Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
Department of Economics
Working Paper
What can bookies teach us about pari-mutuel wagering
reform? Applying Australian lessons to the U.S.
Charles C. Moul Miami University
Joseph M. G. Keller
dunnhumby
August 2011
Working Paper # - 2011-01
What can bookies teach us about pari-mutuel wagering
reform? Applying Australian lessons to the U.S.∗
Charles C. Moul† Joseph M. G. Keller‡
August 2011
Abstract
We consider a policy reform relaxing price controls in American pari-mutuel wager-
ing on horse racing by examining bookie behavior in Australia’s fixed-odds gambling
sector. Descriptive regressions indicate that bookie takeouts (the effective prices of
races) vary substantially and systematically with race characteristics, though in some-
times counterintuitive ways. Estimates of an explicitly reduced form model of bookie
takeout, however, qualitatively match both intuition and prior findings in the litera-
ture. Calibration using these estimates suggests that regulatory reform that permits
racecourses to alter takeout across races would increase variable profit by 3-6%.
Keywords: regulatory reform, gambling, horse racing
JEL codes: D22, L5, L83
Markets all over the world are subject to strict price regulation, and deregulation in those
areas could presumably increase welfare markedly. Unfortunately, the stringent regulations
themselves often prevent the observable variation that would permit empirical analysis to
gauge the potential welfare benefits from deregulation. One such example is American pari-
mutuel wagering on horse racing. This gambling, in which the racetrack takes a percentage
of money wagered (the take-out rate or takeout) and then returns the remainder to winning
bettors, is regulated at the state level and typically involves the takeout being set at a
particular level for different types of wagers. There have been several recent attempts to learn
the responsiveness of bettors to this takeout rate with temporary (Laurel Park, MD, 2007) or
∗This paper is based in part on Keller’s 2010 Master’s exit paper from Miami University. We thank
Max Kaftal for excellent research assistance (in particular for seeking out the data), Tim Ryan (CEO of the
Australian Bookmakers Association) for giving it to him, and Ryan Ruddy for additional research support.
We also thank our many Australian acquaintances with first-hand racetrack experience for sharing their field
research.†Corresponding author: Miami University Farmer School of Business, Department of Economics, Oxford,
OH 45056, USA. Tel: 513-529-2867. E-mail: [email protected]‡dunnhumby, Email: [email protected]
1
permanent (Hialeah, FL, 2010; Tioga Downs, NY, 2010) reductions, but results have thus far
been inconclusive.1 Such regulatory changes and experiments did not permit racetracks to
engage in price discrimination by actively adjusting the takeout across races. Bettor demand
may be such that varying takeout across races, rather than simply selecting a fixed level of
takeout, is another important margin of deregulation. Schmalensee (1981) admittedly shows
that price discrimination by a monopolist that does not increase quantity sold will increase
revenues but decrease welfare. Because states typically levy excise taxes on wagering, though,
shifting to taxes on cumulative takeout and allowing such price discrimination would increase
tax revenues and may allow the reduction of more egregiously welfare-harming taxes. It is
therefore possible that welfare would rise from such a policy change. Both regulatory issues
of takeout level and variability have been prominent in industry suggestions (NRTA Player
Panel Recommendations, 2004). Furthermore, the industry’s ongoing financial troubles from
increased gambling competition and the recent recession have focused both regulators and
racetracks on ways to improve the industry’s viability.
We have no data from U.S. racetracks and such regulatory changes, but we instead
observe bookies from Australian fixed-odds thoroughbred horse race wagering. Bookies in
this context set the takeout implicitly by selecting the level of odds on the field of horses,
and their behavior therefore may be indicative of what a profit-maximizing racetrack would
do if permitted to set and vary its own takeout. Such a comparison hinges on the similarity
between bettors in the two countries. Australians appear to gamble substantially more than
Americans, with Australian per capita losses at least triple those in the U.S.2 This gap,
though, presumably stems at least in part from the comparatively easy access to gambling
in Australia rather than radically different consumer preferences. Moreover, the countries’
common heritages and other resemblances suggest that the comparison may be apt. To the
extent that Australian and American bettors share preferences over race characteristics and
disutility on money spent, our results can shed light on the revenue and welfare impacts for
certain U.S. reforms.
Previous researchers have used variation in takeout rates and amounts wagered to es-
timate price-elasticities for U.S. pari-mutuel gambling (Gruen, 1976; Suits, 1979; Mobilia,
1993; Thalheimer and Ali, 1998; Gramm et al., 2007). These studies have generally found
1Laurel Park halved its takeouts across the board for ten days in August 2007. Hialeah Park lowered its
takeout to 12% for all bet-types in October 2010. Tioga Downs reduced all its takeouts to the state minima
(15% for win/place/show bets) for the 2010 season and maintained those lower levels for the 2011 season.2"The world’s 10 biggest gambling nations - including Canada", Vancouver Sun, 7/11/2011, data taken
from H2 Gambling Capital.
2
that takeout rates are higher than the revenue-maximizing level, with point estimates of
own-takeout elasticities ranging from -1.6 to -3. To our knowledge, no studies have consid-
ered how race characteristics themselves can affect these price-elasticities. This is exactly
the information that the estimates of our reduced form model provide. When combined with
plausible estimates of own-price demand parameters, these estimates allow for the identifica-
tion of the full set of structural parameters and can therefore predict the impacts of different
policy recommendations. We need race-level characteristics to employ this analysis, and we
lean heavily on Coffey and Maloney (2010, henceforth CM) to do so. While that paper uses
data from Churchill Downs (KY) in 1994 to distinguish the incentive effect from selection in
explaining the correlation between performance and reward, it also includes regression results
that show the impact of race characteristics on the amount of money wagered, known in the
U.S. as handle and in Australia as turnover. CM find that handle increases in field size (i.e.,
number of horses) and purse and decreases in dispersion of horse-talent. It is those regres-
sion results (with CM’s summary statistics) that we will combine with the Kentucky-specific
price elasticity estimate from Thalheimer and Ali (1998) in our calibration exercise.
While we build on the above literature, our research is superficially most similar to the
work of Shin (1993) in that we both use the bookmaker’s implied takeout as the dependent
variable and employ race characteristics such as size of field and dispersion of horse-talent as
explanatory variables. Shin (1993), though, frames his empirical exercise as identifying the
prevalence of insider trading, which he posits is the cause of the recurrently observed favorite-
longshot bias in which favorites are underbet and longshots are overbet.3 Our approach, on
the other hand, begins with descriptive regressions and then turns to estimating a reduced
form model that is explicitly derived from a structural model. While the estimates from
the descriptive regression are more useful for predicting the equilibrium impacts of race
characteristics on takeout, the reduced form estimates illuminate the mechanisms by which
those impacts arise.
