Вы находитесь на странице: 1из 22

Abstract

Minimum-distance entry regulation and social welfare

Carlos Cassó
Universitat de Barcelona

September 5, 2010
1 Introduction

Mankiw and Whinston [1986] proves that markets of homogeneous products with free entry
are likely to be in a socially non-e¢ cient equilibrium if business stealing e¤ect is present.
Hence, the prevailing entry regulation policy on markets that are potentially socially in-
e¢ cient is of public interest and a matter of concern for policy makers. The aim of this
paper is to develop and estimate a structural spatial model of entry, in which there is an
entry restriction to new …rms based on the distance among the geographical locations of the
incumbents. By relating the consumers utility to the entry decision of the …rms, these kinds
of policies can be evaluated in terms of social welfare.

The estimation of structural models of imperfect information has been proved to be a


convenient tool for the entry policy analysis. There is a large part of the empirical industrial
organization literature that has been devoted to shed light to this issue. Models like Berry
and Waldfogel [1999] and Davis [2006a] proposes an empirical methodology to captures the
e¤ects of the number of …rms over social welfare. This approaches are useful to evaluate entry
policies that use as a control variable the number of …rms. Even though, to our knowledge
there are no extensions to these models that includes explicitly an alternative way to analyze
entry policies based on a minimum distance between the locations of the …rms.

These restrictions are popular in the drugstore sector in Europe [ÖBIG 2006] and Aus-
tralia [PSA 2010], while in the United States and Canada prevails a free entry regime. As
Borrel and Fernandez-Villadangos [2009] describes, there is an on-going discussion in Europe
on whether the present policies serves to public interest or if they are the result of private
e¤orts to ensures regulatory rents.

This paper provides an extension to spatial models of entry in the line of Gowrisankaran
and Krainer [2006], incluiding the geographical regulation policy into the available informa-
tion for the players regarding the actions of the rest, i.e. a …rm knows that, according to the
regulation, if it decides to enter in a given location the probabiliy that others competitors
choose to enter in its surrounding at a distance smaller than the minimun set by the policy
is zero.The way of introducing the entry restriction in the decision of the players is rather
simple, but novel to these models and gives a tool for the evaluation of policies of this nature.

As an application of the method proposed here, we analyze the consequences of modifying


the entry regulation of the drugstores in Pamplona (Navarra, Spain). Before year 2000, the
minimum distance for opening a new pharmacy was 250 mt and 104 drugstores were opened.
On year 2000, the minimum distances required for starting a new drugstore declined to 150

1
mt, presumably causing that, 86 …rms join the market unitl 2008, (a growth rate of 82:69%
). The fact that the reduction of the minimum distance is followed by a dramatic increase in
the number of pharmacies gives us an interesting scenario for policy evaluation. Using the
methodology proposed here, we’ll try to analyze if the modi…cation of the policy leads to a
more desirable situation in terms of social welfare or the contrary.

For the estimation of the structural parameters a two-step estimation following the
methodology of Bajari, Hong, Krainer and Nekipelov [2009] is proposed. On the …rst stage
we use a multivariate kernel smoothing regression with a Gaussian kernel for estimate the
distribution of the choice probabilities conditioned to the states variables. Later on, the
vector of choice probabilities obtained in the …rst stage is used to recover the structural
parameters maximizing a pseudo-likelihood function. We discuss the conditions in which the
identi…cation of the parameters is posible and gives Monte Carlo evidence for the consistency
of the estimator.

This paper is organized as follows. In section 2 an entry model with minimum-distance


restriction into the choice-probabilities is developed. The model is speci…ed explicitely for
a drugstore market with a constant price margin, but by no means is this methodology is
exclusive for this types of markets. In Section 3 the assumptions that assure the identi…ca-
tion and the estimation of the structural parameters of the model are discussed. Section 4
resents a summary of the dataset used and the particularities of the policy evaluation sce-
nario. Section 6 dicusses and presents the results of the estimation {montecarlo?}. Section
7{simulations!}. Section 8 {Remarks}. Finally, the Appendix A discusses the interpolation
methods applied to the original dataset.

2 Entry model for drugstores

2.1 Free entry model with constant price margin

Consider a geographical space (a country, a state, a city,...etc.) with a …nite number of


locations for a drugstore to enter, ` = 1; :::; L. Each location in ` has a set of demographic
characteristics X = X1;:::; XL that are of common knowledge to all the …rms. Also, assume
that there are L identically and independentely distributed (i.i.d.) vectors of variables which
are private information of the …rms about all the possible locations. The …rms simultaneously
decides wether to enter ` = 1 or not ` = 0 on a particular location.

