Vertical separation increases gasoline prices.
Wilson, Nathan E.
I. INTRODUCTION
The choice and implications of firm structure remain central
questions in industrial economics, and for good reason. Analysis of
recent U.S. Census data suggests that within firm transactions now
create the same amount of value as those between firms (Lafontaine and
Slade 2007). However, despite a rich literature describing the potential
trade-offs involved in choosing between integration and separation,
little is understood about the consequences of a change in form when
both are commonly used in an industry. In large part, this reflects the
difficulty of estimating the marginal impact of an organizational form
as its usage is likely driven by many of the same factors determining
economic outcomes.
I add to the literature by estimating the marginal impact of
vertical separation in retail gasoline markets, where stakeholders have
long been skeptical about integrated organizational forms. Some states
have even limited refiner ownership of retail outlets despite an
economics literature suggesting that such integration can lower prices
by better aligning all parties' incentives (Cooper et al. 2005;
Vita 2000). To account for concern about the endogeneity of form, I use
panel data methods as well as an instrumental variables (IV) strategy.
The combination of the two approaches produces estimates robust to the
presence of time invariant and time varying factors that simultaneously
influence the choice of form and economic outcomes.
After accounting for the endogeneity of form, I find that switching
a station to a vertically separated form increases gasoline prices by
amounts equal to 25%-45% of the average retail margin. These price
effect estimates are slightly smaller in magnitude than the results in
papers using statewide policy changes to identify the impact of
organizational form (Barron and Umbeck 1984; Vita 2000). This difference
makes sense given the distinct estimation strategies. In this paper, I
identify the impact of form at marginal stations, where the refiner is
fundamentally indifferent between forms. In contrast, using policy
changes to identify the marginal effect of form conflates marginal
stations with inframarginal ones, where a freely optimizing refiner
would never choose to vertically separate. Thus, one might expect papers
using policy changes to overstate the marginal impact of form while
accurately estimating the consequences of adopting a law prohibiting a
form. Similarly, estimates of the marginal effect of form will not
necessarily accurately predict the consequences of a sweeping policy
change.
Interestingly, I do not find that vertical separation has a
statistically significant impact on overall fuel sales. I interpret this
as indicating that the high-powered incentives given to local agents at
vertically separated stations induce greater effort to increase local
demand. This can explain the apparently contradictory result of
observing increased prices but not lower sales quantities. Moreover, it
is consistent with past theoretical models of vertical separation in
gasoline retailing (Slade 1996) as well as anecdotal impressions of
refiners' organizational strategies (Kleit 2005).
Overall, my paper contributes to two different literatures. First,
it adds to a small but growing group of papers that disentangle the
effect of form choice from the elements impacting the choice of form.
The papers in this literature suggest that separation can have widely
varying consequences. For example, Novak and Stem (2008) find that
vertical separation induces a trade-off between initial performance and
improvement in the automobile industry, while Forbes and Lederman (2010)
show that vertically integrated regional airlines are better at
minimizing delays and cancellations. In contrast, Kosova et al. (2013)
find no observable difference in outcomes across organizational forms
for hotels associated with a large brand owner. The different findings
in our respective papers suggest that vertical separation induces local
agents to exert effort in different ways in the two industries we
examine.
Second, this paper offers insight into the marginal impact of
vertical separation in an industry where it remains an ongoing source of
controversy. The results suggest that, contrary to some
stakeholders' concerns, vertical integration is associated with
lower prices.
II. VERTICAL SEPARATION AND GASOLINE RETAILING
A. The Impact of Vertical Separation on Economic Behavior
Vertical integration can benefit upstream producers by allowing
them to manage the tension between themselves and their downstream
retailers. Absent formal integration, both upstream and downstream firms
have an incentive to mark-up their prices to their respective customers
(the retailers and retail consumers, respectively). This leads to retail
prices being too high and total quantities being too low relative to the
profit-maximizing levels achieved by an integrated firm. Thus, the
theory of "double marginalization" predicts that vertical
separation should be associated with higher prices and lower quantities
(Spengler 1950).
Given the problem posed by double marginalization, agency theory
can explain why organizational forms other than vertical integration are
sometimes chosen. The theory acknowledges the possibility that
employees' effort is both commercially important and not perfectly
observable. In such situations, a principal (the upstream firm) may be
able to increase profits by tying agents' (retailers) compensation
to performance (Holmstrom and Milgrom 1991; Laffont and Martimort 2002).
However, the impact of vertical separation, when adopted for agency
reasons, on prices and quantities is ambiguous. It depends on how
agents' effort affects profits. For example, if higher effort by
the agent reduces costs (Shelton 1967), vertical separation could have
no, or even a negative, effect on prices. Alternatively, if agents'
efforts increase demand, prices might increase without necessarily also
causing quantities to decline as when prices rise due to double
marginalization.
Agency models that allow for multiple types of efforts show that
agents' effort can be a double-edged sword, especially when the
different types of effort affect the different parties'
asymmetrically. For example, if the principal's interests are
affected by multiple factors (e.g., sales and service quality), but
transaction costs only allow agents' variable compensation to be
linked to one (e.g., sales), then vertical separation may lead to moral
hazard on the part of the agents (i.e., shirking on quality provision).
Costly monitoring may mitigate the problem, but this also reduces the
desirability of using the form (Brickley 1999; Brickley and Dark 1987).
If the costs of monitoring vary across locations, firms might
employ both vertically separated and integrated forms. For example, if
firms fear that a highly incentivized agent will shirk on important
non-sales factors when monitoring is difficult, they will be less likely
to choose vertical separation in those circumstances. Using distance
from headquarters to proxy for monitoring cost, Brickley and Dark (1987)
find evidence for this. (1)
B. Vertical Separation in Gasoline Retailing
Gasoline stations are either exclusively affiliated with individual
refiners (e.g., Exxon or Shell) or are independent and may purchase
unbranded gasoline from whatever refiner they choose. I focus on
refiner-affiliated stations, because refiners' organizational form
choices connect straightforwardly to principal-agent models of vertical
integration (Shepard 1993; Slade 1996). They are also more common,
accounting for 78% of the industry in 2002 according to Kleit (2005).
