Life and Death of Roscas:Leadership, Election and Screening ∗

Olivier Dagnelie†

Instituto de Analisis Economico CSIC

July 2008

Abstract

In this research, we bring new light on the factors influencing the duration of informalsavings groups and their risks of failure using an original data set containing information onliving and dead roscas from Cotonou, Benin. A survival analysis highlights the role of theleadership structure, the influence of democracy inside the group and the importance of thescreening process before accepting new members on the probability of failure. We also putforward how incentives can be decisive as to the success of such groups.

1 Introduction

A Rotating and Savings Credit Association is an informal savings group. At each meeting ofthe group, the fixed contributions are gathered from all the members. The whole amount ofthe collected money, what is called the pot, is then given to one of the members according to apredefined rule of the group: lottery draw, decision by the ruling body, bidding process. Eachmember has to keep contributing until every one has received the pot, which marks the end ofthe cycle. One can immediately realize that members might be tempted to try to get the potat the very beginning of the cycle and then stop contributing. This creates important incentiveproblems. The challenge of this kind of group is therefore to make everyone contribute until theend of the cycle. Should it not be the case, the group would be very likely to collapse. In thisresearch, we intend to underline the factors influencing the survival of roscas.

Members defaulting is a problem of primary importance in the context of roscas and is oftenacknowledged in the literature. In that matter, several papers study the role of the rosca leader

∗I am deeply grateful to Philippe Lemay-Boucher. Thanks are also due for helpful comments to Jean-Marie Baland,Gani Aldashev, Siwan Anderson, Jean-Philippe Platteau and Karl-Ove Moene and seminars and conferences partici-pants at Namur, ISEC, IAE, CSAE Conference 2007 and NEUDC 2006. Financial support from the CRED (Universityof Namur), the National Bank of Belgium and the Barcelona Economics Program of CREA is greatly acknowledged. Ialso whish to thank our team of enumerators: Felicite Chadare, Maurille Gandemey, Shadia Gbaguidi, Calixte Houedey,Euphrem Lankoutin, Pierre Meliho, Raoul Tchiakpe and particularly Charlemagne Tomavo. The encoding process wasdone with the help of Amelie Bodson and Gaetan Dagnelie. Special thanks to Yvonnka Kelly-Dagnelie.†e-mail: olivier.dagnelie@iae.csic.es, address: Instituto de Analisis Economico CSIC, Campus UAB, 08193 Bellaterra

(Barcelona), Spain. tel : +34935806612, fax : +34935801452

1

(president of the group) in the sustainability of the group. Kurtz (1973) evokes the fact that as thepresident receives the first payment1, he must pay off2 the members of the group if somethinggoes wrong and is often the only one to lose money. This is also brought up by Bouman (1995)who claims that in many countries the organizer is paid because running a rosca has almostbecome a profession. As the role of the president is to minimize the risk of default, according tovan den Brink and Chavas (1997), the best incentive scheme is to have him receive the last potmaking him the residual claimant. Handa and Kirton (1999) provide a detailed analysis of therosca leader in which they consider the ’banker’ to be the internal governance structure mini-mizing transaction costs. They confirm his role of risk minimizer as lender of last resort, sinceeither he receives the pot at the end of the cycle or he keeps the pot to cover a potential default.They claim that a sole individual making decisions for the group is more efficient than trying toreach a consensus among the whole group. It is indeed likely that groups run by their founder(founder groups) are well managed as in this case, the president who invested time, effort, rep-utation, etc. will be anxious for her group to flourish. We fear however that, in the absence ofthe right incentives, there is a trade-off between the efficiency gains coming from having onlyone decider and the losses due to misuse of power or mismanagement. We therefore think that,to the exception of founder president, an intermediary solution such as a committee of rulerssharing responsibilities is probably more efficient. One of Handa and Kirton’s main results isthat paying the president for running the rosca diminishes the risk of experiencing problems3

in the group, hence enhancing sustainability4. Even if we do not find an influence of payingthe ruling body on the occurrence of problems, focusing on president groups we observe that itdiminishes the likelihood of group failure (see Table 6).After being loosely mentioned several times in the literature, the issue of randomness of theattribution of ranks has been evoked by Handa and Kirton who show that it diminishes theprobability of experiencing problems in one’s rosca. Actually, random roscas have the advan-tage of being fair and not prone to subjectivity or favouritism regarding rank allocation, theseregularly leading to conflicts inside the group. As Anderson, Baland and Moene (2004) put it,this method of allocation is preferred by the majority of members since each one has the sameprobability of receiving the pot early in the cycle5. It has however the drawback of not optimallygiving the pot to the one the more in need and, according to Anderson et al., of exacerbating theincentives to default. This implies that agents taking part in such groups have to be more vulner-able to social sanctions for the rosca to be sustainable. Our sample does not allow us to confirmthis point as logistic regressions (in Table 5) exhibit that, compared to decision rosca members,random rosca members are wealthier, likelier to be salaried and not to be owner of their house6.Moreover, random roscas resort less to making an inquiry before acepting new members as pre-

1In 36.4% of the roscas of our sample, the time of pot reception of the ruling body is stated in the rules of the group.In 97.7% of these cases, they receive the pot at the very beginning of the cycle.

2Not fulfilling his obligations would tarnish his reputation and prevent him from going on running roscas.310% of the members in their sample experienced problems in their group while 26% did in our sample. Interestingly,

only 1.5% of the members of their sample reported a delinquent president whereas there was 3.3% in our case.4If we admit ’experiencing problems’ seems to be a good proxy of the non sustainability of a rosca, the way the group

tackles problems is also likely to be an important factor influencing the duration of the group. The latter is probablylinked to the organizational design of the group.

5They also prove that multiplicative discounting implies that the majority of members strictly prefers the allocationof ranks to be random.

6One can suspect that, being wealthier and salaried, these members attach less importance to the time of pot recep-tion. These results hold if we restrict the sample to active members of roscas.

2

sented in logistic regressions on the structure of the group (Table 4). Regarding group survival,our empirical analysis shows that if there is a significant relationship between randomness andthe probability of group failure, it is negative, randomness increasing the probability of groupdeath. Results however lack robustness in this regard.

As to the complete falling apart of roscas, references are quite rare. Bouman (1995) quotesthe paper of Wu (1974) bringing up the case of organizers of Huis disappearing with the moneyin Papua New Guinea at the time of independence while Anderson, Baland and Moene (2004)touch on the problems of cheating and the lack of money to contribute. These papers are how-ever of little help regarding the question we are interested in as they do not deal with the specifictopic of group failure and the factors reducing or enhancing the duration of roscas.

