The socioeconomic determinants of agricultural productivity in Brazilian Amazonia

Joshua Skov^*

The agricultural sector of a developing country provides one of the best examples of the gap between best-practice and average-practice technology. Nowhere is that gap more evident than in Brazil, which possesses both a sophisticated modernizing rural entrepreneurial class, and a population of poor, uneducated, often landless rural workers, while tens of thousands of farmers lie along the continuum between the two.

Yet agricultural modernization and productivity growth do not result simply from improved inputs such as better plant genetics and more tractors. The productivity of agriculture, as with every sector of the economy, depends also on more subtle structures of the social context in which production takes place. This paper attempts to measure several such elements of the agricultural context.

This line of inquiry follows logically from even a cursory familiarity with Brazilian agriculture. The obstacles faced by Brazil's rural sector are many: poor infrastructure, inadequate extension services, widespread illiteracy among small farmers, prevalent livestock and plant diseases (in certain regions), and inhospitable or as-yet poorly understood local agroecological circumstances. The solutions to such challenges often have a social as well as purely economic component.

In spite of the availability of these technologies more or less 'off the shelf', we find manifested in Brazilian agriculture an astonishingly wide range of productivity levels, across and within geographical units. This reality leads to a straightforward but complex question: what causes this immense gap between the technologically possible and the current profile of technology in use? In particular, how much of the dispersion is truly technological, and how much is simply due to local variables (such as ecology or infrastructure) or broader socioeconomic ills (such as the availability of capital via credit markets)?

This paper examines these issues in detail with an econometric examination of the nine-state region designated as 'Legal Amazonia'. In Amazonia, these issues are even more pressing than elsewhere in Brazil: in addition to the burden of poverty alleviation and economic growth placed on the agricultural sector, the various Amazonian ecosystems face continued rapid decline due to the pattern of highly extensive land use in the region. Thus, the implementation of best-practice technologies in Amazonia has an added urgency, as the various ecosystems that comprise it are considered to be of importance to all of humankind.

The next section of the paper explains the theoretical foundations of the hypotheses to be tested and the motivation for such questions. In section III I describe the data used and explain the choice of geographical unit of analysis. Next, I report and interpret the results. Section V concludes.

In this section I present a simple framework for thinking of agricultural productivity with the goal of ending up at the reduced form that is the center of the econometric estimation. The reader should bear in mind that, although this paper asks questions about agricultural productivity generally, the focus is on the influence of a set of socioeconomic variables.

Like industrial and service sectors, a modern agricultural sector uses many inputs. A highly general production function for a firm might be of the form

(1) q=f(l,k,t;s)

where l, k, and t are labor, land, and capital, respectively; and S is what is occasionally referred to as ‘social capital’, but which may be thought of more generally as the conditions surrounding agricultural production that contribute to its success, including everything from infrastructure to political stability. At this point, it is probably helpful to think of S as decidedly not firm-specific but instead descriptive of a geographic unit or social stratum of producers.

To capture the specifics of this mysterious S (and move toward a workable reduced form for estimation), consider the production function from equation (1) aggregated for a given geographical area i.

(2) Q_i=f(L_i,K_i,T_i,S_i)

Since production decisions do not take place at the level of a geographical unit (such as a county or state), it is helpful to consider this production function as that of the representative firm for each geographical unit i.

I make several additional assumptions in order to simplify this further for estimation. First of all, I assume a fixed technology, or more precisely, a technology that uses inputs in a fixed ratio. (I defend this apparently outrageous assumption below in detail, but let me mention now that the fixed factor proportions are only by crop, rather than generally.) Second, I consider yield, or Q_i/T_i, as the dependent variable; this, with the first assumption of fixed labor and capital inputs relative to the land input, allows me to consider other variables than those three.

As the mysterious S component is the present focus, I break it into a variety of characteristics that are broadly accepted by non-economists as important features of the social landscape of agriculture: the structure of land tenure and ownership, the size profile of the distribution of agricultural land, and a variety of geographical variables. The resulting equation to be estimated then appears as

(3) (Q/T)_ij = a + b*TENURE_ij + c*FARMSIZE_ij + d*GEOGRAPHY_i + e

where Q/T is the yield, OWNERSHIP, FARMSIZE and GEOGRAPHY are the aforementioned components of S, and i and j are indices for the geographical unit and the crop, respectively.

There are many reasons to look to these socioeconomic sources of agricultural productivity, in addition to farm-level factors. The smooth functioning of markets, for example, has a host of preconditions regarding the availability of information about prices technologies.

