What drives the Implementation of Foreign Large-Scale Land Investments in Developing Countries?

(1)

GEORG-AUGUST UNIVERSITY GÖTTINGEN & UNIVERSITY OF GRONINGEN

MASTER’S THESIS

What drives the Implementation of Foreign Large-Scale Land

Investments in Developing Countries?

- Evidence on Success and Failure from the Land Matrix Global

Observatory

-Author: Lisa Fiedler Matriculation number: 21213362 (University of Göttingen) s2589257 (University of Groningen) E-Mail: fiedler.lisa@gmx.de

Address: Rosengarten 7, 97618 Hohenroth, Germany Supervisor:

Jun. Prof. Dr. Jann Lay

Department of Economics at the University of Goettingen Co-assessor:

Dr. Abdul Azeez Erumban

Faculty of Economics and Business at the University of Groningen

(2)

Abstract

This thesis firstly analyses determinants of foreign large-scale agricultural land in-vestment in developing countries in a gravity model using a cross-sectional data-set from the Land Matrix Global Observatory. Our findings confirm previous research that resource-seeking motives are paramount and that weak land tenure security in the target country is associated to stimulate investments in land. Secondly, we ana-lyse determinants of successful implementation of production of these investments in a binary outcome model by examining the influence of the type of crops intended to produce, farm size, and the type of investor. Our results indicate a negative effect of non-food crops production intention and of increasing large-scale farm size. Stock-exchange listed investors are associated to be most successful in implementing pro-duction compared to investment funds. Thirdly, we address the phenomenon of land speculations by defining it and developing a theoretical framework on the drivers of land speculations.

(3)

List of Tables

1 Number of Projects by Country Pair . . . 23

2 Summary Statistics for Dependent Variables I (all zeros excluded) . . . 23

3 Summary Statistics for the Dependent Variable II . . . 26

4 Intra-Regional Projects (in per cent and number of projects) . . . 29

5 Cumulative Size (in million hectares). . . 30

6 Descriptive Statistics for Independent Variables I (mean comparison) . . . 30

7 Estimation Results from the Negative Binomial Model . . . 33

8 Time Needed to Start Production (in years) . . . 37

9 Descriptive Statistics for Independent Variables II (mean comparison) . . . 38

10 Investors’ Efforts to Consult the Local Community (in per cent and number of projects) . . . 39

11 Type of Former Owner (in per cent and number of projects). . . 39

12 Type of Former Use (in per cent and number of projects) . . . 40

13 Estimation Results from the Probit Model (Average Marginal Effects) . . . 41

14 Predicted Probabilities of Successful Implementation for the Crops Intention (Probit Model) . . . 42

15 Predicted Probabilities of Successful Implementation for the Type of Investor (Probit Model) . . . 43

A1 Classification of Crops . . . 50

A2 Definition of Explanatory Variables and Source for the Analysis of the Determin-ants of Large-Scale Land Investments . . . 51

A3 Empirical Results from the Poisson Pseudo Maximum Likelihood Model . . . 52

A4 Empirical Results from the Zero-Inflated Negative Binomial Model . . . 53

A5 Estimation Results from the Logit Model (Average Marginal Effects) . . . 54

A6 Size Under Contract of Concluded Projects (in hectares) . . . 54

A7 Empirical Results from the Negative Binomial Model with Additional Institu-tional Variables . . . 55

A9 Predicted Probabilities of Successful Implementation for the Crops Intention (Lo-git Model) . . . 56

(5)

List of Figures

(6)

Acronyms

CEPII Centre d’Etudes Prospectives et d’Informations Internationales FAO Food and Agriculture Organisation

FDI foreign direct investments

GAEZ Global Agro-Ecological Zones Data Portal from FAO and International Institute for Applied Systems Analysis

GDP gross domestic product

GIGA German Institute of Global and Area Studies IPD Institutional Profiles Database

LM Land Matrix Global Observatory LSLA large-scale land acquisitions LSLI large-scale land investments LSLS large-scale land speculations ME marginal effects

NB negative binomial

OECD Organisation for Economic Cooperation and Development OLS ordinary least squares

PPML Poisson pseudo-maximum-likelihood RTA regional trade agreement

(7)

1 Introduction

In the last decade, the resurgence of demand for agricultural land in the Global South gained attention in media and academics. In particular, the acquisition of large-scale agricultural land in low-and-middle income countries by foreign investors is nowadays largely discussed.1 This thesis uses the term large-scale land acquisitions (LSLA) to refer to this phenomenon. LSLA, often described as the new "global rush for land", were triggered by a spike in food prices in 2007-2008 and the global financial crisis, which reduced the attractiveness of other assets (Deininger and Byerlee, 2011; Kugelman and Levenstein, 2009). The impacts of these LSLA on the target country and its development are widely discussed in current research: Some scholars point at the threats for the livelihood and the rights of the affected local population (Cotula, 2013; German et al., 2013; Grain, 2008). Some argue that opportunities for rural development and food security exist (Deininger and Byerlee, 2011). Others stress that the recent surge is only a bubble driven by specula-tions on rising land prices and that projects are less likely to start operation (De Schutter, 2011).2 In the following we refer to projects that have the intention to start production as large-scale land investments (LSLI). Those projects that do not have the intention to produce and are driven by speculative motives are referred to as large-scale land speculations (LSLS). However, as the nature and magnitude of this new "global land rush" are still hard to as-sess mainly due to the lack of reliable data, the analysis of the impacts of such land deals remains difficult and vague.3 Yet, before the impacts of LSLA can be analysed, a better understanding of the underlying factors that determine LSLI is pivotal. Hence, an ana-lysis of determinants4 of LSLI is aimed to shed further light on impact questions. These questions include if LSLI may increase the pressure on land, improve agricultural pro-ductivity, and induce spillovers through back-and forward linkages to the economy in the target country.5 For example, if resource-seeking motives of the investors are iden-tified to be a determinant of LSLI, it can be concluded that pressure on land is likely to exacerbate and that the local population may loose access to land. Thus, a negative im-pact of LSLI could be indicated. A positive imim-pact of LSLI could be derived, if countries with a low productivity in agriculture are particularly targeted. Consequently, techno-logy spillovers from the investor country could improve the agricultural productivity.

1 _{The Land Matrix Global Observatory (2014) defines large-scale land acquisition as an acquisition of land}

above 200 hectares. The World Bank defines and provides a list of all low-and-middle income countries. Please refer to: http://data.worldbank.org/about/country-and-lending-groups.

2 _{Summaries of the discussion on impacts are provided by Anseeuw et al. (2012b) and Cotula and}

Vermeu-len (2009).

3 _{This paragraph draws from Anseeuw et al. (2012b), Arezki et al. (2013), Cotula and Vermeulen (2009),}

and Deininger and Byerlee (2011).

