Punctuated decline of human cooperation

Ethical approvalThis research was approved by the Central University Research Ethics Committee (CUREC) at the University of Oxford (reference no. SSD/CUREC1A/10-099). The approval included the collection process and analysis of administrative microfinance data and primary data from human participants collected through semi-structured interviews. The research was performed in accordance with all relevant guidelines and regulations outlined by CUREC. Informed consent was obtained from all participants. The microfinance institution and its clients are de-identified as per the data use agreement.Empirical backgroundThe microfinance institution was founded in 2002 and had a client base of approximately 18,000 borrowers at the time of data collection. In Sierra Leone, 68% of the population was estimated to be living below the national poverty line62. The organization offers small loans to low-income clients with the goal of local poverty alleviation. Group lending is a standard arrangement in developing countries, where potential clients seeking a loan enter into a joint liability contract63. An organization’s motivation for this arrangement is that potential clients often lack sufficient financial collateral, but a group contract may provide a form of social collateral64. Past research has noted that this lending structure creates a natural social dilemma38,65,66. More specifically, the group lending structure can be conceptualized as a threshold social dilemma, as the collective good—access to future credit—is provided only when full repayment is reached2,11,12.Potential group lending clients are instructed to select group members that they know and trust67. Furthermore, loan officers ensure that each member meets basic eligibility criteria. Each client is required to have their own micro-business capable of meeting the minimum loan repayments.Typical micro-business examples include petty trading, food services, barbershops, tailoring and motorbike taxi services. The loan amount ranges from 200,000 to 2,250,000 SLL per member per loan (at the time of data collection, the nominal exchange rate was US $1 = 4,300 SLL). All loans have the same fixed interest rate: 2% per month. Scheduled loan duration ranges from six to twelve months. Group size in the dataset ranges from two to six members, with an average of 4.36 members. Consistent with the organization’s objective to promote financial inclusion, 73.3% of the clients are female.Eligibility for a subsequent loan cycle depends on a group’s repayment performance in the previous cycle. Full repayment is required, but timeliness of payments throughout the whole loan cycle also influences the decision. The institution incentivizes better group repayment with greater potential increases in the subsequent loan amounts. Groups that repay in full and on time receive the standard maximum loan increase for the next cycle. If the loan is paid in full, but with delayed payment(s), the group may be assigned a lesser loan increase, no loan increase or no loan renewal, on the basis of the frequency and severity of the delayed payment(s). The outcome is the same for all group members. Decisions regarding loan renewal and amount are made jointly by the group’s loan officer and loan portfolio manager. Variation in subsequent loan amounts implies that a single provision point for the collective good is a theoretical simplification of the full set of potential group outcomes. However, the core distributional conflict and individual incentive to contribute as little of one’s own resources as possible while reaching each provision point remains consistent.Group loan enforcement operates in two phases: (1) internal enforcement of active loans and (2) institutional enforcement of delinquent loans. Group members are responsible for enforcing on each other during the active phase. Members use a variety of positive and negative enforcement mechanisms, such as encouragement, social pressure, embarrassment and ostracism. Past empirical research in this context has shown that natural variation in the social and spatial structure of joint-liability groups has a critical role in the ability and willingness to apply such social sanctions38. During this phase, the collective good is still achievable; that is, access to future credit. After approximately 30 days of delayed payment, enforcement gradually shifts to the institutional phase. In this phase, the possibility of qualifying for a subsequent loan rapidly declines. The loan is classified as inactive, and the institution initiates formal recovery procedures—typically involving debt collectors and legal threats. These efforts generally continue for at least a year until the group repays or the organization officially records the loan as a write-off and ceases efforts at collection. In the ‘Modelling’ section below, we define cooperative outcome measures using a 30-day window to capture meaningful behaviour attributable to internal group dynamics, distinct from institutional debt recovery efforts.Quantitative materials and methodsQuantitative data collectionThe data were collected by the microfinance organization between 2005 and 2011. During the initial group approval process, the organization records personal, demographic and financial information for each