2
Consider the following functional form to represent the utility that a consumer located
in i 2 ` obtains for going to a hypothetical drugstore located in j 2 ` :

(1) ui;j = + di;j + dhj + i;j

where represents the consumers bene…t from accessing to a drugstore, is the change
in the utility related to the distance dij 2 X between his location i and drugstore in j, is
the e¤ect of having a health center nearby the drugstore (dhj 2 X is the distance from j to
the nearest health center) and ij is a vector of i.i.d. shocks that are type I extreme value
distributed. Hence, the expected quantity formula for consumer i on location j is given by
the standard multinomial logit expression:

+ di;j + dh
e j
(2) si;j ( j = 1; j j X; ) =
X
L
+ dij + dh + di;j + dh
1+e j + je
ij

j=1

j represents the vector of the entry decisions of the potential competitors and =
f ; ; g. Assuming a constant price margin normalized to 1, the expected pro…ts are repre-
sented as follows:

8
< X
L
>
E[si;j ( = 1; j X; )] pi Fj if =1
(3) E[ j (aj; a j ; X; "j ; e] = j j j
>
: i=1
0 if j =0

where pi 2 X is the population of the i location. Fj is the cost for a …rm to operate in j.
Taking into account that …rms may have private information about each location, the cost
function can be modeled as:

(4) Fj = c + s" "j

c is a …xed cost that is public information and "j a private information shock with s"
standard deviation and e is the set of parameters of the expected pro…ts function. In
(3); since the …rm does not know the private information shocks, " j for the others players,
it’s actions are driven by the expected actions of the other players j ( j j X). De…ning
the deterministic part of the expected payo¤ as in Bajari et al. [2009]:

3
X
L
(5) j( j = 1; X; e) = E[si;j ( j = 1; j j X; )] pi c
i=1

and taking into account that if j = 0 the pro…ts are set equal to zero, the optimal entry
decision satis…es,
n o
(6) j( j = 1 j X) = Pr "j j j( j = 1; X; e) + "j >0

With this vector of choice probabilities and the estimation of e, one can infer the con-
sumers welfare and the …rms pro…ts as in Gowrisankaran and Krainer [2006]. Inasmuch, in
this case the price margins are set equal to one and it’s not possible to convert the measure
of the utility nor the costs to money (an extension that includes prices in the utility and
the expected pro…ts function it’s straightforward). Thus, a further welfare analysis must be
based on rates of changes rather than monetary units.

2.2 Minimum-distance regulation

Recalling from (3), the expected pro…ts are a function of the sum of expected market shares
per population. Because the …rms don’t know with probabiity one what will be the entry
decision of the others players, their individual expected pro…ts are not a function of the
direct action of other competitros a j but of it’s choice probailities j( j j X) by the
expected shares:

h
e + di;j + dj
(7) si;j ( j = 1; j j X; ) =
X L
+ di;j + dh + di;j + dh
1+e j + j( j j X)e j

j=1

Consider now a regulation which states that a …rm can enter in a given location j if,
and only if djn r where djn represents the geographical distance between location j and
the location of the nearest competitor n, while r is the minimum distance set by the policy
maker to restric the entry. This type of regulation can be introduced in the formation of the
choice probabilities about the entry decision of the j competitors implied by (6):
( n o
Pr " j j j( j = 1; X; e) + " j > 0 if dj; j r
(8) j( j = 1 j X) =
0 if dj; j <r

4
3 Estimation and identi…cation

It has been proven by Bajari et al. [2009] that the structural parameters e can be estimated
if the following assumpstions holds:

A1 the error terms "j are i.i.d. across actions and agents. Furthermore, the parametric
form of the ditribution comes from a know family.
A2 for al j and all j and X, E[ i (aj = 0; a j ; X] = 0:

A2 is satis…ed by the assumption that the non-entry decision implies zero pro…ts. For
A1 to hold it is necessary to take the assumption that the private information shocks "j are
i.i.d. across locations and across actions. Also the choosing of a reasonable distribution for
the shocks is needed. As the actions aj is a binary variable, assuming a type I extreme value
distribution is a convenient solution.