The size structure of land holding may also have an important impact on the nature of agricultural production, though not always in obvious ways. Scale economies are widely accepted as important in agricultural production, and for crops with mechanized production we might expect to find some evidence that the concentration of large farm production contributes to yields. But the location of many small farms in the same county could also increase yields, especially for those farms, due to better provision of small-farm-specific goods and services, and due to the improved diffusion of small-farm-specific, productivity-enhancing knowledge. This is the sort of improvement in the 'context of agriculture' which I attempt to measure.

Furthermore, the ownership status can have a strong influence on the behavior of farmers. Agricultural history and the development economic literature both provide many examples of how tenant behavior differs from that of owners or sharecroppers (Sadoulet and de Janvry 1995). Alston et al (1996) also demonstrate how uncertain property rights (the case of squatting) can lead to conflict over land. Such influences may differ across crops, and I look for such differences in my analysis.

Nonetheless, large-scale studies of productivity do not normally take account of this factors (e.g Rosegrant and Evenson 1992). The most recent such study for Brazil (Avila and Evenson 1995) includes several measures of infrastructure, but makes no attempt to measure the effect of the profile of land tenure or property rights.

I describe the precise specification in the section IV. For now, let us make sense of the simplifying assumptions.

First, consider the assumption of a fixed-proportions technology for each crop. This strong assumption means that any one of the inputs can be considered a proxy for the use of the three as a whole; thus, using only land T, it is reasonable to estimate production as a function of land and other factors. Equation (3) above does this, as I drop both labor and capital in the expectation that yield – which by definition includes the land input– proxies the proportional use of labor, capital and land simultaneously. (If the fixed-proportions assumption is weak, the variation in one partial production measure (such as yield) implicitly results in part from the variation in other inputs. This reasoning will resurface in section IV when I discuss the econometric results.)

The assumption of fixed technology is reasonable under certain circumstances, e.g. when a certain crop simply is characterized by little scope for substitution among factor inputs. For example, for many perennial crops that grow as fruit on trees or bushes, there is simply no way to substitute another factor for the labor used in planting or harvesting.

Furthermore, in much of Amazonia, plots sizes are so large that land is seldom a constraint, especially in the case of intensive perennial crops. The absence of a land constraint for many certain crops dispells the suspicion that such crops are being grown with land-saving techniques (i.e. with different mixes of inputs, specifically substituting other inputs for land). Since all regressions are by crop, I effectively allow for the possibility that the fixed-proportions assumption is more valid for some crops than others. I discuss this further with the results.

I also ignore one factor that any farmer knows is of the utmost importance: the suitability of land and climate for a given crop, and land quality in general. The absence of agroecological information in my dataset is a serious weakness.

Last, one might complain that the framework described above is decidedly static, as I make no attempt to examine these issues over time. This decision was made in part due to data availability, but also because a time series should not be necessary: if indeed the aforementioned factors are as relevant as I suggest, I should be able to uncover evidence simply from describing the existing pattern of agricultural productivity for a single point in time.

The data for the analysis are drawn mainly from the Agricultural Census of the Brazilian Institute for Geography and Statistics (IBGE) for 1996. Most of the data were available from IBGE’s extensive on-line archives. Although I relied primarily on the Agricultural Census, some information comes from a variety of other IBGE publications.

Two important methodological issues surface immediately: the proper units of analysis, and appropriate levels of aggregation. On the first count, past attempts to look at total factor productivity have attempted to compile multi-crop indices, essentially viewing the agricultural sector as a whole. This approach is faulty, as it implies uniform economic and social conditions not just throughout agriculture but also across crops. Fortunately, I get around this partly by choosing the crop as my unit of analysis. This is, after all, a more appropriate unit of analysis for which to measure of technology: there may be many ways to skin a cat, but there are not so many differences among successful ways to grow coffee. Aggregation of crops also ignores that some crops are grown predominantly by certain types of farm or farmer. Manioc, for example, is rarely grown in a setting of plantation monoculture, while soy and sorghum come predominantly from large, mechanized operations.

Second, what is the correct geographical level of aggregation? Ideally, one would want farm-level data, with information on everything from soil characteristics to labor quality to credit availability. In the absence of such data, I will at least be able to look at the level of the county, which fortunately is usually quite small (the state of Rondônia, for example, has about fifty counties, and just 1.3 million people).

