October 2015
Explaining Cross-Country Productivity Differences in Retail Trade
David Lagakos University of California, San Diego and NBER
ABSTRACT —————————————————————————————————Many macroeconomists argue that productivity is low in developing countries because of frictions that impede the adoption of modern technologies. I argue that in the retail trade sector, developing countries rationally choose technologies with low measured labor productivity. My theory is that the adoption of modern retail technologies is optimal only when household ownership of complementary durable goods, such as cars, is widespread. Because income is low in the developing world, households own few such durables. The theory implies that policies that lead to large increases in measured retail productivity do not necessarily increase welfare. ——————————————————————————————————————–
Email:
[email protected]. For guidance and support on this project I especially thank Matthias Doepke and Lee Ohanian. The paper has benefited greatly from comments by the editor, Sam Kortum, two anonymous referees, and many others, including Andy Atkeson, Steve Berry, Mark Bils, Paco Buera, Ariel Burstein, V.V. Chari, Allan CollardWexler, Alex Galetovic, Christian Hellwig, Berthold Herrendorf, Bart Hobijn, Tom Holmes, Roozbeh Hosseini, Hugo Hopenhayn, Joe Kaboski, Greg Kaplan, Narayana Kocherlakota, Jeff Lin, Igor Livshits, Ellen McGrattan, Gabriel Natividad, Guillermo Ordo˜nez, Richard Rogerson, Jim Schmitz, Gustavo Ventura, Jon Vogel, Mike Waugh, Mark Wright and Randy Wright, plus seminar participants at Arizona State, Chicago Booth, the Federal Reserve Banks of Minneapolis and St. Louis, the Federal Reserve Board, Iowa, Maryland, NYU Stern, Pennsylvania, Rochester, UC Berkeley Haas, UCLA, UCSD, USC Marshall, Wharton, Yale, the 2007 SED in Prague, the Latin American TFP Conference at UCSB, and the 2008 NBER Summer Institute. For help with data sources, I thank Alexis Antoniades, Sofronis Clerides, Karin Murphy, Marcelo Ore V´asquez, Javier P´erez Estrada and Brooke Tosi. For editorial assistance, I thank Rena Henderson and Kathy Rolfe.
1 Introduction Why are observed differences in per-capita income across countries so large? Development accounting studies have concluded that cross-country differences in physical or human capital per worker can explain only a minor part of the observed income gaps. Instead, variations in productivity account for most of the income differences. Unfortunately, aggregate productivity is still poorly understood (Klenow and Rodr´ıguez-Clare, 1997; Prescott, 1998; Hall and Jones, 1999; Caselli, 2005). In this paper, I shed light on aggregate differences in output per worker by explaining productivity differences in one important sector: retail trade. My study of retail trade contributes to an understanding of aggregate productivity differences in two ways. First, the retail sector forms a large fraction of the aggregate, employing just under twenty percent of all workers in a typical country (Bosworth and Triplett, 2004), and is thus directly relevant for understanding the aggregate. Second, my results lead to new implications for why aggregate productivity differences might arise and for the types of government policies that keep productivity low. Specifically, I argue that low measured productivity in developing-country retail trade largely represents optimal choice of technology adoption given these countries’ low income level. This is in stark contrast to arguments that low measured productivity is the result of barriers to entry or technology adoption (Parente and Prescott, 1994, 1999; Herrendorf and Teixeira, 2011); of a lack of competitive pressure (Schmitz, 2005); or of policies that misallocate resources across producers (Hsieh and Klenow, 2009; Restuccia and Rogerson, 2008). My analysis consists of two basic parts. The first part draws on disaggregated data from censuses of retail trade to measure and account for labor productivity differences in the retail sectors of the United States and a set of developing countries. The data show that retail-sector productivity differences are largely accounted for by compositional differences, or, specifically, by differences in the relative use of “modern” retail technologies, such as superstores, and “traditional” technologies, such as mom-and-pop shops. The modern technology is used widely in the United States but much less frequently in the developing world. This is true despite the fact that, in the developing countries, measured productivity of the modern technology is roughly twice as high as that of the traditional technology. The second part of my analysis builds a theoretical framework to help understand cross-country differences in measured productivity and the composition of technologies used in retail trade. The model formalizes two basic channels that have been emphasized in the literature. The first is the role of cars as complementary inputs to modern retail technologies; cars help households economize on shopping time, and this is particularly useful for modern retail technologies, which require large inputs of household shopping time. The second channel is informality among operators of 1
traditional technologies, who can more easily operate informally due to their small size. Informality provides an advantage to traditional technologies because tax evasion lowers the relative price of their output. I begin with a simple version of the model that illustrates the qualitative predictions of the theory. In the model, two retail technologies – modern and traditional – are freely available in all countries. Both use labor to produce output, and the modern technology is more efficient at doing so. Entrepreneurs face taxes on labor input, and those choosing the traditional technology may evade a larger fraction of the taxes. Both technologies require an input of “shopping services” from the households, and the modern technology requires a relatively larger input. Households produce shopping services using time, and cars increase the efficiency of time in producing these services. The main tradeoff that households face is one of price versus time, with the modern retailers offering lower prices but requiring more shopping time. The simple version of the model offers several insights. First, the use of the modern technology is greater whenever car ownership rates are higher, and lower whenever the traditional technology affords more opportunity for tax evasion. Second, measured productivity differences between the two technologies reflect only their relative costs of hiring labor, but not efficiency differences. Instead, efficiency differences are reflected in relative prices. Third, greater use of the modern technology is associated, under certain conditions, with higher consumption but less leisure for the households. These findings suggest that outcomes involving higher measured productivity are not necessarily more efficient, and that policies that raise measured productivity may not raise welfare. To assess the quantitative implications of the theory, I next build a richer version of the model. This quantitative version endogenizes the car-ownership decision, adds capital, and introduces a more realistic set of taxes on producers. I parameterize the model to resemble the United States– in particular in its high employment share at modern retailers and high car-ownership rate. I then compute the model’s prediction when income and tax-evasion rates are changed to match the levels for each developing country. I find that the model explains the bulk of the cross-country differences in retail-sector productivity and the composition of retail technologies employed. Both the cars channel and tax evasion channel are important quantitatively. Decomposing the results suggests that the cars channel explains roughly two thirds of the cross-country differences in the relative use of the modern technology, while the informality channel explains roughly one third. I then use the model to conduct two counterfactual policy experiments. The first simulates reducing the informality rate of the traditional retail segment by one half. The model predicts a 9.6-percent increase in measured productivity in the retail sector as employment moves from the traditional to the modern retail segment. While consumption rises as a result, leisure falls, as households spend more time shopping. The overall welfare gain is just 0.2 percent in consumption equivalents. The 2
second policy experiment simulates a liberalization of the market for imported cars. The model predicts a 3.5-percent increase in measured productivity in the retail sector, as employment again moves to the modern segment. This time, consumption rises and leisure rises slightly due to the more efficient shopping provided by cars, resulting in a 0.6-percent increase in welfare. The overall lesson from the experiments is that measured productivity increases due to policy changes are a poor guide to welfare, and that policies that lead to large measured productivity increases may not lead to large increases in welfare. I conclude by providing two additional types of evidence supporting the role of cars in explaining the prevalence of modern retail technologies. The first looks at geographic variation within the United States and Mexico, based on county-level data on income, car-ownership rates, and employment shares in the modern retail segment. The data show strong correlations among all three variables in both countries, as the theory would predict. The second additional piece of evidence comes from a natural experiment on the island of Cyprus, where policy changes led, over a relatively short period, to large increases in the stock of used cars in use. The data show that modern retail shares rose substantially over the same period, again consistent with the theory. The paper suggests new implications for how economists think about measured productivity differences across countries. One implication is that developing countries may rationally choose some technologies with low measured productivity given that overall efficiency (and, hence, average income) is low. In this vein, my work is similar to that of Acemoglu and Zilibotti (2001), Basu and Weil (1998) and Caselli and Coleman (2006), who argue that developing countries may optimally choose different technologies than richer countries. These studies argue that poor countries may adopt different technologies because of low endowments of skilled labor. My work also shares some of the flavor of the work of Jones (2011), who argues that low productivity at the sector level can lead, through weak linkages between sectors, to lower productivity in the aggregate. Furthermore, my work shares some elements of the “big push” theory of development of Rosenstein-Rodan (1943) and Murphy, Shleifer, and Vishny (1989), in which a prerequisite for each sector to modernize is that all other sectors modernize as well. A second implication of my paper is that certain household durable goods, such as cars, and household time spent acquiring goods and services, are factors of production that are missing from cross-country productivity comparisons. In retail trade, this is relevant since these household inputs are complementary to modern technologies. My theory also would apply to large segments of the service sector, since households spend their time and durable goods as inputs into acquiring market services. In this way, my work builds on the tradition in macroeconomics that assigns home production a central role in driving aggregate phenomena (for example, Benhabib and Wright (1991); Parente, Rogerson, and Wright (2000)). The work most closely related to mine is that 3
of Buera and Kaboski (2012), who look at the role of home production in the rise of the service economy in the United States. To the best of my knowledge, my paper is the first to link home production to differences in technology adoption rates across countries, and to the diffusion of modern retail stores in the United States (Foster, Haltiwanger, and Krizan, 2006; Jia, 2008; Holmes, 2011).1
2 Cross-Country Productivity Differences in Retail Trade In this section, I use new economic census data to account for cross-country productivity differences in retail trade. The work in this section builds on, and is inspired by, the work of the McKinsey Productivity Studies for retail trade, summarized by Baily and Solow (2001) and Lewis (2005). The analysis in the current paper improves on the McKinsey measurements by using publicly available data based on large-scale nationally representative surveys of establishments. More importantly, the current paper also offers a different interpretation of the productivity measures in question (starting in Section 3).
2.1 Measuring Retail Productivity Using Census Data My analysis draws on censuses of retail trade for every developing country in which (to my knowledge) a census of retail establishments has been conducted since 2000, and which satisfy several criteria related to data quality (described below). The data are from Argentina (2004), Brazil (2002), El Salvador (2004), Mexico (2003), the Philippines (2005), and Thailand (2002). For comparisons, I also draw on data from the United States for the same years as the foreign censuses, using retail-sector value added and employment data from the U.S. Bureau of Economic Analysis (BEA.) See Appendix A.1 for a more complete description of each of the censuses. The criteria I use to ensure data quality and comparability are as follows. First, the censuses must be nationally representative and have no restriction on the size of the establishments included. This is important because small establishments employ a large fraction of total employment. Second, the censuses must all define value added and labor input in the same way, as detailed below. Third, they must report value added and labor input by size category of establishment, allowing one to look beneath the surface of the aggregate data. The main limitation of the data is that they are available for only a small number of developing countries. 1 My
work also relates to the literature on international trade focusing on the role of the retail sector (and the nontradable service sector more generally) in understanding movements in real exchange rates (see e.g. Burstein, Neves, and Rebelo (2003) and Burstein, Eichenbaum, and Rebelo (2005) and the references therein).
