Irrigation and rice production evidence in vietnam

Further analyses are conducted to figure out the differences in impact on rice yield among water sources and explore whether effect of using canal for irrigation varies across locations

UNIVERSITY OF ECONOMICS ERASMUS UNVERSITY ROTTERDAM

HO CHI MINH CITY INSTITUTE OF SOCIAL STUDIES

VIETNAM – THE NETHERLANDS PROGRAMME FOR M.A IN DEVELOPMENT ECONOMICS

IRRIGATION AND RICE PRODUCTION:

MASTER OF ARTS IN DEVELOPMENT ECONOMICS

HO CHI MINH CITY, December 2016

UNIVERSITY OF ECONOMICS INSTITUTE OF SOCIAL STUDIES

HO CHI MINH CITY THE HAGUE

VIETNAM THE NETHERLANDS

VIETNAM - NETHERLANDS PROGRAMME FOR M.A IN DEVELOPMENT ECONOMICS

IRRIGATION AND RICE PRODUCTION:

EVIDENCE IN VIETNAM

A thesis submitted in partial fulfilment of the requirements for the degree of

MASTER OF ARTS IN DEVELOPMENT ECONOMICS

By

LE HUU NHAT QUANG

Academic Supervisor:

DR LE VAN CHON

ABSTRACT

This thesis investigates the relationship between irrigation and rice production in the case of Vietnam Further analyses are conducted to figure out the differences in impact on rice yield among water sources and explore whether effect of using canal for irrigation varies across locations on commune canal system The employed dataset is VARHS 2014, which contains information on rice production, inputs use at household level, characteristics of plot managers, quality and irrigation condition of plots, and commune data on infrastructure The thesis uses a cross-section data model with cluster-specific fixed effect and cluster-robust standard error Research results show that irrigation plays an important role in rice production However effect of irrigation varies across water sources, and only effects of using water from canal and river or spring for irrigation are significant Benefit of irrigation

by canal to rice production is not higher than that of using water from river or spring Moreover, there is inequality in water distribution of commune canal system

Keywords: irrigation, canal, rice production, Vietnam

ACKNOWLEDGEMENT

Wholeheartedly, I would like to thank Dr Le Van Chon, my supervisor, for being patient, giving me useful advice and frank comments Without him, I certainly cannot complete my thesis

I am also extremely grateful to Dr Truong Dang Thuy and Dr Pham Khanh Nam for their advice when I get stuck in building research ideas

In addition, I really appreciate all the hard work, advice and encouragement that Mr Do Huu Luat did to help me during the time I doing my thesis I would like to say thank to Mr Le Van Thang, Mr Nguyen Son Kien, Mrs Hoang Quynh Trang and Mrs Vu Thi Thuong for encouraging and helping me to complete my thesis

Finally, from the bottom of my hearth, I would like to say thank to my parents for being there when I need them

ABBREVIATIONS

BLUE Best linear unbiased estimator

CRVM Cluster-robust variance matrix

CSFE Cluster-specific fixed effect

CSRE Cluster-specific random effect

FGLS Feasible Generalized Least Square

GLS Generalized Least Square

OLS Ordinary Least Square

VARHS Vietnam Access to Resources Household Survey VIF Variance Inflation Factor

CONTENTS

CHAPTER 1: INTRODUCTION 8

1.1.PROBLEM STATEMENT 8

1.2.RESEARCH OBJECTIVES 9

1.3.RESEARCH QUESTIONS 9

1.4.SCOPE OF THE RESEARCH 9

1.5.THESIS’S STRUCTURE 10

CHAPTER 2: LITERATURE REVIEW 11

2.1.DEFINITION OF IRRIGATION 11

2.2.THEORETICAL LITERATURE 11

2.2.1 Role of irrigation in agriculture 11

2.2.2 Properties of rice plant and water need for rice production 13

2.2.3 Water sources 15

2.2.4 Analysis model 15

2.2.5 Cobb-Douglas production function 17

2.2.6 Other production functions 19

2.3.EMPIRICAL LITERATURE 19

2.4.HYPOTHESIS TESTING 22

CHAPTER 3: DATA AND METHODOLOGY 24

3.1.DATA SOURCE 24

3.2.MODEL SPECIFICATION 24

3.2.1 Building model 24

3.2.2 Constructing variables 25

3.3.ESTIMATION STRATEGY 30

3.3.1 Cluster-specific random effect model 33

3.3.2 Cluster-specific fixed effect model 35

3.3.3 Choosing between CSRE and CSFE 36

CHAPTER 4: EMPIRICAL STUDY IN IMPACT OF IRRIGATON ON RICE PRODUCTION 37

4.1.VIETNAM RICE PRODUCTION AND PUBLIC IRRIGATION SYSTEM 37

4.2.RICE PRODUCTION AND IRRIGATION OF FARMERS IN VARHS2014 39

4.3.EFFECT OF IRRIGATION ON RICE PRODUCTION IN RURAL VIETNAM 42

CHAPTER 5: CONCLUSION AND IMPLICATION 48

5.1.MAIN FINDINGS 48

5.2.POLICY IMPLICATION 49

5.3.LIMITATION 49

REFERENCES 51

APPENDICES 57

TABLE OF FIGURES

Figure 2.1: Determinants of irrigation frequency 13

Figure 2.2: Analytical framework 23

Figure 4.1: Rice productivity in 2014 of some countries 37

Figure 4.2: Vietnam's annual rice production 38

Figure 4.3: Vietnam's annual harvested area 38

LIST OF TABLES

Table 3.1: Variables and expected sign 28

Table 4.1: Area cultivating rice, ratio of irrigation and ratio of polyculture by number of season growing rice 40

Table 4.2: Average rice productivity by irrigation and water sources 41

Table 4.3: Test results of choosing among OLS, CSRE, and CSFE 42

Table 4.4: Cluster-specific fixed effect regression of log household's rice production 47

CHAPTER 1: INTRODUCTION

1.1 Problem statement

Rice can be grown in a wide range of weather conditions, so that many countries have grown rice, especially in Asia with about 90% rice production of the world (Maclean, Hardy, & Hettel, 2013) In Vietnam, rice is one of the main crops, ratio of land cultivating rice to total land growing cereals is up to 86% (Vietnam GSO, 2015) In 2014, Vietnam was ranked fifth

in the world for rice production, after China, India, Indonesia, and Bangladesh (FAOSTAT) From 1990 to 2014, rice productivity in Vietnam increased by nearly 100% (FAOSTAT) This success is contributed by many factors such as application of new varieties, technology and extension of irrigation system

