The Rise of a Middle Class in East and Southern Africa : Implications for Food System Transformation

We show five points regarding the middle class in developing East and Southern Africa. (1) 55% of the region’s middle class – 37% of the “non-vulnerable” middle class – is rural. (2) 61% to 83% of the middle class’s food is purchased. (3) Processed food occupies 7080% of the class’s food expenditure, with similar shares in urban and rural areas. (4) Perishable products account for 44% to 55% of the class’s expenditure. Policy attention to processing and to food products “beyond-grains” thus needs to be “mainstreamed”. (5) The import share of food expenditure does not rise with income in urban areas.


Introduction
Despite a growing literature on the rise of the middle class in developing countries over the past 15 years (Banerjee and Duflo 2008;Birdsall 2007Birdsall , 2010Easterly 2001;Ncube et al. 2011;Nieftagodien and Van der Berg 2007;Ravallion 2010), as yet there is no detailed analysis of middle-class food consumption behavior. Banerjee and Duflo (2008) analyse household data, including on expenditure, from 13 developing countries but look at food only as one of various broad expenditure categories. Birdsall (2010) also analyses much household survey data without looking in any detail at food. Chikweche and Fletcher (2014) examine the determinants of middle-class purchase decisions in four African countries without focusing on food. Burger et al. (2014aBurger et al. ( , 2014b analyse consumption behavior of the emerging black middle class in South Africa without ever mentioning food. Nieftagodien and Van der Berg (2007) provide the most detail, estimating Engel's curves for overall food share among the emerging black middle class and computing total budget shares by total expenditure decile for grains and meat, among other non-food items. No other authors that we can find examine middle-class expenditure behavior (food or otherwise).
This strikes us as a major gap in the literature, for two reasons. First, since many definitions of the middle class start at the international poverty line of US$2/day in purchasing power parity income (Banerjee and Duflo 2008;Ncube et al. 2011;Ravallion 2010), the lower strata of the lower class are certainly still spending a large share of their income on food but may also be changing in substantial ways the food that they buy. Given that the food and agricultural sector in developing countries remains such a large portion of national income-Sub-Saharan Africa (SSA) holds 15 of the top 20 countries in terms of agriculture's share in gross domestic product (GDP), and food takes up about 50 per cent of the average consumer's expenditure-changes in food consumption driven by an emerging middle class will have important impacts on the structure of these countries' economies and on the policies and investments that are needed to ensure robust and equitable growth. Second, it is well-established that health problems associated with the nutrition transition 1 are emerging rapidly in the developing world (Chopra et al. 2002;Popkin, 1999Popkin, , 2001Popkin, , 2002Popkin, , 2003Prentice 2006;Schmidhuber and Shetty 2005). With diet (along with more sedentary lifestyles, see Popkin 2009) as a major driver of this transition, understanding the evolving food consumption patterns of the middle class should be a high priority for those interested in issues of public health, not just agriculture and food system development.
We begin filling this gap by analysing the food consumption patterns of the middle class in East and Southern Africa (ESA). We use this to draw implications regarding likely structural changes in the region's food systems over the next two to three decades. We ask five questions.
1. Will a continuation of the levels and distribution of growth seen in the region over the past 15 years lead to the emergence of a large middle class in coming decades? Some suggest that much of the growth is so dependent on natural resource extraction and so concentrated that it will have little effect on the emergence of a middle class (Potts 2013). We present results that contradict this scenario for ESA.
2. What food and how much food does the middle class in the region currently eat and how are these factors likely to evolve over the coming decades? The story of the rise in consumption of animal proteins and the decline of starchy staples is well known (Bennett's Law); we document this but, of greater interest, we highlight the already surprisingly high consumption of processed foods in general and highly processed foods in particular, their very high expenditure elasticities, and thus the likelihood of explosive growth in demand for them in coming decades.
3. What is the variation in these emerging dietary patterns over space (rural/urban) and over the income distribution? Conventional wisdom suggests they will be largely limited to higherincome urban households but we find broad dissemination of these patterns in rural areas and among low-income households.
4. Will the rise of a middle class drive a surge in imported food and unsustainable food import bills? While concern about such an outcome runs high (Rakotoarisoa et al. 2011), our results show that the import share of food consumption falls with income in urban areas.
5. At what rate is the rise of the middle class likely to drive an increased market share of modern retail (supermarkets)? Supermarket development in Africa is in its very early stages, and there is controversy about how quickly they will spread (Humphrey 2007;Minten 2008;Timmer 2007, 2012;Traill 2006). What do recent growth and changing consumption patterns suggest about the possibilities of more rapid growth of this sector?
We focus on ESA for two reasons. First, the broadly comparable consumption patterns within the region, most of which is dominated by maize-based cropping systems, allows aggregation of country data with less concern for loss of local detail than if we were focusing on larger or more heterogeneous zones. Second, as the least urbanized region of the continent, documenting the diet transformations unfolding in this region puts a lower bound on what one might find in other areas of the continent.

