Insuring against droughts: addressing issues of trust, transparency and liquidity in the demand for livestock index insurance

ABSTRACT Pastoralists in sub-Saharan Africa are particularly vulnerable to changing climatic conditions and heavily affected by an increasing frequency of severe droughts. To increase drought resilience, livestock index insurance constitutes a promising tool that is aligned to mitigate covariate risks. Uptake rates for these products are, however, relatively low. This article uses choice-experimental data collected in Northern Kenya to analyse preferences for hypothetical livestock index insurance contracts among pastoralists. Results show that pastoralists prefer lower strike levels as well as transparency in the form of regular radio announcements and text messages informing the policyholder about the location-specific index readings. Highly valued are trust enhancing features such as an index certification. We do not find any indication that the option to split the premium payment into two instalments as a way to relax potential liquidity constraints could lead to higher demand.


Introduction
Extreme weather events impose a challenge for many communities in rural sub-Saharan Africa. This is particularly true for the pastoral populations in Northern Kenya, where climate change has reinforced both the frequency and magnitude of droughts (Haile et al., 2019;Herrero et al., 2016;Opiyo et al., 2016;Ouma et al., 2018;Thornton et al., 2009;Uhe et al., 2018). To cope with droughts, members of pastoralist communities oftentimes rely on detrimental strategies such as withdrawing children from school, selling productive assets or producing charcoal, which causes long-term economic repercussions (Janzen & Carter, 2018) and over-exploits existing resources (Opiyo et al., 2015;Speranza, 2010). The development of more sustainable solutions to adapt to climate risks has therefore become an important mission for policy makers, practitioners and academics (Castells-Quintana & McDermott, 2018;Cuni-Sanchez et al., 2019;Haile et al., 2019;Herrero et al., 2016;Opiyo et al., 2015).
Index-based insurance schemes represent a particularly promising tool to alleviate some of the pressure that droughts put on agricultural production (Chantarat et al., 2007;Nieto et al., 2010). As a form of ex-ante risk management, index insurance aims to reduce the vulnerability to poverty before extreme weather events occur. Thereby, payments are triggered based on an external indicator that is correlated with expected losses. A few index-based livestock insurance schemes have recently been introduced to pastoralists in East Africa and although these insurance products are perceived as well designed and have proven to be effective to some extent (Janzen & Carter, 2018;Jensen & Barrett, 2017;Taye et al., 2019), uptake rates are behind their potential and remain weaker than expected (Jensen et al., 2015(Jensen et al., , 2018Takahashi et al., 2020;Taye et al., 2019).
To explain this phenomenon, which is not only found in the context of livestock, but rather across many agricultural contexts globally (Binswanger-Mkhize, 2012; Kunreuther & Pauly, 2004;Kusuma et al., 2019;Matsuda & Kurosaki, 2019;Norton et al., 2014), past research has identified a considerable range of potential barriers of index insurance demand, including high prices, mistrust, a lack of transparency and liquidity constraints (Nshakira-Rukundo et al., 2021;Platteau et al., 2017). Research on strategies that may help alleviate potential barriers through contract characteristics is, however, relatively scarce. This article therefore analyses pastoralists' preferences for index insurance contract design that may address some of these barriers and thereby increase the attractiveness of such products among pastoralists. The design attributes assessed in this article are a certification of the index, a commitment to transparency, the option to pay the insurance premium in two instalments, the strike level, as well as the product's price. It is hoped that some of these design attributes may be able to tackle issues of mistrust, signal product quality as well as reduce liquidity constraints.
To identify preferences for index insurance contracts, this article presents the results of a Discrete Choice Experiment (DCE) that was conducted among 402 members of pastoralist communities in Northern Kenya. The preferences stated by pastoralists allow insights into the potential of the attributes to enhance index insurance demand. Doing so, this article contributes to the existing literature in several ways. While preferences for contracts of agricultural index insurance have gained increasing interest over the past years (Akter et al., 2016;Ali et al., 2020;Castellani et al., 2014;Sibiko et al., 2018), this is the first article concerned with preferences for livestock index insurance. This research gap is considerable, given the challenges that drought events have caused particularly for pastoralists and the fact these communities have historically been marginalized in terms of social services and economic transformation (Amwata et al., 2016). Moreover, to the best of our knowledge, this article is the first to explore the potential of an index certification to increase demand for any type of index insurance product.
The results show that a lower price, lower strike levels, insurance transparency and in particular index certification may be useful strategies to enhance demand. However, we do not find any evidence that the option to spread the payment of the premium represents a promising options to achieve this objective. These results could contribute to the improvement of existing, and to the development of new risk management concepts that are tailored to the preferences and needs of pastoral households.