Until 2011, Vietnam has 254180 km of canal serving for irrigation of 85.5% land cultivating rice (Ha, Nguyen & Nguyen, 2015) However, it also shows many disadvantages in operation, such as weak management, lacking of maintaining and dredging, and low quality

of water (Tran, 2016) This raises the questions about the role of irrigation, and impact of irrigation system in Vietnam on rice production There are some papers related to these issues which are conducted in Vietnam (Walle, 2003; Biltonen, Hussain & Tuan, 2003; Ut, Hossain & Janaiah, 2000), but these studies use dataset in the stage of 1991-2000 Meanwhile, from 2000 to nowadays there have been many changes in Vietnam irrigation system and rice production

The role of irrigation on agriculture and poverty alleviation has been research objective of many economists A large number of studies were conducted in many countries Almost research results show that irrigation helps to increase crops’ productivity and is a solution to reduce poverty (Hussain & Hanjra, 2004) Nonetheless, some of them suggest that impact of irrigation varies across system and even across locations on canal system (Biltonen, Hussain

& Tuan, 2003; Jin, Jansen and Muraoka, 2012; Hussain et al., 2006; Hussain et al., 2004) Moreover, these results are different among countries and regions Hence, it is necessary to investigate the effect of irrigation system on rice production in the case of Vietnam In this thesis, employed dataset is VARHS 2014, which is the latest secondary data and provides very specific information related to agricultural operation of households in Vietnam rural

1.2 Research objectives

This thesis focuses on investigating the impact of irrigation on annual rice production, then compare the effect of using water from canal for irrigation on annual rice production with those of other water sources Finally, the effects of different positions of commune canal system will be assessed

1.3 Research questions

In order to clarify these objectives, corresponding research questions are given:

(1) Does irrigation have a positive effect on annual rice production?

(2) How does impact of irrigation on rice production vary across water sources and locations on commune canal system?

1.4 Scope of the research

The dataset used in this paper is Vietnam Access to Resources Household Survey 2014 (VARHS 2014) This survey is conducted in 12 provinces which include 5 provinces in Northern, 3 provinces in Central Coastal, 3 provinces in Central Highlands, and 1 province in Mekong River Delta For each province, some communes are chosen for survey The data consists of 3648 households with information on many aspects focusing on rural development Information related to agriculture is relatively specific, which includes characteristics of plots, type of plant, inputs, outputs, and irrigation The survey also collected information on household members and communes This paper employs data at household level of farmers who grow rice

In methodology, a Cobb-Douglas production function is used as logarithm form for investigating the impact of irrigation on rice production To estimate the model, the cluster specific fixed effect at commune level with clustering standard error is employed In addition, in order to check the robustness, OLS and FGLS estimation with cluster-robust standard error are conducted Although the thesis can control unobserved variables a commune level, heteroskedasticity and correlation among error terms within commune, it has some limitations, such as endogeneity of input choice, potential endogeneity of irrigation variables, error measurement and omission of variables

1.5 Thesis’s structure

Literature review is presented in the next section The third section is data and methodology The fourth section will be an overview of Vietnam’s circumstance in rice production, irrigation system and impact of irrigation on rice production Finally, the fifth section is main findings, policy implication and limitations

CHAPTER 2: LITERATURE REVIEW

2.1 Definition of irrigation

“Irrigation is the controlled application of water to arable lands in order to supply crops with the water requirements not satisfied by natural precipitation” (United States Department of Agriculture, 1991) Accordingly, irrigated land is land which is supplied water from many sources besides rain or snow Non-irrigated land is called “rainfed land” The term “rainfed agriculture” is also used for farming approach only depends on direct rainfall The main role

of irrigation to provide needed water for plants in their growth stages Furthermore, irrigation has other uses such as controlling weed and avoiding soil consolidation

There are many sources of water can be used for irrigating, they include surface water and groundwater For public irrigation system, canals are built to connect farm with water source, for example, river, spring and lake Many dams are also built to keeping water for dry season

In terms of private irrigation, there is a wide diversity of irrigation methods that farmers can employ such as flood, drip, center-pivot, and sprinkler Methods’ efficiency are various, but choosing an appropriate irrigation system also rely on cost, technique, type of plants and scale

2.2 Theoretical literature

2.2.1 Role of irrigation in agriculture

Initially, it is necessary to discover the relationship among soil, water and plant Water plays

a very important role in plant growth, which can be listed as four main functions: constituent, solvent, reactant, and maintenance of turgidity (Kramer & Boyer, 1995) Accordingly, firstly, water constitutes a vast of plant’s mass, for example, 80 to 90 percent of herbaceous plants and 50 percent of woody plants Secondly, in plant, water acts as the solvent to transport air, nutrients, and organic compounds from cell to cell, part to part Thirdly, water is a reactant in many biochemical reactions of metabolism such as photosynthesis and hydrolysis Finally, water helps plant in maintaining the turgor which is important for cell growth

Plants extract water from soil through their root The process which drives water movement

in plants is transpiration which is the escape of water as gas at leaves and stems (Brouwer & Heibloem, 1986) In order to absorb carbon dioxide from the air for photosynthesis in

daytime, stomata in leaf is constantly opened during this period This promotes water evaporation at leaves, and consequently, plants have to take water steadily from the soil Approximately 95 percent of water which is absorbed by plants goes outside through transpiration, only 5 percent is kept for carbohydrates production The level of transpiration depends on weather conditions (temperature, humidity, sunshine, and wind speed), type of plant, and plant’s growth stage

Besides the function of storing nutrients, soil holds water for plants (United States Department of Agriculture, 1991) This indicates that water, soil and plants have close relations Water holding capacity of soil relies on its chemical and physical characteristics The amount of water which soil can store determines how long the plants have enough water

to use until the next irrigation or rainfall In other words, characteristics of soil determine irrigation frequency In addition, most of water is evaporated at the soil surface, which mainly depends on weather, so that weather is another factor of irrigation frequency

