You are on page 1of 58

Accepted Manuscript

Efficiency, productivity gains, and the size of Brazilian supermarkets

Felipe Neves, Raquel Sampaio, Luciano Sampaio

PII: S0925-5273(17)30422-X
DOI: 10.1016/j.ijpe.2017.12.016
Reference: PROECO 6902

To appear in: International Journal of Production Economics

Received Date: 30 May 2017


Revised Date: 13 November 2017
Accepted Date: 17 December 2017

Please cite this article as: Neves, F., Sampaio, R., Sampaio, L., Efficiency, productivity gains, and the
size of Brazilian supermarkets, International Journal of Production Economics (2018), doi: 10.1016/
j.ijpe.2017.12.016.

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to
our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and all
legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT
Efficiency, Productivity Gains, and the Size of Brazilian Supermarkets

Felipe Neves1, Raquel Sampaio2, Luciano Sampaio3

1
Instituto Federal do Rio Grande do Norte.

PT
2
Escola de Ciências e Tecnologia, Universidade Federal do Rio Grande do Norte.
3
Departamento de Ciências Administrativas, Universidade Federal do Rio Grande do Norte.

RI
U SC
AN
M
D
TE
C EP
AC
ACCEPTED MANUSCRIPT

Abstract

The supermarket sector is an important segment of the Brazilian economy. In recent years,

PT
the role of large retailers and market globalization has transformed this sector in one of the

most competitive in the country. This paper aims to evaluate the technical efficiency and

RI
scale of the Brazilian supermarket sector through Bootstrap Data Envelopment Analysis,

SC
and to measure changes in productivity from 2005 to 2012, applying the Bootstrapped

Malmquist index. A second stage analysis is performed to explain efficiency levels and

U
productivity gains in terms of supermarket chains characteristics. We found that the
AN
Brazilian supermarket chains have low average efficiency levels and, in general, there

was an increase in productivity in the period. The consolidations that took place in this
M

sector can be explained in terms of scales of operation. Small supermarket chains had

lower efficiency levels both in 2005 and 2012, but experienced higher levels of
D

productivity growth in the period. Results of second stage analysis shows that some
TE

supermarket chains characteristics are correlated to productivity gains.

Keywords: Brazilian Supermarket Chains; Bootstrap Data Envelopment Analysis;


EP

Bootstrapped Malmquist Index.


C
AC

1
ACCEPTED MANUSCRIPT

Efficiency, Productivity Gains and the Size of Brazilian Supermarkets

1. Introduction

PT
The supermarket sector is an important segment of the Brazilian economy. In

2012, its share of GDP was 5.5%, which represents approximately 100 billion dollars in

RI
gross sales. The main companies operating in the country are Pão de Açúcar

SC
(controlled by the French group Casino), Carrefour, Walmart Brazil and Cencosud

Brazil. In 2013, these four together had gross sales of BRL 124 billion - 51% of sales

U
from supermarkets in Brazil (ABRAS, 2013).
AN
The Brazilian retail market has undergone significant changes in the recent

years (see a detailed list of events on Table 1) possibly motivated by the expansion of the
M

Brazilian economy1. In 2004, Walmart acquired 118 stores from Bompreço in the

Brazilian Northeastern region. A year later, the group incorporated operations from the
D

Portuguese Sonae, becoming one of the leading chains in the country. In 2012, Pão de
TE

Açúcar group, the largest supermarket chain in Brazil, merged with the French group

Casino.
EP

Due to the progress of international giants in the Brazilian market, regional


C

chains began to join in an effort to gain competitiveness through the exchange of


AC

technologies, increased purchasing power and reduced operating costs. In 2011, seven

regional chains operating in São Paulo, Minas Gerais and Santa Catarina states

established the CoopBrasil, forming a cooperative with gross sales of BRL 2.3 billion.

1
Brazilian gdp experienced an average annual growth of 4.2% between 2004 and 2012.
2
ACCEPTED MANUSCRIPT

These recent changes in the Brazilian sector - mainly the mergers, acquisitions and

cooperatives creation - provoked the interest in investigating the relation between the

optimal scale of operations and the efficiency of supermarket chains. Databases on the

supermarket chains for periods before and after these changes permit a time analysis.

PT
The main goal of this paper is to study efficiency levels of Brazilian supermarket

RI
chains before (2005) and after (2012) these changes have taken place. We will study four

questions, all of which are related to the scale of supermarkets. First, are larger firms more

SC
efficient than smaller ones? Second, are there scale inefficiencies in the data, that is, are

there supermarket chains that are either too big or too small? Third, did productivity

U
increase in this 7-year period? And was the evolution similar between small and big
AN
supermarket chains? Finally, are these results related to other supermarket chain
M

characteristics?

Several studies have evaluated the efficiency and increase in productivity of the
D

retail sector in many countries. The literature in efficiency in retailing have mostly being
TE

based on Data Envelopment Analysis (DEA), sometimes combined with other methods.

For Brazil, Yu and Angelo (2001), Macedo and Ferreira (2004), Didonet and
EP

Lara (2006) and Ferreira et al. (2009) focused on efficiency of supermarket sectors.

Most of these articles are convergent in their results, pointing to a higher level of
C

efficiency in large than in small chains, which would justify the recent changes. However,
AC

they are limited to a one-time period and so they do not capture the evolution of

supermarket chains productivity and its association to their scale change. Furthermore, the

3
ACCEPTED MANUSCRIPT

latest study in Brazil was carried out by Ferreira et al. (2009), using supermarket data from

2005, i.e., before the changes pointed above.

To the best of our knowledge, we are the first work to analyze productivity change

in the Brazilian supermarket sector. Also, most retailing applications did not use bootstrap

PT
to make inference on efficiency scores and/or productivity index. Two exceptions are de

RI
Jorge Moreno and Sanz-Triguero (2011) and Kato (2015) that used a bootstrapped

Malmquist index to study Japanese retailing establishments. Regarding the scale efficiency

SC
analysis, we are not aware of studies for retailing that conducted statistical tests to examine

inefficiencies in firms’ operations. As shown in our results, conclusions based only on

U
“deterministic” estimated efficiency levels might be misleading, and real inference analysis
AN
is necessary to understand the relation between size, efficiency and productivity in this
M

sector.

More specifically, this study evaluates efficiency levels of Brazilian supermarket


D

chains, focusing on the scale efficiency, and its evolution between 2005 and 2012. Output
TE

oriented models of bootstrap DEA, with constant returns to scale (CRS) and variable

returns to scale (VRS) are estimated for 2005 and 2012. For a review of bootstrap DEA
EP

see Simar and Wilson (2008). Then, non-parametric tests (Simar and Wilson, 2002) were

used to conclude if supermarket chains are scale efficient or scale inefficient. For each
C

scale inefficient supermarket chain, we identify if there is decreasing or increasing returns


AC

to scale.

In terms of productivity analysis, we followed Simar and Wilson (1998b) and

Wheelock and Wilson (1999) and used a Malmquist index decomposition analysis to shed

4
ACCEPTED MANUSCRIPT

light on the main driving factors of productivity gains between 2005 and in 2012. And,

finally, a second stage regression analysis was used to explain efficiency scores and

productivity gains for supermarket chains in terms of market presence (number of cities

they operate in, number of states they operate in and total number of stores), number of

PT
brands, and whether or not they are cooperatives or small convenience stores in

RI
comparison to the most usual supermarket chains.

Data was acquired from the Brazilian Association of Supermarkets (ABRAS).

SC
This institution, responsible for representing the interests of the sector in Brazil,

publishes an annual report with information such as number of employees, gross sales

U
and number of boxes for many companies. The database includes from small companies
AN
to international conglomerates that operate in Brazil. The final sample is composed of
M

494 supermarkets listed in ABRAS ranking of 2006 (data of 2005), and 701

supermarkets in the ABRAS ranking of 2013 (data of 2012). For the Malmquist
D

analysis, the sample consisted of 166 supermarkets listed in both periods.


TE

In general, mean efficiency levels are quite low in both years. The biggest firms are

the most efficient both in 2005 and in 2012. Besides, most supermarket chains are not
EP

statistically scale inefficient. Among chains that are scale inefficient, most of them are due

to increasing returns to scale.


C

With respect to the Malmquist index analysis, most supermarket chains gained
AC

productivity in the period, with smaller chains displaying more efficiency gains than larger

chains. This relative advantage of smaller chains is mainly caused by an increase in pure

technology. Finally, some supermarket chain characteristics are relevant to explain

5
ACCEPTED MANUSCRIPT

productivity change, namely being a cooperative, being located in the North and having

only one store in 2005.

In addition to this introductory section, this paper presents, in section 2, a review of

the literature on efficiency analysis of international and Brazilian supermarkets. Section 3

PT
describes the methodological procedures. Section 4 introduces the dataset and section 5

RI
shows and discusses our results. Lastly, section 6 draws general conclusions.

SC
2. Efficiency in the supermarket sector

U
AN
Some studies evaluated the efficiency of retail sectors, including supermarkets, in

various countries. Yu and Ramanathan (2009) used data envelopment analysis, the
M

Malmquist productivity index and Tobit regressions to assess the efficiency of 61

Chinese retailers, including supermarkets, department stores and electronics stores.


D

Using data from the Statistical Yearbook of China Market, for 2002 and 2003, they
TE

estimated DEA models of constant and variable returns to scale and a Malmquist
EP

index analysis. As inputs, the authors used the total number of employees of each unit

and the sales area in square meters. Gross sales revenue and profitability before taxes
C

were chosen as products. Their CRS results revealed that seven supermarkets were
AC

efficient in 2002 and only four in 2003. The Malmquist index indicated that

approximately 38% of the sample obtained an increase in productivity over the

studied period. Tobit regression revealed which characteristics of the retailer were

6
ACCEPTED MANUSCRIPT

significantly related to their efficiency. Among their results, they found that department

stores were more efficient than large retailers.

Exclusively for supermarkets, Sellers-Rubio and Mas-Ruiz (2006) evaluated the

economic and technical efficiency of 100 supermarkets in the Spanish retail sector, from

PT
1995 to 2001. Due to the change of the database over the years, the authors eliminated 64

RI
observations of an original sample of 164 intermediaries. They used two products (gross

sales and operating income) and three inputs (number of employees, number of outlets and

SC
debt). An elevated level of economic inefficiency of the sample was observed.

One year later, the same authors - Sellers-Rubio and Mas-Ruiz (2007) - evaluated

U
96 supermarkets operating in Spain between 1995 and 2003 through a DEA model with
AN
constant and variable returns to scale, and the Malmquist index. As inputs, they used the
M

points of sales number, number of employees and level of indebtedness. As products, gross

sales and net operating income were used. Their results showed an increase in average
D

annual productivity among the companies considered.


TE

Perrigot and Barros (2008) used a two stage DEA to study the efficiency of

supermarkets in France. First, DEA models with constant and variable returns to scale
EP

were estimated, followed by a Tobit regression model. The sample, collected from the

Grand Atlas des Enterprises, consisted of 11 French supermarkets, w i t h d a t a f r o m


C

2000 to 2004, totaling 55 observations. As inputs, the authors used the value of assets,
AC

operating costs and number of employees; as products, profitability and capital turnover.