Potential barriers for index insurance adoption
Index insurance triggers payments based on an external index that is highly correlated with an outcome variable of interest. Predefined payouts are conducted as soon as a certain threshold level of the underlying index is realized. This can be, for example, a certain amount of rainfall, the cumulative temperature over a season or the average livestock mortality in a certain area. Policyholders can expect to receive a compensation without bearing the effort, complexity and time intensity of an ordinary insurance claim (Chantarat et al., 2013). Index insurance is aligned to mitigate covariate risk and eliminate issues of moral hazard and adverse selectiontwo wellknown concepts and barriers in the context of insurance (Janzen & Carter, 2018).
Index-based livestock insurance products have shown to produce positive welfare impacts (Jensen et al., 2015;Taye et al., 2019). Still, demand for index-based livestock insurance remains relatively low. Uptake rates are rarely above 30% of the intended rates and fluctuate strongly (Takahashi et al., 2016). A substantial body of research provides explanations for low uptake rates of index insurance schemes. Detailed overviews are provided by Platteau et al. (2017) and Carter et al. (2017). This article focuses on four key barriers, namely the price, lack of trust, lack of transparency and liquidity constraints.
The price of an insurance is potentially a key determinant of its take-up. The price of a non-subsidized index insurance product is likely to be considerably above the actuarially fair price due to substantial transaction costs and risk premia imposed by the insurer (Carter et al., 2017;Platteau et al., 2017). This alone has the potential to reduce demand substantially, given the high price elasticity of index insurance demand (Cole et al., 2013).
Additionally, households may face liquidity constraints due to imperfect credit markets, which may prevent them from purchasing insurance (Eling et al., 2014;Platteau et al., 2017). It is possible that households without access to credit are actually more attracted to insurance since it offers an alternative strategy to smooth consumption (Gollier, 2003).
However, if potential subscribers experience liquidity constraints precisely at the time when the insurance is offered, financial means for the purchase may just not be available.
The quality and adequacy of contract design represent another key driver of index insurance demand. To reduce basis risk, which refers to the risk of experiencing a livestock loss and yet not receiving any compensation due to imperfect correlation between the index and the losses suffered by the policyholder, and the impact of basis risk on demand, the underlying index and the associated strike level are supposed to be as accurate as possible. This means that the index should be highly correlated with the insured loss and based on transparent data sources that are cannot be manipulated (Jensen et al., 2018). The strike level determines how likely insured individuals will receive compensation. A greater frequency is associated with reinforced demand (Stein, 2018), but the strike level also determines the actuarially fair price of a contract. This leads to a trade-off between payout frequency and premium that requires to be balanced carefully (Chantarat et al., 2013).
A lack of trust can also limit the demand for index insurance. Two aspects can be distinguished here: trust in the insurance product itself and trust in the institutions that are involved (Patt et al., 2009;Platteau et al., 2017). If customers perceive the configuration of the underlying contract and the responding index as trustworthy, they may be more likely to purchase insurance. Trust in the insurance product may be fostered by the experience of payouts and financial relief during an event of loss (Churchill, 2002). This experience can be from direct payouts that oneself experiences or from payouts in one's network (Cole et al., 2013). Should customers experience losses without payouts from the insurance, which is possible due to basis risk or fraud, the insurance product might be considered of low quality and perceived as not trustworthy (Platteau et al., 2017). Subscribers' understanding of the concept of index insurance and the specific insurance contract are therefore crucial. Trust in the provider of insurance is also paramount for demand (Carter et al., 2017;Platteau et al., 2017). If an individual has a history with a financial institution, which involves bad experiences, her levels of trust are likely to be lower. In turn, increasing levels of trust can be established, and subsequently demand enhanced if previous interactions with a positive progression for the agent have preceded (Hill et al., 2013) or if a product has been endorsed by a third party (Cole et al., 2013).