Plants need water for transpiration, soil supplies water to plants, but simultaneously soil loses water through evaporation Therefore, plant’s water need is the water demand for both transpiration and evaporation, and it is called evapotranspiration (Brouwer & Heibloem, 1986) Although in some areas, rainfall in rainy season is adequate for plant growth, drought may occur during the year In low rainfall areas, many types of plants with high water need cannot produce high yield, and even die if water supply only relies on rainfall This indicates that it is necessary to provide enough and in time water to plants to increase crop yield Since agriculture purpose is producing profitable crop, irrigation is an important solution to reduce weather dependence, especially in arid areas (United States Department of Agriculture, 1991)

The determinants of irrigation frequency (to provide enough water for plant’s growth) can be summarized as below:

Figure 2.1: Determinants of irrigation frequency

Source: summarized by author from Brouwer and Heibloem (1986) and United States Department of Agriculture (1991)

In terms of economic perspective, irrigation positively impacts on annual crop yield through increasing output per season and the number of cropping seasons (Hussain & Hanjra, 2004) This appears to be right for annual crops because it can be grown more than one season per year In dry season of some areas, rainfall is too little to cultivate crop without irrigation, land

in this period is left fallow or used for growing other plants which need less water In contrast, if land is irrigated sufficient water, the number of cultivated seasons can be two or even three per year for annual crops

2.2.2 Properties of rice plant and water need for rice production

Rice plant is classified as genus Oryza which include around 23 species There are two species of Oryza which are commonly grown for providing food They are Oryza savita and Oryza glaberrima, which are also known as Asian rice and African rice respectively At the present time, Oryza savita is more popular than Oryza glaberrima since it has higher productivity (Wopereis et al, 2008) Asian rice species can be divided into two group indica

IRRIGATION

FREQUENCY

Plant's water need

Transpiration

Weather conditionsCrop type

Growth stage

of crop

Evaporation

Weather conditions

Soil's properties

Frequency and amount of rainfall

Climate

and japonica Indica is characterized by having non-sticky, long and think grains, while japonica has sticky, short and round grains

The life cycle of rice plant is likely to be divided into three broad stages which are vegetative stage, reproductive stage, and maturity stage (Wopereis et al, 2008) Durations of stages are different Reproductive stage and maturity stage have a fixed duration which are 30-35 days and 30 days respectively Meanwhile the length of vegetative stage is various among varieties and day-length, so that the length of rice’s life cycle depends on duration of vegetative stage Commonly, rice’s life cycle can last from 90 to 180 days or 3-6 months, so rice is annual crop

Rice farming systems can be classified into five groups according to water environment, they are irrigated lowland rice, rainfed lowland rice, deepwater rice, floating rice, and upland rice (Maclean, Hardy, & Hettel, 2013) Particularly, lowland rice is characterized that rice fields are bunded and puddled For irrigated lowland rice, farmers proactively supply water to fields and maintain five to ten centimeters of water, while rainfed lowland rice relies only on rainfall Deep water and floating rice cropping systems are usually applied for flood prone areas Flood in these areas regularly happens with deep flooding for more than ten days and

is not controlled Upland rice is cultivated in dryland, uses only rainfall and is not puddled Two of five rice cultivating systems which are commonly employed are irrigated and rainfed lowland rice, and irrigated lowland rice gives the highest productivity (Maclean, Hardy, & Hettel, 2013) Indeed, over 90 percent of rice produced in the world is cultivated under irrigated and rainfed lowland systems All around the world, total cultivated area of irrigated lowland rice and rainfed lowland rice are approximately 93 million ha and 52 million ha respectively Meanwhile, they respectively supply 75% and 19% of global rice production Although area of irrigated lowland rice is nearly twice more than that of rainfed lowland rice,

it provides rice yield almost four times higher than rainfed lowland rice does

Rice cultivation needs a lot of water, especially for output maximizing purpose Accordingly, the amount of water is needed for rice plants to produce one kilogram of rough rice varies from 500 to 1000 litters, which depends on rice farming systems and is calculated for only transpiration (Haefele et al, 2009) The figure will be much larger if the evaporation is also accounted for, it is 1432 litter on average for one kilogram of produced rough rice (Bouman,

2009) ,and changes from 625 to 1667 litters in lowland rice fields (Zwart & Bastiaanssen, 2004) Meanwhile it is just 900 to 1150 litters for maize on average (Falkenmark & Rockström, 2004; Chapagain & Hoekstra, 2004) In addition, water can be lost through seeping through bunds and percolating down to deeper soil layers Hence, in total a rice field can take from 800 to 5000 litters (2500 litters in common) of water to produce one kilogram

of rough rice, and this amount of water is two to three times higher than that of other major cereals (Bouman, 2009)

2.2.3 Water sources

As mention above, water demand in rice production is great so that quality of water sources for irrigation also play a very important role Fields are near river, spring, lake and pond can use directly water from these sources For other fields, ground water or canal water will be used Since almost fields cannot take water directly from river or lake and ground water is limited, canal system is essential to provide water to crop Water level on river or lake usually largely changes periodically and sometimes is unpredictable Meanwhile, water supplied by canal varies across seasons less than natural sources such as river, lake and well

In some areas canal also is used for drainage (Harris, 2006) However, efficiency of canal relies greatly on quality of canal system and management Moreover, the difference in water distribution among head end, middle and tail end canal may affect crop productivity of fields

in these areas (Hussain & Hanjra, 2004)

In agriculture production, inputs consist of land, labor, seed, fertilizer, pesticide, machinery, and others There are many specific forms of production function such as linear form, Cobb-Douglas production function, and Leontief production function

Linear form of production function:

𝑌 = 𝑎 + 𝑏𝑋1+ 𝑐𝑋2+ 𝑑𝑋3+ ⋯ Cobb-Douglas production function:

𝑌 = 𝑎𝑋1𝑏𝑋2𝑐𝑋3𝑑… Leontief production function is employed in the assumption that inputs are combined in a fixed ratio Accordingly, an increase in one input without respective increase in other inputs will not lead to change in output However, this is not appropriate in agriculture production The form of Leontief production function is as below:

𝑌 = 𝑚𝑖𝑛⁡(𝑎𝑋1, 𝑏𝑋2, 𝑐𝑋3, … )