4 _{We use the terms determinants, drivers, and factors interchangeable.}

5 _{Scientific studies on the impacts of LSLA are scarce. FAO (2013), based on case study evidence, provides}

(8)

Moreover, recent research attempts have revealed that in contrast to the large number of deals that were agreed on and have been signed, the amount of projects that have ac-tually started to set-up agricultural production is considerable lower. We use the term successful project implementation for those projects where the acquisition, the purchase or lease of land, is concluded, and which started producing agricultural goods.6 Failure of implementation refers to projects where the acquisition is concluded as well, but no agri-cultural production is set-up. Deininger and Byerlee (2011) state that for only 20 per cent of the announced projects between 2004-2009 included in their dataset of 14 developing countries, a start in production was reported. Current research has further been stressed that implementation failure is a prevalent issue due to unexpected difficulties, such as technical or financial problems or conflicts with local communities in the target country. Additionally it has been pointed out that implementation of production starts on a small scale compared to the acquired land under contracts.7

Only projects that produce agricultural commodities can achieve spillovers to agricul-tural production or backward and forward linkages to the whole economy. If project implementation is delayed or fails, it still increases pressure on land and may lead to a weakening of land rights and displacements of the local population. Thus, it is of partic-ular importance for the analysis of the impacts of LSLA to understand what determines successful implementation of production and to what extent the set-up of production fails. Thus, the analysis of determinants of successful implementation intends to provide a deeper understanding of the feasibility of LSLI. Furthermore, this analysis is of partic-ular value for target governments to identify and rule out non viable projects.

In addition, it has been argued that those projects whose production has not started yet may be driven in part by speculations on long term rising farmland prices and do not aim to start producing. In particular, minimal purchase and lease prices as well as in-creasing demand in agricultural land may have induced speculations (Anseeuw et al., 2012b; Cotula et al., 2010; De Schutter, 2011). Thus, if a project has not started produ-cing, this can indicate either a failure of implementation of LSLI or a LSLS. Deininger and Byerlee (2011) estimate that 18 per cent of the announced deals have not started pro-duction yet and might be driven in part by speculation. Similar to LSLI that failed to be implemented, LSLS would increase the pressure on agricultural land and may harm the local population. Further, if speculations in land exist to a large extent, a bubble with implications to the world economy could emerge.

6 _{Started producing agricultural goods in this thesis refers to projects which are in the start-up phase of}

production or in operation.

7 _{This paragraph draws from Andrianirina-Ratsialonana et al. (2011), Borras et al. (2011), Cotula (2012),}

(9)

From this, we derive that there is a need to shed light on the drivers of LSLS and to ana-lyse to what extent they exist.

Currently, there are only two studies that analyse the determinants of LSLI, namely Lay and Nolte (2014) and Arezki et al. (2013). Both scientific articles understand LSLI as foreign direct investments (FDI) in agriculture and find out that those are in part different from the determinants of FDI in other sectors. Lay and Nolte (2014) stress that resource-seeking motives are of particular importance, whereas market-resource-seeking motives are negli-gible. Arezki et al. (2013) highlight that countries with poor land governance are targeted, which contradicts the assumption that institutional quality stimulates FDI.

However, both studies do not address the questions what drives successful implement-ation and to what extent do speculimplement-ations exist. This is our motivimplement-ation for an analysis of this research gap. We expand the analysis of Lay and Nolte (2014) and Arezki et al. (2013) by addressing the issues of successful implementation and land speculations. Hence, our main research questions include: What are determinants of successful implementation of LSLI? How are speculations in the context of LSLA defined and in which format do they exist? What drives them, and how can they be evidenced?

By addressing these research questions, we contribute to the literature on LSLA in the following regards: First, determinants of LSLI are analysed, and the findings from Lay and Nolte (2014) and Arezki et al. (2013) are reviewed. Second, based on the findings of Lay and Nolte (2014), hypotheses on determinants of success and failure of LSLI regard-ing their implementation are developed. Third, the phenomenon of speculations in the context of LSLA is elucidated, and hypotheses on drivers of LSLS are developed. Fourth, drawing on data from the Land Matrix Global Observatory (LM)8, we empirically analyse determinants of success and failure of implementing production in the context of LSLI. In addition, descriptive statistics are used to provide indications of the existence and extent of LSLS. Fifth, the results from the analysis are used to derive possible impacts of LSLA. The remainder of the thesis is structured as follows: Chapter two reviews theoretical and empirical findings from relevant literature. Based on this review, research hypotheses are developed for the analysis. Chapter three presents the dataset employed and outlines the methodological approach applied. Chapter four interprets the results from the descript-ive and empirical analysis. Chapter fdescript-ive concludes by summarizing the results from the analysis, discussing methodological difficulties and data limitation, and indicating direc-tions for further research. In addition, possible impacts of LSLA on the target countries are derived from the results of the analysis.

8 _{Official statistics of FDI in agriculture that cover inward and outward flows of emerging and developing}

(10)

2 Literature Review and Hypotheses

This chapter first puts LSLA in the context of FDI and clarifies in which cases LSLA can be considered as LSLI. Second, the literature on the determinants of FDI is reviewed, and hypotheses on the determinants of LSLI are derived based on the findings of Lay and Nolte (2014). Third, the issue of implementation in the context of LSLI is discussed, and hypotheses on the drivers of successful implementation are developed. Fourth, land speculations in the context of LSLA are addressed, and hypotheses on the drivers of LSLS are developed.

2.1 Large-Scale Land Acquisitions and Foreign Direct Investment

This thesis focuses exclusively on foreign interest, lease or purchase, in agricultural land in low-and-middle income countries. Without denying their importance, we will not take national investment in agricultural land as well as land acquisitions for investments in other sectors, such as mining, tourism, conservation, and forestry into account.

Lay and Nolte (2014) clarified that if LSLA have the intention to produce agricultural goods, they can be understood as a specific subset of agriculture FDI as they conform to the Organisation of Economic Cooperation and Development’s (OECD) definition of FDI:

"Foreign direct investment reflects the objective of establishing a lasting in-terest by a resident enterprise in one economy (direct investor) in an enter-prise (direct investment enterenter-prise) that is resident in an economy other than that of the direct investor. The lasting interest implies the existence of a long-term relationship between the direct investor and the direct investment prise and a significant degree of influence on the management of the enter-prise"(OECD, 2008, p.48).

(11)

However, if land is acquired without the aim to set up production and driven by spec-ulations on rising farmland prices, the project does not comply with the OECD’s (2008) definition of FDI and can thus not be considered as LSLI but as LSLS. The phenomenon of LSLS and its drivers will be discussed in section 2.4.

2.2 Determinants of Large-Scale Land Investments

This section reviews relevant literature and derives theoretical considerations on the de-terminants of LSLI. Considering LSLI as agricultural FDI, links the theoretical consider-ations on the determinants of LSLI to the literature on the determinants of FDI. There exists a vast literature on the determinants of FDI in manufacturing and services. The following literature review focuses on the determinants of FDI on the macro-level and on those factors which we consider to be relevant for investments in land for agricultural production in developing countries.