Materials and methods
To examine pastoralists' preferences for livestock index insurance contracts, this research applies a Discrete Choice Experiment (DCE). DCEs represent a method within the stated preferences approach and allow drawing conclusions from previously non-articulated preferences for real choice decisions (Ward & Makhija, 2018). In a choice experimental setting, respondents are confronted with a series of choice sets. These choice sets entail paired alternatives of products or in this case contract versions. The alternatives differ in a certain number of features, so-called choice attributes, which are defined by at least two levels and are represented by either a quantitative or a qualitative variable (Lancsar et al., 2017). Variation of these attribute levels allows for creating a great variety of different contract designs. Subsequently, a variety of choice sets or purchase situations can be established and it can be experimentally tested, how potential policyholders would react to different product or contract designs and how much utility the specific product or contract attributes create (Castellani et al., 2014).
The experimental approach via DCEs can be applied when real market data of revealed preferences is not available (Akter et al., 2016). It is, therefore, particularly suitable for the elicitation of preferences towards novel product designs, as it is the case in this article. Furthermore, DCEs are able to reduce some of the hypothetical bias that is typically associated with stated preference methods (Penn & Hu, 2018).

Choice experiment design
Previous studies using DCE in the context of index-based insurance suggest that such experiments are useful to reveal contract preferences (Akter et al., 2016;Ali et al., 2020;Castellani et al., 2014;Liesivaara & Myyrä, 2014;Sibiko et al., 2018). Using a hypothetical contract design is particularly suitable in the given context since the variation in contract characteristics of existing index-based livestock insurance is very limited. Further, some of the contract characteristics to be analysed in this study are not yet realized in any existing index insurance products. Nonetheless, the general parameters of the insurance contracts used for the DCE can be chosen to resemble existing insurance contracts which ensures that albeit being hypothetical, contracts are credible and as realistic as possible.
General parameters refer to the temporal structure and the payout structure of the contract, which are held constant over the choice sets. These parameters are closely aligned to already existing index-based livestock insurance products that are available in regions close to the study area. The contract covers 1 year with two potential payout periods. The first by the end of the long dry season and the second after the short dry season, which is in March and October, respectively. The payout of the insurance is linear in the average livestock mortality rate in the area that is forecasted based on a Normalized Difference Vegetation Index. Further, the payout structure refers to an insurance contract that covers one Tropical Livestock Unit (TLU) valued at KES 14,000.
The attributes that are included in the DCE are the premium rate, the modality of premium payment, the strike level, transparency and the index certification. An overview of these attributes with their respective levels is given in Table 1.
The first attribute included in the choice set is the premium rate of the hypothetical livestock index insurance contract. The inclusion of the premium rate as a determinant of demand is based on previous theoretical and empirical research (Clarke, 2016;Cole et al., 2013;Karlan et al., 2014;Takahashi et al., 2016Takahashi et al., , 2020. The lowest possible premium rate for a hypothetical insurance contract is KES 1000, followed by KES 1300, KES 1600 and KES 1900 as the highest possible premium rate. These levels are chosen based upon the already existing index-based livestock insurance products. The second attribute of the DCE covers the modality of the premium payment. The second attribute of the DCE covers the modality of the premium payment. Two levels are defined for this attribute: in one level, the hypothetical premiums have to be either paid all at once. In the other level, the policyholder has the option to pay the entire premium at once or spread the payments over two months. Spreading the premium would mean that the pastoralist pays 50% of the premium each month, without changes in the total amount. The choice to include this attribute is based on two rationales. First, pastoralists may value the option to commit themselves to insurance purchase by paying a first instalment, since the splitting option stipulates that no insurance cover is provided and that the first premium is retained by the insurer, if only the first instalment is paid. Similar strategies to provide options for commitment have shown to be effective, for example, in the case of saving accounts (Ashraf et al., 2006). Next to this mechanism, splitting the rate might also respond to liquidity constraints. Pastoralists oftentimes have little financial savings (Karanja Ng'ang'a et al., 2016). Transferring money from one period to another is thus challenging. Past research shows that such liquidity constraints can indeed dampen the demand for insurance (Cole et al., 2013(Cole et al., , 2014Gaurav et al., 2011). Spreading premium payments has therefore recently been suggested as a way of relaxing liquidity constraints (Platteau et al., 2017).