Profit function

Supposing that a firm in a perfectly competitive market has a production function as:

𝑌 = 𝑓(𝑉1, … , 𝑉𝑚; 𝐹1, … , 𝐹𝑛) Where Vi is the ith variable input, Fj is the jth fixed input In short term, profit of firm is:

p is nominal price of output

ri∗ is price of the ith variable input

ri is normalized price of the ith variable input, which equals ri∗/p

V, F, r’ are vectors of Vi’s, Fj’s and ri’s respectively

Suppose that the firm will choose an optimized bundle of variable inputs V to maximize its short-term profit P with a given set of p, r* and F This indicates that profit function is a

function of p, r* and F Let that optimized set of variable inputs be V*, then the profit function can be defined as:

𝜋 = 𝛱(𝑝, 𝑟∗, 𝐹) = 𝑝[𝑓(𝑉∗, 𝐹) − 𝑟′𝑉∗] (1)

Cost function

Suppose a firm can produce an output set Y, and for an output y in Y the firm can use many sets of inputs X(y) Market price of inputs that the firm has to face is vector r = (r1, r2, r3,…) corresponding to vector of inputs x = (x1, x2, x3,…) Hence the cost that firm pays for a set of input x is c = r.x = r1x1 + r2x2 + r3x3 + … Supposing that in order to produce an output y, the firm will choose a set of inputs x which minimizes its cost c, given inputs’ price This shows that cost function is a function of output and inputs’ price Then the cost function can be written as:

𝑐 = 𝐶(𝑦, 𝑟) = 𝑚𝑖𝑛{𝑟 𝑥|𝑥 ∈ 𝑋(𝑦)} (2)

Production function is commonly used because of its simplification, but it suffers from endogeneity of input choice which can be solved by using profit function or cost function (Quisumbing, 1996) However, unlike production function, profit function and cost function require more information, especially price of inputs and outputs This is a disadvantage of these functions, since there are few datasets which have sufficient information Because the dataset VARHS 2014 does not contain adequate price of all inputs, so that this thesis can just employ production function

2.2.5 Cobb-Douglas production function

The Cobb-Douglas production function was first introduced in 1928 by Cobb and Douglas (1928) The original Cobb-Douglas production function has only two inputs which are capital and labor, and has assumption of constant returns to scale The function can be written as:

𝑌 = 𝐴𝐿𝛼𝐾1−𝛼Where Y is output, L is labor and K is capital

1 Lau (1976)

2 Fuss & McFadden (2014)

Since the assumptions of the original Cobb-Douglas production function are not appropriate

in practice, many generalizations have been suggested which are called “Cobb-Douglas type

of function” Two of them are relaxing the assumption of constant returns to scale and the assumption of the number of inputs (Debertin, 2012)

For relaxing assumption of constant returns to scale, the sum of parameters on the inputs is not necessary to equal 1

𝑌 = 𝐴𝑋1𝛼1𝑋2𝛼2

Where 𝛼1+ 𝛼2 can be any number

As the number of inputs is concerned, the function is expanded with more than two inputs:

𝑌 = 𝐴𝑋1𝛼1𝑋2𝛼2𝑋3𝛼3𝑋4𝛼4… This function is more appropriate for analyzing agriculture production because there are many important inputs besides capital and labor, such as seed, fertilizer and pesticide

Debertin (2012) suggests a number of properties of Cobb-Douglas type of function:

(1) Homogeneity in degree of sum of 𝛼𝑖

(2) 𝛼𝑖 is partial elasticity of production for to 𝑋𝑖

(3) Output equals zero when there is at least one input is not used

(4) When increasing an input Xi, output will increase at the level corresponding to the parameter on that input 𝛼𝑖 The value of 𝛼𝑖 can be smaller, equal to or greater than 1

(5) With each set of parameters 𝛼𝑖, the model can only express one production stage

(6) If all 𝛼𝑖 are smaller than 1, there is usually a point where profit is globally maximized with finite inputs X

In order to estimate Cobb-Douglas production function, it can be transformed into logarithmic form or Translog form which correspond to first order and second order Taylor’s series expansion

The logarithmic form of Cobb-Douglas production function:

𝑙𝑛𝑌 = 𝑙𝑛𝐴 + 𝛼1𝑙𝑛𝑋1+ 𝛼2𝑙𝑛𝑋2 + 𝛼3𝑙𝑛𝑋3+ ⋯ = 𝑙𝑛𝐴 + ∑ 𝛼𝑖𝑙𝑛𝑋𝑖

The Translog form of Cobb-Douglas production function:

𝑙𝑛𝑌 = 𝑙𝑛𝐴 + ∑ 𝛼𝑖𝑙𝑛𝑋𝑖+ ∑ 𝛼𝑖𝑗𝑙𝑛𝑋i𝑙𝑛𝑋jThe Translog form gives more flexibility to Cobb-Douglas production function since it accounts for relations among inputs use (Quisumbing, 1996) Nonetheless, similarly to Cobb-Douglas function, it can describe only one production stage (Debertin, 2012) In addition, one of problems of estimating Translog form is possibility of collinearity because of the large number of variables, especially when there are many types of input (Pavelescu, 2011)

2.2.6 Other production functions

Besides Cobb-Douglas production function, Debertin (2012) shows some other production functions are also used in agricultural economics They are Spillman function, transcendental production function, Cobb-Douglas with variable input elasticity, de Janvry Modification, and polynomial forms Spillman function was suggested in 1923-1924, but it has been rarely used since Cobb-Douglas was introduced Except Spillman function and polynomial forms, other production functions were developed from Cobb-Douglas production function to solve the problem of constant elasticity of input Since the parameters on inputs of Cobb-Douglas function are constant, this function cannot represent three stages of production which are increasing in average product, decreasing in average product and negative marginal returns Polynomial forms also have elasticity of input which changes with value of that input, but it

is not similar to Cobb-Douglas function The general disadvantage of these production functions, in comparison to Cobb-Douglas function, is more sophisticated with more parameters need to be estimated

