Employment effects of the Olympic Games in Atlanta 1996 reconsidered.
Feddersen, Arne ; Maennig, Wolfgang
Introduction
In their SEJ (2003) contribution, Hotchkiss, Moore, and Zobay (HMZ)
found significant positive employment effects of a major sporting event,
namely, the 1996 Olympic Games in Atlanta. Their contribution is notable
because it is one of the very few econometric ex-post studies that have
found such positive effects. Research on the economic impact of
professional sport franchises, sport facilities, and sport events have
been performed for more than 20 years--starting with the studies by
Baade (1987) and Baade and Dye (1988)--and the results of this
literature are strikingly consistent. Studies of this nature have always
come to the same conclusion, no matter what geographical units (e.g.,
cities, counties, metropolitan statistical areas, or states) are
examined, no matter what model specification, estimation method, or
dependent variables (e.g., employment, wages, or taxable sales) are
used, and no matter which part of the world is considered (e.g., USA or
Europe); historically, these scholarly analyses contain almost no
evidence that professional franchises, sport facilities, or mega-events
have a measurable impact on the economy. (1) Other studies, particularly
those by Coates and Humphreys (1999, 2001, 2003) and Teigland (1999),
have even indicated significant negative effects. Besides the original
HMZ study, only some few positive exceptions exist including Jasmand and
Maennig (2008) for the Olympic Summer Games in Munich 1972, who find
significant long-term income effects, but exclusively for selected
periods of analysis. Baumann, Engelhardt, and Matheson (2012) analyze
the Salt Lake 2002 Winter Olympics and find a small effect of some
additional 4,000-7,000 jobs, concentrated in the leisure industry, but
little to no effect on employment after 12 months. Rose and Spiegel
(2011) found a 20-40% growth of exports for Olympic host countries, but
Maennig and Richter (2012) demonstrate that this result is due to a
comparison of non-matching countries. Loosely connected to mega-events,
Tu (2005) found significant positive effects from the FedEx Field
(Washington) on real estate prices in the surrounding neighbourhood, as
did Ahlfeldt and Maennig (2008) for three arenas in Berlin, Germany.
Finally, Carlino and Coulson (2004) examined the 60 largest Metropolitan
Statistical Areas (MSA) in the USA and found that having a National
Football League (NFL) team allowed the cities to enjoy rents that were
8% higher but not higher wages. (2)
To test whether the Olympic Games 1996 in Atlanta affected the
(regional) economy in Georgia, HMZ used difference-in-difference (DD)
estimation to compare the differences in outcome before and after an
intervention for groups affected and unaffected by the intervention
(Bertrand, Duflo, & Mullainathan, 2004). More precisely, they used
two variants of the DD approach to analyze (1) a pure level shift
("DD in the intercept") and (2) a pure trend shift ("DD
in the slope") initiated by the Olympics. As a key result, which is
prominently highlighted in their abstract and conclusion, they found a
tremendous boost of employment by 17.2% in Georgia counties (equals
roughly 293,000 additional jobs) that were affiliated with and close to
Olympic activity. (3) As the HMZ model is semi-logarithmic, the
coefficients of the dummy variables are biased. Accounting for this
bias, the employment boost estimated by HMZ is even higher (18.8%),
which could be translated into 324,000 additional jobs in the Olympic
county areas. (4) Furthermore, as a second key result, they attest to a
positive and significant trend shift of an additional 0.2 percentage
points in their Olympic treatment group as compared with other counties
in Georgia as a result of the 1996 Olympics. The HMZ estimate exceeds by
a wide margin Baade and Matheson's (2002) ex-post estimate of
employment gains for the same event that ranged from 3,500 to 42,500
added jobs or the 29,000 jobs added exclusively to some sectors and
limited to the month of the staging of the Olympic Games, as estimated
by Feddersen and Maennig (2012a). (5)
Confronted with the lack of publications indicating substantial
positive economic impacts of such events, one might ask what reasons
might be responsible for the contrasting results obtained in the HMZ
study. To find a rational explanation for the divergent data regarding
the 1996 Olympic Games and the Atlanta metropolitan area as compared to
other sporting mega events analyzed in the literature, we accounted for
several potential sources of bias and reanalyzed the data according to
the methodology of HMZ. That is, we used the same time span (the first
quarter of 1985 until the third quarter of 2000), used the same industry
mix, and followed an endogenous method to determine the "true"
point in time of the intervention. We gratefully acknowledge Julie L.
Hotchkiss, Robert E. Moore, and Stephanie M. Zobay for providing the
original aggregated data and the SAS code.
Methodological Problems of the HMZ Study
DD in the Intercept: Correcting for Underlying Trends The
specification of the level shift DD model of HMZ is as follows:
lnEMP[U.sub.it] = [[beta].sub.1][X.sub.i] +
[[beta].sub.2]VN[V.sub.i] + [[beta].sub.3] POS[T.sub.t] + [[beta].sub.4]
VN[V.sub.i] * POS[T.sub.t] + [[epsilon].sub.it] (1)
Here, ln EM[P.sub.it]is log employment in county i of Georgia in
quarter t. [X.sub.i] is a vector of covariates including an intercept,
quarterly dummies, the industry mix displayed by the employee shares of
industry classes, and the population of each county. As the last two
variables should cover observable differences in the basic endowments of
the counties, they are fixed to values from the year 1990. POS[T.sub.t]
is the intervention variable for all counties. The intervention variable
has a value of zero before the intervention and a value of one following
the intervention. The variable VN[V.sub.i] ("venue and near venue
counties") controls for permanent level differences between the
treatment group and the control group. According to HMZ, the variable
takes a value of one if a county is a venue or near-venue county and
zero otherwise. VN[V.sub.i] * POS[T.sub.t] is an interactive variable
and takes a value of one for the post-intervention period of the
treatment group. This variable should capture in comparison to the
overall post-period level shift POS[T.sub.t]--an additional level shift
affecting the treatment group. Here, and in the following,
[[epsilon].sub.it] is the disturbance term, while other Greek letters
denote coefficients to be estimated.