4
My measure of labor productivity is value added per worker, which is arguably the best simple measure of the value of the “composite bundle of services” produced by the retailer and attached to the goods being sold (Oi, 1992). In each census, value added is defined as the value of goods sold minus the cost of goods purchased for resale. Other purchased intermediates, such as electricity, are excluded, although the value of these inputs is typically a small fraction of all intermediate purchases. In order to facilitate international comparisons, I express value added in each country in international dollars using the purchasing power parity (PPP) exchange rate for GDP from the Penn World Tables (Heston, Summers, and Aten, 2011). In each census, labor input is defined as the total number of paid and unpaid workers plus proprietors. The inclusion of unpaid workers (usually family workers) and proprietors is crucial, as they form a large fraction of retail workers in most developing countries. I report value added per worker in each country for the retail sector as a whole, as well as for two segments: the modern segment and the traditional segment. I define the modern segment as all establishments with 20 or more workers and the traditional segment as all establishments with fewer than 20 workers. The motivation for this cutoff is that larger stores have been shown to use more-advanced distribution techniques and better technologies (see Foster, Haltiwanger, and Krizan (2006) and the references therein) and to have more-efficient management practices (Bloom, Genakos, Sadun, and Van Reenen, 2012).
2.2 Measured Retail Productivity is Low in Developing Countries Figure 1 shows the computations for labor productivity for the retail sector as a whole in each country. Value added per worker in the United States retail sector is normalized to 100. The figure shows that the developing countries all have productivity that is far below the U.S. level. Argentina is the highest, with 29 percent of the U.S. level, followed by Mexico at 27 percent, Thailand at 23 percent, Brazil and El Salvador at 20 percent, and the Philippines at 7 percent. The results suggest that the retail-sector productivity gaps largely mimic those of the aggregate, with large productivity differences separating the United States from the developing world. According to the Penn World Tables, in the same years as the retail census discussed here, per capita incomes in PPP dollars compared to those of the United States were: Argentina at 30 percent, Mexico at 34 percent, Thailand at 20 percent, Brazil at 21 percent, El Salvador at 17 percent, and the Philippines at 8 percent.
5
100 100
80
60
40 29
27 23
20
20
20 7 0
US
Argentina
Mexico
Thailand
Brazil
El Salvador Philippines
Figure 1: Value Added Per Worker in Retail Trade
2.3 Measured Productivity is Much Higher in the Modern Segment I now use the disaggregate data to shed light on what is driving these large sectoral productivity differences. Figure 2 shows labor productivity by retail segment, with the United States retail sector as a whole again normalized to 100. The most dramatic feature to note is that, in each of the developing countries, value added per worker is much higher in the modern segment than in the traditional segment. In Argentina, value added per worker is 50 percent as high as in the U.S. retail sector, compared to just 22 percent in the traditional segment. Mexico is next, at 51 percent and 20 percent in the two segments, followed by Thailand at 47 percent and 20 percent, Brazil at 36 percent and 13 percent, El Salvador at 48 percent and 16 percent, and the Philippines at 18 percent and 6 percent. The figure suggests that each country has access to two different retail technologies, one of which having much higher measured labor productivity than the other.2
2.4 Modern Share of Retail Employment is Low Having measured productivity differences by type of retail segment, I now turn to the question of how prevalent each segment is. Figure 3 shows the percent of retail trade employment in each 2 Unfortunately,
value added per worker by size class of retail establishment cannot be calculated in the United States. The reason is that the main source of data on U.S. retail establishments, the Census of Retail Trade, does not collect information on intermediate expenditures, which are required to construct establishment-level value added.
6
60 51
50
48
50
48
40
36
30 22 20 18
20
18 16 13
10
6
0
Argentina
Mexico
Thailand
Modern
Brazil
El Salvador
Philippines
Traditional
Figure 2: Value Added Per Worker by Type of Retail Segment segment, with the two bars for each country summing to 100 percent. For expositional purposes, I refer to the fraction of employment in the modern segment as the modern retail share. The large differences in modern retail shares across countries are immediately apparent. In the United States, 67 percent of retail employment is in the modern segment, with just 33 percent in the traditional segment. In each of the developing countries, in contrast, the modern retail share is far lower. In Argentina, just 25 percent of workers are employed in the modern segment, compared to 23 percent in Mexico, 19 percent in Thailand, 21 percent in Brazil, and 15 percent in El Salvador and in the Philippines. Taken together, Figures 2 and 3 suggest that composition differences play a prominent role in accounting for the low overall retail labor productivity in these developing countries compared to the United States. In other words, a substantial fraction of the productivity differences in the retail trade sectors of the developing countries is mechanically explained by the low use of the modern technology compared to the United States. The developing countries in question could greatly increase measured output per worker in retail trade by using the United States’s composition rather than their own. So why don’t they?
7
100 85 81 80
77
75
85
79
67 60
40
33 25
23
21
19 20
15
15
0
US
Argentina
Mexico
Modern
Thailand
Brazil
El Salvador Philippines
Traditional
Figure 3: Employment Share by Type of Retail Segment
2.5 Discussion: Why is Modern Retailing Not More Prevalent in Poor Countries? In this section, I discuss two prominent (and contrasting) theories of why modern retail technologies are used less frequently in developing countries than in the United States. One theory is that traditional retail technologies offer an opportunity for entrepreneurs to operate informally, thus earning a price advantage over modern retail technologies, which are larger in scale and cannot evade taxes as easily as smaller, traditional stores (see, e.g., De Soto (1989); Lewis (2005); Robles (2009); Ordo˜nez (2014)). Since informality is particularly common in developing countries, traditional retail establishments are much more prevalent there than they otherwise would be. One can view this, alternatively, as a barrier to technology adoption (Parente and Prescott, 1994, 1999; Herrendorf and Teixeira, 2011) or as a form of misallocation across producers (Hsieh and Klenow, 2009; Restuccia and Rogerson, 2008). Broadly speaking, the policy implication of these theories is that policy makers in developing countries should remove distortions in the retail sector.3 3 One
concrete example of distortions affecting the retail sector are size-dependent policies, which place greater legal barriers in the way of larger establishments or firms than smaller ones. Guner, Ventura, and Xu (2008) show that explicit restrictions limiting the presence of large-scale retailers are prevalent in Japan, France, Italy and Korea are quantitatively important in explaining the limited use of large-scale stores there. Another country that has famously blocked the entry of large-scale retailers is India, which has had a long-standing ban on Foreign Direct Investment (FDI) in retailing. Still, for many other countries, including the set studied in the current paper, laws that directly block large-scale retailers do not seem to be present. In its country reports on the retail trade industry, Euromonitor
8
A second theory is that low car-ownership rates among households in developing countries makes modern stores less attractive there. Because of their large scale of operation, modern stores must locate further than traditional stores from residential centers in order to operate. Cars (and other complementary durables, such as refrigerators) allow households to reduce shopping time at modern stores by cutting down on travel time per trip and/or the number of shopping trips (Fuchs, 1969; Oi, 1992; Bromley, 1998; Goldman, 2000).4 Households without these complementary durable goods find themselves with high time costs of shopping at modern stores. Since households in poor countries elect to buy fewer of these durable goods, fewer modern stores operate. Since low car-ownership rates are natural given the low income levels in the developing world, one can view this theory as a variant of the “appropriate technology” theory of cross-country differences in the type of technologies used (Acemoglu and Zilibotti, 2001; Basu and Weil, 1998; Caselli and Coleman, 2006). Again, broadly speaking, this class of theories suggests that the low use of modern technologies in retailing in developing countries is a natural outcome given low productive efficiency (and income) in general, and that policy makers should concentrate on raising efficiency in the economy as a whole, rather than attempting to remove distortions specific to the retail sector.
3 Simple Model This section presents a simple model that highlights how the explanatory factors of interest – carownership rates and tax evasion – determine the structure and measured productivity of the retail sector. The model predicts that differences in measured labor productivity across retail technology types do not depend on relative efficiencies, but do depend on differences in effective labor costs induced through differential tax evasion. Furthermore, it predicts that the prevalence of the modern technology depends on its relative efficiency, the prevalence of cars, and the extent of differential tax evasion. International reports that laws restricting new entry among certain retail format types are not present in Brazil, Mexico, the Philippines or Thailand; nor are laws banning FDI in retailing (see http://www.euromonitor.com/retailing.) 4 For example, Bromley (1998) argues that in Latin America, “low mobility constrains most consumers to patronising traditional retail facilities which can be reached by foot or bus, whilst the small, affluent car-owning sector of the population comprises the principal clientele of the modern supermarket and planned shopping centres.” Similarly, Goldman (2000) argues that in China, although “appropriate sites [for modern stores] are more easily available in the outlying areas, these areas sparsely populated and large-scale stores there will need to draw customers from larger distances. However, since the ownership of motorized vehicles (cars or motorcycles) is very limited, this is not feasible at present.”
9
3.1 Households The economy is populated by a unit measure of households that have preferences over consumption goods and leisure time. Each household is endowed with one unit of time, which it can use to supply to the labor market, shop for consumption goods, or take leisure. Households spend n units of time working (inelastically, for simplicity) and s units of time shopping for consumption goods (which they choose). Preferences of household i are given by U i = log(ci ) + ψ i log(1 − n − si ),
(1)
where the parameter ψ i represents household i’s value on leisure. The ψ i terms are drawn from a distribution G(·) with support on the positive reals and density unbounded above. Households also differ in whether they own a car, with an exogenous fraction α ∈ (0, 1) being car owners. As will be explained below, car owners have lower transportation costs. In order to acquire consumption goods, the household must purchase the goods and, in addition, supply “shopping services.”5 The households may elect to shop in either the modern or the traditional retail segment. The amount of shopping services required depends on the segment chosen and whether or not the household owns a car. Formally, the shopping input for household i and retail segment j ∈ {m,t} is given by sij =
γ · ν ν
j
j
if i is a car owner
(2)
otherwise,
where ν j denotes the amount of shopping services required at segment j, and γ denotes the time savings of a car. Shopping in the modern segment requires a relatively larger time cost: νM > νT . This assumption is a crude but tractable way of capturing the friction associated with modern retail stores – namely, that their larger scale of operations forces them to locate further from the typical household and, hence, requires a larger input of a household’s travel time in order for market transactions to occur. I also assume that γ ∈ (0, 1), meaning that a car cuts down on the required shopping time in either retail segment. This can be thought of as arising from faster travel time per trip, as well as from economizing on the number of shopping trips. 5 This
formulation builds on the ideas of Becker (1965) and subsequent work by Greenwood, Seshadri, and Yorukoglu (2005) and Buera and Kaboski (2012), in which households produce consumption goods by combining home inputs and market inputs.
10
The budget constraint for household i is given by pt · yti + pm · yim ≤ w · n,
(3)
where p j is the market price in segment j; yij is household i’s quantity purchased in j; and w is the wage rate.