The refiner-affiliated stations observed in my data are divided
into four categories: salary operation, lessee dealer, open dealer, and
jobber/wholesaler. (2) Of these four organizational forms, only salary
operation is vertically integrated. At stations organized this way, all
decision-making authority lies with the refiner. Station personnel have
no performance-based incentives to exert effort. The remaining three
organizational forms are all vertically separated, allocating to local
managers both residual claims to a station's profits and extensive
control rights. The conventional wisdom is that refiners use vertically
separated forms to take advantage of agents' local knowledge and
superior ability to pursue promotions to increase local demand (Kleit
2005). In all vertically separated forms, the local agent's
compensation is tied to a station's sales, and they must procure
their fuel (possibly indirectly) from the refiner. Given the strong
similarity across forms, I follow Shepard (1993) in focusing on the
binary difference between refiner owned and operated stations and
vertically separated ones.
The descriptions of the different organizational forms suggest that
vertical separation in gasoline retailing may affect pricing behavior
for two different reasons. (3) First, as the local agents at vertically
separated stations buy their fuel from the principal (or a middleman who
bought from the principal), their prices may be higher due to double
marginalization. Second, as conventional wisdom in the industry links
vertical separation to demand promotion, an agency model would also
predict that it would lead to higher prices. However, a key difference
between the predictions is that higher prices owing to an increase in
demand should not necessarily be linked to lower sales.
Several past papers have tested the hypothesis that prices are
higher at vertically separated stations. One set of scholars has done
this by exploiting plausibly exogenous variation in state policies. For
example. Vita (2000) shows that states with divorcement laws, which
restrict vertical integration, have higher average prices. Similarly,
Barron and Umbeck (1984) finds that gasoline stations that switched to
vertically separated forms after Maryland passed a divestiture law
increased their prices. That said, while the empirical strategy of
relying on interstate policy differences provides an accurate picture of
the implications of outlawing vertical integration, it may not provide
insight into the marginal impact of form itself. This is because the
estimates are identified off of organizational form changes at
inframarginal stations--that is, ones that an optimizing refiner would
never desire to operate under a vertically separated form--as well as
marginal stations. They thus may overstate the impact of form.
Consistent with this critique, studies addressing the endogeneity
of form using strategies other than policy changes have found more
ambiguous relationships between vertical separation and performance
among gas stations. However, these papers also are subject to some
concerns. For example, Shepard (1993) found only limited evidence of
higher prices at vertically separated stations, but addressed the
potential endogeneity of form just through the inclusion of market and
brand indicator variables. Identifying off of form changes triggered by
a merger, Hastings (2004) found that a vertically separated station had
statistically indistinguishable prices from a vertically integrated
station of the same brand. However, Taylor et al. (2010) present
evidence that the data used by Hastings may have been problematic
insofar as her main results on the merger's impact could not be
replicated using similar data from an alternative source. Taylor et al.
(2010) did not examine the issue of vertical separation in their study.
Thus, the question of how vertical separation impacts economic
behavior in the retail gasoline industry remains very much unanswered.
III. DATA
The data used in this study come from censuses of retail gasoline
markets performed by New Image Marketing. The censuses cover the Denver,
Minneapolis, Toledo, Louisville, and Washington, DC metro areas, and
were collected at uneven intervals between 1996 and 2000. Restricting
the sample to those stations affiliated with refiners leaves 4,687
station-period observations affiliated with 2,774 different unique
station locations. Observations are not evenly distributed across time
periods or states. (4) Pooling the observations, I find that
company-owned and operated stations account for 13% of the sample,
broadly in line with industry-wide summary statistics cited in Kleit
(2005).
During each census, surveyors assessed a large number of station
characteristics, including the prices of different types of gasoline,
the presence of a convenience store, the number of service bays, and the
appearance of the station. (5) In addition, the surveyors asked on-site
staff about ownership and factors such as the total volume of fuel
sales.
To account for variation in the intensity of competition faced by
different stations, I follow Hosken et al. (2008) in using the number of
stations within a 1.5-mile radius of the latitude and longitude provided
in the New Image data. (6) I further control for heterogeneous market
conditions by supplementing the information contained in the New Image
censuses with data on county population collected by the U.S. Census and
average household income (in thousands) collected by the Internal
Revenue Service. (7)
Table 1 shows descriptive statistics for all observations.
IV. ECONOMETRIC METHODOLOGY AND IDENTIFICATION
To identify the impact of vertical separation on prices and sales
volumes, I estimate a series of models with the following linearly
separable general form:
(1) [Y.sub.it] = [F.sub.it] [delta] + [X.sub.it] [lambda] +
[Z.sub.i]v + [[mu].sub.i] + [[epsilon].sub.it],
where i and t index stations and time of observation; Y is the
economic outcome of interest; [F.sub.it] indicates that station i in
time t is operated under a vertically separated form; [X.sub.it] are
time-varying station and market characteristics; [Z.sub.i], are
time-invariant station characteristics; [[mu].sub.i] represents
station-specific heterogeneity; and [[epsilon].sub.it] is the
idiosyncratic error. As in Vita (2000) and Hosken et al. (2008), I
estimate the pricing models in levels, but the results are qualitatively
the same when I employ a log-linear specification. I estimate the volume
models using a log-linear approach, which addresses the greater
dispersion of the sales variable. However, the results are qualitatively
similar when estimated in levels.
Observable controls included in X and Z are the number of
competitors, indicators for the presence of convenience stores and
service bays, a high quality appearance indicator, county population,
and county income. In the volume regressions, I also include the number
of gasoline pumps. To further address the possibility of consistently
different behavior across brands (Hosken et al. 2008) or across regions
(Pinkse and Slade 1998), all regressions also include brand-metro area
and state-date indicator variables. Including these additional sets of
indicator variables means that the estimates are identified by variation
in prices and forms within brands and inside of particular areas at
different time periods. (8)
I estimate Equation (1) in a variety of ways. These correspond to
different assumptions about the relationships between [mu], [epsilon],
the form, and the outcome variables. The worry is that something
unobservable to the econometrician--that is, either [mu] or
[epsilon]--is correlated with both organizational structure and price
(sales). By observing how the estimated impact of organizational form
may change as methods that are robust to weaker assumptions about their
conditional independence, one gains insight into the nature of any
unobservables affecting both the use of different forms and economic
outcomes.
