The role of networks in ex post risk coping strategies

(1)

1

The Role of Networks in Ex Post Risk Coping

Strategies

Ronan Mulligan

University of Amsterdam

Master Thesis Economics

Supervisor:

Prof. Pauline Rossi, PhD

August 2017

Abstract

This thesis analyses the role of networks in ex post risk coping for rural households in Greater Eldoret, Western Kenya. The data used was the Kenya Greater Eldoret Health and Development Survey (2004-2006), a household survey collected over three rounds. A household fixed effects specification was used to determine how transfers respond to income shocks. The first estimation analysed the effect of income shocks on transfers received, with transfers decomposed by location of sender. The second estimation repeated the first estimation but with transfers further decomposed by relationship of sender to receiver, to assess the role of family connections in networks. Finally, a quarterly household fixed effect estimation was done to assess whether migration responds to income shock ex post. Evidence was found that households utilise in village transfers to deal with idiosyncratic shocks, but that covariate shocks reduce the probability of receiving a transfer ex post close to zero. It was also found that transfers from outside a 90km radius from Eldoret were used in response to covariate and idiosyncratic shocks. Proving evidence that households are utilising networks over great geographical distance. Family connections were found to be important for all transfers sent for consumption smoothing purposes and some evidence was found that former household members are sending transfers back to the household after covariate shocks. No evidence was found to support the notion that ex post household migration is used to create network connections.

(2)

Statement of Originality

This document is written by Ronan Mulligan who declares to take full responsibility for the contents of this document.

I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it.

The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

(3)

1 Introduction

Understanding how households manage risk has long been recognised as an important area of research. This is particularly true for developing countries as they are categorised as high-risk environments (Hoogeveen et al, 2004). The methods that are used to deal with risk are vital to understand how households move in and out of poverty. This fact is particularly true for rural communities in the developing world who depend on agricultural production. They face a variety of income shocks that can have direct effect on their welfare and ability to escape poverty. As a response, households have developed sophisticated coping mechanisms to deal with the risk posed by unanticipated shocks (Murdoch 1995). The focus of this thesis is informal insurance arrangements made through networks.

Networks are a prominent method used by households to deal with risk ex post (Fafchamps, 2003; Weerdt and Dercon ,2006; Murgai et al, 2002; Mazzocco et al, 2012). There is evidence to support both the importance of gifts and transfers (Lucas and Stark, 1988; Rosenzweig, 1988), as well as informal loan arrangements within networks for consumption smoothing purposes (Udry, 1994; Townsend, 1995). The motivation for households to do this is to cope with idiosyncratic risk, shocks that are random, independent, and identically distributed across households (Fafchamps, 2008). For example, if all households in a community face the possibility of having their crops destroyed by pests, then by agreeing to give transfers to those affected, the community can insure itself against pest attacks. However, communities may be unable to utilise these arrangements for covariate risk, shocks that are correlated across households (Fafchamps, 2008), as all households in the community will be affected (Hoddinott et al, 2009; Günther, 2009). Therefore, there lies incentive for households to spread their networks over greater distances to better cope with covariate risk.

Understanding networks and informal insurance is also key for policy intervention. Although households in developing countries are able to smooth consumption, many methods may be costly and there is evidence that introducing formal insurance would increase efficiency (Chetty and Looney, 2005; Rosenzweig and Binswanger 1993). Yet such interventions would act as substitutes to informal insurance arrangements and crowd them out (Dercon and Krishnan 2003; Lin et al, 2014), consequently eliminating spill over effects created by networks that could produce many unintended consequences. For example, Rosenzweig and Stark (1989) find that because networks can be formed by marriage through migration, introducing formal

(5)

2

insurance can have significant effects in the market for marriage. Dasgupta (2001) notes that links in networks can create trust, ‘she trusts you, now you trust me, so she now trusts me, and so forth’. Networks can also cause spill overs into urban labour markets. This is because households send members to work in urban centres with the intention for transfers to be sent back to the household. This point is particularly relevant for the context of this thesis, as 70% of Kenya’s urban work force is comprised of migrants (Agesa and Kim, 2001). Therefore, understanding how households use and form networks is key for effective intervention, as introducing formal insurance markets can have many unintended effects.

The aim of this thesis is to understand how households are utilising networks to deal with idiosyncratic and covariate risk. The first hypothesis tested is that: covariate risk will require a network with a greater geographical distance than is needed for idiosyncratic risk. The next aim of the thesis is to determine the role of family connection and former household members within these networks. Finally, this thesis will attempt to determine if households are establishing their own networks through migration of a household member.

The data set used is the Kenya - Greater Eldoret Health and Development Survey 2004-2006, which is a household survey conducted over three rounds. The survey covers primarily rural households who reside near the town of Eldoret. This is appropriate for the research question, given that rural households are often the most exposed to income shocks. The method used for the analysis is a household fixed effects estimation of the effect of income shocks on transfers received.

The first estimations reproduce results found the literature, that idiosyncratic shocks increase transfers within the village but that covariate shocks make it nearly impossible to receive transfers. Evidence was found that households are using networks outside the village to help cope with covariate shocks, but only for transfers received from regions that were outside a 90km radius from Eldoret1. Strong evidence was found that family connections are key for transfers used for consumption smoothing reasons, both from within and outside the village. No evidence was found to support the notion that income shocks increase migration. There was some evidence that former household members play a role in transfers from regions over 90km away for consumption smoothing purposes.

(6)

3

The rest of the thesis is as follows; Section 2 gives an overview and analysis of the existing literature on networks and their use for consumption smoothing. Section 3 details the data used for the thesis, and the limitation of the data. Section 4 describes the methodology of the thesis and details all estimations used. Section 5 reports the empirical results. Section 6 discusses the results and concludes the thesis.

2. Literature Review

Initial work on risk sharing arrangements in rural communities in developing countries characterised them informal but sophisticated. They were found to include implicit contracts and were sometimes found to have functioning interest rates. Udry (1994) reports that households in Northern Nigeria borrow more when they suffer an adverse shock and lend more after a positive shock. Furthermore, repayments of loans were found to be contingent on the realisation of random production and consumption shocks for both borrower and lender, indicating that loans are being used to share risk among households. Udry (1994) states that this is possible as there was no information asymmetry between borrower and creditor as they were in close geographic proximity. Thus, state contingent contracting is possible. In a study of 21 villages in southern India, Townsend (1995) provides evidence that there is a correlation between information and informal insurance arrangements with in the village. However, Townsend also notes that village organisation and structure are also determinates of informal insurance within village. It was also suggested that two villages similar on environment were dissimilar in organisation and outcomes, implying potential room for policy intervention.

The implication of the previously mentioned studies is that information, village structure, and organisation are key for informal insurance arrangements and should be a focus for policy. Another branch in the literature has shifted the focus of networks to family and social connections. This work has built on the idea that information asymmetry issues from a lack of official monitoring are overcome by family relationships and social bonds. Fafchamps (2003) tests this idea in an analysis of risk sharing networks in rural Philippines. Building on the work of Udry (1994), Fafchamps includes gifts and transfers in the analysis and introduces household fixed effects to correct for potential bias. The results show that consumption smoothing is an important motivation for gifts and informal loans, yet households are unable to efficiently share risk at the village level. The reason posited is that loans are not taking place at the village level, but through a network of friend and relatives. Departure from using the village as the unit of analysis for risk sharing is also supported by Weerdt and Dercon (2006). They note that village

(7)

4

level insurance is still possible, provided that every member of the village is also part of a network and that networks overlap sufficiently. The empirical estimates found that full insurance for food consumption at the village level cannot be rejected, yet they do support partial insurance of non-food consumption via networks.

