Diagram of Separation of Mortality Rates
The reason why I cast doubt on the method given by Ruhm (2000) is that he actually
regresses the red line (unemployment rate) on the cross line (aggregate mortality rate).
However, only external mortality is responsive to the economic shocks while the internal one is
probably a random cycle. Therefore, an exaggerated result is inevitable because some non-
economic reason is mistakenly attributed.
My fisher-type test shows the aggregate mortality rate is a random walk, which implies
that the internal death-caused reasons are dominant. The difference on difference test
indicates a significant but small correlation between health and economic condition. That
means people’s health status is still associated with economic condition but is severely
exaggerated due to the random property of internal mortality.
4. Evidences from Micro Data
4.1 Background
In this subsection, I want to introduce where my curiosity of applying unit root test on mortality
rate comes from. This can be identified by the intuition when I was observing the micro data.
By running the fixed-effect estimates of the determinants of total mortality rate, Ruhm
(2000) provides an evidence of a negative correlation between unemployment rate and
mortality rate. In other words, mortality rate is procyclical, which seems contradicts to our
common sense. He thereafter tries to explain why the procyclical phenomenon makes sense by
using the Micro data from Behavioral Risk Factor Surveilance System. Some leading behavioral
responses to the change of economic condition are smoking and drinking, physical activity, diet
and preventable medical care. Mortality rate is likely to be sensitive to these factors.
TABLE V HOW RISKY BEHAVIORS REPOND TO STATE MORTALITY RATE
Smoker Drinker Any physical activity
Daily gram of fat consumes
Medical checkup
State unemployment rate .0064 (.0021) Source: Ruhm. “Are regressions good for your health”, Quarterly Journal of Economics, 2000
As the Table V shown, a one percentage point increase in the state unemployment rate
reduces the predicted number of current smokers by 0.3 percentage point. Physical activities,
diet and medical care are significantly improved during the recession era.
It is true that some behavioral responses to the change of economic condition significantly
influence people’s health status. However, obviously none of these factors leads people’s death
immediately except drunk driving, which is not statistically significant when varying
unemployment rate. For instance, one can drastically increase his/her tobacco use during a
good economic state. But he/she is almost unlike to die during the period of one business state.
Especially when taking the age factor into account, this intuition is more reasonable. Apparently,
Young people’s behaviors are more sensitive to the economic condition. But only after decades
their mortalities can be recorded into data. Their life expectancies were shortened by the
unhealthy behaviors during good states. However to what extent one’s life expectancy is
altered is highly heterogeneous and unable to be measured. Possibly, in a random year, there is
an explosion of mortality rate. However, this is irrelevant to the current economic condition but
related to a collective behavior changes due to good economic condition a few decades ago.
Thus, it reminds me that if a random walk can explain the change of mortality rate because of
the heterogeneous effects of behavioral change on people’s life expectancy.
4.2 Model
Econometric estimates of the determinants of smoking and drinking are summarized in Table
VII. The basic econometric specification is
(6)
where i indicates the individual, H the health input, and other variables are as defined above.
The linear probability model can estimate how state smoking rates respond to the economic
shocks. For drinking-related mortality, I use the same model as Section 2 shows.
4.3 Descriptive Statistics
Summary statistics, weighted by the total resident population in each state by year, are
displayed in Table VI. The drinking-related mortality rate refers to alcohol-caused deaths per
100000 persons. The state and unemployment rates are the weighted average of period (a)
1996-2010, (b) 1984-1992, and (c) 1996-2000. The per capital personal income is the weighted
average of nominal dollar of the same period.