We begin by showing the extent of variation in our observed bookie takeout rates. Even on
the same day and at the same racetrack, the sample standard deviation for takeout is three
percentage points (compared to a mean of seventeen percentage points). Our descriptive
3Broadly speaking, bookies in Shin’s model protect themselves from bettors with inside information on
longshots by offering less favorable odds on those horses than objective probabilities would suggest. Cain et
al. (2003) provide additional empirical support consistent with the thesis. Working against the primacy of
this interpretation, recent research has looked to explain the observed longshot bias in pari-mutuel gambling
as the result of bettor misperception (Sobel and Raines, 2003; Snowberg and Wolfers, 2010) or sequential
information release (Ottaviani and Sorenson, 2009).
3
results linking bookie takeout rates with race characteristics are generally as expected, in
that these takeouts are higher for races with more ex ante even fields, for races with larger
fields, and for late-day races. The inclusion of a race’s purse (prize money to top finishers)
as a proxy for a race’s unobserved prestige and quality, though, yields a surprising result:
observed takeouts are lower for higher purse races. As we find no anecdotal support for the
idea that competition is greater for these races, this apparent demand anomaly is analogous
to the common observation that many items go on sale during periods of high demand.
Explanations for this curious pricing pattern have often centered on an item being a loss-
leader (e.g., Chevalier et al., 2003), but Nevo and Hatzitaskos (2005) offer a novel if more
prosaic explanation that we follow. If high demand periods are characterized not only by
an outward shift of demand but also by an increase in demand elasticity, then firms with
market power may respond to the increased elasticity by lowering markups. Likewise, if
higher purse races attract more elastic bettors, then profit maximizing bookies will respond
by lowering their implied takeout rates.4 The structural model from which we derive our
reduced form model allows for just this sort of demand rotation. The resulting estimates
indicate that the model reconciles the observed negative correlation of purse and takeout
with the expectation that bettor demand should be higher for higher purse races by letting
the slope of the bookie’s residual demand become flatter.
The reduced form parameter estimates are insufficient to conduct any policy experiments,
as most estimates are only identified up to scale. We return to our originally posited rep-
resentative bettor and calibrate the price-disutility parameter from a previously estimated
own-take elasticity for Kentucky and average race characteristics from Churchill Downs,
thereby identifying all parameters of the structural model. Using these parameters, we back
out what bookie variable profits and welfare would have been under hypothetical regulatory
regimes. These results indicate that restricting Australian bookies to common takeout rates
(instead of the observed variable rates) would have reduced bookie variable profits by 3-6%.
This compares to the 7.7% increase in variable profits that our preferred elasticity measure
suggests would occur if Kentucky regulators reduced racecourse takeout for win bets from
16% to 14%.
We next lay out the institutions of pari-mutuel wagering on horseracing in the U.S.
and the hybrid wagering of Australia that allows both pari-mutuel and fixed-odds bets. A
simple model links a bookie’s observed odds to the implied takeout rate, which becomes our
4This can be rationalized by the presence of a core group of price-insensitive bettors (addicts) whose
influence is lessened when more price-sensitive bettors (tourists) attend higher visibility races.
4
primary variable of interest. We introduce our model of a representative bettor’s demand for
horserace wagering and consider the equilibrium supply-side relationships between implied
takeout rates and race characteristics. After discussing the data, we estimate descriptive and
reduced form regressions, and we close with a discussion of our results from the calibration
exercise.
1 Background
1.1 Institutions: American and Australian
Horse race gambling in the U.S. uses the pari-mutuel format exclusively.5 Takeout rates in
the US tend to be set directly by state government or by the state’s regulatory body over
gambling, though some states offer racetracks limited discretion.6 As of 2008, takeout for
win bets ranged from California’s 15.4% to Arizona’s 20% (HANA 2008). Kentucky, the
state which gives us the race characteristics for our calibration exercise, had a relatively low
takeout rate of 16% for win bets. Takeouts for exotic bets on multi-horse outcomes (e.g.,
exacta, quinella, Pick 6) are also regulated and tend to be several percentage points higher
(e.g., Churchill Downs has a takeout of 19% for those non-simple bets). Most states have
excise taxes on handle. For a sense of the magnitudes of handles and the related taxes,
Churchill Downs in the 2009 season had about $649M in handle on which it paid about
$20M in tax.7
In the pari-mutuel format, bettors make wagers on horses in the time leading up to the
race, but, as posted odds are preliminary, these wagers are for an unspecified price. Payouts
are based on final odds and depend on how the handle is distributed across the field of horses
at race time. For simple bets (such as the win-bets that we observe in our Australian data),
the racetrack subtracts a percentage equal to the takeout and then returns the remaining
money to the bettors who placed wagers on the winning horse proportionally to the amounts
wagered. Odds for any given race are therefore determined entirely by how bettors decide
to wager. These wagers can be made at the racetrack or at off-track betting facilities.8
5After gambling was prohibited in almost all American states in the early 20th century, racetrack-operated
pari-mutuel wagering on horse races was reintroduced in many revenue-starved states during the Depression,
with the new condition that excise taxes were placed on handle.6New York, for example, mandates minimum takeout rates of 15% for win/place/show bets, though it
appears that only Tioga Downs since its recent reduction is presently at those minima.7Churchill Downs 2009 annual report, p. 48.8The Interstate Horseracing Act of 1978 (Public Law 95-515) stipulates that off-track betting facilities
are subject to the same regulations regarding takeout as the racetracks themselves and that such facilities
5
Fixed odds gambling in horse racing (as can be found in Australia, Ireland, the United
Kingdom, and several other countries) differs from pari-mutuel wagering in several ways. The
most prominent distinction is the existence of the bookmaker, an individual who is actively
setting odds. As the format’s name implies, odds offered to a bettor are fixed, though these
odds may be changed for subsequent bettors. Foremost to this analysis, the takeout (the
expected percentage of the last dollar bet that is kept by the bookie) is implicitly determined
by the set of odds chosen by the bookmaker and thus can vary across time, racetracks and
races.
Australia, like its Imperial kin, allows gambling using both a fixed odds format and a
pari-mutuel format. There are 379 racetracks in Australia, though only 17 are considered
large. Major racetracks in the same city rarely operate on the same day.9 While Saturday
races tend to draw the largest crowds, Wednesday and Friday (midweek) races are also held.
A number of independent bookmakers (typically 20-40) at these racecourses compete for
bettor business against one another, against the on-site pari-mutuel system, and against
all off-site gambling options. Each bookmaker should be thought of as a three-person team
consisting of the bookie, the penciler (who records odds), and the ledger (who records bettors’
wagers). Depending on the state, the pari-mutuel system is either state-run or operated
under substantial regulation by a for-profit firm. The pari-mutuel takeout rate in New South
Wales (Sydney) and Victoria (Melbourne) varies by bet type; the pari-mutuel takeout rate
for straight win bets is 14.5% in both cities. Pari-mutuel takeout in Queensland (Brisbane)
is regulated differently in that the blended takeout (weighted average of simple and exotic
bet takeouts) cannot exceed 16% over a twelve-month period and no takeout rate can exceed
25%.