Literature identified that locational and organizational choices both determine FDI (Dun-ning, 2009). Locational choice theory focuses on favourable production locations regarding costs and refers to country-level factors, such as factor endowments, geographic prox-imity, common language, institutions, governance, cultural distance, or real wage differ-ential. Gravity model specifications, traditionally used to explain international trade flows, have been widely employed to empirically explain bilateral FDI flows. Based on New-tonian physics the FDI intensity between two countries is influenced by their sizes and proximity. Those models identified economic size of the respective two countries, geo-graphical distance between them and several proxies for transaction costs between coun-tries as determinants for bilateral FDI (Kepaptsoglou et al., 2010; Salvatici, 2013).9 Fur-thermore, Lopez et al. (2009) and Rugman (2005) confirmed that FDI is rather a regional than a global phenomenon. Recently, the concept of multilateral resistance, introduced by Anderson and Van Wincoop (2003), gained prominence in the field of international trade and gravity model specifications. The intensity of bilateral trade and FDI is not only affected by barriers to trade of the respective two countries but also by both trading partners interactions with all other countries (Kepaptsoglou et al., 2010).

Another important factor that determines FDI, and which is particularly interesting in the context of FDI in developing and emerging countries, is the productivity gap between foreign and domestic firms. Investors locate their FDI where their productivity is higher compared to the targeted market in order to exploit a competitive advantage. Bougheas et al. (2008) provide empirical evidence for this link for the acquisition strategy of British firms from 1988 to 1996.

(12)

Moreover, studies focusing on FDI in developing and emerging countries highlight the importance of institutions and economic policy. Busse and Hefeker (2007) find that institu-tional indicators and political risk in the target country are important for the decision to invest. Gastanaga et al. (1998) and Wei (2000) confirm the detrimental effect of corruption and taxes. However, for FDI in natural resources, Aleksynska and Havrylchyk (2013) find evidence that the deterring effect of low institutional quality in the target country is diminished by rich endowments in natural resources. Further studies emphasized the importance of institutional proximity. Inter alia Aleksynska and Havrylchyk (2013) and Bénassy-Quéré et al. (2007) argue that FDI is deterred by large institutional distance.

Organisational choice refers to the choice of a particular trading form: internalisation at home, outsourcing at home, internationalisation abroad, and outsourcing abroad. Inter-nationalisation theory states that international transactions are internalised if the transac-tion costs of the free market exceed internal costs. Changes in the world economy and the nature of FDI such as the technological revolution, the rise of trade in services, the expansion of international vertical specialisation, and the increasing fragmentation of production processes, as well as the availability of new firm-level datasets put the or-ganizational choice of the firm in the focus of research (Helpman, 2006). In addition, the property rights approach links the internalisation decision to incomplete contracts. If in-ternational transactions are governed by incomplete contracts, firms prefer internalising over outsourcing. The ownership of an asset is a source of power and may be misused at the expense of the partner (Antràs, 2003; Helpman et al., 2004). Empirical evidence for the effect of contractability on internalisation is provided by Antràs and Yeaple (2013). They confirm a positive relationship of noncontractabile investments and the prevalence of internalisation for U.S. based firms.

(13)

By contrast, direct investments are necessary. Furthermore, production is internalised to overcome incomplete contracts and to ensure investment returns.

Second, the importance of production factor endowments would be different for LSLI than for FDI in manufacturing. In general terms, favourable factor costs determine FDI. The more abundant a production factor is the lower its price. Lay and Nolte (2014) find out that it is crucial for agricultural production that land and water endowments are avail-able. Capital and labour endowments matter to a lesser extent, because they are mobile and substitutable to a larger extent than land and water. This leads to the conclusion that the abundance of land and water endowments are crucial determinants of LSLI, while labour and capital endowments matter to a lesser extent.

Third, Lay and Nolte (2014) evidence that the motives for LSLI are resource-seeking rather than market-seeking due to those investments being subjective to food and energy secur-ing considerations of the investsecur-ing countries. An important factor is therefore the ease with which products can be exported from the target to the investor country. Therefore, host country experience in commodity exports determines LSLI.

Fourth, the productivity gap is of particular importance for LSLI. Lay and Nolte (2014) find out that investors target in particular host countries with a large share of area with a high yield gap.10 This indicates that those countries are not efficient in agricultural production and an increase in output can be easily achieved. Moreover, low land pro-ductivity implies that land prices are low.

Fifth, regarding institutions, an ambiguous effect is identified for LSLI by Lay and Nolte (2014): Institutional quality in general, such as political stability and a better business en-vironment, is conducive to LSLI. However, the lack of rule of law and corruption are also conducive to LSLI. LSLI take place in an intransparent environment where land rights are vague and insecure for the local population. Investors can take advantage of this and acquire land below market prices with the support of the target government.

Sixth, Lay and Nolte (2014) argue that institutional proximity is of particular importance in the context of LSLI. Investors, who are used to poor institutions in their domestic mar-ket, are attracted to invest in similar institutional environments. Investors from countries with well-defined institutions are assumed to comply with ethical investment principles and guidelines. Those could discard investments in intransparent environments.

Lay and Nolte (2014) further confirm that trade and information costs, agglomeration ef-fects, and infrastructure are equally important for LSLI as for FDI. Arezki et al.’s (2013) findings are in line with Lay and Nolte’s (2014) assumptions.

10_{The yield gap measures the difference between the amount of agricultural products a farm actually}

(14)

Based on these considerations and findings, we derive the following hypotheses with re-spect to the determinants of LSLI:

H1.1: Investors target fertile land and water abundant countries, while labour and capital endowments matter to a smaller extent.

H1.2: The experience of the host country in commodity exports are crucial drivers of LSLI, while market seeking motives are less important.

H1.3: Investors target in particular countries with low agricultural productivity. H1.4: Target country institutions have an ambiguous effect on LSLI.

H1.5: Institutional proximity is an important driver of LSLI.

2.3 Determinants of Successful Implementation of Large-Scale Land

Investments

The previous chapter addressed the underlying factors that determine LSLI and stressed several investor and target country characteristics that influence the decision to acquire land for the purpose of agricultural production. However, descriptive evidence showed that there is a substantial share of projects where the acquisition of land is concluded but production is not implemented.11 Therefore, in this chapter we review relevant literature and develop a theoretical framework on the determinants of successful implementation of production of LSLI.

We assume that the reasons, why production is not implemented are two fold: On the one hand, land is acquired with the intention to set-up agricultural production. Those projects comply thus with the definition of LSLI. The decision to acquire land is based on the in chapter 2.2 outlined determinants. In particular, those investment principles are the availability of crucial production factors, the quality of business environment, the quality of infrastructure, and considerations regarding transaction cost between countries, for example geographical distance. However, after the acquisition of land, there are certain other factors that delay or deter the implementation of production. We discuss these in detail in the remainder of this section. On the other hand, land could also be acquired out of speculation and without the intention to set up production. The phenomenon of LSLS will be defined and its drivers will be discussed in section 2.4.