member of the group. After loan disbursement, the organization records group-level repayment in terms of date and amount. The repayments are self-organized by the group but must be registered at the local branch office. It is not required that all members be present when making a group payment. The microfinance organization provided consent for the analysis and publication of the administrative data in accordance with the data-use agreement, indicating that the research team may disclose and/or publish data such that “the data shall be anonymized and/or aggregated so that no reference is made to individuals’ names, applying to both [microfinance institution] clients and staff.”For each scheduled repayment the organization records whether the payment was made in a single payment or in partial payments on different dates. For example, consider a group that has a scheduled monthly payment of 300,000 SLL due on 6 June 2010. If the group is unable to repay by the due date, the group might pay 100,000 SLL on 8 June 2010 and the remaining balance of 200,000 SLL on 20 June 2010.Once a loan is disbursed, group membership is fixed throughout the duration of the loan cycle. When proceeding to a new loan cycle, there may be minor changes in group membership. The most common change is the removal of a defecting member. Across all loan cycles, 81% of groups maintained the same group size, 18% reduced the group by one or more members, and fewer than 1% of groups increased group membership across cycles. Note that the potential exclusion of a group member from the next loan cycle is an internal group decision. The lending institution only approves a subsequent loan cycle at the group level, that is, it does not differentiate how much each individual contributed to repayment (refer to Extended Data Table 7 and Supplementary Table 10 for a more detailed breakdown of group size over loan cycles). The average days elapsed since repayment of the last round in the previous cycle to disbursement of the next loan is 24.7 days.The statistical analysis and figure generation were conducted using STATA v.18.5, including package Markstat v.2.1, and the software R v.4.2.3. The dataset was cleaned so that loan cycles are not truncated midway by the data window, that is, both the initial disbursement and final scheduled payment fall within September 2005 to January 2011. We include a further four-month data collection buffer (until May 2011) to allow for delayed repayments that were originally scheduled during the data window. The maximum number of cycles completed by any group at this organization was eight cycles at the time of data collection. However, the sample size decreases with each cycle, with few groups having taken more than five cycles (53 out of 1,589 groups). For this study, the dataset is limited to the first five loan cycles with a sample size useful for statistical analysis. The resulting cleaned dataset includes 7,108 borrowers involved in 31,199 scheduled group repayments. Including partial payments, the dataset consists of 47,931 repayment transactions. Extended Data Table 1 provides a summary of descriptive statistics of the quantitative dataset.Modelling approachA group’s cooperative behaviour is modelled using two measures: the financial contribution rate (FCR) and the cooperative effort rate (CER). We model the FCR as the percentage of the monthly scheduled amount paid by the group within 30 days:where \(_}\) is the observed payment amount of group \(i\) in round \(r\) of cycle \(c\), and \(_}\) is the scheduled payment amount for the associated group, round and cycle. Repayment activity may continue to be recorded by the lending institution beyond 30 days. We apply the 30-day boundary to construct a regular time interval that captures variation in a group’s internal cooperative behaviour rather than institutional efforts at debt recovery.Most measures of economic cooperation in a field setting will reflect some degree of an individual’s financial ability to contribute; however, it is important in this context to confirm that changes in repayment are not driven by changes in clients’ financial ability to repay, that is, not a systematic reduction in financial liquidity over rounds. We constructed the CER to further increase our confidence in measuring cooperative behaviour. If a group is willing to cooperate but experiencing a financial difficulty, it may make a smaller partial payment. The measure is based on the number of days overdue before the group makes its first partial or full payment for each month. The organization accepts any minimal partial payment. The data show that partial payments as small as 5% of the monthly amount due were recorded by the organization. The timeliness of the first partial payment reflects a group’s effort in signalling their repayment intentions, rather than their ability to pay the full amount. The CER is calculated as:where \(_}\) is the days overdue for the first partial payment bounded from 0 to 30 for group \(i\) in round \(r\) of cycle \(c\). The CER can be interpreted as follows: if a group makes at least some partial payment by the scheduled due date, the group effort rate equals 100; if the group’s first payment is 15 days overdue, the group effort rate equals 50; if no payment is made within 30 days, partial or full, the