Definitions, data, and methods
There is no commonly accepted definition of the middle class across countries. A basic distinction is between relative approaches, for example, the middle class as all households between 75 per cent and 125 per cent of the median income of a country (Birdsall et al. 2000), and absolute approaches with fixed lower and upper bounds. We follow Banerjee and Duflo (2008) and Ravallion (2010) in using a fixed approach that defines the middle class as those above the international poverty line of US$2/day in 2005 per capita purchasing power parity terms, and below some upper cut-off. We do not follow Birdsall (2010) and her starting point of US$10/day because only about 1 per cent of the region's population currently has incomes above this level and, except under the most optimistic of growth projections, less than 25 per cent will lie above this level even in 30 years' time; using a lower cut-off of US$10/day would provide little to analyse at the present time in Africa. We also show in this paper that consumption behavior changes quite dramatically with rising income well below such a limit. Banerjee and Duflo (2008) use an upper limit of US$10, while Ravallion (2010) uses US$13. We follow Ncube et al. (2011) in using a limit of US$20. We use this higher upper limit because we wish to look forward two to three decades, during which time the currently tiny fraction of the population above this level will likely become far larger, and because by any standard other than that of today's developing world, US$20/day hardly makes one 'wealthy.' To capture diversity in behavior across households, we present our results by income class: US$0-US$2 (the poor), US$2-US$4 (the vulnerable middle class; see Ravallion 2010), US$4-US$10 (the established middle class), US$10-US$20 (the upper middle class), and above US$20 (the upper class).
Beyond the tautology that those in the vulnerable middle class are more likely than others to fall back into poverty, these classes should not be interpreted as anything other than an analytically tractable manner to reflect diversity across income levels within the middle class, and not just between the middle class and others.
Data for the paper come from four sources. (1)  We use these data to characterize current consumption patterns among our income classes, and to generate a set of expenditure elasticities that allow us to project consumption levels and patterns in the region to 2040. 2 (2) The World Bank's PovcalNet database provides distributional data (shares of expenditure by 20tiles) using more than 850 household surveys across 127 developing countries since 1980. We use PovcalNet data from the six countries above to establish recent patterns of inequality in the region's growth for use in the baseline scenario of our projection model. (3) We use the latest (2014) United Nations (UN) data and projections on urban population to establish current rural/urban population shares and as input into the projection model. (4) We use FAOSTAT trade and commodity balance data linked to a detailed commodity breakdown from the LSMS datasets to estimate the net and gross import shares of food consumption patterns for each of our income classes.
We generate a unique categorization of food items in the LSMS datasets, based on a matrix of three levels of processing (unprocessed, low value added processed, and high value added processed) and a dichotomy of perishable/non-perishable. Foods are unprocessed if they undergo no transformation from their original state beyond removal from the plant and (for non-perishables) drying. Processed foods are assigned to the high category if they satisfy at least two of the following three conditions: multiple ingredients; physical change induced by heating, freezing, extrusion, or chemical processes (i.e. more than simple physical transformation); and packaging more complex than simple paper or plastic. The resulting six categories of purchased food allow us to link changing food consumption patterns to key structural changes in the food system-the rise of food processing and the rise in demand for cold chains (based on the perishability of the products). See Appendix B for a listing of the top ten food items (by value) in each of the six categories across all datasets, separately for Ethiopia and the rest of the region. Briefly, the groups are dominated by: 3 • Group 1 (non-perishable unprocessed): legumes, 61 per cent of value • Group 2 (non-perishable low processed): maize meal 36 per cent and milled rice 33 per cent • Group 3 (non-perishable high processed): vegetable oils 26 per cent, breads and biscuits 26 per cent, and food away from home 14 per cent • Group 4 (perishable unprocessed): fresh vegetables 40 per cent, fish 17 per cent, and fruit 12 per cent • Group 5 (perishable low processed): beef 42 per cent, other meat 23 per cent, dried/packaged fish 16 per cent, and processed poultry 13 per cent • Group 6 (perishable high processed): food away from home 57 per cent, milk and milk products 22 per cent, and canned/cooked fish 11 per cent.
The projection model is described in more detail in Appendix A.
3 Key findings 3.1 Will a continuation of recent growth patterns drive the emergence of a middle class? Ravallion (2010) shows clearly that the growth of the middle class across the developing world is strongly tied to the overall rate of GDP growth. Yet the strength of this relationship depends on how growth is distributed across households of different income levels. Potts (2013), among others, questions whether what she qualifies as the highly unequal growth being seen in Africa will drive the expansion of a middle class. We find great variability in the distribution of growth within countries of the region over the past 15 years. Using PovcalNet data, we compute quantile ratios of total percentage growth in per capita expenditure among the top and bottom 5 per cent, 10 per cent, and 20 per cent of the income distributions in Ethiopia, Uganda, Tanzania, Malawi, Mozambique, and South Africa. The period of analysis was from the late 1990s or early 2000s to the latest available data, with period length ranging from five years (Mozambique) to 13 years (Malawi).
