University of Toronto Department of Economics

January 25, 2017

By Diego Restuccia and Raul Santaeulalia-Llopis

Land Misallocation and Productivity

Working Paper 575

Land Misallocation and Productivity†

Diego RestucciaUniversity of Toronto

and NBER∗

Raul Santaeulalia-LlopisMOVE, UAB, and

Barcelona GSE∗∗

January 2017

Abstract

Using detailed household-level data from Malawi on physical quantities of outputs and inputs inagricultural production, we measure total factor productivity (TFP) for farms controlling for landquality, rain, and other transitory shocks. We find that operated land size and capital are essentiallyunrelated to farm TFP implying substantial factor misallocation. The aggregate agricultural outputgain from a reallocation of factors to their efficient use among existing farmers is a factor of 3.6-fold.We directly link factor misallocation to severely restricted land markets as the vast majority of landis allocated by village chiefs and not marketed. In particular, the output gain from reallocation are2.6 times larger for farms with no marketed land than for farms that only operate marketed land.

Keywords: misallocation, land, productivity, agriculture, Malawi, micro data.JEL codes: O1, O4.

†For helpful comments we thank Aslihan Arslan, Nadege Azebaze, Fancesco Caselli, Leandro M. de Magalhaes,Matthias Doepke, Pascaline Dupas, Chang-Tai Hsieh, Chad Jones, Joe Kaboski, Pete Klenow, Evelyn Korn, PerKrusell, Rody Manuelli, Virgiliu Midrigan, Andrew Newman, Chris Pissarides, Nancy Qian, Richard Rogerson,Jose-Vıctor Rıos-Rull, Paul Romer, Nancy Stokey, Kjetil Storesletten, Kei-Mu Yi, and seminar participants atEdinburgh, ENSAI, SITE, SED, NBER Summer Institute, Toronto, IMF Workshop on Inequality, CIDE Mexico,Queen’s, Marburg, Universitat d’Alacant, CEMFI, RIDGE, Wilfried Laurier University, HEC Montreal, OxfordCSAE, World Bank Land & Poverty Conference, ECAMA-IPFRI-IPA, Notre Dame conference on misallocation anddevelopment, UAB, FAO-Roma, Universitat de Barcelona, and LSE. All remaining errors are our own. Restucciagratefully acknowledges the financial support from the Canadian Research Chairs program and the Social Sciencesand Humanities Research Council of Canada. Santaeulalia-Llopis gratefully thanks the Weidenbaum Center on theEconomy, Government, and Public Policy for financial support.∗150 St. George Street, Toronto, ON M5S 3G7, Canada. E-mail: diego.restuccia@utoronto.ca.∗∗Plaza Civica s/n, Bellaterra, Barcelona 08193, Spain. E-mail: rauls@movebarcelona.eu.

1

1 Introduction

A fundamental question in the field of economic growth and development is why some countries are

rich and others poor. The literature has offered many useful perspectives but we build on two. First,

agriculture is relevant because poor countries are substantially less productive in agriculture relative

to non-agriculture compared to rich countries and allocate most of their labor to this sector.1 Second,

the misallocation of factors of production across heterogeneous production units is important in

accounting for differences in measured productivity across countries.2 We exploit a unique micro-

level data from Malawi to measure total factor productivity (TFP) in farms controlling for land

quality and transitory shocks. We find that operated land size and capital are essentially unrelated

to farm TFP, providing strong evidence of misallocation. Quantitatively, factor misallocation has

a substantial negative effect on agricultural productivity, implying a 3.6-fold gain from efficient

reallocation. This gain is much larger than those reported for the manufacturing sector of China

and India (e.g. Hsieh and Klenow, 2009). Importantly, we provide direct evidence linking the bulk

of agricultural productivity losses from misallocation to restricted land markets.

Malawi represents an interesting case to study for several reasons. First, Malawi is an extremely

poor country in Africa, featuring very low agricultural productivity, a large share of employment in

agriculture, and extremely low farm operational scales. Second, the land market in Malawi is largely

underdeveloped. Most of land in Malawi is customary and user rights are allocated locally by village

chiefs. In our representative sample, more than 83 percent of household farms do not operate any

marketed land (either purchased or rented-in).3 Third, the detailed micro data for Malawi allow

1See, for instance, Gollin et al. (2002) and Restuccia et al. (2008).2For instance Restuccia and Rogerson (2008) and Hsieh and Klenow (2009). See also Gancia and Zilibotti (2009)

for an analysis connecting misallocation and technology.3In Malawi, the vast majority of land is either directly or indirectly distributed by the village head (or superior

chiefs that exercise power over a collection of villages). The Customary Land Act grants these local leaders the powerof allowing/banning land transactions (e.g., inheritance) and resolve disputes across villagers related to land limits(Kishindo, 2011; Morris, 2016). This status quo has remained stable since colonial times (Pachai, 1973). While notyet in place, the Malawi Land Bill passed in 2016 aims at reducing these powers.

2

us to precisely connect land constraints at the household-farm level and factor misallocation in

agriculture, providing an important departure from the existing macro development literature.

Our contribution is based on two elements that are essential for the assessment of the quantitative

importance of factor misallocation and its sources. The first element is a measure of farm-level TFP

which requires detailed information on outputs, inputs, and other factors such as land quality and

transitory shocks that are relevant in our context. The second element is the direct identification

from the data of which production units are operating land acquired from the market or not, which

provides the basis to directly connecting misallocation and land markets. It is important to note that

none of these two key elements, measurements of farm-level TFP and information on how operated

land is acquired, is present in recent studies of misallocation in agriculture, such as Adamopoulos

and Restuccia (2014). The data is the 2010/11 Integrated Survey of Agriculture (ISA) for Malawi

collected by the World Bank. This is a large nationally-representative sample of more than 12

thousand households.4 Whereas the dispersion in our measure of farm-level TFP is quantitatively

similar to previous studies in other sectors and countries, our finding is that factors of production

are roughly evenly spread among farmers. That is, operated land size and capital are essentially

unrelated to farm productivity, generating substantially larger amounts of misallocation than found

in other contexts. To assess the aggregate impact of misallocation on agricultural productivity in

Malawi, we consider as a benchmark the efficient allocation of factors across existing household-

farms in the data, taking as given the total amounts of land and capital; and calculate the output

(productivity) gain as the ratio of efficient to actual output. Our main finding is that the output

gain is 3.6-fold in the full sample, that is, if capital and land were reallocated across farms in Malawi

to their efficient uses, agricultural productivity would increase by 260 percent.

A limitation of the empirical literature on misallocation is the weak link with policies and institu-

tions that cause it (e.g. Hsieh and Klenow, 2009), a caveat that also applies to previous studies of

4The ISA data are considered a gold standard for the study of poor countries; see a detailed analysis in de Mag-alhaes and Santaeulalia-Llopis (2015). We provide further discussion in Section 2.

3

land allocations in the agricultural sector (e.g. Adamopoulos and Restuccia, 2014). This limitation

has inspired a promising quantitative literature studying specific policies and institutions, but the

findings are yet elusive in accounting for the bulk of TFP differences across countries (Restuccia

and Rogerson, 2016). In this context, an important contribution of our analysis is to provide a

strong empirical connection between factor misallocation and the limited market for land, show-

ing large productivity losses associated with restricted land markets. The evidence comes from

contrasting the output gains among farmers that have no marketed land to those that operate

marketed land. We find that reallocating factors among farms with no marketed land to reduce the

dispersion in marginal products to the same extent to those farms with all marketed land increases

agricultural output and productivity by a factor of 2.6-fold. This result provides an important first

step in directly identifying (lack of) land markets as a relevant institution generating misallocation

and productivity losses. Our results have important implications for the design of policies and

institutions to promote better factor allocation.5

Our evidence of factor misallocation based on producer-level TFP measures is closely linked to

the seminal work of Hsieh and Klenow (2009) for the manufacturing sector in China, India, and

the United States. Our analysis contributes to this work by providing evidence of misallocation in

the agricultural sector of a very poor country that is less subject to concerns of measurement and

specification errors that cast doubt on the extent of misallocation in poor countries. The evidence

of misallocation we provide is strong because of several reasons. First, it is easier to measure

producer-level TFP in agriculture than in other sectors since output and most inputs are directly

measured in physical quantities (Beegle et al., 2012; Carletto et al., 2013). Moreover, there is less

product heterogeneity in the agricultural sector than in other sectors, and such heterogeneity can be

controlled for in our data by focusing on individual crops. Indeed, the prevalence of maize production

5We recognize that even with detailed and excellent micro data from ISA Malawi, measuring precisely farm-levelTFP is an enormous difficult task. However, we find it reassuring that the gains from factor reallocation are largeeven when comparing groups of farmers in the same country operating with and without marketed land, as thepotential biases of mis-measurement, unobserved heterogeneity, and randomness are mitigated in the comparison.

4

in Malawi allows us to corroborate the extent of misallocation for farms that produce the same crop.

There is also a high response rate in the dataset with very few missing observations, making the

analysis less subject to measurement and specification errors. Second, the micro data provide precise

measures of the quality of inputs (e.g., eleven dimensions of land quality) as well as transitory shocks

such as rain and health shocks that we explicitly control for to measure farm-level TFP. Third, the

large sample size allows us to assess the robustness of our results to within narrow geographic areas

and other relevant characteristics. Fourth, the micro data provide information on land and capital

rental payments that we use to compute land and capital income shares for the agricultural sector

in Malawi that generate virtually identical results to our baseline calibration to U.S. shares. Fifth,

while Hsieh and Klenow (2009) assess the reallocation gains in China and India relative to the

gains in the United States, we are able assess the reallocation gains relative to a set of farmers in

Malawi whose operated land is all marketed, providing a more direct benchmark comparison. Sixth,

the link between the degree of misallocation in manufacturing China and India and its causes is

weak in Hsieh and Klenow (2009). More recently, for the case of China, Song et al. (2011) show

that credit market imperfections are important in explaining resource misallocation between private

and state-owned enterprises. Analogously, we provide a direct link between misallocation and land

markets in agriculture exploiting information on how each plot was acquired. For all these reasons,

we argue that our analysis constitutes the most direct and comprehensive evidence of misallocation

in a poor country.

We also explore the broader consequences of misallocation on structural transformation and in-

equality. The productivity increase from efficient reallocation in Malawi would unravel a substantial

process of structural change, for instance, in the context of a standard two-sector model, a reduction

in the share of employment in agriculture from 65 percent to 4 percent and an increase in average

farm size by a factor of 16.2-fold, making these statistics closer to industrialized-country levels.

These effects would be even larger in more elaborate models that include endogenous investments

5

in productivity, ability selection across sectors, among other well-studied channels in the literature.6

A potential argument that justifies the actual land allocation is the reduction of income inequality.

That is, the distribution of land can operate as an ex-ante redistribution mechanism. Interestingly,

even though the actual allocation of factors is evenly spread across farmers, we find that factor

equalization is ineffective at equalizing incomes in Malawi. Taking the actual allocation of factors

as endowments and decentralizing the efficient allocation via perfectly competitive rental markets,

we show that the income distribution associated with the efficient allocation features much lower

income inequality and poverty. The introduction of rental markets, where operational scales can

deviate from land-use rights, substantially improves aggregate productivity reducing poverty and

alleviating income inequality.

Our paper relates to a growing literature in macroeconomics using micro data to study macro de-

velopment such as Gollin et al. (2014), Buera et al. (2014), Adamopoulos and Restuccia (2015),

Bick et al. (2016), and Santaeulalia-Llopis and Zheng (2016). More closely related to our work is

the macroeconomic study of misallocation in the agricultural sector based on the size distribution

of farms across countries (Adamopoulos and Restuccia, 2014).7 Our first advantage over previous

macroeconomic studies on agriculture is that we are able to recover TFP measures at the farm level

and, hence, assess the quantitative impact of factor misallocation in the agricultural sector in a

specific country, as in Hsieh and Klenow (2009) for manufacturing China and India. Importantly,

our results imply larger productivity losses from factor misallocation in agriculture than in manu-

facturing in poor countries. Our second advantage is that our analysis is able to directly connect

misallocation to restrictions in land markets, providing an important direct link to policy analysis

that is missing in Hsieh and Klenow (2009) and Adamopoulos and Restuccia (2014). Within the

micro development literature, Foster and Rosenzweig (2011) provide evidence of inefficient farm

6See for instance Goldstein and Udry (2008), Lagakos and Waugh (2013), and Adamopoulos et al. (2017).7Recent work in this area include: Adamopoulos and Restuccia (2015) who study the impact of a specific policy—

land reform—on productivity exploiting panel micro data from the Philippines, Chen (2016) who studies the impactof land titles on agricultural productivity across countries, and Gottlieb et al. (2015) who study the implications ofcommunal land on misallocation and productivity in agriculture.

6

sizes in India that may be connected to the low incidence of tenancy and land sales limiting land

reallocation to efficient farmers. Udry (1996) focuses on the intra household-farm reallocation of

factors across wives and husbands for a relatively small sample of farms, obtaining a small role of

misallocation in Burkina Faso. Our results indicate a large role for misallocation. There is a crucial

difference between our work and that of Udry’s. We focus on reallocations across farms instead

of within farms. By exploiting large representative data to study nationwide factor reallocation

across farms, our analysis delivers, arguably, a more accurate picture of the macroeconomic gains

from reallocation.8 Finally, Midrigan and Xu (2014) find that firms’ internal capital accumula-

tion substantially mitigates credit market imperfections after entry. It is important to emphasize

that this channel, while potentially important for other factors, is limited for land which is not

reproducible and accumulates only through transactions. This is consistent with recent empirical

evidence showing that household land holdings barely grow over the lifecycle in Sub-Saharan Africa

(de Magalhaes and Santaeulalia-Llopis, 2015).

The paper proceeds as follows. In the next section and section 3, we describe important elements

of the micro data and construct our measure of farm TFP in Malawi. Section 4 assesses the

aggregate productivity impact of misallocation, linking it with the extent of access to marketed

land. Sections 5 and 6 discuss the broader implications of misallocation on structural change and

inequality. We conclude in Section 7. An on-line Appendix is available at: https://www.dropbox.

com/s/je5wv90trv2nmq3/RS_Online_Appendix.pdf?dl=0.

8There is an important literature in micro development analyzing the role of tenancy and property rights foragricultural productivity such as Shaban (1987), Besley (1995), Banerjee et al. (2002), among others. Whereasthis literature has focused on the impact of endogenous investments and effort incentives on farm productivity, ouranalysis focuses on the impact of resource allocation on aggregate agricultural productivity, taking as given farm-level TFP. Integrating the role of land-market institutions on both factor misallocation and endogenous farm-levelproductivity, is an important and promising area of study that we leave for future research.

7

2 Data

We use a new and unique household-level data set collected by the World Bank, the Malawi In-

tegrated Survey of Agriculture (ISA) 2010/11, see de Magalhaes and Santaeulalia-Llopis (2015).

The survey is comprehensive in the collection of the entire agricultural production (i.e., physical

amounts by crop and plot) and the full set of inputs used in all agricultural activities at the plot

level, all collected by a new and enlarged agricultural module that distinguish ISA from previous

Living Standards Measurement Study (LSMS) surveys. The data are nationally-representative with

a sampling frame based on the Census and an original sample that includes 12,271 households (and

56,397 individuals) of which 81% live in rural areas.9

The survey provides information on household-farm characteristics over the entire year and we focus

our attention to agricultural activities related to the rainy season. The detail on household-farm

agricultural production is excruciating. Information on agricultural production is provided by each

and all crops produced by the household. This is an economy largely based on maize production that

uses 80% of the total land. The total quantity of each crop harvested by each household is available

per plot. We value agricultural production using median at-the-gate prices. In crop production, each

household potentially uses different quantities of intermediate inputs such as fertilizers, herbicides,

pesticides and seeds. This information is also provided by plot. We also apply common median

prices to value these intermediate inputs. As a result, our benchmark measure of household-farm

output is a common-price measure of real value added constructed as the value of agricultural

production (of all crops) minus the costs of the full set of intermediate inputs.10

9The Malawi ISA is part of a new initiative funded by the Bill & Melinda Gates Foundation (BMGF) and ledby the Living Standards Measurement Study (LSMS) Team in the Development Research Group (DECRG) of theWorld Bank. For further details on the Malawi ISA data and the construction of agricultural output and inputs seeAppendix A.

