Unpacking Factors behind the (Low) Uptake of Index-Based Insurance: Quasi-Experimental Evidence from Livestock Insurance in Southern Ethiopia *

Transcription

1 Unpacking Factors behind the (Low) Uptake of Index-Based Insurance: Quasi-Experimental Evidence from Livestock Insurance in Southern Ethiopia * Kazushi Takahashi a, Munenobu Ikegami b, Megan Sheahan c, Christopher B. Barrett c a Institute of Developing Economies-Japan External Trade Organization (IDE-JETRO) b International Livestock Research Institute (ILRI) c Charles H. Dyson School of Applied Economics and Management, Cornell University Abstract: With increased global climate risks, enhancing resilience among the poor through microinsurance is considered an important tool towards sustainable poverty reduction. Although index-based microinsurance, which should be free from the classical incentive problems, has attracted considerable attention, its uptake rates are generally low among developing countries that have introduced it. To understand factors behind the low uptake, we explore the patterns of purchase and non-purchase of index-based livestock microinsurance in southern Ethiopia. The top two constraints to purchasing this insurance product, as perceived by survey respondents, are the lack of understanding of microinsurance products and insufficient funds at the time of enrollment. We conducted quasi-experimental designs to help improve the understanding of the products and the ability to pay through randomly distributing learning kits and discount coupons to subpopulation. Our detailed econometric analysis revealed that the learning kits contribute to improving the knowledge of the products; however, improved knowledge per se in turn does not induce greater uptake. On the other hand, we found that the reduced price through the distribution of discount coupon has an immediate impact on * We are indebted to Wako Gobu, Mohamed Shibia, Birhanu Taddesse for their excellent field assistance as well as Eddy Chebelyon, Samuel Mburu and Oscar Naibei for their careful data management. This work was made possible by financial support provided by Cornell University, the International Livestock Research Institute (ILRI), the Institute of Developing Economies (IDE)-JETRO, JSPS Grant-in-Aid for Scientific Research (B) , and the US Agency for International Development (USAID) Agreement No. LAG-- A through Broadening Access and Strengthening Input Market Systems Collaborative Research Support Program (BASIS AMA CRSP). All views, interpretations, recommendations, and conclusions expressed in this paper are those of the authors and not necessarily those of the supporting or cooperating institutions. 1

2 the increased uptake. Although such one-time price reduction can potentially adversely affect the subsequent demand through reference-dependence, we found virtually no price anchor effects. Keyword: Index-Based Livestock Insurance, Quasi-Experiment, Uptake, Ethiopia JEL Classification: D12, G22, O12 2

3 1. Introduction Climate risks, such as drought, flood, typhoon, and increased variability in weather conditions between seasons, have threatened many rural households livelihoods in developing countries. Access to formal insurance services may be one way to help protect welfare in the event of climate shocks by compensating income and asset losses. A formal insurance market, however, has been either absent or underdeveloped in most developing countries due to classic incentive problems caused by asymmetric information, such as moral hazard and adverse selection, as well as associated high transaction costs. Available self-insurance options to smooth consumption, including ex ante risk mitigation toward more stable but lower return activities and ex post risk coping by selling productive assets in the face of downside shocks, are often costly, jeopardizing the long-term household welfare (Morduch, 1995; Rosenzweig and Wolpin, 1993). Given limited and inadequate self-insurance options, rural vulnerable households have developed mutual assistance mechanisms within local communities, which can partly, albeit not fully, help recover from losses when the nature of the shock is idiosyncratic (Townsend, 1994: Udry, 1994; Dercon and Krishnan, 2000). It is, however, well known that such informal risk sharing mechanisms do not function effectively under covariate, catastrophic natural disasters, where all neighboring community members suffer substantial losses (Barrett, 2011). Protecting the rural poor from such covariate climate risks has been, thus, a major challenge to achieving sustainable rural poverty reduction. Provision of microinsurance, small-scale insurance products aimed at low income people who are generally excluded from more traditional insurance products, has attracted widespread interest as a means to enhancing the resilience of the rural poor 1

4 against covariate climate risks (Churchill, 2006; Mechler et al. 2006; de Bock and Gelade, 2012). In particular, recently introduced index-based weather insurance has attracted considerable attention as it is free from information asymmetry problems. An innovative feature of the index insurance is that indemnity payouts are determined based not on actual losses experienced by policy holders, but on easily observable, objective weather parameters that are highly correlated with expected losses, such as rainfall, temperature, and satellite-measured vegetation level. This feature allows insurers to avoid the significant transaction costs associated with monitoring the behavior and verifying the losses of the insured. While the existence of basis risk, i.e., the discrepancy between realized loss and indemnity payouts predicted by the index, may remain a potential threat to policy holders, a substantial portion of otherwise uninsured climate risks can be mitigated through purchasing the index insurance. Thus, index-based weather insurance is expected to meet a major unmet demand for mitigating covariate weather risks in rural developing contexts. Despite sweeping claims that index-based microinsurance would be the next revolution in development practice (Morduch 2006), evidence to date shows that unexpectedly low uptake, rarely above 30%, causing many to rethink the attractiveness of the product or suggest ways to improve it (De Bock and Gelade, 2012; Miranda and Ferrin 2012; Matul et al., 2013). For example, Binswanger-Mkhize (2012) provides an argument for why index-based insurance will not proliferate. Through a review of the literature, he finds that higher income farmers are already self-insuring against risk by diversifying their income portfolio. Lower income farmers and landless laborers who are unable to diversify optimally would, therefore, be more likely to gain from index-based insurance; however, the cost of doing so generally prohibits uptake. To the effectiveness 2