11_{Descriptive evidence is found by Andrianirina-Ratsialonana et al. (2011), Borras et al. (2011), Cotula}

(15)

The literature on FDI in manufacturing focuses either on the in section 2.2 mentioned market entry determinants or on post-market entry dynamics including survival or per-formance of foreign subsidiaries. None of the two strands refers directly to the imple-mentation of production. Theoretical considerations and empirical evidence on success-ful implementation of FDI is scarce in the literature. The only study that addresses the question of implementation in FDI is Kokko et al. (2003). In this study the authors fo-cus on FDI failure of project implementation in Vietnam. They highlight that unforeseen transaction costs of doing business can reduce the profitability of the investment and thus deter implementation. Moreover, they assume that transaction costs have a stronger effect on investment failure in developing countries than in developed countries due to larger uncertainty and lack of information. In addition, they find that project-specific character-istics, e.g. the size of the investment and the mode of market entry, determine successful implementation.

The literature on LSLA emphasized the issue of implementation of production.12 How-ever, theoretical considerations on the determinants of implementation are scare and em-pirical evidence does not exist. Deininger and Byerlee (2011) provide stylized facts on im-plementation failure derived from case studies from 14 developing countries from 2004 to 2009. They conclude that the low ratio of implemented production can be explained by economic and technical challenges as well as conflicts with local communities. Economic challenges refer to changes in market conditions or reduced profitability. For example, Deininger and Byerlee (2011) explain that in Mozambique not a single biofuel project star-ted to operate due to a decrease in the oil price. Furthermore, investors are confronstar-ted with liquidity and financial problems, for example to access additional capital from inter-national markets. Technical issues comprise, according to Deininger and Byerlee (2011), difficulties in clearing the acquired land, establishing the required internal infrastructure, and linking to required input and sales markets. Conflicts with local communities that deter implementation would represent a particular issue if land rights are not formalized and the target government conducts the acquisition of land. If the local population that used the acquired land for agricultural purpose was not consulted or not involved in the negotiations, and no agreement on the type and amount of compensation was achieved, conflicts that deter implementation of production may occur.

Based on Kokko et al.’s (2003) and Deininger and Byerlee’s (2011) insights, we assume that the following factors influence successful implementation of LSLI: First, specific pro-ject characteristics are assumed to determine successful implementation. The intention which crops will be produced should be considered.

12_{In particular Andrianirina-Ratsialonana et al. (2011), Cotula (2012), Deininger and Byerlee (2011), and}

(16)

To categorise the crops intention we differentiate between the cultivation of food crops, non-food crops, flex crops, and multiple uses as suggested by Borras et al. (2011). Flex crops refer to crops that can be used for both, food and biofuel commodity production, depending on the preference of the investor. Multiple uses production implies the cultiv-ation of at least two different crops intentions, for example one food and one non-food crop. This classification takes into account that in more than one third of LSLA, mul-tiple uses are intended (Anseeuw et al., 2012a). A simple classification in non-food and food is therefore not appropriate.13 The intention to produce flex crops or multiple uses reduces the risks linked to price volatility and commercialisation compared to the exclus-ive production of either food or non-food crops (Anseeuw et al., 2012a). Therefore, the production of multiple uses or flex crops may be implemented to a larger extent than food or non-food production.

Moreover, farm size is assumed to determine implementation. Empirical evidence showed that agricultural production is associated with constant or decreasing returns to scale, once a certain minimum farm size, that guarantees the efficient utilization of machinery capacity is achieved (Arezki et al., 2013). We assume that all projects in the sample em-ployed for the analysis achieved this minimum capacity because it includes only projects with a hectares size over 200. The advantages of large farms, for example being able to access capital from international markets, would be outweighed by the costly man-agement and supervision of workers (Arezki et al., 2013). This implies that production is more likely to be set up on smaller farm sizes, conditioned that the minimum scale requirement has been achieved. However, recent developments in information techno-logy, tillage, and crop breeding may have reduced the costs of labour supervision, labour intensity, and diseconomies of scale (Deininger and Byerlee, 2012). Thus, the competit-iveness of large farms would be increased (Arezki et al., 2013). Moreover, a large farm size indicates a larger investment project. Larger investment projects are associated with investors with larger financial resources (Kokko et al., 2003). Based on the argumentation above, a large farm size is potentially associated with with successful implementation. Second, it is of particular importance to get a better understanding of who invests in LSLI. According to economic theory, private companies are assumed to be more product-ive than state-owned companies in competitproduct-ive markets (Dewenter and Malatesta, 2001). The more productive an investor is, the better its performance in FDI (Melitz, 2003). Thus, it can assumed that private investors are more successful in implementation. However, as LSLI take place in a highly governmental-regulated environment (Lay and Nolte, 2014), a state-owned investors may be more successful in implementing projects than private companies due to closer relations to the target government.

(17)

Moreover, we assume that a public investor has other incentives and motives than a private investor. A public investor has the incentive to secure food and energy supply and is less profit oriented. By contrast, the incentives of a private investor are more profit-oriented (Toulmin et al., 2011). If the profitability of the investment turns out to be less profitable than expected, private investors may not start production.

Further, LSLI target emerging and developing countries, where financial markets are less developed, and doing business may require additional and unexpected capital compared to investments in developed countries (Deininger and Byerlee, 2011). The own financial resources of the investor and the access to international capital markets are thus crucial for successful implementation.

In addition, investor’s efforts to consult with the local community and to obtain their con-sent may also be conducive to implementation. LSLI tend to target countries with poorly defined land rights (Arezki et al., 2013; Lay and Nolte, 2014). Target governments can thus sell land that was used by the local population without the consent or compensation of the local population. This could cause conflicts with the local population, in particular if several actor claim ownership. Investor’s efforts to consult with the local population and to avoid such conflicts are thus pivotal to successful implementation (Deininger and Byerlee, 2011).

Third, the former use and owner of the acquired land play an important role in the set-up of production. If the land was already under agricultural production, in particular under large-scale production, before the acquisition of a foreign investor, the implementation may be relatively easier than in the case where the land was used for other purposes, such as stock farming, forestry, or conservation. The clearing of land causes additional costs and time. Investors may have underestimated those costs. If the land was previ-ously owned by the state or a private large-scale farmer, the transfer of land rights may be easier and faster due to the involvement of less actors than in the case of small-scale or community-owned land.

(18)

Fifth, time specific events, such as political crises, a change of government, or economic shocks in the investor and the target country or on a global scale affect the feasibility of implementation.

Based on these considerations, the following hypotheses for successful implementation of LSLI are derived:

H2.1: Project-specific characteristics, such as the crops intention and farm size, matter for successful implementation.

H2.2: Investor characteristics, including the type of investor, its productivity and finan-cial resources, as well as efforts to consult with the local community are important for successful implementation.

H2.3: The former use and owner of the land are crucial factors for successful implement-ation.

H2.4: Successful implementation is context specific and depends on local factors that are potentially unobservable.