10We use common prices of outputs and intermediate inputs to construct real value added in farms in the samespirit as the real measures of output across countries from the International Comparison Project and the Penn WorldTable.

8

We measure household land as the sum of the size of each cultivated household plot. This includes

rented-in land, which consists of 12.5% of all cultivated land. On average, household farms cultivate

1.8 plots.11 Plot size is recorded in acres using GPS (with precision of 1% of an acre) for 98% of plots

(for the remaining 2% of plots, size is estimated). The operational scale of farms is extremely small.

For each household, we compute the amount of land used for agricultural production regardless of

the land status (whether land is owned, rented, etc.). Hence, we focus on the operational scale

of the household-farm. We find that 78.3% of households operate less than 1 hectare (henceforth,

Ha.), 96.1% of households operate less than 2 Ha., and only 0.3% of households operate more than

5 Ha., see the first column in Table 1. The average farm size is 0.83 Ha.12 The data contains

very detailed information on the quality of land for each plot used in every household. There are

11 dimensions of land quality reported: elevation, slope, erosion, soil quality, nutrient availability,

nutrient retention capacity, rooting conditions, oxygen availability to roots, excess salts, topicality,

and workability. This allows us to control for land quality to measure household-farm productivity.

Regarding land, we emphasize that the land market is largely underdeveloped in Malawi. The pro-

portion of household-farms that do not operate any marketed land is 83.4%. These are households

whose land was granted by a village chief, was inherited or was given as bride price. The remaining

16.6% of farm households operate some land obtained from the market, either rented or purchased,

and the proportion of household-farms whose entire operated land was obtained in the market is

10.4%. Disaggregating the main types of marketed land, we find that 3.0% of household-farms rent-

in land informally (e.g., land borrowed for free or moved in without permission), 9.5% rent-in land

11In rural Malawi, land is the largest household asset representing 44% of household total wealth. House structureis 30%, livestock is 13%, and agricultural equipment and structures (e.g. tools and barns) is 3%, see de Magalhaesand Santaeulalia-Llopis (2015).

12To make a comparison of operational scales in Malawi with other countries, we report the distribution of farmsizes from the World Census of Agriculture in Adamopoulos and Restuccia (2014). These data include all land usedfor agricultural production and corresponds to the year 1990. Note that despite the 20 years difference between ISA2010/11 and the Census 1990 the distribution of land across the two sources is very similar. Comparatively withother countries, the operational scale of farms in Malawi is extremely small, ∼0.70-0.83 Ha., whereas average farmsize is 187 Ha. in the United States and 16.1 Ha. in Belgium. Belgium is a good developed-country reference sincethe land endowment (measured as land per capita) is similar to that of Malawi (land per capita is 0.56 Ha. in Malawiand 0.5 Ha. in Belgium, whereas land per capita is 1.51 Ha. in the United States).

9

Table 1: Size Distribution of Farms (% of Farms by Size)

ISA 2010/11 World Census of Agriculture 1990Malawi Malawi Belgium USA

Hectares (Ha):≤ 1 Ha 78.3 77.7 14.6 –1− 2 Ha 17.8 17.3 8.5 –2− 5 Ha 3.7 5.0 15.5 10.65− 10 Ha 0.2 0.0 14.8 7.510+ Ha 0.0 0.0 46.6 81.9

Average Farm Size (Ha) 0.83 0.7 16.1 187.0

Notes: The first column reports the land size distribution (in hectares) for household farms from the Malawi 2010/11Integrated Survey of Agriculture (ISA). The other columns report statistics from the World Census of Agriculture1990 for Malawi, Belgium, and United States documented in Adamopoulos and Restuccia (2014).

formally (e.g., leaseholds, short-term rentals or farming as a tenant), 1.8% purchase land without a

title and 1.3% purchase land with a title.

In terms of agricultural capital, we have information on both equipment and structures. Capital

equipment includes implements (such as hand hoe, slasher, axe, sprayer, panga knife, sickle, treadle

pump, and watering can) and machinery (such as ox cart, ox plough, tractor, tractor plough, ridger,

cultivator, generator, motorized pump, and grain mill), while capital structures includes chicken

houses, livestock kraals, poultry kraals, storage houses, granaries, barns, pig sties, among others.

To proxy for capital services after conditioning for its use in agricultural activities, we aggregate

across the capital items evaluated at the estimated current selling price.13

In Malawi, a large proportion of the households members, beside the household head, contribute

to agricultural work. Household size is 4.6 with extended families in which several generations live

together in a single household. We use the survey definition of household members as individuals

that have lived in the household at least 9 months in the last 12 months. In terms hours, data are

13This selling price for agricultural capital items, rarely available in previous LSMS data, helps capture potentialdifferences in the quality and depreciation of capital across farms. We provide robustness of our results to alternativephysical measures of agricultural capital such as asset indexes for agricultural equipment and structures in AppendixB. Our results are robust to using these alternative indexes to proxy for agricultural capital.

10

collected at the plot level for each individual that participates in agriculture and by agricultural

activity (i.e., land preparation/planting, weeding/fertilizing, and harvesting). The data provides

information of weeks, days per week, and hours per day employed per plot, activity and individual.

To compute household-farm hours we aggregate the hours of all plots, activities and individuals.

Further, the same information is provided for hired labor and labor received in exchange (for free),

but most household-farm hours consist of family hours. We add hours by hired labor and free

exchange of labor to our measure of total household hours.

Since our data comprises a single cross section of households for 2010-11, it is important to control

for temporary output shocks that may explain variation in output and hence productivity across

households in the data. The single most important temporary shock for farmers is weather. We

use the annual precipitation which is total rainfall in millimetres (mm) in the last 12 months. In

further robustness exercises, we net household-productivity from additional transitory shocks in the

form of health, deaths or food security risks suffered by the household in the last 12 months, and

we also exploit the more recent 2013 round data where about 1/4 of the original households are

interviewed to calculate a more permanent component of farm TFP as an estimated fixed effect in

the panel.

Geographic and institutional characteristics are also recorded for each household-farm. We use

several partitions of these characteristics to conduct robustness exercises. In particular, we use ge-

ographical information on the region, districts, and enumeration area to which household-farms be-

long and institutional characteristics such as the traditional authority (TA) governing the household-

farm or ethno-linguistic characteristics. TAs are relevant for our exercise as chiefs appointed by

TAs perform a variety of functions that include resolving issues related to land and property.14

Finally, we note that the survey response is very high with very few missing observations. Condi-

14Traditional Authorities are part of the administrative structure of Malawi. Any one district is only in one region,any one TA is only in one district. Traditional authorities (and their sub-ordinate chiefs) are ’town chiefs’ formallyrecognized under the Chiefs Act of 1967 and receive a monetary honorarium by the government.

11

tioning on households that produce agricultural output and for which all factor inputs, including the

11 dimensions on land quality, are available and further trimming about 1% of the household-farm

productivity distribution, our sample consists of 7,157 households.15

The data allows us to construct fairly precise measures of real household-farm productivity. The

ISA data represents a substantial improvement with respect to previous LSMS questionnaires. The

detailed information on quantity inputs and outputs reduces substantially the possibility of measure-

ment error and composition bias, making the dataset ideal for our purpose of measuring productivity

at the farm level, assessing the extent of factor misallocation in the Malawian economy, and as-

sessing the extent to which productivity losses due to misallocation are related to imperfections or

frictions in the land and other input markets.

3 Measuring Farm Productivity

We use the micro data from ISA 2010/11 described in the previous section to measure productiv-

ity at the farm level. Constructing a measure of farm total factor productivity (TFP) is essential

in assessing the extent to which factors are misallocated in the agricultural sector. The detailed

micro data for Malawi presents a unique opportunity to assess factor misallocation in agriculture

and its aggregate productivity implications. Fewer assumptions are required to measure micro-level

productivity in our context compared to many studies of misallocation focusing on the manufactur-

ing sector since our data for agriculture comprises physical quantities of outputs and most inputs

whereas additional assumptions are required to separate price and quantity from sales or revenue

data of manufacturing plants.

We measure farm productivity by exploiting the detailed micro data where not only we obtain real

15See further details on the trimming strategy in Appendix B.

12

measures of output and value added in each farm but also control for land quality and a wide array

of transitory shocks. We measure farm-level total factor productivity (TFP) si as the residual from

the following farm-level production function,

yi = siζikθki (qili)

θl , θk + θl < 1, (1)

where yi is real value added, ki is capital, li is the amount of land in operation, ζi is a rain shock,

qi is land quality, and θk,l are the input elasticities. In our analysis, we focus on the allocation

of capital and land across farms, abstracting from differences in labor inputs. For this reason, we

measure value added, capital, and land in the data in per total labor hour terms. This implies that

our residual measure of TFP is not affected by actual differences in the labor input across farmers

in the data.16 In constructing our measure of farm TFP, we choose θk = 0.36 and θl = 0.18 from the

capital and land income shares in U.S. agriculture reported in Valentinyi and Herrendorf (2008).

We later discuss the sensitivity of our results to these factor shares using our micro data for Malawi.

Our measure of output is real value added which takes into account the real amount of intermediate

inputs used in production. This is relevant in Malawi because intermediate inputs are subsidized via

the “Malawi Input Subsidy Program” and the subsidy allocation is based on farmer’s income, with

poorer farmers receiving higher subsidies. According to our preferred measure of productivity, less

productive farmers, which are also poorer farmers, receive a subsidy that is more than 60 percent

their output, whereas more productive farmers (richer farmers) receive a subsidy that is less than 10

percent their output. This implies that a measure of productivity that uses actual expenditures in

intermediate inputs instead of actual quantities of intermediate inputs substantially underestimates

productivity dispersion across farms. In our data, not accounting for distortions in intermediate

input prices underestimates the dispersion in farm TFP by 23 percent.17

16We note, however, that labor hours are misallocated across farms since farm hours are uncorrelated with farmTFP. As a result, our characterization of misallocation is conservative in terms of the potential gains from efficientreallocation.

17Misallocation of intermediate inputs via the input subsidy program is distinct from the misallocation due to

13

We are interested in permanent measures of productivity at the farm level, associated with the

productivity of the farm operator. As a result, two key elements for this measure are to distinguish

between the productivity of the farmer and the productivity of the land under operation and to

abstract from temporary variations in output due to whether shocks. We deal with each of these

components in turn.

There are important differences in the land characteristics operated by farms in our sample. For

the full sample, more than 34 percent of land is high-altitude plains while around 20 percent are

low plateaus and 19 percent mid-altitude plateaus. These characteristics also differ by region where

the Center region is mostly high-altitude plains whereas the South region is mostly low plateaus.

We control for land quality in our measure of productivity by constructing an index as follows. Our

benchmark land quality index q0i is constructed by regressing log output in each farm on the full set of

land quality dimensions described in Section 2. To the extent that farm inputs are positively related

to land quality, our benchmark index attributes too much variation to land quality. We nevertheless

use this index as our benchmark because it is conservative with respect to the productivity effects

of misallocation.18 There is substantial variation in land quality q across households and across

regions, see for instance Figure 1, panel (a); but this variation is not quantitatively substantial

relative to for example variation in TFP across farms. In Table 2 we report the dispersion (variance

of the log) of land quality indexes versus the dispersion in the quantity of land across households

in our data. We find that the dispersion in land quality is large and slightly above that of land

size nationwide, respectively .859 and .749 in terms of the variance of the log. Not surprisingly, the

dispersion in land quality decreases with the size of geographic area: the average dispersion is .833

within regions, .568 within districts, and .147 within enumeration areas.

imperfect land markets we emphasize and as a result we focus on value added measures of output and abstract fromintermediate-input distortions. Nevertheless, it is important to emphasize that intermediate input distortions areanother source of productivity losses in the agricultural sector in Malawi.

18We have considered many alternative land quality indexes with similar results including imposing exogenousfunctional forms, different subsets of land quality dimensions, and different forms of land size and capital controls,see Appendix A.

14

Figure 1: Histograms of Land Quality Index and Rain, Malawi ISA 2010/11

(a) Land Quality Index, qi (in logs)

(b) Rain, ζi (in logs)

Notes: The land quality index is constructed by regressing log output in each farm on the full set of land qualitydimmensions. Median values by region are depicted with vertical lines: Orange (North), green (Center) and lightblue (South).

15

Table 2: Dispersion of Output, Land Size, Land Quality, and Rain, Malawi ISA 2010/11

Full Within Geographic Areas:Sample Regions Districts Enumeration

Areas

Output (yi) 1.896 1.867 1.778 1.649Land size (li) .749 .746 .719 .671Land quality:

Index (qi) .852 .833 .568 .147Index subitems:

Elevation .439 .349 .075 .001Slope (%) .657 .635 .453 .093Erosion .480 .496 .472 .427Soil quality .608 .605 .514 .458Nutrient availability .387 .402 .248 .023Nutrient retention capability .329 .365 .180 .016Rooting conditions .342 .372 .302 .029Oxygen available to roots .097 .094 .105 .010Excess salts .037 .045 .048 .004Toxicity .027 .038 .033 .003Workability .475 .474 .335 .033

Quality-adjusted land size (qili) 1.571 1.531 1.243 .808Rain (ζi)

Annual precipitation (mm) .039 .025 .014 .001Precipitation of wettest quarter (mm) .026 .013 .005 .000

Notes: Output yi, land size (in acres) li, land quality index qi, quality-adjusted land size qili, and rain ζi, arecontinuous variables for which we use the variance of logs as a measure of dispersion. Subitems of land quality, slope(in %) and elevation (in meters), are also continuous variables and we also report the variance of logs as dispersionmeasure. The other subitems of land quality are categorical variables such as soil quality (with categories 1 good, 2fair, 3 poor), erosion (with categories 1 none, 2 low, 3 moderate, 4 high) and nutrient availability, nutrient retention,rooting conditions, oxygen to roots, excess of salts, toxicity and workability that take values from four categories (1low constraint, 2 moderate constraint, 3 severe constraint and 4 very severe constraint). For all categorial variableswe use the proportion of non-mode values in the sample as measure of dispersion. The construction of the landquality index qi is discussed in Section 3. For the last three columns referring to within geographic areas, we reportthe averages across geographic area under consideration (e.g., the variance of output in the ’Regions’ column is theaverage across regions of the variance of output by region).

16

In Malawi, weather shocks, in particular rain, is the single most important source of transitory

shocks to agricultural production—we find that most cropland is rain-fed and 84% of household-

farms do not have any alternative irrigation system. Figure 1, panel (b), reports the density of

rain for the entire population of farmers and, separately, for farmers within regions. We note what

appear substantial differences across household-farms and regions with median values for the Center

and South region that fall below the range of of the distribution of rain in the North region. To

quantify the actual dispersion in rain we report the variance of logs of rain in the bottom of Table 2

for our entire sample and regions. We find that the dispersion of rain is very small compared to

the dispersion in land size and even smaller relative to output. In particular, the dispersion in

annual precipitation is 5.2% that of land size and 2.1% that of agricultural output. The small role

of rain variation further appears within regions, districts, and enumeration areas, as well as using

alternative definitions of precipitation measures, with mean cross-sectional variances that decline

with the size of the geographical area. In short, rain is not a substantial contributor to output

variation across farm households in Malawi.