Dachis, Duranton, and Turner (2011) demonstrate the importance of
controlling for spatial and temporal variation. If the model is
estimated without an adequate control for time-invariant, unit-specific
effects (or unit-invariant, time-specific effects), then these effects
are part of the residuals, and VN[V.sub.i] (POS[T.sub.t]) is endogenous.
Regarding Equation (1), all elements of the vector of covariates
([X.sub.i]) are fixed to their 1990 values and thus time-invariant,
acting as county quasi-fixed effects. To align the regression model with
standard DD models (e.g., Bertrand et al., 2004), the covariate vector
will be replaced by time-invariant effects, which does not affect the
estimated coefficient of interest (VN[V.sub.i] * POS[T.sub.t]). Because
the treatment group definition is based on county boundaries, the venue
and near venue county effects and the county-fixed effects cannot be
isolated separately. Thus, the variable VN[V.sub.i] must be omitted in
favor of county fixed effects. Similarly, a general post-Olympic effect,
POS[T.sub.t], and quarterly fixed effects cannot be separately
identified. Additionally, the three indicator variables based on
isolating seasonal effects, which are introduced by HMZ as part of the
covariate vector, [X.sub.i], are similarly captured by the quarterly
fixed effects. Thus, including county fixed effects and quarterly fixed
effects while dropping VN[V.sub.i], POS[T.sub.t], Quarter2, Quarter3,
and Quarter4 leads to the following equation, which is--with respect to
the variable of interest--equivalent to the original HMZ equation.
ln EM[P.sub.it] = [[alpha].sub.i] + [[phi].sub.t] +
[[lambda].sub.it]VN[V.sub.i] * POS[T.sub.t] + [[epsilon].sub.it] (2)
Furthermore, Dachis, Duranton, and Turner (2011) show that, after
controlling non-parametrically for all purely temporal and purely
spatial variation, this type of DD model requires that there be no
different trends in the growth of employment in the treatment group and
in the control group. In the presence of different employment trends
between the two groups, confounded estimates of the impact of the
Olympic Games might occur if these trends are correlated with the
variable of interest, VN[V.sub.j] * POS[T.sub.t]. To obtain unbiased
estimates of [[lambda].sub.it] using equation (4), the following
constraint must hold:
Cov([[epsilon].sub.it], VN[V.sub.i] * POS[T.sub.t]) = 0 (3)
Figure 1, which is a redraw of Figure 2 from the HMZ paper (p.
695), shows that the sequences of the VNV group and the non-VNV group
are explicitly characterized by an overall trend. If one tests a level
shift after an intervention point for a sequence which is characterized
by a stable trend for the whole observation period, it is unavoidable
that the mean for the post-period will be (significantly) higher than
for the preperiod. Moreover, if the overall treatment group trend is
higher than the control group trend, that is, if the groups are moving
apart, then the pre-period level shift for the treatment group must be
even higher. Thus, it is not surprising that HMZ found a significantly
positive level shift of about 19.6% for the control group and an
additional significantly positive level shift of another 18.8 percentage
points for the treatment group.
[FIGURE 1 OMITTED]
Thus, our first goal was to test whether controlling for
group-specific trends would change the results shown by HMZ. As the
original model contained variables capturing two different level shifts
(control group and treatment group), we must address the problem of the
existence of different trends in the two groups. A basic parametric
trend variable and the non-parametric, time-specific effects cannot be
identified separately, and thus, no general time trend is included in
the model. However, inspired by Figure 1, we control for a different
trend in the treatment group by adding a linear time trend for the venue
and near venue counties, [T.sub.t] * VN[V.sub.i]. [T.sub.t] is a time
trend starting with a value of 85 in the year 1985 and increasing by
0.25 each quarter, and VN[V.sub.i] identifies the Olympic host counties.
Equation (2) contains our first modified regression equation, where all
other variables are defined as above.
ln EM[P.sub.it] = [[alpha].sub.i] + [[phi].sub.t] +
[[lambda].sub.it]VN[V.sub.i] * POS[T.sub.t] + [[rho].sub.it][T.sub.t] *
VN[V.sub.i] + [[epsilon].sub.it] (4)
The determination of the intervention point is crucial. In the best
case, the intervention point is clearly exogenously fixed by an
(unexpected) intervention (e.g., a natural disaster, a war, a
revolution, the implementation of a law, etc.). (6) In other cases, the
point in time when the initial impact occurs is not obvious. Two
different cases are possible. First, ahead of the announcement of the
intervention, remarkable anticipation effects can be found. Second, the
starting date of the measures that are causal for an economic impact
might not be congruent with the announcement date. Furthermore, it might
be possible that the start of several Olympic measures would be
widespread in time. Thus, the starting point of the intervention cannot
be located in time explicitly. The second case might apply to the
Olympic Games 1996 in Atlanta (i.e., the intervention point was not
defined clearly). Particularly, it is not clear which facts might have
caused the intervention effect. The investment in infrastructure and the
staging of the Games including the spending of foreign visitors as well
as growth effects from improved infrastructure or improved image might
be identified as sources of an "Olympic effect" As is
self-evident, many reasonable points in time can be identified, and,
thus, the choice of the exact non-exogenous intervention point is always
debatable. To address this point, we used (according to HMZ) an
endogenous method for all regressions. The models were first estimated
using each year from 1991 to 1998 as the starting points for the
intervention indicator, POS[T.sub.t]. Then, in the next step, the model
with the best fit was chosen. We used the F statistic as the selection
criterion, as in the HMZ study. Here, the endogenously determined
procedure reveals 1994 as the best fit, as in the original HMZ level
shift setup. (7)
Last, as Bertrand, Duflo, and Mullainathan (2004) pointed out, DD
models tend to overestimate the significance of the intervention due to
serial correlations, which might lead to an overestimate of the
significance of the "intervention" dummy. To check for such
problems, we performed an LM test for serial correlation in a
fixed-effects model, as suggested by Baltagi (2001). (8) Note that the
test clearly rejects the null hypothesis of no serial correlation, and
as a result, the standard errors are corrected using an arbitrary
variance-covariance matrix, as recommended by Bertrand, Duflo, and
Mullainathan (2004) in all estimations.