The work of Fafchamps (2003) and Weerdt and Dercon (2006) suggests that there may be more important factors other than village organisation and structure in determining a household’s ability to share risk. Other work in the literature has further suggested that family connections, and social groups are key for risk sharing arrangements. Rosenzweig (1988) in a study of International Crops Research Institute of the Semi-Arid Tropics (ICRISAT) villages finds that households with more family links have transfers that are more responsive to income shocks. Mazzocco and Saini (2012) show that in India risk sharing is rejected at the village level but not at the caste level. Ligon (2004) also provides evidence for two different risk sharing groups differentiated along wealth. Finally, although Grimard (1997) rejects complete risk sharing among ethnic groups in Cote d’Ivorie, there is evidence for partial insurance by individual households with other members of the same ethnic group.

The results of these studies show that networks between friends and relatives, or through societal sub groups, can be more important than connections within the village. These papers have detailed how households utilise networks to manage idiosyncratic risk. However, if the network is not spread over a great enough geographical distance the household will still be at risk to covariate shocks. Hoddinott et al (2009) show that households in Ethiopia were unable to use networks after covariate shocks as few members resided outside the village. Porter (2008) determines that households are able to smooth consumption after an idiosyncratic shock but fail to do so after covariate shocks such as extreme rainfall. Furthermore, Fafchamps (2007) finds that geographical proximity is a major determinant of how networks are formed. Although it is found that the estimates of geographical proximity are correlated with family links and transfers from outside the four sampled villages are not included.

Given that households are unable to use transfers from the local area ex post covariate shocks, there is incentive for networks to span greater geographical distance. Yet distance also raises the costs of establishing and using networks, due to increased moral hazard from a lack of monitoring (Murgai et al, 2002). Although the literature shows that households may establish links over greater distances through migration and remittances. In an empirical study of Kenya, Bigsten (1996) provides evidence that household migration in Kenya is not a permanent move

(8)

5

but a pattern of circular migration in which the migrant leaves to work in an urban area only to return at a later date, while sending back transfers in the meantime. Rosenzweig and Stark (1989) find that households use marriage after migration to from connections between households over great distances.

Given that there are potential gains, are households establishing networks over great distances to manage covariate risk? Furthermore, are these networks a result of household migration? To attempt to answer these questions this thesis will analyse how households utilise their networks ex post idiosyncratic and covariate shocks. The estimation will also decompose transfers received by location and relationship of sender to receiver, allowing determinations of the role of distance and family ties in networks. Finally, estimations of migration responses to shocks will be estimated to attempt to determine if households are creating their own network connections via migration.

3 Data

3.1 Greater Eldoret Health and Development Survey

The data set used in this thesis is the Kenya Greater Eldoret Health and Development Survey (2004-2006). The data is a household survey collected in three rounds by Thirumurthy et al (2006), originally for use in their paper estimating the economic impacts of antiretroviral HIV treatment in Africa

3.1.1 Data Description

The interviews for the data were conducted in Kosirai Division, near the town of Eldoret in western Kenya. The division consisted of an area of 76 square miles with a population of 35,383 individuals making up 6,643 households in 1999 (Central Bureau of Statics 1999). Kosirai Division consists of predominantly rural households, with agriculture and animal husbandry being the primary sources of income. The primary crop cultivated is maize, which is harvested once every year.

The sample size for the data is 826 households across more than 100 villages. The sample consists of two groups, the first is a randomly selected representative sample of Kosirai Division and the second is made up of 250 households who were chosen because they had one or more members who receives treatment for HIV at a clinic in the Mosoriot health centre. Together the groups comprise the 826 households in the sample. The 250 households who have a member receiving HIV treatment could not be separated from the representative sample as

(9)

6

this would be a breach of confidentiality by revealing their health status. Therefore, households with a member receiving treatment for HIV are overrepresented within the sample. Household attrition in the sample was minimal at a rate of 1.4% and was checked by the authors to be random.

The interviews for the household surveys were completed over three rounds in 2004, 2005, and 2006. The interviews were conducted by teams consisting of male and female enumerators. The teams interviewed the household head and their spouse, as well as interviewing a youth in the household. Each interview began by listing all household members and information on characteristics: age, sex, relationship to household head, education, health status and participation in income-earning activities of each member. Critical for the purposes of this thesis, information was collected on income shocks and transfers received. The shocks households were asked about include; climate, loss of agricultural inputs, crops damaged prior to harvest, crops lost while in storage, and death or theft of livestock. The transfers households were asked about include money or goods received by the household that were not expected to be repaid. When goods were received the interviewer asked the respondent for an estimation of value.

3.2 Data Limitations

3.2 1 HIV Overrepresentation

As previously mentioned there is an overrepresentation of households with a member receiving treatment for HIV. Due to confidentiality issues, there is no way to account for this in the analysis. The original paper by Thirumurthy et al (2006) provides summary statistics which detail the observable variables for which the HIV sample differs from the representative sample. These summary statistics are shown in Table 1.

In terms of household structure, the HIV sample is only statistically different in the size of the household. The HIV sample is not statistically different in the number of children under 18, or the number of children living away from home. When looking at household head characteristics 53 % of the HIV sample has a female head, compared to 19% in the representative sample. Furthermore, the HIV sample has a greater percentage of households with single household head, with the HIV sample having 51% households with a single head to the representative samples 22%. There is also a statistically significant difference in the age of household head,

(10)

7

with the HIV sample having a younger head by 2.5 years. Looking to variables that describe asset ownership, the HIV sample appears to be worse off; owning 1.95 less acers of land, 21,360 shillings less of livestock, and having a greater percentage of households who are landless.

Overall, in terms of household structure, the HIV sample only differs in household size. However, there is significant difference in who is the head of the household and the HIV sample appear to be worse off than the representative sample when looking at asset ownership. For comparison, the summary statistics for the complete sample used in this thesis are listed in table 2.

Furthermore, it is worth noting that the summary statistics from Thirumurthy et al (2006) only detail how the HIV sample differs on observable characteristics. There is also the possibility that households in the HIV sample differ on unobservable characteristics. The implications of this sample bias on external validity will be discussed further in Section 5.