5 of the product more generally, Breustedt et al. (2008) found that index-based insurance schemes did not provide statistically significant risk reduction, which lowered the participation into the insurance program among wheat farmers in Kazakhstan. Leblois et al. (2014) use the experience in India to develop a model looking at the ex ante expected utility benefits for risk averse millet farmers of purchasing index insurance and find minimal gains from insurance uptake, especially relative to the costs. Apart from the effectiveness of the product, price and liquidity constraints are often identified as the important factors suppressing demand (Gine et al., 2008; Cole et al., 2013; Karlan et al., 2014). McIntosh et al. (2013) found high uptake among households randomly allocated vouchers that reduced the cost of the product. Also, knowledge of the somewhat esoteric product, especially where no other insurance products are sold, will likely hold potential users back (Skees 2008) and, conversely, that knowledge of the product increases uptake (Gaurav et al., 2011; Cai, et al., 2013; Cai and Song, 2013). Most of the existing studies are rooted in the experience of crop insurance programs from Asia to Latin America to Africa and do not focus on the small number of index-based asset insurance, such as livestock insurance, programs that have emerged in the recent past (McPeak et al., 2010; Chantarat et al., 2012; Janzen and Carter, 2013). To the extent that the livelihood systems, risk mitigation strategies, and the long term welfare outcomes associated with shocks differ between crop-based and pastoral-based production systems, we would expect the demand for and benefits of index-based insurance to similarly diverge. For example, where households derive their income primarily from crops, durable household assets will likely be the household s main store of wealth. While those assets may be drawn down in times of low income, the opportunity to recover income levels will not necessarily be affected by a temporary shock to 3

6 household assets since household assets are decoupled from crop production. On the other hand, where households derive their income primarily from livestock, the income generating engine -- the livestock -- also functions as the household s store of wealth, particularly where household are at least partially transitory. Weather shocks, therefore, may be particularly exacerbated for pastoralist households since direct hits to income earning potential in one year will significantly affect income earning potential in all subsequent years (Chantarat et al., 2012). If households are aware of the fact that the probability of loss disproportionately affects long term income trajectories, we would expect demand for index-based livestock insurance (IBLI) to be greater than index-based crop insurance. Analysis on the levels and determinants of demand for IBLI is just now emerging from products piloted in northern Kenya (Jensen et al., 2014) and Mongolia (Mahul and Skees, 2007). This paper adds to this growing literature using the experience of a new IBLI in pastoral southern Ethiopia. We use two waves of panel data, the first as a baseline and the second as a follow-up where households would have had two opportunities, based on the seasonal nature of the product, to purchase an IBLI policy. The IBLI product under investigation insures pastoralists against livestock mortality that often follows from catastrophic drought. Because severe forage scarcity is common during drought, IBLI is paired with satellite data that tracks local forage conditions which are then used to predict area-average livestock losses and to calculate indemnity payments. Over the course of the IBLI sales period, we introduced two kinds of randomized encouragement designs aimed at improving pastoralists understandings of IBLI (via learning kits, such as comics and skit tapes) and their ability to pay (via discount coupons). Not only were these features included to create incentives for IBLI 4

7 uptake, but also to aid in demand analysis by adding important sources of exogenous variation. Our estimation results indicate that the reduced price of the insurance through the provision of discount coupons significantly increases the uptake of IBLI. Although there is a potential thread that one-shot price reduction serves as price references, which decreases subsequent demand (Dupas, 2014; Fischer et al. 2014), the present study does not find such anchor effects. On the other hand, while the learning kits do positively contribute better knowledge of the products, better knowledge does not appear to induce extended uptake of IBLI substantially. The remainder of this paper is structured as follows. Section 2 explains the study site, sampling framework, and detailed designs of the IBLI product and quasi-experiments. Section 3 discusses descriptive statistics. Section 4 explains estimation strategy, followed by discussion on estimation results in Section 5. Section 6 concludes the paper. 2. Data 2.1. Study area Our study areas are located on the Borana plateau in Oromia regional state 1 in southern Ethiopia (Figure 1). The majority of the population is pastoralist, whose livelihoods depend primarily on livestock. The region is comprised of arid and semi-arid ecological zones with four seasons: long rainy season from March to May; long dry season from June to September; short rainy season from October to November; and short dry season from January to February. Herd migration in search of forage and 1 The largest Ethiopian administrative unit is the regional stage, which is subdivided into zones, then into woredas, and further into kebeles. 5

