Land Markets, Resource Allocation, and Agricultural Productivity∗ Chaoran Chen University of Toronto

Diego Restuccia University of Toronto and NBER

Ra¨ ul Santaeul`alia-Llopis MOVE, UAB, and Barcelona GSE July 2017

Abstract We study the role of land markets on factor misallocation in agriculture using detailed farmlevel micro data from Ethiopia. Land is owned by the state and local authorities distribute land-use rights uniformly among rural residents. While a recent land certification reform is aimed at providing tenure security to farmers, land sales are prohibited and land rentals remain restricted, deriving in large variations in rental activity across individuals and space. We exploit the differences in operational scales generated by land rentals across households, locations, and time. Despite restrictions to land transactions, land rentals are associated with a significant reduction in misallocation and an increase in productivity. We also find that rentals increase the proportion of farms utilizing more capital intensive technologies, further contributing to increased productivity. Despite these positive effects of reallocation, aggregate rental activity remains tenuous and, as a result, there is substantial misallocation in agriculture, an elimination of which can increase aggregate agricultural productivity by 136 percent. Keywords: Productivity, agriculture, land markets, land rentals, misallocation, micro data. JEL classification: O11, O13, O55, Q1.

∗

We thank Stephen Ayerst, Loren Brandt, Rui Castro, Murat Celik, Margarida Duarte, Xiaodong Zhu, and seminar participants at Syracuse, McMaster, and Fudan University for useful comments. All remaining errors are our own. Restuccia gratefully acknowledges the financial support from the Canadian Research Chairs program and the Social Sciences and Humanities Research Council of Canada. Contact: Chaoran Chen, 150 St. George Street, Toronto, ON M5S 3G7, Canada, [email protected]; Diego Restuccia, 150 St. George Street, Toronto, ON M5S 3G7, Canada, [email protected]; Ra¨ ul Santaeul` alia-Llopis, Plaza Civica s/n, Bellaterra, Barcelona 08193, Spain, [email protected].

1

1

Introduction

Agriculture plays an important role in understanding differences in income per capita across countries. First, labor productivity differences between rich and poor countries are much larger in agriculture than in non-agriculture. Second, poor countries allocate a larger share of employment in agriculture.1 There is an important literature addressing differences in agricultural labor productivity across countries.2 Within this literature, there is a recent emphasis on factor misallocation steaming from imperfect land markets as a major obstacle limiting agricultural productivity in poor countries.3 A prevalent feature of land markets in many poor and developing countries is that land is hardly tradable across farmers. Land transactions are either prohibited by law or face high transaction costs, preventing the allocation of land to best uses. Typically, the ownership of land resides with the state or the collective and use rights of land are distributed by local leaders in a fairly egalitarian basis, with all members of the collective assigned an equal amount of land-use rights. As a result, to the extent that households are heterogeneous in their farming ability and that land allocations are not based on these differences, it is not surprising that misallocation is a prevalent feature in poor countries.4 We study the role of land markets on factor misallocation in agriculture using detailed farm-level micro data from Ethiopia. Ethiopia provides an excellent case to study because 1

See, for instance, Gollin et al. (2002) and Restuccia et al. (2008). Among many others, see Gollin et al. (2002), Restuccia et al. (2008), Adamopoulos (2011), Lagakos and Waugh (2013), Adamopoulos and Restuccia (2015), Donovan (2016), Chen (2017b), Gottlieb and Grobovˇsek (2016), and Adamopoulos et al. (2017). 3 See, for instance, Adamopoulos and Restuccia (2014), Banerjee and Iyer (2005), and Restuccia and Santaeul` alia-Llopis (2017). 4 There are important benefits associated with an egalitarian land distribution such as achieving a more equitable distribution of wealth or providing insurance agains risks. But to the extent that there are other instruments to achieve these objectives, our analysis emphasizes the productivity costs of land institutions as an essential input into the overall assessment of these policies. 2

2

while the land institution features state control over the allocation and use of land, a recent land certification reform—aimed to promote tenure security to farmers—derived in large differences in rental activity across households and space.5 We exploit the differences in operational scales generated by land rentals in order to assess the impact of land reallocations on misallocation and productivity across producers, regions, and time. Despite restrictions to land transactions, we find that rentals are associated with a significant reduction in misallocation. For instance, across household farms, the dispersion in the marginal product of land (MPLa) and revenue productivity (TFPR)—two summary measures of misallocation—are significantly lower among farmers who rent in land compared to farmers who only operate their land-use rights. Since land sales are prohibited in Ethiopia, land rentals are the only mean of adjusting operational scales of farms, and hence, land rentals allow more productive farmers to operate closer to their efficient scale. We find, however, that most rentals occur between relatives and friends and, as a result, may not necessarily be directing resources to their best use. To investigate this issue, we further disaggregate farmers with rented land between those who rent land at market rental rates and those who rent at rates substantially lower than the market rate. Indeed, market rentals are associated with much lower misallocation than non-market rentals and no rental farmers. Improved resource allocation through rentals means higher agricultural productivity. The efficiency gain of reallocating resources is lower among farmers operating rented land 5

A communist government in power from 1974 until early-1990s expropriated and redistributed all the rural land in the country and land transactions were prohibited by law. While ownership still resides with the state and many of the restrictions to land transactions remain in place, a series of reforms in the 2000s were implemented to grant land certificates to farmers and to partially allow land to be reallocated across farmers via rentals (up to a limit) of the use rights. Because the implementation of these reforms were decentralized to the level of local governments and the timing and extent has differed across regions, land rentals differ substantially across producers, regions, and time.

3

(2.14-fold) compared to farmers without rented land (2.39-fold). We also quantify the extent of misallocation across regions and correlate it with the prevalence of land rentals across regions. Land rentals differ substantially across regions, among 69 zones (counties) in the data, the percentage of rented land varies from 0 to more than 60 percent. Among 234 woredas (districts), this percentage differs from 0 to 91 percent. These differences in land rentals reflect substantial variation in local land market institutions. Consistent with our previous household-level variation in rentals, we also find that the prevalence of land rentals across regions is negatively correlated with the extent of misallocation. This correlation is significant both at the zone level and at the woreda level. On average, a one percentage point higher land rental is associated with a 0.8 percentage point higher agricultural productivity. Constructing a panel data of farm households from the 2013/14 and 2015/16 surveys, we also show that increased rentals at the household level over time improves resource allocation and productivity. We study the broader consequences of misallocation by relating it with the extent of technology adoption in agriculture. Indeed, we find that land rentals increase the proportion of farms utilizing more capital-intensive technologies, further contributing to increased productivity. Despite these positive effects of reallocation, aggregate rental activity remains tenuous and, as a result, there is substantial misallocation in agriculture, an elimination of which can increase aggregate agricultural productivity by 136 percent. To appreciate the limited role of rentals in the data, we note two features of rentals required to achieve an efficient allocation of resources (an allocation that maximizes agricultural output). First, in an efficient allocation, 73 percent of land would be rented whereas in the data only 10.6 percent of land is rented. Second, in an efficient allocation the top 10 percent most productive farms 4

would operate 95 percent of the rented land, whereas in the data these farms operate only 19 percent of the rented land. Hence, not only there are not enough rentals in the data, but also rentals are much less concentrated among productive farms, clearly indicating a limited role of rentals in directing resources to best uses. This is consistent with the prevalence of personal connections in rental activity in the data. A nice feature of the data is that it contains measures of inputs and outputs in each plot of land operated by the household. Therefore, we exploit this feature of the data in order to provide more context and robustness to our analysis. First, because farm households–our basic unit of observation—operate on average 7 plots of land and our farm-level inputs and outputs are the sum of the inputs and outputs of all the plots operated by the household, we show that shocks and classical measurement error at the plot level are mitigated by aggregation across plots at the farm level. For instance, the standard deviation of plot-level TFP and TFPR are 54 and 51 percent higher than the dispersion in farm-level TFP and TFPR; as a result, reallocation gains would be exaggerated at the plot level as compared to the farm level (338 percent vs. 136 percent in our baseline). We also use the plot-level data to construct alternative measures of farm-level TFP to show that our results are robust to using these alternative measures. Second, whereas farm households produce several different crops, plots are typically used to produce a given crop and therefore, we also explore the plotlevel data to quantify misallocation within crops. We find that the pattern of misallocation within crops is consistent with the aggregate, although we do find that factor misallocation is more severe in cash crops than in food crops. Our paper is closely related to the macro development literature on agricultural productivity referred to earlier. We contribute to this literature by empirically linking land 5

reallocations across producers, across time, and across space to changes in misallocation and agricultural productivity, providing a more direct connection to policies and institutions. Because our focus is on differences in operational scales—resulting from rental activity across farm households, across regions, and across time—our results are not only more directly connected to land institutions, but also less subject to specification and measurement issues that could affect the level and cost of misallocation. Our paper also contributes to the micro development literature studying the role of institutions as an obstacle of economic development.6 We differ from this literature by studying the macro implications of these institutions and emphasizing the aggregate productivity effects of reallocation. The paper proceeds as follows. Section 2 describes the institutional background and the data for Ethiopia. In Section 3, we present the basic framework, the calibration, and the main results on factor misallocation and aggregate efficiency gains. Section 4 presents our main results on the relationship between land rentals and misallocation by exploiting differences in land rentals across households, across regions, and across time. In Section 5, we study the impact of misallocation on technology adoption in agriculture. Section 6 performs robustness checks using plot-level data. We conclude in Section 7. 6

See, for instance, Acemoglu et al. (2001), Banerjee et al. (2002), and Banerjee and Iyer (2005), among others.