Column 2 in Table 1 displays the results from Equation (2). The
coefficient of interest is identical to the original HMZ paper, while
the intercept deviates from the original results due to the different
setup of the county and time fixed effects. This column is shown to
demonstrate that our econometric results are equivalent. Regarding the
modified estimation of Equation (4) presented in the third column, the
inclusion of the basic treatment group trend is justified because the
corresponding coefficient is positive and highly significant. Using the
above-mentioned Halvorsen and Palmquist (1980) transformation of
coefficients, the basic trend for the VNV treatment group was 2.4%.
Obviously, the VNV counties have different characteristics, and seem to
have been more economically active--already before the Games. After
controlling for the basic trend within the employment figures, the
coefficient of the Olympic effect (VN[V.sub.i] * POS[T.sub.t]) became
insignificant (and negative). Thus, the main result of the original HMZ
paper (an employment boost of 18% or 324,000 additional jobs) appears to
have been based on neglecting fundamental trends in the employment data
for Georgia.
DD in the Slope: Allowing for Spline Knots and Simultaneous Level
and Trend Effects In a second regression, HMZ used a DD model to check
for Olympic-related changes in the growth rate of employment for the
treatment group and the control group. The regression equation of the
HMZ paper is as follows:
ln EM[P.sub.it] = [[alpha].sub.i] [X.sub.i] +
[[alpha].sub.2][T.sub.t] + [[alpha].sub.3] [T.sub.t] * VN[V.sub.i] +
[[alpha].sub.4][T.sub.t] * POS[T.sub.t] + [[alpha].sub.5][T.sub.t] *
VN[V.sub.i] * POS[T.sub.t] + [[epsilon].sub.it] (5)
where all variables are described above. The interacted term
[T.sub.t] * POS[T.sub.t] indicates changes in the trend for all counties
in the aftermath of the intervention. Finally, the model contained an
interacted term for trend differences for the treatment group during the
post period ([T.sub.t] * VN[V.sub.i] * POS[T.sub.t]). Positive economic
effects of the Olympic Games should lead to a significant coefficient of
this variable. In a first step, as described above, we replace the
covariate vector [X.sub.i] by county fixed effects. Attributable to the
inclusion of the unit-specific and time-specific effects, some original
variables could not be identified separately and, thus, have been
omitted. In the case of the HMZ trend regression, this involves
[T.sub.t] and [T.sub.t] * POS[T.sub.t]. Furthermore, HMZ included a
variable that relates the population in 1990 with the time trend,
[T.sub.t] * ln(population 1990), which also cannot be separated from the
time-specific effects and has thus been omitted.
Although the idea of modifying the DD approach to capture changes
in the slope is often embraced, the HMZ model is potentially erroneous
because it forced the regression line of the non-treatment and treatment
groups to have the same intercept, both for the pre-intervention and
post-intervention periods. To address these shortcomings, a spline DD
model or a combined spline and level shift DD model seem to be
appropriate.
In a spline model, two regression lines are joined together. The
turning point of the intervention is represented by a spline knot, which
joins the differently-sloped regression lines at a defined point
(Greene, 2003; Marsh & Cormier, 2001)
ln EM[P.sub.it] = [[alpha].sub.i] + [[phi].sub.t] +
[[rho].sub.it][T.sub.t] * VN[V.sub.i] + [[delta].sub.it] ([T.sub.t] - p)
* VN[V.sub.i] * POS[T.sub.t] + [[epsilon].sub.it] (6)
The newly included variable p takes the value of the year of the
(endogenously estimated) intervention. Consequently, prior to the
intervention point, the term ([T.sub.t] - p) * VN[V.sub.i] *
POS[T.sub.t] is zero because POS[T.sub.t] = 0. At the intervention
point, the term ([T.sub.t] - p) * VN[V.sub.i] * POS[T.sub.t] is still
zero due to [T.sub.t] - p = 0. After the intervention, the term
([T.sub.t] - p) * VN[V.sub.i] * POS[T.sub.t] gradually increases to
0.25, 0.50, 0.75, 1.00, and continues in this manner. Thereby, the term
captures the difference in employment growth within the Olympic host
counties from their long-run growth path conditional on the overall
growth pattern measured by the time fixed effects. All variables are
defined as above. Here, the coefficient of interest is [[delta].sub.it].
Equation (6) only allows for changes in the slopes while the
regression lines are knotted together. However, it is imaginable that
the start of investment in the sports venues and in general
infrastructure led to an immediate "jump" in the growth of the
employment figures in the corresponding counties. Because this kind of
effect is neither modeled in the pure level shift model nor in the
linear spline DD model, a combined level shift and spline DD model
should be tested. (9) This model accounts for both changes in the
intercept and changes in the slope without causing a pseudo level shift
while modeling the trend shift.
ln EM[P.sub.it] = [[alpha].sub.i] + [[phi].sub.t] +
[[rho].sub.it][T.sub.t] * VN[V.sub.i] + [[lambda].sub.it]VN[V.sub.i] *
POS[T.sub.t] + [[delta].sub.it] ([T.sub.t] - p) * VN[V.sub.i] *
POS[T.sub.t] + [[epsilon].sub.it] (7)
Again, all variables are the same as before.