8 water during dry seasons is common among pastoralists in this area. The sustainability of pastoralism has been, however, significantly undermined due to recurrent drought, violent conflicts, and other political and economic instability (Desta et al., 2008; Tache, 2008). Among these, drought is by far the greatest cause of livestock mortality in the study areas (Lybbert et al., 2004). Major droughts occurred in almost every 6-7 years in 1973/74, 1983/84, 1991/92, 1999/00, 2005/06, and 2011/12, each causing massive numbers of livestock deaths (Desta et al., 2008; Megersa et al., 2013). There exists a range of customary insurance arrangements, like Debare and Busa Gonofa, that provide informal inter-household transfers in the form of cash or livestock. Yet, many times the entire livelihoods of the community are threatened during drought, rendering traditional risk sharing arrangements weak and insufficient. Indeed, those informal arrangements tend to cover less than ten percent of household loss, usually exclude the persistently poor who need it the most and are generally perceived to be in decline (Lybbert et al., 2004; Huysentruyt et al., 2009; Santos and Barrett, 2011;). In these settings, the returns to a program that insures the pastoral population against drought-induced livestock losses should, therefore, be substantial. 2.2.Design of IBLI To help pastoralists manage the considerable drought-related mortality risk, IBLI was introduced by the International Livestock Research Institute (ILRI) and Cornell University in collaboration with the Oromia Insurance Company (OIC) in The basic product design is similar to that of a previously designed IBLI product in northern Kenya, first rolled out in As in northern Kenya, the standardized Normalized Differenced Vegetation Index (NDVI), a numerical indicator of the degree 6

9 of greenness recorded by satellite, accumulated over one rainy season and the following dry season was used to construct an index (Chantarat et al., 2012; Mude et al., 2012). This index was calibrated for high correlation with average livestock mortality from drought at the woreda level and is linked to a lower-bound threshold that triggers indemnity payments for policy holders if the index falls below the 15 th percentile of historical distribution since IBLI is marketed and sold during two periods per year occurring directly before each rainy season (August-September and January-February) with coverage lasting one year and the potential for two indemnity payouts, one after each dry season (Figure 2). During each sales period, a household decides whether to buy IBLI and if so, how many animals to insure. A premium payment is set to be the value of total insured herd value (TIHV), which varies with species of animals, 2 multiplied by the woreda specific insurance premium rates, which are close to actually fair premium rates and range from 7.5% to 11.0%, across woredas given differences in expected mortality risk. 3 Formally, TTTT = (# of camel insured) 15,000 + (# of cows insured) 5,000 + (# of goats and sheep insured) 700 and PPPPPPP ppppppp = Woreda specific insurance premium rates TTTT. If a household buys IBLI in the August-September sale period, it is insured 2 The values of insured livestock are predetermined as follows: 1 Camel = 15,000 birr, 1 Cow = 5,000 birr, 1 Goat = 700 birr, and 1 Sheep = 700 birr. These values are constant across sales periods. 3 More specifically, woreda specific premium rates are set to be as follows: 9.75% for Dilo, 8.71% for Teltele, 7.54% for Yabello, 9.49% for Dire, 8.58% for Arero, 9.36% for Dhas, and 11.05% for Miyo and Moyale. 7

10 from October to September in the following year and may receive indemnity payout in March and/or October. Note that if a pastoral household buys IBLI not only in August-September sale period but also the following January-February sale period, insurance coverage periods for the two contracts are overlapped from March to September, and it may receive indemnity payouts for the both contracts in October. This insurance design is expected to reduce the cash constraints faced by pastoralists to the extent that it allows them to pay less on more frequent intervals to insure the same values of livestock for the specified period. The feature of two potential payouts in a year and the 15 th percentile trigger level mentioned above make an expected probability of payout occurring once every three and a half years. At present, two sales periods have occurred: the first in August-September 2012 and the second in January-February No indemnity payouts have been made to insured households at this point. 2.3.Sampling framework While IBLI was marketed and sold to any household within the Borana plateau, we study a random sample of households in the region to explore the pattern of IBLI uptake among pastoralists. The baseline survey data were collected in March 2012 before the first IBLI sales period (August to September 2012) with a follow-up survey implemented in April 2013 directly after the second IBLI sales period (January to February 2013). Sampling for the household survey is clustered at the Reera level, the smallest administrative unit after kebele. Sample Reeras (hereafter, we call them as study sites) were selected so as to maximize agro-ecological and livelihood variation across the Boran pastoral area. Reeras inaccessible by vehicle were, however, excluded 8