3.2 Production and Taxation Market production is done in two distinct sectors: the intermediate sector and the retail sector. The intermediate sector produces an intermediate good, X . The retail sector purchases the intermediate good, combines it with retail services, and then sells it to the households. Production technologies in both sectors are operated by perfectly competitive entrepreneurs, and there is unrestricted access to any production technology. The production function of the intermediate sector is: X = E · Lx ,
(4)
where E denotes general efficiency of production and Lx is the labor input employed in production. Throughout, I normalize the price of the intermediate good to be one. The retail sector produces a consumption good, Y , by combining the intermediate good with retail services. Two different technologies are available for producing retail services: modern and traditional, each of which uses labor as an input. The two technologies are best thought of as representing the output of a retail segment, or aggregate of stores in each technology type, rather than as the output of a particular store. Letting j ∈ {m,t} index the technology type, the production function is given by Y j = min[E · Z j · L j , X j ],
(5)
where E · Z j is the efficiency of labor in producing retail services in technology j, and L j is the labor input used in store type j. Of the two, the modern one is more efficient at providing retail services; that is, Zm > Zt .6 The government levies a tax τ L on employment. Entrepreneurs may evade some fraction of this tax, however, depending on which technology they use. Let the tax compliance rates of modern and traditional retail entrepreneurs be denoted φm ∈ [0, 1] and φt ∈ [0, 1], so that the effective tax 6
The assumption of no substitutability between the intermediates and the service provided by the retailer is motivated by the very definition of retail trade. The North American Industrial Classification System (NAICS) defines the retail trade sector as “establishments engaged in retailing merchandise, generally without transformation,” implying a separation between producing the goods themselves and the services added to sell them.
11
rates are φm · τ L and φt · τ L . I assume that φm > φt , which is meant to capture the fact that smaller stores have an easier time evading taxes in general. The profits to entrepreneurs running retail technology j is then given by
π j = VA j − (1 + φ j τ L )wL j ,
(6)
where VA j is value added at segment j, defined as VA j = p jY j − X j . For now, I assume that all tax revenues are spent on a public good that does not provide utility directly to the households, and that there are no other taxes on producers (both assumptions are relaxed in Section 4, to follow).
3.3 Solution to Household Problem The problem of a household is to choose yim and yti to maximize its utility, (1), subject to the shopping input requirement, (2), and the budget constraint, (3). Because goods are perfect substitutes as final consumption, households will make all their purchases at one of the two retail segments. The key tradeoff that households face is one of time savings versus cost savings. Shopping in the modern segment takes more time but offers lower prices (demonstrated below).7 Beginning with car owners, one can show that if both technologies are used in equilibrium, there is a cutoff in ψ i (the taste for leisure) such that households with ψ i above the cutoff always choose the traditional store, while households below the cutoff always choose the modern store. The same is true for households that do not own a car, but with a higher cutoff. Letting ψ˜ (1) and ψ˜ (0) be the cutoffs for households with and without a car, respectively, one can show that
ψ˜ (a) =
log(pm /pt ) log((1 − n − aγνm + (1 − a)νm )/(1 − n − aγνt + (1 − a)νt )).
(7)
As (7) shows, the cutoffs are higher, all else equal, when the relative price in the modern segment is lower, and whenever the relative shopping input in the modern segment is lower. Figure 4 illustrates the solution to the household problem graphically. For households with ψ i below ψ˜ (0), the price savings of the modern segment always outweigh the time savings of the traditional segment, whether or not the household owns a car. For households with ψ i between ψ˜ (0) and ψ˜ (1), those that are car owners prefer the price savings of the modern segment, while the rest prefer the time savings of the traditional segment. Finally, for ψ i above ψ˜ (1), all households prefer the traditional segment. The fact that ψ˜ (0) < ψ˜ (1) follows from the assumption that γ < 1. 7 The
basic tradeoff between shopping time and price paid is consistent with the work of Aguiar and Hurst (2007) and McKenzie and Schargrodsky (2005). Both studies find strong evidence that households that spend more time shopping tend to pay lower prices on average.
12
Modern if car Traditional otherwise
Modern
0
ψ~ (0)
Traditional
ψ~ (1)
Taste for Leisure, ψ i Figure 4: Solution to the Household Problem, Simple Model
3.4 Equilibrium Properties An equilibrium consists of prices and quantities such that households and firms optimize and markets clear. In what follows, I characterize several important properties of the model’s equilibrium. Throughout, I assume that parameters are such that both retail technologies are used in equilibrium, which happens when Zm /Zt > (1 + φm τ L )/(1 + φm τ L ). Derivations are provided in Appendix B.1. Differences in value added per worker reflect relative labor costs The first important equilibrium prediction of the model characterizes the ratio of value added per worker in the modern to traditional segments. Specifically: VAm /Lm 1 + φm τ L . = VAt /Lt 1 + φt τ L
(8)
Equation (8) says that in equilibrium, the ratio of value added per worker by sector depends on relative tax rates, not on relative efficiency. The reason is that, in equilibrium, prices adjust so that the value of the marginal (and average) product of labor equals the marginal cost for each technology. That is, VA j /L j = w(1 + φt τ L ). The ratio of average products, therefore, equals the ratio of labor costs, which are higher in the modern technology the greater is the extent of tax evasion by the traditional one. Note that Zm and Zt do not feature in equation (8). Thus, it is misleading to assume that measured differences in value added per worker are informative about efficiency differences across the technology types. Differences in efficiency are reflected in relative prices Where do efficiency differences between modern and traditional retail technologies show up then, if not in differences in value added per worker? One answer is in price differences between the two
13
segments. In equilibrium, the price of consumption goods in segment j ∈ {m,t} is given by pj = 1+
1 + φ jτ L . Zj
(9)
Thus, retailers in segment j each charge the price they pay for the intermediate good plus a margin that depends on the effective tax rate and efficiency. The margin is increasing in the rate of tax compliance, φ j , and decreasing in the efficiency of production, Z j . The intuition is that either increasing the fraction of taxes paid or decreasing the efficiency of production leads to higher costs of providing retail services. Through competition, this leads to higher prices in equilibrium. Retail-sector value added per worker is determined by composition Value added per worker in the retail sector as a whole can be written as a linear combination of value added by retail segment. Defining retail-sector value added as VAr = VAm + VAt and retail-sector employment as Lr = Lm + Lt , VAr /Lr = E µ (1 + φm τ L ) + (1 − µ )(1 + φt τ L ) ,
(10)
where µ = Lm /Lr is the modern retail share. Thus, value added per worker in the retail sector as a whole is a function of general efficiency of production, the effective tax rates in each sector, and the modern retail share. Composition determined by car ownership and tax evasion What determines the composition of retail technologies used in equilibrium? One can show the following. First, the modern retail share, µ , is higher when car ownership, α , is higher. The reason is that for any equilibrium prices, car owners are more likely to shop in the modern segment. One can see this from the equilibrium cutoffs of (7): cars reduce the relative time cost of shopping in the modern segment but do not change the relative price. Second, µ is higher when the relative rate of tax compliance, φm /φt , is lower. Put differently, µ is higher when the traditional segment has a more level playing field with the modern segment in terms of tax compliance. This can be seen from the equilibrium prices, (17), which are higher in the traditional segment the higher is φt . This leads a larger fraction of households to shop in the modern segment, as can be seen in (7), which leads in turn to a larger employment share in the modern segment. Welfare implications of polices that change composition Now, consider the effects of policies that increase the modern retail share. One prominent example would be greater enforcement of tax collection in the traditional segment. One can show that such a policy increases average consumption but decreases average leisure by increasing shopping 14
time. The reason is that greater tax enforcement in the traditional segment increases the modern retail share, which leads to more efficient retailing but a greater shopping input required of the households. Thus, while such a policy would increase the use of the modern technology and measured productivity in retail trade, it would have an ambiguous effect on the welfare of the average individual.
4 Quantitative Version of Model This section builds a richer version of the model that is better suited to assessing the quantitative implications of the various channels in question and to addressing the link between measured productivity and welfare. This quantitative version has three main differences from the simple version of the previous section. First, it makes the car purchase choice endogenous. Second, it introduces capital into the model, and allows for differences in capital intensity by type of retail technology, which may be important in reality. Third, it introduces a value-added tax, another important tax that traditional producers often evade.
4.1 Households Building on the previous version, consumers now have preferences for automobiles, a, in addition to consumption and leisure: U i = log(ci ) + ψ i log(1 − n − si ) + α · ai .
(11)
Here, α represents the utility of having a car, and a ∈ {0, 1} captures car ownership, with a = 1 if the household owns a car. One can think of α as capturing the direct flow of utility from “Sunday drives” or commuting services, distinct from the car’s value as a shopping device. As in the simple version, the final consumption good is produced by the household good as in (2). Each household is endowed with “ability” η i , which determines the value of its time in producing market output. The η i terms are independent of the ψ i terms and are drawn from a distribution H(·) with non-negative support and mean normalized to one. Households are also endowed with one unit of capital that they can rent to the capital market at rate r. The budget constraint of household i is pa ai + pt yti + pm yti ≤ w · n · η i + r + T, (12) where pa is the price of an automobile and T is a transfer from the government. 15
4.2 Production and Taxation The intermediate sector produces intermediate goods according to: X = E · Lxθx · Kx1−θx ,
(13)
where Lx and Kx are units of labor and capital employed in production, and θx is the labor share in production. Intermediates may either be turned into retail goods by being combined with retail services (described below) or be turned into cars. A car may be produced from A units of intermediates. Again, the price of the intermediate good is normalized to be one. The retail sector produces output according to the production function: θ
1−θ j
Y j = min[E · Z j · L j j · K j
, X j ],
(14)
where L j , K j and X j are inputs of labor, capital, and intermediates of segment j, and θ j is the labor share in production in technology j. The effective tax rates that the two segments face are φm · τ L and φm · τ VA for the modern segment, and φt · τ L and φt · τ VA for the traditional segment. All tax revenues are rebated, lump sum, to the households.
4.3 Solution to Household Problem Household optimization consists of two discrete choices: where to shop and whether to buy a car. Intuitively, since cars are superior goods in the utility function, buying one is optimal only when income is sufficiently high. For a household with efficiency units η i and leisure value ψ i , purchasing a car is optimal if and only if pa 1 η ≥ η˜ (ψ ) ≡ · −r−T , nw 1 − χ (ψ i ) exp(−α ) i
where
i
ψ i 1−n−νm γνm 1−n− ψ i χ (ψ i ) ≡ pm · 1−n−νt pt 1−n−γνm ψ i 1−n−νt 1−n−γνt
(15)
if ψ i ≤ ψ˜ (0) if ψ˜ (0) < ψ i ≤ ψ˜ (1) if ψ i > ψ˜ (1).
