Uptake of Insurance-Embedded Credit in Presence of Credit Rationing: Evidence from a Randomized Controlled Trial in Kenya

Purpose – Drought related climate risk and access to credit are among the major risks to agricultural productivity for smallholder farmers in Kenya. Farmers are usually credit constrained either due to involuntary quantity rationing or voluntary risk rationing. By exploiting randomized distribution of weather risk-contingent credit (RCC) and traditional credit, we estimate the causal effect of bundling weather index insurance to credit on uptake of agricultural credits among rural smallholders in Eastern Kenya. Further, we assess farmers’ credit rationing, its determinants and effects on credit uptake. Design/methodology/approach – The study design was a Randomized Controlled Trial (RCT) conducted in Machakos County, Kenya. 1170 sample households were randomly assigned to one of three research groups: 351 farmers assigned to receive traditional credit; 351 farmers assigned to receive RCC; 117 farmers assigned to receive RCC with subsidies on insurance premium; and 351 farmers assigned to receive no credit. This paper is based on baseline household survey data and the first phase of loan implementation data. Findings – We find that 48% of the households were price-rationed, 41% were risk-rationed and 11% were quantity-rationed. The average credit uptake rate was 33% with the uptake of bundled credit being significantly higher than that of traditional credit. Risk rationing seems to influence the credit uptake negatively whereas premium subsidies do not have any significant association with credit uptake. Among the socio-economic variables, training attendance, crop production being the main household head occupation, expenditure on food, maize labour