11 for logistical and cost reasons. As shown in Figure 1, we have selected 17 study sites for this study, and for the selected study sites, a population and livestock holding census survey was completed by development agents (DAs) who worked in the survey areas as local development officers. Households in the census were then split into terciles based on the number of livestock held to represent wealth classes. Then, 15 percent of households per study site were basically selected for the sample, one third of which were from each of the livestock holding terciles, totaling 528 households across 17 study sites. Due to logistical challenges in March-May rainy season, however, baseline data were collected from only 515 of the selected sample in March Out of those, 475 households were re-surveyed in April 2013 and are extensively used in this study. 2.4.Encouragement design To stimulate uptake of IBLI and construct a treatment-and-control quasi-experimental design, three different encouragement tactics were offered to randomly selected subpopulations during each of the two sales periods. The first component of the encouragement design was created to increase overall awareness of IBLI and to improve knowledge of how the product worked and its benefits. This was done through the use of two tools a comic and a skit which were distributed randomly to households within randomly selected study sites in each sales period. Study sites were categorized into three groups, i.e., those located closer to major livestock markets, those with less rainfall, and those located far from functioning livestock markets. Within each of these three clusters, the sites were randomly assigned comic and skit tape treatments while keeping at least one site as a control. The second component of the encouragement design was the distribution of 9

12 discount coupons which lowered the cost of purchasing IBLI and were expected to improve the chosen households ability to pay. With a coupon, the recipient could purchase IBLI at a discounted rate for the first 15 Tropical Livestock Units (TLUs) 4 insured. In each study site and each sales period, households offered discount coupons were randomly chosen to receive coupons ranging from 10% to 80% in order to manufacture exogenous variation in the effective price faced by prospective IBLI purchasers. 20% of the sample households did not receive any coupon during each sales period. 5 To implement these experiments, DAs were trained to explain and distribute the coupons to the study households either in collective meetings or, less often, in separate personal visits. For disseminating the comic, DAs would read and give the paper version of the comic to treatment recipients, again, either in community meetings or individually. Similarly, the DAs convened group meetings or met households at their home to play the skit via tape. Unfortunately, ILRI staff found that some DAs did not implement these random assignments rigorously in the first sales period, especially for the cartoon and skit tape. Consultants were hired to implement these activities in the second sales period instead of DAs to improve the quality of implementation. 3. Summary statistics Table 1 presents selected descriptive statistics derived from the baseline data 4 1 TLU is equivalent to 1 cow, 0.7 camel, 10 goat, or 10 sheep. 5 Discount coupons were printed in 10% intervals between 10% and 80% with roughly one tenth of the sampled households falling into each interval. In parallel with the household survey for this study, a separate but overlapping herd migration survey was conducted which included 30 households from our larger sample. 15 of those households received 100% discount rates. 10

13 collected in 2012, overall the full sample then separately for those households that have and have not purchased IBLI. We find no statistically significantly difference between these two sets of households across any of the included variables, except for the percentage of income from livestock. As such, the summary statistics described below refer to the overall sample. The average household size is 6.3, with a male-female ratio close to one. The average age of household heads, which are predominantly male, is approximately % of household heads have never attended formal school, and therefore the average amount of completed education is only half a year which is quite low even relative to an average of 4.7 years across all Ethiopian households (McIntosh et al., 2013). However, schooling levels of younger household members have been improving and about half of primary-school aged children (i.e., 7-15 years old) have attended at least some formal schooling. The average monthly consumption per capita is 317 birr, 6 and the poverty incidence, using the $1.25 (purchasing power parity: PPP) per day poverty line, is about 48%. As noted previously, the predominant source of income is livestock, including milk and meat production, which accounts for approximately 60% of total household income. Other income sources, such as crop production and other off-farm activities, play a relatively minor role with limited endowment of land and human capital: only 14.6% of households derive income from crop with the unconditional mean of the share of crop income to the total household income to be 5.7% and cultivated land size to be 2.1 acres. Livestock comprises the overwhelming majority of households non-human asset. The average TLU of animal owned by sample households are 14.7, dominated by 6 1 USD is equivalent to 19.5 Ethiopian birr as of

14 large cattle herds supplemented with goats, sheep, and camel. Given the importance of livestock and limited outside options for protecting against drought risks, it is puzzling to observe the overall low uptake rates of IBLI. Table 2 shows the percent of sample households that purchased IBLI in each sales period as well as the average animal insured, in terms of TLU and TIHV, separately for those who purchased IBLI at both periods, only the first period, only the second period, and never purchased. About 27% of sampled households purchased IBLI during the first period, but declined to 20% in the second period. Only 23 out of 475 households bought IBLI in both sales periods, indicating that purchasing over two consecutive and overlapping periods was uncommon. The number of insured TLU is also considerably small; the unconditional mean is only 0.66 at the first period and 0.47 in the second period, which represents less than 5% of all animal owned. The average TIHV is 3.5 thousand birr in the first sales period and 2.5 thousand birr at the second sales period, respectively, which are close to or only slightly greater than the monthly household income. Those who purchased IBLI during both sales periods tend to insure more animal than those who purchased it at either the first or second period only. Table 3 displays the main reason why survey respondents claim to have not purchased IBLI, reported by respondents in the 2013 survey round. The top two reasons are the lack of cash followed by the lack of knowledge about IBLI, mimicking the major constraints commonly raised across other index-based insurance pilots in the developing world (Gine et al., 2008; Gaurav et al. 2011; Cai et al., 2013; Cole et al., 2013; Platteau and Ontiveros, 2013; Karlan et al., 2014) despite the fact that the previously described random encouragement design employed in this study aimed to mitigate such constraints. Table 4 shows the sources of information when respondents had heard about 12