Recent advances in the literature deal with the sustainability and the stability of the agree-ment between the members of such groups. Anderson, Baland and Moene make clear that with-out social sanctions and contracts, roscas are bound to fail as the first member to receive the pothas no incentives preventing her from leaving the group. As for Ambec and Treich (2007), theypoint out that, if people suffer from self-control problems and contracts are binding7 - whichmight be regarded as a strong hypothesis in such an environment - roscas are stable financialagreements. Basu’s plausible assumption (2006) of rosca members being hyperbolic discoun-ters8, proves that even in the absence of social sanctions and contracting, roscas can be effectivecommitment savings devices.

Beyond the differences related to the order fixing process, and the more trivial differencesrelated to the frequency of the meetings, the amount of the contribution, or the number of mem-bers, what is striking is the extent to which groups can differ due to the ruling structure of thegroup. From our field observations, we can distinguish two kinds of groups: those run by a pres-ident alone (president groups) and those led by a committee consisting of a president, treasurer,secretary, etc. (committee groups). The former case mainly refers to a group the sole function ofwhich is to render a financial service. It is run either by the founder who decided to set up anassociation hoping it would provide her a means of saving and living (founder groups) eitherby a non founder president (non founder groups).9 In the case of committee groups, as most ofthe time they turn to election, becoming a committee member depends on the group’s will. It isconsidered to be an honour which does not bring much advantage beyond social esteem. Thiskind of group fulfills different functions: primarily of savings vehicle but also of socializing op-portunity. These roscas are more cohesive as demonstrated by tests of means difference on thesharing of the leftover funds10, on compulsory attendance, on the prerequisites before joininga committee group and on the proportion of groups having other activities (indemnity funds,dancing and chanting, meals, etc.).

During the survey, we noticed that ex-members of groups run by a president alone oftencomplained about their past experiences in such roscas. Moreover they tended to have lost

7The latter implies defaulting is not possible (or at infinite cost). There is therefore no need for social sanctions in thisframework.

8Recent papers on participation to rosca tend to suggest that it is frequently one of the possible reasons to join. SeeGugerty (2007), Ashraf, Karlan and Yin (2006) and Dagnelie and LeMay (2008) among others.

9Even if most of them are being paid for managing the group, these presidents can not be considered to be profes-sional.

10Most groups claim for a part of the pot to be kept in the cashbox which also contains the fines paid during thecycle. In a majority of committee groups, this amount of cash is redistributed on the occasion of a party at the end of thecycle while this is much less frequent in president groups. This gives evidence that the objectives of these two kinds ofassociations are likely to differ, president group being less cohesive.

3

money in the ending process of the group. That is why we are going to verify if our intuitionsas to the risk enhancing nature of a single person management and importance of incentives areconfirmed by the reality of the field. We begin by providing a brief description of our surveyin section 2. Descriptive statistics about reasons given for the group stopping are displayed insection 3. These lead us to restrict our sample to groups which stopped without consensus ofall the members. Section 4 deals with the typology of groups and addresses the question asto the group differences in terms of functioning and regulation. Aware of selection issues, weinvestigate in Section 5 the kind of people who are attracted to the different rosca organizationaldesigns. In Section 6 we present a survival analysis on the groups which fell apart withoutconsensus of the whole membership. Section 7 therefore concludes that, to the exception ofassociations run by their founder, president groups seem to be likelier to know an inefficientstopping. We also put forward how incentives, election and screening issues play a positive rolein rosca survival.

2 Description of our Survey

We use data we collected in 2004 in the two districts of Vossa and Enagnon located on the out-skirts of Cotonou, about 1.1 million of inhabitants. These areas are known to the city’s authorityto be the poorest. Vossa is located near an inner bay of fresh water and accommodates a com-munity of fishermen. Enagnon encompasses an adjacent slum called Enagnon-Plage which isinhabited by a majority of fishermen living in huts on the beach. Vossa and Enagnon are neardowntown Cotonou where a large part of the inhabitants work and commute everyday. Noformal savings and investment institutions, neither public nor private, such as banks and NGOsare present in these two districts. The choice of these zones goes back to a first mission in 2002,which revealed that many informal groups such as indemnity funds and roscas were activethere.

During the first three months of 2004 we surveyed 497 households: 110 in Vossa and 387in Enagnon, of which 114 are located in Enagnon-Plage. The selection of each household wasdone randomly. The first wave of interviews aimed at creating contacts, getting housing char-acteristics and obtaining information on each member: religion, activity, education, income,expenses, savings, etc. A second round was needed for members of informal groups. Duringthis visit enumerators collected detailed information on all the current and former group(s) theybelong(ed) to. For a maximal accuracy, all members of each household were interviewed sepa-rately throughout the successive waves of our survey so that tricky issues related to expenses orincome were only tackled privately. Particular attention was thus put on confidentiality whichwas strictly followed by our enumerators.

All the 497 households we surveyed represent 2083 individuals, 894 of which are undersixteen. We are thus left with a sample of 1179 individuals divided in 604 women and 575 men.We collected information on 270 members of 242 roscas among which 192 were alive and 50dead.

We only have data on the interrogated individuals and the groups they joined (structure,functioning, problems encountered, etc.) but lack information about the president and the othermembers of the group who were not interviewed. Nevertheless, in an attempt to mitigate theselection bias, we introduce in the regressions using group features as independent variables

4

the group’s mean11 of several key variables of members’ characteristics.As people were interrogated once about their participation in groups which do not necessarilyexist anymore, these are retrospective flow sampling data12.

3 Given Reasons for Group Stopping

From Table 1, one can immediately notice that the first given reason refer to a decision priorto the group stopping. Members of the groups had agreed to stop the group once everybodywould have a motorbike or after a fixed number of cycles. Then it appears that 11 groupsdisbanded following the departure of several of their members who moved to another part ofBenin or to another job. It is apparently totally exogenous to the organizational structure orfunctioning of the roscas. The 13 associations in question can therefore be removed from thesample of dead groups to study as we are interested in inefficient stoppings attributable to thestructure of the group. The next two given reasons concern a decision made by the ruling bodyto stop or interrupt13 the association. It is likely that some of the members have been bothered bysuch a move as the groups ceased their activities without consensus of the whole membership.However, as in all the previously mentioned cases, no member of these organizations seems tohave lost money.

Not a single dead rosca and only four alive groups stated in their rules the duration of thegroup. Moreover, only 7.4 % of the 242 roscas include at least one member who before joininghad decided how many cycles she wanted to take part in. As a very large majority of membersprobably did not want the rosca to stop, it seems reasonable to consider all the 37 dead groupsto have undergone an inefficient stopping.

Furthermore, a close look at Table 1 reveals that 13 roscas run by a president alone and 7 by acommittee caused loss of money to at least one of their members, meaning that 20 out of the 37(54.1%) group failures implied a loss of money for at least one of their members. The differenceof proportion is significant14 at 5%.