With these assumptions, the equation for determining the choice probabilities j( j =
1 j X) (6) turns into the probabilities given by the single agent multinomial logit :

e)
j ( j =1;X;
e e;
(9) j( j = 1 j X) = e)
= j( j( j = 1 j X))
1+e j ( j =1;X;

Given the interdependence of the choice probabilities1 ; a two-step estimation approach


is proposed for the estimation of the structural parameters e and the choice probabilities
j ( j = 1 j X). See Aguirregabiria and Mira [2007], Bajari, Levin and Benkard [2004],
Pakes, Ostrovsky and Berry [2004] and Pesendorfer and Schmidt-Dengler [2003].

Step 1 Observing the entry decisions on the locations ` = 1; :::; L and the respectively
states variables X in each location, one can make a consistent estimation bj ( j = 1 j X)
of j ( j = 1 j X) by using a binary response estimator. One could use a non-parametric
estimation (sieve estimator, kernel smoothing, local polynomial, etc.).

Step 2 Using bj ( j = 1 j X) from the previous step, one can use it to estimate the e
parameters in (9) by maximizing the pseudo-likelihood function using the policy rule stated
in 8, j (e; bj ( j = 1 j X)) :
1
i.e. the decision of a …rm to enter in j depends on the choice probabilities of the rest of the players -j,
and at the same time that the entry decision of -j depends on the choice probabilities of j.

5
b 1X
(10) eP M L = arg max log( j (e; bj ( j = 1 j X)))
e n
b
If we assume that the estimates in eP M L are not sensibles to changes in the policy rule,
b
one can identify the change in the choice probabilities by computing j (epseudo M L ; bj ( j =
1 j X)) for di¤erent values of the minimun-distance parameter r:

A problem that may arise is that for di¤erent values of regulation parameter r we found
the same choice probabilities, in that case the identi…cation is not posible. This will be
the case if the estimated pro…ts are not to variable enough among locations; for example if
the …xed cost estimate bc to high or to low it doesn’t matter how much the variation of the
demand drivers is, the entry decision will roughly be the same for all locations2 .

We avoid the problem of multiple equilibria because the log-likelihood of the standard
logit model is globally concave in the expected pro…ts (thus, for a given set of parameters
there is only one vector of choice probabilities3 ). However, as Aguirregabiria and Mira
[2007] appoints, the assymptotic variance of the estimates on the second step depends on
the variance of the nonparametric estimator used on the step 1. Therefore, the estimate can
be ine¢ cient for a large variance in bj ( j = 1 j X): If this is the case we can use a method
similar to the recursive approach proposed by Aguirregabiria and Mira [2007].

Furthermore, we can also identify the change over a messure of the social welfare in terms
of "quantities"; which in our context are interpreted as number of visits to a drugstore by
the consumer. From market shares (2), the number of visits can be computed using the
numerator expression by the number of peoples in each location, the sum of all the visits
is the social pro…t for a given regulation regime. The social cost is simply the number
of expected …rms multiplied by the estimated …xed cost in terms of quantities. The social
welfare is the di¤erence of the social pro…t and the social cost:

X
L
(11) W = pj qj c E[N ]
j=1

2
Consider the extreme case with 0 …xed cost, in this case all the …rms will decide to enter. And no matter
how much we change r, we allways will found the same decision from all the players and identi…cation will
be imposible.
3
Note that the contrary does not holds. One can have a vector of choice probabilities that comes from
di¤erent sets of parameters, in that case there is a problem of not identi…cation.

6
where pj is the population, qj the quantities (number of visits to the drugstores), c
the …xed cost estimate and E[N ] the expected number of …rms (the sum of the choice
probabilities).

4 Data

For an ilustration of the method proposed here we use the geographical coordinates of the
190 pharmacies on Pamplona (Navarra, Spain) (fuente?) and the coordinates of the18 health
centers of the city given by the Pamplona’s spatial data infrastructure4 . The coordinates of
the pharmacies are associated with a dummy variable that indicates if the entry was before
or afther year 2000, when the deregulation process started. The population for the census
districts of the city is given by the spanish statistical institute, we grid the coordinates of
the city of Pamplona in regular hexagons and interpolate the data of the population in order
to have a more local dimension for competition among …rms (see the Appendix for more
details about the gridding and the interpolation methodology). On table 1 we summarize
the behavior of the variables in consideration:

Table 1: Descriptive Statistics

Statistics Pharmaciesentry<=2000 Pharmaciesentry>2000 Population Health Centers


Mean .028 .023 48.15 .004
Std. Dev. .161 .151 51.88 .070
Min. .000 .000 2.06 .000
Max. 1.00 1.00 341.71 1.00
Sum 98 86 177,482.97 18

Note: These values correspond to the 3,686 potential locations created with the gridding.
* There are 6 pharmacies excluded due to the dimensions of the hexagons in the gridding,
see the Appendix for more detail.