15 IBLI, also collected in the 2013 survey round. DAs and ILRI staff played major roles as information channels, with 86% and 67% of respondents claiming to learn about IBLI from each source respectively. Meanwhile, a non-negligible number of households claimed to obtain information about IBLI through encouragement designs, although some treated respondents did not recognize these tactics as an IBLI information source. To obtain deeper insights into the effect of learning kits, in the 2013 survey, we implemented a quiz to respondents in which eight questions about IBLI were asked, ranging from questions about the insurer, conditions, frequency, and amount of indemnity payout, to simple computations of premiums and payouts under hypothetical scenarios. The number of correct answers tends to be larger when respondents receive the learning kits, either a comic or a skit, and the difference is statistically significant especially when respondents receive them at the second sales period (Table 5). To show the effectiveness of the discount coupon distribution, Figure 3 displays the relationship between TIHV and a household-specific premium rate, where a household-specific premium rate is defined as 7 : HHHHHhooo ssssssss prrrrrr rrrr = (1 dddddddd rrrr) WWWWWW ssssssss ppppppp rrrr. As expected, the IBLI uptake decreases with the household-specific premium rates, suggesting that the demand of IBLI is price elastic and discount coupons may potentially induce uptake. Overall, the encouragement design seems to have contributed to spreading information about the existence of IBLI as well as to inducing uptake in the 7 For households that did not receive a discount coupon, their premium rate is equivalent to the woreda-specific premium rate. Premium rates at the woreda level are the same with sales period, while the discount rates at the household level varied with the sales period. 13

16 study sites. 4. Estimation Strategy In order to study more rigorously the effectiveness of these encouragement designs and a more general set of constraints to IBLI uptake, we turn to regression analysis. We are interested not only in whether or not households choose to buy IBLI in a given sales period, but also how many animals they choose to insure, measured by TIHV, conditional on purchasing an IBLI policy. Since more than half of households do not buy IBLI at all, parameters estimated via Ordinary Least Squares (OLS) would be inconsistent. One standard approach to consistently estimating a model with a continuous dependent variable with censored observations is the Tobit model. The Tobit, however, imposes a rather restrictive assumption that the decision to buy IBLI and decisions about how many TLU to insure are determined by a single process which need not be true, particularly in this context where herd sizes vary dramatically between households. To overcome the restrictive assumptions inherent in the Tobit model, we employ the double-hurdle (DH) model originally proposed by Cragg (1971). The DH model is more flexible than Tobit in that it assumes that the observed demand for IBLI can be decided in a step-wise manner, i.e., first the decision whether or not to buy IBLI in general, followed by the decision on the quantity of animals to insure. 8 The 8 Another approach is the Heckman selection model. As with the DH model, the Heckman selection model takes into account the two-step decision making process. The major difference between the Heckman selection model and DH model is that the former is designed for incidental truncation where zeros are unobserved due to self-selection. In other words, the Heckman model assumes that there will be no zero 14

17 underlying decision-making process of the DH model can be expressed as: d y i i * 1 if di = miα + ξi > 0 = 0 otherwise * * * yi if yi = X iβ + e i > 0 and di > 0 = 0 otherwise (1) (2) where di is a binary indicator variable to describe whether household i bought IBLI, * * y represents TIHV, d and y are the unobserved latent variable, i i i mi and X i are vectors of explanatory variables, and α and β are estimated parameters. Because we observe two separate sales periods, this model can be run separately for each period because coefficients and the underlying decision process may be different given the proliferation of information and changing perceptions about IBLI over time. More specifically, we first estimate equations (1) and (2) for the first sales period then use a slightly modified version of this DH model for the second sales period, considering that the second period choice would be conditional on the first sales period, particularly since the two coverage periods overlap. To be more precise, instead of using the univariate probit model in the first hurdle at the second sales period like equation (1), we apply the recursive bivariate probit model for the second sales period. This modifies the equations (1) and (2) to: d d i1 i2 * 1 if di 1 = mi 1α 1 + ξi 1 > 0 = 0 otherwise * 1 if d = dˆ i2 d i1 + mi2α 2 + ξi > 0 0 otherwise = 2 (3) (4) observations in the second stage, once the first-stage selection is passed, while the DH model allows for the option of deliberate zero observations. Although both models seem relevant in our context, we prefer the DH model because it nests the Heckman model. 15