Table 1: Comparison of Random Sample and HIV Sample Households (Thirumurthy et al (2006))

Random Sample ARV Sample Mean Standard Error Mean Standard Error P-Value Number of Households 503 200 Household Structure Household size 6.04 0.13 5.52 0.13 0.02

Average age of household members

Number of children under 18 3.12 0.09 2.80 0.12 0.06

Number of children living away 1.92 0.12 1.68 0.17 0.28

Household Head Characteristics

Female household head 19% 53%

Single household head 22% 51%

Age of household head 47.9 0.68 45.4 0.64 0.02

Asset Ownership

Quantity of land owned (acers) 6.82 0.47 4.87 0.64 0.02

Value of land owned (shillings) 650,237 44,416 572,772 80,052 0.37

Percent landless 13.4% 26.7%

(11)

8 3.2.2 Data Issues with the Transfer Variable

Within the data there was an issue with how the transfer data was coded. When a household reported a transfer from an individual, the interviewer was to assign the individual a number and list the number in the individual roster. However, if the sender was a child of the household head, who was no longer living in the house, they were assigned a number on the household roster. The numbers on the individual roster began at 1 for the first individual and continued in an ascending order. The assigned numbers on the household roster began at 1 for the household head and continued in ascending order for the current household members and for children who were former household members. This has caused an issue in the transfer data as only the assigned number of the sender is listed and not what roster it is from. Therefore, if a household has a member from the individual roster who has the same number as a former household member on the household roster it is impossible to know who sent the transfer. Some of these clashes could be resolved because some former children on the household roster were too young to send a transfer. Yet 14% of households had transfers where it is impossible to

Table 2: Descriptive Statistics for Households in Complete Sample

Mean S.D. Number of Households 826

Household Structure

Household size 5.56 2.76

Average age of household

members 24.55 11.2

Number of children under 18 2.76 1.97 Number of children living away 1.78 2.64

Household Head Characteristics

Female household head 28.48% Married household head 67.54%

Age of household head 47.02 15.17

Asset Ownership

Quantity of land owned (acers) 6.83 12.3 Value of land owned (shillings) 839,510 1,369,889

Percent landless 18.25%

Value of livestock owned

(12)

9

determine the sender. Therefore, the sample used for the estimations had to be reduced to 711 households who have no clashes in the coding of transfer sender.

The elimination of household from the sample is non-random, as small households with children who have formed new households are the most likely to have clashes. Therefore, as a robustness check the results will be re-estimated, once assigning all clashing transfers to the household roster and then again by assigning all clashing transfers to the individual roster. The effect of this data issue on the results will be discussed in Section 5.

3.3 Description of Variables

3.3.1 Identification of Covariate Shocks

A key component of this analysis is to determine if household use networks differently in response to idiosyncratic risk than to covariate risk. In order to conduct this analysis, it is necessary to be able to identify covariate shocks. However, the survey interview did not ask respondents how widespread the shock they experienced was. Therefore, to identify covariate shocks a threshold of households in a village reporting the same shock was set to determine if a shock was covariate. Any threshold comes with an inherent trade-off between certainty of identification and statistical power. As the threshold is raised the certainty that covariate shock have been correctly identified increases, but there is a loss in statistical power as the sample size reduces. The threshold for the results was 60% as it was determined as the optimal trade off. To determine how sensitive the results are to this threshold the appendix contains the main results with the threshold raised in 10% increments up to 100%. This robustness check will be discussed further in Section 5.

3.3.2 Description of Shocks

As noted in the literature review, households in developing countries are at high risk to income shocks. This notion appears to be accurate within the context of this data set. Figure 1 details the percentage of households that experience one or more of each type of shock between 2003 and 2005. As seen in Figure 1 drought and shocks to livestock are the most prevalent over the specified period, at 57.14% and 56.3% of households reporting occurrences respectively. Drought is often reported at high rates in studies (Dercon et al, 2005; Janvry et al, 2006).

The next two most reported shocks are damage to crops prior to harvest, and loss of agricultural inputs. Both shocks occur as part of agricultural production. The remaining shocks have similar

(13)

10

rates of reporting, and occupy a range of 6.58% to 10.65%. The lowest reported shock is a result of crime, specifically home theft.

As well as being the most reported, drought and shocks to livestock are also the only two shocks to be partially covariate. The breakdown of the levels of each shock is reported in Figure 2. As Figure 2 shows, both shocks have higher rates of idiosyncratic occurrences than of covariate, with livestock shocks having a reporting rate of 50.73% for idiosyncratic occurrences to 12.23% covariate and drought with 43.46% of household reporting an idiosyncratic occurrence to 19.01 covariate.

Figure 2: Rates of Reporting for Shocks with Covariate Component,2003-2005 Figure1: Rate of Reporting for Each Shock, 2003-2005

(14)

11 3.3.3 Transfers by Location

A key component of the analysis involves categorising transfers by the location of where the sender resides. This is done because it is suspected that households will want to establish networks that are sufficiently spatially diverse so that the region in which the sender resides has climatic conditions that are uncorrelated with the households. Therefore, four categories are established. The first is ‘In Village’ and comprises transfers received from within the same village as the household. The second is ‘Uasin Gishu County’ and comprises transfers received from outside the village but still within Uasin Gishu County, the same county in which the sample resides. The third is ‘Bordering Counties’ and is made up of transfers received outside of Uasin Gishu County but from counties that share a border with Uasin Gishu. The final category is ‘Further Regions’, and constitutes transfers received from locations that do not share a border with Usain Gishu, it is comprised of Kisumu, Nakuri, and Nairobi counties and transfers received from outside of Kenya. The categories are summarised in Table 3 and represented graphically in Figure 3.

Table 3: Categories of Location for Transfers Received

Transfer Received Area

Category Description of Category

Distance from Regions Capital

to Eldoret (km) In Village: Sender resides in the same village as receiver - Uasin Gishu County: Senders household not in village but inside Uasin Gishu County

(includes Eldoret) -

Bordering Counties: Sender resides in Nandi, Kakamega, and Kitale counties 39-60 Further Regions: Kisumu, Nakuri, Nairobi counties, and transfers received from

outside of Kenya 90+

(15)

12

Figure 4 shows the rates of reported transfers received between 2003-2005 for each location category. As would be expected, transfers received from inside the village are the most reported with 53.39% of households reporting having received one or more transfers from within the village. Although Bordering Counties send more transfers than Usain Gishu County, with 27.12% to 22.15% respectively, it appears that distance to village and volume of transfers are inversely correlated. This is not a surprise as costs associates with transfers increase with distance. This leads to a trade-off that households face between spatial diversity and the cost associated with sending and receiving transfers

3.3.4 Transfers by Relation of Sender to Receiver

Further to decomposing transfers by location, transfers are also decomposed by the relationship of the sender to the receiver. The first category is ‘Non- Family’, constituting a sender who is not associated to the receiver by family. The second category is ‘Family’, as was determined by whether the receiver had indicated they had met the sender through family. The final category is ‘Former Household Members’, and constitutes senders who used to reside within the receiver’s households.

Figure 5 shows the rates of reporting for each category by region transfers were sent from. Transfers received from family make up most of all transfers. Transfers from non-family are reported by 12.2% of households within the village, but drop off sharply outside of the village

53.39 22.15 27.12 9.69 0 10 20 30 40 50 60

In Village Uasin Gishu County Bodering Counties Further Regions

Perc

en

ta

ge

(16)

13

falling to less than 2% for all other location categories. Transfers from former household members range from 3.67% to 6.74%.

.