4
http://ide.pamplona.es

7
Our intention is to estimate to what extent the deregulation process in Pamplona from
year 2000 to 2008 is convenient to the society. Hence, our relevant players are the potential
entrants from the moment the policy is modi…ed to 2008. In that sense, we shall distinguish
…rms that have already enter in the market from …rms that still haven’t decide wether to
enter or not. A away to meet this particular characteristic of our dataset is to set the choice
probabilities of the …rms that have already enter the market to 1, as the model has none
entry cost this is not an inconvenient. Also, we can use it on the step 1 estimation as a state
varable in order to have a more accurate approximation of the decision of the new …rms.
Taking into account this modi…cation, we are going to estimate the choice probabilities of
the rest of the players given that this …rst group have already decided to enter (and the
minimun-distance policy is already implemented to their sorroundings). So, the decision of
the pre-regulation pharmacies can be considered like a pseudo-state variable, in the sense
that their decision is known by the relevant players but they also represents a component of
the competition.

We ignore the fact that the deregulation process also introduces other types of change
in the entry regulation in order to isolate the e¤ect of the minimun-distance policy. For
a full characterization of the forecasted scenario it would be necesary to include in the
model a mechanism to identify simultaneously the e¤ect of the minimun-distance policy and
other kinds of restrictions (…xed price margins vs ‡exibles or mixed price margins, maximun
number of entries, etc.). Ignoring the other instruments that the policy maker introduces
can be a source of imprecission to the estimate of the social welfare; but, as long as the
policy parameter r have a non-negligible e¤ect over the actions of the player, the direction
that the social welfare takes would be the same5 .

5 Results
5.1 Step I. Choice probabilities estimation
A nonparametric procedure for the …rst step of the estimation o¤ers the advantage of ‡ex-
ibility and does not impose strong assumptions to the speci…cation of the functional form
of bj ( j = 1 j X): The intention of this …rst step estimation is to capture only the e¤ect
of the states variables X on the actions of the players ignoring completely the competition
or strategic e¤ect to select a particular location. To do so, a multivariate kernel smoothing
regression (MVKSR) approach is proposed.
5
Extreme cases like zero price margin or maximum entry close to zero are situations in which the minimun-
distance parameter r cannot be identi…ed.

8
We implement the Nadaraya-Watson6 kernel regression as follows:

X
L
(x Xj )
K h j
j=1
(12) bj ( j = 1 j X) =
X
L
K (x hXi )
i=1

where K is a kernel function, h the bandwith and x is the value of the state variable where
the distribution is being estimated. Using X = fp; pharmaciesentry<=2000 ; dh g; a gaussian
kernel and selecting the optimal bandwith similar to the suggested by Bowman and Azzalini
[1997]7 we have the following estimation which is represented in …gure 1.

Figure 1: Step 1 - Choice Probabilities Estimation


6
See Nadaraya [1964] and Watson [1964].
7
The di¤erence with the Bowman and Azzalini [1997] method is that for the binary variables instead of
using the median wich automatically set the bandwith to zero we use the mean, wich in those cases is a
better measure of the center of the distribution.

9
The darker, in this grayscale representation, the higher the probabilities of entry given
the states variables X. We’ll use this estimation of bj ( j = 1 j X) for maximize 10 on the
second stage.

5.2 Step II. Pseudo-likelihood estimation

Table 2: Step 2 - Pseudo-Likelihood Estimation

Coe¤. (Std.Err.) t-stat


Consumer Pro…t ( ) -6.98 (.126) -55.40
Distance (d) -6.44 (.144) -44.75
h
Distance Health Center (d ) -2.23 (.217) -10.28
Fixed Cost(c) -3.93 (.375) -10.47
Log likelihood -398.46
R2 0.52
Note: Bootstraped standard errors

The estimates in Table 2 appears to be reasonable. The coe¢ cient of distance (d) on
utility is 6:44. According to this result a person with 50 percet chance on going to a given
pharmacy would go with more or least 34 percent probability if the drugstore were 100 mt
away. Moreover, the coe¢ cient that measures the disutility on consumers of not having
a pharmacy near a health center (dh ) is 2:23. This means that a person with 50 percent
chances of going to a pharmacy would go with nearly 45 percent probability if it were 100
mt away from it’s nearest health center. The …xed cost parameter (c) have a value of 3:93,
a negative value is the only criteria we have to say that this estimation is reasonable given
that we don’t have any prices in the model and cannot convert this to monetary units.