2.3 Empirical literature

There are many papers investigating the effect of irrigation on poverty A review of empirical studies conducted by Hussain and Hanjra (2004) shows that the relationship between irrigation and poverty are very strong Irrigation impacts on poverty through both direct and indirect paths In direct path, poor farmers can get benefit from irrigation through higher crop productivity, higher farming profit, choosing higher value crops, and participating in the market In long term, benefit of irrigation will spread throughout the poor people who do not have land

Some researches attempt to use data at macro level to analyze the impact of irrigation on agriculture growth and poverty Indeed, Bhattarai and Narayanamoorthy (2004) use macro data about agricultural production of 14 states of India from 1970 to 1994 to investigate role

of irrigation on Total factor productivity (TFP) and poverty alleviation The results show that irrigation has a positive and significant effect on TFP growth and play an important role in explaining poverty reduction They also suggest that efficiency of irrigation system strongly alters the result of poverty reduction program Butzer, Mundlak & Larson (2002) explore factors of agriculture growth in three Asian countries Each country is separately considered They are Indonesia in the period from 1971 to 1998, Philippines in stage of 1961-1998 and Thailand form 1971 to 1995 A Cobb-Douglas production function is employed with regressand is log of value added The study indicates that the contribution of irrigation to agricultural growth is from 10% to 16% Nevertheless, Walle and Gunewardena (2001) suggest that investigating irrigation’s benefit by using macro data or means basing can lead

to bias since it ignores the heterogeneity among regions and households So that estimating with micro data may prefer

In terms of using micro data, many papers assess the effect of irrigation on likelihood of being poor, and household’s income or expenditure Particularly, Hussain et al (2006) estimate a Logit model with dependent variable is dummy variable which equal 1 if household is poor The study employs primary data in Java, Indonesia in year 2000-2001 and shows that households who access to irrigation system are less likely to be poor than others However, the effect is various across irrigation systems and locations on the system Opeyemi and Babatunde (2014) attempt to evaluate the impact of Kampe irrigation dam in Nigeria They collect data from 140 households which are divided into two strata as treatment and control groups A Logit model is estimated with regressand is dummy variable

of being poor basing on Foster, Greer and Thorbecke poverty index The outcome indicates that the project helps to reduce poverty in the region it is conducted The research results of Berg and Ruben (2006) in Ethiopia show that irrigated land have a positive and much higher marginal effect on household’s expenditure than those of rainfed land Similarly, in China, Huang et al (2005) suggest that irrigation helps to improve household’s total income and reduce poverty

To the best of my knowledge, in Vietnam, recently there is no paper related to irrigation Before that, some studies use data collected in 1992-1993 and 2000 to assess benefit of irrigation to agriculture production and poverty reduction in Vietnam Indeed, Walle (2003) uses Vietnam living standard survey 1992-1993 and employs a profit function The result suggests that both irrigated and rainfed land have a significant and positive impact on crop income, but the effect of irrigated land is twice larger than that of rainfed land Moreover, higher human capital can take larger benefit from irrigation expansion Using primary data collected from farmers under irrigated lowland, rainfed lowland and rainfed upland in both northern and southern of Vietnam in 2000, Ut, Hossain and Janaiah (2000) discover the efficiency of using modern rice varieties They claim that modern rice varieties which is cultivated under irrigated land give more yield and profit than those are grown under rainfed land In addition, Biltonen, Hussain and Tuan (2003) collect data about agriculture production in two irrigation systems, Nam Duong in Red river delta and Nam Thach Han in Quang Tri province They employ a logit model to figure out the effect of irrigation on poverty Qualitative analysis shows that improvement in irrigation performance can lead to higher crop yield, and estimation result suggests that effect of irrigation is not similar for different locations on irrigation system

Studies relating to agriculture productivity show that accessing to irrigation improves crop yield, the benefit consists of land intensity and cropping intensity Accordingly, Ahmad et al (2002) employ a stochastic frontier production function to investigate the role of irrigation on wheat productivity in Pakistan using secondary data of 2228 farmers The study indicates that applying irrigation lead to higher productivity in wheat production, especially water for irrigating is taken from both canal and tubewell gives greatest efficiency In evaluating Zimbabwe European Union micro-project program which financially supports small farms in irrigation, Nhundu, Gwata and Mushunje (2010) assert that farms which access to irrigation have greater productivity in crop production Using long panel data at plot level of farming households in India, Jin, Jansen and Muraoka (2012) assess the impact of irrigation on crop productivity and number of cultivated seasons The estimation is fixed by household level The results show that irrigation has a significantly positive effect on land productivity, but the effect varies across applied irrigation systems which are public or private irrigation The combination of public and private irrigation gives highest benefit They also claim that

increase in land productivity caused by irrigation is mainly come from increase in cropping intensity Similarly, Dhawan and Datta (1992) also suggest strong relation between irrigation and cropping intensity In that study, they use cross-section data, and dependent variable is measured by ratio of gross to net sown area

2.4 Hypothesis testing

Basing on theoretical and empirical literature, three null hypotheses are given respectively to three thesis’s objectives

(1) Investigate the impact of irrigation on rice production

H0: irrigation does not have impact on rice production

H1: irrigation has a positive impact on rice production

(2) Assess differences in impact of water sources for irrigation on rice production

H0: different water sources for irrigation do not have different impact on rice production

H1: different water sources for irrigation have different impact on rice production

(3) Explore differences in impact of locations on canal system for irrigation on rice production

H0: different locations on commune canal system do not have different impact on rice production

H1: different locations on commune canal system have different impact on rice production

An analytical framework is given in Figure 2.2 to show determinants of rice production

Figure 2.2: Analytical framework

Source: performed by author.