In both columns of Table 2, the coefficients of the control
variables are comparable to those of the original HMZ regressions and to
the results shown in Table 1. The basic treatment group trend is 1.9%
for the pure spline specification (Equation (6)) and 2.3% for the spline
and level shift specification (Equation (7)). Both coefficients are
significant at the 1% level. Regarding Equation (6), the variable of
interest, the post-intervention treatment trend ([T.sub.t] - p) *
VN[V.sub.i] * POS[T.sub.t], is insignificant. Following the same
procedure for determining the intervention point in time, as employed by
HMZ--using each year from 1991 to 1998 as starting points for the
intervention indicator--1992 is the preferred intervention data for our
setup. This endogenously determined the intervention point differs from
the original HMZ study, which uses 1994 as the starting point of the
intervention. If we also choose 1994 as the starting point of the
intervention, the coefficient of interest ([[delta].sub.it]) becomes
weakly significant at the 10% level. After allowing for adjustments to
the intercept and the slope in Equation (7), a weakly significant and
negative level shift can be found. After the endogenously estimated
intervention point, no significant post-Olympics trend was found for the
treatment group. (10)
In sum, when selecting the starting point of the intervention
endogenously, the 1996 Olympics in Atlanta have no significant growth
effect in the venue and near venue counties. With a starting point at
1994 instead of the endogenous selection of the starting point, weak
evidence confirming the findings of the HMZ trend shift model can be
found. There is hardly any reliable evidence for a growth boost caused
by the 1996 Olympics.
Methodological Extension
The standard DD methodology, as employed in the previous section,
has been exposed to some critique. In particular, the inflexibility of
the method with regard to the functional form of the intervention can be
seen as a weakness of the standard approach. Recently, more flexible
approaches to identify treatment effects have been developed. (11)
As in most studies of the impact of policy measures or other
interventions, we are confronted with a two-dimensional identification
problem. First, we cannot rule out the possibility of a gradual
adjustment starting with the announcement of the Olympic host city,
e.g., due to the construction of Olympic facilities and related
infrastructure. (12) However, even if significant adjustments take
place, it will not be known a priori when the adjustment process starts
and ends. Second, and even more fundamentally, the treatment might not
be discrete in terms of space. The counties within Georgia might be
affected differently by the 1996 Olympics, as we expect the economic
response to vary with the frequency of Olympic events staged within a
county and the distance of a county with respect to the center of
Olympic activities.
Thus, our empirical strategy should be designed to cope with
unknown functional form, unknown spatial responses, and an unknown
starting date of the Olympic effect. Thus far, some studies have
attempted to deal with more or less flexible treatment definitions. For
example, Ahlfeldt and Wendland (2009) and Gibbons and Machin (2005)
model continuous treatments, whereas McMillen and McDonald (2004) allow
for gradual adjustments. To our knowledge, only a few studies have
allowed for complex continuous patterns with respect to both space and
time (Ahlfeldt, 2010; Ahlfeldt & Feddersen, 2010). Applying such
treatments implies a flexible test of the effect of the 1996 Olympics in
Georgia, as it isolates any underlying relative trends, in addition to
potential anticipation or adjustment processes, and not only compares
areas that are subject to treatment to control areas but also takes into
account the degree to which locations are affected.
Assuming that our outcome variable (employment) can be described as
a "surface" along the dimensions i (location) and t (time),
Dachis, Duranton, and Turner (2011) show that this surface can be
described by a Taylor series expansion. This function, in turn, can then
be decomposed into five parts: (1) a latent employment surface that is
constant in i and t; (2) variation that depends solely on location; (3)
variation that depends solely on time; (4) an interaction effect of both
variations; (5) a zero-mean error term. Clearly, we are mostly
interested in the fourth component, as it reveals the economic impact
and possible adjustments. To detect such an adjustment empirically, we
translate this idea into the following regression-based identification
strategy:
[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (8)
Again, [[alpha].sub.i] denotes county fixed effects, [[phi].sub.t]
quarterly fixed effects, [T.sub.t] a linear time trend,
[[epsilon].sub.it] and the error term. The values of [[gamma].sub.it]
and [[psi].sub.it] are series of coefficients to be estimated. Here, the
county fixed effects capture non-parametrically the proportion of
variation in the employment surface that is solely attributable to
location, while the quarterly fixed effects control non-parametrically
for all variation that depends solely on time. As mentioned before, we
are mostly interested in the vector of interaction variables of the
quarterly fixed effects and the treatment measure ([[phi].sub.t] *
[X.sub.i]). However, after controlling for location-specific and
time-specific effects, estimates of the interaction effects can also be
biased. In particular, if different county-specific employment trends
exist, confounded estimates of the impact of the Olympic Games might
occur if these trends are correlated with, but not caused by, the
Olympics. Therefore, we introduce a set of interactive terms of county
fixed effects ([[alpha].sub.i]) and a quarterly trend variable
([T.sub.t]).
The sequence of the coefficients forms an index of relative
employment within the Olympic treatment area relative to the initial
period, the second quarter of 1990. Following Ahlfeldt (2010), they
provide difference-in-difference estimates, as they differentiate the
(conditional) mean employment over space and time. More formally,
Equation (8) tests for significant deviations from a hypothetical linear
relative growth path, whereas a significant (positive) economic
adjustment should be reflected by a positive deviation from the long-run
path after the treatment becomes effective.
HMZ defined counties in which Olympic venues were located
("venue counties") and counties adjacent to these venue
counties ("near venue counties") as the treatment group
("venue and near venue counties" or "VNV"). In the
following discussion, the HMZ treatment measure will be denoted as
[x.sup.[alpha].sub.i] .