Figure 5 illustrates the household’s optimal behavior. Each household is represented as an element of the plane, with the x-axis representing ψ i and the y-axis representing η i . Households with η i
16
Modern
Traditional
Modern
Efficiency Units, i
η
η~(ψ )
Modern
ψ~ (0)
Traditional Traditional
Taste for Leisure, ψ i
ψ~ (1)
Figure 5: Solution to the Household Problem, Quantitative Version above the threshold η˜ (ψ i ) choose to buy a car, while those below do not. Of those buying a car, the darker-shaded region to the left of ψ˜ (1) are those who shop at the modern store. Intuitively, these are the households whose value of leisure time is sufficiently low that the lower prices of the modern segment outweigh its higher shopping costs (given that they have a car). Of those not buying a car, the lighter-shaded region to the left of ψ˜ (0) are those that shop in the modern segment.8 One important feature of the model is that a household’s efficiency units, ηi , do not affect its choice of where to shop (over and above the car purchase decision.) In other words, the cutoffs ψ˜ (a) do not depend on η i . Households with higher efficiency units have a higher opportunity cost of shopping time in terms of wage income, but this is exactly offset by the decreasing marginal utility of additional goods purchased with those wages. The balanced-growth preferences that drive this feature of the model were assumed in order to match the observation of roughly constant shopping time over the income distribution in the cross-section of United States households.9 8 The
non-linearity of the car purchase decision in ψ is reconciled as follows. For households with ψ values below ψ˜ (0) or above ψ˜ (1), η˜ (ψ ) is decreasing in ψ because as ψ rises, households value the time savings from having a car more. For the remaining households, as ψ rises, the time savings of the traditional store make shopping at the modern store with a car a less attractive option. Hence, the value of a car decreases with ψ . 9 According to my calculations from the 2003-2010 pooled American Time Use Survey (ATUS), the correlation of income and shopping time is essentially zero, as in the model; see Appendix A.4.
17
4.4 Equilibrium Properties One can show that, in equilibrium10, the ratio of value added per worker in the modern to traditional segment is VAm /Lm θt 1 + φm τ L 1 − φt τ VA · . = · VAt /Lt θm 1 + φt τ L 1 − φm τ VA
(16)
Thus, value added per worker can be higher in the modern technology either because it has a lower labor share than the traditional technology, or because tax evasion is easier when operating the traditional technology. Now, this tax-evasion channel works through both higher labor costs and a higher value of output lost to taxes. As before, Zm and Zt are not relevant for differences in value added per worker. In equilibrium, the price of consumption goods in retail segment j ∈ {m,t} is given by pj = 1+
(1 + φ j τ L )θ j · Θ j (w/r) , (1 − φ j τ VA ) · Z j
(17)
θ
where Θ j (w/r) ≡ (w/r)θ j −θx θxθx (1 − θx )1−θx /θ j j (1 − θ j )1−θ j . Thus, producers in segment j each charge the price they pay for the intermediate good plus a margin that is increasing in their effective tax rates on labor and value added, decreasing in their efficiency, and increasing in the function Θ j (w/r), which depends on the relative cost and intensities of labor and capital.
5 Quantitative Analysis In this section, I calibrate the model and conduct several counterfactual experiments. The main experiments begin with the model calibrated to the United States, and then lower (i) the general efficiency of production, E, (ii) the tax rates on labor and value added, τ L and τ VA , and (iii) the tax compliance rates, φm and φt . I then compute the model’s predictions for measured retail productivity and the fraction of employment at modern stores, and compare them to their empirical counterparts. Finally, I use the model to assess the effects of several (potentially welfare-enhancing) policies that developing countries could implement to improve the efficiency of their retail sectors. 10 The
definition of an equilibrium, and equilibrium conditions, are presented in detail in Appendix B.2.
18
5.1 Summary of Calibration I begin by summarizing how each parameter is calibrated; see Appendix B.3 for a more detailed explanation of the calibration and moments targeted. For tractability, I assume that the leisurevalue distribution, G(·), and the efficiency-units distribution, H(·), are distributed log normally. This assumption leaves twenty parameters of the model to calibrate (or normalize). I set values for twelve parameters directly. For the labor shares in production, I set θx to be 0.66, following Gollin (2002), and θm and θt to be 0.40 and 0.62, which are the average labor shares by segment that I estimate from Argentina, Mexico, El Salvador and the Philippines. For γ , I choose a value of 0.48 based on evidence on the time savings of cars. I set n equal to 0.24 to match average time working in the United States. I set the tax rates on labor and value added to be τ L = 0.13 and τ VA = 0.09, which are average labor and sales tax rates in the United States, according to the World Bank. I set the U.S. tax compliance rates, φm and φt , to be one. For the distribution of efficiency units across households, I pick ση equal to 0.97 to match the log variance of U.S. household income, and normalize µη = −0.49 so that E[η i ] equals one, as assumed earlier. Finally, I set the efficiency of the U.S. economy to be E = 1 as a normalization. The remaining eight parameters are calibrated jointly to match eight relevant moments of the data. These parameters are the efficiencies of the two retail segments, Zm and Zt ; the shopping service requirements, νm and νt ; the taste parameter for cars, α ; the mean and variance of the time-value distribution, µψ and σψ ; and the resource cost of building a car, A. The moments matched are: (1) the ratio of prices in modern to traditional stores (0.78); (2) the size of the retail sector (17 percent of aggregate employment); (3) the average shopping time for households in the top half of the income distribution (4.4 hours per week); (4) the average shopping time for households in the bottom half of the income distribution (4.3 hours per week); (5) the share of households that own a car (91 percent); (6) the modern retail share (67 percent); (7) the percent of modern-segment customers that come via car (97 percent); and (8) the average cost of operating a car ($5,600 per year). The resulting parameter values are νm = 0.060, νt = 0.011, α = 0.046, µψ = 1.52, σψ = 0.56 and A = 0.029.
5.2 General Efficiency of Production in the Developing Countries Starting from the model calibrated to the United States, I pick a lower value of E for each developing country under study. I do so to match the country’s GDP per capita at PPP. According to the PWT, compared to the United States, Argentina’s GDP per capita is 30 percent as high, Mexico’s is 34 percent as high, Brazil’s is 20 percent as high, Thailand’s is 20 percent as high, El Salvador’s 19
Table 1: Retail Value Added per Worker, Relative to the United States Argentina Mexico Brazil Thailand El Salvador Philippines
Model
Data
25 34 20 14 12 6
29 27 20 23 20 7
Note: Value added per worker is expressed relative to that of the United States retail sector, which is normalized to 100.
is 17 percent as high and the Philippines’ is 8 percent as high. I pick a value of E for each country that matches these relative income levels.
5.3 Tax Compliance in the Developing Countries I lower the tax rates to match labor and value-added tax rates in each country (see Appendix A.2). I also lower the tax compliance rates, φm and φt , to match my own estimates of the tax compliance rates by retail segment in each developing country. I estimate the tax compliance rate by segment of the retail industry as the fraction of workers whose employer makes contributions to public insurance programs on their behalf. This definition follows that of Azuara and Marinescu (2013), who study the effects on informality of an expansion of payroll-tax-financed public health insurance in Mexico. Table A2 presents my estimated tax compliance rates; Appendix A.3 provides the details of my measurements. For Thailand and the Philippines, for which I could not compute independent estimates of formality rates, I choose values of φm = 0.84 and φt = 0.34, which are the averages of the other countries.
5.4 Quantitative Results Table 1 presents the model’s predictions for value added per worker in the retail sector, as well as their empirical counterparts (presented earlier, in Figure 1.) Overall, the model’s predictions are in line with the data, with far lower value added per worker in retailing in each of the developing countries than in the United States. Argentina and Mexico have the highest value added per worker in retail in both the model and data, at 25 and 34 percent of the U.S. level in the model and 29 and 27 percent in the data. The Philippines has the lowest values, with 6 percent of the U.S. level in the 20
Table 2: Value Added per Worker by Retail Segment, Relative to the United States Modern Argentina Mexico Brazil Thailand El Salvador Philippines
41 44 31 21 21 9
Model Traditional 21 24 17 13 11 5
Modern
Data Traditional
50 51 36 42 48 18
22 20 13 18 16 6
Note: Value added per worker is expressed relative to that of the United States retail sector, which is normalized to 100.
model and 7 percent in the data. The remaining countries have intermediate income levels in the model and the data, with the model spot on at 20 percent in Brazil and below the data in Thailand and El Salvador, at 14 and 12 percent in the model and 23 and 20 percent in the data. Table 2 presents the model’s predictions for value added per worker by retail segment. On two dimensions the model is generally consistent with the data. First, the model predicts that value added per worker is much higher in the modern segment than in the traditional segment, just as in the data. Argentina and Mexico, for example, have values of 41 and 44 percent of the U.S. level in the modern segment, compared to 50 and 51 percent in the data; and have values of 21 and 24 percent in the traditional segment, compared to 22 and 20 percent in the data. Second, the model predicts that the poorer the country, the lower is value added per worker in both retail segments, as in the data.11 Finally, Table 3 presents the model’s predictions for the employment share of the modern retail segment. The first row shows the United States, which is matched exactly at 67 percent, as per the calibration procedure. For the developing countries, the model’s predicted shares are substantially lower in each case and generally in line with the data. The three highest predicted modern employment shares are for Argentina, Mexico and Brazil, at 17, 21 and 22 percent. In the data, these values are 25, 23 and 21 percent. The three lowest are for Thailand, El Salvador and the Philippines, at 16, 10 and 13 percent in the model, compared to 19, 15 and 15 percent in the data. In conclusion, in spite of its simplicity, the model appears to do well in matching the cross-country patterns in value added per worker in the retail sector and by segment, and the share of employment 11 One
limitation of model is that measured productivity differences between the modern and traditional segments are more pronounced in the data than in the model. In the data, the average ratio of value added per worker between the modern and traditional segments is 2.5. In the model, the average ratio is 1.8. Appendix A.5 provides evidence that some of the difference is accounted for by higher human capital per worker at modern retail stores.
21
Table 3: Modern Retail Share, Model and Data United States Argentina Mexico Brazil Thailand El Salvador Philippines
Model
Data
67 17 21 22 16 10 13
67 25 23 21 19 15 15
Note: The modern retail share is defined as the percent of retail employment in the modern segment.
in the modern segment. I next turn to other predictions of the model not targeted directly and ask how the model performs there.
5.5 Assessing the Model’s Other Predictions In this section, I assess the model’s predictions for three other statistics of interest for the developing countries. These are average shopping time, relative prices at modern and traditional stores, and car-ownership rates. It is worth assessing these three predictions since the model’s main tradeoff is one of lower prices versus higher shopping time at modern stores, with car ownership playing a key role in steering households towards modern stores. For average shopping time, the model predicts that households in developing countries spend less time shopping than U.S. households. This is consistent with the finding of Allesandria and Kaboski (2011) that average shopping time is positively correlated with income in the cross-section of countries. The authors document this positive correlation in two different internationally comparable time-use surveys: the Multi-National Time Use Survey (MNTUS) and the European Harmonized Time Use Survey (EHTUS). According to their estimates, a 66-percent drop in income (corresponding to the drop in the model for Mexico) is associated with a drop in average shopping time of 13 percent, with a 90-percent confidence interval of between a 1 percent and a 25-percent drop.12 The model predicts a 4.6 percent lower average shopping time in Mexico than in the United States, and similar values for the other developing countries, which is well within the range of the data. In the model, prices in the modern segment in the developing countries range between 78 percent 12 From the EHTUS, they estimate a coefficient of log shopping time on log GDP per capita of 0.198 with a t-statistic
of 2.83, and from the MTUS, their estimate is 0.208 with a t-statistic of 3.10.