18 y i2 * * y if y = dˆ * i2 i2 θ 2 i1 + X i2β 2 + e i2 > 0 and d i2 > 0 = 0 otherwise (5) where the subscripts 1 and 2 indicate the first and second sales periods, respectively, and d ˆi1 is the projected first stage purchase experience with its parameter to be δ. We allow the correlation between ξ 1i and ξ 2i, and jointly estimate equations (3) and (4) through the recursive bivariate probit model (Wooldridge, 2010). Then, we will run the separate truncated regression as in equation (5) to examine factors determining TIHV at the second sales period, conditional on purchase. Following Cragg (1971), we assume that the first-hurdle error term ξ it and second-hurdle error temε it (e.g., error terms in equations (1) and (2)) are independently and normally distributed with zero mean at each sales period. While covariance between the those errors can be non-zero, Garcia and Labeaga (1996) and Jones (1992) among others show that estimated results are quite similar regardless of whether the assumption of zero covariance was relaxed. To reduce computational burdens, we maintain the assumption of zero covariance between the first- and second-hurdle error terms. Given independent error terms, the log likelihood function for the DH model can be equivalent to the sum of the log-likelihoods of a probit model and truncated regression model (Cragg, 1971; Burke, 2009). Thus, separate regressions for the first hurdle with the probit (for the first period) and bivariate probit (for the second period), followed by the second hurdle with the truncated regression, yield consistent estimates with the DH model described above. Major explanatory variables of interest in the first- and second sales period DH models include the effective price of IBLI faced by each repondent and the knowledge of IBLI. The former is captured by the inclusion of household and sales period specific 16

19 premium rate as well as the dummy to represent whether households received a discount coupon. 9 In the second sales period estimation, we also include the discount premium rate and the dummy for coupon receipient at the previous sales period as additional regressors. This is motivated by recent studies that argue one-off subsidies on a product can reduce future demand as people anchor around the reduced price and become unwilling to pay more for the product later (Dupas, 2014; Fischer et al. 2014). We test whether one-shot price subsidies can immediate impacts on uptake by checking the coefficients of the current period discount rate, while test whether they have longer-term effects by checking the coefficients of the previous period discount rate. Knowledge of IBLI is proxied using the number of correct answers to a quiz about IBLI administered during data collection. The data are derived only from the second wave of the survey because we did not ask the knowledge of IBLI at the time of baseline survey. An obvious concern is that households with greater interest in IBLI know more about the product and are more likely to buy, or the knowledge of IBLI increases after a household bought IBLI, causing an endogeneity problem. To address this, we apply a two-step estimation strategy, where we regress the number of correct answers to the quiz at the first step and then employ the DH models using the predicted number of correct answers as one of the additional repressors at the second step. The treatment dummies of skit tapes and comics, 10 which are purely exogenous by construction, are used as instruments for the knowledge of IBLI. 9 We use the administrative record, instead of respondents report, to measure these data. Since some DAs do no strictly follow the random assignment rule, our estimate is intent to treat with some compliance. 10 As in the previous footnote, we again use the administrative record, instead of respondents report, to measure these data. 17

20 Other controls are all constructed from the baseline survey to avoid potential endogeneity and to provide an ex ante picture of the household before IBLI was introduced (see Appendix 1 for a full list of explanatory variables). These include: (1) risk tolerance dummies elicited through field experiments following Binswanger (1980) 11 ; (2) monthly household income, (3) household livestock holdings, measured in TLU, and a squared term; (4) the value of non-livestock assets, represented by a wealth index computed using the principal component analysis; (5) the size of cultivated land, (6) the household s subjective expected livestock mortality within a year from the baseline survey; (7) dummy variables that capture whether households expect livestock prices to increase or to remain the same within a year from the baseline survey; (8) characteristics of the household and household head, such as household size and age, years of completed education, and gender of household head; and (9) Woreda dummy variables, which function as controls for the woreda-level important unobservables. 12 Clustered standard errors at the study site level are employed for all regressions to derive statistical inference. 11 We do not use an estimated cardinal value of risk preference, such as the midpoints of the imputed constant relative risk aversion (CRRA) intervals, because such a CRRA coefficient imposes strong assumptions on the shape of preferences and may not precisely reflect Arrow-Pratt risk preferences if there exist any threshold effects in underlying wealth dynamics (Lybbert and Barrett, 2011). 12 An important factor that was identified as key determinants of uptake in the literature on index-based insurance, but lacking in this model would be basis risk each household faces. We cannot include a proxy for it, such as distance from the rainfall station (Mobarak and Rosenzweig, 2013) because pastoralists are so mobile that there is no single point that properly reflects the exposure to basis risk at the household level. An alternative way to proxy basis risk may be to use the deviation between the local average livestock mortality rate and the individual-specific mortality rate. Yet such a variable is highly likely to be endogenous and we have no credible instruments. 18

21 5. Estimation Results Table 6 presents the estimated results of the first step regression for factors associated with the knowledge of IBLI. The first sales-period coupon recipients tend to have better knowledge about IBLI. This is presumably because when a discount coupon was distributed, DAs and ILRI staff explain in details what is IBLI and how this coupon can be used, contributing to improving respondents knowledge of IBLI. Besides, the second sales period random encouragement assignments are positively correlated with the number of correct answers, while the first sales period assignments are not, which might be because some DAs did not implement these random assignments rigorously, as explained earlier. Among the other important determinants of IBLI knowledge include education of the household head, which has the positive impacts. The main estimation results on the DH model, incorporating predicted values of the number of correct answers to quiz from the first step regression above, are presented in Table 7, where the dependent variable of probit regressions (i.e., columns (1), (2) and (4)) takes one if the household buys IBLI at the respective period, while the dependent variable of the truncated regressions (i.e., columns (3) and (5)) represents TIHV measured in thousand birr. 13 In each table, column numbers correspond to the equations numbers described above. As apparent, the coefficients of the number of correct answers to quiz are statistically insignificant. This implies that, while our skit tapes and comic improves the knowledge of IBLI, as we saw in the previous table, the impact of knowledge via these 13 We conducted a robustness check by replacing the dependent variable by the number of TLU insured. The estimated results, available upon request, are qualitatively very similar to Table 6. 19