6

2

Institutional Background and Data

2.1

Institutional Background

Ethiopia is one of the poorest countries in the world. According to the World Development Indicators from the World Bank, Ethiopia’s real GDP per capita adjusted by purchasing power parity (PPP) is only about three percent of that of the United States. Agriculture is an important sector in Ethiopia: around 75 percent of employment is in agriculture. Despite being a poor country, the recent economic performance of Ethiopia is positive, with a strong growth rate of real GDP per capita (PPP) of 6.9 percent annually between 2005 and 2015, and as a result it is sometimes referred to as a miracle of economic growth in Africa. Ethiopia is also an interesting country to study because of its historical institutional background related to land policies and the more recent land certification reforms. Current land institutions in Ethiopia are shaped by historical events, but the prevailing characteristic has been the state control over the allocation and use of land. The evolution of land institutions can be divided into three periods. The first period is the imperial period, which comprises from the mid nineteenth century to 1974. During this period, land ownership was usually granted to political supporters regardless of occupation or use in farming, which created a feudal regime. Further emergence of private property during this period resulted in powerful landlords. The second period, from 1975 to 1991, resulted from the severe social injustices created by the feudal regime that lead to a communist regime. A comprehensive land reform, “Land to the Tiller”, was then implemented. The communist government expropriated all the land in the country and redistributed it to all rural households—adjusting for soil quality and family size—in the form of use rights. Land redistributions were frequent, every one to 7

two years, to achieve an equitable allocation of use rights among the local rural populations, and land transactions were strictly prohibited. The third period started with the collapse of the communist regime in 1991 under a market-oriented government that has largely maintained the policies related to land from the previous regime. Essentially, land ownership still resides with the state and households are assigned use rights by local authorities at the village (kebele) or district (woreda) level. Many of the restrictions to land transactions remain in place. However, land certification reforms have been implemented since the early 2000s to mainly promote tenure security by issuing land certificates of use rights. A key aspect of the land reform is that the implementation was decentralized to local authorities so the timing and extent of certification as well as restrictions to rental activity has differed across regions. While severe rental restrictions remain in place, for example only a fraction of use-rights can be rented and the lessee must dwell in the rural area and be engaged only in farming, rental markets have developed at differential pace across regions. In 2013, around 10.6 percent of agricultural land is rented and 24.7 percent of households formally or informally rent in some land for agricultural production. The percentage of rented land differs greatly across zones (counties), from zero rented land in some zones to more than 60 percent in others. We exploit in our analysis the household-level and zone-level variation in land rentals to assess the impact of land reallocations on misallocation and agricultural productivity.

8

2.2

Data

We use household-level data from the World Bank, the Ethiopia Integrated Survey of Agriculture (ISA). The main data are from 2013/14 survey. This survey provides information over the entire process of crop production, including physical measures of farm inputs and outputs. In the original data, 5,262 households are sampled to be representative of the population, among whom roughly 63 percent live in rural areas and participate in agricultural production. Each household is surveyed twice in a year: the first round is during the planting season and the second round is during the harvest season. Almost all farms in Ethiopia are family farms. Therefore, we treat a family farm operated by a household as our basic unit of production. We construct our measures of inputs, outputs, and TFP at the farm (household) level. A farm operated by a household typically consists of several different plots of land; we therefore aggregate the inputs and outputs of these parcels to the household level. We also explore in a robustness section plot-level variation in productivity. We next detail how we measure inputs and outputs from the data. Agricultural output. Farm output is recorded in physical quantities (kilograms) of different crops.7 The most common crops in Ethiopia based on the percentage of households who produce it are maize (54 percent), sorghum (42 percent), tea leaves (38 percent), coffee (28 percent), wheat (24 percent), barley (23 percent), and horse beans (21 percent).8 To aggregate farm production of different crops we use common crop prices. For our purposes, what is key is that aggregate production at the farm level reflects physical variation in 7

Some farmers have not finished harvest at the time of survey. In those cases, they report the percentage of harvest that is pending. We adjust for that to estimate their total harvest. 8 We restrict our analysis to crops only and hence abstract from livestocks as the production cycles of livestocks are usually longer than one year, which is our data period.

9

output and hence valuing output at common prices allow us to compare output across farms that reflect variation in quantities produced. Less important is what common price we use. Since we observe the prices of crops traded at local markets, for each crop, we compute the median price among all transactions and use it as the common price of this crop. The value of crop production of a farm is estimated by multiplying its physical quantity produced with its common price. We then sum up the values of all crops produced by the farm to obtain the value of gross output of each farm. We also use common prices to estimate the value of intermediate inputs used by farms in a similar way. These intermediate inputs include different kinds of fertilizers and seeds. Note that some fertilizers and seeds are from the farmers’ home production; we evaluate these home-produced goods using common market prices as well. Again, the key in these assumptions is that the aggregate measure of intermediate inputs used in a farm tracks as best as possible physical variation in inputs. We subtract the value of intermediate inputs from the value of gross output and the remaining is the value added of a farm. We use this measure of value added in our analysis as the net farm output. Rain. Since we measure productivity from cross-sectional data, it is important to exclude transitory variation in output from value added. In the agricultural production, the most important shock is precipitation. Rainfall information is provided in the data, recorded as the annual precipitation in millimetres, and we can use it to identify the shocks of rainfall. We create 10 dummies representing different levels of rainfall. Then we regress value added on those dummies and obtain the residual of this regression as the value added excluding the shock of rainfall. This is the measure of value added we use in our analysis. Capital. Farm capital has three components: agricultural tools, transportation tools, 10

and some livestocks. Agricultural tools include sickles, axes, pick axes, traditional or modern ploughs, and water pumps. We observe the physical quantity of these tools owned by each farmer, as well as their prices at local markets. Again, we construct common prices, defined as the median of sell prices, to evaluate these agricultural tools. Transportation tools include hand-pushed or animal-drawn carts and bicycles. The price of transportation tools are not directly available in the data, so we estimate their values using local prices from the internet.9 The livestocks used for agricultural crop production are a bit more complicated though. The survey records three most common species in Ethiopia, cattle, goats, and sheep, as well as the usage of these livestocks. In our measure of capital, we include cattle that are for agricultural or transportation purposes only, while we exclude goats and sheep, which are mainly used for meat, wool, or milk. We also observe the prices at which farmers sell their cattle. Then we construct common prices of cattle, male and female separately, to evaluate the value of these cattle. Finally, we sum up the values of agricultural tools, transportation tools, and cattle as our measure of farm capital.10 Labor. The data provide labor input for every plot of land of a farm, in both the planting season and the harvest season. Labor input includes farmers’ family labor, hired labor, and unpaid labor from other households. Family labor is recorded in hours (the data reports hours per day, days per week, and the number of weeks); hired labor and unpaid labor, 9

We assign the prices of transportation tools as follows: one hand-pushed cart is worth about 6 traditional ploughs; one animal-drawn cart is about 9 traditional ploughs; one bicycle is about 17 traditional ploughs. Note that very few farmers have these transportation tools, so excluding them in the measure of capital would only change our results slightly. 10 To deal with a set of farmers who have zero measured capital but report cultivated land and positive production, we follow Adamopoulos et al. (2017) in imputing an amount of capital to all farms representing a common set of very small tools and structures used by farmers that are not recorded in the data. The amount we assign to each farmer is set to equal ten percent of the median capital-land ratio of farms within the zone, times the amount of land input of the farm. We have verified that our results are not sensitive to the size of adjusted capital or to dropping these households.

11

however, are only recorded in days. We assume hired males work the same hours per day as family members, while hired females and children are assigned fewer hours, consistent with their lower wage bills per day.11 We also assume that unpaid labor from other households work the same hours per day as hired workers of the same identity: for example, unpaid males work the same hours per day as hired males. Finally, we construct farm labor input as the sum of hours from all the three kinds of labor for all land plots of this farm in both seasons. We find that out of the total labor input, 78.6 percent is supplied by household members, 11.9 percent by hired labor, and 9.5 percent by unpaid labor from other households. Land. Land input of a farm, or farm size, is the sum of the size of all land plots operated by this farm. For 95.8 percent of land plots, their size is accurately measured by GPS or, if the field was small, by compass and rope at a precision of 0.1 square meters, while the size of the remaining plots is reported by farmers. Farms are in general very small in Ethiopia. The average farm size in our sample is around 1.2 hectares, compared to 169.2 hectares in the United States as reported in 2007 U.S. Census of Agriculture. The farm size distribution is very skewed to very small sizes: 63.2 percent of households in our sample operate farms smaller than 1 hectare, 85.7 percent of households operate farms smaller than 2 hectares, only 1.7 percent of households operate farms larger than 5 hectares. We note that a plot of land is treated as a part of a particular farm if it is operated by that farmer, regardless of whoever has the use rights of the land. In other words, the size of the farm is the operational scale and not the ownership or use rights of land so when we compute farm size we include rented-in land plots and exclude rented-out plots for each household. The land rental market 11

We assume the ratio of hours worked by hired women and hired men equals the daily wage ratio between them. We also estimate hours per day of hired children in the same way.