As predictable, one can note that the consequences of the death of a committee member aremore serious when the president is the only one to lead the rosca. Money problems concern alarge majority of the given causes of group failure. It refers to different cases: irregular paymentsleading to unequal pots, pots in several instalments15 or wrong timing of the pot reception butalso cases where the member decides to leave the group and not to contribute any more after

11For some groups, we possess information from several members. We therefore took the mean by group. These vari-ables, each expressed as a group proportion, are those which appear to be the most significant in our logistic regressionsof group structure on members’ characteristics (Table 5).

12We collected data in 2004 on groups wich were created on different dates - the first ones starting in the 1960’s. Thiscould have produced a biased sample if old roscas were significantly different from younger ones. The flaws of sucha sampling method are taken into account in the survival analysis. We introduced dummies for starting dates and didrobustness checks removing these control variables with similar results. We think this validates our sample.

13These roscas were interrupted and did not start a new cycle since then. Each of them is considered dead by theinterrogated member.

14The significance of the difference remains on a subsample of ex-members of rosca - i.e. members whose rosca diedor who left before the group stopped functioning. Eight losses of money happened in founder groups and five in nonfounder groups, this difference however is not significant.

15It can be highly undesirable if rosca members suffer from self-control problems and use the group as a means todiscipline themselves to save.

5

receiving the pot and last but not least president leaving with the members’ contributions16.Even if these reasons cannot be taken for granted, it seems plausible that the rosca deaths

looked at in the subsample of 37 groups are inefficient and bring about welfare losses for theirmembers.

4 A Typology of Groups

Field observations led us to remark that ex-members of president roscas tended to complainabout their past experiences. This brought us to objectivize those impressions in terms of un-desired mortality of groups and to investigate the organizational structure of groups. Figure1 plotting the smoothed hazard rates split according to the ruling structure of the association,not taking into account covariates, gives the impression that president groups and particularlygroups of which the the president alone is not the founder (non founder groups) have a lowersurvival probability as their hazard rate is higher than committee groups on the whole period.Moreover tests of differences of means reveal that committee and president groups seem verydifferent in terms of functioning which incites us to investigate further the features of these twokinds of groups.

We therefore ran a multinomial logistic regression of the ruling structure on groups’ charac-teristics. As our survey was carried in three different areas, we introduced fixed effects remov-ing the area-specific component from the residuals and eliminating the endogeneity caused byunmeasured area characteristics (Pitt et al., 1999). In order to take into account auto selectionin roscas, six variables of members’ characteristics were added as controls along with dummiesregrouping roscas which started during the same period of time. Furthermore, the design ofour survey was such that the probability for households to be selected in our sample was dif-ferent in the three studied areas which could lead to inconsistent estimates. We thus introducedsampling weigths for our estimates to be independent of the sample design.

From Table 3 and Wald tests run on the coefficients we notice that founder groups are mostlyfemale only and mixed roscas. They are also the least sharing the leftover funds of the cycle,the least using random allocation of turns and the least numerous roscas. This could indicatethat founder presidents, centralizing the decisional power, try to manage their group efficientlyand prefer to limit the number of members to be able to control them more easily. Variablescreated to take into account as much as possible of auto selection issues reveal that the membersattracted to these groups are likely to be richer.

This regression also allows us to notice that the rules of the group are most of the timeonly oral in non founder groups. We guess it blurs the regulation and reduces the ability ofmembers to go against the president’s decisions. Furthermore, attendance to meetings does notseem to be compulsory in these groups contrarily to committee ones. This probably rendersthe members’ cohesion weaker as they do not know each other and might not be aware of thedecisions made by the president. Interestingly, these groups resort more than committee ones towhat we describe as ’Important sanctions’: seizure of goods and recourse to the administrativedelegate for punishing bad behavior.

16Contrarily to common beliefs, a lot of roscas keep money in a cashbox - it is the case for 95.9% of the groups of oursample - coming from fines, deposits, share of the pots, etc. This money is mainly used to make loans, to organize aparty or to pay the wage of the group’s manager.

6

Tests of differences of proportion in Table 2 show that it is possible to rank the different rulingstructures as to the concentration of decisional power. In committee groups, it is more or lessexpected that several people share the responsabilities related to the organization of a group. Itis more surprising that non founder presidents make less decisions than founder presidents17. Apriori it is difficult to know if it must be considered as a dilution of power and hence as negativeor as a more democratic way to manage the group and thus as positive. In president groups, allthe decisions related to taking in or excluding members and changing the order of pot receptiontend to be made by the president only whereas the committee or even all the members wouldbe able to take part in the decision process in committee groups. This concentration of powermay bear important consequences as the group is more vulnerable to the problems related tothe president if he is dealing alone. The latter can indeed die, fall ill, steal the contributions orsimply mismanage the group, all of which can prove detrimental for the group.

We present in Table 4 six regressions, the aim of which is not to uncover causalities but togo beyond tests of differences of means and display conditional correlations between groupfeatures. Election of the ruling body, existence of written rules, sharing of the leftover funds andcommittee groups seem to be associated while a president alone is generally paid (and thereforedoes not share the leftover funds of the group), makes inquiries but is not elected and does notresort to written rules.

Despite being formally paid more often than committee members, the president alone, beingthe only residual claimant on the leftover funds of the rosca, has the ability to grant himselfsome financial advantage18. These pieces of information confirm the impression that the rulingposition is only honorary for committee members while it can offer financial advantages topresidents running the rosca alone. Paying the ruling body seems to imply that attendance tothe group meetings is not compulsory. In this case, the president is supposed to collect themembers’ contributions which do not have to be paid during the meeting. This indicates thatthe associations paying their leader(s) are less cohesive, less oriented towards social functionsand more to pure financial services. Moreover, Table 2 suggests that president groups providetheir members with more economical advantages19. It is interesting to note that the members ofthese associations appear to be richer and that important sanctions are used to deter defaulting.These sanctions appear to be more prevalent in ethnically homogenous groups, which relates toAnderson and Francois (2008). They indeed posit that in those very cohesive groups, it is costlierfor everyone to dismiss one member which leads to resorting to formalism and penalties to keepthe incentives preventing people from defaulting. We believe that imposing important sanctionsrespond to the same logic.

As to the reported problems or roscas, it seems that they are less common in female onlygroups and tend to decrease in groups where the average age is higher. Not surprisingly, roscaswithout problems have a more efficient screening process prior to taking in new members. Ran-

17If the sample is split in four: committee - president and election - no election, the elected committee is the leastconcentrating power option, while differences between president groups are never significant. This could indicate thatthe sample splitting according to founding status is relevant.

18The few following statistics give a clear description of this phenomenon : 87% of the president in committee groupsare not paid while they are 42% in president groups. 72% of the committee groups share the leftover funds at the end ofthe cycle against 20% of the president groups. If we cross those two pieces of information, we see that only 12% of thepresidents alone neither get a wage nor the leftover funds of the cycle whereas 64% of their counterparts in committeeare in the same case.