The horses that are slated to race are known in advance of race day. Opening odds from
the bookmakers are posted approximately 30 minutes before race time, and changes to these
odds are periodically made prior to the posting of the official starting prices. As shown in
McAlvanah and Moul (2011), the takeouts implied by these fixed odds start out relatively
high (∼30%) and tend to fall as the race approaches. This decline occurs as the risk to thebookie of a bettor with inside information falls and as the value to the bettor of fixed odds
wagers relative to pari-mutuel wagers becomes smaller. In the data and throughout this
paper, a wager’s gross odds is the amount for each dollar wagered that is returned to the
bettor in the event of his horse winning. For example, a $1 wager on a winning horse with
be at least 60 miles from the nearest racetrack.9In our sample, two racetracks in the same city are open on only 2 of our 241 Saturdays.
6
listed odds of 4 would pay back $4 (the original $1 plus the $3 of winnings). The price of a
wager is the reciprocal of those odds, and so the above wager would have a price of 0.25.
Two commonly used measures of bookies’ profit potential are the margin and takeout.10
The margin is defined as the amount of a marginal dollar wagered that is retained by the
bookie as a proportion of the amount returned to bettors. The takeout is defined as the
amount of a marginal dollar wagered that is retained by the bookie as a fraction of the total
amount wagered. For example, a 25% margin corresponds to the bookie retaining 20% of
the total amount wagered as takeout and paying out 80%. The connecting formulae between
margin and takeout are thus = +1
and = 1− . Both the margin and takeout should
be weakly positive, else there exists an arbitrage opportunity. While bookie margin has
instructive parallels with Arrow-Debreu prices that sum to more than one as bookies impose
the equivalent of a tax, we prefer the implied takeout in order to facilitate comparisons with
the competing and American pari-mutuel regimes. All of our empirical results are robust to
employing bookie margin instead of implied takeout rates as the dependent variable.
1.2 Transforming observed odds into implied takeout
We now detail assumptions under which the bookmaker’s expected takeout for a race can
be constructed from a set of observed odds. We assume that bettors obtain sufficiently high
recreational utility from gambling as to always wager on a race and that they decide on
which horse to bet on the basis of the expected monetary payoff, assuming risk neutrality.
Our model of bettors is thus a special case of Ottaviani and Sorensen (2010) without private
bettor information so that bettors share common beliefs about race outcomes. Imperfectly
competitive bookmakers set odds , the gross payout to a winner of a $1 wager on horse k
winning the race. Let denote the bookie’s subjective probability of horse k winning the
race. The expected takeout on horse k is thus = 1 − . In expectation, the bookie
retains of every dollar wagered on horse k and pays out . Let denote a bettor’s
subjective probability of horse winning the race. Bettor equilibrium conditions imply that a
bettor is indifferent between a wager on any two horses: = ∀ . These conditionsalso correspond to the bookie maintaining a balanced book, where the bookie is guaranteed
a riskless return. When combined with the fact that subjective probabilities sum to one
(P
= 1), our system contains K equations for K horses. For a given set of observed odds
10An alternative term for the margin is the overround, and alternative, more colorful terms for the takeout
are the juice, the vig (short for vigorish), the edge, and the house edge.
7
in equilibrium, one can uniquely determine the bettor subjective probabilities: =11
.
Note that the converse is not true, as bettor subjective probabilities do not correspond to
a unique set of odds. Bettor subjective probabilities determine only the ratio of odds; for
example, 12=
21, and 1
2=
31for a three-horse race. The bookie has the capacity to fix
the magnitude of the odds for any one horse and thus implicitly the takeout for the race.
Consider the following simple three-horse race example. Substituting =1−
into the
consumer indifference conditions yields(1−1)1(1−2)2 =
21and
(1−1)1(1−3)3 =
31, which simplifies
to 1−11−2 =
1221
and 1−11−3 =
1331. As before, we have more unknowns than equations, and
horse-level takeouts are not uniquely identified by the subjective probabilities. Without
loss of generality, assume the bookie sets the odds on horse 1 and thus determines 1. The
consumer indifference conditions then imply that 2 = 1−(1−1)2112and 3 = 1−(1−1)3113
.
These equations indicate that a bookie maintaining a balanced book cannot set individual
horse-level takeouts independently of each other, due to the inter-linking of odds imposed
by the bettor equilibrium conditions. The expected race-level takeout would thus appear to
depend on , , and for all k.
The expected race takeout is then the sum of individual horse-level takeouts, weighted
by each horse’s fraction of the total amount wagered. In the general case in which bookie
and bettor subjective probabilities may not coincide, we follow the pari-mutuel system and
assume that the fraction wagered on a particular horse coincides with bettor subjective
probability for that horse, so that =P
. Substituting = 1 − and our prior
expression for equilibrium subjective probabilities as the relative share of summed reciprocal
odds yields the formula for takeout:
=X
µ1P 1
¶(1− ) =
µ1P
1
¶X
µ1
−
¶= 1−
µ1P
1
¶(1)
Alternatively, the race-margin is given by
=X
1
− 1 (2)
Intuitively, the extent to which the reciprocal gross odds (i.e., the wagers’ prices) sum to
greater than one signifies bookie’s expected profit margin.
We now link the race takeout (and implicitly the margin) to the bookmaker’s presumed
objective function of expected profits. Let Λ denote the number of dollars wagered on horse
k, and let denote the total amount wagered on a race with a bookmaker. The expected
profit for the race will then be () =P
Λ. Using the prior assumption that the amount
8
of money wagered on a particular horse as a share of the total amount wagered coincides with
bettor subjective probability on that horse (i.e., Λ= ), then () =
P = ∗ .
If the total amount wagered depends on the takeout so that ( ), then the bookie chooses
the level of odds and implicitly the takeout to maximize ( ) ∗ . The bookie’s problem isthus analogous to a revenue-maximizing firm facing a downward-sloping demand curve.
2 Model
We begin by considering the most general case of our bookie’s profit-maximization problem.
Bookmakers choose their margin (and implicitly their takeouts) to maximize their expected
profits conditional on having a balanced book. Letting denote the relevant residual demand
for win bets, the first order condition for the profit maximization problem is
()
= +
= 0 (3)
If race characteristics are exogenous to the bookmaker, then appealing to the Implicit
Function Theorem yields the comparative statics of race characteristics on bookmaker take
at the optimum.
= −
+ 2
2+ 2
2
(4)
The denominator is negative by necessity to ensure a maximum. The sign of (that is, the
sign of a coefficient in our descriptive regressions) then depends on the sign of + 2
.
If the impact of on the slope of the demand of betting is insignificant (i.e., 2
≈ 0),then descriptive estimates
will be the same sign as the impact of race characteristics on
the amount of money wagered . If, however, changes in affect the slope of the demand
curve (i.e., 2
6= 0), may not mimic
in sign.