H2.5: Time specific events influence successful implementation.14

2.4 Large-Scale Land Investment versus Land Speculation

As already indicated in the previous sections of this chapter, not all LSLA may have the intention to set-up agricultural production and are driven by speculative motives. This section addresses the phenomenon of LSLS and develops hypotheses on its drivers. In contrast to an investment, a speculation does not aim to create value added, for ex-ample through an establishment of a business. The intention of a speculation is to earn profits from a future price change of an asset. An asset is purchased at an initial low price with the expectation that prices will rise in the future. The aim of a speculation is to sell the asset later in order to profit from the price difference (Robles et al., 2009; Wahl, 2009). In the context of LSLA, speculations refer to the acquisition of land as a strategic asset in expectation of rising land prices to create the opportunity for a capital gain. The acquired land will be kept vacant and no efforts are undertaken to start agricultural production (De Schutter, 2011; Shrestha, 2011; Triantafyllopoulos, 2010). The underlying motives for land acquisitions are therefore completely different: The speculator expects an increase in the demand for land, whereas an investor expects to earn profits from an enterprise established on the acquired land.

14_{It needs to be noted that we can only test H2.1 and H2.2 empirically. Due to lack of data availability H2.3}

(19)

Evidence for the existence, in particular for the extent, of land speculations in the context of LSLA is scarce. Anseeuw et al. (2012a), Borras et al. (2011), and De Schutter (2011) argue that the the large share of LSLA which has not started production yet leads to the conclu-sion that LSLA are at least partly driven by speculations. However, it is unclear to what extent those not implemented projects represent LSLS. The previous section outlined that there are valid arguments for certain factors deterring successful implementation of LSLI. Thus, speculation is not the unique explanation for a not implemented project.

The current available data does not allow to identify which of the projects that have not started production are speculative and which are not. The literature on commod-ity or housing speculations faces a similar issue, because it is not possible to differentiate between investment and speculation for an individual project. However, to identify the existence of speculations in a market, the literature first defines what a speculation is not: a change in prices due to a change in the business fundamentals, such as changes in consumption demand or physical supply. Second, it is analysed if price changes can be legitimised by changes in the fundamentals. If not, the price differs from the long-term market value and speculations exist.15 Roche and McQuinn (2001) apply this approach to test the existence of speculations in agricultural land in Ireland in the 1970s and 1980s. Thus, the approach is generally applicable to test for land speculations in the context of LSLA. Yet the current available data do not provide information on land prices and their development in the target countries on a cross-country base. Therefore, it is not possible to empirically evidence the existence of land speculations.

Theoretical considerations regarding the drivers of LSLS evaluate both low land purchas-ing prices and expectations about rispurchas-ing land prices as important (Anseeuw et al., 2012b; Cotula et al., 2010; De Schutter, 2011). Certain country characteristics imply that prices for agricultural land are lower in some countries compared to farmland prices in other coun-tries. Target countries of LSLA are characterized by land abundance, imperfect input markets, and high yield gaps. These characteristics are reflected in low land productivity and hence low land prices. Moreover, the lack of land rights and land tenure security al-lows governments to regulate land prices and to purchase or lease land below its market value to attract FDI (Lay and Nolte, 2014). In particular, in Africa, land prices are lower compared to other targeted regions (Anseeuw et al., 2012b).

Expectations about rising land prices are driven by the increasing demand for land. Prices for land are, compared to other assets, more elastic to changes in demand due to the fact that the supply of land is fixed (Foldvary, 1998). Population growth and the spike of food prices in 2007 and 2008, as well as the increasing demand for biofuels and plastics made out of soy beans and corn increased the demand for land for agricultural production.

(20)

The global financial crisis in the same years, and the resulting collapse of bond and equity markets increased the attractiveness of new forms of investment, such as land acquisi-tions (Anseeuw et al., 2012a; Schiller, 2013).

Evidence from Deininger and Byerlee’s (2011) case studies showed that in many target countries of LSLA the approving of a land acquisition is rushed by the target govern-ments to avoid loosing an investment possibility. Therefore, the investor and his inten-tion are not screened properly, and speculative projects may not be detected. In addiinten-tion, screening and monitoring capacities are underdeveloped in many target countries and have only limited resources. These are valid arguments to assume that LSLA are partly driven by speculations and that in particular countries with low land prices and under-developed sreening and monitoring capacities are targets of LSLS.

In contrast to investors, speculators do not aim to set up a functioning production. There-fore, economic and technical feasibility of the projects matter to a lesser extent. Insti-tutional quality, infrastructure, and possible productivity gains in the target country do not determine the decision to speculate. The determinants of LSLS would thus differ in parts from the determinants of LSLI. However, one could also argue that speculators ac-quire land that is most likely to be demanded by investors with the intention to produce agricultural commodities. Thus, economic and technical feasibility of projects could be considered by speculators and the determinants of LSLS would be identical with those of LSLI.

As we already stated in the previous chapter, the motives for land acquisitions are as-sumed to diverge for different investor groups. Private companies, and in particular investment funds, have an interest in capital gains and profits, whereas public companies may have interests in achieving food and energy security. Therefore, public companies are more likely to aim at agricultural production (Toulmin et al., 2011). Private com-panies, and in particular investment funds, are therefore more likely to speculate than state-owned companies.

(21)

Based on these considerations, we develop the following hypotheses for the drivers of LSLS:

H3.1: Low land purchasing prices are crucial drivers for LSLS.

H3.2: Expectations about rising land prices are crucial drivers for LSLS.

H3.3: Economic and technical feasibility of projects may have an impact on LSLS. H3.4: Certain investor types speculate more than others due to different motives.

3 Data and Methodological Approach

This chapter first outlines the conceptual approaches to test the hypotheses developed in the previous chapter. As already discussed in the previous chapter, the current data structure faces the main problem to differentiate between projects that failed to be im-plemented, but which were originally LSLI, from LSLS. Approaches to deal with this issue are discussed. Second, the data used and data limitations are presented. Third, the possible empirical approaches are discussed and the preferred empirical models for the analysis of the determinants of LSLI and of successful implementation are presented.

3.1 Conceptual Approach

To acknowledge the previous hypotheses, we apply three conceptual strategies. Follow-ing the theoretical distinction between the land acquisition and the project implementa-tion outlined in the previous chapter, we differ between these two in our analysis. First, the hypotheses with respect to the determinants of LSLI will be tested. In doing so we fol-low Lay and Nolte’s (2014) bilateral country-level approach. To examine the determinants of bilateral LSLI, they used land demand for any bilateral investor-host pair, measured by the cumulative farm size under contract, as the dependent variable and an augmented gravity model specification. However, we employ the number of projects with a written or oral contract, referred to as concluded projects, as dependent variable.