12

is, however, relatively under-developed in Ethiopia as only 24.7 percent of households rent in any land and most of these rentals are among families and friends. Among the households that operate any rented in land, only 2.3 percent operate on rented land only. The data has information on how farmers acquire land plots that they are operating. The vast majority of land plots feature use rights that are either inherited or granted by a local leader (39.2 percent and 41.5 percent, respectively). On average, 13.2 percent of land plots are rented or borrowed from other households, and this number differs substantially across regions. We use this information on geographical differences in land rentals to assess the impact of land markets on factor misallocation. Land quality. The survey also records land quality and other geographical characteristics for each plot of land. For each plot, we have information on its elevation, slope, terrain roughness, nutrient availability, nutrient retention, rooting conditions, excess salts, toxicity, and workability. The issue is how to combine these measures of land characteristics into one aggregate measure of land quality. We construct a measure of land quality as follows. We regress the log value added per labor hour on these variables indicating land quality, controlling for log capital and land input per labor hour. This regression estimates how each dimension of land quality affects the value added per labor hour. Then we take the coefficients from this regression to evaluate the land quality index q for each farm. This is an upper bound measure of land quality as some inputs may be correlated with the quality of the land and hence is conservative in our analysis of the extent of misallocation. Ethiopia 2015/16 survey. Our results are mainly based on the Ethiopia 2013/14 survey. We also explore the time variation of rentals by using the new round Ethiopia

13

2015/16 survey.12 The data from the newer round survey are available in the same structure as the data from the 2013/14 survey. We construct farm capital, labor, land, and output for the year 2015/16 in the same way we do for the 2013/14 survey. By combining these two rounds of data, we obtain a panel comprising 83.5% of households from the 2013/14 round. We have described our measures of value added, capital, labor, and land. These measures summarize inputs and outputs of farms, and are used in our quantitative analysis in the next section.

3

Framework of Analysis

We first describe the framework for the analysis and our calibration strategy. We then use this framework to quantify the extent of misallocation in agriculture in Ethiopia and compare our results with related studies.

3.1

Framework and Calibration

We start by describing our framework for the analysis, which closely follows Restuccia and Santaeul`alia-Llopis (2017). Consider a farmer with productivity si ∈ S with the following production function: yi = si1−γ [kiα (qi li )1−α ]γ ,

(1)

where yi is the net output of this farm (measured as value added excluding transitory components), ki is the capital input, qi is land quality, and li is the land input. The parameter 12

An earlier round of ISA is available which in principle would allow us to expand the panel dimension, but we are not able to use the earlier data as a key variable (farm gross output) was not recorded consistently.

14

γ governs returns to scale at the farm level, and α determines the share of capital. In our analysis, we focus on the allocation of capital and land across farms, hence, we abstract from the intensive margin of labor supply in our production function.13 For this reason, when we confront this production function with data, we normalize our measures of output (value added), capital, and land in the data by labor hours. In other words, the production function is in per-hour form. The farmer earns profits, which is a fraction 1 − γ of the farm output; therefore, 1 − γ can be interpreted as the labor share. Given the actual inputs and output (value added) of a farm from the data, the farm-level productivity si can be solved out as 

yi si = αγ ki (qi li )(1−α)γ

1  1−γ

.

(2)

We further denote s1−γ as farm-level TFP. i As a benchmark, we solve for the efficient allocations that maximize aggregate agricultural output subject to the total amount of capital K and land L, and the set of farmers who are heterogeneous in their ability si ∈ S. Given the span-of-control technology specified in (1), the efficient allocation of factors among farms is non-degenerate. In particular, we can show that the efficient allocation of capital and land among farms should satisfy

si kie = P K i si

si and lie = P L, i si

(3)

where kie and lie denote the efficient allocation of capital and land (as opposed to the actual 13

A similar approach is followed in Restuccia and Santaeul`alia-Llopis (2017) and Adamopoulos et al. (2017).

15

allocation ki and li in the data). The efficient output (value added) of this farm is given by

yie = si

(K α L1−α )γ P . ( i si )γ

Then aggregate output is given by !1−γ Ye =

X i

yie =

X

si

(K α L1−α )γ ,

i

where Y e is aggregate output from the social planner’s solution. Therefore, it is the maximum output this economy can obtain given the aggregate amount of resources (capital, land, and number of farmers). The efficient allocation has the following two properties: Proposition 1. The efficient allocations of capital and land are proportional to farmer’s ability:

kie kje

=

e si l i , sj lje

=

si . sj

Proposition 2. The efficient allocation equalize the marginal products of capital and land across farms: MPKi = MPKj , MPLi = MPLj , ∀i, j. In general, the actual allocation of capital and land from the data (ki and li ) are not identical to the efficient allocation from the planner’s solution (kie and lie ). The difference indicates resource misallocation. Furthermore, aggregate output in the data Y d = P

i

P

i

yid =

s1−γ (kiα li1−α )γ is lower than Y e . The difference between the actual aggregate output i

Y d and the efficient aggregate output Y e provides a summary statistic of the impact of

16

misallocation on aggregate output and TFP:

e=

Ye > 1, Yd

where e measures the efficiency gain this economy can achieve if factors are reallocated efficiently. We focus on this efficiency gain e as our main measure of misallocation in our analysis. We also construct a summary measure of misallocation as the dispersion of farm-level revenue productivity (“TFPR”), which in our framework is defined as

TFPRi ≡

yid . kiα li1−α

It is straightforward to verify that the efficient allocation (kie , lie ) equates TFPR across farms. Therefore, we use the dispersion of TFPR across farms to measure the extent of misallocation. In fact, it can be shown that the dispersion of TFPR only depends on farm level distortions and is independent on the underlying productivity distribution S of farmers.14 For a quantitative assessment of misallocation, note that in our framework, the only structure we impose is the farm-level production function specified in equation (1). Therefore, we only have two parameters to calibrate: α and γ, governing the factor income shares of the production function. Estimation of factor income shares in agriculture varies in the literature. Valentinyi and Herrendorf (2008) find that in the United States, the capital, labor, and land shares in agriculture are 0.36, 0.46, and 0.18, respectively. Restuccia and 14

See Hsieh and Klenow (2009) and Adamopoulos et al. (2017) for related definitions of revenue productivity.

17

Santaeul`alia-Llopis (2017) use micro data from Malawi and estimate the capital, labor, and land shares to be 0.190, 0.419, and 0.391, respectively. This discrepancy may arise from the fact that Malawi has lower level of mechanization in agriculture compared to the United States. In fact, Chen (2017a) argues that the capital-output ratio (and therefore the capital income share) in agriculture tends to increase as an economy develops. Ethiopia is typically considered to be at a stage of development similar to Malawi. We therefore assign factor shares according to the estimation of Restuccia and Santaeul`alia-Llopis (2017), which results in the parameter values α = 0.327 and γ = 0.581. Given values of α and γ, together with farms’ actual inputs and outputs observed in the data (ki , li , and yi ), we use equation (2) to solve out for farm-level productivity si . We trim farms whose TFP fall in the top or bottom one percentile of the farm TFP distribution as these possible outliers may reflect measurement errors in inputs and outputs in the data.15 We then use equation (3) to solve out for the efficient allocation of capital and labor (kie and lie ), and contrast it with the actual allocation (ki and li ).

3.2

Misallocation and Efficiency Gain

We now use this framework to quantify the extent of misallocation in the agricultural sector in Ethiopia. We start by describing four sets of facts about factor allocation from the data: the patterns of farm inputs, farm output, marginal products of factors, and farm TFPR. We contrast these patterns to the properties of efficient allocation discussed previously. We also compute the efficiency gain associated with factor reallocation as a summary measure of the 15

Trimming observations at the tails of the farm-TFP distribution is conservative as our calculated efficiency gain is increasing in the farm-TFP dispersion.

18

10

15 Farm Productivity (log)

-10

-15

Land Input (log) -10 -5

Capital Input (log) -5 0 5

0

10

Figure 1: Farm-Level Factor Inputs

20

10

15 Farm Productivity (log)

20

Notes: All variables are in log scale for the purpose of illustration. The red dashed lines show the efficient allocation, where the inputs should be proportional to the farm level productivity. The dark blue lines fit the data of our sample. They are horizontal or even negative sloping, indicating severe misallocation: farms with higher productivity are not able to obtain enough inputs as the efficient allocation dictates.

cost of misallocation. Farm inputs. Figure 1 reports farm inputs (land and capital) against farm productivity. Proposition 1 requires that the efficient allocation of land and capital be proportional to farm-level productivity (red dashed line in the figure). The actual allocation in our sample is, however, very different to this efficient allocation: land input is virtually uncorrelated with farm-level productivity and capital input is negatively correlated with farm-level productivity. Therefore, this Figure indicates severe factor misallocation in the agricultural sector in Ethiopia. Farm output. Figure 2 reports the farm output versus the farm productivity. The planner’s solution requires that farm output should be proportional to farm productivity, which is the red dashed line in the figure. The actual farm output is also increasing in farm productivity, but the slope is much flatter. This means that, compared to the efficient allocation, low productivity farms tend to be larger than they should be, and high produc19

-5

Output (log) 0 5

10

Figure 2: Farm-Level Output

10

15 Farm Productivity (log)

20

Notes: Variables are in log scale for the purpose of illustration. The red dashed line shows the efficient allocation, where output should be proportional to the farm level productivity. The dark blue line fits the data of our sample. It is flatter than the red dashed line, indicating that high productivity farms are not producing as much as they would do without misallocation.

tivity farms tend to be smaller. This pattern suggests that farms face distortions that are correlated with their productivity. Marginal products of factors. Figure 3 shows the marginal products of land and capital at the farm level. Recall that Proposition 2 requires that the marginal product of land (capital) should be equalized across farms, independent of the farm level productivity. In our data of Ethiopia, however, we find a strong positive correlation between marginal product of land (capital) and farm level productivity. This is another piece of evidence of misallocation: farms with higher productivity are not able to obtain enough inputs and therefore their marginal products of factors are higher. In other words, farms with higher productivity are facing higher wedges. TFPR. As discussed earlier, if factors are allocated efficiently, TFPR should be the same across farms in our framework. As a result, we use the dispersion in TFPR across farms as a measure of the extent of misallocation. Table 1 reports the dispersion of TFPR. 20

-5

Marginal Product of Land (log) 6 8 10 12 14

Marginal Product of Capital (log) 0 5 10

Figure 3: Farm-Level Marginal Products

10

15 Farm Productivity (log)

20

10

15 Farm Productivity (log)

20

Notes: All variables are are in log scale for the purpose of illustration. The red dashed lines show the efficient allocation, which requires the marginal products of capital and labor to be equalized across farms. In other words, the marginal products should be on the horizontal line without misallocation. The dark blue lines fit the data of our sample. They are positively sloping, indicating that farms with higher productivity should obtain more inputs to improve the overall efficiency.