Having constructed our measure of farm TFP, Figure 2 documents its distribution across all house-

holds in our sample. There are large differences in farm productivity across households that remain

even within regions. The dispersion in farm productivity compares in magnitude to that of the

variation in physical productivity (TFPQ) across plants in the manufacturing sector in the United

States, China, and India reported in Hsieh and Klenow (2009), see Table 3. Whereas the ratio of

physical productivity between the 90 to 10 percentile is a factor of around 15-fold in China and

India and around 9-fold in U.S. manufacturing, the 90-10 ratio across farms in our sample is 10.8-

fold. Similarly, the 75-25 ratio is 3.2-fold in our sample of farms whereas it is around 4.5-fold across

manufacturing plants in China and India.

Table 4 reports the variance decomposition of farm output per hour using our assumed production

function. We note that farm productivity si and to a lesser extent the inputs of capital and land

17

Figure 2: Density of Farm Productivity si (in logs), Malawi ISA 2010/11

Notes: Household-farm productivity si is measured using our benchmark production function, adjusting for rain ζiand land quality qi , yi = siζif(ki, qili) with f(ki, qili) = k.36i (qili)

.18, where yi is farm output, ki is capital, and li island. All variables have been logged.

are the key determinants of output variation across farm households in Malawi, with rain and land

quality playing a quantitatively minor role. For instance, the variance of farm TFP accounts for 76

percent of the total variance of output and the variance of capital and quality adjusted land inputs

for about 20 percent, whereas when rain and land quality are assumed constant across farms, the

variances of these two factors and their contribution to the total variance are fairly similar. The

small quantitative role of land quality for output differences across household farms is robust to

variations in the index which we discuss in Appendix A.

18

Table 3: Dispersion of Productivity across Farms and Manufacturing Plants

Farms Manufacturing PlantsMalawi USA USA China India

Statistic ISA 2010/11 1990 1977 1998 1987

SD 1.19 0.80 0.85 1.06 1.1675-25 1.15 1.97 1.22 1.41 1.5590-10 2.38 2.50 2.22 2.72 2.77N 7,157 – 164,971 95,980 31,602

Notes: The first column reports statistics for the household-farm productivity distribution from the micro datain Malawi. The second column reports statistics for farm productivity in the United Sates from the calibrateddistribution in Adamopoulos and Restuccia (2014) to U.S. farm-size data. The other columns report statistics formanufacturing plants in Hsieh and Klenow (2009). SD is the standard deviation of log productivity; 75-25 is thelog difference between the 75 and 25 percentile and 90-10 the 90 to 10 percentile difference in productivity. N is thenumber of observations in each dataset.

Table 4: Variance Decomposition of Agricultural Output, Malawi ISA 2010/11

Benchmark (ζi = 1, qi = 1)Level % Level %

var(y) 1.896 100.0 1.896 100.0var(s) 1.435 75.7 1.457 76.8var(ζ) .039 2.1 – –var(f(k, ql)) .383 20.2 .343 18.12cov(s, ζ) -.044 -2.3 – –2cov(s, f(k, ql)) .034 1.8 .096 5.12cov(ζ, f(k, ql)) .048 2.5 – –

Notes: The variance decomposition uses our benchmark production function that adjusts for rain ζi and land qualityqi across household farms, yi = siζif(ki, qili) with f(ki, qili) = k.36i (qili)

.18. All variables have been logged. Thevariables are output yi, household-farm productivity si, rain ζi, structures and equipment capital ki, and quality-adjusted land size, qili. The first two columns report results from our benchmark specification where rain and landquality are controlled for. The last two columns report the results abstracting from rain and land quality, i.e. weset ζi = 1 and qi = 1 ∀i. In each case, the column “Level” reports the variance and the column “%” reports thevariance contribution to the total in percentage.

19

4 Quantitative Analysis

We assess the extent of factor misallocation across farms in Malawi and its quantitative impact

on agricultural productivity. We do so without imposing any additional structure other than the

farm-level production function assumed in the construction of our measure of household-farm pro-

ductivity. We then study the connection between factor misallocation and land markets.

4.1 Efficient and Actual Allocations

As a benchmark reference, we characterize the efficient allocation of capital and land across a fixed

set of heterogeneous farmers that differ in productivity si. A planner chooses the allocation of

capital and land across a given set of farmers with productivity si to maximize agricultural output

given fixed total amounts of capital K and land L. The planner solves the following problem:

Y e = max{ki,li}

∑i

si(kαi l

1−αi )γ,

subject to

K =∑i

ki, L =∑i

li.

The efficient allocation equates marginal products of capital and land across farms and has a simple

form. Letting zi ≡ s1/(1−γ)i , the efficient allocations are given by simple shares of a measure of

productivity (zi/∑zi) of capital and land:

kei =zi∑ziK, lei =

zi∑ziL.

We note for further reference that substituting the efficient allocation of capital an land into the

definition of aggregate agricultural output renders a simple constant returns to scale aggregate

20

production function for agriculture on capital, land, and agricultural farmers given by

Y e = ZN1−γa [KαL1−α]γ, (2)

where Z = (∑zig(zi))

1−γ is total factor productivity as the average productivity of farmers and

Na is the number of farms. For our benchmark, we choose α and γ consistent with the production

function used in our measure of farm-level productivity so that αγ = 0.36 and (1− α)γ = 0.18 are

the capital and land income shares in the U.S. economy (Valentinyi and Herrendorf, 2008) — and we

do sensitivity with factor shares directly computed from our Malawi micro data (see section 4.4.2).

This implies γ = 0.54. The total amount of capital K and land L are the total amounts of capital

and land across the farmers in the data. Farm level productivity si’s are given by our measure of

farm TFP from data as described previously.

We illustrate the extent of factor misallocation in Figure 3, where we contrast the actual allocation of

land and capital and the associated factor productivities by farm TFP with the efficient allocation

of factors and factor productivities across farms. In the figure, each dot represents a household

farm in the data whereas the line represents the variable in the efficient allocation. In the efficient

allocation, operational scales of land and capital are increasing in farm productivity so that factor

productivities are constant across farms. The patterns that emerge in comparing the efficient

allocation with actual allocations are striking as the efficient allocation contrasts sharply with the

actual allocation of capital and land in Malawi. The data show that operational scales of land and

capital in farms are unrelated to farm productivity. Figure 3, panel (a), shows the amount of land

operated by each farm against farm productivity using our baseline measure of productivity that

adjusts for rain and quality as described previously. Contrary to the efficient allocation where there

is a tight mapping between land size and TFP, actual land in farms is essentially uncorrelated to

farm TFP, the correlation between land size and productivity is .05. This pattern implies that

the average (or marginal) product of land is not equalized across farms as it would be the case in

21

the efficient allocation. Figure 3, panel (b) documents the marginal product of land (which in our

framework is proportional to the output per unit of land or yield) across farms, which is strongly

positively related to farm TFP with a correlation (in logs) of .77.

Figure 3, panel (c), documents the relationship between the amount of capital in farms by farm

productivity. As with land, capital and productivity across farms are essentially unrelated, the

correlation between the two variables is -.01. And this pattern implies an increasing marginal

product of capital with farm TFP which we document in panel (d) with a correlation between

these variables (in logs) of .76. Although not documented in the figure, the data indicates that the

capital to land ratio is essentially unrelated to farm productivity, with a correlation between these

variables of -.03. This finding suggests that larger farms use more capital but since larger farms are

not more productive on average, the capital to land ratio remains roughly constant with respect to

farm TFP.

That the actual allocation of land across farmers in Malawi is unrelated to farm productivity is

consistent with our characterization of the land market where the amount of land in farms is more

closely related to inheritance norms and redistribution, and access to land is severely restricted in

rental and sale markets so more productive farmers cannot grow their size. Capital is also unrelated

to farm productivity with the capital to land ratio being roughly constant across farm productivity.

Our interpretation of this fact is that restrictions to land markets are also affecting capital allo-

cations echoing de Soto (2000) findings that land market restrictions and insecure property rights

of farmers limit the ability to raise capital for agricultural production.19 Our findings constitute

strong evidence of land and capital misallocation across farmers in Malawi.

19See the related discussion in Besley and Ghatak (2010). We find further direct evidence of the relationshipbetween land markets and access to credit in our micro data for Malawi, see Appendix C for a discussion.

22

Figure 3: Land Size, Capital, MPL and MPK: Actual and Efficient Allocations

(a) Land Size vs. Farm Productivity (b) MPL vs. Farm Productivity

(c) Capital vs. Farm Productivity (d) MPK vs. Farm Productivity

Notes: Panel (a) reports actual and efficient land size in farms li with respect to farm productivity si. Panel (b)reports actual and efficient marginal product of land (MPL) with respect to farm productivity si. Panel (c) reportsactual and efficient capital in farms ki with respect to farm productivity si. Panel (d) reports actual and efficientmarginal product of capital (MPK) with respect to farm productivity si. All variables have been logged.

23

4.2 Aggregate Output Gain

To assess the impact of misallocation on aggregate productivity, in what follows we report the

aggregate output gain defined as the ratio of efficient to actual aggregate output,

Y e

Y a=

Y e∑yai,

where Y e is efficient aggregate agricultural output as defined previously and Y a is actual agricultural

output aggregated from farm-level output yi using the production function in equation (1) with our

measure of farm productivity and the actual amounts of capital and land operated by each farm.20

Because the efficient output takes as given the total amounts of capital, land, and the number of

farmers observed in the data, the output gain is also a TFP gain.

Table 5, panel (a), reports the main results. For the full sample in the entire Malawi, the output gain

is 3.6-fold, that is, if capital and land were reallocated efficiently in Malawi to maximize agricultural

output, output and hence productivity would increase by 260 percent. This is a very large increase

in productivity as a result of a reduction in misallocation compared to the results when evaluating

specific policies which have found increases on the order of 5 to 30 percent or even when eliminating

all wedges in manufacturing in China and India in Hsieh and Klenow (2009) with increases ranging

between 100 to 160 percent. Given that the productivity dispersion across farms in Malawi is similar

to that of manufacturing plants reported in Hsieh and Klenow (2009), a larger reallocation gain

suggests that resources in Malawi are more misallocated than in those other countries. Indeed, the

standard deviation of log revenue productivity in Malawi is 1.05 compared to 0.67-0.74 in the case

of India and China in Hsieh and Klenow (2009). Because we have a large sample, our mean output

gain is tightly estimated. In Table 5, panel (a), we also report the 5 and 95 percentiles of Bootstrap

20The computation of actual output at the farm level abstracts from rain and land quality and hence is comparableto the efficient output described in the previous subsection. Our measure of farm productivity si though is purgedof rain and land quality effects.

24

estimates, which provide a narrow interval, between 3.1 and 4.2-fold output gain.

Table 5: Agricultural Output Gain (Y e/Y a)

(a) Main Results

Bootstrap SimulationsFull Sample Median 5th pct. 95th pct.

Nationwide 3.59 3.55 4.07 3.11

(b) Within Productivity-si Variation

Benchmark 95% 90% 80% 0%Output Gain 3.59 3.40 3.25 3.00 2.48

(c) By Geographical Areas and Institutions

Average Median Max MinGeographic Areas:

Regions 3.40 3.42 3.63 2.87Districts 2.93 2.80 7.10 2.07Enum. Areas 1.65 1.60 11.00 1.04

Institutions:Traditional Authority 3.11 3.48 5.58 1.39Language 3.36 3.88 5.95 1.16

Notes: In panel (a), bootstrap median and confidence intervals are computed from 5,000 simulations obtained fromrandom draws with 100 percent replacement, i.e., each simulation consists of a sample of the same size as the originalsample. See further discussion in Appendix B. Panel (b), see text in Section 4 for details. Panel (c) reports theratio of efficient to actual output when reallocation occurs within three narrower definitions of geographical areas(3 regions, 31 districts and 713 enumeration areas) and two measures of institutional settings/cultural identity (53traditional authorities and 13 languages). We drop enumeration areas with less than 5 household-farm observations.

Factor inputs are severely misallocated in Malawi, implying the large aggregate productivity gains

just discussed. There are two features of factor misallocation in Figure 3, panels (a) and (c): factor

inputs are dispersed among farmers with similar productivity (misallocation in factor inputs within

si productivity types) and factor inputs are misallocated across farmers with different productivity

(which lowers the correlation of factor inputs with farm productivity). We argue that factor input

variation within a farm-productivity type is not due to measurement error as factor inputs such as

land are measured with tight precision. Nevertheless, we can assess the magnitude of the aggregate

25

output gain associated with lower dispersion in factor inputs within a productivity type. We remove

within-s variation in factor inputs by regressing separately log land and capital on a constant and

log farm productivity s and use the estimated relationship to construct measures of factor inputs

that partially or fully remove residual variation. The results are in Table 5, panel (b). At the

extreme, with no within-s variation of factor inputs, there is only gain from reallocation across

productivity types. Even in this case, the output gain is still substantial a 2.5-fold (versus 3.6-fold

in the baseline). This implies that 70 percent of the output gain is due to misallocation of factor

inputs across farmers with different productivity.

We also report the output gain for narrower definitions of geographical areas and institutions such

as language classifications. Table 5, panel (c), reports these results. In the first column, we report

the output gain for the average of regions (i.e., output gain of reallocating factors within a region

for each region and then averaged across regions), for the average across districts (a narrower

geographical definition than region), and for the average across enumeration areas. Enumeration

areas are a survey-specific geographical description and is the narrowest geographical definition

available. Each enumeration area amounts to about 16 households in the data, which given the

average land holdings amounts to about 30 acres of geographically connected land. The table also

reports results by Traditional Authority (TA) and language. Columns 2 to 4 report the Median,

Min, and Max in each case.

What is striking about the numbers in Table 5, panel (c), is that the gains from reallocation

are large in all cases, even within narrowly defined geographical areas, traditional authority and

language. In particular, reallocating capital and land within a region generates output gains of the

same magnitude as in the aggregate (mean output gain of 3.4-fold).21 Similarly, within districts,

the average output gain is 2.9-fold, but the output gain can be as large as 7.10-fold and as low

as 2.1-fold. Even for the narrowest geographical definition, enumeration areas, just reallocating

21The idea is that narrower geographic definitions provide additional land quality controls, see Larson et al. (2014)for a similar strategy.

26

factors within 16 households, average output gains are 1.7-fold and can be as large as 11-fold in

some areas. The output gain within traditional authorities is 3.1-fold in average, with the largest

5.6-fold and median 3.5-fold. Within languages, the average output gain is 3.4-fold, the largest

6-fold and median 3.9-fold.

4.3 The Role of Land Markets

We have found strong evidence that capital and land are severely misallocated in the agricultural

sector in Malawi. We connect factor misallocation directly to the limited market for land in Malawi.

To do so, we use plot-level information about how each plot was acquired and group household-

farms by the share of marketed land that they operate. We find that 83.4 percent of all household

farms operate only non-marketed land and the remaining 16.6 percent operate some marketed land,

with 10.4 percent operating exclusively marketed land. Using this classification of household farms,

we explicitly assess the output gain across farms that differ in the extent of marketed land in their

operation.