Results (Table 1) show that growth was inequality-decreasing (and sharply so) in Malawi, only slightly inequality-increasing in Uganda, and clearly inequality-increasing in Mozambique, South Africa, Tanzania, and Ethiopia. The latter had by far the most unequal growth during the period, but started and even ended with the least unequal distribution. For example, the 5 per cent, 10 per cent, and 20 per cent quantile ratios in Ethiopia at the start of its analysis were 10.0, 6.6, and 4.3, compared to ratios of 66, 40, and 21 in South Africa and 44, 23, and 12 in Malawi. Ethiopia's ratios remained the lowest of the six countries at the end of the analysis, though only slightly lower than those in Tanzania, whose growth during the period was also unequal, but less so than Ethiopia's. South Africa remained far and 5 away the most unequal throughout the period. Growth was typically less unequal when comparing the top and bottom 20 per cent to the top and bottom 5 per cent. Overall, the data suggest that, though recent growth has indeed increased inequality in most countries of the region, it did not do so in all countries, and income grew throughout the distribution of all countries. We captured this distribution and level of growth from PovcalNet, paired them with UN projections of the growth in rural and urban populations over the next 30 years, used the LSMS data from all these countries (except South Africa) to estimate rural:urban total expenditure ratios, and projected the shares of households in each income class out to 2040; the projection thus reflects a continuation of the patterns of growth, in level and distribution, of the past 15 years (Table 2).
Results show that, under these assumptions, the poor will fall from over 70 per cent of the population to under 20 per cent, the middle class will rise from its current 27 per cent to nearly three-quarters, and the share of the vulnerable middle in the overall middle will fall by nearly half, from 73 per cent to only 38 per cent. Thus, even with a continuation of recent unequal growth, a majority of the population should be solidly middle class by 2040 as long as recent levels of growth are maintained.  Table 3 shows the latest food budget shares of each of our six categories plus consumed own production, by income class and rural/urban. For ease of interpretation, it also shows the shares for low-and high-processed foods summed across perishable and non-perishable. Note that processed foods and perishable foods overlap, by design, in our classification scheme. Looking on the left side of the table, unprocessed non-perishable foods have the lowest share among the six categories. Other than own production, this category is the only one whose share falls consistently with income, to well below 1 per cent among the upper class; within the middle class it falls by nearly half from the vulnerable middle to the upper middle. This is consistent with Bennett's Law and expected.

The middle class and the rise of processed and perishable foods
Perishable unprocessed foods have a larger overall share (40 per cent larger than non-perishables, 9.8 per cent vs. 7.0 per cent) and their share more than doubles from the poor to the upper class. This pattern reflects strong rises in the shares of fruit, fish, and eggs, and a very modest fall in the share of fresh produce, from 11.5 per cent among the poor to 9.1 per cent among the upper class (data not shown). This pattern of sharply rising consumption of animal proteins with rising incomes is also widely expected. Note also that the nearly constant budget shares of fresh produce across income classes implies very rapid rises in total per capita expenditure on these items.
The surprising results from this table relate to processed foods. To facilitate interpretation, we also present Table 4, which shows the same data as Table 3 but uses shares of purchased food, not purchased plus consumed own production. Examining the two tables, three results stand out.
First, all processed foods (low-and high-processed together across non-perishable and perishable) constitute nearly 40 per cent of the entire food budget across all households (right side of Table 3), and nearly 70 per cent of purchased food (Table 4). Even more remarkably, the purchased food share of processed foods in Table 4 rises very little with income-the poorest households dedicate nearly as much of their purchased food budget (66 per cent) to processed foods as do the highest income households (80 per cent).
Second, the share of processed foods in purchased food is slightly higher in rural compared to urban areas (71 per cent vs. 68 per cent), which has to count as a major surprise. This finding is driven by (a) the importance of maize meal in rural consumption and (b) the rise of purchased maize meal replacing hand-pounded grain: the share of maize meal in purchased food in rural areas is 8 per cent compared to less than 3 per cent for purchased grain for pounding into meal.
Third, highly processed food has one-third of the purchased food market (Table 4, final column), and shows a sharp rise with income in both rural and urban areas, and also shows little difference in share across rural and urban areas-in each area, the poor dedicate about 30 per cent of their purchased food budget to this category while the upper class dedicates about half. This rise in the share of highly processed food across income classes is driven by sharp rises in nearly every element in this group (Appendix B), especially food away from home, milk, and vegetable oils.
Summing across the perishable food sub-categories in the two tables shows that these foods also see a sharp rise in their budget shares, from 17 per cent of all food (40 per cent of all purchased food) among the poor to 56 per cent of all food (62 per cent of all purchased) among the upper class. From the bottom to the top of the middle class, these shares rise from 24 per cent of all food (43 per cent of purchased) to 43 per cent of all food (53 per cent of purchased). As suggested from the previous discussion, it is the processed categories of perishables, and especially highly processed, that rise the fastest, by a factor of six (in overall food) from poor to upper class, and a factor of more than two from vulnerable middle to upper middle. Purchased food shares of highly processed perishables nearly triple from poor to upper class and increase by 60 per cent from vulnerable middle to upper middle class.