22
and 82 percent of those at traditional stores, depending on the country. Available evidence supports the model’s prediction that modern stores are cheaper than traditional stores. For a set of comparable goods in Thailand, for example, the McKinsey Global Institute reports that modern stores charge, on average, 82 percent as much as traditional stores. For other countries, McKinsey reports a range of price differences. For example, it reports that, in traditional stores, prices are between five- and 15-percent higher in Mexico and up to 30-percent higher in Brazil. In other developing countries not studied in the current paper their findings are similar, with prices in traditional stores between 10 percent to 30 percent higher in Poland and up to 30 percent higher in Turkey (see Lewis (2005) and the references therein.) For car-ownership rates, the model modestly over-predicts the elasticity of car ownership to income. In Mexico, for example, data from the 2000 Census show that 32 percent of households own cars, whereas the model predicts a car ownership rate of 13.5 percent. It is perhaps not surprising that the model cannot match this elasticity very precisely, given how crude the car-ownership decision is in the model. To address this limitation, I conduct an alternative experiment in which I match car ownership rates directly in each country, and find that the model predicts modestly higher modern retail shares than in the baseline experiments. In Mexico, for example, the model predicts a 29-percent modern share when matching car ownership directly, compared to 21-percent in the baseline experiment.
5.6 Decomposing the Results: Cars vs. Tax Evasion In the model, the lower use of modern retail technologies reflects two basic forces. First, carownership rates are lower in the model’s developing countries since aggregate efficiency (and income) is lower. Second, tax-evasion rates are higher in the traditional retail segment than in the modern segment. In this section I present an alternative set of results that considers each effect separately. I find that car ownership explains roughly two thirds of the difference in modern retail shares between the United States and developing countries, and that tax evasion explains roughly one third. These results provide support for both the appropriate technology theory and the misallocation theory, with a larger quantitative role for the former. The “Cars Only” experiment lowers E in each developing country to match its GDP per capita level, but leaves tax rates and tax compliance rates by segment the same as in the United States (i.e. the baseline model). In addition, it lowers A to match the car-ownership rates in Mexico (and, arguably, match car-ownership rates better in the other developing countries.) The second column of Table 4 reports the model’s predictions for the modern retail employment share under this experiment. The model’s predictions are higher than in the baseline model, with modern retail 23
Table 4: Modern Retail Share: Alternative Experiments Experiment Cars Only
Tax Evasion Only
Data
Argentina
39
47
25
Mexico
43
52
23
Brazil
29
64
21
Thailand
30
52
19
El Salvador
26
39
15
Philippines
22
51
15
Note: The modern employment share is defined as the percent of total retail employment in the modern segment. The “Cars Only” experiment changes E in each country but leaves tax rates and tax compliance rates by segment the same, and lowers A in the developing countries. The “Tax Evasion Only” experiment changes tax rates and tax compliance rates by segment but leaves E and A the same.
shares of 39 percent, 43 percent and 29 percent in Argentina, Mexico and Brazil and 30 percent, 26 percent and 22 percent in Thailand, El Salvador and the Philippines. Taking an average across all countries, this amounts to two thirds of the differences with the United States. The “Tax Evasion Only” experiment changes only the tax rates and tax compliance rates by segment but leaves E and A the same. The third column of Table 4 reports the model’s predicted modern retail shares. The model predicts values of 47 percent, 52 percent and 64 percent in Argentina, Mexico and Brazil, and 52 percent, 39 percent and 51 percent in Thailand, El Salvador and the Philippines. Averaging across countries, this amounts to one third of the difference between the developing countries and the United States.13
5.7 Policy Experiments In this section, I use the model to conduct two counterfactual policy experiments, both for Mexico. The first simulates reducing the informality rate in the traditional retail segment by a half. Reducing informality is a key theme echoed by Lewis (2005) and others. The idea is that if entrepreneurs in 13 A
key part of the model’s large quantitive effect of cars on the modern retail share comes from the model’s prediction that households with cars are more likely to shop at modern retailers than households without cars. The model’s conditional probabilities of shopping at a modern store are 87 percent for car owners and 33 percent for noncar owners. Unfortunately, few direct estimates of these probabilities exist. The one study I was able to find (Bromley and Thomas, 1993) uses survey evidence from the town of Swansea in the U.K. to document that “between 71 percent and 86 percent of car owners patronized a superstore for their food shopping; for the carless the percentages vary between 31 percent and 57 percent.” These estimates are in line with the model’s values.
24
Table 5: Counterfactual Policy Experiments Experiment Statistic (Increase)
Reduce Informality Liberalize Car Imports
Modern Retail Employment Share
0.07
0.06
Retail Sector Value Added Per Worker
9.6%
3.5%
Consumption
1.4%
0.5%
Leisure
-0.4%
0.0%
Shopping time
13.0%
-0.4%
Welfare
0.2%
0.6%
Note: Columns 2 and 3 report the increases in the statistics of Column 1 under two different counterfactual policy experiments. Both are computed for Mexico. The first experiment reduces the informality rate in the traditional retail segment to close half the gap in informality with the modern segment. This involves raising φt from 0.37 to 0.58, and is meant to simulate a crack-down on tax evasion among small retail establishments. The second experiment reduces the price of a car by 25% to simulate a liberalization in the market for car imports. The increases in consumption, leisure and shopping time are computed as an average across all model households. Welfare is computed as the consumption increase that makes the average household indifferent between implementing the policies or not.
the traditional segment had to pay more of their required taxes, their prices would be less favorable, and modern stores would become more prevalent. The second policy experiment simulates a liberalization of the car-import market. In the developing world, tariffs or quotas on new-car imports are common and well documented. Perhaps less well known is that a large number of developing countries place strong restrictions on imports of used cars. Policies of this sort typically come in the form of outright bans on used-car imports, prohibitive tariffs, and limitations on the age of used vehicles that can be imported. Pelletiere and Reinert (2002) document the extent of used-vehicle import restrictions in a large number of developed and developing countries and find that these restrictions are widespread and often severe. They report that in 19 developing countries, there are complete prohibitions on importing a used car. In another 27 countries, there are other “substantial restrictions” of various kinds. Table 5 presents the results of the experiments. In the “Reduce Informality” experiment, I change φt from 0.37 to 0.58, which closes half the gap in informality between the traditional and modern segments. The model predicts that the modern retail share rises by 0.07 (from 0.21 to 0.28) and that retail-sector value added per worker rises by 9.6 percent. On the one hand, because of the more-efficient retailing, consumption rises by 1.4 percent. On the other hand, leisure falls by 0.4 percent as households increase shopping time by 13 percent. The overall welfare gain is just 0.2 percent in consumption equivalents.
25
In the “Liberalize Car Imports” experiment, I lower A by 25 percent to simulate a plausible drop in car prices when used-car import bans are eliminated. The model predicts a rise in the modern retail share of 0.06 and an increase in retail-sector value added per worker of 3.5 percent. Consumption rises by 0.5 percent, and leisure rises slightly due to the more-efficient shopping provided by cars. The welfare gain overall is 0.6 percent in consumption equivalents.14 The overall lesson from the experiments is that measured productivity increases from policy changes are a poor guide to welfare. One can see this here, as the first experiment leads to larger measured productivity gains but smaller welfare gains than the second experiment. Similarly, policies that lead to large measured productivity increases, or large consumption increases, may not lead to large increases in welfare. This is especially evident in the first experiment, in which households have sizable consumption increases from more-efficient modern retailing, but spend more of their time shopping.
6 Supporting Evidence This section presents two additional pieces of evidence that support the importance of car ownership in explaining the use of modern retail technologies. The first looks at the cross-section of counties in the United States and Mexico. The second looks at a natural experiment from Cyprus.
6.1 Cross-Section of Counties in Mexico and the United States The cross-section of locales within a given country offers a good opportunity to check the model’s predictions against reality. In this section, I provide some new geographic evidence on income, auto ownership and modern retail segment prevalence using evidence from counties in Mexico and the United States. I focus on these two countries since economic and household census data are collected for these countries at a low (and comparable) level of geographic disaggregation.15 14 Another
important prediction is the model’s elasticity of car ownership to market price. The model predicts an elasticity of around −1.4. McCarthy (1996) surveys estimates of the market price elasticity of demand for cars and finds a range of −0.6 to −1.2, suggesting that the model’s elasticity is modestly higher than the available evidence, though in the same ballpark. 15 For the United States, I construct county-level estimates of median household income and household automobileownership rates from the 2005 American Community Survey (ACS). I combine these with data on of the fraction of retail employment at stores with greater than 20 employees from the 2004 County Business Patterns. For Mexico, I compute county- (municipio) level median household income and automobile-ownership rates from the 2000 Mexican Census of Population, and the share of retail employment at modern establishments using the 2004 Censo Economico available from the Instituto Nacional de Estadist´ıca, Geograf´ıa e Inform´atica (INEGI). Income is expressed in both countries in 2000 PPP dollars. I use all counties for which data are available for all three variables.
26
1 .8 Modern Retail Share .4 .6 .2 0
7
8
9 10 Log Household Income Mexico
11
12
United States
0
.2
Auto Ownership Rate .4 .6 .8
1
(a) Modern Retail Share and Median Household Income
7
8
9 10 Log Household Income Mexico
11
12
United States
(b) Auto Ownership Rate and Median Household Income
Figure 6: Cross-Section of Counties in U.S. and Mexico
27
Figure 6 displays, for the cross-section of counties in Mexico and the United States, the modern retail share and the auto-ownership rate plotted against the log of median household income. Panel (a) displays the modern retail share, while Panel (b) displays the auto-ownership rate. In both panels, the gray crosses represent Mexican counties, while the black dots represent U.S. counties. One can see a clear positive correlation in both Panel (a) and Panel (b). The correlation between the modern retail share and log income is 0.91 overall, 0.46 for just U.S. counties and 0.63 for just Mexican counties. The correlation between auto ownership and log income is 0.88 overall, 0.21 in U.S. counties, and 0.71 in Mexican counties. All correlation coefficients are statistically significant at the one-percent level. Not surprisingly, there is also a strong correlation between auto ownership and the modern retail share. The overall correlation is 0.90; for just U.S. counties it is 0.44; for just Mexican counties, it is 0.59; all correlation coefficients are significant at the one-percent level. These data support the theory in several ways. First, they show that within each country, there is a positive correlation between modern store prevalence and auto-ownership rates, just as the theory predicts. Second, they provide evidence that the correlation works through a single variable – median household income – in all geographic regions taken as a group. Thus, even though Mexico and the United States differ in numerous ways, the income level of a region – regardless of which country it is located in – is a strong predictor of its use of modern retail technologies.