19Economical advantage refer to the cases where members prefer to buy (on credit or not) at the shop of one of theirfellow members or doing business with them or where they claim they met their employer (past or present) in the group.

7

dom roscas are more common in president groups to the exception of those run by their founder.This could be related to the higher decisional power of the latter who prefers to fix the order ofreception and adjust it to members’ requests. Apparently these random roscas use more oftenwritten rules and impose meeting attendance which is related to the fact that people know eachother inside the group but inquiries are less often made before hosting new members.

From these pieces of information, we can conclude that committee groups not only involvetheir members in the management of the rosca but also are more participation oriented andhence more cohesive. As to the president groups, they seem to offer mainly financial servicesand to concentrate the decisional power into the hands of their president. We must turn toa duration analysis to check the impact of these organizational features on the risk of groupfailure.

5 Selection into groups

Even if they are probably constrained by the set of roscas they know and the screening processin place in the groups, it is clear that there is an element of choice as to the rosca people join.Whether they are more attracted to social interactions or only to financial services, whether theyhave a low or high social capital, whether they participate simultaneously in several roscas, allthese characteristics are expected to influence the kind of roscas people will eventually becomea member of.

In this section, we intend to address these questions presenting results of logistic regressionsof the group features on members’ characteristics, run on the whole sample of current or pastmembers’ participations. As errors within households, our sampling unit, are likely not to beindependent, we used cluster effects that take notice of the correlation between observationscoming from the same environment. This produces robust standard errors which would havebeen underestimated without this correction. We also controlled for ethnic affiliation and areafixed effects and used sampling weights.

A multinomial logistic regression (the first two columns of Table 5) where the dependentvariable is split according to the ruling structure of the group conveys that members of com-mittee groups are more often in couple and generally take part in a single rosca contrarily tomembers of president groups. Being in couple broadens the set of people in one’s network andcan therefore be interpreted as an increase in social capital. This is compatible with the ideathat people mainly interested in a financial service will join several president groups. Moreoverthose members participate less in groups where attendance is compulsory. Apparently, the typ-ical member of a non founder group, in comparison to a committee group, is male and single,while he appears to be less educated than the average member of a founder president group. Alogistic regression of the participation in president groups (not presented) describes the averagemember of such group as a young male living on his own, freshly arrived in the neighborhoodand simultaneously taking part in several roscas. It is therefore likely that on average thesemembers belong to smaller social networks which could render them more vulnerable to a neg-ative shock.

Logistic regressions presented in Table 5 suggest that the average member of groups wherethe leader is elected is a mature house owner who takes part in only one rosca. Interestingly,groups inquiring prior to taking in new membership have richer female members. Whereas themembers of groups paying their ruler(s) are younger, have their job for a longer time and simul-

8

taneously participate in several roscas which echoes the findings of Section 4. Apparently, oldermembers are wiser as suggested by the negative significance of the ’age’ variable in the regres-sion explaining the occurrence of problems inside the group. As to the members’ characteristicsof random roscas, they appear to be richer, not to own their house and to get a regular wage,two features they share with those who simultaneously take part in several groups. A regularwage indeed removes much of the uncertainty bearing upon members and allows them to savemore since it greatly diminishes the fear of not being able to repay one’s debt after receiving thepot. Actually, being salaried is highlighted by Banerjee and Duflo (2008) as the probable ”mostimportant difference between the poor and the middle class”.

These pieces of information tend to confirm that president groups are more oriented towardsfinancial services and less towards social events. They also comply with the idea that the agentsattracted to non founder groups could have a lower social capital than the rest of the roscamembers.

6 Survival Analysis

In order to perceive what factors could reduce or enhance the risk of dying for a group, we haveto turn to a survival analysis. Cox proportional hazard model, which is in continuous time,allows to capture the influence of covariates on the risk of dying. As we used retrospective flowsampling and asked about duration in months, it seems more appropriate to work with groupeddata in discrete time. This implies that we have to use a complementary log-log transformationof Cox model20, being of the following form :

log (− log (1− F (t|x))) = x′β + log Λ(t) (1)

We have thus a complementary log-log regression where the baseline hazard, log Λ(t), is de-signed to be fully non parametric with a binary variable for each interval of constant hazard.We created the dummy variables in such a way that there is at least one failure in each interval.As most of the groups of our sample are still alive, we have censored data which we are takinginto account. Given the groups have different starting dates, we introduced dummy variablesgrouping together several points in time when the roscas were created.

With this technique, and by controlling for a set of characteristics, we are going to test theimpact of the organizational structure and rules of the group on the survival probability of beni-nese roscas. However, two sources of endogeneity are identified : non-randomness of the grouprules and membership. As people chose the groups they will join, even if some individuals areprobably constrained in their choice, we are in the presence of auto selection. Because mem-bers are not randomly distributed among groups, differences of public could (and probably do)influence roscas sustainability. The same principle applies to the non-randomness of rules. Aspeople decide which rules and sanctions are to be followed in the group and we cannot cap-ture this effect, we are not able to isolate pure marginal effects.21 In order to partially deal withthe selection issue and capture part of the selection bias, we introduced in the regression six

20See Appendix B for further details.21It is to be noted that in most president groups, members are not involved in the management of the rosca and

therefore do not have their say on its organizational design.

9

variables related to the kind of membership attracted to the group.22

In order to get the purest correlations possible, we also introduce area fixed effects and con-trol variables describing which relationship existed between the people who joined the group atits creation (neighbors, relatives, workmates, etc.). We once again introduced sampling weigthsfor our estimates to be independent of the sample design and consistent.

The results from our regressions are presented as coefficients of a complementary log-logregression (and not as hazard ratios). A negative sign must therefore be interpreted as the neg-ative influence of a variable on the risk of dying and hence enhances the survival probabilityof the group. In regression [1] of Table 6, three variables are particularly significant and inter-esting. The positive sign of ’Run by a president alone’ implies that, controlling for all the othervariables, such a group is more likely to know an inefficient stopping. Whereas the negative signof the variables ’Election’ and ’Inquiry’ means that group where the leader is elected and wherean inquiry is prior to any new membership are less likely to die. These three variables seem tobe the most important factors influencing the failure of groups. This does not mean howeverthat president groups are necessarily bound to fail and that this kind of leadership is deficient.Regression [2] indeed shows that groups run by a president alone who was not elected, beingthe founder of the group, tend not to be more likely to die compared to committee groups. In-troducing this variable does not change much most of the other coefficients to the exceptionof ’President alone’ and ’Election’. In both cases, the magnitude of the effect increases as thepositive effect of being a not elected president is captured in the ’Founder president’ variable.This shows that incentives have a role to play in the management of groups. Moreover, this isconfirmed in regression [3] where the leadership variable is interacted with a dummy taking asvalue one when the ruling body is paid. One can observe that paying the ruling body increasesthe survival probability of president groups, while it hardly influences committee groups as re-vealed by a Wald test of difference on these coefficients. This technique also shows that whenpaid a committee group is less likely to die than a rosca run by a president alone. Columns[4] and [5] of Table 6 present interacted terms of the three variables of interest. In column [4],the reference (the omitted dummy variable) is a group run by a not elected president who isnot the rosca founder. It appears quite clearly that this kind of group represents the worst casescenario as all the dummy coefficients exhibit a negative sign, a large majority of them beingsignificant. Wald tests allow us to point out the best23 structure in terms of survival probability,namely a committee group electing its president and inquiring about potential new membersbefore accepting them. This is confirmed by regression [5] where this kind of group is the ref-erence and all the other dummies are significantly positive and therefore likelier to die. Theworst group structure underlined by this regression and Wald tests is a not elected presidentnot making inquiries before new membership. Even if founder presidents seem to manage welltheir groups, other president groups perform worse than committee groups in terms of survivalprobability. This phenomenon is exacerbated by not electing the ruling body and not inquiringbefore accepting new members.