The third attribute included in the DCE is the strike level of the hypothetical insurance contracts. Again, the strike level refers to the average livestock mortality rate above which payouts are triggered. Hence, it determines the magnitude of loss that policyholders have to manage themselves (Sibiko et al., 2018). As mentioned in Section 2, the strike level must be well defined and reasonable to estimate the index probability distribution for adequate risk exposure analysis and pricing (Chantarat et al., 2013). Put differently, the strike level can influence the accuracy of the contract and thereby affect basis risk.
Whereas for the commercially sold index-based livestock insurance product in Northern Kenya, the strike levels have mostly been set at 15% (Jensen et al., 2016), the levels chosen for the hypothetical contracts are 10%, 15% and 20%. This choice is based on the finding that a variation of the strike level in the range of 10-25% is unlikely to affect the accuracy of the index that underlies an already existing index-based livestock insurance in the region (Jensen et al., 2016). In other words, by varying the strike level within this range, an impact on basis risk is unlikely. Jensen et al. (2016) therefore argue that the strike level is an easily changeable contract parameter of livestock index insurance, which can be adapted to the preferences of index insurance buyer or seller. The idea that farmers may have preferences in this regard is shown by Sibiko et al. (2018), who find that farmers prefer index insurance contracts, in which payouts are triggered at lower absolute strike levels, although insurance premiums are higher. The fourth attribute of the DCE addresses transparency in terms of information flow. It suggests that the hypothetical insurance contract is endowed with the option of receiving information regarding the location-specific satellite readings and whether a threshold for payouts has been reached. The provision of adequate and steady information, and hence transparency, aims to react to negative attitudes towards uncertainty and is supposed to enhance trust in the insurance product as well as in the insurance provider. In line with this argument, previous studies suggest that the insurer's trustworthiness can be a key determinant of insurance enrolment (Cai et al., 2015;Cole et al., 2013;Giné & Yang, 2009;Karlan et al., 2014). Patt et al. (2009) argue that participatory processes and communication encourage building and enhancing trust in the insurance provider and the product itself. Further, it helps the policyholder strengthen confidence in their decisions to purchase insurance. The idea that farmers value regular communication and information from the index insurance provider is also supported by Sibiko et al. (2018).
Three levels are chosen for this attribute: in the first case, the contract does not entail any information services. In the second case, weekly radio broadcasts which inform about the location-specific satellite readings and which communicate whether a threshold for payouts has been reached. In the third case, policyholder do not only have the possibility of obtaining information via radio, they also receive weekly text messages (SMS) via phone, which also entail the above described information.
The fifth and last attribute included in the DCE contains a certification of the index in terms of adequacy and transparency. The inclusion of an index certification in the hypothetical contract is motivated by the importance of trust and perceived quality for index insurance demand. A certification of the index particularly aims to improve the perceived quality of the insurance product as well as the perceived trustworthiness of the provider. This is, subsequently, expected to enhance demand for the livestock index insurance, since the certification mediates certainty and a certain quality standard. Carter et al. (2017) as well as Jensen and Barrett (2017), for example, criticize that quality certification standards for index insurance are neither existent, nor enforced in developing countries, making them a 'hidden trait'. A certification of the underlying index of an insurance has been suggested both explicitly (Patankar, 2011;Platteau et al., 2017) and implicitly (Karlan et al., 2014) as a promising idea to increase demand.
The certification is expected to increase trust by granting two distinct assurances: first, the certification assures that the index is calculated by an independent research institution and cannot be manipulated by the insurer. This addresses the fear that the insurance company could deliberately manipulate the index. This is practically feasible since the data to calculate the underlying index of the existing index-based livestock insurance products are available to the public in near-real time and therefore objectively verifiable (Chantarat et al., 2017). The institution that offers the certificate can therefore easily check the truthfulness of the index reading.
Not only is it crucial what the index certification covers, but also who the issuer of the certificate is. Local governments are potential issuers of the certificate (Patankar, 2011;Platteau et al., 2017). To address potential distrust in the local governments itself, a trusted non-governmental organization (NGO) might also be a reasonable issuer. The NGO is expected to reflect independence from the local governance, but also from the insurance industry. Hence, the attribute entails three levels: first, hypothetical contracts are left without any index certification. Second, the index is certified by the government, and third, the index is certified by a trusted NGO.