Inputs

Land Labor Capital Seed Fertilizer

Other factors

Land fragmentation

Land quality

Agriculture

extension

Commune factors

Climate Infrastructure

Farmer’s characteristic

s

Education Gender

Age

RICE PRODUCTION

Disaster

Other sources

CHAPTER 3: DATA AND METHODOLOGY

3.1 Data source

The dataset used in this paper is Vietnam Access to Resources Household Survey 2014 (VARHS 2014) This survey is conducted in 12 provinces which include 5 provinces in Northern, 3 provinces in Central Coastal, 3 provinces in Central Highlands, and 1 province in Mekong River Delta For each province, some communes are chosen for survey The data consists of 3648 households with information on many aspects focusing on rural development Information related to agriculture is relatively specific, which includes characteristics of plots, type of plant, inputs, outputs, and irrigation The survey also collected information on household members and communes This paper employs data at household level of farmers who grow rice In total, there are 2490 households who cultivate rice corresponding to over 68% of sample size

3.2 Model specification

3.2.1 Building model

As mentioned above, profit function and cost function cannot be employed because of data limitation in inputs’ price, so that this thesis uses production function to accomplish all three research objectives Because of the simplicity in estimating, a Cobb-Douglas production function is used with logarithmic form

A cross section data model is established:

Model 1: For investigating the impact of irrigation on annual rice production

Model 3: For explore the impact of different locations of canal system on annual rice

production

ln(𝑌) = 𝛽0+ 𝛽1ln(𝐻𝑒𝑎𝑑)𝑖+ 𝛽2ln(𝑀𝑖𝑑𝑑𝑙𝑒)𝑖+ 𝛽3ln(𝐸𝑛𝑑)𝑖 + 𝛽4ln(𝑅𝑖𝑣𝑒𝑟)𝑖

+ 𝛽5ln(𝐿𝑎𝑘𝑒)𝑖 + 𝛽6ln(𝑂𝑡ℎ𝑒𝑟_𝑠𝑜𝑢𝑟𝑐𝑒𝑠)𝑖+ 𝛽7ln(𝐼𝑛𝑝𝑢𝑡𝑠)𝑖+ 𝛽8ln(𝐹𝑎𝑟𝑚𝑒𝑟_𝑐ℎ𝑎𝑟𝑎𝑐𝑡𝑒𝑟𝑖𝑠𝑡𝑖𝑐𝑠)𝑖 + 𝛽9ln(𝐶𝑜𝑚𝑚𝑢𝑛𝑒_𝑓𝑎𝑐𝑡𝑜𝑟𝑠)𝑖+ 𝛽10ln(𝑂𝑡ℎ𝑒𝑟_𝑓𝑎𝑐𝑡𝑜𝑟𝑠)𝑖 + 𝑢𝑖

as irrigation variable

In addition, in order to assess the impact of water sources, irrigated land is divided into land size irrigated by canal, river and spring, lake and pond, and other sources Since there are few farm using water from well for irrigation, it is classified as other sources Similarly, land size irrigated by canal is divided into head, middle and end basing on location of plots on commune canal system These variables are calculated in logarithm

In term of expected sign, many mentioned studies show that irrigation improves crop productivity, crop profit, and household’s income, so expected sign of variable irrigated land size is positive Similarly, expected signs of variables, irrigated land size divided by water sources and location on canal system, are positive

Inputs:

Inputs are theory variables of production function In agriculture production, inputs include land, labor, capital, fertilizer and seed (Bindlish & Evenson, 1993; Bindlish et al., 1993; Lockheed, Jamison & Lau, 1980; Saito, Mekonnen & Spurling, 1994.; Moock, 1976) Particularly, land is measured by household’s total area for rice cultivation (square meter) Labor is divided into household’s labor and hired labor Household’s labor is measured as total working days of household’s members for cultivating rice Hired labor is total cost for hiring labor to produce rice (in thousands dong) Total cost for hiring asset, machinery, equipment, means, cattle for producing rice (in thousands dong) and total value of productive assets of household are proxies for capital variable Fertilizer and seed are added into model

as total cost for fertilizer and seed, respectively, to produce rice (in thousands dong) All input variables are used in logarithm form Because these variables are inputs used to produce rice, so that they are expected to have a positive impact on rice production

As mentioned above, one of problems of Cobb-Douglas production function is that output will be zero if any input is not used For logarithm form, the problem now is expressed that log of zero value cannot be determined Hence, to solve this problem, log of zero value is set

as zero This solution is applied for all variables in logarithm form in the model

Farmer’s characteristics:

Many studies explore the impact of farmer’s characteristics, such as age, education, and gender, on agriculture productivity and add them into the model as control variables, but their effects are controversial Quisumbing (1996) reviews eight studies related to gender differences in agricultural productivity The outcome indicates that efficiency of female farmers is not lower than male farmers, except one paper show less efficiency of female farmers but with weak significance Moreover, three in eight papers show that formal education has a significantly positive impact on agricultural productivity, while the rest shows positive but insignificant impact Higher age of farmers is expected to reduce

productivity because of lower strength, but higher age may give more experience which improves agriculture production Particularly, one in eight papers indicates significantly negative impact of head’s age on total value of crop production, and another shows significantly positive impact of age on value of production

This thesis uses age, education and gender of household’s members who are responsible for rice production as control variables Since each household in dataset may own more than one plot which grows rice and managers of plots are different, average measures of age and schooling years of these plot managers are employed, both of them are in logarithm form For gender, two dummy variables are used, they are male plot manager (equal 1 if all rice plot managers in household are male) and male-female plot manager (equal 1 if there are both male and female managers of rice plot in household) The expected sign of formal education is positive However, expected signs of age and two dummy variables, male plot manager and male-female plot manager, are unclear

Commune factors:

Commune factors express characteristics of commune which affect agricultural productivity, they consist of infrastructure (such as availability and quality of road, irrigation system, input market, and information system) and climate (for example rainfall and seasonal climate variance) There are three variables can be used as proxies for infrastructure, they are distance between commune center and main road (km), distance between commune center and district center (km), and percent of concrete commune roads The first two variables are used in logarithm form For climate factor, there is no variable that can represent for rainfall However these variables will be subtracted out of model since cluster specific fixed effect estimation method is employed, which is mentioned in detail later

Other factors:

Other control variables include land quality (Huang et al., 2005; Jin, Jansen and Muraoka, 2012), intra-farm fragmentation of land (Huang et al., 2005; Markussen et al., 2013), visiting agriculture extension (Ahmad et al., 2002; Moock, 1976; Saito, Mekonnen & Spurling, 1994), and hit by disaster (Huang et al., 2005) Accordingly, land quality is measured as ratio

of rice cultivated land, which has higher quality than average quality of land in village, to total rice cultivated land Land intra-farm fragmentation is the fragmentation of land owned

by a farm, it is measured as the number of plots growing rice (in logarithm form) Visiting agriculture extension is added to model as dummy variable which equals 1 if any household’s member visits extension agent at least once in 12 months before interviewed point of time Finally, hit by disaster is measured as ratio of rice cultivated land hit by disaster in 12 months

to total rice cultivated land

Expected signs of variables ratio of land quality, visit agriculture extension and ratio of land hit by disaster are positive, positive and negative respectively In term of number of plots, its expected sign is not clear If land is divided into many plots, difficulty in moving among plots of labor and means will increase However, high intra-farm fragmentation may lead to lower risk and give chance to test new varieties or technology