We posit a representative bettor who has a linear demand for race-time win-bets.11 Let
denote this demand in AUD$ wagered in racetrack i, race k, and period t:
= + − (1 + + ) + (5)
In the above, denotes race takeout (the price of wagering on the race). denotes a
matrix of race characteristics that shift demand, and (a subset of ) denotes a matrix of
characteristics that rotate demand. We allow for two unobserved disturbances ( and ); the
11Limiting our discussion to race-time bets matches the fact that we observe race-time odds and minimizes
the distinction between pari-mutuel and fixed-odds wagers for our cross-country comparison.
9
former captures ex ante determined characteristics that may influence the number and quality
of horses in the field, and the latter captures ex post characteristics. Respective examples for
the two are the race’s prestige and weather on the day of the race. This framework introduces
the obvious concern of omitted variable bias, as higher prestige races may generate larger
and more even fields and the econometrician would then be unable to distinguish bettors’
preferences on the race’s prestige from other observed race characteristics. While bettors
are unlikely to care directly about a race’s purse, purse and prestige are presumably highly
positively correlated. We therefore address this concern by using race-purse as a proxy
for race-prestige ( = ) and henceforth subsume purse into our and
matrices.
The aggregated demands for win-bets can be found bymultiplying this representative con-
sumer’s demand by some fraction of racetrack i’s market population for period t¡()
¢.
Let denote this aggregated demand:
=
¡()
¢− ¡()¢ (1 + ) +¡()
¢(6)
As Louisville has a single major thoroughbred racecourse (Churchill Downs) which only
permits pari-mutuel betting and has no off-track betting facilities nearby, we assume that
= 1 in that case. This demand specification is exactly our endgoal for that market,
and we will employ it when calibrating against the Thalheimer and Ali (1998) own-take
elasticity estimate and the CM summary statistics. Australian betting and its consequent
demand estimates differ in that bettor demand in Australia is divided between pari-mutuel
wagering and fixed-odds wagering and between on-track and off-track wagering. Using data
from the Australian Racing Book (2008) and assuming that tracks do not serve as same-day
competitors to each other, we set = 0084.12
We will further assume that our observed bookmaker at racecourse i in period t is one
of bookies. As the summary statistics will show, the observed bookie takeouts are quite
similar to the pari-mutuel takeouts for the racetrack’s state for Sydney and Melbourne and
higher than the pari-mutuel takeout in Brisbane. When combined with the observation that
racetracks have many (at least 20 and up to 45) bookies, this suggests that bookies engage in
(perhaps tacit) collusion rather than imperfect competition.13 We estimated our model under
12Australian Racing Fact Book 2007/08, pp. 66, Table 86: equals the ratio of summed 2002-07 bookmak-
ers on course totals to sum of comparable total wagering. While it would have been preferable to specify a
model that permitted substitution among the different betting alternatives, we lacked the data to implement
such a model in any meaningful way.13The tourist trap model of Diamond (1971) could also rationalize this result, though the bookies being
10
the differing assumptions of perfect cartel behavior and symmetric Cournot behavior, and
the levels of demand that would rationalize observed takeouts under the Cournot assumption
were implausible (e.g., each bookie-team capturing AUD$72,000 in a eight-race Saturday).
This is driven by observed bookie takeouts being so close to the pari-mutuel takeout and
the assumption that pari-mutuel takeout is either at or above the revenue-maximizing level.
We therefore maintain the hypothesis that our sampled bookie is part of a cartel and limit
our theoretical discussion to that case. The bookie’s residual demand is thus his share of
the market demand.
Our demand specification in (6) then implies inverse demand of
=
µ
1 +
¶µ
¶+
(1 + )−µ
1
() (1 + )
¶ (7)
Our bookie maximizes revenue by setting odds that imply a takeout of
∗ =
µ
1 +
¶µ
2
¶+
2 (1 + )=
µ
1 +
¶µ1
2
¶ + (8)
This is the standard result of revenue-maximization under linear demand, where a monop-
olist will optimally choose a price equal to half of demand’s vertical intercept (choke-price).
Our reduced form can therefore identify the structural parameters that are interacted with
takeout (i.e., demand rotating s) but demand shifting s are identified only up to scale:
=
. Note that neither the population-weighting parameter nor the population variable
appears in this pricing equation; they will arise only in our cross-country calibration exer-
cise. Residuals will be heteroskedastic by construction, and robust standard errors will be
required.
3 Data
The data set, courtesy of the Australian Bookmakers Association, includes near-complete
fixed odds betting information from Saturday races from November 2, 2002, to August
4, 2007, at nine Australian tracks.14 These racetracks lie in three different markets and
states–four in Sydney, New SouthWales; three in Melbourne, Victoria; and two in Brisbane,
Queensland. Odds are taken from a sampled bookie for each racecourse and day. The
data originally contained 5,213 racing starts. Six races were dropped because of apparently
in such close proximity at the racetrack makes a search cost explanation unlikely.14The data did not include eight Saturdays, five of which occurred from November 2002 through January
2003. We thus observe 96.8% of Saturday races between November 2, 2002 and August 4, 2007.
11
erroneous data (for example, all horses having the same odds). Another 190 races were
dropped because they included late scratches.15 The remaining 5,017 observations were then
matched with the races’ total purse value where possible.16 Because purse data were not
available for all races, the final data set includes 4,661 observations. This contrasts favorably
with the sample sizes employed by Shin (1993) and Cain et al. (2003) which respectively
had 136 and a maximum of 1430 observations. Other than purse data being less available
in Victoria than New South Wales and Queensland (82% vs. 97% and 98%), the race
characteristics when purse was and was not observed are similar.
For each race, we observe the date, racetrack, size of field, ordinal placement of race
(e.g., second of day), purse value, and the starting (i.e., final) odds on horses from the sam-
pled bookie. The starting odds were used to calculate bettors’ subjective probabilities, the
bookmaker’s takeout, and various measures of dispersion in the field (e.g., Gini coefficients,
variance of subjective probabilities, entropy). While all dispersion measures yielded similar
results, we will focus on the variance of the logged subjective probabilities (VarLP) as this
is the best match to CM which we use for our later calibration exercise. We operationalize
the race’s ordinal placement by creating an indicator variable Late that equals 1 if the race
is the eighth of the day or later. As late-day races are often the most publicized of the day,
we found that this variable sufficiently captured that form of intraday variation. Finally, we
include a week-based time trend over the sample to capture any secular changes in demand.
Table 1 reports summary statistics for the full Australian sample and broken down by
market, as well as CM’s 1994 sample.17 Of primary interest is the bookie takeouts. Takeouts
differ markedly in levels across markets, with Sydney bookies retaining 13.6% of money
wagered, Brisbane bookies retaining 21.7%, and Melbourne bookies in between with 15.8%.