Implementation of the experiment
As explained earlier, different combinations of attributes need to be created to generate distinct choice sets. In the given setting with five attributes and the respective attribute levels, a full factorial design gives a total of 216 (4 1 × 3 3 × 2 1 ) possible alternatives. Using the software Ngene (ChoiceMetrics, 2012), 18 choice cards with two alternatives per choice card were then created.
An exemplary choice card is presented in Figure 1. On each choice card two unlabelled insurance options and a no-choice alternative are shown, such that the respondent can also choose to make no purchase. With the inclusion of the nochoice option, unintended hypothetical purchase decisions are avoided. It further increases realism since surveyed pastoralists are not forced to choose between the experimentally designed alternatives.
To reduce measurement errors due to participants' fatigue, each individual randomly shows 6 out of the 18 choice cards. Hence, three different versions of the survey were created. Thereby, the order of attributes did not randomly switch for each respondent but stayed the same on each choice set. The structure and visual layout of the choice card was thus always the same and resembled the examples given in the plenary explanations. While this may have introduced some bias with regard to relative attribute importance, we assume that this bias is outweighed by the benefit of having a persistent structure to help the respondents compare the alternatives.

Data collection
The DCE was conducted in Turkana County, located in the northwest of Kenya, from July to August 2018. A two-stage sampling strategy was applied: in the first step, four clusters of villages with approximately 15 villages in each cluster were selected. From each cluster, four to five villages were randomly chosen. In a second step, 20-26 households within each village were randomly invited to participate in the data collection. The final sample includes 402 respondents from 17 villages.
After receiving consent, a prayer was offered by one of the spiritual leaders of the visited village. A small socio-economic survey with each participant followed. Afterwards, we gathered all participants and explained the DCE to the respondents in the group. We first clarified that neither our team nor the university that we represent had any intention to sell any type of product and that participation in the survey had no effect on future eligibility for insurance products. We did, however, mention that similar types of insurances were commercially available in other counties of Northern Kenya and that efforts were currently made to develop similar types of products also for the pastoralists in the region in the near future. We further explained that our research aims at helping to improve the design of such future contracts and that truthful and well thought through answers in the survey and the experiment may help tailor insurance contracts to their preferences and needs.
Since index insurance presents a new tool for risk management among pastoralists, a high degree of unfamiliarity with the concept was assumed. To minimize measurement errors and misinterpretation of the insurance or any of the attributes, we devoted a substantial amount of time explaining index insurance in general and each attribute in particular. One promising method to explain index insurance to pastoralists in rural Kenya was invented and tested by McPeak et al. (2010). Since this method is very detailed, labour intensive and takes one full day, the full method proved to be impractical for the given data collection process. However, some aspects of the method, for example the terminology adapted to fit the pastoralist environment, were used. Furthermore, we utilized some of the concept explanations presented in an educational comic designed particularly for pastoralists in Northern Kenya (Waweru et al., 2011). We then conducted the choice experiment, as explained above with each respondent individually. All communication with participants occurred in the local language and questions were encouraged during all stages.

Econometric strategy
Econometric analysis of DCEs is based on the random utility model (RUM). It is the underling structural model encompassing discrete choice behaviour and derived under the assumption of utility-maximizing behaviour of the decision maker (McFadden, 1973). According to Train (2009), the derivation of choice probabilities can be described as follows: the RUM parses indirect utility U n into an observable, deterministic component V n and an unobservable, stochastic component 1 n for each respondent n: 1 nj is following a random distribution with density f (1 n ). In the given context, an individual n faces a choice of one livestock index insurance contract alternative from a finite set C with J alternatives and with a vector of attributes X nj . The probability that alternative j will be chosen is equal to the probability that the utility gained from its choice is greater than or equal to the utilities of choosing another alternative j in C. Thus inserting (1) and rearranging elements of inequality, the choice probability yields Equation (3) can be interpreted as a cumulative distribution. In particular it describes the probability that (1 nj -1 ni ), i.e. the difference of the amount of unobserved utility between alternative j and the chosen alternative i is lower than (V ni -V nj ), i.e. the difference of the amount of observed utility between the alternatives. After specifying the distribution of the unobserved amount of utility f (1 n ), the choice probability can be adapted to with I(·) becoming 1 if the expression in parentheses is true and 0 otherwise. The choice probability is a multidimensional integral over f (1 n ), the density of the unobserved amount of utility. The specification of f (1 n ), which varies due to different assumptions about the distribution of the unobserved amount of utility, arrives at different discrete choice models. This article employs three different models, namely the Multinominal Logit Model (MNL), the Random Parameter Logit Model (RPL) and the Latent Class Model (LC). The MNL model was the first established discrete choice model (McFadden, 1973) and provides the foundation for many extensions to more sophisticated models (Lancsar et al., 2017). The MNL does, however, have a few restrictive assumptions that are likely to be unrealistic in the context of DCEs (Elshiewy et al., 2017;Lancsar et al., 2017). It is therefore primarily shown as a robustness check. The RPL model is able to overcome some of the restrictions and therefore is likely to represent an overall more appropriate model in the given context. The LC model lastly is applied to further investigate sources of preference heterogeneity in the sample by distinguishing between distinct classes of respondents that reveal different preference patterns. It is possible that, for example, one class of respondents has strong preferences only for the price, the strike level and the payment modality of a livestock index insurance contract, whereas another class has strong preferences for transparency and index certification. The application of the model has shown to be useful in the stage of new product development since it can identify different types of buyers based on their stated preferences (Akter et al., 2016;Birol et al., 2009). The theoretical derivation of the models as well as the iterative process to determine the most suitable model for the here presented analysis are presented in detail in the Appendix, as well as in Train (2009).