Uptake of Insurance-Embedded Credit in Presence of Credit Rationing: Evidence from a Randomized Controlled Trial in Kenya 1.0 Introduction
Rural households from low income countries have listed unfavourable weather conditions as one of the most important risks affecting productivity and household resilience (Cole., 2012).
For instance, analysis of household risks from Malawi by Giné & Yang (2009), from Ethiopia by Dercon and Christiaensen (2011) and from India by Cole et al. (2013) confirm that rainfall variability and shocks are the largest source of risk to agricultural productivity and consumption.Moreover, weather risks and climate shocks are critically important constraints to wealth accumulation, particularly for those in rural areas who are either engaged in agricultural activities or have their livelihoods tied to the well-being of the farming sector (Barrett et al., 2007).Consequently, many farmers from developing countries have low capacity to adapt to weather related risks.Covariate impacts can also have a rippling effect throughout the rural and agricultural economies affecting farm workers, input suppliers, entrepreneurs and workers in agribusiness, and providers of non-tradable goods and services in the rural non-farm economy (Carter et al. 2014).Governments and development agencies are not spared as they face sudden demands for relief, reconstruction, and recovery (Carter et al., 2014;Chantarat et al., 2007;Chantarat et al., 2008;Cummins & Mahul, 2009).
To help farmers cope with the impacts of weather shocks, several interventions have been tried.Most of them have been developed under the Climate Smart Agriculture (CSA) umbrella with adaptive capacity, sustainable intensification and Green House Gases (GHGs) mitigation as the main goals (Branca 2012).They range from conservation agriculture (CA), to stress tolerant crop varieties, pests and diseases management methods, promotion of adaptable alternative crops and animals, capacity building and gender mainstreaming as well as innovative ways of timely sharing the right information with farmers, mostly through mobile phones.Such interventions have been developed and promoted mostly by governments and the donor community and are mostly targeted towards promoting new and efficient technologies.
However, the impacts of weather shocks on agriculture are huge and covariate in nature, and at times catastrophic.As such, the public sector cannot solve them on their own.There is need for market-based interventions to (1) help the farmers afford and adopt such technologies and/or practices, and (2) protect them against weather and climate risks.To provide farmers with risk protection, different insurance contracts have been tried where weather index insurance has been developed to particularly provide hedging against weather shocks . Carter et al. (2014) define Index-Based Insurance (IBI) as a major institutional innovation that could revolutionize access to insurance for millions of smallholder farmers.Similar agricultural insurance products have been offered to farmers for quite some time now especially in the developed and industrialized countries although under high levels of public subsidies ( Turvey 2001;Glauber 2004;Carter et al. 2014) (Smith and Watts, 2009)..For the last two decades, index insurance products have been introduced in several low-income countries at individual (micro) and/or institutional (meso) levels but mostly as pilot and/or experimental projects (Carter et al., 2014).
The impact of weather and climate risks on smallholder farming is further compounded by the fact that farmers lack capital and have limited to no access to credit, especially from formal credit institutions.This undermines their capacity to afford/adopt high yielding technologies.Access to credit has, however, been found to be positively correlated to productivity and resilience among the rural farming households, who are mostly small scale farmers (Muhamma & Jan, 2011).Lack of credit constrains optimal use of inputs and limits marketing opportunities which eventually lead to disinvestment and poverty traps (Barrett et al., 2007;Chantarat et al., 2017;Shee & Turvey, 2012;Shee et al., 2015).
In theory, credit is an important resource that allows farmers to expand their operations or adopt new technologies to achieve higher productivity and profitability (Dzadze at al. 2012).
In practice, the associated financial and liquidity risks that come with credit use constrains its adoption.Risks can not only impact supply-side credit rationing, but also demand side riskrationing (Boucher et al., 2008;Verteramo-Chiu et al., 2014).In principal, banks use restrictive lending policies to avoid loss on investments.Because of risk of default, they are hesitant to not only lend but also increase interest rates on those they deem as risky borrowers.They either give less credit than demanded or reject them altogether (Chodechai, 2004;Stiglitz & Weiss, 1981).This leads to credit/quantity rationing, a non-temporary disequilibrium condition with excess demand for credit than the banks are willing to supply.On the other hand, the banks may require relatively high collateral to cover the risks of the credit to be advanced.This may in return expose farmers to unacceptable risk of collateral loss and hence they voluntarily excuse themselves from the credit markets and hence they are risk rationed (Boucher et al., 2008;Boucher et al., 2009).
Part of the problem can be resolved with agricultural insurance.In theory, index-based insurance provides farmers with risk cover and removes the fear of investing in new technologies while credit, on the other hand, would enable them to access/afford those technologies.On its own, insurance may be ineffective since most of the time farmers are liquidity constrained to even afford the premiums that are usually required to be paid upfront (Chantarat et al., 2017) and(Smith andWatts, 2009).Combining insurance with credit is hence crucial to achieve the dual advantage of risk cover and adoption of technologies.Carter et al. (2011) showed that uptake of improved technology will be low if efforts to link credit and insurance are not made.However, financing in agriculture in low income countries is underdeveloped as evidenced by low to zero portfolios of agricultural loans by financial institutions.For instance, in Kenya, even after the government made efforts to focus on agriculture to eradicate poverty and increase food security, Kenyan banks still hold less than 5% of agricultural loans in their portfolios (Maloba & Alhassan, 2019).
To borrowers, business risks increase financial risks, and to lenders financial risks increase business risks.This natural incompatibility is difficult to resolve but bundling insurance with credit can bridge the credit access gap between the financial institutions and rural smallholder farmers.Bundled, or risk-contingent credit are designed to shift risk from both the farmer and the financial institution to the insurance markets1 .
Substantial effort has been made towards integrating insurance and credit for smallholder farmers albeit at theoretical and experimental level.For instance, Giné and Yang (2009) conducted a randomized field evaluation of traditional and rainfall index insured credits in Malawi and found higher uptake of traditional credit than the insured credit.Karlan et al. (2011) attempted to compare traditional credit with a price contingent credit in Ghana to assess if whether risks have any effect on investment.It was however impossible to evaluate this since uptake for both credit types was high and homogenous.They in the end recommended further research but with other risks, such as rainfall variability, with larger sample sizes, training and longer-term commitments to maintain meaningful presence in a market.Carter et al. (2011) theoretically studied the impacts of insurance-linked credit on financial market deepening and small farm productivity and concluded that for index insurance to work, it must be contractually interlinked with credit needed to capitalize the adoption of new improved technologies.Shee and Turvey (2012) developed a risk contingent credit as an alternative to traditional agricultural credit and working with simulated field data from pulse farmers in India, concluded that it provided downside risk protection against commodity price fluctuations.
This paper is part of a wider study dubbed Satellite Technologies, Innovative and Smart Financing for Food Security (SATISFy) implemented by IFPRI, NRI, Equity Bank, APA Insurance and other partners.It aimed at developing a weather index-insurance linked credit and to evaluate its impacts on promoting resilience and food security for smallholder farmers.
Related work has been done with several theoretical prepositions but limited empirical research (Jensen and Barrett, 2017).For instance, Shee & Turvey (2012) proposed a Risk Contingent Credit (RCC) model which implies a credit instrument bundled with index insurance, transferring part or all of the borrower's liability to the lender or insurer if triggered.Giné and Yang (2009) investigated adoption of an operating loan in Malawi where the payoff was determined by rainfall, and Shee and Turvey (2012) proposed downside risk protection for the pulse farmers in India.A recent review by Marr et al. (2016) asserts that only limited empirical research has been conducted on the impact of bundled products such as combination of insurance and credit.They concluded that the extent to which credit suppliers would react to the insured status of farmers is unknown as well as the farmers' preferences when it comes to a mix of financial products.This study aims at making significant contribution in investigating the determinants of uptake of such a bundled financial product.This paper in particular aims to (1) establish the credit rationing and the drivers of rationing among the rural smallholder households, (2) establish the effect of credit rationing on uptake, (3) assess the uptake of both the traditional credit and RCC and (4) measure the treatment (RCC) effect on loan uptake as well as establish other drivers of traditional credit and RCC uptake.
The remainder of this paper is as follows.In section 2, we present the RCT study design, sampling procedures, data collection, and respondents' descriptive statistics.In section 3, we describe the empirical methodology to identify the determinants of uptake of insuranceembedded credit and the determinants of credit rationing.Finally, in section 4, we present and discuss the estimation results, followed by concluding comments.