They concluded that the French supermarkets were relatively efficient and emphasized

that not all supermarkets that had expanded internationally were efficient.

7
ACCEPTED MANUSCRIPT

Barros and Alves (2004) used a data envelopment analysis and a Malmquist

Productivity index, with output orientation, to access changes in productivity of 47

Portuguese supermarkets. They used cross-section data for the years 1999 and 2000. As

inputs, they used number of full-time equivalent employees; the cost of labor; the

PT
number of checkouts; stock; and other costs. As outputs, they considered yearly sales

RI
and operational results. The main results suggested that scale economies are related to

efficiency. Furthermore, most of retailers did not experience any gain in total

SC
productivity in the referenced period, with only 3 of 47 retailers obtained a Malmquist

index score greater than 1.

U
Barros (2006) examined the efficiency of 22 Portuguese supermarkets operating
AN
from 1998 to 2003 with data envelopment analysis and Tobit regression. The number of
M

employees and total assets of each unit were selected as inputs and gross sales and

operating income as products. Their evidence suggested that large groups were on
D

average more efficient than small ones and that national chains were more efficient than
TE

regional chains. Table 2 shows the main inputs and products used in the efficiency of

supermarket literature in other countries than Brazil, in chronological order.


EP

De Jorge Moreno (2010) first studied efficiency of retailers on eight European

countries, using separate frontiers for each country. Then, in a second step, they used a
C

single frontier to compare firms that were efficient in their countries. Then, they
AC

computed Malmquist index for this group of top firms. As a result, they observed that

more than 50% of DMUs improved in productivity, which is explained in equal terms by

both the catch-up and frontier-shift effects. Companies from Czech Republic and

8
ACCEPTED MANUSCRIPT

Belgium gained productivity in the period, while companies belonging to Spain, Italy,

France and Sweden did not improve.

De Jorge Moreno and Sanz-Triguero (2011) used an order-m approach and

bootstrap Malmquist to estimate efficiency and productivity of the Spanish retail sector.

PT
They found that firms had high inefficiency levels and that mean productivity decreased

RI
over the period of analysis. They also found that bigger firms were more efficient than

smaller firms.

SC
Ghandi and Shankar (2014) studied efficiency, scale efficiency and productivity

performance of 18 Indian retailers from 2008 to 2010. They also used a second stage

U
analysis with bootstrapped Tobit regressions to check if 2010 efficiency scores were
AN
related to firms’ characteristics. They found that firms’ efficiency levels were low and
M

that there was scope to increase efficiency through managerial decisions as well as scale

decisions. While there was gain in productivity from 2008 to 2009, productivity
D

decreased in the following year. They also found that the number of outlets is correlated
TE

to lower efficiency levels and retailers involved in mergers and acquisitions had

significant higher efficiency scores in 2010 than the other retailers.


EP

Kato (2015) estimated bootstrap Malmquist index for Japanese department stores

and supermarkets from 1995 to 2004. Both type of retailers showed stagnant levels of
C

productivity in the period. They also used a second stage analysis to examine the
AC

relationship between productivity performance and retailers’ characteristics. Their

conclusion on the importance of retailers’ characteristics to productivity is not clear,

9
ACCEPTED MANUSCRIPT

since several characteristics had conflicting effects between catch-up and frontier-effects

and among industries.

Specifically, for Brazil, the efficiency in the supermarket sector was analyzed by Yu

and Angelo ( 2001), Souza et al. ( 2010), Didonet and Lara ( 2006) and Ferreira et al.

PT
(2009). All Brazilian studies of the supermarket sector used the ABRAS databases and

RI
adopted gross annual sales as products and the number of boxes (check outs) and the

number of employees as inputs. In addition to these inputs, the first three studies cited

SC
above added the number of stores of the supermarket chains and the last three studies

included the total area in square meters. Souza et al. (2010) and Didonet and Lara (2006)

adopted the four inputs mentioned.


U
AN
Yu and Angelo (2001) compared the efficiency of large and small Brazilian
M

supermarkets, through the CRS model, to estimate the efficiency of 204 supermarkets

contained in ABRAS rankings for the years of 1994 to 1998. Their results indicate that the
D

major chains were more efficient than small chains. The authors explained this difference
TE

by access to technologies such as ERP software (Enterprise Resources Planning), whose

costs of implementation are over one million dollars, raising the level of competitiveness
EP

and restricting the activities of small retailers to small regions such as the neighborhood

where they are located.


C

Didonet and Lara (2006) studied the 50 largest supermarkets listed in the 2005
AC

ranking of ABRAS. Their goal was to compare strategies of supermarkets which adopted

large sales areas with supermarkets that used many shops, but with reduced area.

Furthermore, the authors sought to correlate the size of the store with technical efficiency.

10
ACCEPTED MANUSCRIPT

They estimated supermarkets efficiency with a BCC DEA model and their results showed a

higher level of efficiency for supermarkets that had adopted smaller sales areas. From the

total of 17 efficient supermarkets, 15 supermarkets had less than 1000 square meters’ sales

area, thus being considered of small to medium size.

PT
Souza et al. (2010) used data envelopment analysis to analyze the performance of

RI
100 supermarkets within the position of 201 to 300 at the ABRAS ranking of 2006. The

selection of the sample, according to the authors, was justified by the pursuit of a

SC
homogeneous set of supermarkets, with low dispersion of annual gross sales. They also

used a Mann-Whitney test to check whether organizational performance was unrelated

U
to the size of the stores. Their evidence indicated that the top 10 supermarkets in the
AN
sample had higher performance than the 10 smallest, corroborating other studies. They
M

also suggested that supermarkets with only one shop had efficiency problems related to

scale.
D

Ferreira et al. (2009) analyzed the efficiency of Brazilian supermarkets, focusing


TE

on the top 300 supermarkets listed in the ABRAS ranking of 2006. They stratified the

sample into three groups according to their gross sales levels, with one hundred
EP

supermarkets in each of the small, intermediate and large groups. They found a

relationship between efficiency and the size of supermarkets, with smaller supermarkets
C

being considered the most efficient ones. This general result is the contrary of that found
AC

by Yu and Angelo (2001).

Table 3 presents a summary of the literature cited above for supermarket

efficiency in the Brazilian sector, their input and outputs choices as well as their main

11
ACCEPTED MANUSCRIPT

conclusions. We observed a positive relationship between the gross sales of chains and

their efficiencies. Moreover, the majority of Brazilian studies found that big supermarket

chains are the most efficient ones, thereby emphasizing the importance of scale of

operation, which could explain the recent acquisitions and fusions observed in this country.

PT
It is important to mention that all the related Brazilians studies used only a one-year

RI
database (from ABRAS) and related their efficiency results to the operation scale (or the

chain size). While all studies but Ferreira et al. (2009), observed a positive relation

SC
between these two variables, they did not evaluate the time evolution of their efficiencies

encompassing a period when some of the supermarkets groups changed their sizes.

U
AN
3. Methodology
M

3.1. Bootstrap Data Envelopment Analysis


D

The problem of estimating efficiency levels of supermarket chains is intimately

related to the problem of obtaining an estimate for the set of production possibilities, , in
TE

the supermarket retailing sector. In this setting, each supermarket chain is considered to be

one decision making unit (DMU) that transforms a set of  inputs () into  outputs ().
EP

Based on estimation of a frontier of the production possibilities set and a distance


C

function specification, relative efficiency scores for each DMU can be computed. For

instance, with Debreu-Farell output distance function, efficiency scores for DMU  with
AC

input-output choices ( ,  ) are defined as:

 ( ,  ) = sup > 0: ( ,  ) ∈ .




12
ACCEPTED MANUSCRIPT

Efficiency scores  are equal to 1 for DMUs with production plans ( ,  ) that belong to

the frontier ( ), while inefficient DMUs present efficiency scores bigger than 1, so that

we would have to increase their outputs by a  factor to achieve efficiency.

Alternatively, with Shephard’s distance function, efficiency is measured as:

PT

 ( ,  ) = inf %& > 0: ' , ( ∈ ),
$ &

RI
which implies that Shephard’s efficiency scores are the inverse of Debreu-Farell efficiency

scores, so that 0 < & < 1 for inefficient DMUs and & = 1, for efficient DMUs.

SC
Additional conditions on the production possibilities set, such as assumptions of

U
constant, non-increasing or variable returns to scale, might also be included in the analysis,
AN
and will have an impact on the estimated efficiency scores.

Data Envelopment Analysis (DEA) is a non-parametric technique commonly used


M

to estimate production frontiers and relative efficiency scores. It is based on linear

programming and grounded on the seminal work of Farrell (1957), later disseminated by
D

Charnes et al. (1978).


TE

The model proposed by Charnes et al. (1978), from now on referred to by their

initials (CCR), supposed constant returns to scale. Considering , DMUs, the model
EP

goal is to build a frontier that envelope the data, so that all units are on or below the

frontier. Let - be the technology or production frontier under constant returns to scale. For
C
AC

each DMU, efficiency scores, with the Debreu-Farell distance function, are computed by

solving the following linear program:

13
ACCEPTED MANUSCRIPT

7 7

./0 ( ,  ) = sup 1 |  ≤ 4 56 6 ,  ≥ 4 56 6 for (59 , … , 57 ) such that 56


689 689

≥ 0, for B = 1, … , , C .

PT
The solution of this optimization problem represents an estimator of the frontier of

the constant returns to scale production set, ./0 .

RI
SC
Banker et al. (1984) proposed a model, called BCC, which generalizes the CCR

model by allowing for the existence of variable returns to scale. In this model, the

U
boundary is a convex production frontier, i.e., by varying the range of operation of a
AN
firm it is not always possible to vary the product in the same proportion. The linear

program associated to the BCC model is accomplished by the inclusion of a new


M

restriction on the problem above:

7 7 E

.D0 ( ,  ) = sup 1 |  ≤ 4 56 6 ,  ≥ 4 56 6 for (59 , … , 57 ) such that 4 56


D

689 689 F89


TE

= 1; 56 ≥ 0, for B = 1, … , , C .
EP

Thus, the production frontier with variable returns to scale is always contained
C

within the set of production possibilities of the model with constant returns to scale.

Let .D0 represent the estimator of this frontier.


AC

For a long period, DEA was seen as a deterministic approach, and no inference

were performed on efficiency scores which were computed based on estimated values of

./0 and .D0 . Simar and Wilson (1998a, 2000), recognizing that datasets used for DEA
14
ACCEPTED MANUSCRIPT

analysis were subject to sampling variation, derived a statistical model for efficiency

analysis. Since the underlying data-generating process is quite complex, the authors

suggested that the best way to conduct inference analysis in this setting is by using

bootstrap. However, naïve bootstrapping, which simply resamples over the original data,

PT
does not perform well, because of the enveloping nature of the problem. Indeed, the

RI
estimated frontier is necessarily contained in the real unknown frontier, so there is a clear

direction biasing efficiency score estimates. Also, because the estimated frontier is lower

SC
than the real frontier, the model estimates far too many efficient units, so that bootstrapping

naively across estimated scores would overstate the number of efficient units in bootstrap

samples.
U
AN
Simar and Wilson (2008) presented a detailed revision of alternative better
M

performing techniques, including Simar and Wilson (1998a)’s homogeneous bootstrap,

which we will adopt in this study. Through this methodology, once efficiency scores are
D

estimated, bias corrected efficiency scores are computed and confidence intervals are
TE

obtained with the percentile intervals method.