Expenditure elasticities for processed, perishable, and processed-perishable foods are also high (Table  5). In both rural and urban areas, the top three elasticities, all above 1.0, are for (in order) perishable highly processed, perishable low processed, and non-perishable highly processed. This relationship is robust across the income distribution: the same order of elasticities is maintained across all three terciles of total expenditure in both rural and urban areas (data not shown). Also, bottom tercile households have the highest elasticities in every case. Engel's Law predicts that the poorest households would have the highest overall expenditure elasticities for food; the fact that they also have the highest elasticities for highly processed foods must be considered a major surprise.
Using these elasticities, adjusting them appropriately as incomes rise over time (see Appendix A), assuming a continuation of recent growth patterns, and applying UN projections for rural and urban populations, suggest that purchased highly processed foods will rise to a 26 per cent share of all food from their current 17 per cent, and to nearly 40 per cent of all purchased foods, from their current 32 per cent. Purchased perishable foods will rise to a 36 per cent share of all food from their current 24 per cent, and a 52 per cent share of purchased food from their current 44 per cent. Based on sharply higher total per capita expenditure on all food, per capita expenditure on highly processed foods and all perishable foods will each quadruple, while total market size for each (based on higher populations) will increase by a factor of nearly eight. By any measure, such increases over a period of 30 years represent explosive growth.

Will the middle class drive a surge in food imports?
Concern about Africa's rising food import bill dates at least to 1980 (Eicher 1980). More recently, Rakotoarisoa et al. (2011) address the issue. Popular concern also runs wide about the impact of economic growth and the rise of the middle class on the continent's food import bill (Financial Times 2013; USDA 2013). We conduct two analyses to explore this issue. First, we examine two sets of FAOSTAT data to estimate current import shares for each of the 27 food groups we have defined (Appendix A), then link this to each income class's budget share of each of these 27 groups to compute estimated import shares by income class, separately for rural and urban areas. Second, we use FAOSTAT data on import and export values to update Rakatoarisoa to 2011 (focusing on developing SSA) and pair this with World Bank Development Indicators (WDI) data to econometrically examine the drivers of food imports and how developing SSA compares to Asia in its import levels and trends. We use both these analyses to reach tentative conclusions about food import trends that might be seen in Africa over coming decades.
We use two approaches to tie food imports to current household consumption patterns based on our 27 food groups, as follows.
First, we use FAOSTAT data on import and export value for the five countries for which we have LSMS data. Net and gross imports from this approach were compared to total expenditure figures from the LSMS datasets, weighted by country sampling weights and then by each country's population.
Since the expenditure figures are at retail and imports at CIF (cost, insurance and freight), we assumed an average 50 per cent marketing margin from CIF to retail to compare the two. The simple ratio of adjusted import value to estimated expenditure is an estimate of import share. This approach highlighted wheat, oil crops and vegetable oils, and rice as having the highest import shares. These same food groups show only slight declines in total budget shares as incomes rise, and thus sharp increases in total expenditure. Given their importance, we further investigated these figures with the next approach.
Second, we use FAOSTAT commodity balance data for individual ESA countries for wheat, rice, oilcrops and vegetable oil, with an assumed average 20 per cent oil yield on oilcrops. These data result in weighted average estimated net import shares in the region of 45 per cent for vegetable oil (compared to 67 per cent in the first approach), 46 per cent for rice (compared to 21 per cent), and 52 per cent for wheat (compared to 89 per cent). Note that Ethiopia has the lowest wheat import share from this approach (28 per cent) and the largest population, and thus has a heavy influence on the estimated import share for the region. Table 6 reports results from these approaches by income class and rural/urban residence, using net and gross imports. Key patterns from this table are as follows. First, depending on the approach, net and gross import shares are both about four to six percentage points higher in urban than in rural areas. Second, and perhaps surprisingly, net import shares in urban areas decline with income, meaning that the urban middle class has lower import shares (on a net basis) than the urban poor. Gross import shares rise only very slightly across income in urban areas. This surprising pattern of falling net import shares with income among urban consumers is driven by substitution away from wheat and rice towards meat and other products that, at least currently, have lower import shares. Fourth, because rural population shares are high and import shares rise with income in those rural areas, the middle class overall has a higher import share than the poor. Yet these middle-class food import shares are low (about 3-5 per cent on a net basis, around 20 per cent gross) and, given the income dynamic identified in point two, may not necessarily rise sharply as incomes rise and urbanization proceeds. A simple application of these current import shares for each expenditure class to the projected distribution of population across expenditure classes in 2040 (Table 2) suggests no meaningful change in current import shares; net import shares would fall in a range of 1-3 per cent (compared to the 0.2-2 per cent for 2010 in Table 6) and gross shares would rise only to 17-20 per cent from 16-19 per cent. This is not an alarming scenario.