In my first specification, I make the strong assumption that the
station-specific heterogeneity is uncorrelated with any other
explanatory information. Thus, these models simply account for
observable differences across stations, but preclude the possibility
that there are nonmeasured, station-level factors impacting both the
choice of form and economic outcomes. However, I cluster the standard
errors at the station level to account for the possibility that
different stations are exposed to shocks from different distributions.
My second estimating approach addresses the possibility of
correlation between the persistent unobserved information ([mu]) and the
observable regressors using a method that does not rely solely on within
station variation. I do this because it is comparatively rare for
stations to change forms. (9) Following Kosova et al. (2013) and Mundlak
(1978), I assume that the correlation between unobservable station
heterogeneity and the explanatory variables can be captured through a
parametric function of station-level means of the time-varying
regressors:
(2) [[mu].sub.i] = [[bar.X].sub.i][xi] + [[upsilon].sub.i],
where [[bar.X].sub.i] is the vector of station-level means of
time-varying regressors, and [[upsilon].sub.i] represents time invariant
station information that is uncorrelated with the observables. As there
are few time-varying elements in my data, I include the mean lagged
volume (price) in the price (volume) regressions as well as the
one-period lagged terms directly along with the means of county
population, income, and number of competing stations. (10) The lagged
terms can reasonably be thought to be exogenous at the time the
decisionmaker sets prices.
While the lagged terms used in the Mundlak models help control for
unobserved heterogeneity, they also mean that only stations with
multiple observations are in the sample. However, there are still 730
stations available for use. I account for o, by again allowing the
standard errors to be clustered at the station level, which is a more
general correlation structure than random effects (RE), and hence is a
conservative approach. (11)
In my third specification, I relax the Mundlak method's
parametric assumption about the nature of p, and exploit the limited
number of form switches in the data by estimating station fixed effects
(FE) models. These models fully control for possible correlation between
time invariant station-specific heterogeneity and the other variables.
However, they rely on the limited number of stations that changed form
during my sample period; I observe only 93 such changes. As a result,
they are more sensitive to unobserved shocks. Nevertheless, their
results can provide insight into the validity of the parametric
assumptions made in the Mundlak models.
While the Mundlak and FE models do control for a considerable
degree of station-specific heterogeneity, they nevertheless assume that
form choice is orthogonal to unobserved time-varying station-level
heterogeneity. Thus, the models do not permit the possibility of a
correlation between the idiosyncratic shocks, [[upsilon].sub.it], and
the choice of organizational forms. I address this concern using an IV
strategy. Like Novak and Stem (2008), Forbes and Lederman (2010), and
Kosova et al. (2013), I exploit complementarities in form choice to find
instruments that can be expected to influence the selection of form but
have no separate influence on economic behavior. Specifically, I exploit
refiners' incentive to account for their monitoring costs as well
as station and market characteristics when selecting forms.
Monitoring occurs because of the difficulty involved in writing a
self-enforcing vertically separated contract that covers all margins
impacting the principal's interests but not necessarily those of
the agent. Because principals cannot rely on agents' self-interest
to ensure full compliance, physical monitoring occurs at vertically
separated stations to ensure agents' adherence to clauses governing
factors like appearance. If monitoring costs vary at the station level,
then the utilization of different vertically separated forms should be a
function not just of when local effort is key but also the costs of
using those forms. I follow Kosova etal. (2013) in operationalizing this
insight in a retail context by assuming that the costliness of
monitoring is impacted by the forms that the refiner uses at nearby
stations (i.e., those within the same county). (12) The rationale is
that if the principal already has to send staff to check on similar
agents, then the marginal cost of assessing one more is lower than if
less monitoring already had to take place.
There is strong empirical evidence for the instruments'
validity. For example, Kosova et al. (2013) find evidence consistent
with the hypothesis that the forms of nearby affiliates influence form
choice for hotels, as does Wilson (forthcoming) using the same New Image
data for gasoline retailing. (13) Moreover, in the pricing models, 1
supplement the form instruments with the number of gasoline nozzles as
the past literature on the boundaries of the firm suggests that outlet
size is correlated with vertical integration (Lafontaine and Slade
2007). (14) This allows me to test the validity of the instrumental
variables using Hansen overidentification tests; the null that the
instruments are valid is rarely rejected overall, and is never rejected
when time invariant heterogeneity is also accounted for. Finally, I
estimated several just-identified models, including the other instrument
in the outcome equation, to further check that the instruments are
valid. In these models, the number of pumps was not found to have
significant statistical or economic effects on pricing in models that
include it as independent variable. Similarly, just-identified IV models
that rely solely on the number of nozzles to instrument for vertical
separation produce similar results to those exploiting both nozzles and
the share variable.
Due to concerns about the amount of variation in form within
stations in the data, I only apply the IV approach to the pooled and
Mundlak approaches described above. (15) As noted in Wooldridge (2002),
the IV models can consistently be estimated using standard two-stage
least squares. I continue to cluster the standard errors at the station
level in these models.
V. ECONOMETRIC RESULTS
A. Price Results
I present analyses of the prices of regular, super, and premium
quality gasolines in Tables 2-4, respectively. In each table, Column 1
shows the results of the cross-sectional model. Columns 2 and 3 address
the possible importance of time-invariant heterogeneity using the
Mundlak and FE approaches. Column 4 returns to the cross-sectional
sample but accounts for the endogeneity of form using the IV approach,
while Column 5 combines the IV and Mundlak approaches. The results of
the IV models' first stage are shown in Table B1, which indicates
that the instruments are statistically and economically significantly
correlated with vertical separation. Moreover, the F-statistics for the
first stage are reported in Tables 2-4, and they are well above the
suggested threshold of 10 (Staiger and Stock 1997), suggesting that a
weak instruments problem is unlikely.