We have no compelling explanation for the elevated Brisbane takeouts, though it seems
likely to be related to anecdotal observations that Brisbane bettors are largely domestic
Australians while Sydney and Melbourne tracks receive more wagering from potentially
more price-sensitive southeast Asian bettors. There is substantial variation in takeouts
across markets, but over four-fifths of the variation within markets occurs in races on the
same day at the same track. To the extent that weather is relatively constant within a day,
15Late scratches occur when horses drop out of races after bookmakers publish opening odds but prior to
the start of the race. McAlvanah and Moul (2011) consider how this sort of late change to the field might
lead to deviation from our profit-maximization assumption.16Purse values were obtained from Racing Information Services Australia (RISA), Racing New South
Wales, and Queensland Racing.17CM do not restrict themselves to Saturday races, but they do limit their data set to races that pay out
to four places. As the Kentucky Derby in 1994 paid out to five places, it was excluded from their sample.
12
this strongly suggests that race characteristics play an important role in the takeouts that
bookies set.
We deflate purse to 1994 US $ to facilitate comparisons with the Louisville data.18 Races
at Melbourne and Sydney racetracks offer substantially higher purses than those held at
Brisbane or Louisville tracks, but Brisbane’s average purse is comparable to that of Louisville.
All Australian markets show higher variance in purse than Louisville, though this is lessened
somewhat when the 1994 Kentucky Derby (excluded from CM) is included. Australian races
tend to have larger fields with more variation than the observed Louisville races; this is
driven in large part by the fact that Australian racetracks have a number of fields with at
least 20 horses, while the Louisville sample’s maximum field size is 12. Ex ante dispersion of
the field varies markedly across the three Australian markets, and the Louisville races show
comparable dispersion.
Table 2 presents simple correlations among the observed variables of interest and means
conditioning on a race’s ordinal placement for the entire Australian sample and broken
down by market. Perhaps the most striking figures are the large and positive correlation
coefficients between field size and take. While consistent with Shin’s story of insider trading,
they could also reflect bettor demand for races with more horses (as found in CM) or field
size capturing unobserved race-quality measures. The negative correlation between take and
purse for Melbourne is superficially counterintuitive and appears to run against the results
of CM.19 Even when that correlation is positive, it is of a smaller magnitude than one might
expect. Variable means that condition on whether a race is late in the day indicate that late
races have higher takeouts, larger fields, and (weakly) less dispersion of horse-ability. While
late races have larger purses in Melbourne, data surprisingly indicate that late races have
smaller purses in Brisbane and Sydney.
4 Results
4.1 Empirics
Table 3 displays the descriptive results and t-statistics when takeout is regressed on various
race characteristics. These estimates should be interpreted as the equilibrium impact of the
18The exchange rate in 1994 and 2002-07 was similar and relatively stable at US $0.74 to AU $1. We use
US-CPI data to adjust for inflation.19The negative correlation between purse and takeout for the entire sample is primarily driven by the
fact that Brisbane has low purses and high takeouts while Melbourne and Sydney have high purses and low
takeouts.
13
characteristic on bookie takeout. Given the widely differing levels of takeout across markets,
we include market fixed effects for the full sample, and we also consider market-specific
regressions. Preliminary estimates indicated an increasing and concave relationship between
takeout and field size that was well accommodated by including field size in logs, and so we
proceed using that transformation.
All estimates indicate that takeout falls with purse (though insignificant in the Sydney
regression) and field dispersion (though insignificant in Brisbane regression) and rises with
field size and being late in the day. The time trend polynomial estimates imply that takeout
falls early in the sample, stabilizes and then falls again toward the sample’s end, though
these transition points vary across markets.
We highlight two points from these descriptive regressions. First, coefficients appear to
differ enough across markets to warrant market-specific, rather than pooled, regressions going
forward. Second, with the exception of the negative impact of purse on takeout, these results
are largely consistent with the extant literature. CM find on-site pari-mutuel handle to be
increasing in field size and decreasing in field dispersion. As stated above, Shin (1993) and
the related papers have already documented the positive relationship between field size and
takeout. The purse coefficient, however, stands out. In addition to being counterintuitive, it
appears to contradict our primary purpose for its inclusion, namely that it serves as a proxy
for unobserved (to the econometrician) ex ante race quality. It is this apparent paradox
and the potential resolution proposed by Nevo and Hatzitaskos (2005) that motivates our
particular structural model and its reduced form.
We display our estimates for the reduced form model in Table 4. As discussed above,
our maintained hypothesis is that the observed bookie is part of a bookmaker-cartel, and
so our linear demand implies that the observed takeout is simply one-half of the market
demand’s vertical intercept (i.e., choke price).20 The prior descriptive regressions are merely
special cases of the reduced form in which = 0 for all variables. The nonlinear least
squares estimates for Brisbane and Melbourne reconcile our prior expectations and purse’s
descriptive impact on demand. In both markets, increases in purse shift demand outward
but also increase price sensitivity. Melbourne’s estimates especially showcase the value of the
reduced form model. Those estimates show that, while purse’s net impact is a combination
of countervailing forces (outward shift and more price sensitive), field size’s net impact is
a combination of two forces working in the same direction (outward shift and less price
20Identification of -parameters comes from second-order effects and interactions that were not explored
in the prior simple descriptive results.
14
sensitive). While not shown, when is set to zero in the Brisbane regression, is
positive and highly significant (t-stat ≈ 16).Estimates using Sydney races, though, are less satisfactory. While not significant, es-
timates indicate that increasing purse shifts demand inward and makes bettors less price
sensitive. The latter results could be rationalized, but the former runs counter to both intu-
ition and the results of CM. It is unclear why our model would accommodate Melbourne so
readily and yet struggle with Sydney, even though they both represent high-visibility, high-
quality gambling markets. For our purposes, these results imply that a representative bettor
inferred from Sydney bookie takeouts will be irreconcilable with our Louisville situation. We
therefore focus our calibration exercises on the Brisbane and Melbourne estimates.
4.2 Calibration exercise
We begin by inferring the price-disutility parameter from Louisville’s observed population,
sample means from CM, the Kentucky-specific own-take elasticity of Thalheimer and Ali
(1998), and the -parameters of our reduced form model after deflating purses to 1994 USD.
The linearity of our structural model implies that own-take elasticity (evaluated at sample
means) is
= ∗ ¡− ¡1 + ¢¢ ∗µ
¶(9)
so is calibrated as b = µ −1 + b
¶µ
¶µ1
¶(10)
In addition to the estimates from Table 4, we use a 1994 Louisville population of 676,404,
Thalheimer and Ali’s estimate of = -1.85, the state takeout for win bets of 0.16, and sample
means as found in Table 1.21
Table 5 displays the parameter values that result when using the Brisbane and Mel-
bourne reduced form estimates (3.24 and 3.36, respectively). Given the small Louisville
sample means and comparable estimates of , it is not surprising that the markets’ implied
s are so similar. We then use each value for to compute per capita handle for each of
our observed races. Results suggest that Melbourne’s per capita money wagered is about
double that of Louisville and that Brisbane’s is even greater. This is consistent with other
data showing Australians’ pronounced propensity to gamble, even compared to a U.S. city
21This Louisville population is a linear extrapolation of the 1990 and 2000 Census observations of the
Louisville consolidated city-county.