6.2 Natural Experiment in Cyprus An additional way to test the theory’s predictions is to look across time within a given country. One country in particular – the island nation of Cyprus – offers an informative natural experiment in which the stock of automobiles rose suddenly in response to a policy change. Clerides (2008) documents that Cyprus largely repealed its limitations on the imports of used cars in 1993. Because this policy change occurred for the most part independently of other changes in the regulatory and economic environment, Clerides argues that the Cypriot experience provides a case study in which to learn about the effects of repealing used-car import bans. He finds that after the restrictions were repealed, there was a substantial expansion in the overall car market, led by large increases in used-car imports (almost all from Japan – steering wheels in both countries are on the left side of the car). In 1992, while the bans were still in place, just 7 percent of all first-time car registrations in Cyprus were imported used cars. In 1998, after the ban was repealed, this figure skyrocketed to 72 percent of all first-time registrations. Perhaps not surprisingly, prices of the used-car imports were substantially lower than those of new cars sold of the same make and model. My theory predicts that this expansion in the Cypriot car market should have been accompanied by an expansion in modern retailing. To help assess this prediction, I use data on the share of 28
&!"
*!'!!!" -./01"/2"3456"7894":;5<"8=>4?"
)!'!!!"
%#"
[email protected]":9>CD."8=>4?"
#!'!!!" %!"
E5905F."/2"G5.8>;"H/91594"8." I/659F"-./954":9>CD."8=>4?"
(!'!!!"
$#" &!'!!!" $!" %!'!!!" #"
$!'!!!"
!"
!" %!!!"
$+++"
$++,"
$++*"
$++)"
$++#"
$++("
$++&"
$++%"
$++$"
$++!"
$+,+"
Figure 7: Used Car Stock and Modern Retailing in Cyprus
employment at modern stores and the number of large chain supermarkets in Cyprus around the time of the car-market liberalization. Figure 7 shows the estimated stock of used cars in Cyprus over this period (the bars), the fraction of retail workers at stores with 20 or more workers (the dots), and the number of supermarkets of the five largest chains in Cyprus (the solid line).16 The dramatic increase in the stock of used cars is clearly visible after 1993, when the liberalization took place. The modern retail share, which unfortunately, is not available in all years, rises from around 12 percent in the early 1990s to around 25 percent by 2001. Corroborating these official data, the largest Cypriot supermarket chains expanded over this period, roughly tripling the number of stores. Thus, there was a noticeable increase in the presence of modern retail establishments in the years following the liberalization of used-car imports, as the theory would predict.17 16 I estimate the stock of used cars using yearly data on the sales of used cars from Clerides (2008) plus the assumptions that (i) the car stock depreciates at a rate of 10 percent per year, and (ii) the average yearly sales of used cars before 1989 equaled the average between 1989 and 1993. The official Cypriot retail statistics are publicly available from the Cyprus Department of Statistics and Research of the Ministry of Finance. I obtained the chain supermarket data from the stores themselves. 17 While there are likely to have been other changes in Cyprus around this period that could account for the rise in modern retailing, the most obvious candidate, a sudden rise in income, does not seem to be supported in the data. According to the Penn World Tables, real GDP growth was similar in the five-year period after 1993 to those of the five years before, with an average growth rate of 5.4 percent per year before and 4.6 percent after.
29
7 Conclusion Most theories of cross-country productivity differences emphasize frictions that lead to an inefficient use of production technologies. This paper argues that, in the retail trade sector, the predominance of traditional technologies in developing countries largely reflects optimal behavior. Because income is low in the developing world, households purchase fewer cars and other household durables that reduce shopping time. As a result, entrepreneurs rationally deploy relatively little of the modern technology, which is shopping-time intensive. I support the theory using new disaggregated evidence for a set of developing countries, and with a quantitative analysis of a simple model of home production and retail technology adoption. One implication of the paper is that policies that lead to inefficient production in one sector of the economy might help account for low measured productivity in other sectors. This implication could apply to other services for which frontier technologies require a large scale of operations, and for which low household transportation costs are important. In terms of policy implications, one way that governments can reduce transportation costs is to remove distortions that low car ownership rates. Of course, investments in transportation infrastructure are also necessary in order for cars to reduce transportation costs. Future work should further consider the link between public investments in transportation infrastructure and technology adoption in the service sector.
References ACEMOGLU , D.,
AND
F. Z ILIBOTTI (2001): “Productivity Differences,” Quarterly Journal of
Economics, 116(2), 563–606. AGUIAR , M.,
AND
E. H URST (2007): “Lifecycle Prices and Production,” American Economic
Review, 97(5), 1533–59. A LLESANDRIA , G., AND J. K ABOSKI (2011): “Pricing to Market and the Failure of Absolute PPP,” American Economic Journal: Macroeconomics, 3, 91–127. A MERICAN AUTOMOBILE A SSOCIATION (2008): http://exchange.aaa.com/automobiles-travel/automobiles/.
“Your
Driving
Costs,”
A ZUARA , O., AND I. M ARINESCU (2013): “Informality and the Expansion of Social Protection Programs: The Case of Mexico,” Journal of Health Economics, 32, 938–950. BAILY, M. N., AND R. M. S OLOW (2001): “International Productivity Comparisons Built from the Firm Level,” Journal of Economic Perspectives, 15(3), 151–172. 30
BASKER , E. (2005): “Selling a Cheaper Mousetrap: Wal-Mart’s Effect on Retail Prices,” Journal of Urban Economics, 58(2), 203–299. BASU , S., AND D. N. W EIL (1998): “Appropriate Technology and Growth,” Quarterly Journal of Economics, 113(4), 1025–1054. B ECKER , G. S. (1965): “A Theory of the Allocation of Time,” Economic Journal, 75(299), 493– 517. B ENHABIB , J ESS , R. R., AND R. W RIGHT (1991): “Homework in Macroeconomics: Household Production and Aggregate Fluctuations,” Journal of Political Economy, 99(6), 1166–1187. B IBA , G., F. D. ROSIERS , M. T HERIAULT,
AND
P. V ILLENEUVE (2006): “Big Boxes versus
Traditional Shopping Centers: Looking at Households’ Shopping Trip Patterns,” Journal of Real Estate Literature, 14(2), 175–202. B LOOM , N., C. G ENAKOS , R. S ADUN ,
AND
J. VAN R EENEN (2012): “Management Practices
Across Firms and Countries,” Academy of Management Perspectives, 26(1), 12–33. B OLLARD , A. (2010): “Essays on Industrial Organization in Developing Countries,” Ph.D. thesis, Stanford University. B OSWORTH , B. P.,
AND
J. E. T RIPLETT (2004): Productivity in the U.S. Services Sector: New
Sources of Economic Growth. Brookings Institution Press. B ROMLEY, R. D. F. (1998): “Market-place Trading and the Transformation of Retail Space in the Expanding Latin American City,” in Retailing: Critical Concepts, ed. by A. M. Findlay, and L. Sparks, vol. IV. Routledge. B ROMLEY, R. D. F.,
AND
C. J. T HOMAS (1993): “The Retail Revolution, the Car-less Shopper,
and Disadvantage,” in Retailing: Critical Concepts, ed. by A. M. Findlay, and L. Sparks, vol. IV. Routledge. BUERA , F. J.,
AND
J. P. K ABOSKI (2012): “The Rise of the Service Economy,” American Eco-
nomic Review, 102, 2540–69. BURSTEIN , A., M. E ICHENBAUM , AND S. R EBELO (2005): “Large Devaluations and the Real Exchange Rate,” Journal of Political Economy, 113(4), 742–784. BURSTEIN , A., J. C. N EVES ,
AND
S. R EBELO (2003): “Distribution Cost and Real Exchange
Rate Dynamics During Exchange-Rate-Based-Stabilizations,” Journal of Monetary Economics, 50, 1189–1214. 31
C ASELLI , F. (2005): “Accounting for Cross-Country Income Differences,” in Handbook of Economic Growth, ed. by P. Aghion, and S. Durlauf., 679-741. Elsevier. C ASELLI , F.,
AND
W. J. C OLEMAN (2006): “The World Technology Frontier,” American Eco-
nomic Review, 96(3), 499–522. C LERIDES , S. K. (2008): “Gains from Trade in Used Goods: Evidence from the Global Market for Automobiles,” Journal of International Economics, 76(2), 322–336. D E S OTO , H. (1989): The Other Path. Harper Row, New York. F OSTER , L., J. H ALTIWANGER ,
AND
C. J. K RIZAN (2006): “Market Selection, Reallocation and
Restructuring in the U.S. Retail Trade Sector in the 1990s,” Review of Economics and Statistics, 88, 748–758. F UCHS , V. R. (1969): The Service Economy. Columbia University Press/NBER, New York. G ALIANI , S., AND F. W EINSCHELBAUM (2012): “Modeling Informality Formally: Households and Firms,” Economic Inquiry, 50(3), 821–838. G OLDMAN , A. (2000): “Supermarkets in China: the Case of Shanghai,” in Retailing: Critical Concepts, ed. by A. M. Findlay, and L. Sparks, vol. IV. Routledge. G OLLIN , D. (2002): “Getting Income Shares Right,” Journal of Political Economy, 110(2), 458– 474. G REENWOOD , J., A. S ESHADRI , AND M. YORUKOGLU (2005): “Engines of Liberation,” Review of Economic Studies, 79(1), 109–133. G UNER , N., G. V ENTURA ,
AND
D. X U (2008): “Macroeconomic Implications of Size-Dependent
Policies,” Review of Economic Dynamics, 11(4), 721–744. H ALL , R. E.,
AND
C. I. J ONES (1999): “Why Do Some Countries Produce So Much More Output
per Worker than Others?,” Quarterly Journal of Economics, 114(1), 83–116. H AUSMAN , J., AND E. L EIBTAG (2007): “Consumer Benefits from Increased Competition in Shopping Outlets: Measuring the Effect of Wal-Mart,” Journal of Applied Econometrics, 22(7), 1157–117. H ERRENDORF, B., AND A. T EIXEIRA (2011): “Barriers to Entry and Development,” International Economic Review, 52(2), 573–602.
32
H ESTON , A., R. S UMMERS , AND B. ATEN (2011): “Penn World Table Version 7.0,” Center for International Comparisons of Production, Income and Prices at the University of Pennsylvania. H OLMES , T. J. (2011): “The Diffusion of Wal-Mart and Economies of Density,” Econometrica, 79(1), 253–302. H SIEH , C.-T.,
AND
P. J. K LENOW (2009): “Misallocation and Manufacturing TFP in China and
India,” Quarterly Journal of Economics, 124, 1403–1448. (2010): “Development Accounting,” American Economic Journal: Macroeconomics, 2(1), 207–223. J IA , P. (2008): “What Happens When Wal-Mart Comes to Town: An Empirical Analysis of the Discount Retailing Industry,” Econometrica, 76(6), 1263–1316. J ONES , C. I. (2011): “Intermediate Goods and Weak Links in the Theory of Economic Development,” American Economic Journal: Macroeconomics, 3(2), 1–28. K LENOW, P. J., AND A. RODR´I GUEZ -C LARE (1997): “The Neoclassical Revival in Growth Economics: Has it Gone Too Far?,” in NBER Macroeconomics Annual 1997, ed. by B. S. Bernanke, and J. Rotemberg. MIT Press, Cambridge.