EP

3.2 Scale efficiency

We now proceed to investigate scale efficiencies of supermarket chains. This


C

analysis can give guidance on what are the most reasonable assumption regarding the true
AC

returns to scale of the production frontier. For instance, given the specific dataset in hand,

does the frontier technology exhibit constant or variable returns to scale? Moreover, if the

technology is not globally homogeneous of degree one, then some units are either too large

15
ACCEPTED MANUSCRIPT

or too small. In this case, we would like to identify which DMUS are scale inefficient and

if the source of their inefficiency is due to increasing returns to scale or decreasing returns

to scale.

The first step taken is to test the null hypothesis that the technology globally

PT
displays constant returns to scale against the alternative that it exhibits variable returns to

RI
scale (test 1). If this test is rejected, we proceed to test the null hypothesis that the

technology globally exhibits non increasing returns to scale against the alternative that it

SC
displays variable returns to scale (test 2). Simar and Wilson (2002) proposed several

statistics and investigate their small sample properties for these two tests. They suggested

U
using a ratio of means of estimated distances with CRS and VRS for test 1. Once again,
AN
inference is drawn with Simar and Wilson (1998a) homogeneous bootstrap. Likewise, test
M

2 is performed based on the ratio of means with estimated distance functions for the NRS

and VRS.
D

Similar reasoning was applied to conduct tests 1 and 2 at the local level. If the ratio

of estimated distance functions with CRS and VRS are sufficiently close to 1 for DMU 's
TE

production plan, we conclude that DMU  is not scale inefficient. If local test 1 is rejected,
EP

local test 2 investigates whether there is a significant difference between the ratio of

estimated distance functions with non-increasing returns to scale and VRS. If this ratio is
C

close to 1, then inefficiency is due to decreasing returns to scale, while if this difference is
AC

significantly different than 1, the inefficiency is due to increasing returns to scale.

3.3. Malmquist Index

16
ACCEPTED MANUSCRIPT

A number of indexes have been proposed to measure productivity gains of firms

over time, such as Laspeyres, Paasche, Fisher, Törnqvist, Malmquist and Hicks-Moorsteen

indexes. Total Factor Productivity (TFP) indexes are defined as an aggregate measure of

outputs divided by an aggregate measure of inputs (O’Donnell, 2012a). One of the most

PT
used indexes in the DEA literature is the Malmquist index, probably because it does not

RI
require price data and it is easily interpretable. However, the Malmquist index has lost

some of its popularity, ever since O’Donnell (2012a) showed that it is not always

SC
multiplicatively complete, therefore not being a reliable measure of TFP change. Recent

papers (O’Donnel, 2012b; Arjomandi et al., 2014; Sharma and Dalip, 2014; See and Li,

U
2015; Medal-Bartual et al., 2015; Molinos-Senante et al., 2016) have turned to the Hicks-
AN
Moorsten index, which also does not require price data and it is a proper TFP index
M

In this study, we adopt an output oriented Malmquist index based on Shepard’s

distance function and with respect to frontiers with constant returns to scale. With a single
D

output and under CRS, the output oriented Malmquist coincides with the Hicks-Moorsteen
TE

index as stated by Bjurek (1996). Thus, in our setting, the Malmquist index is a valid TFP

index.
EP

According to Zofio (2007), the advantage of defining the Malmquist index with

respect to frontiers with constant returns to scale is that, even if the underlying technology
C

does not exhibit constant returns to scale, the Malmquist index extends productivity
AC

comparisons over time from technical efficiency to include the role of scale efficiency on

fostering variations in productivity.

17
ACCEPTED MANUSCRIPT

We assume that there is a set of , DMUs each having m inputs denoted by a vector

6 ∈ ℝI and q outputs denoted by a vector 6 ∈ ℝJ over periods 1 and 2. Let ( ,  )9 =

(K9 , K9 ) and ( ,  )L = ( L ,  L ) designate the production plans of DMU  (for

 = 1, . . . , ,) in periods 1 and 2, respectively. In the Malmquist index (MI) analysis, the

PT
efficiencies of DMUs ( ,  )9 and ( ,  )L are evaluated by frontier technologies 1 and 2

RI
with constant and variable returns to scale.

Shephard’s output distance function for firm  with respect to a variable returns to

SC
scale frontier is defined as M (NM ,  M ) = inf$ %& > 0: ( M , $P ) ∈ D0
OQ
M
), where  M is the

R M (NM ,  M ) = inf$ %& > 0: ( M , P ) ∈ /0


U
OQ
production possibility set at period t. Similarly, M
),
$
AN
where /0
M
is homogeneous of degree 1, meaning that the technology exhibit constant

returns to scale. Then, firm  increase in productivity between period 1 and 2, as measured
M

by the Malmquist index, is given by:

R 9 ( L ,  L )
R L ( L ,  L ) 9/L
D


R 9,L ( 9 ,  9 ,  L ,  L )
 =S 9 9 9 ∗ L 9 9 U ,
R ( ,  )
R ( ,  )
TE

which is the geometric mean of productivity change between two periods evaluated with
EP

the base and final period constant returns to scale frontiers.

R 9,L ( 9 ,  9 ,  L ,  L ) > 1, there was progress in the total factor productivity


If 
C

of the DMU  from period 1 to 2, while R 9,L ( 9 ,  9 ,  L ,  L ) = 1


 and
AC

R 9,L ( 9 ,  9 ,  L ,  L ) < 1, respectively, indicate the status quo was maintained and that


there was deterioration in the total factor productivity for this DMU between periods

1 and 2.

18
ACCEPTED MANUSCRIPT

To shed light into sources of productivity change between two periods, various

decompositions of the Malmquist index have been proposed. We follow Simar and Wilson

(1998b) and decompose the MI into four terms2:

R 9,L = WX 9,L ∗ WYX 9,L ∗ YX 9,L ∗ WX 9,L ,




PT
with components defined as:
9
9 ( L ,  L ) 9 ( 9 ,  9 ) L

RI
WX 9,L = Z[\] ^]_ℎ,_ab _ℎa,c] = S L L L ∗ L 9 9 U
( ,  ) ( ,  )

SC
L ( L ,  L )
WYX 9,L = Z[\] ]dd_],_ _ℎa,c] =
9 ( 9 ,  9 )

R L ( L ,  L )

YX 9,L U
L L L
( ,  )
= e_ab] ]dd_],_ _ℎa,c] = 9 9 9
AN
R ( ,  )

9 ( 9 ,  9 )
M

9
R 9 ( 9 , 9 )
R 9 ( L ,  L ) L

h 9 9 9 99 L L k
g ( ,  ) ( ,  ) j
WX 9,L = e_ab] ^]_ℎ,_ab _ℎa,c] = g L 9 9 ∗ L L L j
R R ( ,  )
g ( ,  )
D

j
f L ( 9 ,  9 ) L ( L ,  L ) i
TE

The first two terms, WXK9,L and WYX 9,L , are based only on VRS frontiers and
EP

represent the well know frontier-shift and catch-up effects of Caves et al. (1982)

Malmquist decomposition, respectively. A catch-up effect above 1 indicates progress in


C

relative efficiency from period 1 to period 2, while a catch-up equal to 1 and below 1
AC

indicates no change and decrease in efficiency, respectively. A frontier-shift above 1

indicates progress in the VRS frontier technology around DMU  from period 1 to 2,

2
To present a more compact equation, we have hidden ( 9 ,  9 ,  L ,  L ) from all terms of the equation below.
19
ACCEPTED MANUSCRIPT

while a frontier-shift equal to 1 and below 1, respectively, indicates the status quo and

regress in the VRS frontier technology.

The third and fourth terms, YX 9,L and WX 9,L encompass scale efficiency analysis

by comparing CRS and VRS distance functions. The latter term, WX 9,L , consist of the

PT
geometric mean of two ratios, where each ratio considers a single period input-output

combination. By holding input-output plans of DMU  fixed, each ratio measures how

RI
scale efficiency have changed merely because of scale changes at the best practice frontiers

SC
between the final and base periods, regardless of firm ’s decision. For this reason, it is

called scale technical change. WX 9,L < 1 indicates that VRS and CRS frontiers are closer

U
AN
to each other on the final period then they were on the base period, or that there was a

flattening of the VRS technology. Conversely, WX 9,L > 1 means increasing curvature of
M

the VRS frontier, or a change away from CRS (Simar and Wilson, 1998b).

Finally, YX 9,L computes changes in scale efficiency of DMU . YX 9,L > 1 imply
D

that DMU  is closer to the optimal scale on the second period, with respect to the first
TE

period. This change might either be caused by changes in location of input-output plans or
EP

by changes in the shape of technology.

Statistical inference on the Malmquist index and its components is performed using
C

the smooth homogeneous bootstrap introduced by Simar and Wilson (1999).


AC

3.4. Second stage analysis

20
ACCEPTED MANUSCRIPT

In order to further understand the main drivers of efficiency and productivity gains,

second stage analysis use supermarket chains characteristics to explain the efficiency levels

and Malmquist index.

According to Simar and Wilson (2007), estimation of the relationship between

PT
efficiency levels,  and DMUs characteristics, l , is complicated by the fact that efficiency

RI
levels are not observed but rather estimated by DEA models. Moreover, because DEA is a

benchmark model, relative efficiency scores between DMUs are correlated to each other.

SC
In this setting, the best option to perform inference analysis on second stage is to use

bootstrap. We adopt Simar and Wilson (2007) double bootstrap (algorithm 2) that uses bias

U
corrected efficiency levels, .. , as the dependent variable in the second stage:
AN
.. = mn o + q ≥ 1.

Then, the coefficients of interest, o, are estimated using maximum likelihood assuming
M

that q comes from a normal distributed with zero mean and r standard deviation and left-
D

truncation at (1 − mn o). A parametric bootstrap based on the same assumptions about the
TE

error term distributions yields the construction of confidence intervals by the percentile

method.
EP

Second stage analysis of Malmquist index is simplified by the fact that the
C

Malmquist index and its components are not bounded by one. However, since the
AC

Malmquist index is computed based on distance function to estimated frontiers, there

possibly exists correlation between Malmquist indexes of different DMUS. For this reason,

we use a non-parametric bootstrap and robust ordinary least squares to estimate

21
ACCEPTED MANUSCRIPT

correlations between Malmquist indexes, its components and supermarket chains’

characteristics.