Note also that this scenario is entirely consistent with a sharp rise in per capita food imports in the region.
Our projection model indicates that, under the baseline scenario of a continuation of the recent level and distribution of growth in the region, combined with UN projections on population growth and the urban share of population, per capita cash expenditure on food will rise by 3.4 times from 2010 to 2040. Under this scenario, the slight rise in the net import share of food consumption would lead to a rise of 4.4 times in net per capita food imports. In other words, per capita food imports can rise quite rapidly while remaining a very small net share of food consumption (the 1-3 per cent referenced above).
This analysis, however, ignores the question of whether the production and marketing systems of the region can keep up with the projected dramatic rise in volume and value added of food demand. This question is urgent in light of slow agricultural productivity growth on the continent, the low current value added of food consumed, and the fact that, according to UN projections, there will be fewer than two rural inhabitants per urban inhabitant in 2040, compared to a nearly 4:1 ratio now. The latter fact means that each rural resident will have to feed twice as many urban residents in 2040, and even this assumes no movement out of farming into the rural non-farm sector, which would make the change even larger. Productivity at farm and post-farm levels will have to increase dramatically to avoid an import surge that goes well beyond what we just discussed.
The possibility of such a surge can be seen in Figure 1. To generate the figure we assembled, from FAOSTAT, annual data on per capita value of net food imports since 1980 from all countries in Latin America, developing Asia (Asia minus Japan, Singapore, and South Korea), and SSA. We excluded island nations. We then assembled data from the World Bank WDIs for each country on structural characteristics that should influence the level of imports but not be (strongly or quickly) influenced by those imports. The question to be answered was 'Does SSA import more food than would be predicted from its observable structural characteristics, independent of behavioral/policy/agricultural investment factors?' Specific explanatory variables were: • Real per capita gross national income, in purchasing power parity terms (base = 2010) • The country's urban share in total population • The share of the largest city in total urban population, a measure of the centralization of urbanization • Hectares of arable land per person • Whether the country is landlocked or not (1 = landlocked, 0 = not) • Year, to control for secular trends. We then regressed net per capita real food imports on these variables to generate predicted values for each country, aggregated these regionally, and compared predicted to actual imports. R-squared for the regression was 0.46. The regression left out variables that capture policy and programmatic decisions that influence the productivity of the countries' food systems and thus their ability to produce, process, and distribute the quantity and quality of food demanded by its populace. These could include expenditure on agricultural research and extension, and a range of public investment, including in post-harvest infrastructure such as roads, the electricity grid, and market places, and in education for its workforce. As a result, difference between predicted and actual imports should reflect differences in performance on these variables; imports above (below) predicted levels would reflect inferior (superior) performance relative to the average within the overall sample of countries.
We included Latin America in the regression but exclude it from Figure 1 to highlight the difference in performance between developing Asia and developing SSA. Results are striking. They show that predicted per capita imports in developing SSA have risen slowly but steadily over the period (somewhat more rapidly since the mid-to late 1990s), driven by the temporal pattern of income gains. Actual imports have risen at the same pace but have consistently exceeded the predictions based on observed structural characteristics. This pattern is consistent with the continent's low productivity at farm level and throughout its food system. In contrast, Asia's predicted imports (driven by China) grew dramatically over the period and especially since 2000, driven by the region's exceptionally high income growth. Yet actual imports trended slowly down throughout the period, and were far and increasingly below the predicted levels through the 2000s, suggesting that some mix of policy, programmatic action, and private investment in the food system drove the system-wide productivity gains needed to avoid such an explosion in imports.
Whether Africa repeats Asia's experience or instead sees imports rise rapidly even as a share of consumption depends on whether it adopts the policies and public and private investments that will drive increased productivity throughout its food system. A positive note is that total investment in public agricultural R&D increased 20 per cent between 2001 and 2008, after a long decline. Yet this growth was confined to a small number of countries (Lynam et al. 2012). Crucially, given the rise in demand for value added products documented in this paper, continued and large investments in agricultural R&D must be based on a 'broader policy and strategic framework that encompasses agroindustrial and agribusiness services along with farming' (IFPRI 2011).

The middle class and the rise of modern retail
The rise of a middle class in Africa is likely to fuel growth in the share of supermarkets in the overall food market. For example, Tschirley et al. (2011) show that in Nairobi in 2009, the bottom four quintiles of the income distribution spent an average of only 5 per cent of their food budget in supermarket chains, while the top quintile-with an average income of roughly US$10 per capita per day, spent 26 per cent. In urban Zambia in 2007, the figures were less than 3 per cent for the bottom four quintiles compared to 17 per cent for the top quintile (which had an average per capita income of about US$7/day). In both countries, the top quintile accounted for roughly 80 per cent of all supermarket food sales.