4 Methodology

This thesis will follow Fafchamps (2003) and Ward and Shively (2015) by employing a household fixed effects method. The methodology will build on these papers by looking at transfers by the region in which the sender resides. This section will introduce the models used for each estimation. The first part of this section will detail why a fixed effects estimation was chosen. The second part will describe each of the econometric specifications and why they are used.

4.1 Assumptions and Econometric Specification

An implicit assumption of the empirical analysis is that the decision-making unit is the household. Through this assumption transfers reported by different household members are aggregated to the household level. The member reporting the transfer can be seen as the point of contact for the sender to the household. In terms of household migration, the decision for a household member to migrate is seen as a decision to maximise the household utility function, rather than the individual. Challenges to these assumptions will be analysed further in Section 5.

An issue that occurs in the analysis of income shocks is that households may differ in their ability to manage shocks. For instance, some households may be less able to manage a drought due to a lower ability, and thus report higher rates of shocks. Consequently, shocks may

12.2 0.98 1.83 1.27 43.88 20.39 23.63 8.58 6.75 3.94 4.22 _2.67 0 5 10 15 20 25 30 35 40 45 50

In Villgae Uasin Gishu County Bordering Counties Further Regions

Perc

en

ta

ge

Non-Family Family Former Houshold Members

Figure 5: Rates of Reporting for Each Location Category, Decomposed by Relation of Sender to Household Receiver, 2003-2005

(17)

14

become correlated with unobservable characteristics that cause endogeneity with the dependant variable. However, fixed effects can improve on an OLS estimation by removing much of this omitted variable bias through elimination of time invariant household fixed effects. This relies on the assumption that variables causing bias do not vary over time. Furthermore, the fixed effect estimation also allows for the removal of time variant effects that are constant across households. This will reduce bias in anything that affects transfers for all households between 2003-2005.

4.2 Models 4.2.1 Transfers

To analyse the role of networks in consumption smoothing the effect of income shocks on transfers received by a household is assessed. To do this the following econometric model will be estimated:

𝑇𝑟𝑎𝑛𝑠𝑓𝑒𝑟𝑠𝑖𝑡 = 𝛼1+ 𝛽1𝐶𝑜𝑣𝑎𝑟𝑖𝑎𝑡𝑒𝑆ℎ𝑜𝑐𝑘𝑖𝑡+ 𝛾1𝐼𝑑𝑖𝑜𝑠𝑦𝑛𝑐𝑟𝑎𝑡𝑖𝑐𝑆ℎ𝑜𝑐𝑘𝑖𝑡+ 𝜕𝑖 + 𝜏𝑡+ 𝜀𝑖𝑡 (1)

𝑇𝑟𝑎𝑛𝑠𝑓𝑒𝑟𝑠_𝑖𝑡 is a binary indicator that is 1 when household i receives a transfer in year t, 𝐶𝑜𝑣𝑎𝑟𝑖𝑎𝑡𝑒𝑆ℎ𝑜𝑐𝑘_𝑖𝑡 is a binary indicator that is 1 when household i experiences a covariate

shock in year t, 𝐼𝑑𝑖𝑜𝑠𝑦𝑛𝑐𝑟𝑎𝑡𝑖𝑐𝑆ℎ𝑜𝑐𝑘𝑖𝑡 is a binary indicator that is 1 when household i experiences an idiosyncratic shock in year t, 𝜕_𝑖 is the household fixed effects term, 𝜏_𝑡 is the time effects term, and 𝜀_𝑖𝑡 is the error term. 𝛽₁ is the change in probability of a household receiving a transfer when a covariate shock occurs and likewise 𝛾1 is the change in probability of a household receiving a transfer when an idiosyncratic shock occurs.

A key objective of the analysis is to determine the use of networks over geographical distances. Therefore, Model 1 will be estimated four times for transfers received from; within the village, Uasin Gishu County, ‘Bordering Counties’, and ‘Further Regions’. If 𝛽₁ or 𝛾₂ are positive and significant for transfers within the village, it will be evidence of some form of informal risk sharing arrangements at the village level. If they are positive and significant for any other category it will be evidence that households are utilising geographical distance for risk sharing arrangements.

(18)

15

To further analyse any risk sharing agreements the estimations of Model 1 will be re-done with transfers further decomposed by family connections. This is done because households will often utilise family ties to overcome trust issues between sender and receiver.

Model 1 is useful for demonstrating an uptake, or downturn, in the amount of transfers due to a shock as it provides estimates for changes in probability of receiving a transfer after shock. However, if only households that regularly engage in transfers increase the value received in a transfer ex post, there will not be an observed change in the probability of receiving a transfer, but transfers would still be used for consumption smoothing. To account for this a second approach is used:

𝑙𝑛𝑇𝑟𝑎𝑛𝑠𝑓𝑒𝑟𝑉𝑎𝑙𝑢𝑒𝑖𝑡

= 𝛼₂+ 𝛽₂𝑙𝑛𝐶𝑜𝑣𝑎𝑟𝑖𝑎𝑡𝑒𝑉𝑎𝑙𝑢𝑒_𝑖𝑡+ 𝛾₂𝑙𝑛𝐼𝑑𝑖𝑜𝑠𝑦𝑛𝑐𝑟𝑎𝑡𝑖𝑐𝑉𝑎𝑙𝑢𝑒_𝑖𝑡+ 𝜕_𝑖 + 𝜏_𝑡+ 𝜀_𝑖𝑡

(2)

𝑙𝑛𝑇𝑟𝑎𝑛𝑠𝑓𝑒𝑟𝑣𝑎𝑙𝑢𝑒_𝑖𝑡 is the log of the value of the transfer received by household i in year t, 𝑙𝑛𝐶𝑜𝑣𝑎𝑟𝑖𝑎𝑡𝑒𝑉𝑎𝑙𝑢𝑒𝑖𝑡 is the log of the value lost from a covariate shock,

𝑙𝑛𝐼𝑖𝑑𝑖𝑜𝑠𝑦𝑛𝑐𝑟𝑎𝑡𝑖𝑐𝑉𝑎𝑙𝑢𝑒_𝑖𝑡 is the log of the value lost from an idiosyncratic shock, 𝜕_𝑖 is the household fixed effects term, 𝜏_𝑡 is the time effects term, and 𝜀_𝑖𝑡 is the error term.

Model 2 allows for an estimation of how the size of shocks affects the size of transfers. Thus 𝛽2 represents the percentage change in transfers given a 1% increase in the value lost from a shock a covariate shock, and 𝛾₂ estimates the same for an idiosyncratic shock. Model 2 will be decomposed the same as Model 1. First by geographical distance and then by relationship of sender to the household head.