6 Social Welfare Analysis

The welfare analysis based on this model is a counterfactual one, we shall compare the results
of the action of the players diferent values of the policy parameter: We consider four scenarios

10
for the policy analysis: 1) free-entry (r = 0 mt), 2) the actual situation (r = 150 mt), 3) an
scenario without deregulation (r = 250 mt) and 4) a more regulated scenario (r = 300 mt).
For thouse values of r; we compute the the number of new entries and the social welfare (11)
in "quantities".

Table 3: Welfare analysis

Scenarios (P olicy parameter) Social Welfare Number of new …rms


Free-entry (r = 0mt) 25.63 95.21
Actual regulation (r = 150mt) 15.86 75.65
Non-deregulated (r = 250mt) 11.75 59.13
More restricted (r = 300mt) 11.40 50.65

From table 3 we see that our model predicts the entry of 75.65 new …rms for the period of
consieration with the actual restriction, in comparison with the 86 actual …rms that enters.
This error is justi…ed by the fact that our model is only accounting the e¤ect of the minimun-
distance entry regulation and ignoring the other deregulations. As argued in section 4, this
does not means that we cannot identify the impact, in terms of rate of change, over the
social welfare.

The results are clear adressing the fact that the drugstores market in Pamplona is not
socially inne…cient. The missleading percepction that this types of markets are soccially
inne…cient arise for only accounting the bussines stealing e¤ect and ignoring the market
expantion of every new …rm. This market expantion appears in form of extra utility for
the consumers of having a drugstore near. Models of entry that does not include an explicit
speci…cation for the utility may found positive to mantain this types of regulation in order to
protect the …rms pro…ts, but this would be inne…cient to the society as a whole. As long as
the marginal utility of a new …rm is higher than the …xed cost it is favorable for the society
that new drugstores to enter the market. Our estimation appoints in that direction, the
Monte Carlo experiments in the next section try to prove the consistency of our estimator
in order to garanty this result is not trivial.

11
For the di¤erent scenarios we found a common pattern: the more deregulated is the
market, the more social welfare. According to our estimations the policy changes in year
2000 are favorable to the society, but the free-entry scenario represents a better alternative
for Pamplona. It would be interesting to see if these results can be replicated with data
from other Europeans markets, if so at least in the minimun-distance policy a free-entry
model like the Northamerican has to be considered in the discussion about regulation of the
drugstore market.

7 Monte Carlo evidence

To have an idea over the performance of our estimator we provide some Monte Carlo
evidence. We simulate the vector of choice probabilities choosing a given set of structural
parameters and bootstrap samples of exogenous data. Then, with our estimator we try to
recover the "true parameters" that we have choosen.

Table 4: Monte Carlo Evidence

"True" Value Coe¤. (Std.Err.) t-stat


Consumer Pro…t ( ) -5.00 -4.99 (.131) -38.11
Distance (d) -4.00 -3.95 (.110) -35.98
h
Distance Health Center (d ) -2.00 -2.01 (.202) -9.96
Fixed Cost(c) -4.00 -3.94 (.293) -13.45
Log likelihood -323.86
R2 0.57
Note: Bootstraped standard errors

The estimated coe¢ cients on table 4 correspond to the mean of 1,000 bootstraped sam-
ples, selected uniformly from the original dataset. The results of the experiment indicates
that the estimator perform very well. The di¤erence of our estimate and the "true parame-
ters" are negligible for all the parameters and the low values of the standard errors accounts
for an e¢ cient estimation.

12
Monte Carlo experiment proves the robustness of our pseudo-likelihood estimator and
assures the well identi…cation of the model. With this evidence, it is su¢ cient for assuring
a precise policy evaluation. Moreover, these results con…rm that there is not signi…cative
loss of e¢ ciency from the …rs stage estimation as Aguirregabiria and Mira [2007] suggest it
would happen if the …rst stage estimator have a considerable large variance. Seems to be
the case that the MVKSR approach capture enougth variation to identify the parameters
but not so much to lose e¢ ciency.