In summary, in the next section we present results for the following analysis, in this

order: (1) estimation of efficiency levels of supermarket chains in the first period and

PT
second periods using with a CCR model and with a BCC model with DEA bootstrap; (2)

RI
analysis of scale efficiency for each data period; (3) estimation of bootstrap Malmquist

index and its components for firms present in both data periods; (4) second stage

SC
regression of efficiency levels and Malmquist index and its components in terms of

supermarket chains characteristics. All bootstrap analyses were conducted with 2000

replications.
U
AN
Stata commands teradialbc, nptestrps and simarwilson were used for bootstrap DEA
M

estimation, the analysis of scale efficiency and efficiency scores second stage analysis,

respectively. The R package FEAR 2.0.1 was employed for the bootstrap Malmquist
D

index analysis. For more information, see Wilson (2008) and Badunenko and
TE

Mozharovskyi (2016).
EP

4. Data

We measured the technical efficiency of supermarkets, scale efficiency and the


C

evolution of productivity of Brazilian supermarkets, using the 2006 and 2013 annual
AC

rankings of ABRAS based on data from 2005 and 2012, respectively. ABRAS constructs

these datasets with the support of supermarkets information on infrastructure, financial and

22
ACCEPTED MANUSCRIPT

commercial issues that are voluntarily reported. There are 500 and 709 observations in

the rankings of 2006 and 2013, respectively.

Following Ferreira et al. (2009), we specify, as input variables: the number of

employees, the number of checkouts and store area in square meters. As output, we

PT
considered only the Annual gross sales3. Since data of two different years are merged

RI
together, comprising a period of considerable economic growth associated with high inflation

rates, we have chosen to deflate sales of 2012 with a Brazilian consumer price index, the IPCA.

SC
Table 4 panel A presents descriptive statistics for input and output variables in our

datasets. In 2005, the mean of gross sales of supermarket chains was BRL 137.51 million, while in

U
2012 it had grown to BRL 242.83 million, which corresponds to only BRL 170.93 million, once
AN
inflation is accounted for. This increase in average gross sales was associated to a relative
M

increase in the mean number of employees, which grew from around 692 in 2005 to

860 in 2012. Meanwhile, the average store area and number of boxes have not displayed
D

any prominent change. Besides, standard deviations of all variables are quite large in
TE

comparison to mean values, suggesting high heterogeneity among supermarket chains

in this sector.
EP

Before proceeding with the efficiency analysis, we investigated the presence of

outliers in each database. Following Banker and Chang (2005), a super-efficiency model
C

is estimated, and DMUs with super-efficiency score higher than 1.6 are considered as
AC

outliers. Therefore, 6 and 8 supermarkets were eliminated from the analysis for 2005

3
As shown in Table 2, there is some variation on the selection of inputs in the literature. For this reason, we
have tested the robustness of our findings to the inclusion of the number of stores as input and to the
replacement of total area by the number of stores. Mean estimated efficiency scores are very close, with
maximum variation of 6%.
23
ACCEPTED MANUSCRIPT

and 2012, respectively. Table 4 panel B presents descriptive statistics for the samples

without outliers. A substantial decline in mean values of all inputs and the output are

observed, highlighting the importance of the eliminated observations on the sample.

PT
5. Results and Discussion

RI
5.1 Efficiency Analysis

Relative efficiency scores of supermarket chains in each year were computed using

SC
bootstrap DEA with Debreu-Farell distance functions and constant or variable returns to scale.

Higher scores are associated to less efficient firms. Therefore, we prefer to call these

U
values, “inefficiency” scores. Table 5 summarizes the results.
AN
The average efficiency of supermarket chains was quite small in both data periods4.
M

Out of 494 and 701 chains, only 24 chains were efficient in each year in the model with the

less restrictive assumption of variable returns to scale. By adopting the best practices in
D

this sector, on average, firms could increase their output by around 165% in 2005 and
TE

259% in 2012, maintaining their inputs fixed and with VRS (see bias corrected efficiency

scores). If one assumes that the technology displays constant returns to scale, the
EP

corresponding potential increase in gross sales amount to approximately 186% and 306%

in 2005 and in 2012, respectively. Sellers-Rubio and Mas-Ruiz (2006) and Ferreira et al.
C

(2009) also identified large inefficiencies in supermarket chains.


AC

4
Despite the fact that average efficiencies are already very small, a large number of DMUs present zero dual
multipliers for some inputs. For example, some supermarket chains have zero weight on number of
employees, which in practice would not make much sense. The estimated efficiencies documented here
should be interpreted as upper bound on real efficiencies that would be obtained if we imposed weigh
restrictions on dual multipliers.
24
ACCEPTED MANUSCRIPT

In order to explore possible selection bias in voluntary responses to ABRAS

databases, Table 5 also presents the average results obtained by subgroups of supermarket

chains according to their presence in each ranking. For instance, we want to know if

supermarkets that are present only in the first ranking are consistently less efficient than

PT
supermarkets that submitted their responses in both periods. It could be that those

RI
supermarkets went bankrupt or that less efficient chain managers just do not have time or

desire to submit their information to the ABRAS Association.

SC
In contrast, if we find no significant relation between efficiency and ranking

presence, it might be that simply a random set of managers did not send their information

U
by accident. Similar reasoning applies to supermarkets that are present only in 2012. They
AN
may either be new supermarket chains or old supermarket chains that did not send their
M

information in 2005. In both issues, it is interesting to analyze if there is a systematic

misreport of data by less efficient chains.


D

When comparing mean inefficiency levels of the 166 supermarket chains present in
TE

the databases of the two periods, we find a much lower difference in average inefficiencies

from the initial to the final period of time (1.94 in 2005 and 2.06 in 2012 with VRS and
EP

2.10 and 2.41 with CRS). However, these results relate to the relative efficiency of

supermarket chains compared to other chains for the same period.


C

We also notice that chains present in both years are much more efficient than chains
AC

that only report data in one year, as shown by the absence of overlap in 95% confidence

intervals for inefficiency levels between chains present in both years and chains present in

only one year. By holding their inputs fixed, gross sales could be increased by around

25
ACCEPTED MANUSCRIPT

167% in 2005 and 261% in 2012 for chains that did not report data on both years and

93.6% and 106% for chains that are present in both years with the VRS model. For a

proper analysis of the evolution of the supermarket chains productivity, we will use the

Malmquist index, but we should keep in mind that those results apply to the more

PT
efficient firms that have reported data on both years.

RI
As pointed out in the literature, scale of operation seems to be an important

component of efficiency analysis for supermarket chains. Table 6 presents bootstrap

SC
inefficiency estimates divided by ranking classes of chains in 2005 and 2012. The

biggest companies in terms of gross sales are listed on the top of the ranking.

U
When examining mean inefficiency levels by ranking classes in 2005, it seems
AN
that size is related to efficiency across all ranking classes, with larger chains being more
M

efficient than smaller chains. By taking a closer look into the 95% confidence intervals,

we identify that the smallest firms (“401 to 500” and “301 to 400”) are indeed less
D

efficient than all the other top ranking classes with CRS and the VRS models. However,
TE

efficiency distinctions between the top classes are more difficult to pinpoint. While the

top ranking class (“1 to 100”) is more efficient than all the other classes, at least with
EP

the CRS model, the two other classes have similar levels of efficiency.

For 2012, a very similar pattern emerges, with, once again, bigger chains being
C

more efficient than smaller chains. For instance, inefficiency levels of the first ranking
AC

class (“1 to 100”) range from 1.63 to 2.20 with 95% confidence level, while

inefficiency levels of the third ranking class (“201 to 300”) go from 2.27 to 2.64

assuming variable returns to scale. Especially for the ranking class of the smallest

26
ACCEPTED MANUSCRIPT

chains (“600 or more”), efficiency levels are strikingly low. This ranking class is

composed mainly of stores that did not reported data for 2005.

In general, these conclusions are in line with most of the Brazilian literature on

supermarket efficiency, with the only exception of Ferreira et al. (2009).

PT
RI
5.2 Scale efficiency

To investigate the presence of relevant scale effects in the sector, we conducted a

SC
non-parametric test of returns to scale. First, a global test of returns to scale rejects the

hypothesis that the global technology exhibits constant returns to scale in favor of a

U
technology with variable returns to scale with p-values close to zero for both 2005 and
AN
2012.
M

Then, for each period and for each DMU, two sequential local scale tests were

conducted. Tables 7 and 8 count the number of not statistically scale inefficient and the
D

number of significantly scale inefficient due to increasing and due to decreasing returns to
TE

scale. Table 7 relates scale efficiency indicators to the presence or not of a supermarket

chain on both datasets, while Table 8 classifies scale efficiency indicators by gross sales
EP

ranking classes.

According to Table 7, the vast majority of supermarket chains (385 out of 494 in
C

2005 and 585 out of 701 in 2012) are close enough to their optimal scale not to be
AC

considered statistically scale inefficient, emphasizing the importance of accounting for

sample variation with an inference analysis5. Whenever scale inefficiency was found, it

5
If only the ratio of efficiency with CRS and VRS (test 1) and NRS and VRS (test 2) were used to study
scale efficiency, without a proper inference analysis, one would wrongly conclude that a high number of big
27
ACCEPTED MANUSCRIPT

was mostly due to increasing rather than decreasing returns to scale. In fact, only two

DMUs in 2005 were significantly operating above their optimal range. The remaining 107

chains in the increasing returns to scale area ranked below position 201. None of the 200

biggest firms in terms of gross sales were scale inefficient (see Table 8).

PT
For the subsample of firms with data on both years, only 18 out of 166 firms were

RI
operating below optimal scale and none were above optimal scale in 2005. As a whole,

this evidence suggests that the merger/consolidation phase underwent by this sector was

SC
justified by efficiency considerations.

In 2012, most supermarket chains scale inefficiencies are still due to increasing

U
returns to scale, and they are also concentrated on the lowest ranking classes, all of which
AN
are below the 400th position, and most, bellow the 600th position. In contrast, 12 medium
M

sized supermarket chains, ranked between the 101th and 300th in the 2012 database, are in

the region of decreasing returns to scale.


D
TE

5.3 Malmquist Index Analysis

To check whether or not fusions and acquisitions that have occurred in recent
EP

years targeted an increase in efficiency, the Malmquist index and its four components were

evaluated for supermarket chains that are present in the two-time period of 2005 and
C

2012, although each period frontier is based on data of all observations.


AC

As previously discussed, the Malmquist index sample is composed of long

lasting supermarket chains that also turn out to be more efficient than those reporting

supermarket chains were operating with decreasing returns to scale. For instance, that would represent around
80% of supermarket chains of the first ranking class and 75% of the second ranking class.
28
ACCEPTED MANUSCRIPT

data for only one year. Table 9 shows descriptive statistics of output and input variables

of supermarket chains’ used in this part of the analysis. The comparison of values

between 2005 and 2012 in this table is easier to interpret than that of Table 4, since we

have the same supermarket chains present in both datasets.