Thus, the regressions below take the county as the unit of analysis. I choose the county for a variety of reasons. First of all, it is the smallest socio-political geographical unit available that has some coherence above the household level; data are available for (or can be aggregated to) the official category of ‘micro-region’ or above that to ‘meso-region’. Such areas are meaningful for certain purposes, but generally they are so large that their internal heterogeneity can obscure the variation necessary for an econometric study. Second, the county is the smallest unit of political organization in Brazil, and in particular, the seat of a rural county is almost always the largest (and often the only) urban agglomeration in the county, thus taking on an important economic role.

IBGE classifies the status of farm operators in four ways: owner, renter, sharecropper or 'occupant' with uncertain status. This final group consists largely of squatters, many of whom eventually acquire full property rights to the land on which they squat. Owners are by far the largest of the four categories (see Table 2 below), but even squatting (the smallest group, with just over one percent of the cultivated land for the crops examined below) accounts for more than 50,000 ha in Amazonia. The three variables used below are for renters, sharecroppers and squatters, so all results are in relation to owner-operators.

The two other variables used are for population density, overall and rural. My a priori expectation is that more densely populated rural areas fare better due to more smoothly functioning markets for inputs, especially since Amazonia is quite sparsely populated. Quite anecdotally, I can cite complaints I have heard from ranchers and farmers in the region about the challenges posed by this sparseness.

Fortunately, the available data suggest that not all of this idle speculation is hogwash. The regressions below test the aforementioned hypotheses for two dependent variables and three slightly different groups of dependent variables. The dependent variables are Y (the yield for all farms) and SMMEDY (the yield for small and medium-sized farms taken together); the idea behind this partition is that some of these county characteristics may have quite different effects on the performance of farms of different sizes. The explanatory variables include two variables describing size distribution, three describing land tenure arrangement, and two variables describing the density of settlement:

The regression coefficients below for the land tenure variables (SHRRENT, SHRCROP and SHRSQT) should be seen in relation to land that is occupied by an owner-operator. Owner-operated farms are the norm for the four crops examined below, but there are significant amounts land in each of these categories, as for the size groups:

First, consider rice. A traditional staple of the Brazilian diet, rice is commonly cultivated by small and large farmers alike, and tens of thousands of hectares are in rice cultivation in any given year. Indeed, the regressions below show that at least a small amount of rice is cultivated in more than half of Amazonia’s counties.

The most noticeable result from the regressions is the significant SHRBIG coefficient, for which there is an easy interpretation: in recent decades, rice has increasingly been farmed on large-scale, mechanized farms, so SHRBIG may in fact be measuring the effect of capital and other inputs that are more easily available to larger farms.

However, the presence of large farms also has a significant positive effect on the yields of small and medium farms, and effect for which we need some other explanation. One possibility is that the presence of large rice farms increases the ‘thickness’ of markets for inputs and outputs related to the production of rice, thus making small farms more efficient as well. This is precisely the kind of phenomenon for which I am seeking evidence.

A mystery emerges, however, in the variable RURDENS: whereas one might expect denser agricultural settlements to succeed because of thicker markets as described above, rural population density correlates with lower rice yields. A similar result appears in several of the crops examined subsequently, and I discuss possible explanations below. (The inclusion of RURDENS renders the SHRSM significant in the overall yield regression, but this effect does not last when looking at yields for small and medium farms, so it seems a bit of a fluke.)

Still, note the very small share of variation that is explained by the regressions, especially for the yield of small and medium-sized farms (SMMEDY). Other than suggesting the rather unsurprising presence of economies of scale for large rice-producing farms, these regressions say very little.

The next crop is the inseparable companion of rice in everyday Brazilian cuisine: beans. As with the evidence for rice, the regressions for overall yield Y explain a relatively modest amount of the variation.

The most interesting result is that both forms of tenant farming – renting land (SHRRENT) and sharecropping (SHRCROP) – consistently correlate with lower yields.

The share of land in large farms contributes to yield, but note the difference in the SHRBIG coefficient for the two sets of regressions. The combination of SHRBIG’s positive contribution to overall yields and the coefficient’s insignificance in explaining yields for small and medium-sized farms strongly suggests the presence of scale economies in beans production. This probably does not matter much, since little bean production happens on large farms.

Note also that land in small farms seems to boost bean yields. Recall that the size-distribution variables are in relation to medium-sized farms, so the contribution of SHRSM to SMMEDY (which represents production that is roughly evenly split between small and medium-sized farms) suggests that small farms create a different economic environment from which even medium-sized farms can benefit.

With the preceding crops, manioc completes the trio of basic Brazilian foodstuffs. Also known as cassava, manioc is produced overwhelming by small and medium-sized farms in Brazil, so I report only the three relevant regressions below.