Overall, the results consistently indicate that vertical separation
produces higher prices. The estimated effect of vertical separation is
positive in all of the models. However, the effects are of small
magnitude and not always statistically significant at conventional
levels when time-invariant heterogeneity and/or other possible
correlations between the unobserved information and form choice are not
addressed. As I use methods that are more robust to the presence of
confounding unobservables, the estimated effect of form generally grows,
implying that both time invariant and time varying unobserved factors
influence form choice as well as economic outcomes.
Several findings are particularly interesting. First, comparing the
Mundlak to the FE estimates shows that the former are always smaller in
magnitude. This is consistent with the intuitive idea that the more
efficient Mundlak approach to modeling time invariant characteristics
does not fully account for station heterogeneity associated with both
form choice and outcomes. However, the Mundlak formulation does appear
to capture a significant portion of the variability, suggesting that
this approach has significant merit when data do not permit full FE
models to be estimated.
Second, I find that the coefficients rise further when form is
instrumented for. This suggests that in addition to there being
time-invariant elements affecting both form choice and pricing, there
are also economically significant factors affecting form choice on a
time varying basis. For example, it might be that companies use certain
types of managers in certain types of locations. As the pool of
available staff changed, this could cause firms to update their
organizational form choices while also changing stations'
performances. (16)
Given that the IV-Mundlak models rely on the weakest set of
assumptions, I think that their estimates are the most credible. This
belief is buttressed by C tests of the endogenous variables in the
regressions on regular price, which are equivalent to Hausman tests
comparing an ordinary least squares (OLS) model to the IV model.
Depending on gasoline type, the Column 5 estimates indicate that
vertical separation leads to price increases of 6-9 cents. As Hosken et
al. (2008) and Kleit (2005) report that retail margins average 20 cents
or less, my estimates imply that the choice of form can change margins
by 25%-45%. It is interesting to compare these results to those of Vita
(2000), who finds that, on average, states with divorcement laws had
regular unleaded gas prices 2.6 cents higher than those without. Insofar
as 20% (or fewer) of stations are company-owned in my sample, my results
imply that the average price of regular gasoline would increase by
around 1 cent if inframarginal stations behave the same as marginal ones
and stations already operating under vertically separated forms did not
respond. This smaller estimate of the policy impact of mandatory
vertical separation is consistent with the idea that switching the form
of a vertically integrated station on the margin between the two forms
has a smaller impact than switching the form of a station that a freely
optimizing decision-maker would never separate. (17)
It is also interesting to compare my results with those of Kosova
et al. (2013), who apply a similar identification approach to studying
vertical separation in the hospitality industry. Whereas I find that
vertical separation increases prices, they find no behavioral impact of
form choice. This apparent inconsistency can be resolved by reflecting
on what agency theory says about the choice of form. When effort goes to
fostering demand--as the prior literature has suggested for the gasoline
industry--it should lead to higher prices (whatever the role of double
marginalization). However, when effort goes to reducing costs--as
Shelton (1967) (though not Graddy 1997) found, for example, in
fastfood--it should have the opposite effect. Thus, the fact that we
estimate different effects for vertical separation suggests that local
agents' effort plays a different role in gasoline retailing than in
the hotel industry.
Though not necessarily statistically significant, the results for
the control variables reported in Tables 2-4 are broadly in line with
intuition and the prior literature. For example, I generally find that
the number of nearby stations exerts downward pressure on price though
this effect is not always of statistical or economic significance.
Intuitively, these effects are smaller and less precisely identified
when time-invariant factors are controlled for since markets'
structures did not change often during my short sample. Finding that the
presence of a convenience store is negatively correlated with gasoline
price is consistent with the idea that there are demand
complementarities between retail and gasoline sales. While intuitive, I
believe this result is novel within the empirical literature (though
hypothesized in Slade 1996). Similarly, I find that service
capabilities--plausibly inversely correlated with gasoline demand--are
associated with higher prices.
Although my results imply that prices would go up if vertical
integration were prohibited, it is not clear how much--if any--consumer
welfare would be lost were vertical integration to be forbidden. This is
because the higher prices at vertically separated stations could stem
from agents' efforts to increase local demand as opposed to double
marginalization. Thus, even if the marginal estimate of form did
generalize to inframarginal stations, we cannot quantify overall welfare
effects from any policy changes. As noted in Cooper et al. (2005), this
is a familiar problem when assessing the welfare effects of vertical
separation. However, while I lack the data to quantify the relative
contributions of different factors, I can explore the issue by seeing if
output varies across organizational forms.
B. Volume Results
Table 5 presents the results of regressions considering the log of
total volume of fuel sold using the cross-sectional, Mundlak, FE, and
two IV specifications. When I make strong assumptions about the
conditional independence of form in the cross-sectional and Mundlak
models, I find that vertical separation is associated with lower volumes
of fuel sold. However, when these assumptions are relaxed, I find
effects that are much smaller in magnitude and no longer statistically
significant. Thus, the marginal vertically separated station sells
approximately the same amount of fuel as a vertically integrated station
once the endogeneity of form is controlled for even though it has higher
prices. This result is consistent with agents' effort increasing
consumer demand and doing so by a sufficiently large amount that it more
than accounts for whatever impact double marginalization may be having.
The apparent lack of change in quantity raises the question of why
a refiner would ever not use a vertically separated form at these
marginal stations. After all, it appears that vertically separated
stations' variable profits increase. There are several possible
explanations. One might be that the refiner is unable to capture any of
these returns, and hence does not take them into account when making
their form choices. Alternatively, one might think that refiners can
capture some of the higher returns, but that adopting a separated form
also creates offsetting costs to the brand. Some tentative evidence for
this explanation may be found in Wilson (forthcoming), which shows that
vertically separated stations tend to choose lower qualities.