PT
The real average gross sales for this group of supermarket chains increased by

RI
around 1.67 times while the number of boxes, total area and the number of employees

increased about 1.47, 1.42 and 1.74 times between 2005 and 2012, respectively. If the

SC
average output increased proportionately more than the average inputs, we should

expect an average increase in productivity growth. However, it is not that simple to draw

U
conclusions based only on average values, since Table 9 also highlights the presence of
AN
strong heterogeneity in this sector. Indeed, standard deviations are more than 4 times
M

bigger than mean values for all variables in both years. Since the Malmquist index allows

us to study the evolution of productivity for each supermarket chain, it allows us to


D

identify which specific firms experienced productivity growth or decline between 2005 and
TE

2012.

Table 10 displays the results of bootstrap Malmquist index estimation. To


EP

associate productivity gains and size, supermarket chains were divided into the

following four groups based on sample quartiles of gross sales from 2005: (1) Small, for
C

chains with sales up to BRL 11.5 million; (2) Small to Medium, for chains with sales
AC

between 11.5 and 32.1 million; (3) Medium to Large, for chains with sales between 33.1 and

83.4 million; and (4) Large, for chains with sales higher than 88 million. Besides, we

29
ACCEPTED MANUSCRIPT

have also looked at the top 5% chains in terms of gross sales (more than BRL 405.1

million), which are potentially more affected by recent developments in this sector.

On average, supermarket chains increased their productivity by a factor of 1.25

between 2005 and 2012, with the 95% Confidence Interval ranging from 1.20 to 1.30.

PT
Thus, supermarket chains improved their average productivity by around 20% to 30% in

RI
seven years.

Out of 166, 104 (62%) DMUs increased their productivity in the 7-year period and

SC
only 20 (12%) experienced a decrease in productivity as measured by the Malmquist index.

To further analyze if supermarket chains with significant increase in productivity changed

U
their production plans between 2005 and 2012 in a different way than those that
AN
experienced a significant increase in productivity, Figure 1 presents a radar graph of ratio
M

of average 2012 values of variables used in the efficiency model (gross sales, number of

outlets, number of employees and total area) with respect to 2005 values of the same
D

variables. DMUs with significant productivity improvement in this period (“MI > 1”),
TE

increased their inputs by a smaller factor than the remaining DMUs (“MI < 1” and MI = 1),

while, at the same time, being able to expand their average gross sales by a larger amount
EP

(2.20 for “MI > 1” in comparison to 1.75 for “MI = 1” and 1.63 for “MI < 1”).

With respect to supermarket chain sizes, all four groups had a significant increase in
C

mean productivity, but there occurred a higher productivity gain for smaller firms, since
AC

average Malmquist index was higher for the “Small” group, followed by the “Small to

Medium group” and the “Medium to Lange” group, while there was no significant

difference between the “Medium to Large” and the “Large” group.

30
ACCEPTED MANUSCRIPT

Looking into the source of these productivity gains, we now move to decomposition

of the Malmquist index into four terms: pure efficiency change, pure technological change,

scale efficiency change and scale technology change.

Panel B of Table 10 indicates no average pure efficiency change between 2005 and

PT
2012, even though 72 DMUs had a significant decrease in efficiency and 28 had a

RI
significant increase in efficiency, both measured with respect to the VRS frontier. Also,

there was no clear distinction between groups of firms with respect to the pure efficiency

SC
change.

Panel C of Table 10 shows that pure technological progress has benefited almost

U
everyone, given that 119 observations had a significant increase and no observation had a
AN
significant decrease in pure technological change. Additionally, the average increase in
M

pure technological effect (1.28), is almost of the same size as the average Malmquist index

(1.25). Again, “Small” chains variable returns to scale frontier-shift improvement was the
D

most prominent of all, with an average of 1.55.


TE

Scale productivity analysis is divided in two terms, one relating to the fact that

DMUs changed their scale to be closer to the optimal scale, named scale efficiency change,
EP

and a second term, known as scale technology change or scale bias of technical change,

which captures changes in the optimal scale that goes into the direction of the scale
C

employed by the firm. Panels D and E of Table 10 presents bootstrap estimates of these
AC

effects.

The overall increase in efficiency is not explained by a converging process toward

optimal scale, as the average scale efficiency change is lower than 1, though not

31
ACCEPTED MANUSCRIPT

statistically significant at 5%. However, expect for the “Small” group, there was a

significant decline in scale efficiency for the other size groups, meaning that most

supermarket chains moved away from the optimal scale between the base and the final

year. In fact, 92 of all firms showed a decrease in scale efficiency and only 18 showed an

PT
increase in scale efficient. Out of these 18 chains, 15 are classified into the “Small” group

RI
and 3 in the “Small to Medium” group. No chains on the “Medium” and “Medium to

Large” moved closer to the optimal scale.

SC
Average scale technical change of supermarket chains was 1.10, significantly above

1. Thus, there was a global flattening of the VRS frontier taking it closer to the CRS

U
frontier in 2012 with respect to 2005. Here, only the “Small to Medium” and the “Medium
AN
to Large” groups have a scale technical change statistically significant. In fact, 13 out of
M

the 15 chains with significant decrease in scale technical change are “Small” and the other

2 belong to the “Small to Medium” group.


D

In general, scale efficiency change and scale technical change basically offset each
TE

other and leave a small insignificant increase in productivity due to changes in scale. In

fact, the product of YX 9,L and WX 9,L , corresponds to the change in returns to scale
EP

measure of Ray and Desli (1997) Malmquist decomposition. Average returns to scale for

all groups are quite close to 1 and statistically insignificant for all groups of firms.
C

In summary, the productivity increase observed by a significant increase in the


AC

Malmquist index is mainly driven by pure technological progress. Pure efficiency did not

statistically contribute to increase in productivity, while scale efficiency change and scale

technology almost cancel each other out, implying in insignificant returns to scale change.

32
ACCEPTED MANUSCRIPT

Noteworthy, small chains frontier-shift improvement was the most prominent of all,

implying that small supermarkets chains have approached large and medium chains in

terms of efficiency in the period.

PT
5.4 Second stage analysis

RI
Using a second stage analysis, we explored if bias corrected efficiency levels and

Malmquist components are correlated with some chains characteristics. The second stage

SC
analyses for efficiency scores were performed using Simar and Wilson (2007)’s method.

Since the Malmquist index and its components are not subject to truncations, i.e., they are

U
not restricted by 1, their second stage analyses were based on robust ordinary least squares
AN
with non-parametric bootstrap.
M

Chains’ explanatory variables include dummy variables for cooperative (dcoop)

and convenience store (dconv) chains, and chains’ market dispersion as measured by
D

chain’s number of stores (nstores), number of brands (nbrands), number of cities (ncities)
TE

and number of states (nstates). Following Didonet and Lara (2006) and Souza et al.

(2010), we also included mean store size (size) defined as total area divided by number of
EP

stores and a dummy for chains with only one store (d1store). To capture regional market

particularities, dummy variables for four Brazilian regions (dN, dNE, dCW and dS) were
C

included in the regressions. We chose the South East, which is the richest Brazilian region,
AC

as the reference category. Only four supermarket chains were present in more than one

region and only two supermarket chains were present in all regions. The four chains

present in more than one region are singled out by a dummy variable called dnational.

33
ACCEPTED MANUSCRIPT

Information about chains’ characteristics was mostly gathered at their websites,

except for the number of stores that came from ABRAS databases. We were able to collect

data from 2012 for 165 chains (one of them went bankrupt in 2014). Data for 2005 were

harder to obtain, and there is missing information for 22 chains, including some of the

PT
biggest chains. Besides, for chains with non-missing data, there is almost no change in type

RI
(cooperative or convenience stores), number of states, number of cities and number of

brands from 2005 to 2012. For this reason, we chose to focus on explaining 2012

SC
efficiency levels and, for the second stage regressions of the Malmquist index, we have

only included the 2012 values of these variables.

U
Table 11 displays descriptive statistics of explanatory variables of our second stage
AN
analysis. Supermarket chains have substantially expanded their business between 2005 and
M

2012, with the average number of stores increasing by around 60% and the number of

chains with only one store decreasing from 33% to 25%. Meanwhile, the average store size
D

increase by approximately 5.8% in this period.


TE

The large majority of chains in this dataset operate in the Southern and

Southeastern regions of the country, which are the richest ones and with populations
EP

spatially distributed over a larger number of non-capital cities. Besides, less than 10% of

the observations are classified into convenience stores or cooperatives.


C

Regarding market dispersion, the average supermarket chain is present in around


AC

six cities located in one or two states and with mostly one brand. In fact, 147 chains

operate in only one state, 15 chains operate in two states and only three chains operate in

three or more states in 2012. Regarding the distribution of brands, 145 chains has only one

34
ACCEPTED MANUSCRIPT

brand, while 15 chains have two brands and 5 have three or more brands. To specify a

more flexible model, instead of nstates, we used dummy variables for 2 states (d2states)

and 3 or more states (d3states), leaving chains present in only one state as the reference

category. Likewise, dummies for 2 (d2brands) and 3 or more brands (d3brands) were

PT
used in place of nbrands.

RI
Table 12 presents Simar and Wilson (2007)’s estimates of the relationship between

explanatory variables and bias corrected inefficiency scores based on Debreu-Farell

SC
distance functions with CRS and VRS models for 2012. The only variable that is

significantly correlated to inefficiency levels is store average size (size2012). An increase

U
of 100 square meters is associated to a reduction in inefficiency levels of around 0.0477
AN
with variable returns to scale, which is quite small considering that the average inefficiency

level with this model is 2.062 (see Table 5)6. Consonant results are obtained with the CRS
M

model.
D

Table 13 shows estimates of Malmquist index and its components with explanatory
TE

variables. Considering that the dependent variables measure productivity change between

2005 and 2012, we have added the average percent change in store size (Dsize) and the
EP

increase in number of stores (Dnstores = nstores2012-nstores2005), instead of the

corresponding time period variables.


C

The variables that have a significant correlation with the Malmquist index are the
AC

dummy for convenience stores chains, the dummy for cooperatives, the dummy for chains

that had only one store in 2005 and the dummy for the Northern region.

6
Results for 2005 efficiency levels based on 143 observations are similar to those of 2012.
35
ACCEPTED MANUSCRIPT

Everything else equal, cooperatives had a MI approximately 0.22 lower than regular

supermarket chains, which is explained by a significantly lower catch-up effect (0.12) and

lower scale change effect (0.07). At the same time, convenience store chains are more

efficient than traditional supermarket chains, with 0.20 higher MI.

PT
As expected from our previous discussion relating chain sizes and the MI, chains

RI
that had only one store in 2005 have a MI 0.37 points higher than the other chains. They

present higher pure technology and scale change, but there was a flattening of the VRS

SC
frontier around their production plans. Finally, chains in the North of the country

experienced lower productivity improvement than chains located in the South-East.

U
AN
6. Conclusions
M

This study aimed to evaluate the efficiency of operations in supermarket chains


D

in Brazil and its evolution during the period 2005 to 2012 and to associate it to the
TE

chains’ sizes. Some of our results corroborate previous studies for the sector and more

specifically for the Brazilian market. In particular, average efficiency is low, showing that
EP

it is possible to increase output considerably from existing production inputs, especially

for small chains, which are less efficient than big chains.
C

Regarding the major consolidations that took place in this sector, these
AC

institutional changes could be explained by the lack of inefficiencies due to decreasing returns

to scale. The biggest supermarket chains were seemingly operating at or close enough to

36
ACCEPTED MANUSCRIPT

their optimal levels, and profits increase could be obtained by incorporation of small

supermarket chains that were operating on increasing returns to scale.