L EWIS , W. W. (2005): The Power of Productivity: Wealth, Poverty and the Threat to Global Stability. University of Chicago Press. M C C ARTHY, P. S. (1996): “Market Price and Income Elasticities of New Vehicle Demands,” Review of Economics and Statistics, 78(3), 543–547. M C K ENZIE , D.,
AND
E. S CHARGRODSKY (2005): “Buying Less, But Shopping More: The Use
of Nonmarket Labor During a Crisis,” Economia, 22(2), 1–43. M INNESOTA P OPULATION C ENTER (2014): “Integrated Public Use Microdata Series, International: Version 6.3 [Machine-readable database],” Minneapolis: University of Minnesota. M URPHY, K. M., A. S HLEIFER , AND R. W. V ISHNY (1989): “Industrialization and the Big Push,” Journal of Political Economy, 97(5), 1003–1026. O I , W. Y. (1992): “Productivity in the Distributive Trades: The Shopper and the Economies ofMassed Reserves,” in Output Measurement in the Service Sectors, ed. by Z. Griliches, chap. 4. University of Chicago Press/NBER, Chicago. ˜ , J. C. L. (2014): “Tax collection, The Informal Sector, and Productivity,” Review of O RDO NEZ Economic Dynamics, 17(2), 262–286. 33
PARENTE , S. L., AND E. C. P RESCOTT (1994): “Barriers to Technology Adoption and Development,” Journal of Political Economy, 102(2), 298–321. (1999): “Monopoly Rights: A Barrier to Riches,” American Economic Review, 89(5), 1216–1233. PARENTE , S. L., R. ROGERSON ,
AND
R. W RIGHT (2000): “Homework in Development Eco-
nomics: Household Production and the Wealth of Nations,” Journal of Political Economy, 108(4), 680–687. P ELLETIERE , D.,
AND
K. A. R EINERT (2002): “The Political Economy of Used Automobile
Protection in Latin America,” World Economy, 25, 1019–1037. P RESCOTT, E. C. (1998): “Needed: A Theory of Total Factor Productivity,” International Economic Review, 39(3), 525–551. R ESTUCCIA , D., AND R. ROGERSON (2008): “Policy Distortions and Aggregate Productivity with Heterogeneous Establishments,” Review of Economic Dynamics, 11(4), 707–720. ROBLES , M. (2009): “Aggregate Effects of Imperfect Tax Enforcement,” IFPRI Discussion Paper 00845. ROSENSTEIN -RODAN , P. (1943): “Problems of Industrialisation of Eastern and South-Eastern Europe,” The Economic Journal, 53(210), 202–211. S CHMITZ , J. A. J. (2005): “What Determines Productivity? Lessons from the Dramatic Recovery of the U.S. and Canadian Iron Ore Industries Following Their Early 1980s Crisis,” Journal of Political Economy, 113(3), 582–625. T HE U.S. D EPARTMENT OF T RANSPORTATION (2004): “Summary of Travel Trends: 2001 National Household Travel Survey,” http://nhts.ornl.gov/2001/pub/STT.pdf. YOUNG , A. (1995): “The Tyranny of Numbers: Confronting the Statistical Realities of the East Asian Growth Experience,” Quarterly Journal of Economics, 110(3), 641–680.
34
Appendix A Data Appendix A.1
Censuses of Retail Trade
As described in Section 2, the census data that I employ come from every developing country (those with less than one-half of the United States’ income per capita) in which a census of retail trade has been conducted in the last ten years, and for which comparable data exists. The four compatibility criteria are as follows. First, the data must have allowed me to isolate retail trade from wholesale trade. Second, the census must not have excluded establishments below a certain size. Third, the data must have reported inputs and outputs by size category of store. Fourth, the labor input data must have included unpaid workers as well as paid, since unpaid work is prevalent in retailing. The censuses satisfying these criteria are: Brazil - Data come from the 2002 Pesquisa Anual de Com´ercio conducted by the Instituto Brasileiro de Geograf´ıa e Estat´ıstica. El Salvador - Data come from the 2005 Censo Econ´omico, Com´ercio, available from the Economic Ministry, Office of Statistics and the Census. Mexico - Data come from the 2005 Censo Econ´omico conducted by the Instituto Nacional de Estadist´ıca, Geograf´ıa e Inform´atica. Philippines - Data come from the 2005 Annual Survey of Philippine Business and Industry conducted by the Philippines National Statistics Office. Thailand - Data come from the 2002 Business Trade and Services Survey conducted by the National Statistical Office. In Thailand, I use a cutoff of 25 workers to define the modern segment, which is the closest available. Defining the modern segment as establishments with 10 or more workers or 50 or more workers produces qualitatively similar results. United States - Data come from the BEA Value Added by Industry Accounts. The underlying source of the data is the Economic Census, conducted every five years by the Bureau of the Census. For comparability with other countries, I use total full-time and part-time employees as my measure of labor input.
A.2
Tax Data
The tax rates on labor and value added come from the World Bank’s Doing Business database, which can be found here: http://www.doingbusiness.org/. The tax rates are calculated assuming that the employer is “representative” along a variety of dimensions: for example, the business is domestically owned, has sixty employees, and operates in the country’s largest city; see the webpage for more on their methodology. Taxes on labor are defined as the sum of statutory tax rates levied on payrolls and paid by the 35
Table A1: Country-Specific Variables Tax Rates Argentina Mexico Thailand Brazil El Salvador Philippines
GDP/L 0.30 0.34 0.20 0.21 0.17 0.08
τL 0.26 0.27 0.05 0.38 0.16 0.11
τ VA 0.21 0.16 0.07 0.22 0.13 0.12
Note: GDP/L is GDP per capita expressed at PPP from the PWT; τ L and τVA are the estimated taxes on labor and value added reported by the World Bank’s Doing Business database.
employer. These include contributions to social security programs, payroll taxes, or other taxes on hiring labor at the federal, state or local level. Taxes on value added are defined as any federal, state and local taxes on sales minus intermediate costs, or, in the United States, just on sales. For Brazil, I exclude the Imposto Sobre Produtos Industrializados, which covers only manufacturing, and not retail trade.
A.3
Estimated Tax Compliance Rates
I construct tax compliance measures by retail segment for Argentina, Brazil and Mexico, the countries for which the relevant data are available. For Argentina, I draw on data from the 2004 Encuesta Permanente de Hogares, which is a nationally representative survey of urban Argentine households. For Brazil, I use the 2002 Pesquisa Nacional por Amostra de Domicilio, which is a nationally representative survey of Brazilian households. For Mexico, I use the 2004 Encuesta Nacional de Empleo, which is a nationally representative survey of the Mexican labor force. For each country, I define the modern segment using the closest available cutoff to the 20+ worker cutoff used earlier. This amounts to cutoffs of 11 or more workers in Argentina and Brazil, and 16 or more workers in Mexico. Table A2 presents the estimated tax compliance rates by retail segment. I combine my estimates for Argentina, Brazil and Mexico with those of Galiani and Weinschelbaum (2012) for El Salvador, which are similar except that they are for the entire economy, not just retail trade, and use a size cutoff of 5 or more workers to define the modern segment (see Table 3, pg. 827). In the modern retail segment, tax compliance rates are high in all four economies. In Argentina, 84 percent of retail workers at modern establishments report that their employers contribute to social insurance programs on their behalf. In Brazil and Mexico, these percents are 82 percent and 78 percent. In El Salvador, this figure is 84 percent. Among traditional retailers, in contrast, tax compliance rates are much lower: 35 percent, 57 percent and 37 percent in Argentina, Brazil and Mexico, respectively, and just 7 percent in El Salvador.
36
Table A2: Estimated Formality Rates by Retail Segment Argentina (2004) Mexico (2004) Brazil (2002) El Salvador (2003)
Modern 0.84 0.78 0.82 0.84
Traditional 0.35 0.37 0.57 0.07
Note: Formality rates are defined as the fraction of all employees by segment whose employer contributes to social security programs on their behalf. Source: Author’s calculations for Argentina, Brazil and Mexico, and Galiani and Weinschelbaum (2012) for El Salvador.
A.4
Shopping-Time Data
An important component of the calibration procedure is the average shopping time by household income group. The calibration procedure targets average shopping time in the bottom half and top half of the income distribution, and in doing so it captures the fact that shopping time is largely uncorrelated with income. This section provides more detail about the statistics that are targeted, and demonstrates that the lack of correlation between income and shopping time in consistent with two broad alternative definitions of shopping time. Table A3: Shopping Time by U.S. Income Group 1st 2nd 3rd 4th Average income $10,508 $29,399 $52,939 $125,740 Shopping time (broad) 4.2 4.3 4.4 4.5 2.5 2.5 2.7 2.7 Shopping time (narrow) Note: Average income is computed using the 2000 U.S. Census. Shopping time is measured as average hours per week, and is computed from the pooled 2003-2010 American Time Use Surveys.
Table A3 presents average shopping time in hours per week by quartile of the U.S. income distribution for two measures of shopping time. Both are calculated using the 2003-2010 pooled American Time Use Surveys. Shopping time (broad) consists of all time spent on consumer purchases, travel related to consumer purchases, comparison shopping and researching purchases. This is what the calibration procedure targets. Shopping time (narrow) consists of all time spent on consumer purchases on groceries, food and gas and all travel related to consumer purchases. As the table shows, shopping time is very similar across income groups by both definitions.
A.5
Human Capital and Average Hours Worked by Retail Segment
One possible additional reason why value added per worker is higher in the modern retail segment is that, on average, workers in the modern segment have higher human capital or work more hours 37
than their counterparts in the traditional segment. This section draws on data from household surveys to assess the importance of this hypothesis. The data are available for Argentina, Brazil and Mexico, as before, and use the same sources described as in Section 5.3 above. As a frame of reference, I include the same calculations for the United States, using data from the 2000 census. In each country, I measure average years of schooling and average hours worked for workers reporting that their primary employment is in the retail sector. I further classify workers into the modern and traditional segments using the employer size cutoffs described in Section 5.3. To compute average human capital per worker, I assume that each year of school increases human capital by ten percent, which is consistent with estimates from a broad set of countries (see, e.g., Hsieh and Klenow (2010) and the references therein.) Table A4: Human Capital and Average Hours Worked by Retail Segment
Argentina (2004) Brazil (2002) Mexico (2004) United States (2000)
Years of Schooling Human Capital Hours Worked Modern Traditional Ratio Modern Traditional 12.6 10.4 1.20 45.4 48.2 12.6 10.8 1.19 51.5 52.9 11.4 9.3 1.24 43.3 42.9 12.4 12.3 1.01 34.6 35.6
Ratio 0.94 0.97 1.01 0.97
Note: Human capital is computed assuming a ten-percent return per year of schooling. Source: author’s calculations.