As before, the coefficients on the controls are generally of
intuitive signs and magnitudes. For example, the presence of a
convenience store is consistently associated with higher sales volumes,
while the presence of service bays is associated with lower sales. Both
findings are consistent with hypotheses about demand complementarities
among the different products if one accepts the hypothesis that
consumers searching for service care are not simultaneously searching
for fuel. (18)
VI. CONCLUSION
I contribute to the growing literature addressing the behavioral
impact of endogenously chosen organizational forms by exploiting unique
data on gasoline stations. Across a range of specifications, the data
provide robust evidence that vertical separation increases prices
without affecting sales quantities once the presence of unobserved
confounding factors are accounted for. This suggests that vertical
separation induces local agents to expend effort in order to increase
demand. The magnitude of my estimated price increases is somewhat
smaller than previous work utilizing state-level variation in the costs
of vertical integration to identify the impact of organizational form.
These identify effects off of inframarginal as well as marginal
stations, and thus may overstate the impact of form itself. The finding
that at least part of the price difference reflects efforts to increase
demand complicates welfare analysis. Overall, the paper's results
should be of interest not just to scholars but also policy-makers as
many refiners are making the decision to get out of direct operation of
retail gasoline stations (MSNBC 2008).
ABBREVIATIONS
FE: Fixed Effects
IV: Instrumental Variables
OLS: Ordinary Least Squares
RE: Random Effects
doi: 10.1111/ecin.12203
APPENDIX A: DATA
A1 NEW IMAGE VARIABLE DESCRIPTION
Below, I provide the name and description provided by New Image of
those variables used in the analysis and the method by which they were
transformed (if appropriate).
* Organizational Form: Categorical variable corresponding to the
answer to the following question. TYPE OF OPERATION)(TOO)--Overall
status of operation, ask respondent to identify:
1. No building or doesn't sell gasoline
2. Lessee dealer building and facility owned by major/non major oil
company, business owned by dealer. [I reordered this as Type 2.]
3. Salary operation building and facility owned by major/non major
oil company. Personnel paid by company. [I reordered this as Type 1, so
that salaried operations represented the baseline.]
4. Open Dealer--Land and operation owned by individual who is
supplied product by major/non major oil company.
5. Jobber/Wholesaler Operation owned by a local company that owns
several operations in the area. (EXP distributor) or a franchise/chain
organization (EXP a convenience store chain)
* Regular Unleaded Price: Numerical variable corresponding to
non-constrained answer to the following question. OCT REGULAR
UNLEADED)(UO)--Price Reg Unleaded)(RUP)
* Super Unleaded Price: Numerical variable corresponding to
non-constrained answer to the following question. OCT MIDGRADE
UNLEADED)(MO)--Price mid UnleadedXMUP)
* Premium Unleaded Price: Numerical variable corresponding to
non-constrained answer to the following question. OCT SUPERXSO)--Price
Super Unleaded)(PUP)
* Volume: Numerical variable corresponding to nonconstrained answer
to the following question. MONTHLY VOLUME)(GV)--Enter average number of
gallons sold in one month, (last completed month)
* C-Store: Dummy variable which takes value of 1 if an answer other
than 0 chosen for the following question. INTERIOR C-STORE
APPEARANCEKINAP). As it appears to consumer.
* No snack shop
* Outstanding (top 10%)
* Excellent
* Better than average
* Equal to average
* Below average
* Poor
* Unacceptable (bottom 10%)
* Service Bays: Dummy variable which takes value of 1 if a number
other than 0 chosen for the following question. SERVICE
BAYSXNOSB)--Total number of service bays. If not in operation mention in
comments.
* Appearance: Dummy variable which takes value of 1 if the answer
to the following question takes the value of 1 or 2. APPEARANCE OF
BUILDING)(AOB)
* N/A
* Outstanding (top 10%)
* Excellent
* Better than average
* Equal to average
* Below average
* Poor
* Unacceptable (bottom 10%)
* Nozzles: Numerical variable corresponding to nonconstrained
answer to the following question. GASOLINE NOZZLES)(GN)--Total number of
gasoline only nozzles. Do not include diesel or kerosene.
APPENDIX B: ADDITIONAL TABLES
TABLE B1
First Stage of IV Models
Full Mundlak
b/se b/se
Share Separated 0.281*** 0.420***
(0.049) (0.069)
Nozzles -0.004*** -0.001
(0.001) (0.002)
Competitors 0.001 0.001
(0.001) (0.003)
C-Store 0.022* 0.030*
(0.011) (0.016)
Service Bays 0.090*** 0.058***
(0.014) (0.021)
Appearance -0.098*** -0.041
(0.015) (0.030)
Population 0 0.002
(0.000) (0.002)
Income 0 -0.004
(0.001) (0.004)
Lagged Volume 0
(0.001)
Brand-Metro Effects Yes Yes
State-Date Effects Yes Yes
Observations 4,298 1,615
Notes: Table presents the estimated coefficients ("b") and standard
errors ("se") of the different models. All models contain
state-date and brand-metropolitan area fixed effects.
* p < 0.10, ** p < 0.05, *** p < 0.01.
TABLE B2
Price Effects of "Vertically Separated" Indicator Variable for
Additional Models
Type of Full Brand-State County-Date IV County
Gasoline Categories Date
Regular 1.097*** 1.815*** 1.820*** 1.168***
(0.358) (0.492) (0.528) (0.409)
Super 1.906*** 2.208*** 1.115*** 1.770***
(0.590) (0.656) (0.361) (0.515)
Premium 1.911*** 4.326** 6.381*** 5.341**
(0.569) (1.811) (2.319) (2.462)
Notes: Table presents the estimated coefficients ("b") and standard
errors ("se") of the different models. All models estimated using
the Mundlak sample and controls.
* p < 0.10, ** p < 0.05, *** p < 0.01.
REFERENCES
Barron, J. M., and J. R. Umbeck. "Effects of Different
Contractual Arrangements: The Case of Retail Gasoline Markets."
Journal of Law and Economics, 27(2), 1984, 313-28.
Brickley, J. A. "Incentive Conflicts and Contractual
Restraints: Evidence from Franchising." Journal of Law and
Economics, 42(2), 1999, 745-74.