Table 6 reports the aggregate output gain for farms that operate with no marketed land, with some

marketed land as either rented in or purchased, and with only marketed land. The output gain is

much higher for those farms with no marketed land than with some marketed land. For instance,

the output gain for those farms with no marketed land, which comprises 83.4 percent of the sample

of farms, is 4.2-fold, slightly larger than the 3.6-fold output gain for the entire sample. But for

farms with some marketed land, which comprises the remaining 16.6 percent of the sample farms,

the output gain is 2-fold. The output gain of factor misallocation is reduced by more than half when

farms have access to some marketed land. The output gain is even smaller among the set of farmers

whose entire operated land is either rented in or purchased. These farmers comprise 10.4 percent

of the sample and have an output gain of 1.6-fold versus 4.2-fold for farms with no marketed land.

27

These results provide direct and solid evidence of the connection of misallocation to land markets.22

Table 6 also decomposes the output gains among farmers by the type of operated marketed land:

(a) rented-in informally, for example land borrowed for free or moved-in without permission; (b)

rented-in formally such as leaseholds, short-term rentals or farming as tenant; (c) purchased as

untitled; and (d) purchased as titled. The vast majority of farmers with some marketed land are

formally renting-in, 9.5 percent of the sample out of the 16.6 percent with some marketed land. The

output gain is fairly similar for farmers operating land rented-in formally or informally, an output

gain of 1.72-1.73-fold. The lowest output gain is recorded for farms with operated land that was

purchased with a title, 1.39-fold versus 4.15-fold for farms with no marketed land. Interestingly,

farmers with purchased land without a title record a fairly large output gain, 5.13-fold, somewhat

larger than the output gain for farmers without marketed land. We note however that the sample

of farms with purchased land is fairly limited, only 3.1 percent of the sample with 1.8 percent of

farms with purchased untitled land and 1.3 percent of farms with purchased titled land.

Even though we have showed a strong connection between land markets and the degree of output

loss generated by misallocation, farms with marketed land are still far from operating at their

efficient scale which suggests that land markets are still limited even for farms with access to

marketed land. For instance, the correlation between land size and farm TFP is .14 for farms

with no marketed land, .25 for farms with some marketed land, and .30 for farms that operate all

marketed land. Having access to marketed land implies that farmers can command more inputs

and produce more output. Also, operating marketed land is associated with greater access to other

markets (e.g., credit) and with other indicators of economic development. Farmers with marketed

land are substantially more educated than farmers with no marketed land, women in these farms

are more empowered in terms of labor force participation and market wages, more of these farmers

22The distribution of farm TFP, which we report in Figure C-1 in Appendix C, is slightly shifted to the right forfarms with marketed land compared to farms with no marketed land, however, the output gain difference betweenthe two groups remains large even when controlling for farm productivity.

28

Table 6: Land Markets—Output Gain for Farms with Marketed vs. Non-marketed Land

By Marketed Land Share By Marketed Land TypeNo Yes All Rented Purchased

(0%) (> 0%) (100%) Informal Formal Untit. Titled

Output Gain 4.15 1.97 1.57 1.72 1.73 5.13 1.39

Observations 5,962 1,189 746 215 682 126 97Sample (%) 83.4 16.6 10.4 3.0 9.5 1.8 1.3

Notes: The output gain is calculated as the ratio of efficient to actual output separately for subsamples of farmhouseholds defined by the share of different types of marketed land used. The share of marketed land is defined fromthe household-farm level information on how land was obtained, see Section 2. Each column refers to a particularsubsample. The first column reports the output gain for the subsample of household farms that do not operate anymarketed land. The second column refers to the subsample of household farms operating a strictly positive amountof marketed land, either purchased or rented-in. The third column refers to the subsample of household farms forwhich all their operated land is marketed land. The last four columns disaggregate the results by the main typesof marketed land: (1) rented informally, i.e. land borrowed for free or moved in without permission; (2) rentedformally, i.e. leaseholds, short-term rentals or farming as a tenant; (3) purchased without a title; and (4) purchasedwith a title. There is only 1% of households with marketed land whose type is missing in the Malawi ISA data.

are migrants, a fewer proportion live in rural areas, and a larger proportion of these farmers invest

in intermediate inputs and technology adoption. We report these dimensions of farm differences by

the type of operated land in Appendix C. One potential concern with our characterization is the

possibility that access to marketed land is driven by farm productivity. We find however very small

differences in productivity between farms with and without marketed land. These differences are

too small compared to what it would be if access to marketed land was purely driven by frictionless

selection in farm TFP. Even though on average farms with some marketed land are 25 percent more

productive than farms with no marketed land, the distribution of household-farm productivity is

very similar between the group of household-farms with marketed land compared to the group of

farms without marketed land, with respective variances of 1.19 and 1.17. Nevertheless, we can assess

the gains from reallocation holding constant farm-level TFP. To do so, we regress household-farm

output gains from reallocation on household-farm productivity and a dummy variable that controls

for whether the farm operates marketed land. Specifically, we estimate using OLS the following

relationship: lnyeiyai

= cons+ψ1 ln si +ψ21market, where ψ2 is captures the effect of marketed land on

29

the farm output gain controlling for farm productivity si. We use actual output as weights so as to

reproduce the output gains of reallocation in our nationwide benchmark. We find that ψ2 = −.39,

that is, operating marketed land decreases the output gain by 39 percent holding productivity

constant.

Finally, to gain further perspective of the productivity gains from reallocation, we note that Hsieh

and Klenow (2009) report the gains from reallocation for the manufacturing sectors in China and

India relative to those in the United States. The idea is that the U.S. data is not necessarily absent

of frictions or measurement and specification errors, but that more misallocation in China and India

than in the United States is indicative of more distortions in those countries. Our decomposition

of the degree of misallocation across groups of households farms with or without marketed land

provides a more direct reference for the gains of reallocation, that is, reallocation gains attained

by household farms without marketed land relative to farms operating marketed land. Taking

advantage of the fact that we can identify which household farms have marketed land, we restrict

the gains of reallocation of household-farms that do not operate marketed land to be at most the

gains attained by households farms that operate marketed land. The output gain of farms with no

marketed land relative to the output gain of farms with only marketed land is 2.6-fold (4.2/1.6).

This implies a reallocation gain that is three times the one obtained in Hsieh and Klenow (2009)

for the manufacturing sector of China and India relative to the United States. Further, this gain

may be a lower bound, as it is unlikely that there is no misallocation among farms operating only

marketed land given the extent of land reallocation frictions in Malawi. Nevertheless, the fact that

we use a group of household farms within Malawi as benchmark reference mitigates the effects of

differential institutional features and frictions (beyond land markets access) that are more prevalent

when using the United States or another country as reference in comparison with Malawi.

30

4.4 Discussion

4.4.1 Farm Size vs. Farm Productivity

A common theme in the development literature is the view that small farms are more efficient

than large farms. This view stems from a well-documented inverse relationship between yields

(output per unit of land) and farm size (e.g. Berry and Cline, 1979). A key concern is whether the

yield by farm size is a good proxy for farm TFP. Our data for Malawi provide an opportunity to

directly assess the relationship between the yield and farm TFP. Consistent with the literature, in

the Malawi data, the yield is negatively related with farm size.23 As documented earlier, yield and

farm TFP are strongly positively correlated across farms but this is because the allocation of land

is not related to productivity (misallocation) so many productive farmers are constrained by size.

But in an efficient allocation in our framework, there would be no relationship between yields and

farm TFP as the marginal product of land would be equalized across farmers.24 To appreciate the

distinction between farm TFP and farm size, note that land productivity of the largest 10 percent of

farms relative to the smallest 10 percent of farms is 0.34 (a 3-fold yield gap). This characterization

contrasts sharply with our earlier documentation that land productivity is strongly increasing with

farm TFP, see Figure 3 panel (b): land productivity of the 10 percent most productive farms relative

to the 10 percent least productive farms is a factor of 58-fold.

Whereas the pattern of land productivity by size suggests reallocation of land across farm sizes—

from large to small farms— the pattern of land productivity by farm TFP suggests reallocation

across farmers with different productivity—from less productive to more productive farmers. We

23See Appendix D, Figure D-1 which documents the relationship between land productivity (yield) and farm sizein our data. Land productivity declines with farm size, which conforms with the finding in the inverse yield-to-sizeliterature.

24The yield is a measure of farm productivity commonly used in studies of agricultural development, e.g., Bin-swanger et al. (1995). If the yield and farm TFP are uncorrelated in an efficient allocation, changes in policies andinstitutions that allow a better allocation of land would tend to, other things equal, reduce land productivity ofthe expanding farms and increase land productivity of the contracting farms, potentially masking the productivitybenefits of the reforms.

31

find that the aggregate output and productivity gains of these two forms of reallocation are dramat-

ically different. As discussed earlier, reallocating factors across farms with different productivity

can increase agricultural output and TFP by a factor of 3.6-fold. In contrast, reallocating factors

across farms by size, taking the yield-by-size as a measure of productivity, accrues a gain in output

of only 26 percent. And this gain may be an upper bound as the implementation of reallocation

by size may lead to further amounts of misallocation. For instance, land reforms are often pursued

to implement reallocation of land from large to small holders that are also accompanied by further

restrictions in land markets in order to bias redistribution to smallholders and landless individu-

als. This redistribution can lead to productivity losses instead of gains as it was the case in the

comprehensive land reform in the Philippines (see, Adamopoulos and Restuccia, 2015).

4.4.2 Further Robustness

We conduct robustness checks to alternative factor shares and technology parameters, specific crop-

type production, human capital and specific skills, and transitory health and other shocks.

Technology parameters We note that by restricting agricultural factor shares to those in the

United States in our baseline measure of farm productivity we aim at limiting the impact of factor

market distortions in Malawi on the calibrated shares. However, the U.S. agricultural sector is not

necessarily absent of frictions which, in turn, might differ from those in Malawi. To address this

issue, we use household-farm rented-in land and capital payments reported in our micro data to

compute, respectively, the land and capital share of income for the agricultural sector in Malawi.

We use these factor shares to redo our measurement of productivity and reallocation gains in

Malawi. The results are reported in Table 7. Our benchmark results that use the U.S. values of

the agricultural sector, are reproduced in the first column. The second and third columns provide

two alternatives. In the second column, Full Sample, we measure factor shares of output using

32

the average rental rates of land and capital. We compute these rates as the ratio of factor rental

payments to the factor stocks from the sample of farms that rent-in all of their capital and land.

Then we use these rental rates to impute the land and capital rental income for all farms in the

entire sample. This implies an average capital share of income of 0.19 and an average land share

of income of 0.39. With these factor shares, the output gain in the full sample is 3.57-fold, nearly

identical to the 3.59-fold in our baseline calibration. In the third column, Renting Sample, we

assume that the land and capital shares of output are the averages from the renting sample only,

that we then apply to all farms in the sample. In this case, factor shares are lower which imply

a smaller but still substantial output gain of 3-fold. We conclude that our results do not depend

crucially on factor shares using Malawi data, even if we use factor income shares from households

that have access to rental markets in Malawi.

Table 7: Reallocation Results with Factor Shares From Malawi Micro Data

Baseline Micro Malawi ISA DataCalibration Full Sample Renting Sample

(θk, θl)= (0.36,0.18) (0.19,0.39) (0.09,0.33)

Output gain 3.59 3.57 3.04

Notes: The baseline calibration selects factor shares based on U.S. data from Valentinyi and Herrendorf (2008). “FullSample” measures land and capital shares using the average rental rates of land and capital computed as the ratioof factor rental payments to factor stocks from the sample of farms that rent in all of their capital and land. Weuse these rental rates to impute the land and capital rental income per farm in the entire sample by multiplying therates by the respective farm-level stocks. “Renting Sample” measures factor shares using the average ratios of factorrental payments to value added using the renting sample only and then applying these averages to the entire sample.The sample of farm-households used is the same in all specifications.

Rather than factor shares of capital and land, the results are more sensitive to the implied income

share of labor. This is because in our framework, the share of labor is determined by the extent of

decreasing returns in the farm production function γ. In the misallocation literature, there is not

a lot of guidance as to the exact value of this parameter. A large quantitative literature considers

values between 0.8-0.85 (e.g. Restuccia and Rogerson, 2008), whereas monopolistic competition

33

frameworks such as that in Hsieh and Klenow (2009) consider a preference curvature parameter

that maps into decreasing returns of 0.67 (for an exposition of this mapping see Hopenhayn, 2014).

The value for γ ranges from 0.54 in our baseline calibration to U.S. shares to 0.58 and 0.42 using

Malawi data. Hence, our calibration and estimates are lower than the values typically considered

in the literature. For each value of γ, we recompute farm productivity and the output gains from

reallocation. While the output gains become small for small values of γ, in the range of plausible

values, from 0.4 to 0.8, the output gains range from 3-fold to more than 10-fold. That is, using

values of γ closer to those considered in the manufacturing (and other) sectors would increase our

output gains from reallocation. We have also explored a constant elasticity of substitution (CES)

production technology and find that our results remain fairly robust to more extreme elasticity of

substitution between capital and land.

Crop type production In our data, most of the agricultural land is devoted to maize production

(around 80 percent), see also FAO (2013).25 Since optimal farm operational scale may differ across

crop types, we investigate whether the output gains are different across farms producing different

crops. Since most farms produce both maize and non-maize crops, we consider reallocation among

farms that produce mostly maize and those farms that produce mostly non-maize. We report the

results in Table 8 where the first column reports the output gain in the nationwide benchmark,

the next two columns report output gains for farms that produce mostly maize (whether the maize

share is positive or above the median of all farms), and the last two columns report output gains

for those farms that produce mostly non-maize (wether the non-maize share is positive or above

the median). The output gain among farms that produce maize is somewhat smaller (2.7-fold for

farms with maize production above the median versus 3.6-fold in the benchmark), which implies

that the farms with non-maize production have output gains that are larger (4.1-fold for farms that

25Maize, cassava, and potatoes are the main crops produced in Malawi in terms of volume with roughly equalamount in tonnes, but it is maize that accumulates the vast majority of cropland. The importance of maize furtherappears in terms of the household diet in Malawi where about 50% the average daily calorie intake is obtained frommaize, 8.4% from potatoes, and 5.8% from cassava, see FAO (2013).

34

produce non-maize above the median). We conclude that crop composition and the dependence of

maize production in Malawi are not critical in determining the large negative productivity impact

of factor misallocation.

Table 8: Reallocation Results by Crop Type

Maize Share Non-Maize ShareBenchmark >0% > Median >0% > Median

Output gain 3.59 2.83 2.71 3.82 4.13

Notes: The table reports output gains within households producing maize and non-maize. Because most householdfarms produce both maize and non-maize crops (around 74 percent), the table also reports output gains withinhousehold-farms whose production of maize is above median production. The median share of maize productionwithin the group of household-farms that produce a strictly positive amount of maize is 83.1% and the median shareof non-maize production within the group of household-farms that produce a strictly positive amount of non-maizecrops is 66.2%.

Other robustness results In Appendix D we report further robustness results related to dif-

ferences in human capital across farm operators and other specific skills, as well as adjusting our

measure of farm TFP to other transitory shocks (other than rain) such as health and death shocks,

food security risk, marital status, distance to markets and the availability of other income sources.

Overall, we find that the output gain of factor reallocation to be large in all cases. More impor-

tantly, ISA conducted a second wave of interviews for roughly one-fourth of the original sample in

2013. That is, there is a panel for almost 3,000 household farms in Malawi for 2010 and 2013. This

provides us with the opportunity to recover for each household an alternative measure of produc-

tivity defined as the estimated fixed effect of individual farm TFP across waves. We find that the

fixed effects from the panel roughly capture 87 percent of the original cross-sectional TFP variation.

Importantly, this new measure of TFP based on fixed effects is, again, essentially unrelated to land

and capital. The output gain from efficient reallocation in 2010/11 using this alternative measure

of farm TFP is 3.1-fold, fairly close to the 3.6-fold output gain with our benchmark measure of farm

TFP. Hence, our results are fairly robust to alternative measures of productivity that isolate more

35

permanent components of farm TFP.