4.2.2 Migration

Further to analysing how households utilise networks, this thesis also analyses migration responses to shocks. This is done to determine if networks are formed by migration responses to income shocks. The ideal method to do this analysis would be to use a yearly household fixed effects regression. However, the data on migration is incomplete for waves 1 and 2, meaning the only available migration data is for wave 3. Given that there is a low level of covariate shocks for 2005, migration is analysed by shock nature. Consequently, the chosen specification is a quarterly household fixed effects regression for 2005:

(19)

16

𝑀𝑖𝑔𝑟𝑎𝑡𝑖𝑜𝑛𝑖𝑡 = 𝛼3+ 𝜔1𝑁𝑎𝑡𝑢𝑟𝑒𝑆ℎ𝑜𝑐𝑘𝑖𝑡+ 𝜙1𝑁𝑎𝑡𝑢𝑟𝑒𝑆ℎ𝑜𝑐𝑘𝑖𝑡−1+ 𝜕𝑖 + 𝜏𝑡+ 𝜀𝑖𝑡

(3)

𝑀𝑖𝑔𝑟𝑎𝑡𝑖𝑜𝑛_𝑖𝑡 is a binary indicator for household i in quarter t that is 1 when a household has a member who has moved away from the household. 𝑁𝑎𝑡𝑢𝑟𝑒𝑆ℎ𝑜𝑐𝑘𝑖𝑡 is a vector of binary indicators for household i in quarter t for different types of shocks that are 1 when a household experiences that shock. 𝑁𝑎𝑡𝑢𝑟𝑒𝑆ℎ𝑜𝑐𝑘_𝑖𝑡−1 is a vector of binary indicators for household i in quarter t-1 for different types of shocks that are 1 when a household experiences that shock. Including a lagged value for shocks allows for the possibility that shocks have a delayed effect on migration. For instance, after experiencing a shock it may take a household a couple of months to organise and prepare for a member to migrate.

4.2.3 Livestock Sold

The final part of this analysis is to compare networks to other methods of consumption smoothing. To achieve this, two estimations using binary indicator variables are used. The first estimates the effects of shocks nature on transfers received:

𝑇𝑟𝑎𝑛𝑠𝑓𝑒𝑟𝑠𝑖𝑡 = 𝛼4+ 𝜔2𝑁𝑎𝑡𝑢𝑟𝑒𝑆ℎ𝑜𝑐𝑘𝑖𝑡+ 𝜕𝑖+ 𝜏𝑡+ 𝜀𝑖𝑡

(4)

𝑇𝑟𝑎𝑛𝑠𝑓𝑒𝑟𝑠_𝑖𝑡 represents a binary indicator variable that is 1 when a household receives a transfer. 𝑁𝑎𝑡𝑢𝑟𝑒𝑆ℎ𝑜𝑐𝑘_𝑖𝑡 is a vector of binary indicators for household i in quarter t different types of shocks that are 1 when a household experiences that shock.

The second estimates the effects of shocks, decomposed by nature, on livestock sold:

𝐿𝑖𝑣𝑒𝑠𝑡𝑜𝑐𝑘𝑆𝑜𝑙𝑑_𝑖𝑡 = 𝛼₅+ 𝜔₃𝑁𝑎𝑡𝑢𝑟𝑒𝑆ℎ𝑜𝑐𝑘_𝑖𝑡+ 𝜕_𝑖 + 𝜏_𝑡+ 𝜀_𝑖𝑡

(5)

𝐿𝑖𝑣𝑒𝑠𝑡𝑜𝑐𝑘𝑆𝑜𝑙𝑑_𝑖𝑡 represents a binary indicator that is 1 when a household sells one or more units of livestock. Through the comparison of Models 4 and 5, an analysis on how household form strategies for different shocks can be conducted.

(20)

17

5. Results

5.1 Transfers

Table 4 details the household fixed effects estimation for the effect of an income shock on transfers received, with dependent and independent variables as binary indictors. Column (1) shows that a covariate shock decreases the probability of receiving a transfer within the village by 7.9%, significant at the 5% level. Furthermore, with a constant of 7.98%, the net probability of receiving a transfer within the village after a covariate shock is 0.08%. An Idiosyncratic shock increases the probability of receiving a transfer by 5.54%, significant at the 5% level. The results indicate that households are using within village transfers to deal with idiosyncratic risk ex post. Since covariate shocks affect a majority of households within the village, households are unable to insure each other ex post and in village insurance arrangements breakdown.

Looking to transfers further away from the village there is no statistically significant positive effect for transfers received from within Uasin Gishu County, or from counties that border it. All coefficients for both categories are negative and none are statistically significant. A potential reason that no effect is observed is that households do not consider these regions spatially diverse enough for risk coping purposes, as climatic conditions may be too correlated with climate conditions near Eldoret.

Table 4: Effect of an Income Shock on the Probability of Receiving a Transfer

Dependent Variable:

Binary Indicator for Transfers Received

(1) (2) (3) (4)

In Village Uasin Gishu County Bordering Counties Further Regions Covariate -0.0790** -0.0294 -0.0407 0.0370** (0.0380) (0.0252) (0.0273) (0.0173) Idiosyncratic 0.0554** -0.00212 -0.0128 0.0178* (0.0230) (0.0141) (0.0174) (0.0102) Constant 0.0798*** 0.0697*** 0.0672*** 0.0161* (0.0167) (0.0110) (0.0129) (0.00841) Observations 2,133 2,133 2,133 2,133 R-squared 0.127 0.023 0.036 0.019 Number of hnumber 711 711 711 711

Households FE YES YES YES YES

Time Effects YES YES YES YES

Errors clustered at the household level and are shown in parentheses *** p<0.01, ** p<0.05, * p<0.1

(21)

18

Column (4) in Table 4 represents transfers received from counties for which their capital is over 90km away from Eldoret. Covariate shocks increase the probability of receiving a transfer by 3.7%, significant at the 5% level. This result provides evidence that households are utilising these regions for ex post consumption smoothing. Furthermore, idiosyncratic shocks appear to have a smaller positive effect on the probability of receiving a transfer from ‘Further Regions’, though it is only significant at the 10% level. The results from Table 4 indicate that households are using transfers from outside the village to manage risk ex post. Furthermore, it appears that households are using transfers from ‘Further Regions’ at least in some part to compensate for the inability to use in village transfers post covariate shock.

While Table 4 details the increase in probability of receiving a transfer, Table 5 represents the effect of an increase on the value lost from a shock on the value received from a transfer. Column (1) of Table 2 details this specification for transfers received from within the village. A 1% increase in the value lost from an idiosyncratic shock increases the probability of receiving a transfer from within the village by 0.053%, significant at the 5% level. This result provides further evidence that households are utilising within village transfers as part of their ex post risk coping strategy for idiosyncratic shocks.

The estimations for Uasin Gishu county and bordering counties are listed in Columns (3) and (4) respectively. Again, there is no statistically significant effect for either category, for either covariate or idiosyncratic shocks. Also all coefficients are negative. This reaffirms the results described in Table 4, that households are not using transfers from these counties for any sort of ex post risk coping purpose.

Table 5: The Effect of the Value Lost from an Income Shock on the Value of Transfers Received

Dependent Variable: Log of Value of Transfers Received

(1) (2) (3) (4)

In Village` Uasin Gishu County

Bordering

Counties Further Regions log Covariate -0.0685* -0.0216 -0.0290 0.0509*** (0.0359) (0.0260) (0.0260) (0.0196) Log Idiosyncratic 0.0527** -0.00920 -0.0178 0.0184* (0.0207) (0.0135) (0.0153) (0.0101) Constant 0.644*** 0.608*** 0.590*** 0.110 (0.141) (0.0965) (0.108) (0.0771) Observations 2,113 2,113 2,113 2,113 R-squared 0.121 0.024 0.035 0.023 Number of hnumber 711 711 711 711

(22)

19

The final column in Table 2 represents the results for ‘Further Regions’. As shown a 1% increase in the value lost from a covariate shock increases the value received from a transfer by 0.0509%, significant at the 1% level. This is more definitive evidence that households are utilising transfers from ‘Further Regions’ for consumption smoothing after a covariate shock. This would suggest that households are increasing the value of transfers from these regions to compensate for the fact that in village transfers are unavailable after covariate shocks.