Table A4 shows the results. In all three developing countries, average schooling per worker is higher in the modern segment than in the traditional segment. Workers in the modern segment have, on average, 12.6 years of schooling in Argentina and Brazil and 11.4 years in Mexico. Workers in the traditional segment have 10.4, 10.8 and 9.3 years, respectively. This amounts to 1.2, 1.19 and 1.24 times as much human capital per worker in the modern segment as in the traditional segment. Thus, in these three countries, workers in modern stores have roughly 20 percent more human capital than their counterparts in traditional stores. Thus, skills can explain a modest portion of the higher value added per worker in the modern segment. My findings here are similar to those of Bollard (2010), who takes a thorough look at a similar issue in manufacturing plants in India. He documents that wages are much higher in large plants than in small plants, and finds that differences in observable worker skills can explain some, but far from all, of the difference in average wages. Differences in hours worked by store type are small on the other hand. In Argentina and Brazil, workers in modern stores tend to work slightly less on average than workers in traditional stores, with ratios of 0.94 and 0.97 average hours worked in the modern to traditional segment. In Mexico, average hours are 1.01 times as high in modern as traditional stores. The results in this table suggest that, at least for these three countries, differences in human capital per worker by retail segment account for a factor 1.2 of the ratio of value added per worker in the modern to traditional segment, while differences in hours worked are negligible. Thus, differences in the composition of workers by retail segment do help close some of the gap in value added per worker.
38
B Model Appendix B.1 Derivation of Results of Simple Model The equilibrium price of each retail segment, (9), follows directly from the first-order condition for segment j, which is that (p j − 1)EZ j = (1 + φm τ L )w, combined with the first-order condition for the intermediate producer, which is that E = w. The same is true of the expression for value added per worker by segment, (8). The comparative statics in µ , the modern retail share, are derived as follows. Note that output at the modern segment is given in equilibrium by G(ψ˜ 1 ) G(ψ˜ 0 ) Ym = wn α , + (1 − α ) pm pt with a similar expression for Yt . Now consider the effect of an increase in α . Prices do not change by (9) (and since w = E); nor do the cutoffs ψ˜ 0 and ψ˜ 1 by (7). It follows, then, that Ym increases. The reason is that pm < pt by assumption on parameters, and ψ˜ 0 < ψ˜ 1 by (7). One can show that Yt must fall as well, which implies that µ increases. Now consider the effect of a decrease in φm or an increase in φt . Either change leads to a decrease in the relative price at the modern store, and, hence, an increase in both ψ˜ 0 and ψ˜ 1 . By the expression for Ym above, it follows that Ym rises, and, by similar reasoning, Yt falls. This implies that µ increases. To understand the welfare implications of policies that increase µ , consider a policy that increases φt . As shown above, this would increase µ , and by (10), this would increase aggregate productivity in the retail sector. To see the effect on average consumption and leisure, note that equilibrium average consumption is given by G(ψ˜ 1 ) 1 − G(ψ˜ 1 ) G(ψ˜ 0 ) 1 − G(ψ˜ 0 ) C = wn α + + (1 − α ) + . pm pt pm pt When φt increases, the relative price at the modern segment must fall. By (7), this means that ψ˜ 0 and ψ˜ 1 increase. Since pm < pt , this implies that C increases. Now consider what happens to average shopping time, which, in equilibrium, equals S = αγ [νm G(ψ˜ 1 ) + νt (1 − G(ψ˜ 1 ))] + (1 − α ) [νm G(ψ˜ 0 ) + νt (1 − G(ψ˜ 0 ))] . Using the fact that an increase in φt means that ψ˜ 0 and ψ˜ 1 increase, it follows that S increases as well since νm > νt . Thus, leisure falls, and the welfare effect of an increase in φt is ambiguous.
B.2 Equilibrium of Quantitative Version An equilibrium is defined as prices, {w, r, pm , pt , pa }, quantities {ci , yim , yti , ai , sim , sti } for all i, and aggregates {Ym ,Yt , A, Lm , Lt , Lx , Km , Kt , Kx , Xm , Xt , Xa } such that three conditions hold. First, given 39
prices, the quantities solve the households’ problem. Second, given prices, the quantities solve the firms’ problems. Third, markets for Ym , Yt , A, X , labor and capital all clear. The market clearing condition for Ym is that Ym =
Z ψ˜ (0) Z η˜ (ψ ) T + r + wη n 0
0
pm
dH(η ) dG(ψ ) +
Z ψ˜ (1) Z ∞
T + r − pa + wη n dH(η ) dG(ψ ), pm η˜ (ψ )
0
where the first integral represents the quantity purchased by households without cars but choosing the modern segment, and the second integral represents the quantity purchased by households with cars and choosing the modern segment. Similarly, the market-clearing condition for Yt is that Yt =
Z ∞ Z η (˜ψ ) T + r + wη n ψ˜ (0) 0
pt
dH(η ) dG(ψ ) +
Z ∞ Z ∞ ψ˜ (1)
T + r − pa + wη n dH(η ) dG(ψ ), pt η˜ (ψ )
where the first integral is the quantity purchased by households without cars and choosing the traditional segment, and the second integral is the quantity purchased by households with cars but choosing the traditional segment. The market-clearing condition for autos is that the total quantity of the intermediate good used to produce cars equals the number of intermediates needed per car times the fraction of households demanding a car: Z Z ∞
Xa = A
0
∞
η˜ (ψ )
dη dψ .
The market-clearing condition for X is that the number of intermediates produced equals the demand for intermediates by the modern retail segment, the traditional retail segment, and the automobile sector: X = Xm + Xt + Xa , where Xm = Ym and Xt = Yt . The market-clearing conditions for labor and capital are: E[η i ] n = Lm + Lt + Lx and 1 = Km + Kt + Kx .
B.3 Additional Explanation of Calibration Procedure Labor Shares in Production. To discipline the labor shares in the retail sector, θm and θt , I use the retail census data described in Section 2.1. For each retail segment, I measure the labor share as the estimated cost of labor divided by value added. As Gollin (2002) points out, a key challenge in measuring labor shares is to account properly for self-employed workers, who do not receive wage income directly. This is particularly important in the retail sector, as unpaid workers form a large fraction of all workers in developing countries. To estimate the labor cost of unpaid workers, I follow Gollin (2002) and Young (1995) and assign unpaid workers the average labor cost of paid workers. I then compute the total estimated cost of labor as the sum of total labor costs plus the imputed labor costs of unpaid workers. Table A5 reports the estimated labor shares by retail segment for Argentina, Mexico, El Salvador and the Philippines, the countries that have the data necessary to make the calculations described above. Interestingly, in all of these countries, the modern segment has a lower estimated labor share than the traditional segment. For the modern segment, the estimated labor shares are in the range
40
Table A5: Estimated Labor Shares by Retail Segment Argentina (2004) Mexico (2004) El Salvador (2005) Philippines (2005)
Modern 0.38 0.41 0.42 0.39
Traditional 0.59 0.60 0.65 0.66
Note: Labor shares are computed as the ratio of labor cost to value added, and are computed using the retail censuses described in Section 2.
of 0.4, with Argentina and Thailand lower, at 0.38 and 0.39, and Mexico and El Salvador higher, at 0.41 and 0.42. Similarly, traditional stores have estimates in the range of 0.6, with Argentina again lowest at 0.59, Mexico following at 0.60, and El Salvador and the Philippines the two highest at 0.65 and 0.66. Reassuringly, the estimates for all four countries are similar by segment. Therefore, I set θm to be 0.40 and θt to be 0.62, which are the average estimates across the four countries. Time Savings of a Car. For the time savings of a car, γ , I choose a value of 0.48, as follows. I assume that owning a car saves shopping time in two ways. First, it reduces the travel time per shopping trip, and second, it reduces the number of shopping trips (because the household can buy in bulk). Assuming that total shopping time satisfies total shopping time = # of trips × (in-store time per trip + traveling time per trip), one can back out the total time savings of owning a car by computing (1) the travel-time reduction of a car versus public transportation; (2) the reduction in the number of trips with a car versus public transportation; and (3) the fraction of per-trip shopping time consisting of traveling. To compute travel-time savings, I use data on commuting time, which is widely studied. The U.S. Department of Transportation (2004) reports in its National Highway Transportation Survey (NHTS) that in 2001, an average commute by private automobile covered 35.2 miles per hour, whereas an average commute by public transport covered 19.6 miles per hour. This suggests that public transport is 56-percent as fast as travel by car for a given commuting trip. For the fraction of shopping time that represents travel, I compute, using the 2003 American Time Use Survey (ATUS) that travel is exactly 50 percent of total shopping time. Finally, for trip reductions by car, the NHTS reports that 86 percent as many shopping trips are taken with a car as with public transportation. I assume that this reflects the car’s advantage in economizing on the number of shopping trips. Given that 91 percent of households are car owners, this implies that car owners make 61 percent as many shopping trips as households without cars (since (0.86/0.91)/(1 −0.86)/(1 −0.91) = 0.61.) Putting these numbers together, we see that shopping time with a car is a fraction 0.61 × (0.5 × 1 + 0.5 × 0.56) = 0.48 of shopping time without a car. Other Parameters Calibrated Directly. For the average time spent working, n, I choose a value of 0.24 to match the average working time per capita in the United States for 2000, calculated using U.S. Census data from the Integrated Public Use Microdata Series (IPUMS) (Minnesota Population Center, 2014). This fraction comes from the ratio of average weekly time spent working (among 41
all individuals 18 or older), which is 26.9 hours, divided by 112 hours per person, which represents all non-sleep time. For the tax rates on labor and value added, I choose values of τ L = 0.13 and τ VA = 0.09, which are the values reported by the World Bank’s Cost of Doing Business database for the United States. Appendix A.2 provides a more detailed description of how these tax rates are calculated in all the countries. For the tax compliance rates, φm and φt , I choose values of one for each, which represents full compliance. In the counterfactual experiments to follow, I lower this value in accordance with tax compliance rates in developing countries. For the distribution of efficiency units across households, I pick ση to match the log variance of household income in the 2000 U.S. Census, which is 0.97. I then set µη so that E[η i ] equals one, as assumed earlier. Since E[η i ] = exp(µη + 0.5ση2 ), this implies setting µη = −0.49. Finally, I set the efficiency of the economy to be E = 1 as a normalization. Jointly Calibrated Parameters. The target for the ratio of prices of modern to traditional stores is 0.78, which comes from a study of Hausman and Leibtag (2007), who estimate that the average product in a select basket of food items is priced 78-percent as high in the modern retail segment, consisting of “supercenters, mass merchandisers, and club stores,” as in traditional grocery stores. This estimate is consistent with the work of Basker (2005), who finds that Wal-Mart prices are in the range of 61 percent to 83 percent as high as those of competing grocery stores, depending on the product. For the fraction of aggregate employment in the retail sector, I target a value of 0.17, which is the value for the U.S. in 2000, according to the BEA. For the average shopping times among households in the top and bottom halves of the income distribution, I calculate these to be 4.4 and 4.3 hours per week, respectively, using the 2003-2010 American Time Use Surveys; Appendix A.4 provides more details on these calculations. For the modern retail share, I target the value of 0.67 found in Section 2.4 of the paper. For the fraction of modern-segment customers that come via car, I target a value of 0.97, which comes from a study of customer shopping patterns by Biba, Rosiers, Theriault, and Villeneuve (2006). For the cost of operating a car, I target a value of $5,600, which is the estimated yearly cost of operating the average car, according to the American Automobile Association (2008).
42