Brickley, J. A., and F. H. Dark. "The Choice of Organizational
Form: The Case of Franchising." Journal of Financial Economics,
18(2), 1987, 401-20.
Cooper, J. C., L. M. Froeb, D. O'Brien, and M. G. Vita.
"Vertical Antitrust Policy as a Problem of Inference."
International Journal of Industrial Organization, 23(7), 2005, 639-64.
Forbes, S. J., and M. Lederman. "Does Vertical Integration
Affect Firm Performance? Evidence from the Airline Industry." RAND
Journal of Economics, 41(4), 2010, 765-90.
Graddy, K. "Do Fast-Food Chains Price Discriminate on the Race
and Income Characteristics of an Area?" Journal of Business and
Economic Statistics, 15(4), 1997, 391-401.
Hastings, J. S. "Vertical Relationships and Competition in
Retail Gasoline Markets: Empirical Evidence from Contract Changes in
Southern California." American Economic Review, 94(1), 2004.
317-28.
Holmstrom, B., and P. Milgrom. "Multitask Principal-Agent
Analyses: Incentive Contracts, Asset Ownership, And Job Design."
Journal of Law, Economics, and Organization, 7(special issue), 1991,
24-52.
Hosken, D. S., R. S. McMillan, andC. T. Taylor. "Retail
Gasoline Pricing: What Do We Know?" International Journal of
Industrial Organization, 26(6), 2008, 1425-36.
Kleit, A. N. "The Economics of Gasoline Retailing: Petroleum
Distribution and Retailing Issues in the U.S Energy Studies Review,
13(2), 2005, 1-28.
Kosova, R., F. Lafontaine, and R. Perrigot. "Organizational
Form and Performance: Evidence from the Hotel Industry." Review of
Economics and Statistics, 95(4), 2013, 1303-23.
Laffont, J. J., and D. Martimort. The Theory of Incentives: The
Principal-Agent Model. Princeton, NJ: Princeton University Press, 2002.
Lafontaine, F., and M. Slade. "Vertical Integration and Firm
Boundaries: The Evidence." Journal of Economic Literature, 45(3),
2007, 629-85.
Meyer, D. and J. Fischer, "The Economics of Price Zones and
Territorial Restrictions in Gasoline Marketing." Federal Trade
Commission Bureau of Economics Working Paper, 2004.
MSNBC. Exxon to Sell All of Company's Gas Stations. 2008.
Accessed January 15, 2014. http://www.nbcnews.
com/id/25126563/ns/business-oil_and_energy/t/exxonsell-all-companys-gas-stations/#.VL0MIyyrGM0
Mundlak, Y. "On the Pooling of Time Series and Cross Section
Data." Econometrica, 46(1), 1978, 69-85.
Novak, S., and S. Stern. "How Does Outsourcing Affect
Performance Dynamics? Evidence from the Automobile Industry."
Management Science, 54(12), 2008, 1963-79.
Pinkse, J.. and M. E. Slade. "Contracting in Space: An
Application of Spatial Statistics to Discrete-Choice Models."
Journal of Econometrics, 85(1), 1998, 125-54.
Shelton, J. P. "Allocative Efficiency vs.
'X-Efficiency': Comment." American Economic Review,
57(5), 1967, 1252-8.
Shepard, A. "Contractual Form, Retail Price, and Asset
Characteristics in Gasoline Retailing." RAND Journal of Economics,
24(1), 1993, 58-77.
Slade, M. E. "Multitask Agency and Contract Choice: An
Empirical Exploration." International Economic Review, 37(2), 1996,
465-86.
Spengler, J. J. "Vertical Integration and Antitrust
Policy." Journal of Political Economy, 58(4), 1950, 347-52.
Staiger, D., and J. H. Stock. "Instrumental Variables
Regression with Weak Instruments." Econometrica, 65(3), 1997,
557-86.
Taylor, C. T., N. M. Kreisle, and P. R. Zimmerman. "Vertical
Relationships and Competition in Retail Gasoline Markets: Empirical
Evidence from Contract Changes in Southern California: Comment."
American Economic Review, 100(3), 2010, 1269-76.
Vita, M. G. "Regulatory Restrictions on Vertical Integration
and Control: The Competitive Impact of Gasoline Divorcement
Policies." Journal of Regulatory Economics, 18(3), 2000, 217-33.
Wilson, N. E. "The Impact of Vertical Contracting on Firm
Behavior: Evidence from Gasoline Stations." US Federal Trade
Commission, Bureau of Economics Working Paper No. 310, 2012.
--. Forthcoming. "Local Market Structure and Strategic
Organizational Form Choices: Evidence from Gasoline Stations."
International Journal of the Economics of Business.
Wooldridge, J. M. Econometric Analysis of Cross Section and Panel
Data. Cambridge, MA: MIT Press, 2002.
(1.) Lafontaine and Slade (2007) survey many other papers also
finding this relationship.
(2.) The precise definitions are shown below in Appendix A. For
more discussion of the different forms, see also Kleit (2005), Meyer and
Fischer (2004), Shepard (1993), and Slade (1996).
(3.) An earlier version of this paper formally demonstrated this
using a theoretical model (Wilson 2012).
(4.) Consistent with the fact that the different states have
different laws affecting refiners' ability to own and operate
stations, the relative usage of the organizational forms in the
different market areas varies significantly. Details on the breakdown
can be seen in Wilson (2012).
(5.) The station characteristics were generally recorded as
categorical variables. For the sake of parsimony, I transform these into
binary indicator variables. However, results for the variables of
interest are robust to including indicator variables for all categories
as can be seen below in Table B2. The precise definitions for the New
Image controls are given below in Appendix A as well as how they were
transformed for use in the econometric analysis. In addition, I
generated qualitatively similar results using the New Image variable
capturing the number of branded competitors visible from the station
instead of the number of stations in the zipcode.
(6.) The point estimates in the regression models are almost
identical when I simply use the number of stations within a zipcode.