Revenue Productivity (LogTFPR) 4 6 8 10 12

Figure 4: Farm Revenue Productivity

10

15 Farm Productivity (log)

20

Notes: Variables are in log scale for the purpose of illustration. The red dashed lines show the total factor revenue productivity (TFPR) of farms associated with the efficient allocation. The dark blue line fits the TFPR of our sample. The correlation between log TFPR and log TFP is 0.90 indicating that TFPR is not equalized across farms.

21

The standard deviation of (log) TFPR is 1.02 in our sample. As a comparison, Hsieh and Klenow (2009) find this statistic to be 0.63 and 0.67 in the manufacturing sector of China and India. The 75 – 25 difference is 1.29 in our sample, compared to 0.82 and 0.81 of China and India. The 90 – 10 difference is also larger in our sample. This comparison indicates that the extent of misallocation is severe in Ethiopia. Severe factor misallocation in Ethiopia is largely associated with a uniform allocation of land-use rights and obstacles to reallocation of these rights as described in the institutional background for Ethiopia. Figure 4 shows that farm TFPR—a summary measure of implicit wedges facing each farmer—tends to be increasing with farm productivity, confirming our earlier characterization that more productive farms unable to operate at larger scales are facing larger implicit distortions. The correlation between log TFPR and log farm TFP is 0.90, indicating that correlated implicit distortions constitute a strong source of misallocation. Table 1: Revenue Productivity (TFPR) S.D.

75 – 25

90 – 10

Obs.

Ethiopia 2013/14: Whole sample

1.02

1.29

2.55

2,672

Hsieh and Klenow (2009) United States, 1997 China, 2005 India, 1994

for the manufacturing sector: 0.49 0.53 1.19 194,669 0.63 0.82 1.59 211,304 0.67 0.81 1.60 41,006

Notes: We compute TFPR as kα ly1−α . The table shows statistics on log TFPR. S.D. is standard deviation, 75 – 25 is the difference between the 75th and 25th percentiles, and 90 – 10 is that between the 90th and 10th percentiles.

Efficiency gain. We estimate the gain in aggregate output and productivity associated with an efficient reallocation of resources. The efficiency gain—which equals the ratio between the maximum output (Y e ) and the actual output (Y d )—is 2.36 in Ethiopia.16 This 16

We also compute the standard deviation of the efficiency gain through bootstrap. The standard deviation

22

means that reallocating resources from the actual to the efficient allocation across the existing farmers, aggregate output will increase by 136 percent. This efficiency gain is comparable to that in the literature in other contexts. For instance Hsieh and Klenow (2009) estimate that the efficiency gains are between 1.9 to 2.3-fold in the manufacturing sectors of China and India, whereas Adamopoulos et al. (2017) estimate the efficiency gain in the agricultural sector in China in the 90’s to be around 1.8-fold. The efficiency gain is, however, smaller than the 3.6-fold gain for agriculture in Malawi reported in Restuccia and Santaeul`alia-Llopis (2017). Note that this does not mean factor misallocation is less severe in Ethiopia than in Malawi. The efficiency gain depends on both misallocation and the dispersion of farm TFP (var(log TFPi )) (and more generally the joint distribution). In fact, the lower efficiency gain in Ethiopia is mostly due to a smaller dispersion of farm TFP (0.87) than that of Malawi (1.20). Within-Zone Misallocation.

There are four levels of administrative divisions in

Ethiopia: regions (states), zones (counties), woreda (districts), and kebele (wards). We have the location information of each farm down to the kebele level. We have in total 2,672 observations, located in 10 regions, 69 zones, and 234 woredas. Due to sample size, we mainly focus our analysis at the zone level since we have a reasonable number of zones and a relatively large number of observations within each zone. We decompose the nationwide efficiency gain of 2.36 into a component of reallocation within zones. To calculate the within zone reallocation, we first treat each zone as a closed economy with given factor endowments of capital (Kz =

P

i∈z

ki ), land (Lz =

P

i∈z li )

and a fixed set of farmers. We then calculate

the efficient output for each zone following the same procedure as the efficient nationwide is 0.13, meaning that the efficiency gain is tightly estimated.

23

output: Yze = Sz1−γ (Kzα L1−α )γ , where Sz = z

P

i∈z

sz,i is the sum of productivity of individual

farms within this zone. Yze represents the maximum output this zone can obtain given its resources. We then aggregate this output to the country level to obtain the aggregate effiP cient output Yˆ e = z Yze . This Yˆ e measures the output this economy can obtain when all within-zone misallocation is eliminated, while there is no reallocation across zones. Hence, by comparing this level with Y e and Y d as defined in Section 3.1, we can calculate the contribution from within-zone misallocation. By only eliminating all within-zone misallocation, the efficiency gain is

Yˆ e Yd

= 1.91. This means that eliminating within-zone misallocation

accounts for log(1.91)/ log(2.36) = 76 percent of the overall efficiency gain. The remaining 24 percent is accounted for by reallocating resources across zones.

4

The Role of Land Markets

Given the substantial evidence of severe factor misallocation in Ethiopia, we now try to understand how poor institutions associated with restricted land markets affect the extent of misallocation. We focus on land market institutions as a recent literature has shown that land misallocation can be an obstacle limiting agricultural productivity in poor countries.17 To assess the importance of land markets on misallocation and agricultural productivity, we exploit farm-level and regional-level differences as well as time variation in the extent of land market activity that affect land reallocations and agricultural productivity. 17

See, for example, Adamopoulos and Restuccia (2014), Adamopoulos and Restuccia (2015), Restuccia and Santaeul` alia-Llopis (2017), Chen (2017b), among others.

24

4.1

Farmers With/Without Rented Land

Land market frictions can distort the allocation of inputs across farms and lower agricultural productivity. We use the micro data from Ethiopia to quantify the impact of land reallocations on misallocation and agricultural productivity. Recall that the institutional setting regarding land in Ethiopia is such that farmers have roughly similar amounts of user rights of land and land sale transactions are prohibited. Therefore, to the extent that farms are heterogeneous in their productivity, rental activity is essential in generating operational scales of farms that are closer to efficient or to put it differently, rental activity is the only means of reducing dispersion in the marginal product of land across farmers. However, despite a comprehensive land certification reform intended to provide tenure security to farmers, the land market remains highly underdeveloped in Ethiopia. In our 2013/14 sample, there are 12,583 parcels of land among which 41.5 percent are inherited, 39.2 percent are granted by local leaders, and only 13.2 percent are rented or borrowed from other households.18 Rented parcels are on average slightly larger than non-rented parcels (0.51 ha versus 0.37 ha). In our sample, 75.3 percent of all household farms operate non-rented land only; 24.7 percent of households formally or informally rent in or borrow some land for production. As a fraction of total cultivated agricultural land, rented land represents only 10.6 percent. In our analysis, we divide households into two groups: (a) the set of farmers with no rented land and (b) the set of farmers that operate some rented land. To the extent that rented land allows farmers to operate closer to their efficient scale, this classification highlights the role of land rentals in determining factor misallocation and 18

The data explicitly record how a parcel of land operated by the household was acquired. We consider a parcel of land to be rented if it is recorded as “rent” or “borrowed for free” in the data.

25

Table 2: Land Markets, Misallocation, and Efficiency Gain Whole Sample -

No Rental (0%)

Some Rental (>0%)

S.D. (log TFPR) S.D. (log MPLa)

1.02 0.94

1.05 0.99

0.90 0.78

Efficiency gain

2.36

2.39

2.14

Observations Sample (%)

2,672 100

2,011 75.3

661 24.7

Notes: S.D. (log TFPR) is the standard deviation of the log revenue productivity defined in the text. S.D. (log MPLa) is the standard deviation of the log marginal product of land. For each group, we calculate the efficiency gain within this group in the same way as we do for the whole sample.

efficiency gains. Table 2 reports the dispersion of log TFPR—a measure of the extent of misallocation— for the two groups separately. The dispersion of TFPR is larger for farmers with non-rented land only. For example, the standard deviation of log TFPR is 1.05 for farmers without rented land while it is 0.90 for farmers with some rented land. The standard deviation of log MPLa is also substantially lower for farmers with rented land. These differences are significant at the one percent level.19 We also report in Table 2 the efficiency gain separately for these two groups. The efficiency gain among farmers who operate non-rented land only is 2.39, slightly larger than the aggregate efficiency gain of 2.36. The efficiency gain is 2.14 among farmers who operate some rent in land. This comparison provides direct evidence that land rentals help improve factor allocation and agricultural productivity. Although land rentals alleviate resource misallocation, the extent of misallocation is 19

We obtain the significance level through bootstrap. Specifically, we randomly draw from the data each observation with replacement for the same number of observations as in our sample and compute the efficiency gain. We repeat this process a large number of times to obtain a distribution of efficiency gains which we use to obtain the confidence intervals.