group effort rate equals 0. In combination, the FCR (based on amount paid) and the CER (based on timeliness) offer greater insight to a group’s cooperative behaviour.Loan durations in the dataset are 6, 8, 10 or 12 months. For plotting purposes, the rounds of loans with different durations were rescaled to the minimum duration. For example, for a 12-month loan, the first and second payments are grouped as round 1; the third and fourth payments are grouped as round 2; and so on. A consistent base makes it visually easier to compare cooperation rates for the aggregated dataset. Importantly, we do not use rescaled rounds in the regression models when statistically testing longitudinal trends.Econometric modelsTo model longitudinal cooperative behaviour, we estimate group fixed-effects regressions of the form:$$_=_+_c+_r+_c\times r+_+_^ }\gamma +_$$ (3) where \(_}\) is the cooperation rate of group \(i\) in round \(r\) of cycle \(c\), \(_\) is a group fixed effect, \(_}^ }\) is a set of controls and \(_}\) is the error term. We also run regressions that are identical to equation (3) except that they treat cycle as a fixed effect rather than a continuous variable.To estimate the size of the restart effect, we estimate group fixed-effects regressions that compare round-to-round changes in cooperation rates across all rounds of a cycle to the change in cooperation rates between the last round of a cycle and the first round of the subsequent cycle, including observations only from groups that participate in both cycles:$$_}=_+__}+_+_+_}^ }\gamma +_}$$ (4) where \(\Delta _}\) is the change in the cooperation rate from round \(r-1\) to round \(r\) for group \(\) in restart number \(t\); \(n\) is an indicator that the observation corresponds to a restart change (that is, a change from the last round of a cycle to the first round of the subsequent cycle); \(_\) is a restart number fixed effect; \(_\) is a group fixed effect; \(_}^ }\) is a set of controls; and \(_}\) is the error term. Observations corresponding to restart number 1, for example, comprise all round-to-round changes in cycle 1 plus the change from the last round of cycle 1 to the first round of cycle 2.We also fit random-effects panel regression models to examine the group factors (for example, group size, proportion female, monthly sales, business type diversity) and loan factors (for example, loan amount, loan officer, seasonal variation) correlated with cooperation rates (Supplementary Table 9). Columns ‘average cycle 1’ estimate a simple ordinary least squares regression:where \(_\) is group \(i\)’s average contribution or effort rate in cycle 1 and \(_^ }\) is a vector of group and loan characteristics of group \(i\) in cycle 1 (specifically, group size, female proportion, proportion married, average number of children, average monthly sales, the s.d. of the monthly sales, average business equity size, s.d. of business equity, index of business diversity within the group, proportion of group members engaged in petty business, an indicator that at least a member of the group is engaged in service business type, loan size, rainy season and microfinance institution branch controls). For observations with missing covariate data, values were imputed using mean substitution. The ‘RE all rounds’ columns in Supplementary Table 9 estimate the random-effects panel equation:$$_=_+_r+_+_r\times _+_+_^ }\gamma +_$$ (6) where \(_\) is the contribution or effort rate of group \(i\) in round \(r\) of cycle \(c\); \(_\) is a cycle fixed effect; \(_\) is a group random effect; and all of the other terms are as defined above.Consistently for the different specifications, greater group size is significantly associated with lower contribution and effort rates, whereas the proportion of female individuals in the group is significantly associated with higher effort rates. Perhaps counterintuitively, larger average monthly sales are significantly associated with lower contribution and effort rates. We do not interpret these correlations as causal, as it is possible that the microfinance institution uses these variables when deciding whether and what kind of loan to give to the group.Qualitative materials and methodsQualitative data overviewWe conducted 73 in-depth semi-structured interviews: 64 interviews with group lending clients and 9 interviews with members of the lending institution staff. On average, client interviews lasted 39 min and staff interviews lasted 1 h and 34 min. Interview time totalled approximately 56 h. The interviews were conducted by the principal investigator between 5 April 2011 and 6 June 2011, contemporaneous to the collection of the quantitative data.The client interviews provide detailed descriptions of the cooperative dynamics in the borrowers’ own words. The staff interviews are valuable for understanding the organization policies and their practical implementation. Collectively, the interview data perform two key functions in the study: (1) they help to contextualize the cooperative dilemma, to ensure proper understanding of the quantitative patterns and their appropriate interpretation; (2) they provide insight to the behavioural mechanisms underlying the longitudinal trends. The qualitative data may be viewed as serving both a ‘confirmatory’ function, that is, ensuring that the results are not