(7.) See http://www.census.gov/popest/counties/ and
http://www.irs.gov/taxstats/article/0..id = 120303,00.html,
respectively.
(8.) To a large degree, states and metropolitan areas overlap in
the data. The only exception is for the Washington, DC area, where the
metro area includes all of the observations from DC, VA, and MD. The
results are qualitatively similar when I include more and less
parsimonious sets of controls. In particular, employing county-date or
brand-state indicator variables does not change the results as can be
seen in Table B2.
(9.) Station characteristics generally do not change over time and
market conditions evolve slowly. For example, in the data used in the
analysis, the coefficient of variation for the dummy indicating the
presence of service bays is less than 0.18.
(10.) As noted above, I sometimes observe stations at 2-year
intervals. In such cases, the lagged term is simply the most recent
previous observation. This clearly injects a certain amount of noise
into the variable. Insofar as this makes it less likely to capture
something meaningful, I see the approach as conservative.
(11.) In practice, the choice between clustering and RE has almost
no effect on the estimates or their statistical significance. Details
are available upon request.
(12.) I focus on the number of affiliated outlets in the county as
opposed to zipcode, because I believe that conditional on traveling from
their headquarters to a given county, it costs salaried employees of the
principals little to travel between zipcodes to monitor different
stations. Consistent with this argument, my results are qualitatively
similar when I focus on brands' shares at the zipcode level.
(13.) In addition, I examined the extent of geographic variation in
forms' usage. I found non-negligible variation within and across
metropolitan areas (within and across brands) in the extent to which
vertically separated forms are used. Maps showing this are available
upon request.
(14.) To check that scale was not associated with pricing, I
experimented with models that did not use it as an instrument, and
consistently found that my results were qualitatively unchanged.
(15.) IV models that included FE were generally of the predicted
sign and magnitude. I do not report these results as their precision is
highly dependent upon whether or not the full sets of FE controls are
included.
(16.) Furthermore, the conclusion that vertical separation induces
an increase in prices is robust to variation in empirical specification.
This can be seen in Table B2. which shows the impact of vertical
separation on the different price variables for four additional models.
All of the estimates are in line with those presented above. Column 1
shows the results of Mundlak models with controls for all categories of
station characteristics to address any concern that some categories are
disproportionately linked to different form types. Column 2 examines
whether brands behave differently within the different market areas by
including brand-state--as opposed to brand-metropolitan area--fixed
effects. The difference between state and metropolitan area only affects
VA, DC, and MD. Column 3 shows the results of Mundlak models that
include more granular county-date level controls for market
heterogeneity. Finally, Column 4 shows the results of Mundlak-IV models
with county-date controls. Although not shown, I also estimated similar
models that excluded certain states' observations. In particular,
models that dropped DC and MD produced analogous results as did ones
relying only on VA stations.
(17.) These back-of-the-envelope comparisons rely on nominal
prices.
(18.) Alternatively, Shepard (1993) notes that newer stations tend
not to have service bays. If consumers prefer newer stations all else
equal when shopping for gas. this could produce the estimated negative
correlation between sales volume and service bays.
NATHAN E. WILSON, The opinions expressed here are those of the
author and not necessarily those of the Federal Trade Commission or any
of its Commissioners. An earlier version of this paper previously
circulated with the title "The Impact of Vertical Contracting on
Firm Behavior: Evidence from Gasoline Stations." I am extremely
grateful for suggestions from the editor, Luke Froeb. and those of the
anonymous referees. In addition, this paper benefited tremendously from
comments from Dan Hosken as well as other FTC colleagues, Itai Ater,
Renata Kosova, Scott Masten. Tom Lyon, and attendants of the 2011
International Industrial Organization and 2012 Western Economic
Association Conferences. The usual caveat applies. Wilson: Bureau of
Economics, Federal Trade Commission, 600 Pennsylvania Ave, Washington.
DC 20580. Phone 2023163485. E-mail
[email protected]
TABLE 1
Descriptive Statistics
Variable Observations Mean Standard Minimum Maximum
Deviation
Regular 4,299 116.71 13.73 79.90 167.90
Super 4,296 126.70 13.15 86.90 186.90
Premium 4,299 134.86 12.52 88.90 193.90
Volume 4,535 104.99 50.32 10.00 400.00
Competitors 4,687 9.32 5.68 0.00 32.00
(C-Store) 4,687 0.72 0.45 0.00 1.00
(Service Bays) 4,687 0.39 0.49 0.00 1.00
(Appearance) 4,687 0.15 0.35 0.00 1.00
Nozzles 4,535 18.13 9.95 2.00 60.00
Pop. COOOs) 4,687 619.02 294.32 40.99 1109.63
Income ('000 s) 4,687 57.87 14.59 35.47 96.69
TABLE 2
Regression of the Price of Regular Unleaded Gasoline on Organizational
Form
OLS Mundlak FE
b/se b/se b/se
Separated 0.452 1.267*** 2.399*
(0.281) (0.378) (1.449)
Competition -0.064*** 0.009 -0.055
(0.016) (0.057) (0.049)
C-Store -0.688*** -0.261
(0.208) (0.253)
Service Bays 0.587*** 0.471*
(0.201) (0.253)
Appearance -0.395** 0.921
(0.183) (0.296)
Population 0 -0.021 0.009
(0.000) (0.030) (0.014)
Income 0.104*** 0.054 0.102***
(0.008) (0.083) (0.026)
Lag Volume 0.007
(0.011)
State-Date Effects Yes Yes Yes
Brand-Metro Effects Yes Yes Yes
Observation 4,299 1,616 4,299
F Separated
IV IV-Mundlak
b/se b/se
Separated 5.998*** 5.556***
(1.592) (1.797)
Competition -0.069*** 0.001
(0.017) (0.057)
C-Store -0.814*** -0.527**
(0.220) (0.258)
Service Bays 0.027 0.189
(0.260) (0.278)
Appearance 0.173 -0.777**
(0.250) (0.320)
Population 0 -0.04
(0.000) (0.031)
Income 0.101*** 0.057
(0.008) (0.083)
Lag Volume 0.006
(0.011)
State-Date Effects Yes Yes
Brand-Metro Effects Yes Yes
Observation 4,298 1,615
F Separated 30.356 18.459
Notes: Table presents the estimated coefficients ("b") and standard
errors ("se") of the different models. Mundlak models all contain
means of lagged volume, number of competitors, population, and
income. All models contain state-date and brand- metropolitan area
fixed effects.