22 encouragement devices on IBLI demand may not be substantial. Household demand of IBLI is clearly sensitive to the price the household faces. In each sales period, the household-round specific premium rates consistently and negatively affect the decision to purchase as well as the value of animals insured, except for the purchase decision at the second sales period (column (3)). Marginal impacts 14 of the premium rates are 1.39 and 1.12 at the first and second sales periods, respectively, implying that that a decrease in the premium rate by one percent is associated with increases in the value of animal insured by 1.39 and 1.12 thousand birr at the first and second sales periods, respectively. It is also important to observe that in the second sales period, the price of the previous period has no effect on the current decision to purchase, indicating that there is no anchoring effect, as suggested by other literature (Dupas, 2014; Fischer et al. 2014). Controlling for the price, its demand is more or less affected by whether households receive a coupon or not, although the direction of impacts is mixed; in the first period, coupon recipients are more likely to purchase IBLI, while in the second period, they tend to insure less value of animals. On the one hand, coupon recipients may be more encouraged by DAs and ILRI staff to buy IBLI when they receive the coupon. On the other hand, however, self-motivated people, who are willing to buy IBLI even without the discount coupon, tend to buy more IBLI than those motivated by the research team. These compelling effects may cause an opposing effect at the different sales period. Note that the discounted premium rates may have income effects aside from the pure price effect. Although we expect such income effects are small in our context as 14 These marginal impacts reflect partial effects of the discount premium rate on the value of insured animal conditional on IBLI purchase. 20

23 the reduction in the premium rate is minor relative to the total household income, we cannot rule out that possibility. However, if there is any positive income effect through the random encouragement designs, we may see that the coefficient on income also positively affects the uptake. Our results show only one out of five cases that the income has systematically positive impacts on the decision to purchase (column (5)) and its statistical significance is weak even at the 10% level. This may partly support our view that the price effect will be a primary reason for the significant impact of the discounted premium rate on uptake, rather than possible income effect. As shown in Table 3, insufficient livestock holdings is one of the important reasons households report for not purchasing IBLI. However, through regression analysis, we find that total household TLU does not affect the decision to purchase IBLI or not, whereas the estimated results from truncated regressions in column (2) and (5) show that livestock holding is statistically significantly associated with the value of insured animal conditional on having an IBLI policy. 15 Interestingly, their relationship is not monotonic, but inversed-u shaped. That is, the value of insured animals increases with the initial number of TLU holding up to a certain threshold, but decreases with it afterwards We also estimate the entire models by replaing TLU and its squared term by the proportion of income from livestock to see whether those households with diversified livelihood portfolios are more likely to buy IBLI. The results, available upon request, show that the proportion of income from livestock is positively correlated with the probability of buying IBLI and the value of insured animal conditional on IBLI purchase, reflecting that a household with more diversified income portfolio is less likely to rely on IBLI as a risk-coping mechanism. 16 The fact that there is a peak demand of IBLI at the middle of livestock class seems to be consistent with the view of multiple herd size equilibria, where the demand of IBLI is highest among pastoralists with herd size slightly greater than the threshold in order 21

24 Cultivated land size and wealth index show largely expected signs, indicating that wealthier people are more likely to insure more animals. The results are robust for both the first and second sales period. Households who expect livestock price to rise are more likely to buy IBLI and tend to insure greater value of animal. On the other hand, risk preference dummies are largely statistically insignificant. Although theory suggests risk-averse households buy insurance more, existing studies on index-based insurance products show mixed results (Gine et al., 2008). We expect this is because households that purchase IBLI now face two risks: the risk of drought and livestock death and the risk that IBLI will provide a payout commensurate with the losses they experience. These two competing risks may cancel each other out in the demand estimation. Other socio-economic characteristics of the household and household head also appear to affect IBLI demand. First, it seems that household size is negatively associated with the value of animal insured at the second sales period. On the other hand, male-headed households are more likely to buy larger insurance policies at the first sales period. Although we a priori expected the education of the household to be positively associated with IBLI demand, it turns out to be opposite. We can also infer something about the dynamic pattern of IBLI uptake using the coefficient on the first round IBLI purchase experience in the second round purchase model and the correlation of the error terms (rho) between equation (3) and (4). 17 Interestingly, the IBLI purchase experience from the first period has positive impacts on purchase at the second period in column (4), indicating that those who initially bought to avoid being trapped into poverty traps (Lybbert et al., 2004; Carter and Barrett, 2006; Chantarat et al., 2012). 17 The percentage of correctly predicted is 75.1% given the break point at