(22)

A further approach to solve the identification issue of LSLS would be to analyse the de-terminants of LSLI on a lower geographical scale, for example for each target region sep-arately. In regions with a low share of no-production projects, one can assume that LSLS is a minor or no issue. Hence, by analysing those regions separately, one could confirm the determinants of LSLI. However, the current data available in terms of numbers of observations is too limited to consider target regions separately, and the results would therefore not be reliable. Moreover, an analysis if and how the determinants of LSLI dif-fer for difdif-ferent target regions would be of interest. It would be further of interest to analyse if the determinants of LSLI differ for certain groups of investors. For example, between investors from high-income and low- or middle-income countries or between public and private investors. Again, we face the issue of limited data availability.

Second, to analyse the determinants of successful implementation of LSLI we use a project-level approach. In-production and no-production is considered as a binary outcome. Hence, the probability of in-production rather than no-production is examined. Again, we face the issue that LSLS cannot be excluded from the sample. However, to be able to analyse the determinants of successful implementation we have to assume that all pro-jects in the sample have the intention to produce and comply with the investment determ-inants outlined in H1.1-H1.5. Thus, in the empirical evaluation of the next chapter, we use the identification hypothesis that only true investors (and not speculators) are present in the sample.

An attempt to identify LSLI subtract from LSLS would be to consider regions or even countries with a small share of no-production projects separately. In those regions or countries one can assume that LSLS is a minor or no issue. However, the current avail-able data is too limited for such an analysis. The degrees of freedom in the empirical analysis of sub-samples of particular regions and countries are too small.

(23)

3.2 Data

This section first describes the data-set obtained from the LM and the data management applied. Second, the additional data included as explanatory variables for the analysis of the determinants of LSLI is presented. Third, the explanatory variables used for the analysis of the determinants of successful implementation of LSLI are outlined, the issue of lack of data availability is discussed, and an approach to overcome this problem is presented.

3.2.1 The Land Matrix Data-Set

For the empirical analysis, we use data from the LM as of April 14, 2014. The data-set is compiled from various sources, including research papers and policy reports by interna-tional and local organisations and NGOs, personal information contributed through the Global Observatory’s website, field-based research projects, official government records, company websites, and media reports. The LM records transactions that include a trans-fer of rights to own, control or use land through sale, lease or concession, that cover an area of 200 hectares or larger, that have been initiated since the year 2000, and that target low-and-middle income countries.16 Due to lack of transparency and limited data avail-ability in the context of LSLA, the LM notices that several biases may be introduced into the data and recommends that any trends evident in the analysis of the data should be considered as indicative rather than as given. Potential bias may originate for example from the fact that media and research focus on certain regions, certain investors, and cer-tain sectors. Projects out of this focus may be under represented in the data-set (Land Matrix Global Observatory, 2014).17

We use a subset of the whole LM data set and applied the following data-management: First, we limited the sample to land acquisition for agricultural purposes. Second, because we are interested in foreign land acquisitions, domestic investments are excluded. Third, while the LM also provides data on the intentions to acquire land and failed negotiations, we consider only concluded projects, projects which are based on an oral agreement or written contract.18 We exclude all concluded but abandoned projects.

16_{The LM uses the classification of low-and-middle income countries from the World Bank. Please refer to:}

http://data.worldbank.org/about/country-and-lending-groups.

17_{The Land Matrix Global Observatory (2014) points out further sources of biases.} _{Please refer to:}

http://www.landmatrix.org/en/about/#are-there-biases-in-the-data.

18_{There are two variables in the LM that describe the status of a deal: the negotiation and the}

(24)

Further, we split the sample of concluded projects into two sub-groups: The first one includes projects that entered the production stage, meaning that the project is in a start-up phase or in operation. The second grostart-up includes no-production projects whose production has not started yet, or no further information regarding implementation is available. Fourth, we exclude all small island states from the sample.19 Fifth, we du-plicate/triplicate the observations of those projects that have two/three investors from different countries, because the dataset does not provide information on the the primary investor. Our final sample consists of 781 concluded projects, 528 in-production, and 253 no-production projects. There is one observation for each concluded project, which indic-ates a cross-sectional data structure.

3.2.2 Variable Specification for the Analysis of Determinants of Large-Scale Land In-vestments

This section presents the variables that we included to the data set, based on the empir-ical evidence and theoretempir-ical considerations in section 2.2, for the country-level analysis of the determinants of agricultural FDI. Those will be used as explanatory variables in the country-level analysis:20 To investigate which and how different production factors mat-ter, we include the stock of land, measured by agricultural area in 1000 hectares per capita from FAOStat, the share of available land, measured by the available forest land, grass-land and woodgrass-land as percentage of total grass-land from the Global Agro-Ecological Zones Data Portal from FAO and International Institute for Applied Systems Analysis (GAEZ), and water availability, measured as actual renewable water resources per capita from FAO Aquastat. Per capita values are used to account for the size of the country, because larger countries are associated to have larger values than smaller countries. In addition, we in-clude the availability of labour, measured as active population in agriculture as a share of total working-age population from FAOStat, and the availability of capital, measured as a country’s rank of getting credit from the World Bank Doing Business. The lower the rank the better the access to credits.21

19_We _apply _the _United _Nations’ _definition _of _small _island _states. _Please _refer _to:

http://sustainabledevelopment.un.org/index.php?menu=1520.

20_{As the LM reports LSLA since the year 2000, we include, if possible, the additional variables from the}

year 2000. If no or too limited data was available from the year 2000, e.g. for land tenure insecurity, the data originates from other years. Table A2 in the Appendix provides a list of all variables included from other data sets.

21_{It would be further interesting to differentiate between between financial capital in general terms, such}

(25)

To test whether experience in commodity exports is crucial, we include net food imports per capita from FAOStat. Negative values imply that a country is a net exporter of agricultural commodities, positive values indicate that it is a net importer. To measure market size, we include total population, and gross domestic product (GDP) from the World Development Indicators (WDI).

To examine the effect of institutional quality and the effect of similar institutions, we use the institutional indicator constructed by Lay and Nolte (2014) based on data from the World Governance Indicators (WGI).22 The composite indicator ranks from 0 to 1, with 0 indicating low institutional quality. Proximity of institutions is measured with a dummy that is 1 if the difference in the institutional indicator between the investor and the target country is between -0.25 and 0.25.

In a further specification of explanatory variables, we include land tenure insecurity and the functioning of political institutions index from the Institutional Profiles Database (IPD). The indices rank from 0 to 4, with 0 indicating very low land tenure insecurity and poor func-tioning of political institutions, respectively. Moreover, two dummy variables to measure similarity of land tenure insecurity and functioning of political institutions are included. The dummies are 1, indicating similarity, if the difference between the target and the in-vestor country is between -1 and 1.

To assess the importance of productivity and the productivity gap, we include the share of area with a ratio of actual and potential yield over 70 per cent from GAEZ. This meas-urement indicates the share of area with low yield gap and therefore efficient production. We also include the share of area with a ratio of actual and potential yield below 40 per cent. This measurement indicates the share of area with high yield gap and therefore inefficient production.