Structure of randomized control trial
The study was a Randomized Controlled Trial (RCT) conducted in Machakos County, Kenya.
This paper is based on baseline household survey data and the first phase of implementation which took place from April 2017 to June 2018.Five sub-counties were selected in consultation with Equity Bank and provincial administration officials.Thirteen locations, which serve as the clusters for this project, were then selected from the five Sub-counties.Sub counties and locations/cluster selection was purposive considering Equity Bank's coverage and capacity to deliver the proposed products.Later, at the onset of the baseline survey, six villages per location were randomly selected.This was done from villages sampling frame comprising of lists of villages per location which were provided by the Chiefs, the local administrators in charge of locations.Thereafter, the sub-chiefs and village elders from the selected villages provided the research team with household sampling frames including the names of all households under their jurisdiction.Fifteen households per village were randomly selected from the sampling frame giving us 90 households per location and 1170 in total.The baseline survey was implemented from May 16 to June 2, 2017.This was conducted by a team of 30 trained enumerators and supervisors using a pretested questionnaire.Data collection was done electronically using a questionnaire programmed into Tablets using CSPro.
The RCT protocol is illustrated in Figure 1.Following the baseline survey, selected farmers were invited to training events which involved games and roll-playing to explain RCC, costs and benefits, payment structures and limitations.The training events were conducted at the location/cluster level (90 households) and targeted either the household head or spouse.It was during this training that households were randomly assigned to the experimental groups by drawing a chip from an urn (Figure 2) 2 .Treatment included one group to be offered traditional credit (N=351) and a second that would be offered RCC (N=351).A third group, used as a control, received no credit offer (N=351).To gather some insights into the demand elasticity of RCC, a subset of 117 RCC members were offered subsidies of 25%, 50% and 75% of the effective costs of insurance premiums.The training events and lotteries were conducted between September 19 th -29 th 2017.
2 This was done as public lotteries whereby each participant blindly drew a chip once from the urn and the outcome was announced publicly and entered against his/her name in the households' lists.Those who were not in the meeting were picked for by the chiefs, sub-chiefs, village elders or someone who knew them.There were 27 chips written Control, 27 written Normal credit, 27 written RCC, and 3 chips of each subsidy categories (25%, 50% and 75% subsidy on premium).This adds up to 90 chips which were just enough for the selected households in a location/cluster.Although the trigger value and probability distributions differed by sub-county, the actual premiums averaged about 12% across districts.With a 25% loading factor imposed by the insurer, the yield as a percentage of the loan amount was set at 15% for each sub-county, even though each of them had a distinct trigger against which indemnities were to be calculated.