26

still severe even among farms that operate some rented land. For example, the correlation between farm size and farm productivity is almost zero among farmers without rented land and only 0.08 among farmers with some rented land, whereas the efficient allocation entails a perfect correlation between farm size and farm productivity. Therefore, even though land rentals help to direct land to more productive farms, overall farms are operating far from the efficient allocation, which suggests that land markets are still limited and subject to various frictions. This should not be entirely surprising since as discussed in our institutional background, substantial restrictions on rentals remain in place, such as only a fraction of land-use rights can be rented and farmers must still reside in the rural area and be engaged in agriculture. In addition, rental activity may be hindered by other imperfections such as weak legal institutions. Another way to characterize the limited role of rentals in the data is by noting two features of rentals required to achieve an efficient allocation of resources. The first property is that in the efficient allocation, 73 percent of the land would be rented whereas in the data only 10.6 percent of land is rented. The second property is the concentration of rentals. In the efficient allocation, the top 10 percent most productive farms would operate 95 percent of the rented land, whereas in the data these farms operate only 19 percent. Hence, not only there are not enough rentals in the data, but also there are much less concentrated among productive farms, clearly indicating a limited role of rentals in directing resources to best uses. We consider the following regression at the household level in order to control for farmlevel TFP when comparing misallocation between farmers operating with and without rented land: ξi = β0 + β1 log TFPi + β2 Di + εi , 27

where ξi is our measure of farm-level misallocation, Di is a dummy variable where Di = 1 indicates that the farmer i operates some rented land. The first measure of farm-level misallocation is the absolute value of log farm-level revenue productivity (TFPRi ) relative to the economy-wide average. We consider misallocation as the absolute value since any deviation from the economy-wide average—positive or negative—represents misallocation. d

We define the economy-wide average of TFPR as TFPR = K αYL1−α . Hence, our first measure   i of misallocation is log TFPR . The second measure of farm-level misallocation which TFPR relates more specifically to land is the log farm-level marginal product of land (MPLai )   Yd i , where MPLa = (1 − α)γ . Using relative to the economy-wide average, log MPLa L MPLa each of these measures of misallocation as the dependent variable ξi , we run ordinary least square (OLS) regressions using our whole sample. The estimated coefficients of β2 are −0.15 and −0.11, both significant at the one percent level. Therefore, consistent with our previous analysis, operating some rented land significantly reduces misallocation even after controlling for farm-level TFP.

4.2

Market Rentals versus Non-Market Rentals

A nice feature of our data is that there is detailed information on rentals, including the payment for rentals—both in cash and in kind—and from whom the land is rented. In general, despite the land certification reform discussed earlier, the land market continues to function poorly in Ethiopia. For instance, not only there is a small percentage of farm households that operate rented land, but also the vast majority of land rentals occur between relatives and friends. Our data show that among the 661 households who operate some rented

28

land, 48 percent (316) rent from relatives and 36 percent (240) rent from friends. This suggests that personal connections may be a key determinant of rental transactions rather than willingness to pay through a market process. If this is the case, then reallocations through rentals may not be as efficient in directing resources to best uses as could otherwise be and hence would not have as strong as positive impact on agricultural productivity. We focus on rental payments providing a key indicator of whether land rentals are market based: if the rental price of land is substantially below the market rate, then it is likely that the rental is not market based. We construct the rental rate as the ratio between the rental payment, including both cash and in-kind, and the actual output of a parcel of land. In our data, the median payment of rent-in land is about 52% of the output of a parcel of land. This is consistent with Otsuka et al. (1992), who argue that in poor countries farmers’ payments to landlords are roughly half of their output. Based on this rental rate, we classify rentals whose rates are below half of the median rate, i.e., whose rates are below 26% of the output from the parcels, as non-market rentals.20 According to this specification, a farm household may operate with a mixture of market rentals and non-market rentals. We consider a household operating a market rental as long as there is a positive amount of market rentals (since at the margin the household is comparing the marginal revenue versus marginal cost of the additional land). We consider a farm household operating non-market rentals if all parcel rentals are non-market. Therefore, we categorize farms operating land rentals into those who only have non-market rentals and those who have some market rentals. Using this breakdown, we compare market rentals versus non-market rentals. Recall that in our framework both the marginal product of land (MPLa) and revenue 20

We emphasize that considering different thresholds for the market rates deliver roughly similar results.

29

Table 3: A Detailed Investigation of Land Rentals Standard Deviation (log) (1) (2) (3)

Obs.

MPLa No rental Some rental Non-market rental only Some market rental

0.99 0.78 0.79 0.76

0.88 0.71 0.72 0.71

0.48 0.38 0.39 0.37

2,011 661 173 488

TFPR No rental Some rental Non-market rental only Some market rental

1.05 0.90 0.94 0.87

0.92 0.81 0.85 0.79

0.38 0.34 0.36 0.33

2,011 661 173 488

Control variables: Land quality Zone fixed effect Farm productivity

Yes No No

Yes Yes No

Yes Yes Yes

– – –

Notes: The table reports the standard deviation of (log) marginal product of land (MPLa) and (log) revenue productivity (TFPR) among farmers of different rental categories: Farmers with rentals are further classified into non-market rentals only (where rental rates are below 50 percent the median rental rates) and market rentals. Column (1) reports the original standard deviation, controlling for land quality. We also regress farm-level MPLa (or TFPR) on zone fixed effect and farm productivity, and then we obtain the residuals and calculate the standard deviations based on the residuals. The results are reported in Columns (2) and (3).

productivity (TFPR) should be equalized among farmers in the efficient allocation, so dispersion of MPLa and TFPR represent measures of misallocation. Table 3 reports the dispersion of log MPLa and log TFPR among farmers classified according to market and non-market rentals with different controls. Focusing on the dispersion of MPLa, as we discussed earlier, the dispersion is smaller among farmers who operate some rented land. But the dispersion is substantially lower for farmers operating market rentals as it would be expected. Interestingly, the dispersion of MPLa is larger among farmers who only rent in land at below market rates (non-market rental) than that for all farmers operating some rental land. The

30

dispersion of TFPR shares the same patterns—it is substantially lower for market rentals and higher than some rental among farmers operating non-market rentals. These results suggest that non-market rentals are not as effective as market rentals in terms of improving resource allocation. The described patterns pertain to dispersion measures controlling for differences in land quality across operated land by households in column (1) of Table 3. Another potential concern is that farm output depends on farm productivity as the rent payment ratio may be biased downwards if the farmer is very productive, or on specific institutions at the zone level where the farmer resides. To address these issues, we assess differences in resource allocation between market and non-market land by controlling for land quality, farm TFP and zone fixed effects through the following regression:

ξz,i = β0 + δz + β1 log TFPz,i + εz,i ,

where ξz,i is the log of MPLa or TFPR of farm i, δz is the zone fixed effect, and TFPz,i is the farm TFP. Note that our measure of farm output has already explicitly controlled for the land quality differences, so further adding the land quality index qi in the regression makes no difference to the results. From the regression, we obtain the residuals and then use the residuals to compute the standard deviation. The results are reported in columns (2) and (3) of Table 3, where we add these controls. We find that our results are robust to adding these controls, in particular, the dispersion of MPLa and TFPR are always significantly lower for farmers operating marketed rental land. Note that although not as effective as market rentals, non-market rentals still improve 31

resource allocation relative to non rentals since the dispersion of MPLa and TFPR are in general smaller among farmers with non-market rentals compared to farmers with no rentals. We also find that misallocation tends to be more severe among farmers who pay their rent in cash compared to farmers who pay in kind. This result suggests that there may also be frictions in the output market as well.