Two other noticeable pieces of information are worthwhile mentioning. Imposing importantsanctions24 seems to be an efficient deterrent to bad behavior as this variable is robustly negative

22Note that none of these were significant in our survival analysis and that our very robust results do not depend onthe presence of these variables.

23One can also affirm that an elected committee making no inquiry seems to be less likely to die than a group run byits founder.

24This variables is defined as a dummy taking 1 when the group resorts to seizure of goods or to the administrative

10

across the regressions. Male only groups seem riskier in terms of survival probability as thecoefficients of mixed and female only groups exhibit a negative sign and are often significant.

A look to hazard rates (computed without covariates) and to the baseline hazard of regres-sion [5]25 on Figure 2 reveals that the death probability of a rosca increases, reaches its peak aftera certain time, around two years, and then declines. This supports the idea that once these in-formal groups have completed more than two cycles, the most difficult times seem to be behindthem. It is indeed probable that defaulting members and / or ruling body manifest themselvesshortly after the creation of the group and precipitate the failure of the association.

6.1 Robustness Checks

Our results are robust whatever the specification of the regressions. Beyond the different choicesof covariates presented in Table 6, in other regressions based on columns [1],[2] and [5], we re-moved the different controls for areas, starting dates, selection effects and relationship betweenfirst members of the group. In all cases, the effects underlined before are significant. As furtherrobustness checks, we also removed simultaneously all controls and / or the sampling weights;this leading to no reverse conclusion. We also run cox regressions in continuous time with sim-ilar results.

As final check, we introduced normally distributed unobserved heterogeneity by runningcomplementary log-log panel regressions. Once again, the risk increasing nature of a not electedleader alone (to the exception of founder president) making no inquiries is confirmed.

All this tends to demonstrate the quality and robustness of our results which appear not todepend on a particular specification.

7 Conclusion

Our empirical evidence shows that, among the different kinds of groups observed in Cotonou,roscas run by a president alone are more likely to fail. This is even truer if they are not run bytheir founder. These associations are characterized by less participatory functioning and a moreconcentrated decision process which could influence the sustainability of the group.

The duration analysis confirms the important role of incentives, screening, leadership anddemocracy inside the group. It is not surprising to notice that incentives have an impact onthe survival of roscas. Whether they are paid, elected or the founder of the association, it isin the leaders’ best interests to manage efficiently the group. From another point of view, onecan consider that a democratic management of the group, involving members in its operation,provides the latter with the right incentives and improves cohesion. As to the screening issue, itis expected that being cautious before accepting new members reduces the probability of groupfailure. It happens through at least two channels. Firstly, if the screening process is successful,only members satisfying the criteria imposed by the rosca can join it which implies that theirsupposed quality for the group is higher. Secondly, once affiliated with such groups, individualsface a higher cost of defaulting in terms of reputation and, taking into account the probability

delegate to punish bad behavior.25The same exercise was run on regression [5] after removing the controls for starting dates. Baseline hazard estimates

follow the same pattern.

11

of group failure, in terms of expected savings. One can therefore assume they will do their bestnot to renege their obligations towards the group.

Regarding democracy, we can relate informal organizations and country states and evokethe literature on political economy. The latest results on the effect of democracy on growth sub-stantiates its positive impact as mentioned among others by Persson and Tabellini (2008). Besleyand Kudamatsu (2007) show that democracies on average perform better than autocracies butthey also underline that the variance among autocracies is higher and that the latter can exhibithigh growth rates. As to Bruckner and Ciccone (2007), they put forward that democracies bet-ter resist to conflicts, civil wars, than non-democratic countries in period of low growth whichcould also prove relevant for informal groups. As democratic leaders depend on the welfare oftheir electorate, it would be logical that they do their best for their members to be satisfied bytheir management.

If some individuals are systematically denied access to the safer groups in terms of survivalbecause of their social capital or lack of financial resources - what we cannot exclude -, thisconstitutes an additional reason for their expected savings being low. In the presence of suchsegmentation, interventions by NGO’s or public authorities would be justified to improve thesavings conditions of these doubly disadvantaged people.

APPENDIX

A Description of our Survey

A.1 Geographical Description

Vossa’s 63 hectares are encircled by stagnating waters and swamps which represent an impor-tant vector of disease. This district has not yet been divided into plots, though a long termproject was launched during our stay. Vossa is let to itself: the authorities of Cotonou have notyet paved any of its roads even its principal axis. The recurrent and important problem of floodshas not been dealt with even if it critically paralyses the area during a few months every year.Enagnon, a dense slum located on the Atlantic Ocean shore, has also received low attention andimportant sanitary problems have not been tackled yet. Half of its superficies of 60.1 hectareshas been divided into plots in 1998.

A.2 Survey Methodology

We selected households according to a random process. In Enagnon we succeeded in obtaininga map of the city and performed a simple selection of a lot according to an implemented randomprocess. In these two districts, many households live on the same lot in semi-detached rooms.Enumerators selected one room on a lot according to a clock-wise direction varying from lotto lot (for the first lot of the day they selected the first room clock-wise, for the second one thesecond room clock-wise and so on). In Enagnon-plage and Vossa we used a pseudo-randomprocess by which every tenth lot according to a specific direction was picked and then roomselections were done in a similar fashion to Enagnon. Overall only 3 households categoricallyrefused to be surveyed and were replaced by other randomly selected households. Enumeratorswere asked to pass several times and at different moments of the day, until contacts were estab-lished in such a way that none of the selected households were skipped. The most qualified of

12

our enumerators also acted as a supervisor and visited many households already interviewedin order to check the accuracy of the responses. Other than that we analysed every completedquestionnaire closely. Several appointments were held with every team of enumerators and incase of incoherence or lacking answers we regularly sent them back on the field. Questionnairesoften needed successive rounds of checks until final approval. As mentioned above we stressedthat the interview with every single household member had to be carried in his/her sole pres-ence in order to get as precise and reliable information as possible. Fear of divulging informationin front of other family members would have led individuals to lie or to refuse answering. Onaverage our four teams of two enumerators completed two questionnaires a day. By taking intoaccount intra-household secrecy, the survey was considerably lengthened by requiring specificappointments with each adult member. Another time consuming factor was the detailed partof our questionnaire concerning groups: we often needed more than an hour for a single group.We compensated every household for their time by donating 1500 CFA francs. Finally, withtwo previous missions, in 2002 and 2003, about eighty group interviews were carried out. Weattended regular meetings or met members of their governing body in order to acquire a betterunderstanding of their functioning.