dependent on a single type of data, as well as a ‘complementary’ function, filling the interpretative gaps inherent to the quantitative dataset68.Sample selectionThe sample of clients for interviews was drawn from the overall quantitative dataset, using a two-stage cluster random sampling, plus a purposive enhancement41. In the first stage of the random sampling, we used simple randomization of groups, after restricting the population of potential groups based on two criteria: (1) we geographically restricted the pool to groups that were administered at the lending institution’s principal branch. This was implemented for practical efficiency of interview logistics; (2) we restricted the pool to groups that had been engaged in borrowing within the last six months. This was implemented to reduce recall bias during the interviews. This resulted in 35 joint liability groups drawn by simple randomization from the subpopulation. Supplementary Table 6 provides descriptive statistics of the client interview sample at the group level.In the second stage of the random sampling, we selected one member per group to be interviewed using simple randomization within the group. We then enhanced this sampling design by implementing a purposive sampling of an additional member from within the randomly selected groups. The choice of whether to conduct another interview and with which specific member was on the basis of the content provided in the first member’s interview, following the researcher’s discretion regarding which additional group member’s perspective would provide the most valuable information. For example, if the first interviewee indicated that a specific member X had been the main source of cooperative disruption in the group, member X was selected for a direct interview to hear his or her version of the events. The intent of additional within-group interviews was to cross-validate the initial interview, collect potentially contradictory data, and understand a complex phenomenon from different points of view69. This resulted in 29 further interviews, producing a total of 64 client interviews. Extended Data Table 3 provides descriptive statistics of the client interview sample at the individual level. The two-stage cluster random sampling with the additional purposive sampling used here promotes both external validity, in the first and second stage through the random selection of a representative sample, and internal validity, through the triangulation of multiple interviewee perspectives.Interviews were also conducted with a non-random sample of nine staff members of the lending institution—including three loan officers, two information and accounting officers, two loan portfolio managers, and two executive directors—regarding organization policies, practices in the field, and the organization’s record keeping process. Supplementary Table 7 provides descriptive statistics of the staff interview sample. The interview content was instrumental in interpreting the quantitative dataset, providing context, and offering cross-validation of borrowers’ descriptions of the group dynamics from an external perspective. Furthermore, to assess the generalizability of the cooperative behaviour at this lending institution as compared to other organizations in Sierra Leone, interviews were conducted with staff at three independent lending institutions in Sierra Leone offering group loans. The interviews indicated that the organizational practices and group dynamics at the primary institution were highly similar to those at the other lending institutions.Qualitative data collectionInterviews were conducted in-person in Sierra Leone by the principal investigator. The interviews were conducted at a location preferred by each client or staff member to promote their ability to speak candidly (for example, in their home or place of work). Before each interview was conducted, the research purpose and use of the interview data was explained to the client or staff member and a physical document of informed consent was reviewed together. All participants provided informed consent by signature or thumbprint. To safeguard participant confidentiality, personally identifiable information was redacted from the interview data.All client interviews were conducted with a professional translator from Sierra Leone present. The translator was independent from the lending institution and presented no conflict of interest. Choice of interview language was at the preference of the client. The languages used by the clients during the interviews were Krio, Temne, and English. Extended Data Table 3 includes the language predominately spoken at each interview. All staff interviews were conducted in English. During all the interviews, field notes were taken by the principal investigator and audio was recorded for later transcription and analysis.Supplementary File 1 provides the semi-structured interview protocol for clients, an outline that provides a topic structure, a list of prompts and example questions to direct the conversation. The value of the semi-structured interview format in this context is the ability to adapt flexibly to the interviewee responses, to pursue interesting topics in greater depth, and uncover unexpected or disconfirming evidence70. The interview guide consisted of the following primary