4.3

Land Rental Markets across Locations

Our analysis of land rentals has so far focused on individual farmers across all Ethiopia. We now exploit the differences of land rentals across geographical locations. As discussed earlier, Ethiopia is an interesting context to study the role of land markets on misallocation since starting in the early 2000s a land certification reform has granted individual farmers a certificate to land-use rights, aiming to improve tenure security of farm land. Note that although better tenure security is a necessary condition for a well-functioned land rental market, it is not sufficient. Restrictions on land rentals are, however, still present after the reform. Land rentals are only partially allowed up to a limit and under certain conditions. The reform was implemented by local governments with different timing and rules, instead of being implemented at the national level (Deininger et al., 2008). Therefore, the land rental market depends on how the reform was implemented locally, and we exploit the variation of implementation across locations to study the impact of land rentals. This approach follows an important existing literature exploiting regional differences using data from other countries, such as Banerjee and Iyer (2005), Deininger et al. (2011), and de Janvry et al. (2015). For this analysis, we mainly focus on results at the zone level, although our results also hold at

32

1

Efficiency Gain (Log) 2 3

Figure 5: Efficiency Gains at the Zone Level

.01

.05 .1 .25 Rental Percentage

.5

1

Notes: There are 56 zones in this Figure. We only include zones with positive percent of rental land and more than 10 observations. We also trim zones with the highest and lowest efficiency gain as they may reflect potential measurement errors. The size of the circles indicate the number of observations in each zone. The regression line is also accuracy weighted by the number of observations in each zone. The slope of the regression line is -0.10 and is significant at the one percent level.

the woreda level. )γ ) and actual output (Yzd = We calculate efficient output (Yze = Sz1−γ (Kzα L1−α z

P

i∈z

yid )

for each zone in the data as discussed previously. The ratio ez = Yze /Yzd > 1 represents the efficiency gain at the zone level. We also compute the percentage of land rentals Rz in each zone, defined as the ratio between the size of rented land and total land size. This ratio measures directly the land rental market in each zone. As alluded earlier, the land certification reform was implemented by local governments with different timing and rules, so the variation of Rz across zones contains information on how the reform impacted land market activity across zones. Figure 5 illustrates the relationship between the efficiency gain of each zone ez and the

33

1.5

.5

.5

Dispersion of Log MPL 1

Dispersion of Log TFPR 1

1.5

Figure 6: The Dispersions of TFPR and MPLa at the Zone Level

.01

.05 .1 .25 Rental Percentage

.5

1

.01

.05 .1 .25 Rental Percentage

.5

1

Notes: We calculate and plot the dispersions of farm level revenue productivity (TFPR) and marginal product of land (MPLa) within each zone. There are 56 zones in this Figure. We only include zones with positive percent of rental land and more than 10 observations. We also trim zones with the highest and lowest efficiency gain as they may reflect potential measurement errors. The size of the circles indicate the number of observations in each zone. The regression line is also accuracy weighted by the number of observations in each zone. The slopes of the regression line are -0.08 and -0.07, respectively, and both are significant at the one percent level.

percentage of land rentals in each zone Rz . We only include zones with positive land rentals and more than ten observations.21 Clearly, efficiency gains ez are negatively correlated with land rentals Rz : a zone with more land rentals tend to have a lower efficiency gain (and therefore a lower degree of misallocation). In particular, the elasticity weighted by the number of observations in each zone is -0.10 , which is significant at the one percent level. Note also the large variation in rental market activity across zones, from less than 1 percent to more than 50 percent. Similarly, Figure 6 shows that the dispersion of farm revenue productivity (std(log TFPRi )) and the marginal product of land (std(log MPLai )) also decrease with the percentage of land rentals in each zone. Similar to the household-level analysis, we consider the following regression to address 21

For illustration, we also trim zones with the highest and lowest efficiency gain as these outliers cloud out the figure and may reflect potential measurement error.

34

for potential differences in farm TFP dispersion across zones:

log ez = β0 + β1 log(Rz ) + β2 log TFPz + β3 VzTFP + εz ,

(4)

where Rz is the percentage of rented land in zone z, TFPz is the zone-level TFP defined as TFPz =

P

i∈z

si


1−γ

, and VzTFP is the dispersion of farm TFP within zone z defined as

VzTFP = std(log(TFPz,i )). Both TFPz and VzTFP are independent of distortions. We include the zone level efficient TFP and the dispersion of farm TFP in the regression to control for the endogeneity of Rz . We further apply accuracy weight to the regression, where the weight equals the number of observations in each zone. Table 4 shows the parameter estimates of this regression. We estimate the parameter β1 to be -0.05, which is the elasticity between the efficiency gain and land rentals:

β1 =

∆ez Rz . ∆Rz ez

Having estimated this elasticity, we can calculate the efficiency gain from a one percentage point higher rental rate: ∆ez = β1 ∆Rz

ez . Rz

(5)

At the zone level, on average 10.7 percent of land is rented; the average efficiency gain is 1.85.22 Applying equation (5) we conclude that a one percentage point increase in land rental reduces the efficiency gain by 0.83 percentage points. Recall that the efficiency gain is simply defined as the ratio between efficient to actual output; therefore, the result can also 22

To be consistent with the log-log regression, we calculate the geometric mean instead of the arithmetic mean of these two variables.

35

Table 4: Land Rentals and Misallocation across Locations Dependent variable: Coefficient on rentals (β1 ) R2 Observations

Efficiency Dispersion Gain in log TFPR -0.048 (0.028) 0.53 56

Dispersion in log MPLa

-0.045 (0.014) 0.79 56

-0.028 (0.018) 0.63 56

Notes: Each zone is treated as an observation. We estimate the regression specified in equation (4) at the zone level. In particular, the regressions estimate the relationship between the log of regional efficiency gain on the log of rental percentage, controlling for the zone level TFP. The results are in the first column of the table. We also repeat this same regression with the zone-level dispersion of log TFPR and the dispersion of log MPLa as dependent variables and the results are in the second and third columns. Regressions are accuracy weighted by the number of households within each zone. Standard deviations are in the parentheses.

be interpreted as more rentals increasing output and productivity by 0.83 percentage points. We also replicate the analysis at the woreda level. The results are quantitatively similar to our zone-level analysis.23 We also note that the zone-level efficiency gain only eliminates within-zone misallocation, which as we discussed earlier, is on average about 76 percent of the economy-wide efficiency gain. Effectively, land rental activity affects zone-level productivity which in turn affects across-zone reallocation. Therefore, the results on productivity from these regressions are conservative as they abstract from the potential across-zone reallocation effects. We also replicate the regression specified in equation (4) using our two summary statistics of misallocation as the left-hand-side variable: the dispersion of log revenue productivity (TFPR) and the dispersion of log marginal product of land (MPLa) within each zone. The results are in Table 4. We also find that more rentals of a zone are associated with smaller dispersions of TFPR and MPLa, and therefore reduce misallocation. 23 Note that we do not focus the analysis at the woreda level because although we there are more woredas, there are much fewer observations per location (only around 10).

36

The regression analysis across locations confirms our characterization of rentals across households that a more active land market is associated with better resource allocation and higher agricultural productivity, consistent with existing theoretical and empirical studies. Our results also highlight the importance of comprehensive land reforms in developing countries that address not only tenure security—which has been the main focus in most reform episodes—but also the tradability of the land to promote better resource allocation. A land reform which grants secure property rights to farmers and at the same time facilitates land transactions to decouple land use from land rights, can substantially boost agricultural productivity.

4.4

Variation in Rentals over Time

We exploit the availability of recent data from the Ethiopia 2015/16 survey which together with the data from the Ethiopia 2013/14 survey allow us to construct a panel. Although the time dimension is relatively short to gain a full insight into general trends, we find that at the aggregate level, there is little improvement on the land rental markets in Ethiopia. In the 2013/14 survey round, 13.2 percent of all land plots are rented or borrowed from other households, and this number is merely 13.5 percent in the 2015/16 survey round. This indicates that there is little variation of land rentals at the aggregate level during this time period. Similarly, we find relatively small changes over time in rental activity at the zone level. However, we do find more relevant changes in rental activity across farm households and, hence, we focus on farm-level comparisons between the two waves of data. We focus on the balanced panel and separate farmers into two groups: (a) farmers whose

37

market rentals, as a percentage of total cultivated land, increased over time and (b) the farmers whose market rentals did not increase.24 We separately calculate the dispersion of the marginal product of land (std(log MPLai )) for each group and for each year. We find that, for farmer whose market rentals increased, the dispersion of MPLa decreases from 0.76 to 0.70, whereas for farmers whose market rentals did not increase the dispersion in revenue productivity increases from 0.83 to 0.85. This suggests that increased market rentals are associated with lower misallocation relative to no increases in rentals. The difference of the trends between groups, which is (0.85 − 0.83) − (0.70 − 0.76) = 0.09, is significant at the one percent level (standard deviation obtained from bootstrap). Similarly, we also calculate the dispersion of the marginal product of land (std(log TFPRi )) and find that it decreases from 0.82 to 0.76 for farmers whose market rentals increased, while the dispersion remains similar, at 0.89, for farmers whose rentals did not increase. The difference between two groups is statistically significant at the ten percent level. The previous group comparison reflects averages across groups. We consider a more disaggregate analysis by estimating the following farm-level regression polling data from both years with a balanced panel:

ξzit = γz + λt + βdzt + εzit ,

where ξzit is our measure of misallocation for farm i in zone z, γz is zone fixed effect, λt is year fixed effect, and dzt is an indicator which equals one in the second period if market 24

Note that for each farm, we separately estimate TFP for each year, and then we use the geometric average of its TFP over two years as our measure of farm TFP in this cross-time analysis. See Section 6.1 for details on how we obtain this measure of farm TFP (which is alternative measure 4 in Section 6.1).

38

Table 5: Rentals and Misallocation over Time

Dependent variable:

TFPRi

MPLai

β

-0.067 (0.032) 4,422 0.19

-0.084 (0.030) 4,422 0.12

Observations R2

Notes: The estimated coefficient β describes the the relationship between misallocation (log relative TFPR or log relative MPLa) and increased market rentals at the zone level. Regressions are accuracy weighted by the number of households within each zone. Standard deviations are in the parentheses.

rentals increased over time in zone z and zero otherwise. As before, we consider two measures of farm level misallocation ξzit : relative revenue     TFPRzit MPLazit productivity log TFPRt and relative marginal product of land log MPLat , both of which are defined in Section 4.1. The results are reported in Table 5. The estimation of β is negative and significant for both TFPR and MPLa, again confirming our previous results that more rentals over time are associated with a lower degree of misallocation.