Descriptive results
Descriptive statistics of the sampled respondents are presented in Table 2. The average of 1.1 years of education point towards a poor overall level of education in the study area. The average herd size owned by the household is 33 TLU, which reinforces the claim that livestock is a centrepiece of the respondents' livelihood. In terms of financial endowments, only 32% of the surveyed pastoralists report to have cash savings available. This indicates that pastoralists might face liquidity constraints, a circumstance which is supposed to be addressed through the option to pay the premium payment in two instalments. Respondents' trust towards financial institutions presents a mixed picture. Approximately 39% of the respondents report to have little or no trust towards such institutions. This implies that mistrust could be an issue for a considerable proportion of the pastoralists in the region, but it does not seem to be omnipresent. As further shown, almost 80% of the surveyed  pastoralists name drought or insufficient rain as the main reason for livestock loss, whereas disease, raiding and conflict are mentioned only by 15%, 5% and 1%, respectively. This tendency supports the importance of addressing drought-related issues faced by pastoralists.

Results of the discrete choice experiment
This section provides results obtained by the discrete choice model estimations. From the full sample of 402 respondents, choices of only 381 respondents are considered since 21 did not differentiate their responses, meaning that for each of the 6 choice sets faced the same contract option was selected. This behaviour is also known as straightlining and can represent a source of systematic bias (Loosveldt & Beullens, 2017). Eventually, 6858 observations are generated since each of the 381 respondent was confronted with 6 choice cards that include 3 options each (381 × 6 × 3).
Results of the MNL model are reported in the first column of Table 3. All coefficients are statistically significant such that all attributes can be assumed as relevant in the decision-making process. However, Hausman specification tests reveal the assumption of independence of irrelevant alternatives, which underlies the MNL, to be violated. In the following, we therefore focus on the results of the RPL model and the LC model estimates.
The results of the RPL model are reported in the second column of Table 3. Assessing the overall goodness-of-fit, the RPL model is a considerably better fit compared to the basic MNL model. Both the AIC and the BIC are lower and the Pseudo R 2 is higher. While the mean parameter for the option to split the payment turns out not to be statistically significant, all the other mean parameters are statistically significant and keep the same sing as in the basic MNL model. Hence, the relevance of the payment modality cannot be confirmed, whereas every other attribute seems to matter in the decision-making process. Further, the negative sign of the alternative specific constant (ASC) and significance at the 1% level indicate that respondents have a positive general attitude towards livestock index insurance contracts. The ASC is intended to capture systematic influences of the individual alternatives that are possibly not modelled adequately. In this specification, the ASC is defined such that it describes how the pastoralist value the nochoice option, when observed factors are controlled for.
The mean parameter for the premium rate attribute is negative, which is a direct consequence from the log-normal distribution that was assumed for the premium rate attribute (see supplementary material) and is statistically significant at a 5% level. As already mentioned, the mean parameter for the payment modality attribute turns out to be insignificant in the RPL model. Hence, no clear statement can be drawn with respect to a preference for the payment modality. Turning towards the strike level attribute, the mean parameter is negative and statistically significant at the 1% level. Hence, holding other attributes constant, pastoralists seem to prefer contracts with lower strike levels. This implies contracts with an average livestock mortality rate of 10% to be preferred over contracts with 15%, which are, in turn, preferred over a contract with 20% average livestock mortality.
Considering the transparency attribute, the mean parameters for both levels, weekly radio broadcasts and weekly text messages in combination with weekly radio broadcasts, attain positive signs. This outcome suggests that regular communication from the insurers side, which informs policyholders about the location-specific satellite readings and which communicates whether a threshold for payouts has been reached, is valued by the pastoralists. However, the coefficient of the weekly text messages in combination with radio broadcasts is smaller in magnitude implying that contracts offering only radio broadcast are more likely to be chosen over contracts offering both communication channelsholding other attributes constant.