* p <0.10, ** p<0.05, *** p<0.01
TABLE 3
Regression of the Price of Super Unleaded Gasoline on
Organizational Form
OLS Mundlak FE
b/se b/se b/se
Separated 0.652* 2.043*** 4.079*
(0.349) (0.540) (2.238)
Competition -0.038* 0.073 -0.135**
(0.021) (0.069) (0.067)
C-Store -0.031 -0.063
(0.281) (0.316)
Service Bays 0.910*** 0.543*
(0.271) (0.315)
Appearance 0.205 -0.947**
(0.232) (0.413)
Population 0 -0.045 -0.006
(0.000) (0.033) (0.015)
Income 0.142*** 0.185* 0.083***
(0.010) (0.102) (0.032)
Lag Volume -0.005
(0.014)
State-Date Effects Yes Yes Yes
Brand-Metro Effects Yes Yes Yes
Observation 4,296 1,616 4,296
F Separated
IV IV-Mundlak
b/se b/se
Separated 2.726 8.557***
(1.923) (2.283)
Competition -0.038* 0.061
(0.020) (0.069)
C-Store -0.11 -0.473
(0.284) (0.326)
Service Bays 0.681** 0.11
(0.326) (0.349)
Appearance 0.399 -0.731
(0.289) (0.460)
Population 0 -0.075**
(0.000) (0.032)
Income 0.141*** 0.190*
(0.010) (0.101)
Lag Volume -0.006
(0.014)
State-Date Effects Yes Yes
Brand-Metro Effects Yes Yes
Observation 4,295 1,615
F Separated 30.356 18.459
Notes: Table presents the estimated coefficients ("b") and standard
errors ("se") of the different models. Mundlak models all contain
means of lagged volume, number of competitors, population, and
income. All models contain state-date and brand- metropolitan area
fixed effects.
* p < 0.10, ** p < 0.05, *** p < 0.01.
TABLE 4
Regression of the Price of Premium Unleaded Gasoline on Organizational
Form
OLS Mundlak FE
b/se b/se b/se
Separated 0.559 2.171*** 2.59
(0.368) (0.598) (2.740)
Competition -0.083** -0.053 -0.185***
(0.022) (0.072) (0.068)
C-Store -0.37 -0.179
(0.295) (0.371)
Service Bays 0.960*** 0.687**
(0.267) (0.339)
Appearance -0.226 -0.915*
(0.250) (0.484)
Population 0 0.011 0.015
(0.000) (0.039) (0.020)
Income 0.158*** 0.084 0.046
(0.011) (0.109) (0.0370
Lag Volume 0.003
(0.016)
State-Date Effects Yes Yes Yes
Brand-Metro Effects Yes Yes Yes
Observation 4,299 1,616 4,299
F Separated
IV IV-Mundlak
b/se b/se
Separated 3.545* 7.710***
(2.118) (2.458)
Competition -0.083*** -0.062
(0.021) (0.071)
C-Store -0.476 -0.563
(0.296) (0.366)
Service Bays 0.635* 0.294
(0.326) (0.357)
Appearance 0.057 -0.748
(0.323) (0.510)
Population 0 -0.019
(0.000) (0.036)
Income 0.157*** 0.091
(0.011) (0.107)
Lag Volume 0.003
(0.016)
State-Date Effects Yes Yes
Brand-Metro Effects Yes Yes
Observation 4,298 1.615
F Separated 30.356 18.459
Notes: Table presents the estimated coefficients ("b") and standard
errors ("se") of the different models. Mundlak models all contain
means of lagged volume, number of competitors, population, and
income. All models contain state-date and brand- metropolitan area
fixed effects.
* p<0.10, ** p < 0.05, *** p < 0.01.
TABLE 5
Regression of (Log) Sales Quantities on Organizational Form
OLS Mundlak FE
b/se b/se b/se
Separated -0.256*** -0.155*** -0.024
(0.029) (0.045) (0.063)
Competition 0.001 0.001 0.007***
(0.001) (0.005) (0.002)
C-Store 0.100*** 0.087***
(0.019) (0.028)
Service Bays -0.111'** -0.135***
(0.018) (0.025)
Appearance 0.215*** 0.119*'*
(0.018) (0.037)
Nozzles 0.027*** 0.022***
(0.001) (0.001)
Population 0.000*** 0 0
(0.000) (0.003) (0.002)
Income -0.001 0.012* 0.001
(0.001) (0.006) (0.003)
Lag Price 0.006
(0.004)
State-Date Effects Yes Yes Yes
Brand-Metro Effects Yes Yes Yes
Observation 4,298 1,383 4,298
F Separated
IV IV-Mundlak
b/se b/se
Separated -0.061 -0.049
(0.199) (0.172)
Competition 0.001 0
(0.001) (0.005)
C-Store 0.094*** 0.082***
(0.020) (0.028)
Service Bays -0.129*** -0.143"*
(0.026) (0.029)
Appearance 0.234*** 0.128***
(0.027) (0.038)
Nozzles 0.027*** 0.022***
(0.001) (0.002)
Population 0.000*** 0
(0.000) (0.003)
Income -0.001 0.012'
(0.001) (0.006)
Lag Price 0.006
(0.004)
State-Date Effects Yes Yes
Brand-Metro Effects Yes Yes
Observation 4,298 1,383
F Separated 33.402 34.949
Notes: Table presents the estimated coefficients ("b") and standard
errors ("se") of the different models. Mundlak models all contain
means of lagged price, number of competitors, population, and
income. All models contain state-date and brand- metropolitan area
fixed effects.
* p < 0.10, ** p< 0.05, *** p< 0.01.