5 Implications for Structural Change

Increased productivity from an efficient allocation of resources would trigger broader implications

in the economy. We assess the implications of increased productivity in agriculture on structural

transformation associated with the reallocation of labor across sectors. A TFP increase of a 3.6-

fold factor in the agricultural sector would produce a process of substantial structural change in

the Malawian economy. This reallocation would produce broad impacts in the economy through

well-known features such as the potential selection effects associated with the movement of labor

from agriculture to non-agriculture and dynamic investment effects such as additional investments

farmers would make to exploit increased farm size, capital and human capital accumulation, among

many others.

To provide a simple characterization of these broader implications of reallocation, we consider an

extension of the previous analysis to allow for a non-agricultural sector.26 Recall that the aggregate

production function for agriculture in the efficient allocation is given by equation (2). To simplify

the analysis, without much loss of generality, we abstract from capital by assuming that α = 0.

Hence, the aggregate production function in agriculture is given by:

Ya = ZLγN1−γa ,

where Na is the fraction of employment in agriculture and γ = 0.54. The non agricultural production

function is given by Yn = A(1−Na). We assume that preferences are such that consumers not only

have a minimum consumption of agricultural goods a, but also this level is a satiation point so per

26For a more elaborate analysis of misallocation in agriculture in the context of a two-sector economy seeAdamopoulos and Restuccia (2014).

36

capita consumption of agricultural goods is given by a and any income above the one required for this

amount of agricultural consumption is spent on non-agricultural goods.27 We continue to consider a

benevolent social planner that chooses the allocation of labor across sectors to maximize consumer’s

welfare. Given our assumptions about preferences, the solution to this allocation problem has a

simple form and is given by:

Na =( a

ZLγ

) 11−γ

.

Note that the solution is such that an increase in productivity in the agricultural sector Z reduces

the share of employment in agriculture. In particular, the change in employment in agriculture can

be easily calculated using the above equation from the change in TFP in agriculture raised to the

power 1/(1 − γ). We report the results in Table 9. With γ = 0.54, an increase in productivity

of 3.6 implies a decrease in the employment share in agriculture of 16.2-fold. In other words, the

share of employment in agriculture in Malawi would decrease from the actual 65 percent to only

4 percent which is close to the average for rich countries, see for instance Restuccia et al. (2008).

This tremendous reallocation of labor from agriculture to non-agriculture implies that average

farm size would increase by a factor of 16.2-fold (recall that in our previous one-sector analysis

of misallocation, by construction the gains in agricultural productivity of moving to the efficient

allocation involved no change in average farm size). It is also easy to see that the increase in labor

productivity in agriculture Ya/Na is given by a factor of 16.2-fold.28 Hence, for our quantitative

experiment, the entire increase in agricultural labor productivity is reflected in an increase in average

farm size and none in an increase in the yield.

We emphasize that the effects just described abstract from other potential sources of amplification.

For instance, the potential role of selection into the ability of farmers that stay in agriculture.

This feature that has been found quantitatively important in amplifying the productivity increases

27See for instance Gollin et al. (2002) for a model with such preferences.28Note that agricultural output per unit of land or yield Ya/L remains the same because for the preferences we

consider, the increase in productivity is exactly offset by the decrease in agricultural labor so total agricultural outputremains the same.

37

Table 9: Misallocation, Productivity, and Structural Transformation

Benchmark IncreasedProductivity

Productivity in Agriculture (Z) 1.00 3.60Share of Employment in Ag. (Na) 0.65 0.04Average Farm Size (L/Na) 1.00 16.2Labor Productivity in Ag. (Ya/Na) 1.00 16.2

Notes: The “Benchmark” refers to the actual allocation in Malawi. The “Increased Productivity” refers to theefficient reallocation which in the full sample increases total factor productivity in agriculture by a factor of 3.6-fold.The table reports the effects of increased TFP in agriculture on output per unit of land in agriculture (yield), theshare of employment in agriculture, average farm size, and aggregate labor productivity in agriculture when factorsare allowed to be reallocated across agriculture and non-agriculture.

in the agricultural sector (e.g., Lagakos and Waugh 2013 and Adamopoulos et al. 2017). Also,

complementary investments such as mechanization, improvements in land quality, or the adoption

of modern technologies could further increase productivity in the agricultural sector. Overall, we

find that the increase in agricultural productivity due to the more efficient allocation of factors can

unravel a substantial process of structural change and productivity growth in agriculture that can

dramatically change the face of the Malawian economy.

6 Implications for Economic Inequality

Unlike the actual distribution of factors in the Malawian economy which is fairly flat across farmers

with different productivity, the efficient allocation implies a substantial increase in the dispersion

of operational scales, in terms of the distribution of operated capital and land across farmers. This

massive redistribution of factors across farmers may lead to concerns over distributional implica-

tions, specially if the actual allocation of factors reflect policy and institutional choices in place to

alleviate poverty and distributional concerns, see de Magalhaes and Santaeulalia-Llopis (2015).29

29For example, de Magalhaes and Santaeulalia-Llopis (2015) show that the bottom 20% and top 20% of the landdistribution hold a similar share of aggregate consumption in rural Malawi, respectively 20% and 25%.

38

We study the distributional implications of the efficient operational scales. Table 10 reports the

actual and efficient distribution of factors, output, and income across farmers by productivity.

Whereas the actual distribution of land across farm TFP is fairly flat, with most farms operating

less than 2 acres of land, the efficient distribution implies that the top quintile of farm TFP operates

on average almost 6 acres, representing 97 percent of the total land. This implies a substantial

redistribution of factors to achieve higher levels of aggregate productivity. We note that despite

relative equalization of factor inputs across farmers, the actual distribution of income is widely

dispersed in Malawi, in fact as dispersed as the distribution of productivity. For instance, taking

agricultural output as a measure of farm income, the ratio of top to bottom quintile of income is a

factor of 34-fold even though the ratio of capital and land factors is within a factor of 1 to 2-fold.

To put it differently, equalizing access to land across households does not necessarily translate into

income equalization as these farmers differ substantially in their productivity.

To gauge the distributional income effects of factor redistribution we pursue the following coun-

terfactual. We consider the actual distribution of factors as endowments and allow the efficient

allocation to be achieved via perfectly competitive rental markets. Given the competitive rental

rates of capital and land in this decentralized solution, we compute the income associated with the

efficient allocation as:

endowment income = rk(ka − ke) + rl(l

a − le) + ye,

where (ka, la) are the actual allocation of capital and land which we take as endowment and (ke, le)

are the efficient allocations, rk and rl are the prices that support the efficient allocation as a

competitive equilibrium, and ye is efficient output. Table 10 reports the results for this measure

of endowment income and compare the income inequality to that of the actual income, which we

assume is approximated by actual agricultural output ya.

39

Table 10: Actual and Efficient Distribution of Land, Capital, MPL and MPK

Productivity PartitionQuintiles

1st 2nd 3rd 4th 5th var(lnx)

Farm productivity (si) .75 2.10 3.72 6.39 21.50 1.435

Land (li)Actual 1.19 .87 1.01 1.03 1.99 .749Efficient .00 .01 .04 .14 5.91 6.782

Capital (ki)Actual 55.93 25.35 21.84 24.70 26.71 1.820Efficient .04 .32 1.10 3.60 149.52 6.782

MPL (∝ Yield)Actual 4.21 11.00 17.82 29.10 82.04 1.485Efficient 76.30 76.30 76.30 76.30 76.30 .000

MPKActual .73 2.19 3.94 7.25 24.54 2.154Efficient 6.03 6.03 6.03 6.03 6.03 .000

Output (yi)Level:

Actual .14 .39 .69 1.20 4.78 1.824Efficient .00 .05 .18 .60 25.06 6.782

Percentage of total:Actual 2.01 5.46 9.57 16.67 66.26 –Efficient .02 .20 .71 2.33 96.71 –

Agricultural incomeLevel:

Actual .14 .39 .69 1.20 4.78 1.824Efficient 4.28 2.22 2.17 2.56 14.65 1.228Income gain 23.70 3.88 2.27 1.58 1.97 –

Percentage of total:Actual 2.01 5.46 9.57 16.67 66.26 –Efficient 16.55 8.58 8.41 9.88 56.56 –

Notes: Households are ranked by farm productivity si. Land, capital, and output are in per hours terms. MPL is

the marginal product of land and MPK the marginal product of capital computed for each bin as, MPL = .18∑

i yi∑i li

and MPK = .36∑

i yi∑i ki

, and the variance is computed with respect to household-farm specific ratios MPLi = .18 yi

li

and MPKi = .36 yi

ki. Actual income is agricultural output whereas efficient income is computed assuming actual

allocations are the endowments and the efficient allocation is achieved via perfectly competitive rental markets.Income gain is the efficient to actual income ratio.

40

Not only farmers in the lower end of the productivity (and income) distribution benefit the most

from the increase in the return to factors, but also overall inequality declines. For instance, the

overall variance of the log income decreases from 1.8 with actual income to 1.2 to efficient income.

More dramatic are the changes in income across the richest and poorest households. Whereas

the ratio of income between farmers in the top and bottom quintiles is a factor of 34-fold in the

actual allocation, this ratio is 3.4 in the efficient allocation, that is income inequality among these

farmers falls by a factor of 10-fold. Moreover, the ratio of efficient to actual income increases for

all household farms but this increase is larger for the poorest households: 23.7-fold increase for the

first quintile, 3.9-fold for the second quintile and only 2-fold for the top quintile.

Well-functioning rental markets for capital and land to achieve the efficient allocation of operational

scales can lead to substantial increases in agricultural productivity as well as dramatic reductions

in inequality levels and poverty.

7 Conclusions

Misallocation of land and other complementary factors in the agricultural sector is enormous in

Africa. Using detailed nationally representative household-farm level data for Malawi, we show

that a reallocation of land (and capital) to their efficient uses increases agricultural output and

productivity by a factor of 3.6-fold. We show these large gains in the agricultural sector in Malawi

are not due to differentials in TFP dispersion compared with other sectors or more developed

countries, but to severe factor misallocation. We find that land size (and capital) are essentially

unrelated to household-farm level TFP, which is perhaps not surprising given the egalitarian nature

of land-use distributions and weak property rights over land. These restrictions imply that 83% of

the total operated land in Malawi is not marketed. Indeed, our analysis provides a strong empirical

connection between factor misallocation and restricted land markets. The productivity gains from

41

reallocation are 2.6 times larger for farms with no marketed land than for farms with marketed land,

which is roughly three times the reallocation gains found in the manufacturing sector of China and

India in Hsieh and Klenow (2009).

We show that the increase in aggregate productivity from the efficient reallocation of factors could

trigger of a profound process of structural change by which the agricultural sector in Malawi could

approach levels of farm size and sectoral employment shares of the industrialized world. This result

implies that factor misallocation in the agricultural sector is of first-order importance to understand

productivity differences and economic development across countries. Further, in a counterfactual

exercise we show that the introduction of rental markets where operational scales can deviate from

land ownership not only increases aggregate productivity to its efficient level but also decreases

income inequality; large but unproductive farmers are better off by renting-out land to small but

productive farmers. These findings point to a pressing need to facilitate the reallocation of land to

the more productive farmers without necessarily altering the ownership structure. This requires the

development of well-functioning rental markets and well-defined property rights over land. What

policies and institutions are best in promoting a better allocation of resources across farmers is of

crucial importance for future research.

Looking ahead, while our analysis takes the distribution of land and productivity across farmers

as given and asked about the efficiency gains of reallocation, it may also be of interest to study

the dynamic implications of misallocation for productivity whereby a reduction in misallocation

encourages the more productive farmers to grow, utilize modern inputs (mechanization, chemical

seeds, intermediate inputs), and invest in farm productivity. See Restuccia and Rogerson (2016) for

a discussion of the importance of the dynamic implications of misallocation. Further, while the gains

from the reallocation of agricultural inputs across wives and husbands within farm households are

small compared to the gains from reallocation across households, the quantitative role of women—

who are potentially more restricted in land (and capital) inputs than men in many parts of the

42

world—in understanding productivity differences across farms remains a very open but important

question. We leave these interesting and important extensions of our analysis for future research.

References

Adamopoulos, T., Brandt, L., Leight, J., and Restuccia, D. (2017). Misallocation, Selection andProductivity: A Quantitative Analysis with Panel Data from China. Working Papers tecipa-574,University of Toronto, Department of Economics.

Adamopoulos, T. and Restuccia, D. (2014). The Size Distribution of Farms and InternationalProductivity Differences. American Economic Review, 104(6):1667–97.

Adamopoulos, T. and Restuccia, D. (2015). Land Reform and Productivity: A Quantitative Anal-ysis with Micro Data. manuscript, University of Toronto.

Banerjee, A. V., Gertler, P. J., and Ghatak, M. (2002). Empowerment and efficiency: Tenancyreform in west bengal. Journal of political economy, 110(2):239–280.

Beegle, K., Carletto, C., and Himelein, K. (2012). Reliability of Recall in Agricultural Data. Journalof Development Economics, 98(1):34–41.

Berry, A. and Cline, W. (1979). Agrarian Structure and Productivity in Developing Countries.Baltimore: Johns Hopkins University.

Besley, T. (1995). Property Rights and Investment Incentives: Theory and Evidence from Ghana.Journal of Political Economy, pages 903–937.

Besley, T. and Ghatak, M. (2010). Property Rights and Economic Development, volume 5 of Hand-book of Development Economics, chapter 0, pages 4525–4595. Elsevier.

Bick, A., Fuchs-Schundeln, N., and Lagakos, D. (2016). How do Average Hours Worked Vary withDevelopment? Cross-Country Evidence and Implications. Working Paper 21874, National Bureauof Economic Research.

Binswanger, H. P., Deininger, K., and Feder, G. (1995). Power, Distortions, Revolt and Reform inAgricultural Land Relations. In Chenery, H. and Srinivasan, T., editors, Handbook of DevelopmentEconomics, volume 3 of Handbook of Development Economics, chapter 42, pages 2659–2772.Elsevier.

Buera, F. J., Kaboski, J. P., and Shin, Y. (2014). Macro-Perspective on Asset Grants Programs:Occupational and Wealth Mobility. American Economic Review, 104(5):159–64.

43

Carletto, C., Savastano, S., and Zezza, A. (2013). Fact or Artifact: The Impact of Measure-ment Errors on the Farm SizeProductivity Relationship. Journal of Development Economics,103(C):254–261.

Chen, C. (2016). Untitled Land, Occupational Choice and Agricultural Productivity,. AmericanEconomic Journal: Macroeconomics, forthcoming.

de Magalhaes, L. and Santaeulalia-Llopis, R. (2015). The Consumption, Income, and Wealth ofthe Poorest: Cross-Sectional Facts of Rural and Urban Sub-Saharan Africa for Macroeconomists.World Bank Policy Research Working Paper, WPS7337.

de Soto, H. (2000). The Mystery of Capital: Why Capitalism Triumphs in the West and FailsEverywhere Else. Basic Books. New York.

FAO (2013). Malawi BEFS Country Brief. Technical report, Food and Agriculture Organization ofthe United Nations.

Foster, A. D. and Rosenzweig, M. R. (2011). Are Indian Farms Too Small? Mechanization, AgencyCosts, and Farm Efficiency. Unpublished Manuscript, Brown University and Yale University.

Gancia, G. A. and Zilibotti, F. (2009). Technological change and the wealth of nations. AnnualReview of Economics, 1:93–120.

Goldstein, M. and Udry, C. (2008). The Profits of Power: Land Rights and Agricultural Investmentin Ghana. Journal of Political Economy, 116(6):981–1022.