In terms of the index certification attribute, sign and significance of the mean parameters reveal that both levels, a certification issued by a local government, but also by a trusted NGO are preferred over livestock index insurance contracts without any certification. Compared to the basic MNL model, coefficients increased in magnitude in the RPL model, but significance of both coefficients remains at the 1% level. These estimates suggest that pastoralists value livestock index insurance contracts in which practices of the insurer and radio announcements regarding the current index reading in each zone are certified through a trusted third party. However, the coefficient for the NGO certification is higher in magnitude implying that contracts entailing this certification type are more likely to be chosen over contracts entailing a certification by a local government ceteris paribus.
Considering the standard deviation parameters, which are presented in the lower part of the second column of Table 3, all parameters appear to be significant. This points towards a considerable preference heterogeneity. Therefore, an LC model is applied with the aim to determine potential sources of preference heterogeneity. As explained in the previous section, the LC model can be applied to identify different classes, or groups of respondents that exhibit different preference patterns. Thus these classes have homogeneous preferences within the class, but heterogeneous preferences between classes. The results of the LC model are shown in Table 4. Comparing the goodness-of-fit measures, namely the BIC and the AIC, for models with 2 up to 10 preference classes, the ideal number of distinct classes appears to be three. As reported in Table 4, the average probability of belonging to the first class is 46.6%, while it is 40.1% for the second and 13.3% for the third class. To obtain a quantitative measure of how well the model does in differentiating the specific preference classes, the average of the highest posterior probability of respondents' class membership is calculated. Since this average is estimated to be 90%, it can be stated that the model does very well in distinguishing among different underlying taste patterns for the observed choice behavior. The coefficient for the premium rate attribute is assumed to be class-invariant and, thus, the same for all three classes. Like in the RPL model, the coefficient is negative, but only significant at the 10% level. Further, among all classes, the ASC has a negative coefficient implying a positive attitude for livestock index insurance contracts in general. However, in the second class the coefficient turns out be insignificant, which renders a clear statement for this class difficult.
For respondents assigned to the first class the model suggests that, next to the premium rate, the strike level, the payment modality and the certification are relevant contract features, whereas transparency does not seem to create decisive utility. Like in the RPL model, the sign of the strike level attribute is negative, implying a preference for lower levels. The positive sign for both certification options reveals preference over no certification, although, considering the magnitude of the mean parameter, preference for a certification issued by an NGO seems to be even stronger, which is in line with results of the RPL model. Counter-intuitively, respondents seem to experience disutility when offered the option to pay the amount over a 2-month-period.
In the second class, only the coefficients for the certification attribute turn out to be statistically significant. Hence, for respondents of this class, only a certification can generate utility, whereas for all the other contract features no clear preferences can be observed. In line with results of the RPL model, preferences for a certification issued by an NGO seem to be even stronger compared to a certification issued by the local government.
In the third class, all attribute levels reveal statistically significant coefficients, such that all attributes seem to be relevant in the decision-making process. Except for the preference for paying the entire premium at once, results are in line with the estimates of the RPL model. Hence, lower strike levels, index certification and transparency are valued. Considering the latter attribute, it seems that the third class is the only one significantly preferring this contract feature. Thereby, communication via radio is even more preferred over no communication than via radio and text messages.
In a further step, we compare general and context specific characteristics among the three classes to explore potential sources of preference heterogeneity. Results of these comparisons, which are conducted via t-tests, are shown in Table 5. The classes are very similar in terms of gender distribution, age, educational levels, household sizes and risk aversion. Our results therefore cannot confirm the findings of Akter et al. (2016) who found gender and in particular gender differences in trust towards insurance institutions to be a key source of preference heterogeneity.
Noteworthy differences between the classes can be found for average income and herd size, which are highest for the second class and lowest for the first class. Thus income and herd size might be a source of preference heterogeneity. Furthermore, we find that trust in financial institutions is smaller in class one compared to class two, which can be linked to the generally high preference for an index insurance among class one. Overall, however, no clear picture regarding respondents' characteristics and the class affiliation can be identified and it needs to be stated that the comparison of the average characteristics does not provide any causality with regard to the preference patterns.