A summary of variables and their expected signs is as below:

Table 3.1: Variables and expected sign

Dependent variable: household’s total amount of rough rice in 12

months (kg)

Independent variables

Irrigation

2 Area of irrigated land at head end canal (log) Positive

4 Area of irrigated land at tail end canal (log) Positive

6 Area of land irrigated by river or spring (log) Positive

8 Area of land irrigated by other sources (log) Positive

Inputs

2 Household members' working days to produce rice (log) Positive

6 Cost of renting asset, machinery, equipment, means and cattle (log) Positive

Farmer’s characteristics

2 Average schooling years of plot managers (log) Positive

Commune factors

1 Commune’s infrastructure

2 Commune’s climate

Other factor

1 Ratio of land with better quality than average in the village Positive

Source: performed by author.

3.3 Estimation strategy

Before going to estimation strategy, it is necessary to figure out assumptions of classical linear regression model and some problems the model may have

Assumptions of classical linear regression model (Gujarati, 2015):

(1) Parameters are linear

(2) Independent variables are non-stochastic

Other problems are heteroskedasticity and autocorrelation If the assumption of homoscedasticity is not held, the estimators are still unbiased and consistent, but they are not efficient under OLS estimation, and produce inaccurate results in hypothesis testing (Gujarati

& Porter, 2009, p 374) Cross-section data tends to be suffered by heteroskedasticity than time-series data since cross-section data contains many households or firms with different

characteristics and sizes If variance of 𝑢𝑖 is known, weighted least square estimation can be applied to solve heteroskedasticity Nonetheless, variance of 𝑢𝑖 is seldom available, so other approaches are required regardless of variance of 𝑢𝑖 One of them is White’s heteroskedasticity-corrected standard errors (White, 1980) This method gives higher standard errors than those of OLS estimation, so t-statistic values will be smaller There are many test to check whether heteroskedasticity exists However, for fixed effect model Green (2012, p 338) suggest a modified Wald test for heteroskedasticity within cluster of residuals Null hypothesis of this test is that all variances of ui within cluster are constant If H0 is rejected, there is heteroskedasticity in fixed effect regression

Similarly to heteroskedasticity, under autocorrelation OLS procedure still produces unbiased and consistent estimators, but it is not efficient (Gujarati & Porter, 2009, p 423) Therefore,

F test and t test will give inaccurate outcome Autocorrelation can be pure autocorrelation or caused by model mis-specification For pure autocorrelation, Generalized Least Square procedure or approach of Newey and West (1987) which is extended from White’s heteroskedasticity-corrected standard errors can be used to solve the problem The latter one

is employed for large sample

Another important problem which receives much attention is correlation between error term and regressors, it is also known as endogeneity Unlike heteroskedasticity or autocorrelation, under OLS estimation this problem leads to biased and inconsistent estimators (Gujarati & Porter, 2009) The main reason for this problem is omitting variables which correlate with both dependent and independent variables It commonly happens because of data limitation and there are many variables which cannot be measured There are some solutions to this issue, the best solution which is recommended is using instrument variables Appropriate instrument variables are variables that can mainly explain endogenous variable variation, but they do not directly correlate with dependent variable However, this method gives much difficulty because data is limited and it is hard to find good instrument variables Hence, many methods are suggested to use information from the model as instrument variables, such

as GMM model In estimating production function, Levinsohn and Petrin (2003) provide an estimation method of using investment and intermediate inputs as instrument variables for inputs use to solve endogeneity of input choice Nevertheless, this method is not appropriate

in agriculture production Finally, fixed effects can be also used to partly deal with this problem

For this thesis’s model, omitting variables that correlate with both irrigation and rice production can make estimated coefficients of irrigation variables biased and inconsistent These omitted variables may be related to ability and personality of farmers, climate (such as total amount of rain and its distribution during a year), and quality of irrigation system Because of data limitation, these variables cannot be added into model However, it can be partly solved by using cluster specific fixed effect approach which will subtract commune variables from the model For subtracting ability and personality of farmers, a panel data model with fixed effect estimation can be applied Nevertheless, it may not work well since VARHS data are surveyed every two year which ability and personality may change a lot In addition, variation of land and irrigation over time is not much, so that panel data with fixed effect can make these variables insignificant

Finally, since VARHS data is not simple random sampling it may give some problems Cameron and Trivedi (2005) claim that stratified and clustered sample can lead to difference

in distribution of (𝑦𝑖, 𝑥𝑖) among strata and correlation among households within cluster However, if our purpose is estimating the impact of regressors on regressand rather than predict for population and stratifying is not basing on regressand, the variation of sampling rates can be ignored Thus, stratification is not the case for this thesis’s objectives, but clustering does The correlation among households within cluster is caused by some reason First, surveyed households may live in the same block Second, it is existence of unobserved cluster specific variables that affect all households in cluster Finally, unobserved variables may impact all households in the same province or region such as policy, culture and climate Since these unobservable variables may correlate with both dependent and independent variables, the estimators under OLS procedure will be biased and inconsistent In addition, within cluster correlation among error terms makes estimators under OLS approach inefficient

There are two ways to solve the impact of these unobserved variables, they are cluster specific random effect (CSRE) and cluster specific fixed effect (CSFE) CSRE model is used

in the situation that there is no correlation between unobserved cluster variables and independent variables However, it is not the case of this thesis’ circumstance since as

mentioned above unobserved commune variables are likely to affect both inputs use, irrigation decision and produced output Thus, CSFE model is needed to be applied because

it will subtract cluster-invariant variables, but some tests will be used to test whether CSFE is necessary