[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (9)
where the vector VNV is composed of 37 elements, which are
identified by their county codes within the state of Georgia according
to the FIPS system. Thus, VNV can be identified by the elements: 011,
013, 029, 045, 051, 053, 057, 059, 063, 067, 077, 085, 089, 097, 103,
111, 113, 117, 121, 135, 137, 139, 145, 151, 157, 187, 195, 213, 215,
217, 219, 221, 247, 255, 263, 297, 311.
As an alternative treatment variable, we delineate a more
conservative indicator variable, which includes only the ten venue
counties and not the near venue counties. For this purpose, a vector VEN
is defined containing the elements 051, 059, 063, 067, 089, 121, 135,
139, 215, 247.
[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (10)
Table 3 shows that the distribution of the Olympic events is
unequally distributed among the ten venue counties. For example, nearly
80% of the capacity-weighted Olympic events took place in the City of
Atlanta ("Fulton County"), followed by Clark County with
approximately 11%. The other counties each hosted between 0.5% and 3.5%
of the capacity-weighted events.
We allow for differences in the magnitude of the effect for the
different venue counties by using a third alternative Olympic treatment:
a dummy, of which the magnitude is set equal to a county's
percentage share of the overall spectator capacity. Consequently, the
indicator variable [x.sup.c.sub.i] is defined as follows:
[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (11)
Finally, geographic spillover effects should be analyzed. As
Atlanta could be referred to as a quintessential suburban U.S. city, the
majority of its people do not live near the city center but in the
suburbs. Often, these suburbs are located in neighboring counties, as
counties in Georgia are rather small. Thus, it could be expected that
regions surrounding the Olympic core region would benefit from its
primary impact. The treatment measure [x.sup.a.sub.i] can be seen as an
effort to capture spillover effects by declaring venue and near venue
counties to be equally weighted as a treatment group.
In a fourth alternative Olympic treatment, we consider the idea
that positive effects occur in the Olympic center and spread into the
neighboring regions, gradually decreasing with increasing distance from
the event area.
We measure distance as the straight-line distance in kilometers
between the geographic centroid of Fulton County and the centroid of
county i, dj 121. For Fulton County, we follow studies from economic
geography (Crafts, 2005; Keeble, Owens, & Thompson, 1982) and
generate an area's internal distance measure based on the surface
area,
[d.sub.int] = 1/3 [square root of ([area.sub.121]/[pi])] (12)
where [d.sub.int] is Fulton county's internal distance, equal
to one-third of the radius of a circle of the same area as Fulton county
in square kilometers ([area.sub.121]).
Following the gravity equation literature in international trade,
we model the spatial depreciation of the spillover effects as a power
function of the distance from Fulton County to county i, that is,
[d.sup.[tau].sub.i,121]. We apply a distance elasticity of [tau]=1,
which is in the center of the range of values estimated in empirical
economic geography (Anderson & van Wincoop, 2003; Redding &
Venables, 2004) and is commonly used in applied research (e.g., Redding
& Sturm, 2008). Thus, the fourth alternative treatment measure
[X.sup.d.sub.i] takes the following form:
[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (13)
where [d.sup.-1.sub.int] is the inverse of the internal distance of
Fulton county calculated according to equation (12), and
[d.sup.-1.sub.i,121] denotes the inverse of the distance from county i
to Fulton County (FIPS 121).
The estimation results for the y t index are presented in Figure 3.
Additionally, the corresponding 90% confidence intervals are displayed
as well as a linear and a smoothed lowess (locally weighted regression)
trend. Note that the non-linear trends are generated as a smoothed
function of the estimated time-varying non-parametric treatment
coefficients. These non-linear trend lines are, according to Ahlfeldt
(2010), similar to an alternative approach by McMillen and Thorsnes
(2000, 2003) using semiparametric regression techniques. The results
depicted in panel [a] relate to the treatment measure [x.sup.a.sub.i],
which corresponds to the original HMZ treatment group definition. Panel
[b] refers to the narrower treatment group definition of venue counties
only ([x.sup.b.sub.i]). Panel [c] represents a treatment group
definition ([x.sup.c.sub.i]) based on the percentage of session-weighted
capacity, and panel [d] shows the coefficient index of the continuous
treatment measure [x.sup.d.sub.i] based on distances to the Olympic
center. As we assume that no considerable anticipation effects occurred
before the announcement, the quarter prior to the election of the 1996
Olympic as the host (1990Q2) has been assigned as the base treatment
period. Moreover, previous quarters are not considered here because, due
to the flexible nature of this approach, they are not needed to
differentiate the pre- and post-effects, such as in a standard DD
approach.
[FIGURE 2 OMITTED]
Panel [a], which is based on the original HMZ treatment group
definition, displays an index of the estimated coefficients of the
"treatment measure"-"time dummy" interactive. Until
the beginning of 1995, the non-linear trend line shows a negative trend
in the effect series. From that date on, the trend line turns into a
positive sloped trend but is still negative compared with the linear
relative growth path. Regarding the 90% confidence interval, the
estimated coefficients are significant between 1993 and 1999. The
results depicted in the other panels show a similar pattern, whereas the
use of the venue county treatment measure [x.sup.b.sub.i] presented in
panel [b] reveals mostly positive but small and insignificant
coefficients of the [[psi].sub.it] index. Common to all panels of Figure
3, but revealing a different sharpness of the effect, is that starting
near the beginning of 1995, the clear negative slope of the coefficient
index turns into a positive slope. This effect ends near the end of 1996
for all four panels. Taking into account that the estimated coefficients
are mostly negative and sometimes also insignificant at the 10% level,
hardly any evidence for a short-term Olympic effect can be found.