However, it should be noted that the effect of idiosyncratic shocks on the value received from in village transfers, and the effect of covariate shocks on the value received from transfers from ‘Further Regions’, appear relatively inelastic to income shocks. A contributing factor for this could be that households are utilising multiple consumption smoothing mechanisms after a shock in tandem with transfers. Another reason is that there may be an issue with reporting. This arises because respondents were asked to estimate the value lost from a shock, which can be very difficult to estimate. Whereas, transfers are much easier to report accurately as they were often given in cash. Furthermore, respondents may over report shocks as negative events are much easier to remember than events like receiving a transfer.

5.2 Family Estimations

The estimations shown in Tables 4 and 5 were repeated with transfers decomposed into transfers received by family and non-family. Table 6 reports these estimates with dependent and independent variables as binary indicators. As seen transfers from family largely follow the estimates from Table 6. Column (1) shows that covariate shocks have a statistically significant negative effect of 8.85% on transfers from family, and idiosyncratic shocks have a positive effect of 4.66% on transfers from family. The effect from ‘Further Regions’ is significant at the 10% level. Conversely, Table 6 shows that transfers from non-family appear to have no observable effect and are not statistically significant for any region. The results from Table 6 indicate that all transfers are being driven by family times. These results to reaffirm the literature that determined that family ties are key to network connections.

The pattern noted for Table 6 is present for Table 7 also. There is a statistically significant estimation for increases in the value of transfers received in the village after an idiosyncratic shock of 0.042% for transfers received from family. Furthermore, a 1% increase in a covariate shock increases the value of a transfer received from ‘Further Regions’ by 0.041% for family members, which is statistically significant at the 5% level. As with Table 6 there are no statistically significant results for any transfers received for non-family transfers.

(23)

20

Table 6: The Effect an Income Shock on the Probability of Receiving a Transfer, with Sender Decomposed by, Non-Family, Family, or Former Household Member

Dependent

Variable: Binary Indicator for Transfer Received

In Village Uasin Gishu County Border Counties Further Regions

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)

Non-Family Family Former Household

Member

Non-Family Family Former Household Member Covariate 0.00953 -0.0885*** -0.0188 -0.00835 -0.0211 -0.00265 -0.00533 -0.0354 -0.0165 0.00810 0.0289* 0.0117 (0.0196) (0.0338) (0.0157) (0.00569) (0.0245) (0.00968) (0.00612) (0.0269) (0.0129) (0.00805) (0.0155) (0.0103) Idiosyncratic 0.00881 0.0466** 0.00120 0.000918 -0.00304 0.00342 0.00113 -0.0139 -0.00592 0.00237 0.0154 0.00887 (0.0114) (0.0210) (0.00925) (0.00188) (0.0139) (0.00566) (0.00445) (0.0169) (0.00693) (0.00178) (0.0101) (0.00589) Constant 0.0204** 0.0594*** 0.0212*** 0.00547** 0.0643*** 0.0110** 0.00524 0.062*** 0.0192*** 0.00269 0.0134 0.00622 (0.00895) (0.0152) (0.00630) (0.00241) (0.0108) (0.00438) (0.00322) (0.0126) (0.00451) (0.00214) (0.00818) (0.00497) Observations 2,133 2,133 2,133 2,133 2,133 2,133 2,133 2,133 2,133 2,133 2,133 2,133 R-squared 0.009 0.121 0.009 0.003 0.023 0.006 0.001 0.036 0.007 0.003 0.016 0.006 Number of households 711 711 711 711 711 711 711 711 711 711 711 711

Households FE YES YES YES YES YES YES YES YES YES YES YES YES

Time Effects YES YES YES YES YES YES YES YES YES YES YES YES

(24)

21

Table 7: The Effect of Value Lost from an Income Shock on the Value of Transfers Received, with Sender Decomposed by, Non-Family, Family, or Former Household Member

Dependent

Variable: Log of Transfers Received

In Village Uasin Gishu County Border Counties Further Regions

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)

Member

Non-Family Family Former Household Member Log Covariate 0.00905 -0.0775** -0.0119 -0.0105 -0.0111 0.00347 -0.0204 -0.0204 -0.00486 0.00977 0.0412** 0.0207* (0.0192) (0.0318) (0.0158) (0.00652) (0.0252) (0.0107) (0.0255) (0.0255) (0.0132) (0.00964) (0.0174) (0.0117) Log Idiosyncratic 0.0109 0.0418** -0.00091 0.000367 -0.00956 0.00166 -0.0160 -0.0160 -0.00664 0.00547 0.0130 0.00605 (0.0102) (0.0188) (0.009) (0.00142) (0.0133) (0.00562) (0.0149) (0.0149) (0.00642) (0.00387) (0.0094) (0.00544) Constant 0.153** 0.491*** 0.191*** 0.0494** 0.559*** 0.0923** 0.527*** 0.527*** 0.159*** 0.00782 0.103 0.0492 (0.0776) (0.127) (0.0563) (0.0203) (0.0945) (0.0366) (0.105) (0.105) (0.0398) (0.0298) (0.0723) (0.0464) Observations 2,113 2,113 2,113 2,113 2,113 2,113 2,113 2,113 2,113 2,113 2,113 2,113 R-squared 0.009 0.114 0.008 0.004 0.024 0.006 0.035 0.035 0.006 0.005 0.020 0.007 Number of households 711 711 711 711 711 711 711 711 711 711 711 711

Households FE YES YES YES YES YES YES YES YES YES YES YES YES

Time Effects YES YES YES YES YES YES YES YES YES YES YES YES

(25)

22

Transfers received from former household members are statistically significant at the 10% level, with a coefficient of 0.0207. This suggests that transfers from former household members are contributing to the overall risk coping strategy from ‘Further Regions’. The results of Table 7 provide further evidence that family connections are key for transfers sent for consumption smoothing purposes.

The results from Tables 6 and 7 show that transfers received from family are primarily driving the results seen in Tables 4 and 5. Furthermore, there is evidence that transfers from former household members are contributing to transfers from ‘Further Regions’ post covariate shock. This provides evidence that households are utilising family bonds to overcome moral hazard issues that arises from informal insurance. It is also notable that family connections are driving within village transfers as well. This indicates that insurance between non-family connections is not enforceable either within the village or at any geographical distance.