Future supermarket growth in Africa will be driven by the evolution of demand by consumers for supermarket services and by the supply of these services by firms (Reardon et al. 2003). Key drivers on the supply side are policies regarding foreign direct investment (FDI), and policy and investment factors that determine the general cost of doing business in the country. Policies that hinder or promote private investment, whether foreign or domestic, will hinder or promote food system transformation, including the rise of supermarkets. Opening to FDI in Latin America was a key policy change that allowed international food retailing companies to target the urban populations of the continent and spurred the rapid growth of modern retail in that region. Economic liberalization in SSA in the mid-1990s was a necessary trigger for what supermarket expansion there has been to date on the continent. FDI has responded to this openness: between 1998 and 2011, the percentage growth of FDI in SSA exceeded that in East Asia by nearly 5 per cent and was more than three times world average growth. Contrast this with 1970 to 1998, when total percentage growth of FDI in the region was only 14 per cent as large as the world's growth, and 3.4 per cent and 5.7 per cent, respectively, as large as the growth in East and South-east Asia and South Asia. Multinationals from the West, from China, Brazil, and India, and within SSA (primarily but not only South Africa) increasingly see the African consumer market as a growth market, in contrast to years past when investment in Africa was primarily oriented towards international exports.
The reduction of bureaucratic impediments to doing business can be important to attract more local and foreign investment. Public infrastructural investment in water, sanitation, transport, and electricity also are needed if supermarkets are to be able to reduce their costs and provide superior quality of service at prices that most consumers can afford.
Several demand-side factors need to come together to see rapid supermarket expansion. One is urbanization, which has been occurring at generally high rates in the region for several decades. When urbanization and income growth occur together, as they have since the late 1990s, they drive major changes in consumer behavior that favor supermarket development. Among these changes is an increase in the opportunity cost of time, especially for women. Convenience thus becomes of greater value to urban residents. This can work in favor of supermarkets for households that have the ability to buy larger quantities of food at less frequent intervals, which is associated with ownership of vehicles (or at least access to good public transport) and refrigeration. For poorer households without access to vehicles and refrigeration, however, shopping in diffuse informal outlets of the traditional marketing system can be far more convenient (see below for results from Zambia on locational convenience). Public investment in the electricity grid, road infrastructure, and public transport thus also affect the demand for supermarket services. The distribution of income growth-in particular growth strategies that raise incomes for the poorest-can also lead to more rapid growth of a broad market for supermarket services. Finally, urban consumers with growing incomes tend to become more educated, raising awareness of the need for sanitation and a preference to shop in the clean environment that most supermarkets provide.
Two demand-side analyses of shopping in supermarkets are Tschirley et al. (2011) and Neven et al. (2006). Tschirley et al. (2011) analyse the data from four cities of Zambia described above, plus similar data from Nairobi, Kenya. Neven et al. (2006)  Two additional results for Zambia from Tschirley et al. (2011) are noteworthy. First, supermarket chains may have more difficulty gaining market share in large urban centers than in smaller towns; conversely, supermarket expansion may be more rapid in the rapidly growing segment of smaller cities and towns ) than has already been seen in large capital cities. Second, distance to various retail outlets in Zambia has an important influence on choice of outlet. This puts supermarkets at a disadvantage compared to the traditional sector in competing for the low-income consumer, one which they are trying to address by creating a denser network of smaller format stores.
These results, together with the changes in consumer demand patterns discussed above, clearly suggest that an economically growing and urbanizing SSA that remains open to private investment will see continued growth of supermarkets. Recent investments on the continent by Walmart and Carrefour show that large international retail corporations have Africa in their strategic plans. The key question is the rate at which supermarkets will take over market share. The answer matters because the socalled 'traditional' marketing system that currently dominates wholesale and retail food trade in the region suffers from serious and growing problems of congestion and lack of sanitation due to woefully inadequate investment over the past 40-50 years. In developing expectations on this issue, one must first recognize the very rapid growth in overall market food demand and thus the growth that supermarkets must experience just to keep pace and not lose market share: our projections in this paper suggest this growth will average about 6.5 per cent per year in value terms. 4 Dihel (2011) reports that supermarket sales in East Africa are forecast to grow at 10-11 per cent per year over the next five years, after growing between 7 per cent per year (in Kenya) and 15 per cent per year (in Rwanda) over the previous five years. If current total food market shares of supermarkets in the region are, as we suggest, around 10 per cent, and if they are able to sustain growth rates of 10 per cent per year over the next 30 years (compared to our projected 6.5 per cent annual growth in overall food demand), then they will reach a 26 per cent market share at that time. Slightly higher growth rates sustained over this long period of time mean far larger shares: supermarket growth of 12 per cent per year leads to a market share after 30 years of 45 per cent, and sustained 15 per cent growth would result in complete supermarket takeover of the food market in that time! This exercise makes it clear that it is impossible to predict with any certainty what supermarket shares will be that far out. As Reardon et al. (2013) note, many factors can lead to sudden more rapid growth, and as we see above, small changes can make huge differences over time. Second, however, much empirical evidence shows that food systems tend to maintain strong diversity in retail outlets, even in Brazil and other countries of Latin America with far higher incomes than ESA will reach even in 30 years (Booz Allen Hamilton 2003;Farina et al. 2005). Note that currently in South Africa-which today has mean real incomes higher than what we project for ESA in 2040 under the baseline scenario-supermarkets hold a 50 per cent share in the national food market. 