TABLE VI VARIABLES USED IN ANALYSIS OF INDIVIDUAL DATA
Variables Mean Standard deviation
Smoking rate 20.99 3.65 Drinking-related mortality per 10000 people (1) 43.73 8.06 Drinking-related mortality per 10000 people (2) 35.56 8.17 Tobacco use 18.19 1.49 Per capita personal income (in thousands) (a) 43.14 6.44 Per capita personal income (in thousands) (b) 26.32 5.46 Per capita personal income (in thousands) (c) 37.53 6.4 State unemployment rate in % (a) 5.67 1.86 State unemployment rate in % (b) 6.54 1.94 State unemployment rate in % (c) 4.63 1.18 National unemployment rate in % (a) 5.60 1.62 National unemployment rate in % (b) 6.51 0.83 National unemployment rate in % (c) 4.60 0.50 All variables are weighed by state populations. The data of drinking mortality rates are from Stinson, F.S., Nephew, T.M., Dufour, M.C., and Grant, B.F. 1996. U.S. Alcohol Epidemiologic Data Reference Manual. The data of smoking rate and tobacco use are provided by T Behavioral Risk Factor Surveillance System Survey Data (BRFSS) from the period of 1996-2010 and 1996-2000 respectively. The unemployment rates are provided by Bureau of Labor Statistics, Local Area Unemployment Statistics (LAUS) program. Income data are constructed using the data from U.S. Census Bureau, Current Population Survey, Annual Social and Economic Supplements. Population statistics are provided by Population Division, U.S. Census Bureau.
4.4 Reexamination of risky behaviors
A. Smoking and drinking
Smoking rate has a significantly negative relationship with state unemployment rate. One
percentage increases in unemployment rate leads to 0.16 percentage decrease in smoking rate.
However, does it enough to show that the increasing consumption of smoking can largely
explain the deteriorated health condition during economic expansion?
To show this, I run another regression to find the relationship between smoking rate and
mortality rate. The result exhibits that the smoking rate and mortality rate has a significant
positive relationship and one percentage rise in smoking rate is associated with 0.18
percentage increase in mortality rate. Then by using simple algebra I can conclude that if
unemployment rate goes up by 1%, the mortality rate driven solely by rising tobacco
consumption decreases about 0.03% (0.16*0.18). If compared with 1.2% decrease in total
mortality rate, this number is quite small.
What about drinking? By intuition we know that the death caused by drinking during
economic expansion is more noticeable than smoking. This is because of drunk driving. During
1984-1992, the annual alcohol-related driving fatalities are 22500 on average. Drunk driving,
unlike smoking even binge drinking, probably leads to death in next minute. Thus, even though
Ruhm (2000) argues that the drinking rate is not very procyclical, some treatment is still needed.
The regression result shows that one percentage increase in state unemployment rate
leads to 1.7% decrease in drinking death rate. The result is significant and quite big. In addition,
as the Figure III shows, during the period of 1984-1992, the alcohol-related death shares a same
pattern as total mortality rate, both of which is procyclical. Nevertheless I still have to do some
adjustment on the number. By calculation I get among all people who died during 1984-1992, 5%
is related to drinking. This means that if unemployment rate rises by 1%, the mortality rate
driven solely by drinking-rated reason decreases about 0.08% (1.7*0.05). First of all, this
number is consistent with my previous guess that drinking-rated death is more than smoking-
rated death triggered by economic reason. Moreover, 0.08% is still very small relative to the 1.2%
in total. If we add the coefficients on smoking and drinking up, 0.11% (0.03+0.08) can only
explain 10% of decrease in total mortality rate caused by good economic condition. Besides, to
find the relationship between unemployment rate and mortality rate, I use the data from 1984-
1992. However, the annual death rate caused by drunk driving has decreased over 70%. This
means that 0.08% is actually overestimated because about 29% of drinking-related death can
be explained by drunk driving.