5.3 Migration Responses.

To examine the relationship between migration and income shocks a quarterly fixed effects regression was estimated with the dependent variable of household out migration and independent variables of income shocks by nature. These estimations were done with all variables as binary indicators, as shown in Table 82_{. At first glance Table 8 shows that all} statistically significant coefficients are negative. This means that shocks are only having a negative effect on household migration responses. Column (2) of Table 8 shows the effect of a shocks lagged one quarter on migration. As shown flood and home theft shocks are reducing the probability of sending a migrant by 10.3% and 5.57% respectively. These results are not incompatible with the literature. As Ward and Shively (2015) note, that when choosing whether to send a migrant after a shock a household’s decision will be affected by the change in marginal productivity of labour after a shock. For example, if a flood significantly damages a household’s farm then a household may choose to delay the planned migration of a member to increase labour supply to the farm.

The results of Table 8 fail to find evidence that networks are established through migration ex post. However, it is possible that households are establishing networks ex ante and migrants are sending transfers back to the household contingent on household needs. Further, given that

2_{Estimations of the log of the number of migrants on the log of value lost from a shock were also done and produced results with the same}

(26)

23

there is evidence that transfers from ‘Further Regions’ are partially made up of transfers from former household members, the results observed would be compatible with ex ante network formation.

5.4 Shock Nature and Livestock Sold

Table 9 gives estimates for the effects of shocks, decomposed by nature, on transfers received from different locations. Table 10 gives estimation for the effects of shocks, decomposed by nature, on livestock sold. Both tables have dependent and independent variables as binary indicators. Tables 9 and 10 allow for insight into how households are using networks compared to smoothing consumption via their own means.

Table 8: The Effect of Shocks on Household Out Migration

Dependent Variable:

Binary Indicator for Household Having One or More Migrants

(1) (2)

Same Period as Shock Shocks Lagged One Quarter Drought 0.0112 -0.0167 (0.0213) (0.0232) Rain -0.0491* -0.0174 (0.0255) (0.0476) Flood -0.0359 -0.103*** (0.0473) (0.0381) Crops -0.0375 0.00856 (0.0260) (0.0284) Inputs 0.0533 0.00124 (0.0553) (0.0508) Storage 0.0345 -0.0147 (0.0719) (0.0542) Livestock -0.000670 0.0194 (0.0174) (0.0184) Home Theft 0.0502 -0.0557** (0.0409) (0.0268) Constant 0.0587*** 0.0579*** (0.00593) (0.00612) Observations 4,715 4,715 Number of households 1,186 1,186 R-squared 0.010 0.010 Households FE YES YES

Time Effects YES YES

(27)

24

Table 9: The Effect of an Income Shock, Decomposed by Nature, on the Probability of Receiving a Transfer

Dependent Variable: Binary Indicator for Transfer Received

(1) (2) (3) (4)

In Village Uasin Gishu County

Bordering Counties Further Regions

Covariate Drought -0.0559 0.000620 -0.0515 0.0516** (0.0491) (0.0300) (0.0356) (0.0258) Idiosyncratic Drought 0.0331 -0.00371 -0.0124 0.00991 (0.0311) (0.0211) (0.0215) (0.0128) Rain -0.0259 -0.108*** -0.0561 0.0135 (0.0588) (0.0329) (0.0464) (0.0336) Hail 0.0188 0.0265 -0.00752 0.00323 (0.0546) (0.0392) (0.0381) (0.0251) Flood -0.0585 0.0215 -0.0624* 0.00380 (0.0612) (0.0430) (0.0361) (0.0283) Crops -0.0134 0.0190 -0.00776 0.0209 (0.0381) (0.0271) (0.0288) (0.0128) Inputs 0.0945** -0.0307 0.0135 -0.00437 (0.0440) (0.0264) (0.0302) (0.0221) Storage -0.00589 0.0228 0.00257 0.00330 (0.0577) (0.0401) (0.0427) (0.0220) Home Theft 0.0229 0.0442 -0.0310 0.0410** (0.0458) (0.0330) (0.0320) (0.0189) Covariate Livestock -0.0557 -0.0184 0.0331 0.0105 (0.0554) (0.0428) (0.0389) (0.0278) Idiosyncratic 0.0796*** -0.00376 0.0201 0.00529 Livestock (0.0266) (0.0179) (0.0184) (0.0112) Constant 0.0814*** 0.0679*** 0.0619*** 0.0173** (0.0155) (0.0106) (0.0116) (0.00715) Observations 2,113 2,113 2,113 2,113 R-squared 0.136 0.031 0.042 0.024 Number of hnumber 711 711 711 711

(28)

25

Beginning with drought shocks, households appear to be dealing with idiosyncratic drought by selling livestock, as seen in Column (1) of Table 10. An idiosyncratic drought shock increases the probability of selling one or more units of livestock by 9.71%. There is no statistically significant effect for idiosyncratic drought for transfers received from any region. However, for covariate drought the only statistically significant effect is an increase in the probability of receiving a transfer from ‘Further Regions’ of 5.16%. The reason for this effect is that climatic conditions of transfer senders residing in ‘Further Regions’ are the most likely to be uncorrelated with climatic conditions of the receiver. However, it is less clear as to why households do not appear to utilise livestock sales post covariate drought as they do

Table 10: The Effect of Income Shocks, Decomposed by Nature, on the Probability of Selling One or More

Units of Livestock

Dependent Variable: Binary Indicator: Livestock Sold (1) Covariate Drought 0.0797 (0.0492) Idiosyncratic 0.0971*** (0.0303) Flood -0.0224 (0.0581) Rain 0.00713 (0.0587) Hail 0.0989* (0.0559) Crops 0.0584 (0.0392) Inputs 0.0118 (0.0471) Storage -0.00245 (0.0589) Home Theft -0.00983 (0.0439) Covariate Livestock 0.0470 (0.0562) Idiosyncratic 0.0354 (0.0282) Constant 0.151*** (0.0159) Observations 2,458 R-squared 0.062 Number of hnumber 826 Households FE YES

Time Effects YES

Errors clustered at the household level and are shown in parentheses

(29)

26

idiosyncratic drought, given that there is no statistically significant effect. It is possible that covariate drought is also affecting the local markets for livestock, but further investigation is needed to give a proper mechanism.

Households appear to be utilising within village transfers to deal with idiosyncratic shocks to livestock, with an increase in the probability of receiving a within village transfer of 7.96%, as seen in Column (1) of Table 9. Households, may be adopting a different strategy for idiosyncratic livestock than idiosyncratic drought because shocks to livestock will increase the marginal value of the remaining livestock making in village transfers more appropriate for consumption smoothing. There is no observed consumption smoothing for covariate livestock shock from either Table 9 or 10.

The final two shocks households appear to be utilising their networks for are loss of agricultural inputs and loss from home theft. Shocks to agricultural inputs create an increase in the probability of receiving in village transfers by 9.45%. A potential reason households may favour in village transfers for input shocks is the fact that shocks to inputs occur at the start of the agricultural production cycle. Therefore, households may need cash quickly to replace inputs fast to ensure the production cycle is unaffected. If selling livestock is a more timely and costly process, in village transfers may be more useful. Shocks resulting from home theft appear to be increasing the probability of receiving a transfer from ‘Further Regions’. This is most likely because when households are a victim of crime it can cause an altruistic response from other households within their network. However, there could be other mechanisms causing the observed response to shocks to inputs and home theft, Further research into why households chose transfers or selling assets in response to specific shocks could determine which mechanisms are driving the observed responses.