8 Conclusion
References
Aguirregabiria, V. and Mira, P.: 2007, Sequential simulation-based estimation of dynamic
dicrete games, Econometrica 1(75), 1–53.
Bajari, P., Hong, H., Krainer, J. and Nekipelov, D.: 2009, Estimating static models of
strategic interactions, NBER Working Paper .
Bajari, P., Levin, J. and Benkard, L.: 2004, Estimating dynamic models of imperfect com-
petition, Econometric Society .
Berry, S. T. and Waldfogel, J.: 1999, Free entry and social ine¢ ciency in radio broadcasting,
RAND Journal of Economics 30(3), 397–420.
Borrel, J.-R. and Fernandez-Villadangos, L.: 2009, Assesing exccess pro…ts from di¤erent
entry regulation, Serie de documents de treball de a XREAP 2009-3, Xarxa de Ref-
erencia en Economia Aplicada .
Working Paper.
Bowman, A. W. and Azzalini, A.: 1997, Applied Smoothing Techniques for Data Analysis,
New York.
Davis, P. J.: 2006a, Spatial competition in retail markets: Movie theaters, RAND Journal
of Economics 4(37), 964–982.
Goovaerts, P.: 1997, Geostatistics for Natural Resources Evaluation (Applied Geostatistics),
Oxford University Press, New York.
Gowrisankaran, G. and Krainer, J.: 2006, The welfare consequences of atm surcharges:
evidence from a structural entry model, NBER Working papers .

13
Hengl, T.: 2009, A Practical Guide to Geostatistical Mapping, 2 edn, O¢ ce for O¢ cial
Publications of the European Communities.
Issacks, E. H. and Srivastava, R. M.: 1989, Applied Geostatistics, Oxford University Press,
New York.
Mankiw, N. G. and Whinston, M. D.: 1986, Free entry and social ine¢ ciency, RAND Journal
of Economics 17(1).
Matheron, G.: 1969, Traité de géostatistique appliquée, Mémoires du Bureau de Recherches
Géologiques et Minières. 14.
Nadaraya, E.: 1964, Theory of probability and its applications, 9.
Pakes, A., Ostrovsky, M. and Berry, S.: 2004, Simple estimator of the parameters of dicrete
dynamic games (with entry/exit samples), National Bureau of Economic Research
Working Papers (10506).
Pesendorfer, M. and Schmidt-Dengler, P.: 2003, Identi…cation and estimation of dynamic
games, CEPR Discussion Papers (3965).
PSA: 2010, Issues Paper on the Future of Pharmacy, Pharmaceutical Society of Australia.
Watson, G.: 1964, Smooth regression analysis, Sankhya - The Indian Journal of Statistics
26, 359–372.
ÖBIG: 2006, Surveying, assessing and analysing the pharmaceutical sector in the 25 eu
member states, Report, O¢ ce for O¢ cial Publications of the European Communities,
Brussels.

A Appendix: Griding and Kriging

The most dissagregated data about the population in Pamplona are given by the Instituto
Nacional de Estadistica (INE), they provide the number of persons that lives in the 137
census districts of the city. The distribution of the districts is far from being regular, both in
size and in the number of habitants. As the Table 5 resumes, the distribution of the census
districts doesn’t seems to be suitable for our estimation. Particularly, the area of this objects
are in most of the cases to large to be considered as a localization for just one …rm to enter.

14
Min Max Mean Std. Dev.
Area (Km2 ) 0.0115 3.2083 0.1831 0.3614
Number of drugstores 0.0000 9.0000 1.3942 1.5018

Table 5: Summary Statistics of the Census Districts

Also, as the estimation of the choice probabilities ^ i (ai = 1 j s) for the entry decision of
the …rms depends on the local information of the states variables, the irregularity and the
size of the census districts fails in re‡ecting the local nature of the states variables that it’s
needed. Hence, we would rather use a more local dimension for the data and approximate
the information within, than using the original disposition of the census districts.