B Complementary log-log regression

Cox’s proportional hazard model is of the form:

λ(t|x) = λ0(t)ex′β (2)

where λ(t|x) is the hazard rate at time t conditional on a vector x of covariates.If we want to take into account the discrete structure of our survey data, we have to trans-

form the continuous time Cox model as showed below.If we take a cumulative version of the proportional hazard model (where Λ(t|x) is the integratedhazard function at time t depending on covariates x), we get:

Λ(t|x) = Λ(t)ex′β (3)

By definition,Λ(t|x) = − log (1− F (t|x)) (4)

which simplifies to (1):log (− log (1− F (t|x))) = x′β + log Λ(t)

The baseline hazard, log Λ(t), is designed to be fully non parametric with a binary variable foreach interval of constant hazard.

13

References

[1] Ambec, S., Treich, N., 2007. Roscas as financial agreements to cope with self-control prob-lems. Journal of Development Economics, 82(1), 120–137.

[2] Anderson, S., Baland, J.-M., Moene, K.-O., 2004. Sustainability and organizational designin informal groups, with some evidence from Kenyan Roscas. BREAD Paper.

[3] Anderson, S., Francois, P., 2008. Formalizing Informal Institutions: Theory and EvidenceFrom a Kenyan Slum, in: Helpman E. (Ed.), Institutions and Economic Growth. November2008, Harvard University Press.

[4] Ashraf, N., Karlan, D., Yin, W., 2006. Tying Odysseus to the Mast: Evidence from a Com-mitment Savings Product in the Philippines. Quarterly Journal of Economics 121, 635–672.

[5] Banerjee, A., Duflo, E., 2008. What Is Middle Class about the Middle Classes around theWorld? Journal of Economic Perspectives 22(2), 3–28.

[6] Basu K., 2006. Hyperbolic Discounting and the Sustainability of Rotational Savings Ar-rangements. mimeo.

[7] Bouman, F., 1995. Rotating and Accumulating Savings and Credit Associations: A Devel-opment Perspective. World Development 23, 371–384.

[8] van den Brink, R., Chavas, J.-P.,1997. The Microeconomics of an Indigenous African Insti-tution: The Rotating Savings and Credit Association. Economic Development and CulturalChange 45, 745–772.

[9] Bruckner M., Ciccone, A., 2007. Growth, Democracy, and Civil War. mimeo, UniversidadPompeu Fabra.

[10] Dagnelie O., LeMay, P., 2008. Rosca Participation in Benin: A Commitment Issue. mimeo,UFAE-IAE working paper.

[11] Gugerty, M. K., 2007. You Can’t Save Alone: Commitment in Rotating Savings and CreditAssociations in Kenya. Economic Development and Cultural Change 55, 251-282.

[12] Handa, S., Kirton, C. 1999. The Economics of Rotating Savings and Credit Associations:Evidence from the Jamaican ’Partner’. Journal of Development Economics 60, 173–194.

[13] Jenkins, S., 2004. Survival Analysis. Draft book manuscript.

[14] Jenkins, S., 2004. Introduction to the Analysis of Spell Duration Data. ISER, University ofEssex.

[15] Kurtz, D. V., 1973. The Rotating Credit Association: An Adaption to Poverty. Human Or-ganisation 32, 49–58.

[16] LeMay, P., 2006. Insurance for the Poor: the Case of Informal Insurance Groups in Benin.mimeo, CRED, University of Namur.

14

[17] Persson, T., Tabellini, G., 2008. The Growth Effects of Democracy: Is It Heterogenousand How Can It Be Estimated?, in: Helpman E. (Ed.), Institutions and Economic Growth.November 2008, Harvard University Press.

[18] Pitt, M.M., Khandker, S.R., McKernan, S.-M., Abdul Latif, M., 1999. Credit Programs forthe Poor and Reproductive Behavior in Low Income Countries: Are the Reported CausalRelationships the Result of Heterogeneity Bias ?. Demography 36, 1–21.

[19] Wu, D. Y. H., 1974. To kill three birds with one stone: The rotating credit associations of thePapua New Guinea Chinese. American Ethonlogist 1, 565–583.

15

Figure 1: Smoothed hazard rate by ruling structure.

16

Figure 2: Smoothed hazard rate and nonparametric baseline hazard.

17

Total President Alone CommitteeLoss of Loss ofMoney Money

Consensus stopping 0.04 0 0 0.04 0Members’ departure 0.22 0.06 0 0.16 0Non consensus stopping 0.10 0.04 0 0.06 0Death of committee members 0.08 0.06 0.04 0.02 0Problems related to money 0.54 0.30 0.22 0.24 0.14No respect of the rules 0.02 0 0 0.02 0Sum 1 0.46 0.26 0.54 0.14Observations 50 23 13 27 7

Source: Data collected by the author

Table 1: Given Reasons for Group Stopping.

18

Mean Committee President DifferenceExistence of written rules 0.645 0.757 (0.039) 0.414 (0.074) 0.344 ***Compulsory attendance 0.453 0.563 (0.049) 0.228 (0.053) 0.335 ***Sharing of the leftover funds 0.582 0.783 (0.037) 0.170 (0.050) 0.613 ***Election of the ruling member 0.591 0.770 (0.047) 0.225 (0.044) 0.546 ***Member must be known before joining 0.591 0.666 (0.046) 0.438 (0.070) 0.228 ***Member must be proposed for membership 0.703 0.772 (0.040) 0.560 (0.071) 0.213 ***Member must have their spouse’s consent 0.381 0.440 (0.051) 0.261 (0.068) 0.179 **Last number of members 28.53 32.34 (2.40) 20.70 (1.46) 11.64 ***Number of members at the beginning 20.46 22.32 (1.54) 16.57 (1.29) 5.75 ***Contributions paid during the meeting 0.721 0.810 (0.041) 0.538 (0.070) 0.272 ***Paid ruling body 0.333 0.171 (0.041) 0.664 (0.062) -0.493 ***Activity of indemnity fund 0.129 0.173 (0.04) 0.039 (0.018) 0.133 ***Economical advantage 0.116 0.062 (0.020) 0.225 (0.064) -0.162 **Loss of money 0.359 0.230 (0.097) 0.591 (0.127) -0.361 **