The results of Tables 9 and 10 appear to show that households are responding differently but appropriately to various shocks. Furthermore, it demonstrates that the level of shock is key in households using spatially diverse networks, with transfers from ‘Further Regions’ being the primary response to covariate drought.

5.5 Robustness Checks

To check how sensitive the results are to the identification of covariate shocks the threshold, which was 60% for the results, is raised in 10% increments, the tables are shown in the appendix. For in village transfers the same conclusions made for the main results could be made while setting the covariate threshold at 70%, but would fail to hold at 80% or greater.

(30)

27

‘Further Regions’ transfers appear to be less robust to increasing the threshold. The only estimates that retain significance are the estimates from Table 5 at 70% threshold. The fact that Tables 4 and 5 reproduce results predicted from the literature for in village insurance arrangements, that covariate shocks make in village transfers unusable, suggests that this trade-off is acceptable.

To assess how robust the results are to the issue of clashing transfers, the estimations were redone first by assigning unresolved transfers to the household roster and then to the individual roster. The estimations are reported in the appendix. The results for inside village transfers are robust to reassigning transfers. The results for ‘Further Regions’ do not hold for binary indicator estimates, but do hold for logarithmic estimates of Table 5. This suggests that the value of transfers is responsive to the value lost from a shock, but the effect on the probability of receiving a transfer is unclear. Furthermore, after reassigning transfers the estimates are lower than reported in Table 5 suggesting an overestimation in the magnitude of the effect. The same pattern is seen for family transfers; the results for in the village are robust to transfer assignments, but the transfers for ‘Further Regions’ are only robust for the logarithmic estimations3. The results pertaining to the logarithmic estimations of transfers sent from former household members from ‘Further Regions’ hold for transfer assignment to the household roster, but not for the individual roster. Consequently, there can be less certainty about the validity of this estimation.

6 Conclusion

The aim of this thesis was to determine how households utilise networks to cope with the risk posed by unanticipated income shocks. Specifically, whether households will utilise networks that are spatially diverse for consumption smoothing after covariate shocks. The results indicate that covariate shocks render village level insurance almost unusable, but that within village transfers are part of the ex post response to idiosyncratic shocks. Also, households are using networks to deal with covariate risk, but only if the sender resides outside a 90km radius cantered at Eldoret. Family connections were found to be key to any use of networks for consumption smoothing purposes and some evidence was found that former household members play a role in sending transfer from ‘Further Regions’, though the estimates were

3_{It should be noted that the logarithmic estimations for family transfers from ‘Further Regions”, after clashing} transfers re-assignment are not significant at the 5% level any more. The t value for assignment to the household roster is 1.95 and for individual roster is 1.77.

(31)

28

only significant at the 10% level. There was no evidence found to support the hypothesis that households establish their own connections via ex post migration.

These results are limited in scope in terms of external validity. This arises from sample issues, caused by an over representation of households with a member who is receiving treatment for HIV. The results could be specific to this sample. Furthermore, it is unknown how having a member of the household receiving treatment for HIV interacts with transfers. The results are also contingent on how aptly covariate shocks have been identified. The chosen threshold was determined to be the optimal trade-off between correct identification and retention of statistical power, and the conclusion drawn from the main results hold when the threshold for covariate identification is raised to 70%, but fail to hold thereafter. A restriction in the data because of an error in how the transfers data was coded, that resulted in an inability to properly identify the sender for some transfers. The main results were robust to assigning unresolved transfers to the different rosters, but they do indicate that the magnitude of some estimations may be overstated. Only the logarithmic estimations for transfers received by family members were robust to transfer reassignment.

The econometric specification allowed for the removal of household fixed effect that were correlated with the dependent variables. However, for these effects to be totally removed from the estimations they must be time invariant. Also, the methodology of the thesis assumes that all decisions are made at the household level, this allowed for transfers to be aggregated to the household level. However, there are challenges to this assumption as research within development economics has found evidence for split management of finances within households, particularly between spouses. For example, Qian (2008) finds evidence that increasing female specific income improves survival rates for girls in the household, while increasing male specific income has the opposite effect. Qian interprets the results as evidence for intra-household bargaining instead of a of a unitary household model.

The results of the thesis reaffirm much of the evidence found in the literature. Specifically, that households are using in village transfers for managing idiosyncratic risk but are unable to be used for covariate risk. They also expand on the literature of networks by demonstrating that there is a threshold distance, found to be a radius of 90km from Eldoret, which when crossed households will use networks to cope with covariate risk. Furthermore, the result that households are also using transfers over great distances has clear policy implications. An intervention of private insurance will clearly have unintended spill over effects into other areas

(32)

29

of the economy, by crowding out the observed informal arrangements. Also, any externalities generated by these networks would be eliminated. Further work could asses the benefits, or costs, of these network externalities, as such research would be vital for any effective policy.

This thesis further demonstrates how key family connections are in overcoming moral hazard issues associated within informal insurance mechanisms. The results are more consistent with a model of reciprocal gifts and transfers, rather than a system of informal insurance with loans between village members. What is less clear from the results is how household migration plays a role in the creation of network connections. Further research into how networks are established, specifically on the role of ex ante migration could give a more complete picture on the role of networks. It could also provide a link between the findings of this thesis and Bigsten’s (1996) work on household circular migration patterns for the same context.

(33)

30

7 Bibliography

Agesa, R. U., & Kim, S. (2001). Rural to Urban Migration as a Household Decision: Evidence from Kenya. Review of Development Economics, 5(1), 60-75.

Bigsten, A. (1996). The Circular Migration of Smallholders in Kenya. Journal of African Economies, 5(1), 1-20.

Central Bureau of Statistics (1999). Kenya 1999 Population and Housing Census. Nairobi: Central Bureau of Statistics

Chetty, R., & Looney, A. (2006). Consumption Smoothing and the Welfare Consequences of Social Insurance in Developing Economies. Journal of Public Economics, 90(12), 2351-2356.

Dasgupta, P. (2001). Social Capital and Economic Performance: Analytics. Stockholm: Beijer international institute of ecological economics.

Dercon, S., & Krishnan, P. (2003). Risk sharing and public transfers. The Economic Journal, 113(486), 86-94.

Dercon, S., Hoddinott, J. & Woldehanna, T. (2005). Shocks and Consumption in 15 Ethiopian Villages, 1999-2004. Journal of African Economies, 14(4), 559-585.

Fafchamps, M. (2011). Risk Sharing Between Households. Handbook of Social Economics, 1255-1279.

Fafchamps, M., & Gubert, F. (2007). The formation of risk sharing networks. Journal of Development Economics, 83(2), 326-350.

Fafchamps, M., & Lund, S. (2003). Risk Sharing Networks in Rural Philippines. Journal of Development Economics, 71(2), 261-287.

Grimard, F. (1997). Household consumption smoothing through ethnic ties: evidence from Cote dIvoire. Journal of Development Economics, 53(2), 391-422.

Günther, I., & Harttgen, K. (2009). Estimating Households Vulnerability to Idiosyncratic and Covariate Shocks: A Novel Method Applied in Madagascar. World Development, 37(7), 1222-1234.

The role of networks in ex post risk coping strategies