A.1 Griding

Our ideal dataset would use the information of the census districs in a fashion that the
…nal data presents the following properties: a) be small enougth for it to account the local
variability of the data and at the same time minimize the number of …rms per location, b)
consist of objects of equal size and c) has number of locations computationally tractable.
A way to get closer to our ideal dataset is to create a grid for the city of Pamplona with
equal size polygons, and then chose the size of the polygons that accounts for a) and c). We
prefer to use hexagons over rectangulars polygons because the distance between the center
of each hexagon and the center of all adjacent hexagon is constant. By comparison, in a
rectangular grid the distance from the center of each square cell to the center of the four
diagonal adjacent cells it shares a corner with is greater than the distance to the center of
the four adjacent cells it shares an edge with. This property of the hexagonal griding is
desirable because in our speci…cation of the utility function 1 the distance among locations
is an important factor. On Figure 2 and 3 we show the original census districts and our
de…nition of a spatial location using an hexagonal grid. The resulting grids have a length
from the center to the respective vertices of 57:87 mt and of 50:11 mt from the center to the
center of a given side.

Using the hexagonal grid now there are 3,686 potential locations for the …rms to locate
in Pamplona, wich is more realistic than the 137 census distritcs. The number of drugstore
per location now is nearly zero (mean = 0:05) and only 6 hexagons have more than one
(two) drugstores located in it, wich is better than having …rms in virtually every location

15
Figure 2: Census Districts of Pamplona

Figure 3: Hexagonal grid for Pamplona

of the map, see Table 5. If we reduce the radio so we don’t have thouse 6 polygons with
more than one (two) drugstores, the number of hexagons created increase dramatically and

16
it becomes a serious computational limitation for the further estimation, keep in mind that
for the estimation the dimensions of the states variables s increase exponentially due to the
simultaneity of the decisions of the players and the complete public information assumtion.
Considering that this 6 cases are the exception rather than the rule, they will be treated just
as is it was only one big drugstore located.

A.2 Kriging

Now that we have a new dimension for the data, we need to approximate the population
data to a local level. For that we consider the kriging methodology. The kriging is a
technique widely used on geoestatistic that tries to interpolate the value of a random …eld
at an unobserved location from observations of its value at nearby locations.

A.2.1 General Equations of the Ordinary Kriging (OK)

The predictions of the standard version of the ordinary kriging (OK) as presented on [Hengl
2009] are based on the model:

(13) Z(s) = + "0 (s)

where is the global mean and "0 (s) is the spatially correlated stochastic part. The
predictions of this models are made with the following expresion:

X
n
T
(14) z^(so ) = wi (so ) z(si ) = 0 z
i=1

where 0 is the vector of kriging weights (wi ), z is the vector of n observations at primary
locations. Since the novel work of [Matheron 1969], the method of estimating the weights in
an objective way realys on the semivariances, de…ned as:

1
(15) y(h) = E[(z(si ) z(si + h))2 ]
2
17
where z(si ) is the value of a target varable at some sampled location (the population
in our case) and z(si + h) is the value of the neighbor at distance si + h. The plot of all
semivariances versus their separator distances is called the experimental variogram. But,
as we don’t know what will be the semivariance for the points to be predicted we …t the
parameters of a theoretical variogram model: such as linear, spherical, exponential, circular,
Gaussian, Bessel, power and similars ?.

A.2.2 Variogram model Fitting

The theoretical variogram model is a function that describes the degree of spatial dependence
of a spatial random …eld. The parameters that decribes a variogram are: the nugget, the
sill and the range (n; s; r). n describes the jump of the variogram at the discontinuity at the
origin, s is the limit of the variogram tending to in…nity lag distances and r is the distance
in which the di¤erence of the variogram from the sill becomes negligible. Our criteria for
estimating the parameters will be the goodness of …t. For each functional form (linear,
spherical...etc.) we shall estimate the set of parameters (n; s; r) that minimize the sum of
the squared errors of the estimation. The Figure 4 shows the values of the experimental
variogram and the theoretical model …tted using the data of the population. The Table 6
summarizes the estimation results for the best model according to the goodness of …t criteria.

Estimated Parameters
Nugget 1,254.9467
Sill 6,880.8600
Range 1,416.2100
r2 0.9944

Table 6: Estimation of Spherical Variogram Model

18
Figure 4: Spherical Variogram Fitting

A.2.3 Population Interpolation

Using the variogram estimation we can now have an approximation of the weights for the
prediction of the kriging model in the Equation 14. The Figures 5 and 6 shows the population
using the census districts and the kriging aproximation. One can see that the kriging inter-
polation is more realistic than the original disposition of the data, which imply an uniform
distribution within each census district for the population.

19
Figure 5: Original data of the population of Pamplona

Figure 6: Kriging interpolation of the population of Pamplona

20

Вам также может понравиться