The lending processes
The RCT provided 'access' to credit for 819 farmers but providing access did not necessarily mean the farmer would accept the offer and use the credit.Since many farmers had no prior interaction with the local lender, Treatment farmers were invited to mandatory meetings where they received further training on the bank's loan processes and, where necessary, opened a bank account.This was a bank-led process with the bank applying conventional rules in due diligence to determine credit worthiness.
To ensure that credit was to be used only in production processes the loan amount was not advanced in cash.Instead, farmers were provided vouchers which they used to collect inputs from local Agrovet supply shops within their communities.The Agrovets had been selected earlier in consultation with the farmers.
The maximum loan amount was set at Ksh 10,000, or $100, which was deemed sufficient to provide improved inputs for maize production for one acre for one season.For the RCC group, the Ksh 1,400 insurance premium (14%) was added to the base loan3 .The premium was automatically transferred to the local insurance firm4 .The interest rate applied to the loan balance was 14%/year5 .

Credit rationing and risk rationing
We adapted and modified the direct elicitation method proposed by Boucher et al., (2009) and applied by Verteramo-Chiu et al., (2014) to capture households' rationing status and the process is illustrated by Shee et al (2018).We asked the respondents a set of credit-related questions and used the responses to rate and group them into three groups namely: (1) unconstrained also called price rationed consisting of farmers who either were being approached by banks to take credit, would normally receive full credit they applied for or did not have a demand for credit since they are not capital constrained, (2) quantity rationed implying a supply-side rationing where farmers either get less credit than their demand or they are denied credit altogether and (3) risk rationed comprising of farmers who excuse themselves from the credit markets due to fear of losing collateral or other implications of defaulting on loan repayment.Out of the total sample, 48% were unconstrained, 11% were quantity rationed and 41% were risk rationed.As illustrated in Figure 3, random assignment resulted in a reasonable consistency of rationing typologies across treatment and control groups.
Figure 3: Credit rationing groups disaggregated into control and treatment categories

Households' socio-economic characteristics
Socio-economic characteristics of the sample are provided in Table 1.Consistent with randomized treatments, we find that traditional credit and RCC households were similar on all the baseline characteristics as shown by p-values columns of Table 1.48% of the households were credit unconstrained, 11% were quantity rationed and 41% were risk rationed.71% of the sampled households were represented in the initial training that was offered to explain the products as well as offer some financial literacy and tips on good maize agronomic practices.
66% of the household heads were practicing crop production as their main occupation while only 2% were into livestock production as main their occupation.79% of the household heads were male with an average age of 56 years and 8.6 years of formal education, which implies they had cleared at least primary school education.The average household size was 6 (5.8) persons, which is typical of rural households in Kenya (Bongaarts & Casterline, 2013).
Dependency ratio was 80% implying that only 20% of the household members were of the working age, defined as anyone above 14 years old and below 65 years old.On average, they were spending Ksh 2,280 (USD 22.80) weekly on food, Ksh 4,560 (USD 45.60) monthly on other non-food but frequent items such as fuel, transport, etc, and Ksh 89,920 (USD 89.92) annually on other non-food low frequency items such as clothes, fees, medication, etc.They had access to 4 acres of land whereby most of it (3.83acres) was own land, 0.14 acres were rented in and 0.03 was borrowed land.Generally, renting and borrowing in land is not a very common practice in the research areas.Half of this land was under maize.They had an annual maize production of 833kg per household.This translates to roughly 9 bags of 90kg each, which is the most common item used for storage and measuring cereals in Kenya.Maize took most of the labour in man-days with an average of 190-man days per year compared to beans which took 28 days only.35% of the total farm labour was hired labour with the rest coming from family and exchange labour.
Respondents report that they spent roughly Ksh 4,000 (USD 40) annually on fertilizer.
All animals/livestock in the household summed up to roughly 7 tropical livestock unit (TLU) 6 , which was valued at Ksh 86,190 (USD 861.90) and earned the households Ksh 13,540 (USD 135.40) annually.Households were represented in at least two social groups within the community.Access to extension in this area is very low with an average score of 1.12 out of possible 8.On a subjective welfare scale of 1 (very bad) to 5 (very good), the households scored an average of 2.78 ~3 which puts them at the middle of the scale.This means that, on average, respondents viewed their economic condition as being neither good nor bad in the local context.
To determine risk attitudes, we included in the baseline survey a simple experimental game following Binswanger (1980).First, respondents were rewarded with Ksh 300 (equivalent to average daily wage for the region) for their time.They were then asked if they wanted to risk their reward by participating in a risk game with potentially higher or lower returns, depending on the outcome of tossing a coin once.99% agreed to participate in the risk 6 To estimate TLU ownership per household, we mostly relied on Chilonda & Otte (2006) livestock units conversion table.The TLU column was calculated as a product of the number of animals owned by the households per species and its respective exchange ratio/coefficient in the TLU conversion table.Some species in the data collected were however not present in Chilonda & Otte's (2006)