TABLE VII FIXED-EFFECT ESTIMATES OF THE DETERMINANTS OF RISKY BEHAVIORS
Smoking rate Mortality rate
(a) (b) (a) (b)
State unemployment rate -0.1604 -0.1767 Smoking rate 0.0018 0.0018 (0.0534) (0.0541) (0.0009) (0.0009) Personal income -1.6125 Personal income -0.0250 (0.9048) (0.0214)
Drinking death rate 1 Drinking death rate 2
(a) (b) (a) (b)
State unemployment rate -0.0169 -0.0169 -0.0158 -0.0157 (0.0018) (0.0018) (0.0020) (-0.0020) Personal income 0.0000
(0.0398) -0.1447
(0.0442)
Cigarettes per day Mortality rate
(a) (b) (a) (b)
State unemployment rate -0.1810 -0.1808 Cigarettes per day 0.0002 0.0002 (0.1407) (0.1410) (0.0016) (0.0016) Personal income 0.2751
(1.1726) Personal income -0.0137
(0.0259)
All specifications include vectors of year and state dummy variables. Robust standard errors, estimated assuming observations are independent across years and states but not within states in a given year, are displayed in parentheses.
Figure III Comparison between Total Mortality Rate and Alcohol-Related Death
B. Workload
If the behavioral change on smoking or drinking patterns cannot explain the large negative
figure, what about stress? An increase of working hours during economic expansion exerts
more responsibilities for workers. This in turn deteriorates people’s health. Conversely, during
recessions, people tend to have less workload, which subsequently reduce their stress. In
practice, Sparks, Cooper, Fried, and Shirom (1997) argue that a rising workload has a negative
effect on people’s health status by using meta-analysis to examine the relationship between
the length of the working week and health symptoms. White and Beswick (2003) and Siegrist
and Rodel (2006) draw the same conclusion.
I want to cast doubt on this argument by three insights. First, intuitively speaking, both
internal and external stress may influence people’s health. Working hours are external stress.
Conceivably this exerts negative effects on people’s health. How about internal stress? When
people are losing job, they meanwhile lose their steady income for living. It apparently makes
people stressful. No one feels happy if they do not know how to fund for the food tomorrow.
People in that case tend to be self-motivated. They start searching for job actively. This process
does take time and energy. However, during bad era, finding a job is considerably difficult. It
not only increases their workload of preparing for processes, but also discourages them. When
they are so discouraged that forgoing the efforts, they will not be seen as job seekers.
Unemployment rate in this case does not count them.
Furthermore, indeed, if people have to work beyond the typical 8 hours per day, extra
workload does make people stressful. However, a modest workload could help people arrange
their daily routine regularly. Especially for young people, unemployment does not mean that
they can have a good rest. In contrast, their schedule might be entirely disorganized, which
threatening their health. Thus, increasing working hours have an effect in combination of extra
responsibility and good schedule. Which one is more influential is mysterious. However,
obviously they may cancel each other out.
Besides, let us consider some related empirical result. Aguiar, Hurst and Karabarbounis
(2011) explore how foregone market work hours are allocated to other activities over the
business cycle of 2003-2010. We focus on the most recent great recession. During the period of
2008-2010, the aggregate unemployment rate was rising from around 5.8% to around 9.6%.
The data shows that the market work hours fell by 7%. This is consistent with the statistics
given by the Bureau of Labor Statistics (BLS). A simple calculation shows that one percentage
increase of unemployment rate is associated with a 1.84% decrease in working hours. This
means that a representative worker who is working about 40 hours per week has extra 44
minutes leisure hours during a recession when the unemployment rate increases by 1 percent. I
do not have empirical result to show how this 44-minute difference per week alters people’s
health condition. However, this is obviously a small number in practice.
Moreover, this number could be shrinking if we take how people allocate the extra hours
during a recession into account. Aguiar, Hurst and Karabarbounis (2011) shows that roughly 30%
of the foregone market working hours are reallocated to non-market production (excluding
child care), about 32% of foregone work hours are allocated to some regular housework,
including cooking, cleaning, shopping, and child care. People cannot feel much more relaxed if a
very large fraction of forgone working time is allocated like this. Comparably speaking, sleeping
only account for 20% and health care and civic activities is even less.
C. Diet, physical activities and medical care
Among all behavioral changes in response to the economic shocks, I choose smoking and
drinking to do further research. This is because a drastic increase in consumption of them may
significantly deteriorate people’s health. Diet, for instance, do help people keep good health
indicators (blood pressure, blood sugar, etc.). However, they are long-term determinants of
people’s life expectancy. In a short-run, they are obviously less important since my dependent
variable is mortality rate. Medical checkup, likewise, do not directly improve or jeopardize our
health.