The choice of control variables is crucial for the specification of the empirical model. The selection of controls should be based on economic theory to reduce the possibility of omit-ted variable bias (Kennedy, 2003). Therefore, we add the traditional gravity variables geo-graphical distance, common official language, colonial relationship, and regional trade agreements (RTAs) from CEPII. To control for infrastructure, we include telephone lines per 100 people from WDI. To account for agglomeration and learning, we construct a variable from the LM that counts all projects concluded in a particular target country, and a variable that counts all projects of a particular investor country. To test if LSLI are sim-ilar to general FDI, a regional rather than a global phenomenon, we include a dummy that is 1, if the investor and the target country originate from the same geographical region.23

22_{In this composite indicator, the WGI indicators, voice and accountability, political stability, government}

effectiveness, regulatory quality, rule of law, and control of corruption are included. The composite indicator is constructed through principal component analysis.

(26)

3.2.3 Variable Specification for the Analysis of Determinants of Successful Imple-mentation

In this section we present the variables that we use, based on the theoretical considera-tions in section 2.3, as explanatory variables for the project-level analysis of the determin-ants of successful implementation of LSLI from the LM data. To test the effect of the crops intention we construct a categorical variable that differentiates between multiple uses, food crops, non-food crops, and flex crops. To examine the influence of farm size on suc-cessful implementation, we consider the size of area under contract in hectares. To assess the effect of different types of investors, we construct a categorical variable that differenti-ates between the following types of investors: state or semi-state owned, stock-exchange listed, investment funds, private companies or individual entrepreneurs, and other. The LM provides data on the former owner and use, as well as on community consulta-tion. However, the number of observations is too small to include those in this empirical analysis. No information is provided if more than one actor claim ownership of the land acquired. Furthermore, there is no information available with regard to specific investor characteristics, such as investors’ own financial resources, their access to international fin-ance, or their access to technology dictating their project-specific productivity. Moreover, no data is available on local factors, e.g. the quality of local markets for inputs, such as fertilizer and pesticides. As we assume that these factors influence implementation of production, we are aware of that the empirical analysis may face omitted variable bias. An approach to reduce it, is to include dummies for each investor and each host country. Even though, these dummies are not a perfect representation of project- and investor-specific characteristics, they still work as a catch-all element, reflecting the average con-ditions in the country - and thence proxying for the missing elements. For example, the financial resources and productivity of the investment company are incorporated by in-vestor country variables.24 Further, if dummies for each investment company would have been used, the above outlined project specific explanatory variables could not have been included into the model due to high multicollinearity with those dummies. Therefore, we argue that the inclusion of investor and target country dummies is a plausible approach to deal with the problem of omitted variable bias under the current state of research and available data.

24_{We are aware that any individual company c might largely differ from the average of the country j where}

(27)

3.3 Empirical Approach to the Analysis of Determinants of Large-Scale

Land Investments

This section discusses the empirical approach employed for the analysis of the determ-inants of LSLI. As already outlined in section 3.1, for the analysis of the determdeterm-inants of LSLI, we employ a gravity model specification, which has been successfully used to analyse the determinants of bilateral FDI (Salvatici, 2013). The simple form of the model can be expressed as:

Tij =β0Y_iβ1Y_jβ2D_ijβ3 (1)

Tij is the intensity of FDI between the investor country j and the target country i, which is proportional to the economic sizes of country j (Yj), and country i (Yi), and inversely proportional to their physical distance Dji. β0, β1, and β2are unknown parameters. The model can be easily expanded, to include further variables representing any other factor stimulating or preventing FDI.

Conventionally, a log-linearised model, estimated by ordinary least squares (OLS), is used for the empirical analysis. In order to empirically estimate a log-linearised gravity model, a random disturbance term ηij, which is assumed to be independent and identically dis-tributed, is added to equation (2) and the multiplicative form of the theoretical model is transformed into a linear empirical specification by taking logarithms on both sides of the model (Silva and Tenreyro, 2006):

(28)

To solve these issues, Silva and Tenreyro (2006 and 2011) propose to use a Poisson pseudo-maximum-likelihood (PPML) estimator which makes it possible to estimate the gravity model in its multiplicative form. Moreover, the PPML can deal with the heteroskedasti-city of the error terms and does not have the issue of zero logarithms. Besides a critique from Martínez-Zarzoso (2013), the approach achieves academic acceptance.25 Further-more, Burger et al. (2009) suggest to use negative binomial (NB) and zero-inflated (ZI) models as an alternative approach to estimate a gravity equation in its multiplicative form. We discuss these models in detail below as we opt for those models also because of statistical reasons.

Under the framework of the gravity model, we use the following notation to describe the dataset. nji is the number of projects of investor country j in target country i with j = {1,..,C}, where C is the number of all countries in the world that are possible investor countries and with i = {1,..,P}, where P is the number of low-and-middle income coun-tries that are possible target councoun-tries.

We construct a bilateral count dataset of njiwith i = {1,..,P}and j= {1,..,C}. This dataset includes all investor-host country pairs excluding combinations where the host country is not a low or middle income country. First, to analyse the determinants of LSLI, we consider the number of concluded projects of investor country j in target country i as de-pendent variable. Second, to check the robustness of those determinants, we refer to the number of in-production as dependent variable. Therewith, LSLS are excluded from the sample.

The data set consists of 17,930 possible investor-target country pair combinations. 110 possible target countries are included. 64 of them are target countries of concluded pro-jects, and 57 are target countries of projects that have started production. Further 164 possible investor countries are included. 57 of them are investor countries of concluded projects, and 48 are investor countries of in-production projects. For both dependent vari-ables, the amount of zero observations in the sample is with over 90 per cent considerably large (see Table 1). Whereas for the number of concluded projects 279 observations are unequal to zero, for the number of in-production projects 219 are non-zero (see Table 2). Most country pairs have only a few projects, while some have a large number of projects. For example, the average number of projects concluded is two, while the maximum num-ber of projects concluded equals 45.

25_{Arezki et al. (2013), Prehn et al. (2012), Fally (2012), Arvis and Shepherd (2011), and Burger et al. (2009)}

(29)

Table 1: Number of Projects by Country Pair 0 1 2 3 4 5 6-10 11-20 >21 Total Number of 17,651 179 41 14 12 9 15 7 2 17,930 concluded projects Per cent 98.44 1 0.23 0.08 0.07 0.05 0.08 0.05 0.02 100 Number of 17,711 145 33 12 9 6 8 5 1 17,930 in-production projects Per cent 98.78 0.81 0.18 0.07 0.05 0.03 0.06 0.03 0.01 100

Table 2:Summary Statistics for Dependent Variables I (all zeros excluded)

Obs. Mean Median Std. Dev. Min Max

Number of concluded projects 279 2.480287 1 4.099325 1 45

Number of in-production projects 219 2.109589 1 2.805115 1 29

The selection of explanatory variables for the augmented gravity model specification, that is based on the evidence from the literature and presented in chapter 3.2, leads to the following estimation equation that applies to both dependent variables:

nji =β0+β1CharacteristicsPairji+β2CharacteristicsTargeti (3)

+β3CharacteristicsInvestorj+eji

Pairwise LSLI are regressed on a set of variables from the target country i, the investor country j, characteristics of the target and investor pair ji, and an error term eji.