Empirical framework
Since the assignment of treatments was randomized, the empirical strategy to establish the drivers of credit uptake avoids problems of endogeneity and justifies a distributional assumption of normality (Angrist & Pischke, 2008;Cameron & Trivedi, 2010;Cole et al., 2013;Giné & Yang, 2009;Karlan et al., 2011;Soderbom et al., 2015).Consequently, we investigate credit uptake with the binomial Probit model.
We first estimate a univariate Probit model to measure the treatment (RCC) effect on the uptake (uptake = 1, otherwise = 0), using treatment dummies as predictor variables.Because of the randomization process these treatment variables are the only source of uncommon exogenous variation and are thus valid instruments (Angrist & Pischke, 2008;Soderbom et al., 2015).
The Probit model follows Soderbom et al., (2015): (1) Where  is loan uptake which takes a value of 1 if a household accepted the credit offer and 0 otherwise;  is a factor variable representing treatment dummies which takes a value of 1 if a household 'i' was either in RCC or RCC plus subsidies and 0 if in traditional credit; and  is the unknow treatment effect which we seek to estimate.The symbol Φ identifies a Probit estimator stochastic error term () that is assumed to be normally distributed.A similar model was applied by (Cole et al., 2013) and Karlan et al. (2011) in their investigations into the determinants of uptake of credit bundled with index insurance policies.
Second, we likewise estimate a similar univariate Probit model to establish the effect of credit rationing on uptake: (2) In this model,  is a factor variable that takes the value of 1 if a household 'i' is either quantity or risk rationed and 0 if unconstrained. is the unknown effect of rationing on the uptake that we seek to establish.
Third, to determine the drivers of uptake, we extend the models by relaxing restrictions on individual household characteristics and estimate the multivariate binomial Probit model: (3) where  is a vector of household and farm characteristics (1,2, …, k) added to the model to examine the drivers of uptake and  1 ,  2 ,  3 , …   are their estimated coefficients.
Finally, to determine the determinants of credit rationing, we estimate a multinomial Probit model of the form: Where  represents our categorical outcome variable with j categories, i.e. price, quantity and risk rationed,  is a k element vector of the determinants of rationing included in the estimation, and Φ indicates the multinomial normal distribution assumption of the error term and hence a multinomial Probit model.

Credit Uptake
Based on the conventional wisdom about linked or bundled credit, our results are quite surprising.Overall, 819 farmers were offered credit, either traditional or RCC.Despite a perception of heightened demand for credit, only 33% (267) accepted the credit offer.Two factors likely played a significant role in their decisions.The first was that the loans were restricted to agricultural practices and could not be used for any other purpose.The second factor could be a week or two delay in loan distribution which may have discouraged some farmers from taking a loan.
The larger surprise was the determination that the difference in loan uptake was not substantially different among the treatments with 30% uptake among the traditional credit farmers and an average of 32% among the RCC categories (RCC with and without subsidies).
Among the RCC farmers, those offered insured credit with a 25% subsidy on premium had the highest uptake at 37% followed by RCC without any subsidy and with 50% subsidy which tied at 35%.The lowest uptake was with RCC combined with 75% subsidy at barely a quarter (23%).These results suggest that credit demand is more inelastic than we thought.An overriding consideration is that the majority of participants had never received a formal loan before and so there was significant pent-up demand for credit of any type.In other words, the shadow price of the liquidity constraints facing these farmers was sufficiently higher than the cost of credit even with the additional charge of RCC, which corroborates some existing literature such as Barry et al (1981).
Further, we compare uptake among the credit rationing groups.The unconstrained and quantity rationed farmers had markedly higher uptake than risk rationed farmers, and this was significant at 5%.The unconstrained and quantity rationed farmers' uptake tied at 35% while risk rationed uptake was at 29%. (Fig 4) Figure 4: Uptake descriptive -by treatment categories and rationing group