Discussion
The results presented in the previous section have several implications for the design of index insurances for livestock, some of which deserve further discussion. A first implication is that having the option to spread the payment of the premium into two instalments does not seem to be a promising strategy to increase demand. Although there is no concrete empirical evidence that splitting premiums over 2 months is valued in the given context, this result is rather counter-intuitive, since one would expect individuals to gain utility from at least having the option to split the payment, instead of being obliged to pay the entire premium at once. This implies that either liquidity constraints are negligible, or spreading the payment over 2 months is not sufficient to address liquidity constraints.
A second implication is that lower strike levels are preferred over higher strike levels ceteris paribus. Again, this is intuitive since lower strike levels imply that the expected payout of the insurance increases. However, this also means that insurance companies are likely to demand a higher price for the an insurance with lower strike levels. Whether an actuarially fair index insurance with low strike levels is preferred over an actuarially fair index insurance with high strike levels therefore presents an important question in this context. Since the estimation of an actuarially fair premium of an index insurance for the study region goes beyond the scope of this paper, we leave this question to further research. A third implication is that the provision of regular information is a promising tool that could make livestock index insurance more attractive and foster the demand thereof. It is assumed that transparency achieved by the regular information mitigates ambiguity aversion, but also enhances trust in the insurance product and the provider. The most sensible channel to spread information appears to be radio broadcasts, whereas additional text messages seem to be somewhat less attractive to pastoralists. This is in line with previous research (Sibiko et al., 2018).
The fourth and most important implication of this research is that the inclusion of an index certification could constitute a very promising novel approach to increase demand for livestock index insurance products. The assumed primary mechanism is increased certainty and trust in the product's quality and in the insuring institution. The idea of an index certification was already raised by Patankar (2011) and Platteau et al. (2017), but this research presents first evidence that this concept could have considerable potential.
Moreover, results suggest that an index certification is most attractive to pastoralists when issued by a trusted NGO. However, contracts including a certification issued by the local governments could also be successful. Agents involved in livestock index insurance development should thus engage in collaborations with third parties, especially acknowledged NGOs, to develop index certifications that are accepted by the target group.

Conclusion
In recent years, livestock index insurance has emerged as a promising tool for pastoralists to adapt to climate risks and to increase drought resilience. Demand for such insurance products remains, however, low. Several non-price factors such as risk and ambiguity aversion, but also mistrust, a lack of concept or contract understanding as well as liquidity constrains are assumed to play a crucial role in uptake decisions by a growing body of literature.
Based on survey and choice experimental data collected in Northern Kenya, this article investigated pastoralists' preferences for different contract attributes, which are designed to address some of these barriers. It is argued that stated preferences for specific configurations of the attributes could contribute to the improvement of existing as well as to the development of new livestock index insurance products that are tailored to pastoral preferences and needs and that are, eventually, demanded by those.
To reveal how the surveyed pastoralists react to different contract designs, choice models were applied. Results suggest that pastoralists prefer contracts with the lowest possible price that provide transparent information, that include an index certification and that have the lowest possible strike level. The results also suggest that three groups with heterogeneous preference can be distinguished: one group only considers a low strike level as well as an index certification to be relevant contract features. For a second group, it is only decisive that contracts entail an index certification, whereas a third group prefers contracts with a low strike level, a commitment to transparency and an index certification. In this latter case, transparency seems to be even more important than the index certification. We, therefore, support previous claims which advise to keep contract attributes flexible, or to offer different contract designs to address the specific preference patterns (Ceballos & Robles, 2020).
Eventually, more research is needed, especially in terms of identifying mechanisms and contract features that can address liquidity constraints. Further empirical findings could add to the knowledge base of suitable contractual designs in different contexts and should consider dynamics of purchase decisions as well as the implications of non-stationarity of the index base due to climate change (Daron & Stainforth, 2014;Takahashi et al., 2020). That said, we believe that the implications of this article can provide inspiration and impetus for the configuration and design of livestock index insurance contracts. Given that pastoralists in Northern Kenya are particularly vulnerable to droughts, improving their attraction to such products is of high policy relevance. As the existing livestock index insurance products are planned to be further developed, and new livestock index insurances are in planned to be implemented in different parts of sub-Saharan Africa, the here presented research might come at just the right time to be transformed into action.

Disclosure statement
No potential conflict of interest was reported by the authors.

Funding
This research was financially supported by the German Research Foundation (DFG) through grant number RTG1666 (GlobalFood).

Notes on contributors
M. Linhoff is an economist and works in the field of private sector development. She holds a M.Sc. in International Economics from Göttingen University, Germany and has worked in several development cooperation projects from national and international organizations.
O. Mußhoff is Professor of Farm Management at the Georg-August-University Göttingen, Germany. He has worked on a broad range of research questions in the field of agricultural economics, including modelling of entrepreneurial decisions, investment and finance, risk management as well as experimental impact analysis of agricultural policy measures. He holds a PhD in Agricultural Economics from the Humboldt-University in Berlin, Germany.
M.C. Parlasca is a Senior Researcher at the Center for Development Research (ZEF) at Bonn University, Germany. His work focuses on food system analyses and the implications of technologies for farmers in developing countries. He holds a PhD in economics from Göttingen University, Germany.