BPPG response - Options for Defined Benefit schemes - 19Apr24.pdf
Group Based Index Insurance
1. Rainfall Insurance and Informal Groups
Evidence from a Field Experiment in Ethiopia
Stefan Dercon1 , Ruth Vargas Hill2 , Ingo Outes-Leon1 ,
Alebel Bayrou3 , Daniel Clarke1 , Alemayehu Seyoum Taffesse2
1 University of Oxford
2 International Food Policy Research Institute
3 Ethiopian Development Research Institute
Clermont-Ferrand, June 2011
2. Basis risk and risk-sharing groups
• Risk-sharing within groups is commonly practiced in rural Africa.
• Groups find it hard to manage risks that affect all group members
simultaneously, such as catastrophic weather events.
• Can index insurance be used as a tool to transfer large covariate shocks
(extreme shortfalls in rain) away from groups, whilst encouraging group
members to share smaller agricultural risks among themselves?
3. Research program
• Long-run research project in Ethiopia to look at this question. In
particular, focusing on funeral insurance mutuals.
• First year (2009): experimental games played with individuals and with
groups–observed how information and decisions were made on building
block index insurance contracts by groups.
• Second year (2010): sales of index insurance to groups in a small number of
villages–all contracts, policies and marketing was the same, but training was
randomized to test what happens when sharing basis risk is suggested.
• Third year (2011): (i) sales of index insurance to groups or individuals in
randomly selected villages, (ii) randomization of commitment to ex-ante
rules in group contract villages.
• Next year (we hope): pilot a number of ex-ante rules perhaps in combination
with savings or credit product to mitigate group level basis risk.
4. Research program
• Long-run research project in Ethiopia to look at this question. In
particular, focusing on funeral insurance mutuals.
• First year (2009): experimental games played with individuals and with
groups–observed how information and decisions were made on building
block index insurance contracts by groups.
• Second year (2010): sales of index insurance to groups in a small number of
villages–all contracts, policies and marketing was the same, but training was
randomized to test what happens when sharing basis risk is suggested.
• Third year (2011): (i) sales of index insurance to groups or individuals in
randomly selected villages, (ii) randomization of commitment to ex-ante
rules in group contract villages.
• Next year (we hope): pilot a number of ex-ante rules perhaps in combination
with savings or credit product to mitigate group level basis risk.
5. Groups and demand
Groups can increase demand for insurance for a number of reasons:
• Share basis risk
• Groups might make better decisions:
• Group leaders are more financially educated
• Might be best placed for understanding insurance products and explaining
them to member farmers
• Reduce transaction costs in purchasing insurance and making claims.
• When groups are used as intermediaries they can increase levels of
trust in insurance products.
All marketing and retailing in this study was through groups, so our focus is
on the group potential for mitigating basis risk.
6. This study
• Rainfall deficit index policies are marketed through pre-existing
risk-sharing groups: funeral societies called iddirs.
• All products were an individual contract retailed through the iddir. Iddir
leaders managed the payment of premiums and signed contracts on
behalf of individual members that had signed up.
• All marketing and training about these products was through the iddir.
Iddir leaders were trained on the product, and were asked to select
additional members of the iddir to also receive training.
• We randomize the content of the training provided to iddir leaders:
• Training A: Focused on the individual benefits of insurance, and illustrated
how to choose the right policies for an individual farmer.
• Training B: Focused on the group benefits of insurance, and illustrated how
to choose the right policies for a group, and how groups could enable risk
sharing.
7. Why might training increase demand
Let the basis risk associated with an index product for farmer i in community j
be denoted by bij . The idiosyncratic component of the basis risk that is
unique to that farmer is εij , and the component farmer i shares with others in
community j is µj .
Why would the training increase demand?
• If mutual insurance groups are not already sharing all εij encouraging a
mutual insurance network to internalize the idiosyncratic part of basis
risk (essentially “crowding in" more informal insurance) will reduce the
basis risk from bij to µj .
• The relative magnitude of εij to bij will determine how large an effect
such mutualization will have on demand. At the extreme, if εij is
negligible, there will be no impact from offering insurance to groups
instead of individuals.
8. Other explanations?
• If the training sessions which suggested sharing εij resulted in better
understanding of the product, increased demand may result.
• Training sessions which suggested sharing may have made the collective
nature of drought more salient and thereby increasing demand for formal
insurance products that were reliable in times of such covariate shocks.
• Suggestions to share policies may have encouraged individuals in the
group to divide policies amongst themselves, and the increased
divisibility that resulted could increase demand.