Interestingly, although small supermarket chains were the ones with lower

efficiency levels in 2005, they experienced the strongest growth in productivity during

PT
this period, a growth that is mostly explained by an improvement in pure technology

RI
effects. In fact, only 12% supermarket chains showed a significant decrease in

productivity, which indicates that, in general, the supermarket sector gained productivity

SC
in the 7-year period.

Productivity improvements are associated to location and to chain type, with lower

U
increase in productivity for chains located on the Northern region and for cooperatives in
AN
comparison to regular supermarket chains. In addition, there does not appear to be a
M

relation between market presence, number of brands, chain’s types and efficiency levels or

productivity gains.
D

The main limitations of this study relate to the availability of data and data
TE

collection method. The fact that ABRAS is answered spontaneously by supermarkets

casts doubt into the representativeness of our results. Besides, input and output
EP

variables are aggregated by company name, and thus it is not possible to compare

efficiency of various stores in the same group. Gathering more detailed data would
C

enable us to have a better understanding of factors driving efficiency of retailers and,


AC

ultimately, of the prospects of this sector.

37
ACCEPTED MANUSCRIPT

References

ABRAS, 2013. Ranking Abras 2013. Revista Superhiper 39 (442), 80–84.

Arjomandi, A., Valadkhani, A., & O’Brien, M. (2014). Analysing banks’ intermediation
and operational performance using the Hicks–Moorsteen TFP index: The case of Iran.
Research in International Business and Finance, 30, 111-125.

PT
Badunenko, O., & Mozharovskyi, P. (2016). Nonparametric frontier analysis using Stata.
The Stata Journal, 16(3), 550-589.

RI
Banker, R. D., Chang, H., 2005. The super-efficiency procedure for outlier
identification, not for ranking efficient units. European Journal of Operational Research

SC
175 (2), 1311–1320.

Banker, R. D., Charnes, A., Cooper, W. W., 1984. Some models for estimating technical
and scale inefficiencies In Data Envelopment Analysis. Management Science 30 (9),

U
10781092.
AN
Barros, C. P., 2006. Efficiency measurement among hypermarkets and supermarkets
and the identification of the efficiency drivers. International Journal of Retail &
Distribution Management 34 (2), 135–154.
M

Barros, C. P., Alves, C., 2004. An empirical analysis of productivity in a Portuguese


retail chain using the Malmquist Index. Journal of Retailing and Consumer Services 11
(5), 269278.
D

Bjurek, H. (1996). The Malmquist total factor productivity index. The Scandinavian
TE

Journal of Economics, 303-313.

Caves, D. W., Christensen, L. R., & Diewert, W. E. (1982). The economic theory of index
numbers and the measurement of input, output, and productivity. Econometrica: Journal of
EP

the Econometric Society, 1393-1414.

Charnes, A., Cooper, W. W., Rhodes, E., 1978. Measuring the efficiency of decision
making units. European Journal of Operational Research 2 (6),
C

429–444.
AC

Cooper, W. W., Seiford, L. M., Tone, K., 2007. Data Envelopment Analysis: a
comprehensive text with models, applications, references and DEA-solver software.
Springer.

de Jorge Moreno, J. (2010). Productivity growth of European retailers: a benchmarking


approach. Journal of Economic Studies, 37(3), 288-313.

38
ACCEPTED MANUSCRIPT

de Jorge Moreno, J., & Sanz-Triguero, M. (2011). Estimating technical efficiency and
bootstrapping Malmquist indices: Analysis of Spanish retail sector. International Journal of
Retail & Distribution Management, 39(4), 272-288.

Didonet, S. R., Lara, J. E. (2006). Eficiencia productiva y estrategias en la distribución


comercial: el caso de los supermercados brasileños. In: Anais do IX Simpósio de
Administração da Produção, Logística e Operações Internacionais.

PT
Fare, R. S., Grosskopf, S., Lindgren, B., Roos, P. (1994). Data Envelopment Analysis:
Theory, methodology and applications, 253–272.

RI
Färe, R., Grosskopf, S., & Roos, P. (1996). On two definitions of productivity. Economics
Letters, 53(3), 269-274.

SC
Farrell, M. J. (1957). The measurement of productive efficiency. Journal of the Royal
Statistical Society: Series A 120 (3), 253–290.

U
Ferreira, M. A. M., Venancio, M. M., Abrantes, L. A., 2009. Análise da eficiência do setor
de supermercados no Brasil. Economia Aplicada 13 (2), 333–347.
AN
Gandhi, A., & Shankar, R. (2014). Efficiency measurement of Indian retailers using data
envelopment analysis. International Journal of Retail & Distribution Management, 42(6),
500-520.
M

Kato, A. (2015). Productivity and Characteristics of Firms: An application of a


bootstrapped data envelopment analysis to Japanese firm-level data. Journal of Asia-
D

Pacific Studies (Waseda University) No.


TE

Medal-Bartual, A., Molinos-Senante, M., & Sala-Garrido, R. (2015). Assessment of the


Total Factor Productivity Change in the Spanish Ports: Hicks–Moorsteen Productivity
Index Approach. Journal of Waterway, Port, Coastal, and Ocean Engineering, 142(1),
04015013.
EP

Molinos-Senante, M., Sala-Garrido, R., & Hernández-Sancho, F. (2016). Development and


application of the Hicks-Moorsteen productivity index for the total factor productivity
assessment of wastewater treatment plants. Journal of Cleaner Production, 112, 3116-3123.
C

Perrigot, R., Barros, C. (2008). Technical efficiency of French retailers. Journal of


AC

Retailing and Consumer Services 15 (4), 296–305.

O’Donnell, C. J. (2012a). An aggregate quantity framework for measuring and


decomposing productivity change. Journal of Productivity Analysis, 38(3), 255-272.

39
ACCEPTED MANUSCRIPT

O'Donnell, C. J. (2012b). Nonparametric estimates of the components of productivity and


profitability change in US agriculture. American Journal of Agricultural Economics, 94(4),
873-890.

Ray, S. C., & Desli, E. (1997). Productivity growth, technical progress, and efficiency
change in industrialized countries: comment. The American Economic Review, 87(5),
1033-1039.

PT
Kumar Sharma, S., & Dalip, R. (2014). Efficiency and productivity analysis of Indian
banking industry using Hicks-Moorsteen approach. International Journal of Productivity

RI
and Performance Management, 63(1), 57-84.

See, K. F., & Li, F. (2015). Total factor productivity analysis of the UK airport industry: A

SC
Hicks-Moorsteen index method. Journal of Air Transport Management, 43, 1-10.

Sellers-Rubio, R., Mas-Ruiz, F. ( 2006). Economic efficiency in supermarkets:


Evidences in Spain. International Journal of Retail & Distribution Management 34 (2),

U
155–171.
AN
Sellers-Rubio, R., Mas-Ruiz, F. ( 2007). An empirical analysis of productivity growth
in grocery retailing: Evidence from Spain. International Journal of Services Industry
Management 18 (1), 52–69.
M

Sharma, V., Choudhary, H. ( 2011). Measuring operational efficiency of retail stores in


Chandigarh Tri-City using DEA. Journal of services research 10 (2),
100–115.
D

Souza, M. A. F., Macedo, M. A. S., Ferreira, M. S. (2010). Desempenho organizacional no


TE

setor supermercadista brasileiro: uma análise apoiada em DEA. Revista de Gestão 17


(2), 151–167.

Simar, L., & Wilson, P. W. (1998a). Sensitivity analysis of efficiency scores: How to
EP

bootstrap in nonparametric frontier models. Management Science, 44(1), 49-61.

Simar, L., & Wilson, P. W. (1998b). Productivity growth in industrialized countries.


Université Catholique de Louvain. Center for Operations Research and Econometrics
C

[CORE].
AC

Simar, L., & Wilson, P. W. (1999). Estimating and bootstrapping Malmquist indices.
European Journal of Operational Research, 115(3), 459-471.

Simar, L., & Wilson, P. W. (2000). A general methodology for bootstrapping in non-
parametric frontier models. Journal of Applied Statistics, 27(6), 779-802.

40
ACCEPTED MANUSCRIPT

Simar, L., & Wilson, P. W. (2002). Non-parametric tests of returns to scale. European
Journal of Operational Research, 139(1), 115-132.

Simar, L., & Wilson, P. W. (2007). Estimation and inference in two-stage, semi-parametric
models of production processes. Journal of Eeconometrics, 136(1), 31-64.

Simar, L., & Wilson, P. W. (2008). Statistical inference in nonparametric frontier models:

PT
recent developments and perspectives. The measurement of productive efficiency (H.
Fried, CAK Lovell and SS Schmidt Eds), Oxford University Press, Inc, 421-521.

RI
Wilson, P. W. ( 2008). FEAR 1.0: A software package for frontier efficiency analysis
with R. Socio-Economic Planning Sciences 42, 247–254.

SC
Yu, A. S. O., Angelo, C. (2001). Performance of Brazilian supermarkets: a comparative
analysis between large and small store chains. Journal of Small Business and Enterprise
Development 8 (4), 339–348.

U
Yu, W., Ramanathan, R. ( 2009). An assessment of operational efficiency of retail
firms in China. Journal of Retailing and Consumer Services 16 (2), 109–122.
AN
Zofio, J. L. (2007). Malmquist productivity index decompositions: a unifying framework.
Applied Economics, 39(18), 2371-2387.
M
D
TE
C EP
AC

41
ACCEPTED MANUSCRIPT

Table 1: List of relevant mergers and acquisitions that occurred between 2004 and 2012.
Year Chain Mergers and acquisitions
2004 Wal-Mart Brasil In 2004, Walmart acquired 118 stores from Bompreço in the Northeast Brazil.

PT
2004 Abv Comércio de Alimentos Ltda The Chain acquired MaxBom Supermarket.
2005 Grupo Pão de Açúcar Casino and Pão de Açúcar start a Holding called Companhia Brasileira de Distribuição.

RI
2005 Wal-Mart Brasil The group incorporated operations from the Portuguese Sonae, becoming one of the
leading chains in the country.
2006 Comercial Unida de Cereais The chain was generated by the merger between Comercial Unida and Rissul.

SC
2007 Grupo Pão de Açúcar The Pão de Açúcar group starts a joint venture with the Assai Atakarejo.
2007 Gbarbosa Comercial Acquired by the Cencosud Group.

U
2007 Carrefour Carrefour makes an agreement with the Portuguese group Sonae to acquire 10 stores in
the Big hypermarket chain in the state of São Paulo.

AN
2007 Supermercado Bh Com. De Alim The Super Nosso chain negotiated three of its units with the group BH Supermarkets.
2009 Supermercado Gimenes Partially acquired by Ricoy Group.