The negative effects of the tenancy arrangements that we noticed in the beans regressions appear here as well, and with a greater magnitude: a one percent increase in the share of land operated by either renters or sharecroppers is associated with approximately a one percent decrease in the manioc yield. The prevalence of squatting also brings yields down; this is particularly intuitive, since squatters are likely to have little human and financial capital, and thus should be expected to experience lower yields, as they rely on their own inefficient production of a simple foodcrop such as manioc.

The conundrum, however, is that the presence of manioc production on large farms has a significant positive effect on SMMEDY, which is the manioc yield for small and medium-sized farms. Such an externality is not so intuitive, though it may be related to some unusual characteristic of the few locations with large-farm manioc production.

The final staple to consider is corn, which like rice is commonly cultivated on farms of all sizes. The following regression results for corn lend additional support to most of our suspicions from the previous crops. There are also some additional hard-to-explain results.

Once again, both kinds of tenancy drag down yields, as does squatting. Here the differences are as large as or greater than with previous crops: an additional one percent of corn-producing land in the hands of either sharecroppers or renters corresponds to more than a one-percent decline in corn yields.

As with rice, SHRBIG raises Y considerably, probably reflecting either economies of scale or greater access of large-scale operators to inputs. SHRBIG also appears as significant in explaining SMMEDY, perhaps suggesting that the presence of large-scale corn production may help small and medium-sized producers.

If you want to pay really close attention, it is worth noting that both SHRSM and SHRBIG explain somewhat less in the SMMEDY regressions. In the case of SHRBIG, it is surprising that large-farm corn production explains any of small and medium-sized farms’ yields; in the case of SHRSM, it is unclear why the effect on overall yields would be greater than for SMMEDY, which includes the farms represented by SHRSM. We can suggest the existence of positive small-farm externalities for large-farm corn production, though I find it hard to imagine the form such externalities would take.

The share of land in small farms correlates with higher yields, even more for overall yield than for SMMEDY. The variables for rural and overall population density again have a negative effect on both kinds of yield, as in previous regressions (though the magnitudes involved are quite small, as in previous examples).

The reader may notice that the crops tested above are all annuals. I tried the same specifications on a series of perennials, including cacau, coffee, orange, cupuaçu, papaya and pineapple. Unfortunately, I found almost no econometric results, either in support of or contrary to the hypotheses advanced above. This is perhaps due in part to the smaller number of observations (in most cases, fewer than one hundred) for many of the perennial crops; after all, Amazonia has only 627 counties, and few of them grow significant quantities of a wide range of crops.

It is worthwhile at this point to augment the foregoing discussion by clarifying what these results might mean. First consider the limitations in the data and the level of analysis. As I mentioned earlier, household data would be preferable to the aggregate data used here. Additionally, there are some glaring omissions, such as the lack of agroecological variables. In the future, I hope to refine the current framework in a variety of ways. I hope eventually to look at the impact on yields of overall land tenure structure and distribution, rather than simply the tenure profile and farm-size distribution by crop. Such a modification makes sense if we believe that a local economy can be amenable to farms of certain sizes in ways that are farm-size-specific, rather than crop specific. Another obvious extension would be the use of agroecological data, or historical land use data, in order to take into account the effect of soil fertility and its decline. And of course, in the absence of farm-level data on inputs other than land, proxies for capital and labor would allow me to focus on the socioeconomic variables.

However, given the current emphasis – or, conversely, the absence of variables such the ones I consider – in broad quantitative studies of agricultural productivity, it is remarkable that such a simple study as this one could uncover fairly consistent results for the tenancy variables. For corn, beans and manioc, tenancy and squatting consistently have a negative effect on yield. The slight differences in specification have no effect on this result, either for overall yield or for the yield of small and medium-sized farms.

The major shortcoming, either theoretical or econometric, shows through in the results for the population density variables. As explained in section III, I expected the density measures to contribute to yields, as the nearness of cultivators of a given crop to each other would add to their ability to produce. The opposite result suggests either that this agglomeration effect is negative (which I doubt) or, more likely, that density is highly correlated with an important omitted variable. One candidate is the otherwise omitted agroecology: it is possible that, on average, a more densely populated rural area has been settled for a longer time, and that soil fertility is lower as a result. But this is only a guess, and clearly the effect I had in mind did not appear in the econometric analysis. My only real defense of the density ideas is that I still have to test for density effects at other levels of aggregation.