9. Context
• Highly heterogenenous farming practices and outcomes–some evidence
that this is partly due to risk-coping–but suggests the ratio of εij to bij
could be quite high.
• May also mean that the most appropriate index for group j = 1 will be
quite different from the most appropriate index for group j = 2. Flexibility
in the index contract taken by the group may be desirable. It may also be
the case that µj is non-negligible.
• Deficit rainfall is by far the highest source of risk in the areas we are
working, and the most covariate. Too much rain is also a problem:
• 2009 survey 44% of farmers reported serious losses in wealth and
consumption due to drought in last 4 years, and 22% report losses due to
too much rain and floods.
• Major non-rainfall yield risks—pests, disease, livestock or birds
destroying the crop—are quite observable and do not appear as highly
covariate as rainfall risk.
• Take reports of risks to crop production over five years for households. Run
fixed effects regression with time dummies. R-squared relative to rainfall
shocks: disease=0.27, insect=0.38, livestock and birds=0.04.
10. Context
Funeral insurance mutuals are widespread throughout rural Ethiopia, in the
areas where we are working almost all households are members of at least
one (often more than one). Characteristics of these societies:
• Although starting as informal associations asking for contributions at the
time of a funeral, many now require up-front payments (perhaps
supplemented at the time of death)–looks like insurance.
• Payments are equal regardless of age, wealth or family size–strong
preference for equality and simple rules. Wealthier farmers who want
more funeral coverage join more iddirs.
• Membership is largely defined by geographical location. Regular
meetings and mandatory funeral attendance put a physical constraint on
how far away your iddir can be. Members are neighbours.
• Iddirs have evolved to performing other functions: health insurance,
livestock death, fire, loans in the event of harvest losses, loans for inputs.
Not all members will pay for other benefits (for example only livestock
owners would pay the additional fees for livestock death coverage).
• Coverage that implies an unobservable event: multi-purpose health
insurance, harvest losses etc is often given as a loan rather than a grant.
11. Policies offered
• Nyala Insurance S.C. introduced an individual index-based rainfall
insurance in rural Ethiopia.
• The policies took the form of monthly coupons whereby a fixed payout
would be due if the monthly rainfall fell short of a particular precipitation
target (Hill and Robles 2011)
• Policies were calibrated using the historic data from the local rainfall
station.
• Six policies were introduced:
• Two policies for each of the rainy season months: July, August and
September.
• ‘Severe Shortfall’: For a premium of 100 Birr, the farmer could receive a
payment of 500 Birr with a chance of 1/5.
• ‘Very Severe Shortfall’: For a premium of 50 Birr, the farmer could receive a
payment of 500 Birr with a chance of 1/10.
12. Timeline of activities
• January: listing and survey of iddirs in selected kebeles.
• early May: iddirs randomly allocated to training A (individual) or B (group
sharing).
• mid May: Leaders of iddirs attended training A or B. Nominated
members to receive individual training.
• beginning June: demand forms collected and policies issued.
• mid-end June: household survey– sampling frame of the household
survey was constituted by the full memberships of iddirs that took part in
the training exercises.
• beginning July to end September: insurance coverage.
13. Federal structure of iddirs
• Some iddirs have developed a federated structure whereby a large iddir
has several smaller “sub-iddirs" underneath it.
• We did not always know the full federated structure but restricted the list
of iddirs to include only iddirs of 100 members or more–this excluded
most sub-iddirs.
• All iddir leaders of the large iddir (typically a committee of 3-5), and
leaders of any of the sub-iddirs within it were eligible to attend the
training session.
14. Federal structure and mixing of training types
• Excluding iddirs of 100 members or less did not perfectly exclude all
sub-iddirs and some of the iddirs on our list were sub-iddirs.
• For these sub-iddirs there was a possibility that they were allocated to
one training type whilst their main iddir to another. At the main iddir level
these iddirs were mixed–some leaders had type A training and some
type B training.
• All analysis is conducted using the type of treatment leaders in the main
iddir as a whole attended. There are thus three treatment groups:
• All leaders received training A
• All leaders received training B
• Some leaders received training A and some training B
• The probabilities of allocation to treatment varied depending on whether
or not an iddir had one of its sub-iddirs on the list.
• No sub-iddirs on the list, probability of allocation to treatment A or B was 0.5
• Iddirs with one sub-iddir on the list, probability of allocation to treatment A or
B was 0.25 and mixed was 0.5.