M
2010 Cooperativa Mista São Luiz Acquisition of the Econômico Supermarket.
2010 Supermercado Ideal Merger with the Casa Branca group.

D
2010 Irmãos Bretas, Filhos e Cia Acquired by the Cencosud Group.
2010 Companhia Sulamericana de Generated by the merger between São Francisco Supermarket and Cidade Canção

TE
Distribuição Supermarket.
2011 Cencosud Acquired Serrana Empreendimentos e Participações Group and Prezunic Group.
2011 Companhia Sulamericana de Acquired its first units in the State of São Paulo through the acquisition of 2 stores of the
EP
Distribuição Passarelli Supermarket chain.
2011 Coopbrasil Seven regional chains operating in São Paulo, Minas Gerais and Santa Catarina
C

established the cooperative CoopBrasil.


2012 Grupo Pão De Açúcar Casino became the sole controlling shareholder of Grupo Pão de Açúcar, Brazil’s leading
AC

retailer
2012 Zoni Supermercados Ltda The group incorporated five stores from the Centro Supermercados.
2012 Cooper Itajaí Cooper acquired five stores from the Breithaupt chain.
ACCEPTED MANUSCRIPT

Table 2: Summary of studies of the retailing sector of countries other than Brazil

PT
Study Sample Inputs Outputs Main conclusions
Barros and Alves (2004) 47 Portuguese retail stores Number of employees, cost Yearly sales and operational Mostly no gain of productivity in

RI
from 1999 to 2000 of labor, number of results. one year. Scales of operation are
checkouts, stocks and other highly associated to efficiency
costs. levels.

SC
Barros (2006) 22 Portuguese supermarkets Total assets and number of Gross sales and earnings. Large groups are on average more
from 1998 to 2003. employees. efficient than smaller groups.
National groups are more efficient

U
than regional groups.
Sellers-Rubio and Mas-Ruiz 100 intermediary Spanish Number of outlets, number Gross sales and earnings. High level of inefficiency in the

AN
(2006) retail sectors from 1995 to of employees and the level Spanish retail sector.
2001. of indebtedness.
Sellers-Rubio and Mas-Ruiz 96 Spanish supermarket Number of stores, number of Gross sales and earnings. Increase in average annual

M
(2007) chains from 1995 to 2003. employees and the level of productivity among firms analyzed.
indebtedness.
Perrigot and Barros (2008) 11 French retailers from Number of employees, value Turnover and profitability. These 11 French supermarkets are

D
2000 to 2004. of assets and operating costs. relatively efficient.
Yu and Ramanathan (2009) 61 Chinese retailers from Sales area and number of Gross revenues and profits 38% of the sample increased in

TE
2000 to 2003. employees. before taxes. productivity.
de Jorge Moreno (2010) Over 1 000 European Number of employees, fixed Gross sales, operational Half of the firms increased their
retailers from 1998 to 2006. assets, cost of labour and results productivity in the period, due both
EP
stock. to technical efficiency change and
technical progress.
de Jorge Moreno and Sanz- Spanish retail sector from Personnel costs, fixed assets Gross sales Decline in productivity over the
C

Triguero (2011) 1997 to 2007. and intermediate period. Positive relationship


consumption. between size and efficiency.
AC

Ghandi and Shankar (2014) 18 Indian retailers from Cost of labour, capital Profit and sales. Average efficiency scores are low.
2008 to 2010. employed Productivity increased from 2008
to 2009 and then declined from
2009-2010.
Kato (2015) Over 3 000 Japanese Labour measured in man- Total sales Productivity did not change much
ACCEPTED MANUSCRIPT

retailers (department stores hours, intermediate costs over the period.


and supermarkets) from and Fixed assets.
1995 to 2004.
Table 3: Summary of studies of the Brazilian supermarkets' sector

PT
Study Sample Inputs Outputs Main conclusions
Supermarkets present in Number of outlets, number
Large chains are more

RI
Yu and Angelo (2001) ABRAS rankings from 1994 of employees and number of Gross annual sales
efficient than small chains.
to 1998. stores.
50 largest supermarkets Number of outlets, number Chains with small sales area

SC
Didonet and Lara (2006) listed by ABRAS ranking of of employees, number of Gross annual sales are more efficient than
2005. stores and total area chains with large store areas.
Bottom tier supermarkets are
300 largest supermarkets of Number of outlets, number

U
Ferreira et al. (2009) Gross annual sales the most efficient. Overall
ABRAS ranking of 2006. of employees and total area.
efficiency is low.

AN
Supermarkets ranked from Number of outlets, number Very large chains are more
Souza et al. (2010) 201 to 300 at ABRAS of employees, number of Gross annual sales efficient than very small
ranking of 2006. stores and total area. chains.

M
D
TE
C EP
AC
ACCEPTED MANUSCRIPT

Table 4: Descriptive statistics of input and output variables for Brazilian supermarket chains in 2005 and
2012.

Panel A: original data

Standard
Year Variable Mean deviation Min Max Obs

PT
Gross sales in millions
(BRL) 137.51 1.06 0.1 16198.96 500
2005 Number of outlets 93.34 622.68 1 9184 500

RI
Area in m2 11795.65 87988.94 70 1206254 500
Number of employees 691.75 4276.93 2 62803 500
Gross sales in millions
242.84 2402.00 0.07 57233.63 709

SC
(BRL)
Deflated gross sales in
170.93 1690.76 0.05 40286.41 709
millions (BRL)
2012 Number of outlets 91.52 688.24 1 14993 709

U
2 12766.82 127269.90 30 2962008 709
Area in m
Number of employees 860.28 6905.82 1 159093 709
AN
Panel B: data without outliers

Standard
Year Variable Mean Min Max Obs
deviation
M

Gross sales in millions


106.40 784.71 0.1 12546.23 494
(BRL)
2005 Number of outlets 75.85 473.75 2 7902 494
D

2
Area in m 9491.34 70254.52 144 1170021 494
Number of employees 572.94 3266.61 4 50112 494
TE

Gross sales in millions


163.74 1090.34 0.07 25932.91 701
(BRL)
Deflated gross sales in
115.26 767.49 0.05 18254.02 701
millions (BRL)
EP

2012 Number of outlets 71.12 401.71 1 9358 701


Area in m2 8683.46 62759.44 100 1533191 701
Number of employees 642.86 3523.21 2 82341 701
C
AC

45
ACCEPTED MANUSCRIPT

Table 5: DEA bootstrap supermarket chains' inefficiency estimates with the constant returns to scale model (CRS) and variable returns to scale model (VRS)
for 2005 and 2012 data periods.

PT
CRS model VRS model
nº of Bias Lower Upper nº of Bias Lower Upper

RI
Year Sample Obs Inefficiency efficient corrected Bound Bound Inefficiency efficient corrected Bound Bound
dmus inefficiency 95% 95% dmus inefficiency 95% 95%
All

SC
494 2.859 9 2.999 2.884 3.195 2.421 24 2.646 2.472 2.996
chains
Chains
also

U
166 2.101 6 2.200 2.121 2.334 1.936 11 2.102 1.978 2.323
present
2005 in 2012

AN
Chains
not
328 3.242 3 3.403 3.271 3.630 2.666 13 2.922 2.722 3.336
present

M
in 2012
All
701 4.059 6 4.371 4.127 4.803 3.243 24 3.587 3.343 4.091

D
chains
Chains

TE
also
166 2.414 1 2.558 2.445 2.762 2.062 4 2.270 2.120 2.563
present
2012 in 2005
EP
Chains
not
535 4.569 5 4.933 4.649 5.437 3.610 20 3.996 3.722 4.565
present
C

in 2005
AC
ACCEPTED MANUSCRIPT

Table 6: DEA bootstrap supermarket chains' inefficiency estimates with the constant returns to scale model (CRS) and variable returns to scale model (VRS) for 2005 and 2012
data periods by classes of ranking positions.

PT
2005 CRS Model 2005 VRS Model
Ranking nº of Mean Bias Lower nº of Mean Bias Lower Upper
Mean Upper Bound Mean
classes Efficient Corrected Bound Efficient Corrected Bound Bound
Inefficiency 95% Inefficiency

RI
2005 DMUs Inefficiency 95% DMUs Inefficiency 95% 95%
1 to 100 1.75 3 1.84 1.77 1.95 1.64 9 1.82 1.68 2.11

SC
101 to 200 2.05 5 2.15 2.07 2.29 1.98 5 2.13 2.03 2.30
201 to 300 2.04 1 2.13 2.05 2.25 1.98 1 2.11 2.02 2.26
301 to 400 2.44 0 2.53 2.46 2.65 2.22 3 2.37 2.26 2.54

U
401 to 500 6.22 0 6.56 6.27 7.06 4.40 6 4.93 4.50 5.96

AN
All Chains 2.86 9 3.00 2.88 3.19 2.42 24 2.65 2.47 3.00

M
2012 CRS Model 2012 VRS Model
Ranking nº of Mean Bias Lower nº of Mean Bias Lower Upper
Mean Upper Mean
classes Efficient Corrected Bound Efficient Corrected Bound Bound

D
Inefficiency Bound 95% Inefficiency
2012 DMUs Inefficiency 95% DMUs Inefficiency 95% 95%

TE
1 to 100 2.04 1 2.18 2.07 2.39 1.57 10 1.80 1.63 2.20
101 to 200 2.32 0 2.44 2.35 2.63 1.96 1 2.12 2.00 2.33
201 to 300 2.47 3 2.63 2.50 2.86 2.22 3 2.41 2.27 2.64
EP
301 to 400 2.47 2 2.63 2.51 2.85 2.36 2 2.59 2.43 2.85
401 to 500 2.88 0 3.07 2.92 3.31 2.72 0 2.98 2.79 3.29
C

501 to 600 3.68 0 3.97 3.75 4.34 3.21 1 3.55 3.31 3.98
AC

601 or more 12.48 0 13.58 12.71 15.13 8.61 7 9.60 8.90 11.27
All Chains 4.06 6 4.37 4.13 4.80 3.24 24 3.59 3.34 4.09
ACCEPTED MANUSCRIPT

Table 7: Summary of scale efficiencies for supermarket chains by their presence in datasets of 2005 and
2012.

Significantly Significantly
Not statistically
inefficient inefficient
Year Sample scale
due to due to Total
inefficient
IRS DRS

PT
All chains 107 2 385 494
2005 Chains also present in 2012 18 0 148 166
Chains not present in 2012 89 2 237 328
All chains 103 13 585 701

RI
2012 Chains also present in 2005 3 4 159 166
Chains not present in 2005 100 9 426 535

U SC
AN
M
D
TE
C EP
AC

48
ACCEPTED MANUSCRIPT

Table 8: Summary of scale inefficiencies for supermarket chains by classes of ranking positions.