Gollin, D., Lagakos, D., and Waugh, M. E. (2014). Agricultural Productivity Differences acrossCountries. American Economic Review, 104(5):165–70.

Gollin, D., Parente, S., and Rogerson, R. (2002). The Role of Agriculture in Development. AmericanEconomic Review, 92(2):160–164.

Gottlieb, C., Grobovsek, J., et al. (2015). Communal land and agricultural productivity. Technicalreport.

Hopenhayn, H. A. (2014). Firms, misallocation, and aggregate productivity: A review. AnnualReview of Economics, (0).

Hsieh, C.-T. and Klenow, P. J. (2009). Misallocation and Manufacturing TFP in China and India.The Quarterly Journal of Economics, 124(4):1403–1448.

Kishindo, P. (2011). The Village Head and the Problem of Role Relevance in the Context ofDeclining Rural Land Availability in Malawi. Working Papers 2, Centre for Social Research,Chancellor College, Zomba, Malawi.

Lagakos, D. and Waugh, M. E. (2013). Selection, Agriculture, and Cross-country ProductivityDifferences. The American Economic Review, 103(2):948–980.

44

Larson, D. F., Otsuka, K., Matsumoto, T., and Kilic, T. (2014). Should African Rural Devel-opment Strategies Depend on Smallholder Farms? An Exploration of the Inverse-ProductivityHypothesis. Agricultural Economics, 45(3):355–367.

Midrigan, V. and Xu, D. Y. (2014). Finance and Misallocation: Evidence from Plant-Level Data.American Economic Review, 104(2):422–458.

Morris, B. (2016). An Environmental History of Southern Malawi: Land and People of the ShireHighlands.

Pachai, B. (1973). Land Policies in Malawi: An Examination of the Colonial Legacy. The Journalof African History, 14(4):681–698.

Restuccia, D. and Rogerson, R. (2008). Policy Distortions and Aggregate Productivity with Het-erogeneous Plants. Review of Economic Dynamics, 11(4):707–720.

Restuccia, D. and Rogerson, R. (2016). The Causes and Costs of Misallocation. manuscript,University of Toronto.

Restuccia, D., Yang, D. T., and Zhu, X. (2008). Agriculture and Aggregate Productivity: AQuantitative Cross-Country Analysis. Journal of Monetary Economics, 55(2):234–250.

Santaeulalia-Llopis, R. and Zheng, Y. (2016). The Price of Growth: Consumption Insurance inChina 1989-2009. Working papers, Barcelona GSE.

Shaban, R. A. (1987). Testing Between Competing Models of Sharecropping. The Journal ofPolitical Economy, pages 893–920.

Song, Z., Storesletten, K., and Zilibotti, F. (2011). Growing Like China. American EconomicReview, 101(1):196–233.

Udry, C. (1996). Gender, Agricultural Production, and the Theory of the Household. Journal ofPolitical Economy, 104(5):1010–46.

Valentinyi, A. and Herrendorf, B. (2008). Measuring Factor Income Shares at the Sector Level.Review of Economic Dynamics, 11(4):820–835.

45

*NOT FOR PUBLICATION

Online Appendix: “Land Misallocation and Productivity”

By Diego Restuccia and Raul Santaeulalia-Llopis.

1

Contents

A The Data: The Malawi Integrated Survey of Agriculture, 2010/2011 3

A.1 Agricultural Production, Intermediate Inputs, and Value Added . . . . . . . . . . . . . . . . . 3

A.2 Land Size and Land Quality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

A.3 Rain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

A.4 Hours . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

A.5 Capital Equipment and Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

B Trimming Strategy and Measurement Error 10

B.1 Trimming Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

B.2 Dealing with Measurement Error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

B.2.1 Potential Recall Bias for Agricultural Production and Hours . . . . . . . . . . . . . . . 11

B.2.2 Proxying Agricultural Capital with an Asset Index . . . . . . . . . . . . . . . . . . . . 12

B.2.3 Further Accuracy Assessments: Bootstrapping Our Household-Farms . . . . . . . . . . 14

C Further Insights on the Role of Land Markets 15

C.1 Distribution of Farm TFP by Marketed Land . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

C.2 Demographics, Human Capital, Female Empowerment, and Welfare . . . . . . . . . . . . . . 16

C.3 Land Markets and Access to Credit and Other Markets . . . . . . . . . . . . . . . . . . . . . 18

D Discussion on Farm Size and Robustness 23

D.1 Farm Size . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

D.2 Human Capital and Specific Skills . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

D.3 Productivity Net of Transitory Health and Other Shocks . . . . . . . . . . . . . . . . . . . . . 24

2

A The Data: The Malawi Integrated Survey of Agriculture, 2010/2011

The Malawi 2010-2011 Integrated Survey of Agriculture (ISA) is part of a new generation of

household surveys funded by the Bill & Melinda Gates Foundation (BMGF) and led by the

Living Standards Measurement Study (LSMS) Team in the Development Research Group

(DECRG) of the World Bank to improve the quality and policy relevance of household-

level data on agriculture in Sub-Saharan Africa. This way, the Malawi ISA incorporates

an extended and comprehensive ’Agricultural Questionnaire’ on agricultural production and

factor inputs, including land quality and rain.

An important aspect of the Malawi ISA for the evaluation of macroeconomic policies is

that it is a nationally-representative survey. To ensure representativeness, the survey follows

a stratified 2-stage sample design. First, 768 enumeration areas (EAs) are selected with

probability proportional to size (PPS) within each district of the 27 districts of Malawi. The

sampling frame is based on the 2008 Population and Housing Census (PHC) that includes

three major regions of Malawi, namely North, Center and South; and is stratified into rural

and urban strata. The urban strata include the four major urban areas: Lilongwe City (the

current capital city), Blantyre City (capital of the South region), Mzuzu City (capital of

the North Region), and the Municipality of Zomba (old colonial capital). Second, random

systematic sampling is used to select 16 primary households and 5 replacement households

from the household listing for each sample EA.

A.1 Agricultural Production, Intermediate Inputs, and Value Added

The most important aspect of the Malawi ISA is that it provides detailed information on

physical quantities of all farm agricultural production and the full set of inputs used in

that production. In this Appendix we discuss the construction of agricultural production,

intermediate inputs, and value added at the household-farm level.

Each household agricultural production are reported separately for crop produced by

season. We focus on the rainy season that captures about 93% of the entire agricultural

3

production in Malawi. Denote by ψi the type of crop i.1 Denote the total quantity of

nonpermanent crop-ψi harvested by household z per season s by yψi,s,z (AG:g13). The

information on harvested crop is available per plot, that is, yψi,s,z =∑

d yψi,d,s,z where yψi,d,s,z

is the amount of crop-ψi harvested by household z in season s and plot d. We find that there

are up to 5 types of crops that are potentially harvested per plot. These data are collected

for each plot cultivated by the household-farm.2 The quantity produced of a given crop is

in most cases reported in Kilograms (more than 90 percent of observations). We follow a

standard conversion procedure to Kgs for the small set of cases in which units are reported

differently.3 We then use a common at-the-gate price to estimate the value of production for

both sold and unsold production.

In terms of intermediate inputs, each household z incurs in input costs per season and

plot associated with fertilizers, herbicides, pesticides, and seeds. Fertilizers are the most

important intermediate input in terms of monetary costs. The main types of inorganic

fertilizer used in Malawi are 23:21:0+4S/Chitowe and Urea, that account for almost 57% of

fertilizer use. Inorganic fertilizer is used in 61% of all plots and organic fertilizer is used in

10.7% of all plots.4 To attribute a value for each type of fertilizer, we use the prices paid

by those households that purchased fertilizer without coupons, i.e., without interference

of the Malawi Farming Input Subsidy Program (FSIP). We use the median price for all

households.5 More than 96 percent of all fertilizer purchases are reported in Kg., while

the rest are reported in liters, buckets, wheelbarrows, ox carts, or others. We again use a

1Information on all typical nonpermanent crops φi is available. These crops are: maize (local, compos-ite/OPV, hybrid, hybrid recycled), tobacco (Burley, flue cured, NNDF, SDF, oriental, other), groundnut(chalimbana, CG7, mani-pintar, mawanga, JL24, other), rice (local, faya, pusa, TCG10, IET4094, kilo-mbero, etc.), ground bean, sweet potato, Irish (Malawi) potato, wheat, finger millet (mawere), Sorghum,peral millet (mchewere), beans, soybeans, pigeonpea (nandolo) cotton, sunflower, sugar cane, cabbage, tana-posi, nkhwani, therere/OKRA, tomato, onion, pea paprika, other). For exposition simplicity, we use rainyseason variables name codes (AG:b-i) in the Agricultural questionnaire. The procedure for the simba (dry)season name codes (AG:j-o) is analogous, replacing ’c’ for ’j’ and so on.

2A plot is defined as a continuous piece of land on which a unique crop (or mixture of crops) is grownunder a uniform consistent crop management system. These questions are generally asked to the person thatmakes the economic decisions on the plot (see AG:d01-d02).

3We use a ’price conversion’ unit procedure to convert all crop-units pairs into Kgs, see de Magalhaesand Santaeulalia-Llopis (2015).

4Pesticides/herbicides are used in 2.2% of all plots.5As we discuss below, we use common prices so that our measure of productivity does not inherit nominal

variation from subsidies across households.

4

standard conversion procedure to deal with the small set of observations where the reporting

unit is different from Kg. We then value intermediate units using a common median price

per unit for each type of intermediate input.

Finally, we construct our measure of output as agricultural value added at the farm level.

That is, to define output for each household-farm we subtract to the household’s agricultural

production valued at common at-the-gate prices the value of intermediate inputs, also valued

at common prices.

A.2 Land Size and Land Quality

Our preferred measure of household land is the sum of the size (in acres) of each household’s

plot used for cultivation. We include rented-in land (6.9% of all plots) in household land size.

For the vast majority of plots, 98%, the acres per plot are recorded using GPS with precision

of 1% of an acre. In the remaining 2% of plots, size is self-reported with an estimate from

the household. This leaves virtually no room for error in our measure of land input, see a

detailed assessment in Carletto et al. (2013). The data also contain detailed information on

the quality of land for each plot used in each household. We consider all 11 dimensions of

land quality available: elevation, slope, erosion, soil quality, nutrient availability, nutrient

retention capacity, rooting conditions, oxygen availability to roots, excess salts, topicality,

and workability.6 Our benchmark land quality index is defined per household as the predicted

value of output (net of rain effects which we discuss in the next Section) generated by the

joint behavior of all dimensions of land quality controlling for capital and land size, see the

first column in Table A-1.7 We also explore alternative definitions of the land quality index

in Table A-1 that depend on the number of land quality dimensions that we incorporate and

on the way we control for capital and land. A reassuring aspect of our land quality index,

defined from physical measures (e.g., erosion, soil quality, etc.), is that it is positively related

6 The slope (in %) and elevation (in meters) are continuous variables while the rest of land qualityvariables are categorical such as soil quality (1 good, 2 fair, 3 poor), erosion (1 none, 2 low, 3 moderate, 4high ), nutrient availability, nutrient retention, rooting conditions, oxygen to roots, excess of salts, toxicityand workability (1 constraint, 2 moderate constraint, 3 severe constraint and 4 very severe constraint).

7For households with more than one plot we take a weighted average of each of these 11 dimensions usingthe size of plot as weight.

5

to land prices, see Table A-2, in particular, for the plots that have been obtained through the

market for which we obtain a correlation of .503. Finally, we perform a robustness analysis

of our reallocation results with respect to our entire set of land quality indexes without

substantial changes in our findings, see Table A-3.

Table A-1: Land Quality Index and Its Dimensions

Land Quality Index

Bench. Alternative Definitions:Dimensions: q0 q1 q2 q3 q4 q5

Elevation X X X X X XSlope X X X X X XErosion X – X – X –Nutritient Availability X – X – X –Nutritient Retention X – X – X –Rooting Conditions X – X – X –Oxygen to Roots X – X – X –Excess Salts X – X – X –Toxicity X – X – X –Workability X – X – X –

Additional Controls:Capital X X – – X X

(Const.) (Const.)

Land Size X X – – X X(Const.) (Const.)

A.3 Rain

We use the annual precipitation which is the total rainfall in millimetres (mm) in the last

12 months. Figure A-1 plots the effects of rain, grouped in percentiles, on output. We find

a clear positive and concave relationship between rain and output, that is, the effects of

rain on output become less strong as rain increases. Throughout, our benchmark measure

of output (value added) is net of the rain effects. As rain might be more relevant in some

months than others we also tried an alternative measure of rain, the wettest quarter within

the last 12 months. We find that this alternative measure of rain effects does not alter our

6

Table A-2: Land Quality Index and Land Price, Malawi ISA 2010/11

Land Quality Index

Bench. Alternative Definitions:q0 q1 q2 q3 q4 q5

Full Sample:Effects on Land Price .128 .271 .089 .197 .123 .312

Non-Marketed Land:Effects on Land Price .116 .265 .076 .182 .105 .301

Marketed Land Share >0%:Effects on Land Price .188 .236 .142 .227 .228 .311

Marketed Land Share 100%:Effects on Land Price .275 .308 .225 .292 .332 .372

Purchased Land, Untitled:Effects on Land Price .191 .246 .158 .199 .223 .299

Purchased Land, Titled:Effects on Land Price .503 .763 .296 .673 .536 .885

Notes: The table reports the OLS estimated coefficient β from the following specification: ln qi = cons +β ln pi + 1region where qi is the land quality index, pi is the price of land computed as the ratio betweenthe self-reported estimated land value (under the hypothetical scenario in which the owner sells the land)and the quantity of land, and 1region is a set of regional dummies. We estimate this relationship using ourbenchmark measure of land quality and 5 alternative measures of land quality defined in Table A-1. In eachcase we consider alternative samples that include the full sample, farms with non-marketed land, farms witha positive marketed land share, farms with only marketed land, farms that have purchased untitled land andfarms that have purchased titled land.

Table A-3: Output Gain, Robustness to Land Quality Index

Bench. Alternative Definitions of Land Qualityq0 q1 q2 q3 q4 q5

Output gain 3.59 3.42 3.90 3.53 3.44 3.42

7

results.

Figure A-1: The Effects of Rain on Output per Hour

Notes: Rain in percentiles and quadratic behavior.

A.4 Hours

In Malawi, not only the household head but also a large proportion of the households mem-

bers, which average 4.57 per household, contribute to agricultural work. The household

head is identified as the person who makes economic decisions in the household (e.g., use of

production or transfers). We define household members as individuals that have lived in the

household at least 9 months in the last 12 months. These household members potentially in-

clude family (e.g. children, spouses, siblings, and parents) and also non-relatives (e.g. lodgers

8

and servants).8 Individual information about each household member’s (including children)

extensive and intensive margins of labor supply is collected: (i) weeks worked, (ii) days per

week, and (iii) hours per day. This information available in the ’Agriculture Questionnaire’ is

provided retrospectively by plot and by agricultural activity covering the entire agricultural

production. Agricultural activities consist of land preparation/ planting, weeding/fertilizing,

and harvesting.9 The hired labor and free/exchange labor (e.g. exchange labourers or assis-

tance for nothing in return) number of days worked by men, women and children are also

collected by plot, activity and season. The detailed information on individual agricultural

hours through the entire year avoids the seasonal component of labor supply; that is, we

do not rely on data on labor supply related to ’last week/month’ behavior. Our benchmark

measure of household labor supply is aggregate hours of all individuals (household members

and nonmembers) supplied in all plots cultivated by the household in the rainy season.10

A.5 Capital Equipment and Structures

Agricultural capital equipment includes implements (i.e., hand hoe, slasher, axe, sprayer,

panga knife, sickle, treadle pump and watering can) and machinery (e.g. ox cart, ox plough,

tractor, tractor plough, ridger, cultivator, generator, motorized pump, grain mill, and oth-

ers). Agricultural capital structures includes chicken houses, livestock kraals, poultry kraals,

storage houses, granaries, barns, pig sties, etc. To measure the capital stock per item we

use the estimated current selling price of capital items (“If you wanted to sell one of this

[ITEM] today, how much would you receive?”) after conditioning on its use (“Did your

household use the [ITEM] during the last 12 months?”). We construct the household agri-

cultural capital stock by aggregating across all agricultural items. The use of the selling

price (not available in previous LSMS data) avoids the cumbersome use of the age of capital

8Individual characteristics of each household member are collected. They include sex, age, language,religion, educational attainment, health behavior and status, marital status, societal system, geographicalvariables, migration characteristics, and so on.