15. Tests of balance
• Data from baseline survey of iddirs and characteristics of members
collected in the household survey that are unlikely to have changed as a
result of the training (characteristics of the household head, land
ownership etc.) to test balance across training type:
• of 39 variables tested 3 are significantly different at 5% or less.
16. Assessing compliance
• We compare training allocation and training attendance to assess
compliance.
• We find there was one training team that did not stick to the protocol–half
of the non-compliance we observe comes from this team alone. We test
robustness of our results to including and excluding this data (excluding
2 kebeles)
• Aside from this we observe 92% compliance with training allocation at
the iddir levels. 6 out of 71 iddirs went to the wrong training session, not
clear why:
• 1 iddir switched from A to B
• 1 switched from mixed to B
• 1 switched from A to mixed
• 3 switched from B to mixed
• We test the robustness of our results to endogenous switching by
estimating the ITT and LATE also.
17. Description of take-up
Training A Training B Mixed
Leader 0.37 0.53 0.31
Trained member 0.33 0.53 0.44
Untrained member 0.02 0.03 0.00
Total 0.21 0.34 0.23
18. ATE estimates
Farmer trained (tij ) 0.317*** 0.299*** 0.295***
(0.059) (0.055) (0.055)
Training B (gij ) 0.140* 0.006 -0.045 -0.031
(0.077) (0.030) (0.038) (0.043)
Mixed training -0.037 -0.020 -0.042 -0.017
(0.078) (0.020) (0.040) (0.046)
gij ∗ tij 0.190* 0.242** 0.242**
(0.108) (0.098) (0.098)
Mixed ∗tij 0.073 0.141 0.141
(0.090) (0.088) (0.087)
Constant 0.233*** 0.020 0.225*** 0.222***
(0.045) (0.020) (0.081) (0.077)
Basic characteristics No No Yes Yes
District fixed effects No No No Yes
Observations 77 332 329 329
R2 0.059 0.224 0.348 0.351
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
19. ATE estimates, reduced sample
Farmer trained (tij ) 0.317*** 0.292*** 0.285***
(0.059) (0.056) (0.056)
Training B (gij ) 0.140* 0.006 -0.048 -0.028
(0.079) (0.030) (0.039) (0.045)
Mixed training -0.027 -0.020 -0.046 -0.027
(0.092) (0.020) (0.047) (0.049)
gij ∗ tij 0.190* 0.246** 0.250**
(0.108) (0.098) (0.098)
Mixed ∗tij 0.083 0.123 0.117
(0.103) (0.103) (0.101)
Constant 0.233*** 0.020 0.209** 0.209**
(0.047) (0.020) (0.093) (0.089)
Basic characteristics No No Yes Yes
District fixed effects No No No Yes
Observations 71 292 290 290
R2 0.053 0.228 0.346 0.352
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
20. ITT and LATE
ITT LATE
Farmer trained (tij ) 0.315*** 0.296*** 0.293*** 0.292*** 0.268*** 0.264***
(0.056) (0.054) (0.053) (0.066) (0.062) (0.061)
Training B (gij ) 0.004 -0.025 -0.010 0.011 -0.025 0.002
(0.027) (0.037) (0.040) (0.033) (0.048) (0.055)
Mixed training -0.019 -0.026 -0.025 -0.020 -0.030 -0.030
(0.018) (0.050) (0.058) (0.020) (0.052) (0.061)
gij ∗ tij 0.162 0.198** 0.193* 0.178 0.219* 0.209*
(0.104) (0.098) (0.098) (0.129) (0.125) (0.126)
Mixed ∗tij 0.185* 0.219** 0.218** 0.209 0.248** 0.249**
(0.111) (0.109) (0.108) (0.127) (0.121) (0.119)
Constant 0.019 0.190** 0.198** 0.019 0.187* 0.196**
(0.018) (0.094) (0.091) (0.020) (0.096) (0.093)
Basic characteristics No Yes Yes No Yes Yes
District fixed effects No No Yes No No Yes
Observations 292 290 290 292 290 290
R2 0.226 0.347 0.351 0.218 0.335 0.343
Robust standard errors in parentheses
21. Number and value of policies bought
Number of policies Value (Birr)
ATE ITT ATE ITT
Farmer was trained (tij ) 0.342*** 0.345*** 19.677*** 20.142***
(0.083) (0.075) (5.694) (5.205)
Training B (gij ) -0.019 0.002 -1.683 -0.270
(0.055) (0.049) (3.852) (3.547)
Mixed training 0.012 -0.010 0.644 -0.180
(0.069) (0.084) (4.699) (5.781)
gij ∗ tij 0.308** 0.231* 16.104* 11.860