15
where wagering is relatively easy. Interacting these per capita values with the city popu-
lations using on-track bookies yields the average amount wagered with bookies each race.
Multiplying those aggregate amounts by the implied takeout gives a measure of aggregate
cumulative takeout (variable profit) by race. To the extent that bookies fail to achieve our
profit-maximizing ideal, this figure is an overstatement, but it may also be understated as
bookies capture rents on bets well before racetime. We then compute a rough measure of
per-bookie variable profit by dividing this aggregate figure by the typical number of bookies,
finding that the average three-person bookie-team in Brisbane earns $940 per race and the
average Melbourne bookie-team $900 per race. Over our sample’s complete years (2003-06),
these figures suggest that a Brisbane bookie-team who worked every Saturday race for a year
at one of the racetracks would earn a gross income of about $170,000, from which associated
expenses and track fees would have to be paid. Melbourne racetracks have races on fewer
days, and a comparable bookie-team would earn a gross income of about $100,000. These
implied figures seem somewhat high for three individuals, but, without information on the
fees that bookies must pay in order to be on-site, it is difficult to gauge the extent of the
overstatement.
Our recovered structural parameters also enable us to consider how their implied impacts
compare to the elasticity estimates of CM. Table 5 displays CM’s 95% confidence intervals
for purse elasticity, field size elasticity, and dispersion hemi-elasticity. Neither Brisbane’s
estimates nor Melbourne’s estimates match especially well, as both imply purse effects well
below and field size effects well above those of CM. For both elasticities, however, Brisbane
is much closer to Louisville than Melbourne is. On the dispersion measure, Brisbane’s
implication is within the confidence interval while Melbourne’s is not. When combined with
the fact that the variation in the Louisville sample is most similar to that of the Brisbane
sample, we cautiously judge that implications from the Brisbane estimates are likely to have
more relevance for our Louisville question.
Levels and elasticities are interesting side-tests, but our primary concern is the predicted
impact of allowing a U.S. pari-mutuel racetrack the flexibility to vary its takeout across
races.22 We address this by considering the converse question of what would happen if our
bookmaker cartel were restricted to charge the fixed profit-maximizing takeout
=1
2∗P
( + )P (1 + )
(11)
22These percentage impacts are therefore generally robust to any proportional overstatements that the
prior figures may display.
16
We compare the outcomes under this counterfactual to the previously inferred outcomes
under perfectly variable takeouts. The implication of these results is then the change that
would result if constant pari-mutuel takeouts were already set at the unit-elastic level.
Our first counterfactual considers the comparison between all tracks in a market facing a
common fixed takeout, and the second considers the case of when each track may have its own
fixed takeout. Consistent with Schmalensee (1981), moving from a fixed to variable takeout
in either market raises cumulative takeout, substantially decreases consumer surplus, and
marginally decreases welfare.23 There appears to be minimal heterogeneity across racetracks
that would lead to markedly different outcomes under the two counterfactual scenarios.
Brisbane impacts are uniformly smaller (in absolute value) than Melbourne’s, and, given our
prior arguments, we will focus on the Brisbane predictions. Allowing flexible takeouts raises
cumulative takeout by 3%, lowers consumer surplus by over 8%, and lowers welfare by 1%.
To put those figures into context, the Thalheimer and Ali (1998) elasticity estimate of -1.85
implies that dropping Kentucky’s takeout for win bets from 16% to 14% would raise handle
by 23.1% and cumulative takeout by 7.7%.
Similar to other states, the state of Kentucky levies an excise tax on handle on live races
of 3.5% (1.5%) for large (small) tracks, where $1.2M of daily average handle is the size-
threshold (KRS 138.510). Our linear functional form implies that moving from a fixed to
variable takeout will have no impact on cumulative handle, and so the current excise tax
regime would yield no gains to the state from a reform that increases takeout flexibility.
If, however, Kentucky were to tax cumulative takeout instead of handle, then some of the
gains would go to the state government where they could displace or prevent other taxes
with higher negative welfare consequences. Given the presently fixed takeouts, this change is
largely semantic, in that the 3.5% excise tax on a large track’s money wagered is equivalent
to a 21.875% tax on cumulative takeout on simple win/place/show bets.
5 Conclusions
While bookies have no role in pari-mutuel wagering in the U.S., we have provided a model
to link our Australian estimates to potential reforms of the American horse racing industry.
Our estimates highlight the value of the incorporation of theory into empirical work and also
provide more support for the idea that many observations that appear paradoxical within a
23Our linear functional form implies that moving from a fixed to variable takeout will have no impact on
handle.
17
model of perfect competition can be readily reconciled in a model that allows market power.
Finally, our estimates give some idea of the impacts that would follow reform that grant
racetracks flexibility in setting takeouts.
Economists are broadly interested in cases when deregulation might improve welfare,
and pari-mutuel wagers on horse racing, while a small industry, illustrates how cross-country
study may provide critical empirical variation to achieve answers to relevant questions. Given
the possibility that proposed flexibility would lower consumer surplus and even welfare, we
hesitate to advocate too strongly for it. Nevertheless, the American horse racing industry is
in a sufficiently parlous state that a proposal that may increase track profits and viability
as well as government coffers without reducing welfare is worth some consideration.
References
[1] Australian Racing Fact Book 2007/08 (2008), Jake Howard (ed.), Australian Racing
Board: http://www.australianracingboard.com.au/uploading/arbFinalBook.pdf
[2] Cain, Michael, David Law, and David Peel (2003). “The Favourite-Longshot Bias, Book-
maker Margins and Insider Trading in a Variety of Betting Markets,” Bulletin of Eco-
nomic Research, 55(3): 263-73.
[3] Chevalier, Judith A., Anil K. Kashyap, and Peter E. Rossi (2003). “Why Don’t Prices
Rise During Periods of Peak Demand? Evidence from Scanner Data,” American Eco-
nomic Review, 93(1): 15-37.
[4] Churchill Downs Incorporated Annual Report. 2009: http://www.annualreports.com/
partners/Report/17423
[5] Coffey, Bentley, and Michael Maloney (2010). “The Thrill of Victory: Measuring the
Incentive to Win,” Journal of Labor Economics, 28(1): 87-112.
[6] Diamond, Peter (1971). “A Model of Price Adjustment,” Journal of Economic Theory,
3: 156-68.
[7] Gramm, Marshall, C. Nicholas McKinney, Douglas H. Owens, and Matt E. Ryan (2007).
“What Do Bettors Want? Determinants of Pari-Mutuel Betting Preference,” American
Journal of Economics and Sociology, 66(3): 465-91.
18
[8] Gruen, Arthur (1976). “An Inquiry into the Economics of Race-Track Gambling,” Jour-
nal of Political Economy, 84(1): 169-78.