In general, it is assumed that count data follow a Poisson distribution rather than a nor-mal distribution which would be more adequate to apply a linear model such as OLS. Additionally, a linear model does not assure that the predicted values are non-negative. They can take any real value: positive or negative, integer or not. In count data, the ob-servations can only take positive integer values or zero. For the Poisson model, which is estimated by maximum likelihood, to ensure that the mean µ >0, the parametrisation of the mean is:

µ =E[y|x] = exp(x0β) (4)

(30)

However, count data are often not equidispersed. The majority of count data exhibit over-dispersion. The two major reasons for overdispersion are heterogeneity and the excess of zeros. Heterogeneity refers to the phenomenon that the coefficient estimates are not the same for all individuals in the sample. Excess of zeros refers to the phenomenon that the data contains more zeros than expected if the data is Poisson distributed (Kennedy, 2003).

The PPML estimation relaxes the equidispersion assumption. The PPML estimator is widely used for count data, which makes it applicable for our analysis.26 However, for a consistent PPML estimation the conditional mean, which coincides with the model spe-cification, has to be correctly specified (Cameron and Trivedi, 2013).

Another approach to deal with overdispersion in count data is to use a NB model. In contrast to the PPML that addresses over- or underdispersion by departing from a com-plete distributional specification, the NB model applies a distributional specification that allows a more flexible modeling of the variance to address overdispersion. The NB model is obtained by introducing unobserved heterogeneity into the Poisson model and by in-cluding an additional error term that is gamma distributed with mean unity and variance

α. The conditional mean is the same as for the Poisson and the PPML model, but the

vari-ance is specified as a function of the conditional mean µ and the overdispersion parameter

α. Different NB models can be created by assuming α to be different functions of µ. The

most applied specification, the NBII model, considers α as a fixed parameter (Kennedy, 2003). The moments of the NBII are µ = E[y|x]and VAR[y] = µ+αµ2. The NB models

are, unlike the PPML, not consistent if the variance specification is incorrect. However, the quadratic variance of the NBII model is a very good approximation in a large variety of overdisperesed count data and works therefore well in practice (Cameron and Trivedi, 2013; Wooldridge, 2010).

As a final extension, ZI models are used to address the issue that more zero observations are evident than the Poisson model or the NB model would predict for count data. ZI models combine a binary variable c with a standard count variable y. If c takes the value 0, then y = 0. If c takes the value 1, y takes the count values 0,1,2,.... Therefore, zeros are created from two processes, namely as a realization of the binary and of the count process. There are two latent groups: an always-zero group, i.e. never will trade with each other, and a not-always zero group that has the probability to have positive counts, i.e. that potentially will trade with each other in the future. The ZI model estimates first a logit or probit regression for the always-zero group. Second, a Poisson or NB model is estimated for the not-always zero group (Long and Freese, 2001; Winkelmann, 2008).

(31)

To decide which model is most appropriate for our data, the following strategy is applied: First, the graphical inspection (histogram) as well as the skewness/kurtosis tests for nor-mality reject the null hypothesis that the dependent variables, the number of concluded projects and the number of in-production projects, are normally distributed. Thus, it can be concluded that a linear model (OLS) would not be appropriate. Further, the results indicate that the data is skewed to the right. The excess of zeros could thus be an issue. Second, descriptive inspection of the mean and the variance of the dependent variables as well as a statistical test for overdispersion are used to test whether the data is equid-ispersed (Burger et al., 2009; Trivedi and Colin, 2010). The results indicate that the data is overdispersed. Thus, it can be concluded that the data is not poisson distributed and that a Poisson maximum likelihood model is not appropriate.

Third, the heteroskedastic-robust RESET test checks the correct specification of the con-ditional mean of the NB model and the PPML (Silva and Tenreyro, 2006; Wooldridge, 2010). The test confirms that the conditional means are correctly specified in the PPML and the NB model (see p-values that are reported at the bottom of Table 7 and A3 in the Appendix).

Fourth, the likelihood ratio test as well as the Akaike and the Bayesian information cri-terion can be used to examine the fit of the PPML and the NB model (Trivedi and Colin, 2010).27 All three, reported at the bottom of Table7 and A3 in the Appendix, confirm that the NB model has a better fit to the data compared to the PPML.

To test whether a zero-inflated model is favoured over a non-zero inflated model, to ad-dress the issue of excess zeros, the Vuong statistic can be applied (Burger et al., 2009; Trivedi and Colin, 2010). The Vuong statistic, reported at the bottom of Table A4 (column 1) in the Appendix, favours the NB model over the ZI NB model for the preferred model specification.

Based on these findings, we estimate the number of pairwise land acquisitions with a NB model as the preferred model. As robustness checks, we report a PPML and a ZI-NB model.

Further, we use a second model specification which measures institutional quality with two other variables, land tenure insecurity and political institutions, as a robustness check. It needs to be noted that with this second model specification, the number of observations drop considerably, and the results are thus less consistent.

To account for multilateral resistance, we follow the approach of Lay and Nolte (2014) and add a third model specification that includes investor and target country-fixed effects.

27_{When data is analysed with a PPML or a NB model, an equivalent statistic to the R-squared from the}

(32)

3.4 Empirical Approach to the Analysis of Determinants of Successful

Implementation

For the project-level analysis of the determinants of successful implementation, we fol-low the approach of Kokko et al. (2003) who use a binary outcome model to estimate the determinants of investment project failure. Therefore, we employ the following data structure: The dichotomous variable p_s/ndescribes whether a project has started produc-tion or has not started producproduc-tion. p_s/n is labelled 1 in case the production started, 0 in case that there is no-production. The project-level data sample consists of 781 observa-tions, 528 in-production, and 253 no-production projects (see Table 3).

Table 3:Summary Statistics for the Dependent Variable II

ps/n Freq. Percent

no-production (0) 253 32.39

in-production (1) 528 67.61

Total 781 100

With this data structure we are able to estimate the probability of in-production rather than no-production determined by a set of explanatory variables. We assume the exist-ence of a continuous variable y∗ which is linear dependent on a vector of explanatory variables x and a vector of coefficients α. This gives the single-index model: y∗ = x0α+e.

Although y∗can not be observed, the outcome,y, can be observed:

y =    1 if y∗ >0 0 if y∗ <0 (5)

This latent variable models yields:

Pr(y=1) = Pr(x0α+e>0) = Pr(−e <x0α) = F(x0α) (6)

F(.)is the cumulative distribution function of e. If −e is standard normally distributed,

it yields a probit model. If −e is logistically distributed this yields a logit model. If F(.)

is assumed to be linear, a linear probability model is applied (Cameron and Trivedi, 2013; Wooldridge, 2010).

What drives the Implementation of Foreign Large-Scale Land Investments in Developing Countries?

MASTER’S THESIS