Mean S.E. Difference S.E.President decides to accept new members (mean = 0.339)Committee group [1] 0.093 (0.033) [1] - [2] = -0.821 (0.052) ***Founder President [2] 0.914 (0.040) [1] - [3] = -0.688 (0.078) ***Non founder President [3] 0.782 (0.070) [2] - [3] = 0.132 (0.080)President decides to exclude members (mean = 0.277)Committee group [1] 0.101 (0.036) [1] - [2] = -0.666 (0.099) ***Founder President [2] 0.775 (0.092) [1] - [3] = -0.382 (0.100) ***Non founder President [3] 0.491 (0.094) [2] - [3] = 0.284 (0.131) **President decides to change the order of reception (mean = 0.177)Committee group [1] 0.048 (0.019) [1] - [2] = -0.503 (0.108) ***Founder President [2] 0.551 (0.106) [1] - [3] = -0.302 (0.088) ***Non founder President [3] 0.351 (0.086) [2] - [3] = 0.201 (0.136)Standard errors in parentheses, correction for sampling weights*** significant at 1%, ** significant at 5%, * significant at 10%Source: Data collected by the author

Table 2: Tests of difference of means over the ruling structure.

19

Founder Pdt Non founder PdtComposed of women 9.99 ***(3.33) 1.77 (1.08)Composed of women and men 12.6 ** (5.94) .573 (.767)Same ethnical group .945 (2.29) 1.39 * (.821)Existence of written rules -5.96 (3.64)-2.02*** (.739)Election of the ruling member -62.7 ***(11.3)-.633 (.88)Compulsory attendance -.116 (3.84)-1.24 * (.69)Sharing of the leftover funds -9.17 ***(2.58)-2.03*** (.669)Inquiry before accepting members 4.51 * (2.44) 1.14 (.814)Member must be known before joining 1.73 (2.02) .808 (.856)Member must be proposed for membership -.0129 (2.75)-1.65 * (.905)Member must be living in the area 1.12 (1.4) -.364 (.678)Member must have their spouse’s consent -1.71 (1.64)-.567 (.732)Paid ruling body 3.8 * (2.08) 1.71 ** (.807)ln(monthly contribution) 1.79 (2.41) .463 (.396)Last number of members -.158 ***(.055)-.004 (.022)Important sanctions -3.53 (4.45) 2.3 ** (1.13)Worse sanctions after pot reception 1.16 (4.09) -.11 (1.01)Random rosca -2.82 * (1.68) .435 (.742)

Group proportion of members in couple .555 (2.81)-.065 (.753)Group proportion of salaried members -2.64 (4.17)-.024 (.749)Group mean of members’ age .101 (.066) .029 (.039)Group mean of members’ income .026 ***(.009) -.01 (.009)Group proportion of house owners 1.13 (1.96) .403 (.689)Group mean of members’ duration in the area -.001 (.007)-.001 (.003)

Area fixed effects yesControl for group starting date yesRobust standard errors in parentheses, correction for sampling weights*** significant at 1%, ** significant at 5%, * significant at 10%Source: Data collected by the author

Table 3: Multinomial logistic regression of the roscas’ ruling structure on group characteristics.

20

Pdt alone Election Inquiry Paid Sanctions Problems RandomComposed of women 2.75 ** (1.09) .164 (1.06) -2.18 * (1.12) 1.49 (.983) -1.8 (1.41) -2.75 ***(1.06) -.061 (.811)Composed of women and men 1.29 (.921) -.352 (.815) -1.96 ** (.784) 1.09 (.801) -1.34 (.951) -.99 (.643) -.757 (.642)Same ethnical group 1.56 * (.911) -1.28 (.797) -1.15 (.721) .567 (.813) 2.38 ***(.914) .0891 (.851) .354 (.561)Run by a president alone -2.91***(.707) 1.61 ** (.77) 1.23 (.832) 1.29 (1.03) .337 (.878) 1.24 * (.719)President: founder of the group -.865 (.885) -.297 (.762) -1.98 (1.37) -1.05 (1.21) -2.02 ** (.931)Election of the ruling member -3.08***(.804) .76 (.747) -1.55 ** (.771) .41 (.654) .41 (.745) .668 (.723)Existence of written rules -1.95 ** (.854) 2.03 ***(.748) 1.34 * (.71) .005 (.572) -.728 (.835) -.994 (.631) 1.18 ** (.575)Compulsory attendance -.086 (.763) .831 (.559) -.384 (.745) -.9 * (.529) .761 (.634) -.332 (.524) .927 * (.477)Sharing of the leftover funds -2.37***(.602) -.732 (.693) 1.25 ** (.605) -2.78 ***(.648) 1.33 (1.01) 1.03 ** (.486) -.434 (.552)Inquiry before accepting members 1.34 * (.804) .663 (.785) 1.44 (.902) .842 (.862) -1.18 * (.631) -1.78***(.595)Mb must be known before joining -.291 (.848) .361 (.589) .474 (.646) -1.28 ** (.594) 1.29 ** (.594) -.922 * (.502) 1.05 ** (.521)Mb must be proposed for membership-.833 (.734) .586 (.868) 3.36 ***(.704) 1.18 (.9) .001 (1.03) -.116 (.668) .598 (.491)Mb must be living in the area -.264 (.704) .476 (.651) -.161 (.646) 1.15 * (.641) -.325 (.61) 1.16 * (.614) .551 (.503)Mb must have their spouse’s consent -.535 (.682) .975 (.648) 1.87 * (1.03) -.447 (.628) .874 (.719) -.016 (.547) .611 (.529)Important sanctions .763 (.918) 1.07 (.715) .902 (.808) 2.71 ***(.752) -.0771 (.55) .711 (.647)Worse sanctions after pot reception .105 (.753) -.772 (.778) -.743 (.707) -2.17 ***(.775) 6.2 ***(1.15) .37 (.51) -.215 (.6)Last number of members -.014 (.015) .014 (.015) .007 (.015) .001 (.01) -.006 (.007) .006 (.007) -.007 (.007)ln(monthly contribution) .585 (.382) -.551 * (.305) -.128 (.31) .512 (.314) .026 (.514) .279 (.257) -.55 ** (.261)Paid ruling body 2 ** (.799) -1.09 (.804) 2 * (1.13) 1.87 ** (.783) .073 (.592) .239 (.597)Random rosca .027 (.587) 1.24 ** (.536) -2.19***(.721) .04 (.602) 1.06 (.953) -.255 (.614)