Effect of treatment on the uptake
To estimate the effect of RCC on up-take, we estimated binary univariate and multivariate Probit models with uptake = 1 and declined credit = 0 on the left hand and treatment dummies on the right.Traditional credit was considered as the base category.When we ran a univariate model, uptake of all RCC credit types was not significantly different from traditional credit.
However, with a multivariate model, which controls for endogeneity and hence more precise, we find a positive and significant effect of RCC without subsidy on uptake.This implies that holding all other factors constant, farmers offered RCC without subsidy were more likely to take the credit than farmers offered traditional credit.This has a marginal effect of 0.068 indicating that RCC without subsidy increased uptake by roughly 7%.Uptake of RCC combined with different levels of subsidies on premium was not significantly different from the uptake of traditional credit as shown in Table 2.We had also tried to compare uptake of RCC with and without subsidies and still the difference was not statistically significant.It is possible that among the up-takers, demand and motivation for credit overshadowed pricing conditions.(Table 2).We later conducted a sensitivity analysis where we repeated the model with baseline controls but dropping the subsidy groups.We see consistent result on the treatment effect on uptake where RCC without subsidies increased uptake likelihood by 6.6%~7% and is still significant at 10%.Annex:1 As mentioned, we found these results puzzling.The point of the sub-experiment, albeit smaller than the whole with only 39 subsidy offers for each of three subsidy levels (N=117), was to obtain some insight into the elasticity of the risk premium when linked to credit.These results suggest that not only is the demand for insurance-linked credit highly inelastic, but that the added value of linking insurance to credit may not be as strong as we had originally thought.
Subsequent work, as yet unpublished, employed a Choice Experiment on a random sample of the original subjects.Those results suggest that the demand is not as inelastic as this subexperiment implies.Nevertheless, these results remind us that the enthusiasm that often accompanies results from theory ought to be dampened when that theory is put to practice.

Effect of credit rationing on up-take
As already reported in section 4.1, credit up-take was similar among unconstrained and quantity rationed households, both at 35%, while the uptake among the risk rationed was 29%.To statistically examine the effect of credit rationing on uptake, we estimated a binary Probit model with uptake on the left hand and rationing on the right.Unconstrained was the omitted base category for the rationing groups.We find that quantity rationing has a positive effect on uptake although this is neither substantial (negligible marginal effect) nor statistically significant.However, the positive sign implies that farmers who are quantity rationed are potentially more likely to take credit if offered compared to their unconstrained counterparts.
This was in line with our priori expectation and the existing theory that quantity constrained farmers have positive notional demand for credit, but they receive less or zero supply.Offered a credit facility, as was done with this RCT, this group of farmers will eventually have effective demand.
As expected, risk rationing has a negative effect on uptake, and this was statistically significant at 10%.This implies that risk rationed farmers were less likely to take the offered credit compared to their unconstrained colleagues.From the marginal effect column, being risk rationed reduced the probability of uptake by 6.3%.This again is in line with existing theory that even when provided with an income enhancing credit contract, a risk rationed farmer chooses to withdraw from the credit market and instead undertakes a less return but more certain/less risky venture.This is due to the fear of the risk of losing collateral or undergoing other defaulting implications.(Table 3) RCC was designed to solve this pre-existing risk rationing condition.However, this may not have been achieved with this first implementation phase since the implementing lender required collateral in line with any traditional loan offered by the bank.This will be explored further in subsequent research, but it is worth noting at this point that although risk-rationed farmers accepted traditional credit at a lower rate than RCC, they did nonetheless accept traditional credit.The essential argument in Boucher et al (2009) is that while farmers may have a demand for credit they do not, or cannot, act upon it, resulting in distortions in the rural credit system.Verteramo- Chiu et al (2014) support this view.In our RCT we use the same direct elicitation and identification approach used in these studies so the relatively high uptake of traditional credit by the risk-rationed group comes as some surprise.One possible explanation is a finding in Verteramo- Chiu et al (2014) that the credit demand elasticity of riskrationed farmers increases as interest rates fall.The base interest rate of 14% actually applied to traditional loans in this RCT may have breached the high inelasticity that comes with higher interest rates on informal loans and rotating savings and credit associations (ROSCA) in the Machakos region which had interest rates as high as 100% according to farmers in focus groups.The relatively low interest rate of 14% could therefore encourage even risk-rationed farmers to borrow.A promising avenue to pursue is an exploration of risk-rationing in the presence of binding liquidity constraints.A starting point might be the proposition that for some risk-rationed farmers, the shadow price of binding liquidity constraints is so high, that the utility gain from relaxing those constraints exceeds the disutility that comes from a natural aversion to borrowing.