25 IBLI are more likely to continue to buy IBLI in the second period, all else equal. This is presumably due to learning from doing in which the experienced insured recognize the benefit of IBLI more, leading to continuous purchase. Alternatively, this may reflect that those who were willing to purchase IBLI split the purchasing timing into two times such that each premium payment is reduced to mitigate the impact of cash constraints, as was intended in the design. Second, the correlation coefficient of the first and second period purchase is negative, implying that those who were expected to buy it in the first period actually buy it with delay, although its effects are statistically insignificant. This may presumably because pastoralists recognize the benefit of IBLI by observing the neighbors behavior through learning from others. In addition, as we saw previously, the discount price of the previous sales period does not have any lasting impacts by serving as price anchors. Thus, to make pastoralists familiar with IBLI, it may be effective to motivate them by reducing the price once, and once they experienced IBLI, they will continuously demand it even without the price subsidy in the subsequent period. These results together provide suggestive evidence that adopting innovative products, like IBLI, is a dynamic process involving information and expectation updating that affect the speed and path of their diffusion. 6. Conclusion Index-based insurance is recognized as a promising means of protecting the poor from the losses associated with climate shocks. Attempts have been made worldwide in the past decade to introduce innovative index-based weather insurance products. They have, however, commonly suffered low uptake rates. We presume that one of the reasons of such low uptake is because the existing insurance programs cover 23

26 only income loss. Using the index-based insurance for livestock implemented in southern Ethiopia as a case, which covers loss of asset as well as a major livelihood source, this paper attempts to explore factors underlying the demand of asset insurance that cover longer-term welfare. We particularly focus on the role of knowledge about the products and price in decisions to uptake. To cleanly identify causal relationships, we randomly assigned learning kits and discount coupons to the subpopulation under study. We found that uptake rates of IBLI are not as high as expected, well below 30%, which was comparable to other existing index-based insurance. The estimation results show that financial education through the provision of skit tapes and comic improves the knowledge of the product, but the deeper understandings of IBLI seem not to significantly induce expanded uptake. Although the lack of understanding is often identified as a key constraint on the adoption of the index-based insurance in the literature, and our survey respondents also reported it so, our empirical evidence do not support its argument. More importantly, providing incentives, such means as discount coupons, effectively and substantially increase the uptake rates without lowering the future demand by serving as price reference. We also find supporting evidence that once pastoralists get familiar with the products, they tend to continue to purchasing IBLI presumably through learning by doing. It is, therefore, advisable that the provision of these incentives or other interventions to encourage uptake be implemented in the early phase of diffusion to fully exploit such learning effects, which may in turn induce rapid diffusion and as a result substantially enhance the resilience of the climate vulnerable area. 24

27 Reference Barrett, Christopher B. (2011). Covariate Catastrophic Risk Management in the Developing World: Discussion, American Journal of Agricultural Economics, 93(2), pp Binswanger, Hans P. (1980). Attitudes toward Risk: Experimental Measurement in Rural India. American Journal of Agricultural Economics, 62(3), pp Binswanger-Mkhize, Hans. P. (2012). Is There Too Much Hype about Index based Agricultural Insurance? Journal of Development Studies, 48(2). Pp Burke, William.J. (2009). Fitting and Interpreting Cragg's Tobit Alternative Using Stata, Stata Journal, 9(4), pp Breustedt, Gunnar., Bokusheva, Raushan., and Heidelbac, Olaf. (2008). Evaluating the Potential of Index Insurance Schemes to Reduce Crop Yield Risk in an Arid Region, Journal of Agricultural Economics, 59(2), pp Cai, Jing., de Janvry, Alain., and Sadoulet, Elisabeth. (2013). Social Networks and the Decision to Insure, mimeo. Cai Jing., and Song, Changcheng. (2013). Do Hypothetical Experiences Affect Real Financial Decisions? Evidence from Insurance Take-up, mimeo. Carter, Michael.R., and Barrett, Christopher B. (2006). The Economics of Poverty Traps and Persistent Poverty: An Asset-Based Approach, Journal of Development Studies 42(2), pp Chantarat, Sommarat., Mude, Andrew. G., Barrett, Christopher. B., and Carter, Michael. R. (2012). Designing Index-Based Livestock Insurance for Managing Asset Risk in Northern Kenya, Journal of Risk and Insurance, 80(1), pp Churchill, Craig. (2006). Protecting the Poor: A Micro-Insurance Compendium. Geneva: International Labour Office. Cole, Shawn., Giné, Xavier., Tobacman, Jeremy., Topalova, Petia., Townsend, Robert., and Vickery, James. (2013). Barriers to Household Risk Management: Evidence from India, American Economic Journal: Applied Economics, 5(1), pp Cragg, John G. (1971). Some Statistical Models for Limited Dependent Variables with Application to the Demand for Durable Goods, Econometrica, 39(5), pp De Bock, Ombeline., and Gelade, Wouter. (2011). The Demand for Micro-Insurance: A Literature Review," ILO research paper 26. Dercon, Stefan., and Krishnan, Pramila. (2000). In Sickness and in Health: Risk Sharing within Households in Rural Ethiopia. Journal of Political Economy, 108(4): pp , 25