In panel [c], which is based on a treatment measure according to
the percentage of session-weighted capacity, a peak can be observed in
the third quarter of 1996--the quarter in which the Olympic Games were
held. The coefficient is positive but insignificant. In conclusion, no
clear evidence for the existence of a lasting effect of the 1996 Olympic
Games can be found using a very flexible non-parametric identification
strategy. Furthermore, only weak evidence of a rebound effect of the
slope of a nonlinear trend between the beginning of 1995 and the end of
1996 can be attested.
Conclusions
This paper investigated the regional economic impact of the 1996
Olympic Games in Georgia. Our starting point was the analysis of
Hotchkiss, Moore, and Zobay (2003) in the SEJ, which, contrary to many
regional impact analyses of sports facilities, sports franchises, and
sporting-events, found significant and outstandingly strong employment
effects caused by the 1996 Olympic Games in Atlanta.
In our view, the original HMZ analysis is subject to two problems.
First, the level shift DD model used in the HMZ paper did not take
underlying trends into account. Second, the trend shift DD specification
of the original paper was characterized by an unintentional pseudo-level
shift. After addressing the first concern, we were unable to reject the
hypothesis that there was not a significant level shift in the
employment figures caused by the 1996 Olympics. In particular, the
prominently highlighted key result of HMZ, a tremendous boost of
employment by more than 17% in counties that were affiliated with and
close to Olympic activity, disappears completely after the introduction
of differing trends. Second, after modifying the HMZ trend shift
regression to control for time fixed effects and spline trend shifts, no
significant growth effect of the 1996 Olympics can be found in the venue
and near venue counties. Weak evidence confirming the findings of the
HMZ trend shift model can only be found if we do not apply the
endogenous selection criteria for the starting point introduced by HMZ
in the original paper. By fixing the start of the treatment period to
1994, a positive trend effect of less than 1%, which is only significant
at the 10% level, can be identified. Finally, we test on the basis of
four non-parametric models, which consider different complex continuous
treatment patterns with respect to space and time, to take into account
criticisms of the standard DD approach used in section 2 and in the
original HMZ study. Once again, we have to reject the hypothesis of a
long-term and persistent employment boost caused by the Olympics. Only
statistically weak evidence for a small rebound effect in Olympic venues
between the beginning of 1995 and the end of 1996 can be attested.
As a general conclusion, no persistent and significant level
shifting or trend shifting impact of the 1996 Olympics on the regional
economic development in Georgia can be confirmed by using any of three
different standard DD specifications and four different flexible
non-parametric DD specifications.
Thus, in light of the presented findings, the economic effects of
the 1996 Atlanta Olympic Games are in line with almost all other
scholarly ex-post analyses of mega sporting events, which have never
found any positive economic effects comparable to those reported by HMZ
in the original study. (13)
References
Ahlfeldt, G. M. (2010). Blessing or curse? Appreciation, amenities
and resistance to urban renewal. Regional Science and Urban Economics,
41(1), 32-25.
Ahlfeldt, G. M., & Feddersen, A. (2010). From periphery to
core: Economic adjustments to high speed rail. London School of
Economics & University of Hamburg, LSE research online, 29430.
Ahlfeldt, G. M., & Maennig, W. (2008). Impact of sports arenas
on land values: Evidence from Berlin. The Annals of Regional Science,
44(2), 205-227.
Ahlfeldt, G. M., & Wendland, N. (2009). Looming stations:
Valuing transport innovations in historical context. Economics Letters,
105(1), 97-99.
Anderson, J. E., & van Wincoop, E. (2003). Gravity with
gravitas: A solution to the border puzzle. American Economic Review,
93(1), 170-192.
Atlanta Committee for the Olympic Games. (1997). The official
report of the Centennial Olympic Game (Vol. 1, Planning and Organizing).
Baade, R. A. (1987). Is there an economic rationale for subsidizing
sports stadiums? Heartland Policy Study, 13.
Baade, R. A., & Dye, R. F. (1988). An analysis of the economic
rational for public subsidization of sports stadiums. Annals of Regional
Science, 22(2), 37-47.
Baade, R. A., & Matheson, V. A. (2002). Bidding for the
Olympics: Fool's gold? In C. P. Barros, M. Ibrahimo & S.
Szymanski (Eds.), Transatlantic sport: The comparative economics of
North American and European sports (pp. 127-151). Cheltenham, U.K. and
Northampton, Mass.: Edward Elgar.
Baltagi, B. H. (2001). Econometric analysis of panel data (2 ed.).
New York, NY: Wiley & Sons.
Baumann, R., Engelhardt, B., & Matheson, V. A. (2012).
Employment Effects of the 2002 Winter Olympics in Salt Lake City, Utah.
Journal of Economics and Statistics, 232(3), 308-317.
Bertrand, M., Duflo, E., & Mullainathan, S. (2004). How much
should we trust differences-indifferences estimates? Quarterly Journal
of Economics, 119(1), 249-275.
Brakman, S., Garretsen, H., & Schramm, M. (2004). The strategic
bombing of German cities during World War II and its impact on city
growth. Journal of Economic Geography, 4(2), 201-218.
Carlino, G., & Coulson, N. E. (2004). Compensating
differentials and the social benefits of the NFL. Journal of Urban
Economics, 56(1), 25-50.
Carlino, G., & Coulson, N. E. (2006). Compensating
differentials and the social benefit of the NFL: Reply. Journal of Urban
Economics, 60(1), 132-138.
Coates, D., & Humphreys, B. R. (1999). The growth effects of
sport franchises, stadia, and arenas. Journal of Policy Analysis and
Management, 18(4), 601-624.
Coates, D., & Humphreys, B. R. (2001). The economic
consequences of professional sports strikes and lockouts. Southern
Economic Journal, 67(3), 737-747.
Coates, D., & Humphreys, B. R. (2003). The effect of
professional sports on earnings and employment in the services and
retail sectors in US cities. Regional Science and Urban Economics,
33(2), 175-198.