The drivers of uptake
In this section, we present the drivers of uptake from the Probit multivariate estimation described in section 4.2 above.Later, we dropped credit rationing and repeated the model to eliminate endogeneity that may arise since the same household characteristics may influence both rationing and uptake.The results from both models were however consistent.To avoid falling into the dummy-trap problem, we omitted the first categories for both treatment and rationing groups; namely traditional credit and unconstrained respectively.
The direction of the relationship between the rationing groups and uptake still holds as explained earlier but this time, after including household characteristics covariates, even risk rationing is not statistically significant.Compared to unconstrained households, quantity rationed households were 4.5% more likely to take the credit offered while risk rationed households were 4% less likely to take.
Training attendance increased uptake probability by 19% and was significant at 1%.A household was 7.4% more likely to take up the credit if the household head's main occupation was crop production.This was significant at 5%.In the estimation excluding rationing, uptake probability increased by 3% with a Ksh.1000 (USD 100) increase in weekly food consumption.
Land ownership was negatively correlated to uptake where increment in own land size by one acre reduced uptake probability by 0.5% when rationing was included, and 0.4% when rationing was excluded from the estimation.Additionally, an increase of hired farm labour by 1% increased uptake likelihood by 0.1% and this was significant at 10%.Revenue from livestock had a positive effect on uptake.An increment of annual revenue from animals by Ksh 1000 increased uptake likelihood by 0.04%.This may not be substantial but was significant at 5%. (Table 4)

Drivers of credit rationing
To determine the factors that drive credit rationing among the households, we used multinomial Logit and multinomial Probit models with rationing as the output variable and household characteristics as the predictor variables.The price rationed or unconstrained group was the base category and therefore we interpret our results in reference to it.We find consistent results from both estimators with small differences in estimated coefficients.Since the coefficients are from non-linear models and marginal effects are not included in the table, we only interpret the nature and significance of the relationships but not the magnitude.
Uptake difference between the unconstrained and the quantity rationed was significantly affected by only 4 out of 16 factors, two positively and two negatively correlated.
Quantity rationing increased with dependency ratio and weekly food expenditure while the same reduced with household dietary diversification index and subjective welfare.This implies that households with more dependants than works and those who spend most of their income on food were more likely to be denied credit or get less than demanded.We reason that, as the number of dependants in a household increases while holding income constant, the proportion of income that goes to food increases.This leaves the household with less surplus funds which can be put into loan repayment and hence their demand for credit, at best, is likely to be met with less supply, or at worst, be turned down.The situation is likely to get worse if dependency increment and income reduction occur simultaneously, which is not foreign to rural households especially with death of heads, spouses and/or other working household members.Dietary diversification and subjective economic welfare are both indices of wellbeing of a household and as expected, as a household's wellbeing improves, the chances of their credit demand being met also increases.These findings also confirm that in the appraisal process, banks collect information related to households' size, income and welfare indicators and consider them when making credit-worthiness verdicts.
On the other hand, risk rationing was significantly affected by 10 out of 16 factors selected in the model.The households whose heads were into agriculture as their main occupation, either crop production or livestock rearing were less likely to be risk rationed, which implies that they were more likely to apply for credit.In rural areas, farming is the main source of income.Those not into farming as their main occupation are mostly into low income casual jobs and therefore have no income generating ventures to apply credit against and for.
A few are into other businesses which still rely on the welfare of the farming households.Risk rationing also reduced with household head's level of education but increased with his/her age.
This indicates that households with younger and more educated heads were less likely to be risk rationed and hence would have an effective demand for credit.Risk rationing increased with dependency ratio, suggesting that households with more dependents than workers were more likely to be risk rationed than households with fewer dependants.Risk rationing also reduced with dietary diversification index and expenditure on food.This insinuates that households who spent more on food and whose consumed more diversified foods were less likely to be risk rationed.This is unlike quantity rationing which increased with food expenditure meaning that households spending more of their income on food do apply for credit more than those who spend less, but they are more likely to be denied credit more than those who spend less on food.Further, risk rationing reduced with land ownership, hired farm labour and household's representation and participation in social groups and/or other social networks within the community.(Table 5)