(0.133) (0.128) (8.386) (7.789)
Mixed training ∗tij 0.112 0.283 8.314 17.291
(0.144) (0.173) (9.899) (11.362)
Constant 0.205 0.200 12.362 12.157
(0.147) (0.147) (10.292) (10.258)
Basic characteristics Yes Yes Yes Yes
District fixed effects Yes Yes Yes Yes
Observations 290 290 285 285
R2 0.265 0.264 0.219 0.222
Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
22. Who purchased insurance (ATE)
Farmer was trained (tij ) 0.279***
(0.057)
Iddir leader 0.056
(0.082)
Training B (gij ) -0.023
(0.045)
gij ∗ tij 0.261**
(0.118)
gij ∗ tij ∗ iddir leader -0.062
(0.150)
Mixed training -0.009
(0.047)
Mixed ∗tij 0.158
(0.102)
Mixed ∗tij ∗ iddir leader -0.058
(0.140)
Constant 0.208**
(0.082)
R2 0.352
23. Discussing insurance
(1) (2) (3) (4) (5) (6)
Talk Talk Talk Number above 50 1 to 15
Training B (gij ) 0.103* 0.105* 0.108 -9.575* -0.05* 0.129**
(0.057) (0.059) (0.065) (4.889) (0.03) (0.061)
Mixed training -0.047 -0.032 -0.022 -8.994* -0.05* 0.120*
(0.066) (0.072) (0.074) (4.922) (0.03) (0.066)
Constant 0.797*** 0.925*** 0.935*** 17.390*** 0.05* 0.797***
(0.040) (0.119) (0.122) (4.788) (0.03) (0.055)
Basic char. No Yes Yes No No No
District f.e.s No No Yes No No No
Observations 198 198 198 161 161 161
R2 0.024 0.100 0.102 0.037 0.029 0.033
Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
25. Salience of covariate nature of basis risk
Is the rainfall measured at the weather station a good measure of the rainfall
on your field?
Farmer was trained (tij ) 0.242*** 0.243***
(0.068) (0.077)
Training B (gij ) 0.281** 0.335***
(0.117) (0.118)
Mixed training -0.053 -0.031
(0.084) (0.101)
gij ∗ tij -0.335*** -0.353***
(0.107) (0.110)
Mixed training ∗tij 0.020 0.049
(0.110) (0.117)
Constant 0.245*** 0.423***
(0.065) (0.127)
Basic characteristics No Yes
District fixed effects No Yes
Observations 332 329
R2 0.058 0.104
Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
26. Training and purchases
(1) (2) (3) (4)
Discussed Decided Shared Bought because
with others jointly policies of others
Farmer was trained (tij ) -0.417 0.125 -0.040 0.320
(0.513) (0.348) (0.292) (0.463)
Training B (gij ) 0.000 -0.000 -0.000 1.000
(0.711) (0.482) (0.404) (0.643)
Mixed training -0.143 0.115 -0.085 -0.120
(0.144) (0.097) (0.082) (0.129)
gij ∗ tij 0.010 -0.062 -0.057 -1.007
(0.724) (0.491) (0.412) (0.654)
Constant 1.000* 0.000 1.000*** 0.000
(0.503) (0.341) (0.286) (0.454)
Observations 83 83 82 84
R2 0.039 0.051 0.017 0.048
Standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
Mixed training ∗tij dropped because of collinearity
27. Conclusions
• Demand was substantially increased when groups were exposed to
training that encouraged sharing of payouts within a group.
• One mechanism for this higher level of take-up may come from the ability
of groups to mitigate some of the basis risk inherent in these products.
• Data we have is consistent with this view and suggest that if farmers are
increasing informal risk sharing it is being done in small groups of
selected farmers.
• Future work:
• What would be the magnitude of the effects in less cohesive groups, or
groups that are not as familiar with formalized risk-sharing?
• Do sharing rules have to be formalized at the time of insurance purchase?
What kind of rules can members credibly commit to?
• How much of εij can be shared? Observable events, binary events. State
contingent loans for cases where moral hazard is likely?
• What about µj ? Can savings or contingent credit help groups insure this
across time?
• There are potentially other advantages to group contracts (reduced per unit
transaction costs, increased trust in insurer), what is the combined effect of
offering insurance through groups?