5 Thus, if a range of 30-50 per cent is a reasonable expectation for supermarket share in 30 years' time, the most germane conclusion is simply that the traditional marketing system will remain a major actor in ESA food systems for decades to come, even with robust economic growth and expansion of the middle class. Figure 2 summarizes the implications of our findings for likely changes in the structure of the agrifood system in ESA over the next 30 years. The gross import share in the top portion of the figure is the mean of the two gross shares in Table 5. The striped portions in the top of the figure come from the detailed analysis of LSMS consumer expenditure data in the five countries analysed in this paper, and the classification of those expenditures by our six processing/perishability categories, plus consumed own production. These same elements in the bottom portion of the figure come from baseline projection model results for 2040, while the import figures in the bottom come from the analysis discussed above. Shares of the 'traditional' and 'modern' sectors at retail reflect best judgment based on review of the literature and data as discussed in the previous section. The key points from the figure are (1) the stagnant share of food imports between 2010 and 2040 (subject to the caveats discussed above), (2) the rise of purchased processed foods, from 36 per cent of all food to 50 per cent, (3) the even more rapid rise, within this group, of highly processed purchased foods, from 17 per cent to 26 per cent, (4) the rise of purchased perishable foods from 13 per cent to 22 per cent, and (4) an estimated rise in the share of food purchased in supermarkets (the modern sector) from perhaps 10 per cent today 6 to 30-50 per cent in 2040. Two points that the figure does not show are also of great importance. First, under a baseline scenario of a continuation of recent growth patterns (in level and distribution), the value of the overall food market will increase by over five times, and the value of purchased food by nearly seven times. Second, the share of urban consumers in the purchased food market in ESA is already 52 per cent-this in the least urbanized area of the continent-and will rise to 67 per cent by 2040 under the scenario used in this paper. We highlight six implications from this anticipated future of vast increase in the size of the market, its rapid shift to urban areas, the profound changes in the composition of what is consumed, and the likelihood that the share of the modern sector will at least triple while still leaving a large share in the hands of the traditional sector.

Implications for agrifood system transformation in Africa
First, marketing infrastructure, both in the expanding modern sector and the badly under-provisioned traditional sector, needs to expand dramatically to handle the vast increase in food flowing through markets. Policy and practice related to private investment-essentially the costs in time and money imposed on local or foreign firms to invest-will be one prime determinant of whether capacity can expand sufficiently. Policy related to 'public' marketing infrastructure-especially wholesale markets-will also be important. In particular, more flexible approaches to ownership and management of such market places are badly needed. Public investment, especially in transport and electrical infrastructure and in public-private partnership fashion in market places, will also be central.
Second, these public policies and investments must drive productivity growth with a whole agrifood system focus-from farm to retail-if Africa is to avoid (as Asia successfully did) rapid growth in imports as a share of total consumption. The food supply problem is not just a farm production problem, and this will be increasingly true as the rise of processed and perishable foods increases the value added share of total food demand.
Third, the rise of processed foods, and of meat and dairy, will feed an acceleration of the nutrition transition documented by Popkin and others unless strong preventive measures are taken; participants in the Bellagio Declaration on the health implications of the nutrition transition concluded unanimously 'that prevention is the only feasible approach to addressing this epidemic of nutritionrelated chronic diseases. The cost of their treatment and management imposes an intolerable economic burden on developing countries' (Popkin 2002: 102). These preventive initiatives must focus on diet change (more healthy processed foods, since convenience and taste will be increasingly demanded by the rising middle class; leaner meat in smaller amounts) and increased physical activity, as these are the worldwide drivers of the transition, linked to income growth and urbanization. See Haggblade et al. (2013) for a review of policy and programmatic efforts under way.
Fourth, the emerging 'quiet revolution' in traditional food marketing systems documented broadly by Reardon et al. (2013) and in Ethiopia by Minten et al. (2013) is likely to feature (a) proliferation of wholesale markets in secondary towns, following the strong decentralized urbanization that is being seen on the continent ), (b) continued 'dis-intermediation' in supply chains  robust investment response to the demand for processed foods and in cold chains for fresh perishable foods. Regarding the latter, Reardon et al. (2013) survey the evidence and suggest that the so-called 'missing middle' in the African processing sector is in fact being filled by aggressive local investment, especially in dairy in Kenya, cassava in West Africa, middle-sized maize milling throughout ESA, poultry production throughout the continent but especially in Nigeria, and the rise of Zartech (chicken), beef in Zambia and Mozambique, and diverse and rapidly rising processing for breakfast food and snacks in most countries of the continent.