9We conduct robustness to recall bias in agricultural production and hours in Appendix B.10We have conducted robustness analysis under different specifications of hours that include using a differ-

ent set of individuals (e.g., only head hours, and only members hours) or different units (e.g., weeks per year,or days per year). Further, to control for human capital, i.e., the fact that not all hours might contributethe same to agricultural production, we construct household efficiency units by weighting individual hoursusing the wages of hired labor by age and sex groups as weights. We find that our results are robust to thesealternative specifications.

9

to impute current value from the value at the time of purchase which requires recalling and

depreciation assumptions by asset’s age. We discuss further issues on the measurement of

capital, including alternative measures such as asset indexes, in Appendix B.

B Trimming Strategy and Measurement Error

We discuss our trimming strategy and measurement issues associated with agricultural out-

put and factor inputs.

B.1 Trimming Strategy

There are very few missing observations in agricultural production, a blessing for a survey

in a poor country. Our understanding from the World Bank field managers in charge of

the data collection is that this is due to the fact that respondents took the Malawi ISA as

“official”. We focus on those households with nonnegative agricultural production, i.e, 9,538

household farms. Then, to retrieve household farm productivity (see Section ??), we need

data on all household farm inputs, i.e., hours, capital, land, all eleven dimensions of quality of

land, and rain. These data are available for 92% of all observations. Further, our measure of

output is value added, that is, agricultural production net of intermediate inputs (i.e., seeds,

fertilizers, herbicides and pesticides). This leaves room for potential non-positive values of

output that we drop. The resulting sample consists of 7,202 household farms.11 Last, after

visual inspection, we further trim from the distribution of household-farm productivity about

1% of the sample.

B.2 Dealing with Measurement Error

While the Malawi ISA survey response rate is high and there is little room for measurement

error in our estimates of land size (as each plots’ size is GPS measured, see Appendix A.2),

there are other potential sources of measurement error associated with agricultural produc-

11Using lower than median prices to value intermediate inputs increases our sample size. Ultimately, thereduction of intermediate prices to zero implies our measure of output is simply agricultural production inwhich case all output observations are non-negative. In this case our reallocation exercise implies an aggregateoutput loss of .2562 which is larger than our benchmark output loss. That is, dropping non-positive valueadded observations biases slightly down our reallocation gains.

10

tion and other agricultural inputs that we discuss next.

B.2.1 Potential Recall Bias for Agricultural Production and Hours

In rural settings the underreporting of agricultural production is a recurrent issue for sur-

vey data.12 There are two aspects of the Malawi ISA design that help mitigate and study

this issue. First, the ISA applies an intense data collection strategy to capture agricultural

production in which in many instances the survey provides internal consistency checks (e.g.,

households are asked total sales, and also sales by crop and by plot; the interviewer must

check in situ that the two sums coincide or otherwise re-interview). Second, the ISA not only

collects data on agricultural production but also close to “gold standard” data on consump-

tion that includes food consumption (in physical units) from own production with 7 days

recall.13 This provides a unique opportunity to externally validate agricultural production

using consumption data. In this context, a reassuring result is that in rural household-farms

that do not sell their agricultural production and have little or no consumption purchases

(i.e., about 50% of the entire rural sample), the reported agricultural production and the

reported consumption net of transfers imply very similar quantities, which suggests a small

scope for measurement error (from recall or elsewhere) in agricultural production, see de Ma-

galhaes et al. (2015).

Not only agricultural production is collected retrospectively, but also hours. To further

investigate the basis for potential recall bias in the collection of production and hours, we

note that in Malawi there is only one main harvest associated with the only rainy season.

The rainy season, which starts in late March/early April and goes through July, captures 93

percent of annual agricultural production.14 Hence, given that ISAs data collection rolls over

12 months, when households are asked about their last collected harvest for the rainy season,

12A similar issue arises in the measurement of income for self-employed households in both rich and poorcountries, see Deaton (1997).

13The 7 days recall consumption is found to be a good reliable compromise compared with more expensiveand intense consumption data collection that can be set as benchmark or “gold standard”, see the recentassessments in Beegle et al. (2010) and Gibson et al. (2015).

14The remaining production attained in the dry (Dimba) season and in terms of permanent crop. Further,the hours input is also more intensively used in the harvest season that in pre-harvest activities such asplanting, weeding and fertilizing.

11

this might happen right after harvest or several months after it (e.g., some household farms

asked in December might have collected their last harvest from the rainy season nine months

before).15 To assess the potential effects of recall on our reallocation results, we re-conduct

our entire analysis for only the households farms that report agricultural production shortly

after harvest. Our results are reported in Table B-1. We find that if we restrict our sample

study to those households that report agricultural production within four months after their

main harvest was collected, the output (productivity) gain is 3.7-fold which is slightly larger

than the 3.6-fold output gain in our benchmark specification. This finding suggests that our

results are robust to recall bias.

Table B-1: Robustness to Potential Output Recall Bias

Post-HarvestBenchmark Months

Output Gain 3.59 3.68

B.2.2 Proxying Agricultural Capital with an Asset Index

Our benchmark measure of agricultural capital (i.e., equipment and structures) uses the self-

reported selling price of each of the agricultural assets in production during the year under

study.16 This implies that part of the variance in agricultural capital across households might

be due to differences in the price of capital. If correctly measured, the selling price of capital

should capture the relative price of investment associated with that particular asset and its

depreciation since the year it was purchased.17 But the use of the selling price might be

limited if reported by households in areas where capital sales and purchases are infrequent

or in which markets for those assets are not fully developed.

We assess how much variation there is in agricultural capital across households when, as

opposed to our benchmark measure of agricultural capital, we use a physical asset index based

15For the case of Malawi ISAs 2010/11 there are actually 13 months, from March 2010 to March 2011.16We proxy the value of agricultural capital services by the value of the agricultural capital stock.17Under competitive markets the inverse of this price is productivity, i.e., the quality of capital.

12

solely on the ”yes/no” answers (i.e., dummy variables) about the ownership of agricultural

assets. The idea is that the use of simple ”yes/no” answers reduces potential measurement

error introduced by the households’ self-reported inference on the price of capital. Precisely,

we construct a standard asset index for agricultural capital following Filmer and Pritchett

(2001) and McKenzie (2005) as:

k =

nk∑k

βk

(xk − xkσk

),

where β′β = 1 with β = (β1, ..., βnk)′, and xk and σk are, respectively, the mean and standard

deviation of the asset xk. We include the full set of nk = 25 agricultural assets collected in

the Malawi ISA 2010 and specified in Appendix A.5. The agricultural capital index for a

farm i, i.e., ki =∑nk

k βk

(xi,k−xkσk

), is the principal component score for that farm.18

In terms agricultural capital dispersion, we find that the variance in logs of our benchmark

measure of agricultural capital (i.e., through the self-reported selling price) is 1.82, while the

variance associated with the asset index, k1, is 1.491. If instead of using only information

on whether a farm owns (and uses) a given agricultural asset, we use further information

on the quantity of assets of a given type used in production, we find that the variance

associated with this alternative asset index, k2, is 1.181. This conforms to the notion that

part of our benchmark agricultural capital measured through selling prices captures more

than differences in the amount of physical capital some of which may be quality differences.

It is possible, however, that some variation in our benchmark measure of capital cap-

tures not only genuine variation in the quality of capital, but also some noise in the data

collection or in the household’s assessments of the price of their own capital. To investigate

this, we re-conduct our reallocation exercise using the agricultural asset indexes described

above as proxies for agricultural capital services. Due to the smaller variation of the capital

indexes across households, the associated dispersion of farm productivity is larger under the

indexes than under our benchmark measure of capital. Indeed, with k1 the household farm

18Note that(xi,k−xk

σk

)has zero mean and variance λ, where λ is the largest eigenvalue of the correlation

matrix of x.

13

productivity displays a variance (in logs) of 1.607, and of 1.551 with k2, while this dispersion

is 1.435 in our benchmark. Further, while the correlation of our benchmark capital measure

with its associated household farm productivity is -.015 (non-significant), the correlation

of k1 and k2 with their associated household farm productivities are -.167 and -.089 (both

significant). That is, the asset indexes drive the allocation of capital further away from

the efficient allocation relative to our benchmark measure. This implies much larger output

gains of 17.5-fold under k1 and of 16-fold under k2 (versus 3.6-fold in our benchmark), see

Table B-2.

Table B-2: Robustness to Agricultural Capital Measured with an Asset Index

Asset Indexes

Benchmark k1 k2

Output Gain 3.59 17.54 16.00

Notes: The agricultural asset index k1 is constructed using ”yes/no” information for 25 types of agricultural

assets used in production. The agricultural asset index k2 further incorporates the quantity per agriculturalasset used in production.

B.2.3 Further Accuracy Assessments: Bootstrapping Our Household-Farms

We show that our results are robust to potential inaccuracies associated with sample size.

This is largely due to the fact that our sample size is relatively large. We find a nationwide

output gain 3.58-fold, a bootstrap median gain of 3.55-fold, and 5th and 95th bootstrap

percentiles of 4.07 and 3.11. These estimates are computed from 5,000 simulations obtained

from random draws with 100 percent replacement, that is, each simulation consists of a

sample of the same size as the original sample, and clustered at enumeration area—i.e., the

smallest geographical area available in the Malawi ISA. Our bootstrap results imply that our

reallocation findings are precise. We report in this Appendix the entire bootstrap distribution

of output (productivity) gains in Figure B-1. The histogram reveals a distribution of output

gains that are tight and close to normal with a slight skewness to the left. Precisely, with

a median of 3.55 gains, the tails of the output gains basically die out well before reaching a

factor of 2.5 gains from below while there is still some mass to the right of 4.5 gains from

14

above.

Figure B-1: Histogram of Output Gains: Bootstrapped Simulations

Notes: Results from 5,000 simulations clustered at enumeration area.

C Further Insights on the Role of Land Markets

An interesting aspect of our results is that the gains from reallocation are more than twice as

large for farms with no marketed land than for farms with some marketed land (purchased

or rented in). This result highlights the importance of the role of land markets for aggregate

output (and productivity). In this Appendix, we investigate further the differences between

the farms that operate marketed land and the farms that do not.

C.1 Distribution of Farm TFP by Marketed Land

One potential concern with our characterization of the output gains across farms with and

without marketed land is the possibility that access to marketed land is driven by farm

productivity. However, our data indicates very small differences in productivity between

farms with and without marketed land. These differences are too small compared to what it

would be if access to marketed land was purely driven by frictionless selection in farm TFP.

15

Even though on average farms with some marketed land are 25 percent more productive

than farms with no marketed land, the distribution of household-farm productivity is very

similar between the group of household-farms with marketed land compared to the group of

farms without marketed land, with respective variances of 1.19 and 1.17. Figure C-1 shows

the close overlap of these distributions. Nevertheless, we assess the gains from reallocation

holding constant farm-level TFP with fairly similar quantitative results.

Figure C-1: Density of Farm Productivity si (in logs) by Marketed Land, Malawi ISA 2010/11

Notes: Household-farm productivity si is measured using our benchmark production function, adjusting forrain ζi and land quality qi , yi = siζif(ki, qili) with f(ki, qili) = k.36i (qili)

.18, where yi is farm output, ki iscapital, and li is land. All variables have been logged. The sample is divided between those that operateland acquired in the market and those that do not.

C.2 Demographics, Human Capital, Female Empowerment, and Welfare

In terms of demographics, we focus on the characteristics of the household head, see Table C-

1. We find that 95.7% of farms with no marketed land are in rural areas, against 79.3%

of those farms with some marketed land, and 71.3% of those farms operating only land

obtained from the market. Household heads in farms with no marketed land tend to be

slightly older, an average of 43.3 years of age, although the oldest group of farmers are those

that purchased land, either untitled, 44.2 years, or titled, 44 years. The share of female heads

in the household-farm population decreases with marketed land, 19% of females operating

in farms with no marketed land whereas 16.4% in farms operating some marketed land.

Further, the proportion of migrants heads increases with marketed land, 19% of migrants

16

in farms operating no marketed land and 32.5% of migrants in farms operating marketed

land.19

Land markets are also related to human capital. Years of schooling of household heads

increase with marketed land, 4.9 years of schooling for in farms operating no marketed land,

7.1 years in farms operating some marketed land, and 7.7 years of schooling in farms that

only operate marketed land.

In terms of production, marketed land is associated with higher production (valued at

constant at-the-gate prices), higher agricultural value added, higher farm productivity si,

higher land productivity (i.e., yield), and higher capital productivity (i.e., capital-to-output

ratio). All these numbers are largely driven by farms with purchased land. For example, pro-

duction is 223% higher for farms with purchased titled land than for farms with no marketed

land, productivity is 47% higher, and capital productivity 12% higher. Interestingly, while

land quality qi also increases in farms with marketed land, on average by 42% compared with

the land in farms with no marketed land, we find that for farms with land purchased with

title, land quality decreases by 14% relative to the land quality of farms with no marketed

land.

Female empowerment is also considered a measure of development. We find that access

to land markets tends to enhance female empowerment in terms of female labor force partici-

pation and female wages. For households that operate only non-marketed land, female labor

force participation is 26.9%, whereas this figure raises to 43.8% for household that operate

some marketed land and 50.7% for households that only operate marketed land.20 Female

empowerment is also reflected in wages, as females that work in the market in households

that operate some marketed land earn wages that are 168% higher than wages of females in

household-farms that operate non-marketed land. Females wages in households that operate

only marketed land are 234%. The largest female labor force participation and wages are

observed in the household-farms that operate land purchased with a title, with a female

labor force participation of 55.5% and female wages 339% higher than female wages in farms

19Most migration occurs from rural to rural areas, see de Magalhaes and Santaeulalia-Llopis (2015).20Labor force participation is defined as participation in nonagricultural marketed activities including own

businesses, informal work, or working for a salary (including wages from other household farms).

17

with no marketed land.

Finally, land markets are also highly related to household welfare. We find that con-

sumption in adult-equivalent units (nondurable consumption and food consumption) also

increases with marketed land.21 In terms of nondurable consumption, households that op-

erate some marketed land consume 40% more per adult-equivalent than households with no

marketed land, 51% more in households operating only marketed land, and 76% more in

households operating land purchased with a title. In terms of food consumption the dif-

ferences are minor, partly because the basket composition of those with marketed land has

more weight towards non-food consumption. Consumption of maize per adult equivalent (in

Kg.) is higher in farms operating non-marketed land than farms operating land purchased

with a title.

To summarize, land markets are associated with other indicators of economic development

such as human capital, female empowerment, migration, and welfare.

C.3 Land Markets and Access to Credit and Other Markets

We interpret the facts that farm productivity is unrelated to land holdings and that the

capital to output ratio is roughly constant with farm productivity with the notion that

restrictions to land markets affect the allocation of capital, as also argued in de Soto (2000).