[9] Hialeah, Announcement: “Hialeah Cuts Takeout Rate to 12% for All Bets,” 10/6/2010:
http://www.bloodhorse.com/horse-racing/articles/59227/hialeah-cuts-takeout-rate-to-
12-for-all-bets
[10] Horseplayers Association of North America, “Track Ratings 2008 by Track Name”:
http:// www.horseplayersassociation.org/hanatrackratingsbytrackname.html
[11] Laurel Park, Announcement: “Laurel Park’s Summer Meet to
Feature Ten Days at 10 Percent” (press release), 7/18/2007:
http://www.laurelpark.com/NR/rdonlyres/430A0FAE-003A-405E-9840-
4271D226F225/18111/LPJuly07News.pdf, press release
[12] Laurel Park, Results: “Jury Still Out on Recent Takeout-Reduction Experiments,” Tom
LaMarra, 9/17/2007: http://www.bloodhorse.com/horse-racing/articles/40837/jury-
still-out-on-recent-takeout-reduction-experiments
[13] McAlvanah, Patrick, and Charles Moul (2011). “The House Doesn’t Always
Win: Evidence of Anchoring among Australian Bookies,” working paper:
http://ssrn.com/abstract =1540614
[14] Mobilia, Pamela (1993). “Gambling as a Rational Addiction,” Journal of Gambling
Studies, 9(2): 121-51.
[15] National Thoroughbred Racing Association Players’ Panel Recommendations,
2/19/2004: http://www.horse-races.net/library/article-ntrapanel.htm
[16] Nevo, Aviv, and Konstantinos Hatzitaskos (2005). “Why Does the Average Price Paid
Fall During High Demand Periods?” NBER Working Paper No. w16511.
[17] Ottaviani, Marco, and Peter Norman Sorensen (2009). “Surprised by the Parimutuel
Odds?” American Economic Review, 99(5): 2129-34.
[18] Ottaviani, Marco, and Peter Norman Sorensen (2010). “Noise, Information, and the
Favorite-Longshot Bias in Parimutuel Predictions,” American Economic Journal: Mi-
croeconomics, 2(1): 58-85.
19
[19] Schmalensee, Richard (1981). “Output and Welfare Implications of Monopolistic Third-
Degree Price Discrimination,” American Economic Review, 71(1): 242-7.
[20] Shin, Hyun Song (1993). “Measuring the Incidence of Insider Trading in a Market
forState-Contingent Claims,” Economic Journal, 103(420): 1141-53.
[21] Snowberg, Erik, and Justin Wolfers (2010). “Explaining the Favorite-Long Shot Bias:
Is it Risk-Love or Misperception?” Journal of Political Economy, 118(4): 723-46.
[22] Sobel, Russell S., and Travis Raines (2003). “An examination of the empirical derivatives
of the favourite-longshot bias in racetrack betting,” Applied Economics, 35(4): 371-85.
[23] Suits, Daniel B. (1979). “The Elasticity of Demand for Gambling,” Quarterly Journal
of Economics, 93(1): 155-62.
[24] Thalheimer, Richard, and Mukhtar M. Ali (1998). “An Economic Analysis of a
Parimutuel Racetrack-Racebook,” Working paper for Equine Industry Program, Uni-
versity of Louisville.
[25] Tioga Downs, Announcement: “NY Board approves Tioga
Downs take-out reduction” (press release), 3/26/2010:
http://xwebapp.ustrotting.com/absolutenm/templates/?z=1&a=36219
[26] Tioga Downs, Results: “Tioga Downs take-out reduc-
tion yielding positive results” (press release), 6/21/2010:
http://xwebapp.ustrotting.com/absolutenm/templates/?z=1&a=37537
[27] Vancouver Sun, "The world’s 10 biggest gambling nations - including Canada",
7/11/2011: http://www.vancouversun.com/world+biggest+gambling+nations+includ
ing+Canada/5084481/story.html
20
Table 1: Summary Statistics of Australian (2002‐07) and Louisville/Churchill Downs (1994) data
‐‐ Louisville data from Coffey and Maloney (2010), Table 1a
All‐AUS Brisbane Melbourne Sydney Louisville
# Tracks/Races 9/4661 2/1672 3/1261 4/1728 1/566
Pari‐mutuel TOb‐‐‐ ~16%/25%c
14.5% 14.5% 16%
Implied bookie TObmean 17.12% 21.74% 15.84% 13.59% ‐‐‐
std‐all 5.28% 3.87% 4.04% 3.79% ‐‐‐
std‐by track 3.87% 3.86% 3.99% 3.78% ‐‐‐
std‐by track‐yr 3.71% 3.72% 3.86% 3.61% ‐‐‐
std‐by track‐day 3.23% 3.22% 3.24% 3.24% ‐‐‐
Betting poold mean ‐‐‐ ‐‐‐ ‐‐‐ ‐‐‐ 153,410
std ‐‐‐ ‐‐‐ ‐‐‐ ‐‐‐ 110,410
Pursed mean 54,460 29,829 71,724 65,695 28,748
std 116,370 45,477 132,543 144,191 26,122
min 11,525 13,648 13,130 11,525 7,320
max 1,881,251 612,002 1,792,082 1,881,251 233,950
Field Size mean 11.0 11.9 11.0 10.2 10
std 2.9 2.9 2.8 2.7 2
min 4 5 4 4 4
max 21 20 21 20 12
VarLPe mean 0.90 0.79 0.67 1.16 1.01
std 0.52 0.37 0.35 0.62 0.38
min 0.022 0.035 0.077 0.022 0.09
max 4.96 2.67 2.91 4.96 2.28
Latef mean 0.15 0.12 0.16 0.17 ‐‐‐
a Louisville dataset excludes 1994 Kentucky Derby. If Derby (Purse=$500,000, Field Size = 14 with 11 unique win‐bets, VarLP = 0.344)
included, Purse's mean and standard deviation rise to $29,579 and $32,743, but other variables' means and standard deviations are
essentially unchanged.
b All takeouts apply to win‐bets.