The clustering problem becomes more complicated when considering the number of clusters Let C is the number of clusters, and Nc is the number of observations within cluster Then there are two situations which are few clusters (C is small and Nc → ∞) and many clusters (Nc is small and C → ∞) In VARHS dataset, observations can be classified by province level (12 provinces), district level (138 districts) and commune level (500 communes) Clustering

at province level is not appropriate since there is high heterogeneity within province related

to social-economic condition, infrastructure, irrigation system, farming practices and even climate Similarly there is also high heterogeneity within district Meanwhile, commune level gives more homogeneity and correlation among households in agriculture production Moreover, clusters cannot be too small because they do not give enough information for estimation, each cluster has to have at least two observations Clustering at commune level can satisfy this condition Hence, clusters should be communes Since there are many communes and the number of observations in each commune is about 7 on average, it should

be classified as many-clusters situation So that, the situation of few clusters will not be mentioned in this thesis

3.3.1 Cluster-specific random effect model

Suppose that there is a sample of N observations with C clusters Then a cluster-specific effects model is written as:

𝑦𝑗𝑐 = 𝑥′𝑗𝑐𝛽 + 𝛼𝑐+ 𝑢𝑗𝑐where: the script jc denote the jth observation in the cth cluster, j = 1, 2, …, Nc, c = 1, 2, …, C

𝑥𝑗𝑐 is a K × 1 vector of regressors 𝛼𝑐 is cluster specific effect which changes across clusters, and it is assumed that 𝛼𝑐⁡~⁡[0, 𝜎𝛼2] 𝑢𝑗𝑐 is assumed to have zero mean and constant variance Under a CSRE model, 𝛼𝑐 is assumed to distribute randomly and independently with regressors 𝛼𝑐+ 𝑢𝑗𝑐 is also supposed to be correlated within cluster, 𝐶𝑜𝑣(𝛼𝑐 + 𝑢𝑗𝑐, 𝛼𝑐 +

𝑢𝑘𝑐) = 𝜎𝛼2 with j ≠ k

If the true model is CSRE model, OLS estimation can give consistent estimators but it is not efficient (Cameron and Trivedi, 2005) Hence, the default variance of OLS estimators needs

to be adjusted The issue can be solved by using cluster-robust variance matrix (CRVM) estimation (White, 1984; Liang & Zeger, 1986) CRVM of OLS estimator can be estimated as:

Source: Cameron and Miller, 2015

In order to test for cluster effect, Breusch and Pagan LM (Lagrange Multiplier) test can be employed, this test is a two-sided test using residuals of OLS regression without cluster-

specific effect (Breusch & Pagan, 1980) Nonetheless, Moulton (1987) suggest an one-sided

LM test that is more powerful than Breusch and Pagan LM in term of cluster-specific effect The null hypothesis of this test is H0: 𝜎𝛼2 = 0, with H1: 𝜎𝛼2 > 0 If H0 is rejected, CSRE is more appropriate than OLS without cluster-specific effect This test is calculated as:

2(∑𝐶 𝑁𝑐2 𝑐=1 − 𝑁)1/2(∑ (𝑁𝑐𝑢̅𝑐)

2 𝐶

𝑐=1

∑ ∑𝑁𝑐 𝑢̂𝑗𝑐2

𝑗=1

𝐶 𝑐=1

− 1)

Where 𝑢̂ is residuals of OLS regression without cluster-specific effect LM statistic follows standard normal distribution

3.3.2 Cluster-specific fixed effect model

Again, a cluster-specific effects model is written as:

𝑦𝑗𝑐 = 𝛼𝑐 + 𝑥′𝑗𝑐𝛽 + 𝑢𝑗𝑐Under CSFE, both regressors and the intercept 𝛼𝑐 need to be estimated In the situation of few clusters, 𝛼𝑐 can be estimated by using cluster dummy variables However, when there are many clusters, within-cluster estimation should be applied A within-cluster estimation is written as:

To choosing between CSFE and OLS without cluster-specific effect, a simple F test can be used with null hypothesis is H0: 𝛼1 = 𝛼2 = ⋯ = 𝛼𝑐 If H0 is rejected, CSFE regression should be used instead of OLS without cluster-specific effect

3.3.3 Choosing between CSRE and CSFE

Although CSFE estimation produces consistent 𝛽 coefficients when there is potential correlation between unobservable cluster-invariant variables and independent variables, it has disadvantages that it reduces variation of variables and cannot determine coefficients of cluster-invariant variables Hence, a test should be used to test whether OLS with CRVM estimation or Feasible GLS is valid If the test shows that these procedures are invalid, CSFE should be applied In order to test the assumption of no correlation between 𝛼𝑐 and independent variables, a Hausman test can be conducted The test is basing on differences between regressors of CSRE and CSFE regressions:

𝑇𝐻𝑎𝑢𝑠𝑚𝑎𝑛 = (𝛽̂𝐹𝐸− 𝛽̂𝑅𝐸)′[𝑉̂[𝛽̂𝐹𝐸] − 𝑉̂[𝛽̂𝐹𝐸]]−1(𝛽̂𝐹𝐸− 𝛽̂𝑅𝐸)

Source: Cameron and Trivedi, 2005

Where: 𝛽̂𝐹𝐸 is vector of estimated coefficients of CSFE regression, 𝛽̂𝑅𝐸 is vector of estimated coefficients of CSRE regression Since cluster-invariant variables are dropped in CSFE regression, these variables are not contained in the formula of Hausman test This test follows

𝜒2 distribution If the null hypothesis is rejected, CSRE regression is not appropriate, then CSFE regression needs to be used

However, this test is valid in the situation that CSRE regression is full efficient If there is heteroskedasticity in 𝛼𝑐 or 𝑢𝑗𝑐, or correlation among 𝑢𝑗𝑐 in the same cluster, another test is required Hausman test can be replaced by a Wald test for the auxiliary OLS regression with cluster-robust variance (Wooldridge, 2010)

The auxiliary OLS regression is written as:

𝑦𝑗𝑐− 𝜃̂𝑦̅𝑐 = (1 − 𝜃̂)𝛼 + (𝑥𝑗𝑐 − 𝜃̂𝑥̅𝑐)′𝛽 + (𝑥𝑗𝑐 − 𝑥̅𝑐)′𝛾 + 𝜀𝑗𝑐Only cluster variant variables are included A robust Wald test is conducted with null hypothesis is 𝛾 = 0 If H0 is rejected, CSFE regression should be used