Panel A: 2005

PT
Not
Ranking Statistically Statistically
statistically
classes inefficient inefficient Total
scale
2005 due to IRS due to DRS
inefficient

RI
1 to 100 0 0 100 100
101 to 200 0 0 100 100

SC
201 to 300 5 0 95 100
301 to 400 45 1 54 100
401 or more 57 1 36 94

U
Total 107 2 385 494
AN
Panel B: 2012

Not
Ranking Statistically Statistically
M

statistically
classes inefficient inefficient Total
scale
2012 due to IRS due to DRS
inefficient
D

1 to 100 0 0 100 100


101 to 200 0 6 94 100
TE

201 to 300 0 6 94 100


301 to 400 0 0 100 100
401 to 500 3 1 96 100
EP

501 to 600 31 0 69 100


601 or more 69 0 32 101
Total 103 13 585 701
C
AC

49
ACCEPTED MANUSCRIPT

Table 9: Descriptive statistics of input and output variables for chains present in both rankings.
Standard
Year Variable Mean Min Max Obs
deviation
Gross sales in millions (BRL) 163.04 921.37 0 11731.76 166
Number of outlets 113.23 524.54 2 6586.00 166
2005 2
Area in m 14749.01 91175.92 200 1170021 166
Number of employees 845.52 3969.34 5 50112 166

PT
Gross sales in millions (BRL) 386.04 2040.62 1.80 25932.91 166
Deflated gross sales in millions (BRL) 271.73 1436.38 1.27 18254.02 166
2012 Number of outlets 166.23 747.46 2 9358 166

RI
2
Area in m 20991.17 119734.20 240 1533191 166
Number of employees 1474.25 6544.06 5 82341 166

U SC
AN
M
D
TE
C EP
AC

50
ACCEPTED MANUSCRIPT

Table 10: Summary of the Malmquist index and its components for chains present in both rankings
and for subgroups based on gross sales.

PANEL A: Malmquist Index


Mean Mean nº of dmus nº of dmus
nº of dmus with
Mean Lower Upper with with Obs
Group insignificant
Bound bound significant significant
change
95% 95% decrease increase

PT
All chains 1.25 1.20 1.30 20 42 104 166
Small 1.67 1.60 1.76 5 7 29 41

RI
Small to Medium 1.17 1.12 1.20 5 13 24 42

Medium to Large 1.09 1.05 1.12 3 13 25 41

SC
Large 1.09 1.04 1.11 7 9 26 42
Top 5% 1.13 1.08 1.15 1 0 8 9

U
PANEL B: Pure Efficiency Change
nº of
AN
Mean Mean nº of dmus
nº of dmus with dmus
Mean Lower Upper with Obs
Group insignificant with
Bound bound significant
change significant
95% 95% decrease
increase
M

All chains 0.98 0.84 1.07 66 72 28 166


Small 1.07 0.95 1.17 8 22 11 41

Small to Medium 0.93 0.84 1.00 11 23 8 42


D

Medium to Large 0.95 0.85 1.03 18 15 8 41


TE

Large 0.97 0.73 1.10 29 12 1 42


Top 5% 1.10 0.59 1.35 8 0 1 9
EP

PANEL C: Pure Technology Change


nº of
Mean Mean nº of dmus nº of dmus
dmus
Mean Lower Upper with with Obs
Group with
C

Bound bound insignificant significant


significant
95% 95% change decrease
increase
AC

All chains 1.28 1.15 1.40 47 0 119 166


Small 1.55 1.39 1.68 1 0 40 41

Small to Medium 1.27 1.18 1.36 1 0 41 42

Medium to Large 1.16 1.06 1.25 16 0 25 41

Large 1.16 0.96 1.32 29 0 13 42


Top 5% 1.12 0.78 1.40 9 0 0 9
51
ACCEPTED MANUSCRIPT

PANEL D: Scale change


nº of
Mean Mean nº of dmus nº of dmus
dmus
Mean Lower Upper with with Obs
Group with
Bound bound insignificant significant
significant
95% 95% change decrease
increase

PT
All chains 0.92 0.79 1.01 56 92 18 166
Small 1.08 0.91 1.16 22 4 15 41

Small to Medium 0.92 0.84 0.98 12 27 3 42

RI
Medium to Large 0.85 0.75 0.92 6 35 0 41

Large 0.82 0.66 0.97 16 26 0 42

SC
Top 5% 0.77 0.51 0.99 4 5 0 9

PANEL E: Scale technology change

U
nº of
Mean Mean nº of dmus nº of
dmus
Mean Lower Upper with dmus with Obs
AN
Group with
Bound bound insignificant significant
significant
95% 95% change decrease
increase
All chains 1.10 1.00 1.22 64 15 87 166
M

Small 0.96 0.91 1.04 25 13 3 41

Small to Medium 1.09 1.02 1.16 16 2 24 42


D

Medium to Large 1.18 1.09 1.29 6 0 35 41


TE

Large 1.20 0.99 1.40 17 0 25 42


Top 5% 1.22 0.80 1.53 7 0 2 9

a Groups are defined based on quartiles of gross sales of 2005.


C EP
AC

52
ACCEPTED MANUSCRIPT

Table 11: Descriptive statistics of explanatory variables of chains present in both datasets.

Standard
Year Variable Mean Min Max Obs
deviation
nstores2005: Number of stores 12.88 57.73 1 542 165
2005 size2005: Average store size 1398.29 1319.28 45.54 10500 165
d1store2005: Dummy for 1 store 0.33 0.47 0 1 165

PT
nstores2012: Number of stores 20.72 116.87 1 1377 165
size2012: Average store size 1479.56 1398.90 50.75 11000 165
d1store2012: Dummy for 1 store 0.25 0.44 0 1 165

RI
nstates2012: Number of states 1.33 2.10 1 21 165
ncities2012: Number of cities 6.07 27.43 1 290 165
nbrands2012: Number of brands 1.19 0.74 1 9 165

SC
2012 dconv: Dummy for Convenience store chains 0.01 0.11 0 1 165
dcoop: Dummy for Cooperatives 0.08 0.27 0 1 165
dN: Dummy for North 0.03 0.17 0 1 165

U
dNE: Dummy for North-East 0.11 0.31 0 1 165
dCW: Dummy for Center-West 0.07 0.26 0 1 165
AN
dSE: Dummy for South-East 0.41 0.49 0 1 165
dS: Dummy for South 0.38 0.49 0 1 165
dnational: Dummy for national chains 0.02 0.15 0 1 165
M
D
TE
C EP
AC

53
ACCEPTED MANUSCRIPT

Table 12: Second stage regression results for the 2012 bias corrected inefficiency scores with variable (VRS)
and constant (CRS) returns to scale based on robust ordinary least squares.
2012 VRS 2012 CRS
bias corrected bias corrected
VARIABLES inefficiency levels inefficiency levels
size2012 -0.000477*** -0.000416***

PT
(9.53e-05) (7.26e-05)
nstores2012 -0.00249 0.00121
(0.00265) (0.00196)

RI
dconv 0.371 0.285
(0.869) (0.762)
dcoop -0.142 0.0581

SC
(0.278) (0.231)
d1store2012 0.0788 -0.107
(0.158) (0.142)
ncities2012 0.00715 -0.00318

U
(0.0128) (0.00984)
AN
d2states -0.147 0.0497
(0.296) (0.240)
d3states 0.711 1.188
(2.678) (2.029)
M

d2brands -0.334 -0.168


(0.270) (0.223)
d3brands -0.803 -0.685
D

(0.629) (0.462)
dN 0.222 0.564
TE

(0.508) (0.414)
dNE -0.0806 0.00821
(0.240) (0.212)
EP

dCW 0.121 0.115


(0.291) (0.254)
dS 0.233 0.289**
C

(0.155) (0.138)
dnational -0.517 -0.966
AC

(2.383) (1.808)
Constant 2.742*** 2.988***
(0.164) (0.142)
sigma 0.764*** 0.719***
(0.0559) (0.0474)
Observations 165 165
Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1

54
ACCEPTED MANUSCRIPT

Table 13: Second stage regression results for the Malmquist index and its components based on robust
ordinary least squares.

Malmquist Pure Pure Scale


VARIABLES Index Efficiency Technology Scale Techonology
Dsize -0.000949 -0.00137* 0.000657** 0.000732** -0.000160

PT
(0.00140) (0.000767) (0.000313) (0.000344) (0.000169)
Dnstores -0.000789 -0.00131 0.000577 0.000343 -0.00101
(0.0161) (0.00815) (0.00355) (0.00342) (0.00267)

RI
dconv 0.197 0.0817 -0.146 0.0356 0.217
(6.283) (3.607) (1.662) (1.380) (1.169)
dcoop -0.223*** -0.108* -0.0437 -0.0738** 0.0464

SC
(0.0826) (0.0610) (0.0758) (0.0300) (0.0392)
d1store2005 0.348** 0.0522 0.236*** 0.162*** -0.148***
(0.151) (0.0913) (0.0347) (0.0236) (0.0197)

U
ncities2012 0.00117 0.00349 -0.000956 -0.00138 0.00222
(0.0152) (0.00899) (0.00659) (0.00441) (0.00470)
AN
d2states -0.109 -0.0253 -0.00774 -0.0658* 0.0149
(0.132) (0.0819) (0.0896) (0.0349) (0.0481)
d3states -0.0982 -0.106 -0.0120 -0.0664 -0.0197
(5.061) (1.794) (1.785) (1.287) (1.087)
M

d2brands -0.120 -0.0259 -0.0599 -0.0419** 0.0162


(0.109) (0.0749) (0.0378) (0.0208) (0.0254)
D

d3brands -0.116 -0.0740 -0.0524 0.0221 -0.0372


(1.010) (0.157) (0.320) (0.119) (0.255)
TE

dN -0.297** -0.116 -0.0270 -0.123* 0.0470


(0.126) (0.0738) (0.0853) (0.0711) (0.0480)
dNE 0.459 0.317 0.0148 0.00212 0.0156
(0.505) (0.297) (0.0548) (0.0430) (0.0351)
EP

dCW -0.0108 0.00618 0.0197 -0.00528 -0.0132


(0.123) (0.102) (0.0487) (0.0265) (0.0305)
dS 0.0167 0.0315 0.0133 -0.0178 0.0109
C

(0.0835) (0.0544) (0.0298) (0.0205) (0.0169)


dnational 0.0571 0.0261 -0.0487 0.0439 -0.00213
AC

(0.176) (0.109) (0.0920) (0.0412) (0.0486)


Constant 1.146*** 0.948*** 1.206*** 0.882*** 1.136***
(0.0778) (0.0507) (0.0335) (0.0190) (0.0207)
Observations 165 165 165 165 165
R-squared 0.114 0.072 0.367 0.457 0.408
Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1

55
ACCEPTED MANUSCRIPT

Gross Sales
(output)
2.50

PT
2.00
1.50
1.00

RI
0.50 MI < 1
Employees
0.00 Outlets (input) MI =1
(input)

SC
MI > 1

U
Area(input)
AN
Fig 1: Ratio of average values of output (gross sales) and inputs (number of outlets, number of employees and
total area) in 2012 with respect to average values of 2005. Based on the Malmquist index, supermarket chains
are divided into 3 groups: those with significant productivity gains (MI >1), those with insignificant
M

productivity gains (MI = 1) and those with significant decrease in productivity (MI = 1).
D
TE
C EP
AC

56

You might also like