4.5

Discussion

Our emphasis has been in connecting misallocation with restrictions to land markets in Ethiopia. Even if the land certification reforms have been successful in providing tenure security as their primary objective, we have documented that strong restrictions in rentals remain in place and that even if not everywhere enforced, rental activity remains tenuous with most rentals occurring between relatives and friends and hence potentially not directing land to best uses. But to the extent that there may be other frictions in the Ethiopian economy—such as poor infrastructure which makes difficult access to markets in remote rural locations—that may be driving the misallocation we document, it is relevant to assess 39

the extent to which the land market is the dominant source of misallocation in the data as opposed to other frictions. To this effect, we exploit the availability of variables related to distance to market as a proxy to product market distortions and assess the extent to which these variables are related to the farm-specific measures of distortions. In particular, we estimate the correlation between the log of distance to the nearest market and log TFPR for each household farm and find that this correlation is weak although significant, around -0.10. We find similar results when we estimate the same correlation using distance to the nearest city or distance to the nearest major road. Since product market distortions are more likely to be common among farmers within a specific location, this is consistent with our earlier finding that most of the misallocation in Ethiopia occurs within locations (zones) rather than across locations.

5

Land Rentals and Farm Capital

We are interested in the connection of misallocation with other decisions by farmers that may affect productivity.25 For example, misallocation—in this context the inability of farmers to operate on larger farm scales—may discourage the adoption of more advanced technologies or investment in farming. Chen (2017a) argues that land market frictions, especially land segmentation, can substantially delay mechanization and technology adoption in poor countries since modern technologies are profitable only when the operational scales are large enough. Therefore, land rental markets can potentially alleviate land segmentation and encourage farmers to use capital to substitute for labor. 25

See, for instance, Restuccia and Rogerson (2017) for a discussion of the broader consequences of misallocation.

40

The data allow us to study how land rentals affect the usage of farm capital in the stage of land preparation. Land preparation is an activity that lends itself to the adoption of capital to substitute for labor since it is power intensive but not control intensive (Pingali, 2007). In our data, we observe three different ways that farmers prepare their land for the planting season: using tractors (either own or rented), using livestocks, or using human power. Ethiopia is a country at a very low stage of development, hence in our 2013/14 sample we only observe about five percent of farmers using tractors. About 58 percent of farmers use livestocks—where two percent use both tractors and livestocks—and the remaining farmers use human labor only. To study what affects land preparation, we consider a Probit regression among households which has the indicator of using capital (either tractors or livestocks) of preparation on the left-hand-side of the equation. Specifically, denote this indicator of using capital as K: Ki = 1 means household i uses capital in land preparation. To help illustrate the problem, consider the following equivalent latent variable model. Suppose there exists an auxliary random variable K∗ specified as

K∗i

 TFPR  i = β0 + β1 Di + β2 log TFPi + β3 FarmSizei + β4 log + εi . TFPR

Then K can be viewed as an indicator for whether this latent variable is positive: K = 1 if K∗ > 0. In this regression, Di is an indicator whether farmer i rent in any positive amount of land or not. We also control for farm TFP (TFPi ), farm size, and farm TFPR (relative to the economy-wide average) as a summary measure of farm misallocation. Intuitively, higher farm TFP and larger farm size facilitate adopting the capital-intensive technology, 41

while higher farm-level distortion reduces the return to adopting it. Our key parameter of interest, β1 , is estimated to be positive and significant (βˆ1 = 0.13 with a standard deviation of 0.07), meaning that rentals do help promote adopting a more capital-intensive technology. In order to contextualize this estimate, consider a “representative farm” with average farm TFP, average farm size, and average distortion. Our estimate implies that such a farm operating rental land is 4.5% more likely to use capital in land preparation. The estimation of the other parameters also confirm our conjecture that farm TFP and farm size are positively associated with using capital in land preparation, while the farm-level distortion is negatively associated with using capital. We also explore the time variation of using capital in land preparation. Denote Ki,2015 = 1 if farm i does not use capital in land preparation in year 2013/14 while it starts to use capital in year 2015/16 (newly adopted) and Ki,t = 0 otherwise. Similarly, define ζi,2015 = 1 if farmer i increases rentals in year 2015/16 compared to 2013/14 and ζi,t = 0 otherwise. We consider the following Probit regression:

∗ Ki,t

 TFPR  i,t = β0 + β1 ζi,t + β2 log TFPi,t + β3 FarmSizei,t + β4 log + εi,t , TFPRt

∗ where Ki,t is the corresponding latent variable of Ki,t .26 The key parameter, β1 , is again

estimated to be positive and significant (βˆ1 = 1.40 with a standard deviation of 0.06), meaning that a farmer that increases market rentals is more likely to switch to capital in land preparation. Again, a “representative farm” with average farm TFP, farm size, and distortion, that increases market rental is 33% more likely to switch to capital as the means 26

We do not include year fixed effects in this regression since they are not separately identified.

42

to land preparation.

6

Robustness and Extensions

A merit of the dataset we use is that it records inputs and outputs for each plot operated by all households. We exploit this feature to check the robustness of our results to different measures of farm TFP and to study the role of crop composition and land quality on the extent of misallocation.

6.1

Robustness

An important concern in the misallocation literature is the possibility of measurement errors in individual inputs and outputs in driving dispersion in marginal products across producers. Also important is the measure of productivity estimated from farm inputs and outputs, which could potentially suffer from classical measurement error. Before we try to potentially address these issues, we emphasize that our main focus in this paper is on the comparison across groups of farmers who differ on whether they operate rented land or not, the comparison across regions who differ in the extent of land rentals, and the comparison of household changes in rented land over time, while the relative size of measurement errors should be similar across household groups, regions, and over time. Therefore, while the level of dispersion in marginal products may reflect some measurement error, differences in the dispersion of marginal products across differing groups (location and time) should not be as affected by measurement error. Our quantitative results on the economy-wide efficiency gain are, however, sensitive to

43

the measure of farm productivity and measurement errors in farm inputs and outputs. Note that we measure output, labour input, land input, and capital input in physical quantity; therefore, we require fewer assumptions than the typical revenue-based productivity analysis, e.g. Hsieh and Klenow (2009). Nevertheless, in this section, we provide several alternative ways to measure farm productivity, and conclude that our quantitative results are fairly robust to these alternative measures of farm productivity. Recall that our basic unit of production is the farm household. Since in the data a household operates a farm which contains several plots of land, in the case of Ethiopia an average of more than 7 plots, we have aggregated inputs and outputs to the household level, assigning a unique farm productivity to a household. An important implication of this aggregation at the farm level is that plot-level shocks and classical measurement errors on inputs and outputs are averaged out. To illustrate the importance of this feature of our analysis, we contrast our results with the alternative where the unit of production is each plot of land. In this alternative setup, each plot of land operated is assigned a unique productivity computed from its inputs and outputs as done previously for the farm household. In particular, the land and labor inputs and the value of output (value added) are available for each plot of land (a field in the data). The capital stock is, however, slightly more complicated. Capital stocks are measured at the household level; furthermore, it is reasonable to assume that the capital (for example, a plough) of a given household can be used in multiple plots owned by this household. As a result, we cannot directly assign capital to different plots using our data. To deal with this problem, we assume the plot-level capital inputs are proportional to the size of plots: a larger plot uses more capital than a smaller plot. Then we split the household level capital into different plots according to this rule. After 44

obtaining plot-level inputs and outputs, we consider the plot as the basic unit of analysis and calculate the plot-level productivity as we do in Section 3. Consistent with a large micro development literature, there is a lot of plot-level variation in productivity even within farm households. In particular, the dispersion in plot-level TFP is 54 percent higher than the dispersion in farm-level TFP. (See Figure 7.) The dispersion in plot-level revenue productivity is also about 51 percent higher than the dispersion in farmlevel TFPR. As a result, the efficiency gain from reallocation is much higher at the plot level (338 percent) than at the farm level (136 percent). Hence, aggregation of plot-level inputs and outputs within farm households as done in our baseline results is essential in providing a more robust characterization of the extent and consequences of misallocation in agriculture.27 We can also exploit the plot-level data to explore alternative productivity measures at the farm level. Rather than aggregating inputs and output of all the plots operated by the household, we can take the mean, median, or the second highest value of these plot-level productivity as alternative measures of household-level productivity. Once we have these alternative measures, we aggregate inputs and outputs to the household level, and redo our previous analysis using these alternative measures of household productivity, instead of our baseline measure calculated from household-level inputs and outputs. In particular, in Section 3, we have assigned each household a unique household-level productivity si . Suppose this household i has several plots of land, and we have calculated plot-level productivities 27

We note that a key feature of the land institution we emphasize is the weak connection between land use and productivity reflected in a nearly zero correlation between land input and farm TFP. This pattern of misallocation is not much different when characterized at the plot level, for instance, the correlation between log TFPR and log TFP is 0.93 when the unit of production is the plot and 0.90 when the unit of production is the farm household.

45

Figure 7: Distributions of Farm and Plot Level Productivity

Notes: The figure reports the distributions of farm-level and plot-level TFP. The dispersion in plot-level productivity, represented by the standard deviation of log TFPQ, is 1.34, whereas for the farm-level productivity is 0.87.

sij . We can now construct the following three alternative measures of household-level productivity: 1

s1i = (Πj sij ) J ,

s2i = Medianj (sij ),

s3i = max2j {sij }.