c Pari‐mutuel blended takeouts in Brisbane cannot exceed 16% over 12‐month period and cannot exceed 25% for any bet‐type.d Betting pool and all purses in 1994 US $ (multiply by 1.637 for 2002 AUS $)e VarLP=Var(ln()) where is bettor subjective probability implied by observed oddsf Late is indicator for race being eighth or later in day
Table 2: Correlations and conditional meansa
All markets (n=4661)
Take Purse Field Size VarLP Trend Late=0 Late=1
Take 1.00 ‐0.09 0.55 ‐0.25 ‐0.10 Take 16.79% 18.98%
Purse ‐0.09 1.00 0.14 0.12 0.01 Purse 0.090 0.084
Field Size 0.55 0.14 1.00 ‐0.07 0.02 Field Size 10.8 12.6
VarLP ‐0.25 0.12 ‐0.07 1.00 0.08 VarLP 0.90 0.85
Trend ‐0.10 0.01 0.02 0.08 1.00
Brisbane (n=1672)
Take Purse Field Size VarLP Trend Late=0 Late=1
Take 1.00 0.02 0.38 ‐0.04 ‐0.08 Take 21.46% 23.79%
Purse 0.02 1.00 0.32 0.12 0.10 Purse 0.050 0.040
Field Size 0.38 0.32 1.00 0.06 0.10 Field Size 11.8 13.2
VarLP ‐0.04 0.12 0.06 1.00 ‐0.04 VarLP 0.79 0.79
Trend ‐0.08 0.10 0.10 ‐0.04 1.00
Melbourne (n=1261)
Take Purse Field Size VarLP Trend Late=0 Late=1
Take 1.00 ‐0.07 0.55 ‐0.16 ‐0.18 Take 15.47% 17.79%
Purse ‐0.07 1.00 0.17 0.16 0.00 Purse 0.115 0.128
Field Size 0.55 0.17 1.00 0.03 ‐0.03 Field Size 10.7 12.5
VarLP ‐0.16 0.16 0.03 1.00 0.16 VarLP 0.68 0.62
Trend ‐0.18 0.00 ‐0.03 0.16 1.00
Sydney (n=1728)
Take Purse Field Size VarLP Trend Late=0 Late=1
Take 1.00 0.08 0.68 ‐0.19 ‐0.15 Take 12.99% 16.48%
Purse 0.08 1.00 0.19 0.10 ‐0.02 Purse 0.112 0.085
Field Size 0.68 0.19 1.00 ‐0.04 ‐0.02 Field Size 9.7 12.3
VarLP ‐0.19 0.10 ‐0.04 1.00 0.12 VarLP 1.18 1.05
Trend ‐0.15 ‐0.02 ‐0.02 0.12 1.00
a Purses in 2002 AUS $Ms, implied take on win‐bets, VarLP=Var(ln()) where is bettor subjective probability implied by
observed odds, Late is indicator for race being eighth or later in day
Table 3: Descriptive regressions (OLS) on implied win‐bet takeout T (in percentage points)a
Sample All‐AUS Brisbane Melbourne Sydney
#Tracks/Races 9/4661 2/1672 3/1261 4/1728
E(T) 17.12% 21.74% 15.84% 13.59%
b /t/ b /t/ b /t/ b /t/
Purse ‐1.13 2.38* ‐3.21 3.64** ‐2.68 6.40** ‐0.18 0.60
ln(FS) 7.73 22.41** 6.28 17.12** 8.41 24.56** 8.39 25.50**
VarLP ‐0.88 7.02** ‐0.32 1.37 ‐1.19 5.12** ‐0.87 9.52**
Late? 1.28 8.22** 1.53 6.03** 0.87 3.95** 1.32 6.57**
Trend ‐8.72 7.31** ‐8.91 5.56** ‐3.89 2.45* ‐12.58 10.39**
Trend2 7.34 7.85** 7.98 5.77** 4.80 3.50** 9.15 8.85**
Trend3
‐1.89 8.51** ‐2.10 6.05** ‐1.62 4.67** ‐2.02 7.80**
Intercept ‐‐‐ 9.35 10.00** ‐1.92 1.92 ‐0.01 0.02
R2 0.6538 0.2172 0.4116 0.5382
a Results using all markets employ market fixed effects; purse deflated to 2002 AUS $Ms; Trend reflects number of weeks since
start of sample (divided by 100); all t‐statistics reflect White correction and those using all‐market sample also reflect clustering
at the racetrack level and use of Student t distribution; * and ** indicate 95% and 99% levels of significance
Table 4: Reduced Form estimates (NLLS) on implied win‐bet takeout T (in percentage points)a
B = Pop*(X ‐ (1+Z)T + ) Structural demand T* = (X/(1+Z))(1/2) + U, where = / Profit‐maximizing takeout
Sample Brisbane Melbourne Sydney
#Tracks/Races 2/1672 3/1261 4/1728
E(T*) 21.74% 15.84% 13.59%
b /t/ b /t/ b /t/
Purse 0.85 3.20** 0.66 4.53** ‐0.043 1.67
ln(FS) ‐0.100 1.72 ‐0.110 3.17** ‐0.193 9.73**
Purse 31.32 3.98** 12.55 10.62** ‐1.76 1.02
ln(FS) 5.73 0.89 10.55 2.84** 4.31 1.65
VarLP ‐0.53 1.09 ‐2.06 4.01** ‐0.93 3.88**
Late? 2.34 2.83** 1.30 2.71** 1.29 2.27*
Trend ‐14.20 3.37** ‐5.91 1.76** ‐13.47 4.39**
Trend2 12.78 3.40** 7.48 2.49* 9.78 4.11**
Trend3 ‐3.37 3.45** ‐2.54 3.17** ‐2.16 3.81**
Intercept 23.48 3.51** 2.08 0.43 11.13 3.76**
R2 0.2180 0.4172 0.5464
a Reduced form assumes bookmakers acting as cartel; purse deflated to 2002 AUS $Ms; Trend reflects number of weeks since
start of sample (divided by 100); all t‐statistics reflect White correction; * and ** indicate 95% and 99% levels of significance
Table 5: Calibration results
Reduced form model re‐estimated using purse in 1994 US $Ms and with Take in percentage
T* = (X/(1+Z))(1/2) + U, where = / Profit‐maximizing takeout
B = Pop*(X ‐ (1+Z)T + ) Structural demand
Matched against Thalheimer & Ali's (1998) KY‐specific own‐take elasticity = ‐1.85 ‐1.85 = = Pop*(‐(1+Z))*(T/B), so = (‐/(1+Z))*(1/T)*(B/Pop)
Louisville Brisbane Melbourne
‐‐‐ 3.24 3.36
Louisville observations/Australian implications for win bets (in 1994 US $)a
Per capita $ wagered 0.23 0.56 0.43
Total $ wagered 153,410 84,010 134,750
Total $ retained 24,546 18,798 22,545
Per‐bookie $ retainedb‐‐‐ 940 902
Retention rate 16% 22.38% 16.73%
Louisville s from 1994 (Pop = 676404), 95% confidence intervalsc
Purse [0.270, 0.431] 0.119 0.014
Field Size [0.226, 0.574] 1.046 1.826
VarLP [‐0.227, ‐0.043] ‐0.076 ‐0.307
Counterfactuals from takeout flexibility around optimal retention rate
Case 1: All tracks in market face same and constant takeout rate over entire sample
% change in cumulative takeout 3.17% 6.61%
% change in consumer surplus ‐8.37% ‐15.57%
% change in welfare ‐0.98% ‐1.97%
Case 2: Each track faces constant takeout rate over entire sample
% change in cumulative takeout 3.17% 6.39%
% change in consumer surplus ‐8.36% ‐15.15%
% change in welfare ‐0.98% ‐1.91%
a Louisville money wagered is total amount wagered with Churchill Downs, Australian money wagered is implied amount
wagered with on‐site bookmakers b Assumes 20 (25) bookmakers in each of Brisbane (Melbourne) racetracks c Estimates taken from Coffey and Maloney (2010), Table 3. VarLP parameter is hemi‐elasticity.