Furthermore, my research focused on the relationship between mortality rate and
economic condition. Therefore, I want to single out some behaviors driven by economic factors
which are subject to change during business cycle. In other words, if people can find some free
alternatives during a bad state, for example running around your garden instead of going to
gym, the correlation would be weakened even though a significant coefficient was produced.
That’s why I excluded the physical activities because people are easily accessible to free
alternatives.
5. Conclusion
This study casts doubt on the strong relationship between macroeconomic conditions and
health. Previous and my studies show a one percentage point rise in the state unemployment
rate is associated with 0.6-1.2% percent decrease in total mortality by fixed-effect model.
However, when converting the dependent and independent variables into stationary series, the
correlation shrinks about 50-80%. A fisher-type panel unit-root tests reveal that this is probably
due to the mortality rates are essentially random walk.
Micro data of smoking rates and alcohol-related death give some evidences on the unit-
root concern. Even though the linear probability and fixed-effect models prove that the risky
behaviors lead to more fatalities in expansion era, they are too small to explain such a negative
relationship. Stress may be another assumption to interpret the procyclical nature of the
mortality rate. However, some related researched have already shown the average saving
working hours are not influential enough to alter people‘s health status. Even if people have
less working responsibilities in recession era, they tend to allocate a large fraction of the saving
hours to non-market production and housework, which consume energy as well.
Are recessions good for your health? Maybe the answer appears to be yes. However, the
effect is far less than previously predicted. Further research may find more evidences by
focusing more on age distribution and some specific causes of death.
Reference
[1] Aguiar, Hurst, Karabarbounis (2011) “Time use during recessions.”
[2] Breitung J (2000) “The local power of some unit root tests for panel data.” In B.Baltagi (ed.),
Nonstationary Panels, Panel Cointegration, and Dynamic Panels, Advances in Econometrics, 15.
JAI: Amsterdam, 161-178
[3] Berniell (2012) “The effects of working hours on health status and health behaviors.”
[4] Cho, Cooley (1994), “Employment and hours over the business cycle”, Journal of Economic
Dynamics and Control 18 (1994) 411-432.
[5] Choi In (2001) “Unit root tests for panel data.” Journal of International Money and Finance
20: 249 – 272.
[6] Diego Romero-Ávila and Carlos Usabiaga (2007) “Unit Root Tests, Persistence, and the
Unemployment Rate of the U.S. States,” Southern Economic Journal, 698-716
[7] Levin A, Lin CF Lin, Chu J (2002) “Unit root tests in panel data: asymptotic and finite-sample
properties.” Journal of Econometrics 108: 1-24.
[8] Im KS, Pesaran MH, Shin Y (2003) “Testing for unit roots in heterogeneous panels.” Journal
of Econometrics, 115: 53-74.
[9] Maddala GS, Wu S (1999) “A comparative study of unit root tests with panel data and a new
simple test.” Oxford Bulletin of Economics and Statistics 61: 631-652.
[10] Sparks, K., C. Cooper, Y. Fried, and A. Shirom (1997): "The effects of hours of work on
health: A meta-analytic review," Journal of Occupational and Organizational Psychology, 70(4),
391-408.
[11] Siegrist, J., and A. Rodel (2006): "Work stress and health risk behavior," Scandinavian
journal of work, environment & health, 32(6), 473.
[12] Ruhm, C. (2000): "Are Recessions Good for Your Health?" Quarterly Journal of Economics,
115(2), 617-650.
[13] Ruhm, C (2003): “Good times make you sick," Journal of Health Economics, 22(4), 637-658.
[14] Ruhm, C (2005): "Healthy living in hard times," Journal of Health Economics, 24(2), 341-363.
[15] Schwert, G (1987) “Effect of model specification on tests for unit roots in macroeconomic
data”, Journal of Monetary Economics 20 (1987) 73-103
[16] White, J., and J. Beswick (2003): "Working long hours,"