Coates, D., & Humphreys, B. R. (2008). Do economists reach a
conclusion on subsides for sports franchises, stadiums, and mega-events?
Econ Journal Watch, 5(3), 294-315.
Coates, D., Humphreys, B. R., & Zimbalist, A. (2006).
Compensating differentials and the social benefits of the NFL: A
comment. Journal of Urban Economics, 60(1), 124-131.
Crafts, N. (2005). Market potential in British regions, 1871-1931.
Regional Studies, 39(9), 11591166.
Dachis, B., Duranton, G., & Turner, M. A. (2011). The effect of
land transfer taxes on real estate markets: Evidence from a natural
experiment in Toronto. Journal of Economic Geography, 12(2), 327-354.
Davis, D. R., & Weinstein, D. E. (2002). Bones, bombs, and
break points: The geography of economic activity. American Economic
Review, 92(5), 1269-1289.
Davis, D. R., & Weinstein, D. E. (2008). A search for multiple
equilibria in urban industrial structure. Journal of Regional Science,
48(1), 29-65.
Edmans, A., Garcia, D., & Norli, 0. (2007). Sports sentiment
and stock returns. Journal of Finance, 62(4), 1967-1998.
Feddersen, A., & Maennig, W. (2012a). Mega-events and sectoral
employment: The case of the 1996 Olympic Games. Contemporary Economic
Policy.
Feddersen, A., & Maennig, W. (2012b). Sectoral labour market
effects of the 2006 FIFA World Cup. Labour Economics, 19(6), 860-869.
Galster, G., Tatian, P., & Pettit, K. (2004). Supportive
housing and neighborhood property value externalities. Land Economics,
80(1), 33-54.
Gibbons, S., & Machin, S. (2005). Valuing rail access using
transport innovations. Journal of Urban Economics, 57(1), 148-169.
Greene, W. H. (2003). Econometric analysis. Upper Saddle River, NJ:
Prentice-Hall.
Halvorsen, R., & Palmquist, R. (1980). The interpretation of
dummy variables in semilogarithmic equations. American Economics Review,
70(3), 474-475.
Hotchkiss, J. L., Moore, R. E., & Zobay, S. M. (2003). Impact
of the 1996 Summer Olympic Games on employment and wages in Georgia.
Southern Economic Journal, 69(3), 691-704.
Humphreys, J. M., & Plummer, M. K. (1995). The economic impact
on the State of Georgia of hosting the 1996 Summer Olympic Games.
Mimeograph. Athens, Georgia: Selig Center for Economic Growth,
University of Georgia.
Imaizumi, A., Ito, K., & Okazaki, T. (2008). Impact of natural
disasters on industrial agglomeration: A case of the 1923 Great Kanto
earthquake. University of Tokyo CIRJE F-Series, 602.
Jasmand, S., & Maennig, W. (2008). Regional income and
employment effects of the 1972 Munich Summer Olympic Games. Regional
Studies, 42(7), 991-1002.
Keeble, D., Owens, P. L., & Thompson, C. (1982). Regional
accessibility and economic potential in the European community. Regional
Studies, 16(6), 419-432.
Lozano, F. A. (2011). The flexibility of the workweek in the United
States: Evidence from the FIFA World Cup. Economic Inquiry, 49(2),
512-529.
Maennig, W., & Richter, F. (2012). Exports and Olympic Games.
Journal of Sports Economics, 13(6), 635-641.
Marsh, L. C., & Cormier, D. R. (2001). Spline regression
models. Sage University Paper Series on Quantitative Applications in
Social Science, 07-137. Thousand Oaks, CA: Sage.
Matheson, V. (2006). Mega-events: The effect of the world's
biggest sporting events on local, regional, and national economies.
Department of Economics, College of the Holy Cross.
McMillen, D. P., & McDonald, J. (2004). Reaction of house
prices to a new rapid transit line: Chicago's Midway line,
1983-1999. Real Estate Economics, 32(3), 463-486.
McMillen, D. P., & Thorsnes, P. (2000). The reaction of housing
prices to information on superfund sites: A semiparametric analysis of
the Tacoma, Washington Market. In T. B. Fomby & R. C. Hill (Eds.),
Advances in econometrics: Applying kernel and nonparametric estimation
to economic topics (Vol. 14, pp. 201-228). Stamford, CT: JAI Press.
McMillen, D. P., & Thorsnes, P. (2003). The aroma of Tacoma:
Time varying average derivatives and the effect of a Superfund site on
housing prices. Journal of Business and Economic Statistics, 21(2),
237-246.
Redding, S. J., & Sturm, D. M. (2008). The costs of remoteness:
evidence from German division and reunification. American Economic
Review, 98(5), 1766-1797.
Redding, S. J., & Venables, A. J. (2004). Economic geography
and international inequality. Journal of International Economics, 62(1),
53-82.
Rose, A. K., & Spiegel, M. M. (2011). The Olympic effect.
Economic Journal, 121(553), 652-677. Skogman Thoursie, P. (2004).
Reporting sick: Are sporting events contagious? Journal of Applied
Econometrics, 19(6), 809-823.
Teigland, J. (1999). Mega-events and impacts on tourism; The
predictions and realities of the Lillehammer Olympics. Impact assessment
and project. Appraisal, 17(4), 305-317.
Tu, C. C. (2005). How does a new sports stadium affect housing
values? The case of FedEx Field. Land Economics, 81(3), 379-395.
Authors' Note
We thank two anonymous referees for their helpful comments.
Endnotes
(1) Cf. Matheson (2006) or Coates and Humphreys (2008) for an
overview.
(2) In a comment, Coates, Humphreys, and Zimbalist (2006) showed
that these results are not robust, for example, to the exclusion of
extreme outliers. However, see also the reply to this comment by Carlino
and Coulson (2006).