Fifth, however, this response depends crucially on public policy and investment (as discussed above) and introduces huge distributional questions. For the processing sector, the question is whether and how local micro-processors can be helped to grow to become small, and local small and medium processors can be assisted to compete with the large-scale operators (OLAM, Export Trading, East Africa Tiger Brands, a rising tide of Brazilian and Asian firms) that are already investing aggressively on the continent without creating a policy environment that cuts off needed overall investment.
Finally, the rise of supermarkets on the continent will increasingly have five system-wide and interrelated effects (this is likely to happen sooner in places like Kenya and Zambia than in countries of West and Central Africa, or in Ethiopia for example). First, through their operational efficiencies they will eventually drive lower prices throughout the food system, to the benefit of consumers (Minten and Reardon 2008). This is especially likely for the processed and semi-processed goods such as maize meals, wheat flour, bread, oils, meat, fish, and dairy, that typically make up 85 per cent of these stores' sales.
Second, supermarkets are likely to drive consolidation and increased scale of operation in the processing and wholesaling sectors in their push to 'squeeze costs out of the system.' The main question is when and how quickly this will occur, as the process has been quite slow to date but could pick up pace very rapidly with the changes we document.
Third, supermarkets are also likely to reduce the number of smaller independent shops and drive them towards niche markets as they (the small shops) attempt to earn higher profits on declining volumes. In this way, supermarkets can drive peripheral diversification in the food system as existing small retailers search out new markets in an attempt to remain in business.
Fourth, robust evidence indicates that smallholders are largely excluded from the supermarket procurement system, despite much talk of corporate social responsibility and real attempts to include them. Whether this exclusion becomes a policy and political problem depends on the rate of market capture by supermarkets and by the rate of transformation of the farm sector-if robust growth pulls farmers rapidly off the land into other economic activities, the negative impact of this exclusion will be ameliorated.
The fifth and final systemic effect that supermarkets can eventually have in African food systems is reduction of food safety problems through the creation of more hygienic shopping environments and better ability to mainstream food safety practices throughout their supply chain by demanding higher standards from their suppliers. As this happens, traditional markets will need to modify their own practices to avoid even more rapid loss of market share. In this way, competition among these two channels can drive improved practices throughout the system.

FSZ:
Staple consumption patterns vary across the continent depending in part on agro-ecological conditions and related cropping patterns, influenced also by history. For example, the share of maize in total food consumption ranges only from 3 per cent to 6 per cent in West and Central Africa, but from 11 per cent to 21 per cent in ESA. Cassava's share ranges from 21 per cent to 44 per cent in West, Central, and East Africa, but is only 6 per cent in Southern Africa and 3 per cent in the Sahel. Yam consumption shares are well over 10 per cent in Coastal West Africa, Nigeria, and the Horn of Africa, but nowhere else on the continent do they exceed 1 per cent. These sharp differences in staple consumption patterns suggest that the trajectory of change in consumption patterns may also differ across zones. Understanding what these differences might be is one important element in any forwardlooking exercise. Haggblade et al. (2012) systematized these differences to define ten FSZs across the continent. We

Expenditure classes:
We use the income classes from Ncube et al. (2011) in this analysis, based on real per capita purchasing power parity (PPP) income as of 2010 (base = 2005). Landscan population data and PovcalNet data from World Bank 8 were used to create 20-tiles of income for each FSZ and to allocate population from the three FSZs to each income class; 20-tiles were then consolidated into the five income classes.

Food item aggregation and estimation of budget shares and expenditure elasticities:
We used LSMS datasets from five countries and seven years to develop two distinct food item aggregations and to compute budget shares and estimated elasticities. South Africa data were used only in the 22 expenditure elasticity estimates. The two food item aggregations were distinguished by (1) processing level and perishable/non-perishable as explained in this paper and Appendix B, and (2) 27 food groups.
Bennett's Law states that expenditure elasticities decline as total expenditure rises. Properly estimating by how much these elasticities decline with income becomes important when projecting consumption patterns over 30 years, during which time projected incomes will rise well beyond current levels. To generate reliable estimates for our purposes, we used LSMS data from all five countries plus South Africa. In summary, the approach took advantage of the wide variation of incomes across the LSMS datasets to estimate log-linear relationships between total expenditure and expenditure elasticities of demand for each food group, separately by income terciles in rural and urban areas in each FSZ. Elasticities for the projections were then selected using these relationships evaluated at midpoint total expenditure from each expenditure class. The essential gains from this approach are that (1) the regression captured the non-linear relationship that typically exists between elasticities and income, and (2) it did so over a range of income that, due to the inclusion of South Africa, included the highest projected incomes in the region. Finally, we use LSMS data from all countries but South Africa to compute food budget shares and total budget shares for each of the categories explained above.
Both midpoint arc elasticities and Tobit-Engel elasticities were estimated. Models were run with each and delivered very similar results, with the only meaningful difference in 2040 budget shares emerging for consumed own production: models based on Tobit-Engels elasticities projected larger declines in this item than did the models based on midpoint arc elasticities. Projection results mentioned in the paper are based on the average of model results from each set of elasticities.