Although this inquiry is interesting in itself as a test of both the diffusion of technology and the importance of related variables in agricultural productivity, it also has enormous potential policy relevance. Over the past decade and a half, Brazilian policy-makers have vastly reduced the government funds going to agriculture; a fresh look at the importance of the variables considered in this paper could assist the sensible allocation of these scarce resources. And the global importance of the Amazon rainforest underscores the need to understand how technology – especially already existing methods – can reduce forest conversion through agricultural intensification.

If this paper presents evidence for anything at all, it is for the notion that land tenancy is associated with lower agricultural productivity. Fortunately, tens of thousands of hectares of farmland in Brazil are officially ‘unproductive’, and thus present an excellent opportunity to improve productivity. Land reform, often viewed as passé, could perhaps still make an economic contribution to agriculture in the Amazon.

This paper is an attempt to look past the traditional determinants of agricultural output in order to measure what I have called 'socioeconomic' components of the agricultural context. I have attempted to use variables with no necessary relation to agricultural production to explain yield, a basic agronomic measure of success. While the results are mixed, there is some evidence that these variables matter; in particular, the evidence suggests that ownership provides somewhat better economic performance than tenancy and squatting.

This focus is not a complaint about Brazilian agriculture, as that sector has much of which to be proud. Brazil has managed to become one of the world's breadbaskets, ranking as a major (or the largest) exporter of soybeans, coffee, orange juice concentrate, sugar, processed beef, peanuts, black pepper and a variety of horticultural crops. Indeed, best-practice technologies for many crops and livestock animals offer not just a way out of rural poverty, but a path to prosperity for the successful medium- or large-scale farmer. Unfortunately, modernization as it has happened in Brazil has brought about a great deal of social upheaval. Mechanization and the consolidation of rural land have marginalized a large share of the rural population; this group has, for the most part, not shared in the growth. This paper is a rough attempt to measure the economic consequences of this marginalization.

Some of these results should make us pause before giving unequivocal praise to agricultural modernization, or at least the form it has taken in Brazil. The growth of large farm enterprises has displaced many workers from their lands throughout Latin America, and the expansion into Brazil’s frontier regions in Amazonia has done little to alleviate rural poverty (Cunha and Sawyer 1997). Critics of agricultural modernization’s particular nature usually point to the social inequities that have accompanied Brazil’s agricultural modernization. The tentative results of this paper indicate that such undesirable social side-effects can carry negative economic consequences as well.

Alston, Lee J., Gary Libecap, and Robert Schneider (1996). "Violence and the Assignment of Property Rights on Two Brazilan Frontiers," in The Political Economy of Conflict and Appropriation, Michelle R. Garfinkel and Stergios Skaperdas, eds. Cambridge University Press, New York.

Avila, Antonio Flavio Dias, and Robert E. Evenson (1995). Total Factor Productivity Growth in Brazilian Agriculture and the Role of Agricultural Research. Presented at the XXXIII Brazilian Congress of Rural Economics and Sociology, July 31-August 3.

Brunt, Liam (1999). "Technology, Climate and Wheat Production in the Industrial Revolution, 1700-1860." Paper presented to the Economic History Seminar at UC Berkeley, March 1.

Carter, Michael R., and William Chris Walker (1989). "The Evolution of Agrarian Structure in Latin America: an Econometric Investigation of Brazil." Mimeo.

Cunha, Aercio S., and Donald Sawyer (1997). "Agricultural Growth, Sustainability, and Poverty Alleviation in the Brazilian Amazon," in Sustainability, Growth, and Poverty Alleviation, ed. Stephen R. Vosti and Thomas Reardon.

Instituto Brasileiro da Geografia e Estatística (1996), Censo Agropecuário (agricultural census data available on-line at http://www.ibge.gov.br).

Mueller, Bernando, Lee Alston, Gary D. Libecap, and Robert Schneider (1994). "Land and property rights and privatization in Brazil. Quarterly Review of Economics and Finance, v34, p. 261-280.

Rosegrant, Mark W., and Robert E. Evenson (1992). "Agricultural productivity and sources of growth in Southeast Asia," American Journal of Agricultural Economics, v74, p. 757-761.

Sadoulet, Elisabeth, and Alain de Janvry (1995). Quantitative Development Policy Analysis. Johns Hopkins University Press, Baltimore.

Vosti, Stephen R., and Thomas Reardon (1997). Sustainability, Growth, and Poverty Alleviation: A Policy and Agroecological Perspective. Johns Hopkins University Press (for International Food Policy Research Institute), Baltimore.