(3) Because HMZ found that the evidence for a wage effect of the
1996 Olympics is unclear (p. 703), we will concentrate on the employment
effects in the following.
(4) In a semi-logarithmic model, for a dummy parameter b the
percentage effect is equal to e-1 (Halvorsen & Palmquist, 1980). In
the following analysis, all coefficients in tables are uncorrected,
while coefficients interpreted in the text are corrected accordingly.
(5) It should be mentioned, however, that these estimates are based
on diverging geographical areas. For example, the HMZ study is based on
37 counties in Georgia, whereas the study by Baade and Matheson (2002)
is based on the Atlanta MSA, which reflects 20 counties within Georgia.
The HMZ estimate also exceeds the ex-ante projections of the Olympic
organizers of the 77,000 added jobs.
(6) Examples for exogenous interventions analyzed in other research
contexts are the German division and re-unification (Redding &
Sturm, 2008), the allied strategic bombing during World War II (Brakman,
Garretsen, & Schramm, 2004; Davis & Weinstein, 2002, 2008), and
earthquakes (Imaizumi, Ito, & Okazaki, 2008).
(7) An anonymous referee pointed out that choosing a specification
that maximizes fit tends to maximize the chance to find an effect that
is statistically significant. Note that our conclusions do not change
for alternative intervention points. Using the HMZ setup the post
intervention VNV effect is highly significant for all intervention
points. Furthermore, using altering intervention points together with
our setup controlling for the trend differences in equation (4) reveals
insignificant coefficients for our variable of interest except for 1991
as intervention point. For this year, this coefficient is weakly
significant on the 10%-level but negative.
(8) The LM test statistic is L[M.sub.5] = [square root of
(N[T.sup.2]/(T-1)[??][[??].sub.-1]/[??][??])], which is asymptotically
distributed as N(0,1).
(9) For a comparable DD approach, see Galster, Tatian, and Pettit
(2004).
(10) If we again replace this endogenously determined intervention
point by the intervention year 1994 from the original HMZ study, the
variable of interest ([T.sub.t] - p) * VN[V.sub.i] * POS[T.sub.t]
becomes weakly significant at the 10% level and positive, but at the
same time, the post-intervention level shift for the treatment group
becomes significantly negative, counteracting the positive trend effect.
(11) For examples, see, e.g., Ahlfeldt (2010), Ahlfeldt and
Feddersen (2010), Dachis, Duranton, & Turner (2011), Feddersen and
Maennig (2012b), and McMillen and Thorsnes (2000, 2003).
(12) Due to the difficult-to-predict host city election process by
the International Olympic Committee (IOC), we assume that no
considerable anticipation effects have occurred.
(13) Note that some studies found sports events to have an impact
on individual's behavior (Edmans, Garcia, & Norli, 2007;
Lozano, 2011; Skogman Thoursie, 2004). We owe this comment to an
anonymous referee.
Arne Feddersen [1] and Wolfgang Maennig [2]
[1] University of Southern Denmark
[2] University of Hamburg
Arne Feddersen is an associate professor in the Department of
Environmental and Business Economics at campus Esbjerg. His research
interests include sports economics, applied regional economics, and
media economics.
Wolfgang Maennig is a professor and the chair for Economic Policy
in the Department of Economics. His research interests include sports
economics, transportation economics, and real estate economics.
Table 1. Difference-in-Difference in the Intercept
Equation (2) Equation (4)
Intercept 8.4269 *** 7.9294 ***
(0.0141) (0.0943)
T x VNV -- 0.0239 ***
(0.0044)
#VNV x POST 0.1719 *** -0.0160
(0.0361) (0.0153)
[R.sup.2] 0.9929 0.9932
adj. [R.sup.2] 0.9927 0.9930
L[M.sub.5] 88.3557 88.36278
Intervention year 1994 1994
F statistic 19.4133 *** 20.0372 ***
Notes: *** p<0.01, ** p<0.05, * p<0.10. Robust standard errors,
which are computed using an arbitrary variance-covariance matrix,
as suggested by Bertrand, Duflo, & Mullainathan (2004, pp. 270-272),
are provided in parentheses.
Table 2. Difference-in-Difference in the Slope
Equation (6) Equation (7)
Intercept 8.0411 *** 7.9503 ***
(0.1083) (0.1128)
T x VNV 0.0185 *** 0.0229 ***
(0.0051) (0.0053)
#VNV x POST - -0.0357 *
(0.0186)
#VNV x (T-p) x POST 0.0068 0.0050
(0.0056) (0.0057)
[R.sup.2] 0.9932 0.9932
adj. [R.sup.2] 0.9930 0.9930
L[M.sub.5] 88.3526 88.3371
Intervention year 1992 1992
F statistic 20.1217 *** 19.8968 ***
Notes: *** p<0.01, ** p<0.05, * p<0.10. Robust standard errors,
which are computed using an arbitrary variance-covariance matrix,
as suggested by Bertrand, Duflo, & Mullainathan (2004, pp. 270-272),
are provided in parentheses.
Table 3. Frequencies of Olympic Sporting Events in Venue Counties
County FIPS Overall Number of % of Overall
Capacity Ticketed Sessions Capacity
Fulton, GA 121 11,273,100 411 78.7%
Clayton, GA 63 105,600 11 0.7%
Rockdale, GA 247 480,000 15 3.4%
Muscogee, GA 215 140,800 16 1.0%
Hall, GA 139 276,800 16 1.9%
Clark, GA 59 1,537,600 32 10.7%
DeKalb, GA 89 72,000 12 0.5%
Gwinnett, GA 135 440,000 16 3.1%
SUM 14,325,900 529 100.0%
Notes: Overall capacity is calculated by multiplying stadium
capacity by the number of ticket sessions. Calculations based
on Atlanta Committee for the Olympic Games (1997, pp. 539-544).