elements: (1) introductions and informed consent; (2) background/warmup; (3) group lending verification; (4) social connections within group; (5) client perception of the dilemma structure and responsibilities; (6) repayment process; (7) cooperative motivation and incentives; (8) free-riding and/or cooperative behaviour; (9) loan enforcement; (10) behavioural, financial and structural changes over time; and (11) closing.Accounting for positionality and mitigating biases were central to the design and conduct of the interviews70. In this context, the perceived dynamic between group lending clients and an external researcher shapes the nature of the conversation and requires the practice of reflexivity to promote impartial data. For example, to avoid misinterpretation of his role, the interviewer would clarify at the outset of each interview that he was not acting on behalf of the lending institution and that the discussion would have no effect on the client’s standing with the lending institution. To help to build participant comfort, the protocol was structured to begin with relatively easy questions before progressing to more challenging questions regarding group behaviour. If client responses appeared to be culturally embedded, the interviewer would further probe the topic to ensure that the answers would be appropriately interpreted.To reduce recall bias, interview questions focused on the current or most recent loan cycle. The protocol included probes designed to elicit concrete details and examples from the clients. For example, printed copies of client and group information, loan characteristics, and repayment data as recorded by the lending institution were brought to each interview by the researcher. The researcher familiarized himself with this information before each interview. The information could then be used during the interviews to reduce social desirability bias if a client was misrepresenting their group’s behaviour as overly positive. The information also served the benefit of verifying the accuracy of the formal repayment data and probing for any systematic inconsistencies in the repayment recording process. To further reduce social desirability bias, the sampling strategy incorporated additional member(s) of the same lending group, describing the same event but from different perspectives. We found the third-person accounts valuable as they were more open to describing socially undesirable behaviour.Qualitative data coding and analysisAudio recordings of the interviews were manually transcribed verbatim by the research team. The data were managed and coded using NVivo software (v.14.24.1). Pseudonyms are used in the direct quotes to protect client and staff confidentiality. Anonymized identifiers for joint-liability members were created at the group and client levels. The identifiers (for example, G34.C63) consist of a group number (G1–G35) and a unique client number (C1–C64). Anonymized identifiers for staff are designated as S1–S9. We randomly assigned the staff identifiers across roles to avoid compromising confidentiality.We analysed the transcribed interviews to identify themes and concepts concerning clients’ perceptions, motivations, and direct experiences of cooperation and defection in the joint liability groups71. Codes were primarily derived from existing literature, but also included ‘open codes,’ adaptable to unexcepted themes and disconfirming evidence72. Codes were organized into preliminary lower- and higher-order themes. All authors reviewed the codes and themes to assess their relevance. Revisions were made through multiple group discussions, following an iterative and reflexive process. The analytic method was abductive, recursively iterating between data and existing theory, to develop themes that coherently reflect the patterns present in the data70,71.Conceptually, when we refer to cooperative motivation and cooperative effort, we draw on a common distinction between motivation (as an internal process or psychological force that initiates and sustains goal-directed behaviour) and effort (as the observable exertion or intensification of mental or physical activity towards that goal)51,73,74,75,76. The conceptual distinction is informative, as a group member may have the motivation to cooperate (for example, strong fear of social embarrassment), but may be externally constrained and not able to exhibit observable effort (for example, she is sick and unable to contribute). In this study, we solicit cooperative motivations directly from clients in interviews and analyse their observable behaviour in the lending group as a signal of cooperative effort. We are able to analyse cooperative effort both with the interview data, using first-hand accounts of group member behaviour, as well as with the quantified measure of cooperative effort applied across the entire dataset over time.Codified themes and direct quotes are available in the Supplementary Information, organized according to the following higher-order themes: (1) cooperative dilemma structure and client understanding in the field; (2) extent and evidence of free-riding; (3) self-reported motivations for cooperation; and (4) changes in behaviour over time.Reporting summaryFurther information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

View Original Article