Group proportion of mbs in couple -.037 (.615) .87 (.584) -.525 (.572) 1.4 ** (.595) .057 (.637) -.143 (.46) -.345 (.462)Group proportion of salaried mbs .206 (.893) -.956 (.835) -.737 (.765) -.622 (.794) -.0511 (.831) -.449 (.643) 1.15 (.754)Group mean of members’ age .038 (.037) .017 (.029) .029 (.023) -.034 (.034) -.02 (.032) -.087 ***(.025) .026 (.02)Group mean of members’ income .004 (.003) .0001 (.003) .005 (.006) .004 ** (.002) -.007 ***(.003) .001 (.002) .005 (.005)Group proportion of house owners .904 (.749) 1.82 ** (.715) -.191 (.748) .195 (.58) -.191 (.871) .884 (.685) -1.89***(.613)Group mean of mbs’ duration in area -.003 (.004) -.008***(.002) -.003 (.002) -.0007 (.002) .0003 (.003) .001 (.002) .001 (.002)

Area fixed effects yes yes yes yes yes yes yesControl for group starting date yes yes yes yes yes yes yesControl for starting mbs’ relationship yes yes yes yes yes yes yes

Pseudo R-squared .594 .566 .526 .507 .69 .323 .406Robust standard errors in parentheses, correction for sampling weights*** significant at 1%, ** significant at 5%, * significant at 10%Source: Data collected by the author

Table 4: Logistic regressions of group structure on groups’ characteristics

21

Multinomial logistic LogisticFounder Non Founder Election Inquiry Paid Problems Random Sim. Part.

Female -.584 (.802)-1.79 ** (.705) .607 (.699) 1.5 ** (.721) .608 (.819) -.422 (.632) -.84 (.69) -.761 (1.32)In couple -1.78 ** (.805)-2.02*** (.698) 1.37 ** (.633) .57 (.701) .0454 (.643) .524 (.511) -.322 (.618) -.66 (1.03)Female * Couple 1.42 (1.1) 2.01 ** (.784) -.916 (.805) -.405 (.88) .173 (.865) -.0717 (.736) 1.13 (.814) 2.33 (1.52)Number of dependents .071 (.104) .014 (.139) .171 (.116) .046 (.105) -.002 (.11) -.059 (.109) -.104 (.1) .221 (.143)House owner -.722 (.681) .219 (.617) 1.02 * (.528) .141 (.516) -.259 (.481) .502 (.525) -.856 * (.486) -1.26*(.702)Age -.046 (.028)-.048 (.036) .046 * (.024) .004 (.021) -.088***(.026) -.053 **(.023) .021 (.022) -.016 (.042)ln(income) .294 (.36) -.109 (.391) -.04 (.285) .95 ***(.312) .341 (.277) .132 (.269) .741 **(.319) .515 (.495)Primary degree .616 (.699)-.893 (.675) .378 (.563) .503 (.587) .282 (.571) -.831 (.511) -.336 (.525) .876 (.67)Salaried -.814 (.771)-.636 (.778) .096 (.534) .139 (.591) -.458 (.66) -.577 (.508) .983 * (.564) 1.37 *(.781)Simultaneous participation 2.62 *** (.76) 2.26 ** (.921) -1.85***(.601) -.891 (.786) 1.27 * (.652) .741 (.745) -.838 (.57)Job duration .002 (.003) .002 (.002) -.004 * (.002) .002 (.002) .006 ***(.002) .0002 (.002) .002 (.002) -.001 (.003)Duration in the area -.002 (.002)-.002 (.002) -.001 (.002) -.002 (.002) -.002 (.002) .002 (.001) .003 * (.001) .003 (.002)

Area fixed effects yes yes yes yes yes yes yesControl for ethnic affiliation yes yes yes yes yes yes yes

Pseudo R-squared .201 .236 .165 .185 .157 .156 .161Robust standard errors in parentheses,*** significant at 1%, ** significant at 5%, * significant at 10%Source: Data collected by the author

Table 5: Logistic regressions of group structure on members’ characteristics.

22

[1] [2] [3] [4] [5]Composed of women -1.63 * (.948) -1.67 (1.09) -1.63 * (.968) -1.33 (.859) -2.01 (1.31)Composed of women and men -1.35 ** (.583) -1.35 ** (.617) -1.23 ** (.596) -1.28 ** (.546) -1.74 ** (.746)Same ethnical group -.941 (1.17) -1.24 (1.13) -1.01 (1.17) -.825 (.954) -1.33 (1.39)Run by a president alone 1.93 ** (.806) 2.46 *** (.821)President: founder of the group -1.21 (.946) -1.58 * (.919)President * Paid -1.26 * (.7)Committee * Paid -2.74 *** (1.03)Committee * No paid -2.43 ** (.983)President * Election * Inquiry -2.65 *** (.981) 2.36 * (1.41)President * Election * No inquiry -1.02 (.948) 3.79 ** (1.76)President * No election * Inquiry 2.77 * (1.55)President * No election * No inquiry 5.83 *** (1.81)Committee * Election * Inquiry -4.97 *** (1.3)Committee * Election * No inquiry -3.68 *** (1.01) 1.3 * (.776)Committee * No election * Inquiry -2.99 *** (.991) 2.08 ** (.934)Committee * No election * No inquiry -1.98 * (1.12) 3.05 ** (1.36)Election of the ruling member -1.35 *** (.48) -1.73 *** (.494) -1.12 ** (.435)Existence of written rules -.184 (.7) -.0333 (.71) -.0727 (.718) -.0797 (.732) -.000121 (.803)Compulsory attendance .776 (.573) .958 * (.576) .845 (.589) .814 (.589) .799 (.745)Inquiry before accepting members -1.56 *** (.598) -1.53 ** (.634) -1.61 *** (.597)Member must be known before joining .558 (.589) .413 (.568) .436 (.56) .23 (.54) .559 (.6)Paid ruling body -.84 (.596) -.693 (.541) -.705 (.638) -.698 (.746)ln(monthly contribution) -.159 (.193) -.15 (.202) -.161 (.188) -.159 (.253) -.0505 (.255)Last number of members .00397 (.0105) .0053 (.00921).00394 (.00936).00639 (.00972) .00997 (.00999)Important sanctions -1.33 * (.679) -1.21 * (.701) -1.17 * (.635) -1.13 (.754) -1.38 ** (.657)Worse sanctions after pot reception -.683 (.565) -.52 (.561) -.748 (.548) -.676 (.614) -.347 (.649)Random rosca .969 (.593) .8 (.549) .788 (.602) 1.08 * (.61) .701 (.674)

Area fixed effects yes yes yes yes yesControl for group starting date yes yes yes yes yesControl for starting mbs’ relationship yes yes yes yes yesControl for selection effect yes yes yes yes yesNumber of dead groups 37 37 37 37 37Robust standard errors in parentheses, correction for sampling weights*** significant at 1%, ** significant at 5%, * significant at 10%Source: Data collected by the author

Table 6: Complementary log-log regressions with dummies for baseline hazard control.

23