Conclusions
Weather related risks, particularly drought, is among the major hurdles facing smallholder farmers in Kenya.This is further compounded by capital constrictions, credit inaccessibility and existing credit rationing and risk rationing conditions.There has been much work done towards developing and promoting technologies and practices to help farmers cope with drought, but inadequate capital and lack of weather risk hedging make it hard for small holders to adopt them.To manage drought risk and to improve access to credit, we developed a marketbased solution that bundles rainfall index insurance and agricultural credit and implemented a randomized controlled trial with 1170 households.We assess farmers' credit rationing, its determinants and effects on credit uptake.We also assess the effects of bundling weather index insurance on credit uptake as well as the determinants of credit uptake, including the influence of farm level characteristics.
We find that 48% of the households were price-rationed/unconstrained, 41% were riskrationed and 11% were quantity-rationed.With reference to price-rationed or unconstrained households, quantity rationing probability increased with dependency ratio and weekly food expenditure while the same reduced with household dietary diversification index and subjective welfare.Risk rationing probability on the other hand increased with age of the household head and dependency ratio and reduced with crop production and livestock rearing being the household head's main occupation, household head's education level, dietary diversification index, weekly expenditure on food, land ownership and hired labour.
The average credit uptake rate was 33% with the uptake of bundled credit being significantly higher than that of traditional credit.Premium subsidies did not affect uptake significantly insinuating that the need for credit among those who decided to take was high and the price differences in credit and insurance did not affect their decision.Risk rationing seems to influence the credit uptake negatively.RCC was designed to particularly deal with risk rationing where farmers fail to take credit due to fear of losing their asset or hampering consumption to pay for credit in case their investment fails to yield due to drought.Other nonrandom socio-economic factors which were significantly associated with uptake are, training attendance, crop production being the main household head occupation, expenditure on food, hired labour, livestock revenue and access to credit which had a positive correlation to credit uptake and land ownership which was negatively correlated to uptake.
While these results provide very important insights into credit behaviour and demand of subsistence farmers, our study came with some surprises.The first is our finding that farmers self-identified as being risk-rationed accepted traditional credit at a lower, but only slightly so, rate than RCC.Risk-rationing was a key driver of the RCT with the premise being that RCC would be substantially more attractive than traditional credit.Our reliance on risk-rationing as a motive might have been over-stated.Our sub-experiment which provided insurance subsidy to 117 farm households had a near-insignificant effect suggesting that the demand for the insurance component of RCC is highly inelastic.We have suggested that the opportunity costs of binding liquidity constraints may have been so high that the costs of traditional or riskcontingent credit were low in comparison, and that this liquidity effect might dominate the riskrationing effect.Further research along these lines is warranted.

Figure 1 :
Figure 1: Randomized Control Trial design

Table 1 :
Binswanger's (1980)take risk were then provided with five coin-tossing options to identify where a household stands from being extremely risk-averse to preferring risk or being neutral.FollowingBinswanger's (1980)risk aversion classification, the average constant relative risk aversion score was 0.4 which indicate that on average, the households were moderately riskaverse.Comparing baseline means between control and treatment categories table.In such cases, and since TLU conversions are based on animals' weight, we used the coefficients of the animals in the table whose weight is closest to them.game

Table 2 :
The effect of treatment on credit uptake

Table 3 :
Effect of risk rationing on credit uptake

Table 4 :
Drivers of credit uptake for all categories combined

Table 5 :
Factors of credit rationing -multinomial Logit and Probit with unconstrained as the base category Annex 1: Treatment effect on credit uptake -Sensitivity check by dropping subsidies