The first measure s1i uses the geometric mean of the plot-level productivity to approximate the household level productivity; the second measure s2i uses the median of the plot-level productivity as an approximation; the third measure s3i uses the second highest value as an approximation. These measures are based on the assumption that a household should have the same productivity across plots. Therefore, the variation of plot-level productivity within a household may reflect potential measurement errors or misspecification. We take the mean, median, or the second highest value of plot-level productivity to minimize these measurement errors. 46

Table 6: Alternative Measures of Farm-Level Productivity Measure of farm productivity si Baseline Alternative: 1 (1) s1i = (Πj sij ) J (2) s2i = Medianj (sij ) (3) s3i = max2j {sij } 1 (4) s4i = (si,2013 si,2015 ) 2

Rank Correlation with Baseline si

Efficiency Gain

Observations

–

2.36

2,672

0.72 0.71 0.57 0.76

2.59 2.60 2.67 2.06

2,672 2,672 2,672 2,672

Notes: For each farm household, we construct four alternative measures of productivity: the geometric mean, the median, and the second largest value of the plot-level productivity, and the geometric mean of farm productivity between 2013 and 2015. We compute the Spearman’s ranking correlation of these alternative measures with our baseline farm-level productivity used in Table 2. The results are reported in the first column, and are all significant at one percent level. We replace each farm’s productivity using one of our alternative measures and re-compute the efficiency gain as in Table 2. The results are reported in the second column.

We also explore the panel structure of the data to construct a fourth measure of farm productivity. Recall that we can separately calculate farm productivity for each year in the panel. Assuming that the ability of farmers is stable over this short period of time (and so is farm productivity), we calculate the geometric mean of farm productivity across years as the “persistent component” of farm productivity and use it as the alternative fourth measure of farm productivity to calculate the efficiency gain in the year 2013/14. Note that not all farms in the 2013/14 wave are in the 2015/16 wave and so for these farms we use the 2013/14 productivity level in order to keep the sample of farms the same as in our cross-section analysis. These alternative measures of farm productivity are highly correlated with our baseline farm productivity constructed in Section 3. The Spearman’s rank correlation between our baseline measure of household-level productivity and these three alternative measures are 0.72, 0.71, 0.57, and 0.76, respectively, and they are all highly significant at the one percent 47

level. Having calculated these alternative measures, we replace the household level productivity si in Section 3 with these three alternative measures and re-compute the efficiency gains. Note that we keep the input unchanged as in Section 3 while we recalculate outputs using these new measures of farm productivity to be consistent. The results are in Table 6. The efficiency gains corresponding to these three alternative measures are 2.59, 2.60, 2.67, and 2.06, respectively. All these four numbers are fairly close to our original estimation of efficiency gain (2.36). Therefore, we conclude that measurement errors are unlikely to be driving our quantitative results.

6.2

Misallocation within Crops

Farmers in Ethiopia cultivate a variety of crops with maize, sorghum, tea leaves, coffee, wheat, barley, and horse beans among the most produced crops. Our production function specification is common across farm households who may be producing different crops and hence differences in the specification of production can generate dispersion in marginal products across farm households due to crop composition. In order to address this potential concern, we explore the extent of misallocation within each crop using our plot-level data. The data records the crop cultivated in each plot operated by a household. We then focus on an individual crop indexed by c. We keep all land plots cultivating crop c, aggregate inputs and outputs of these plots to the household level, and then repeat the analysis in Section 3 to calculate the economy-wide efficiency gain. Table 7 reports the results for ten different crops, which are the most widely cultivated in

48

Table 7: Misallocation within Crops

Crop Maize Sorghum Tea Leaves Coffee Wheat Barley Horse Beans Ensete Haricot Beans Kale

Number of Farms (%)

Cultivated Land (%)

Dispersion in TFPRi

Efficiency Gain

54.0 41.7 38.3 27.8 23.9 22.6 21.0 19.7 17.3 17.1

13.5 17.3 5.5 7.8 2.7 2.5 2.2 6.1 6.1 1.7

1.15 1.00 0.94 1.53 1.09 1.18 1.17 1.47 1.45 1.48

2.25 2.31 1.80 4.17 2.10 2.28 2.80 2.04 3.48 7.01

Notes: The ten crops listed are the most common in Ethiopia. Column 1 reports the percentage of household farms cultivating at least a plot with a particular crop. Column 2 reports the percentage of land used to cultivate a given crop. The last two columns report the dispersion of TFPR and the efficiency gains, as defined in Section 3, when we only focus on farm plots of a single crop.

Ethiopia. The efficiency gains within crops are largely in line with our results in Section 3. Misallocation is less severe within food crops, such as barley, sorghum, and wheat, while more severe within cash crops, such as coffee, horse beans, and kale. This is because the dispersion of farmers’ ability is larger in those cash crops. Intuitively, more productive farmers may choose to specialize in cash crops, while less productive farmers may cultivate food crops. This is consistent with the analysis in Adamopoulos and Restuccia (2015).

6.3

Differences in Land Quality

Another concern is whether there are any land quality differences between rented and nonrented land. For example, it may be that rented land is of higher quality. We assess this possibility using our plot-level data. We construct land quality at the plot level in the same way we constructed land quality at the household level in Section 3 and then compare land 49

quality between rented and non-rented plots. The Welch’s t-test shows that land quality q is only about 3 percent higher among the rented land plots than the non-rented plots, and the difference is not statistically significant.

7

Conclusion

We studied the impact of land markets for factor misallocation and agricultural productivity in Ethiopia. While land markets remain largely restrictive in Ethiopia, a land certification reform mainly intended to promote tenure security among farmers, derived in substantial differences in land rental activity across households, across space and across time. We exploited the impact of these differences on factor misallocation and agricultural productivity. We found that rentals substantially reduce misallocation and increase productivity in all these instances (across households, regions, and time). For instance, at the regional level, a one percentage point increase in rentals is associated to 0.8 percentage points higher agricultural productivity. We also established a significant positive relationship between rentals (and hence reduced misallocation) and the adoption of more capital intensive technology in agriculture. We have used our detailed data at the plot level to corroborate the robustness of these results to potential measurement error in inputs/outputs, to crop composition, and to land quality differences between rented and non-rented land.

50

References Acemoglu, D., Johnson, S., and Robinson, J. A. (2001). The colonial origins of comparative development: An empirical investigation. American Economic Review, 91(5):1369–1401. Adamopoulos, T. (2011). Transportation costs, agricultural productivity, and cross-country income differences. International Economic Review, 52(2):489–521. Adamopoulos, T., Brandt, L., Leight, J., and Restuccia, D. (2017). Misallocation, selection, and productivity: A quantitative analysis with micro data from China. Working Paper. Adamopoulos, T. and Restuccia, D. (2014). The size distribution of farms and international productivity differences. American Economic Review, 104(6):1667–1697. Adamopoulos, T. and Restuccia, D. (2015). Land reform and productivity: A quantitative analysis with micro data. University of Toronto Working Paper. Banerjee, A. and Iyer, L. (2005). History, institutions, and economic performance: The legacy of colonial land tenure systems in India. American Economic Review, 95(4):1190– 1213. Banerjee, A. V., Gertler, P. J., and Ghatak, M. (2002). Empowerment and efciency: Tenancy reform in West Bengal. Journal of Political Economy, 110(2):239–280. Chen, C. (2017a). Technology adoption, capital deepening, and international productivity differences. Working Paper. Chen, C. (2017b). Untitled land, occupational choice, and agricultural productivity. American Economic Journal: Macroeconomics, forthcoming. de Janvry, A., Emerick, K., Gonzales-Navarro, M., and Sadoulet, E. (2015). Delinking land rights from land use: Certification and migration in Mexico. American Economic Review, 105. Deininger, K., Ali, D. A., and Alemu, T. (2008). Assessing the functioning of land rental markets in Ethiopia. Economic Development and Cultural Change, 57(1):67–100. Deininger, K., Ali, D. A., and Alemu, T. (2011). Impacts of land certification on tenure security, investment, and land market participation: Evidence from Ethiopia. Land Economics, 87(2):312–334. Donovan, K. (2016). Agricultural risk, intermediate inputs, and cross-country productivity differences. University of Notre Dame Working Paper. Gollin, D., Parente, S. L., and Rogerson, R. (2002). The role of agriculture in development. American Economic Review: Papers and Proceedings, 92(2):160–164. Gottlieb, C. and Grobovˇsek, J. (2016). Communal land and agricultural productivity. Working Paper. 51

Hsieh, C.-T. and Klenow, P. J. (2009). Misallocation and manufacturing TFP in China and India. Quarterly Journal of Economics, 124(4):1403–1448. Lagakos, D. and Waugh, M. E. (2013). Selection, agriculture, and cross-country productivity differences. American Economic Review, 103(2):948–980. Otsuka, K., Chuma, H., and Hayami, Y. (1992). Land and labor contracts in agrarian economies: Theories and facts. Journal of Economic Literature, 30(4):1965–2018. Pingali, P. (2007). Agricultural mechanization: Adoption patterns and economic impact. Handbook of Agricultural Economics, 3:2779 – 2805. Restuccia, D. and Rogerson, R. (2017). The causes and costs of misallocation. Journal of Economic Perspectives, 31(3). Restuccia, D. and Santaeul`alia-Llopis, R. (2017). Land misallocation and productivity. University of Toronto Working Paper. Restuccia, D., Yang, D. T., and Zhu, X. (2008). Agriculture and aggregate productivity: A quantitative cross-country analysis. Journal of Monetary Economics, 55(2):234–250. Valentinyi, A. and Herrendorf, B. (2008). Measuring factor income shares at the sectoral level. Review of Economic Dynamics, 11(4):820–835.

52