The first panel in Table C-2 shows how access to land markets relates to access to credit

markets. First, we find that there are not substantial differences in the self-reported need

for a loan across households with different access to land. We find that 63.8% of households

that do not operate any marketed land need a loan, and this number decreases slightly to

59.3% for households operating only marketed land, and to 57.7% for households operating

purchased land. Second, conditional on needing a loan, we do observe substantial differences

in loan applications rates with 27.7% of households with no marketed applying for a loan

versus 41.1% of households with some marketed land applying for a loan, that is, almost

a 50% higher rate of loan application for farms with marketed land than for farms with

21We use measures of annual deseasonlized consumption in de Magalhaes and Santaeulalia-Llopis (2015)which crucially include own production valued using consumption prices adjusted by region and season.

18

no marketed land.22 Third, there are also substantial differences across households with

different access to land markets in terms of the success rates of being granted the loan.

We find that 61.1% of households that apply for a loan and operate no marketed land

succeed in obtaining a loan, whereas the success rate for farms that only operate marketed

land is 69%. Note that since we observe which household-farms choose to apply for a loan

and those that choose not to, the success rates that we compute control for self-selection.

Putting together the application and success rates we find that the proportion of household

farms who need a loan and get it unambiguously increases with access to land markets.

This proportion is 16.9% for households that do not operate any marketed land, 26.2% for

households that operate some marketed land, and 27.8% for households that operate only

marketed land. That is, access to credit is associated with access to marketed land. Finally,

this relationship between land markets and credit is also present in terms of the amount

borrowed. Within households that succeed in getting a loan, household-farms that do not

operate any marketed land borrow on average USD68, whereas this figure is USD154.8 for

farms with some marketed land and USD194.6 for farms that operate only marketed land.

The second and third panel in Table C-2 display the relationship between land markets

and labor and intermediate inputs markets. Unlike the case of credit markets, for labor we

cannot control for self-selection (e.g., households are not asked about the willingness to hire

labor and its success in doing so). Nevertheless, in terms of the proportion of hired labor in

total farm hours we observe a clear positive relationship between access to land markets and

labor markets. For farms that operate only non-marketed land, the amount of hired labor

represents 4.6% of total farm hours while this figure increases to 15.9% for farms with some

marketed land, and to 19.5% for farms that only operate marketed land. The intermediate

inputs market on fertilizers and seeds portrays a similar picture. Household-farms that

operate some marketed land use 51% more fertilizers and 21% more seeds than farms that

operate only non-marketed land, and those with land purchased with a title use 156% more

fertilizers and 54% more seeds than farms that operate only non-marketed land. Finally, in

22Households that need a loan but self-select into the group that does not apply for a loan justify thischoice mostly by the fact that they do not know any lender and also because they believe their applicationwould be turned down.

19

terms of technology adoption on fertilizers, we barely observe any differences in the use of

organic fertilizer, i.e., low technology across groups with different access to land market, but

we do observe substantial differences in the use of inorganic fertilizer, i.e., high technology,

with household farms that operate some marketed land using 54% more inorganic fertilizer

than farms with no marketed land. Similarly, in terms of technology adoption on seeds

quality, access to land markets reduces the use of seeds from local maize, i.e, low technology,

and increases the use of hybrid maize, i.e., high technology.

To summarize, our evidence suggests that the access to land markets is positively related

with potentially more developed credit, labor and intermediate inputs markets. Further, in

the context of intermediate inputs, access to land markets is also related to the adoption of

better technologies such as inorganic fertilizer and hybrid maize seeds.

20

Table C-1: The Role of Land Markets: Household-Farm Demographics, Human Capital,Female Empowerment, and Welfare

Marketed Land Share Rented PurchasedHousehold-Farm Characteristics: 0% > 0% 100% Untit. Titled Untit. Titled

Demographics:B Rural (%) 95.7 79.3 71.3 70.6 81.5 87.3 77.3B Age 43.3 40.7 41.7 41.9 39.3 44.2 44.0B Females (%) 25.3 16.4 17.9 21.8 15.1 18.2 18.5B Migrants (%) 19.0 32.5 37.1 37.6 27.3 46.4 30.4

Schooling Years 4.86 7.11 7.72 7.65 6.96 5.99 7.07

Production (Index):B Agricultural Production 100 137 117 75 128 209 223B Agricultural Value Added 100 140 120 81 132 199 227B Productivity, si 100 125 118 102 130 125 146B Land Productivity: yi/li 100 107 105 108 110 106 107B Capital Productivity: yi/ki 100 107 103 78 118 106 113B Land Quality, qi 100 142 121 127 158 151 86

Female Empowerment:B Female Labor Force Part. (%) 26.9 43.8 50.7 51.0 41.7 21.7 55.5B Female Wages (Index) 100 268 334 310 239 156 439

Welfare:B Nondurable Exp. (USD) 534 749 808 772 707 722 942B Food Exp.(USD) 380 481 508 516 471 420 497B Maize Consumption (kg) 269 291 323 233 335 249 218

Obs. 5,962 1,189 746 215 682 126 97

Notes: The share of marketed land is defined from the household-farm level information on how land wasacquired, see Section ??. Nondurable expenditures, food expenditures and maize consumption (in Kg)are transformed into an adult equivalent units using the equivalence scales in Krueger and Perri (2006).Expenditures are in USD using the exchange rate in 2010, i.e., we divide Kwachas by 152.

21

Table C-2: The Role of Land Markets: Household-Farm Access to Credit and Other Markets

Marketed Land Share Rented PurchasedHousehold-Farm Characteristics: 0% > 0% 100% Untit. Titled Untit. Titled

Credit Market:B Need a Loan (%) 63.8 61.9 59.3 63.7 63.3 61.9 57.7B Application Rate (%) 27.7 41.1 40.4 28.5 45.4 40.7 36.0B Success Rates (%) 61.1 63.8 69.0 60.5 63.9 61.2 66.6B Amount Borrowed (USD) 68.0 154.8 194.6 201.1 119.4 289.0 145.3

Labor Market:B Hired Labor (% of Farm Hours) 4.6 15.9 19.5 15.9 15.5 10.7 17.6

Intermediate Inputs Market/Technoloty Adoption:

B Fertilizer:Total Fertilizer (USD) 64.5 97.9 86.7 62.3 94.6 119.1 165.2Low Tech: Organic Fert. (USD) 3.5 3.6 3.7 1.8 3.7 5.4 6.1High Tech: Inorganic Fert. (USD) 60.8 94.2 84.8 60.4 90.7 113.7 158.9

B Seeds:Total Seeds (USD) 15.6 19.0 15.2 13.8 18.4 21.9 24.1Low Tech: Local Maize (USD) 3.9 3.8 3.4 3.7 3.7 3.7 4.0High Tech: Hybrid Maize (USD) 10.5 11.7 11.4 10.3 12.0 12.3 12.6

Number of Observations 5,962 1,189 746 215 682 126 97

Notes: The share of marketed land is defined from the household-farm level information on how land wasacquired, see Section ??. Expenditures are in USD using exchange rate in 2010, i.e., we divide Kwachas by152.

22

D Discussion on Farm Size and Robustness

D.1 Farm Size

Figure D-1 documents the relationship between land productivity (yield) by farm size in our

data. Land productivity is slightly declining with farm size, which conforms with the finding

in the inverse yield-to-size literature. For instance, the land productivity of the largest 10

percent of farms relative to the smallest 10 percent of farms is 0.34 (a 3-fold yield gap).

Figure D-1: Land Productivity by Farm Size, Malawi ISA 2010/11

Notes: Household-farm land size is the sum of all plots per household farm and marginal product of land(MPL) is constructed using our benchmark production function, i.e., MPL = .18 yi/li. All variables havebeen logged.

D.2 Human Capital and Specific Skills

Table D-1 reports robustness results with respect to human capital and specific terrain skills.

In each case, we perform the efficient reallocation exercise but within each education and skill

type. With respect to schooling levels of the household head, the output gains are large at all

schooling levels and roughly constant across the human capital spectrum. For no schooling

farmers the output loss is 0.31 and between 0.26 to 0.28 for farmers with primary and more

than primary schooling. Output losses may also be related to specific skills such as how to

operate farms in different terrain roughness. Similarly to our findings for human capital, the

output losses are large even within specific terrain roughness, 0.27 for high altitude plains

23

Table D-1: Robustness to Human Capital and Specific Skills

Output Loss within Schooling GroupsSchooling groups: No Schooling Dropouts Primary > Primary

Output gain 3.22 3.33 3.81 3.58

Output Loss within Terrain-specific SkillsTerrain-roughness High Altitude Low Mid-Altitude Mid-AltitudeSkills: Plains Plateaus Plateaus Mountains

Output gain 3.69 3.71 3.03 2.74

Notes: The table reports robustness exercises within schooling groups and within types of terrain roughnessfor which specific skills might be required. The sample distribution across education groups defined as highesteducational degree completed is: no schooling 25%, primary school dropouts 45%, primary school graduates23%, and more than a primary school degree 7%.

and 0.36 for mid-altitude mountains.

D.3 Productivity Net of Transitory Health and Other Shocks

We compute the output losses associated with farm productivity measures that are net

of health and other transitory shocks. To do so we regress our benchmark measure of

productivity on a set of controls that capture potential sources of transitory shocks. Then,

we use the residual from each of these regressions as the new measure of productivity. We

find that the output gains with these alternative measures of productivity are fairly similar

to our benchmark measure of farm TFP.

We describe in detail the variables that we use to adjust farm productivity for transitory

shocks. We use six specifications that differ in the number of controls. The implications

for the dispersion of farm productivity are reported in the first row of Table D-2. First,

we include measures of health risk that include illness and injuries affecting the household

head and present in the past two weeks.23 These illnesses/injuries include, for example,

malaria, diarrhea, stomach ache, respiratory problems, dental problems, eye problems, burns,

23These data are available at the individual level for each household member. While we focus our analysison the household head, we find that the introduction of the health status of the rest of household membersdoes not change our results.

24

fractures, wounds, and poisoning. Note that these illnesses/injuries potentially last more

than the two weeks reference period and can affect annual productivity.24 We also include

hospitalizations over the past 12 months at both medical facilities (i.e., hospitals or health

clinics) as well as at traditional healers dwellings. We also include health expenditures over

the past 12 months either for prevention or treatment. Prevention predominantly includes

regular checkups and doctor visits and treatment includes the costs of overnight stays at

medical facilities or traditional healer dwellings, transportations costs, and food costs at those

facilities. In terms of prevention, we also include the presence of bed nets and insecticide use

to prevent malaria infection.25 The result of netting health shocks from our farm productivity

measure reduces the variance of productivity from our benchmark of 1.435 to 1.4226.

Second, the data records the deaths of any member of the household over the past two

years. Information on the age of the deceased is available. We introduce death shocks as

the deaths suffered by the household over the past 24 months for all individuals above 11

years old.26 Controlling for death shocks, we find this further reduces the variance of farm

productivity but not in any significant manner to 1.4225.

Third, the households also report whether in the last 12 months they have been faced

with a situation in which they did not have enough food to feed the household. We find that

this food insecurity episodes occur for 48.7% of all households. The main reason for food

insecurity is no food stocks as a consequence of drought/poor rain (33.55%), lack of farm

input (3.37%), and small land size (10.65%). Netting our measure of productivity from food

security risks drops the variance of farm productivity further to 1.4137.

Finally, we also introduce additional controls that potentially affect productivity through

the course of the year such as marital status and the performance of other income sources

(i.e., labor market income, business income, capital income, and net transfers), or through

24We are not interested in including chronic diseases as we consider them part of the permanent componentof farm productivity.

25We find that 19% of all households heads suffer some illness/injury present in the past 2 weeks in ruralareas and 15% in rural areas, but only around 17% of households are treated for those illnesses/injuries.Also, 3% of household heads are hospitalized in the past 12 months in both rural and urban areas. However,health expenditures barely represent 1% of total consumption in both rural and urban areas.

26Most deaths are infant deaths of children below 1 year of age (17.81%), below 2 years of age 25.84%,and below 5 years of age 39.42.

25

their transitory nature previous to a reallocation policy such as the household distance to

roads, population centers, input suppliers, auctions, markets, and border posts. We find

that the sequential introduction of each of these additional controls further reduces, though

not in substantial manner, the variance in farm productivity to a final variance of 1.4073.

These small reductions in the variance of farm productivity do not affect substantially the

magnitude of the output losses reported, in fact, the output losses are slightly higher than

in the baseline measure of productivity that abstracts from these potential transitory effects

on farm output and productivity.

26

Table D-2: Productivity Net of Health and Other Transitory Shocks

Household-Farm Productivity:Specs. Net of Further Transitory Shocks

(1) (2) (3) (4) (5) (6)

var(si) 1.4226 1.4225 1.4137 1.4119 1.4092 1.4073

Illness/Injury (last 2 wks) X X X X X XHospitalizations (last 12m.)

B Formal: X X X X X XB Traditional Healer: X X X X X X

Health Expend. (USD) (last 12m.)B Prevention X X X X X XB Treatment X X X X X X

Malaria (last 12m.)B Nets X X X X X XB Insecticide X X X X X X

Death (last 2 years) – X X X X X

Food Security (last 12m.) : – – X X X X

Marital Status – – – X X X

Distance to:Roads – – – – X XPop. Center – – – – X XInput Supplier (Admarc) – – – – X XAuction – – – – X XMarket (Boma) – – – – X XBorder Post – – – – X X

Other Income Sources:Labor – – – – – XCapital – – – – – XBusiness – – – – – XLivestock Produce – – – – – XNet Transfers – – – – – X

27

References

Beegle, K., DeWeerdt, J., Friedman, J., and Gibson, J. (2010). Methods of householdconsumption measurement through surveys: Experimental results from tanzania. PolicyResearch Working Paper, The World Bank.

Carletto, C., Savastano, S., and Zezza, A. (2013). Fact or Artifact: The Impact of Mea-surement Errors on the Farm SizeProductivity Relationship. Journal of DevelopmentEconomics, 103(C):254–261.

de Magalhaes, L., Koh, D., and Santaeulalia-Llopis, R. (2015). Consumption and Expen-ditures in Poor Countries: A Life With No Humps and Rural-Urban Insurance Traps toGrowth. Unpublished Manuscript Bristol University and Washington University in St.Louis.

de Magalhaes, L. and Santaeulalia-Llopis, R. (2015). The Consumption, Income, andWealth of the Poorest: Cross-Sectional Facts of Rural and Urban Sub-Saharan Africa forMacroeconomists. World Bank Policy Research Working Paper, WPS7337.

de Soto, H. (2000). The Mystery of Capital: Why Capitalism Triumphs in the West andFails Everywhere Else. Basic Books. New York.

Deaton, A. (1997). The Analysis of Household Surveys. A Microeconometric Approach toDevelopment Policy. The Johns Hopkins University Press.

Filmer, D. and Pritchett, L. H. (2001). Estimating Wealth Effects Without ExpenditureData-Or Tears: An Application to Educational Enrollments in States of India . Demog-raphy, 38(1):115–132.

Gibson, J., Beegle, K., DeWeerdt, J., and Friedman, J. (2015). What Does Variation in Sur-vey Design Reveal about the Nature of Measurement Errors in Household Consumption?Oxoford Bulletin of Economics and Statistics, 77(3):466–474.

Krueger, D. and Perri, F. (2006). Does income inequality lead to consumption inequality?evidence and theory. Review of Economic Studies, 73:163–193.

McKenzie, D. (2005). Measuring Inequality with Asset Indicators. Journal of PopulationEconomics, 18(2):229–260.

28