Survey Trust, Experimental Trust and ROSCA Membership in Rural Cameroon Alvin Etang, David Fielding and Stephen Knowles1 April 2008 1 Department of Economics, University of Otago BWPI Working Paper 44 Brooks World Poverty Institute ISBN : 978-1-906518-43-1 Creating and sharing knowledge to help end poverty www.manchester.ac.uk/bwpi Abstract We analyse new experimental and survey data from rural Cameroon, where the level of trust is much higher than is typical of Africa. We find the level of trust exhibited by individuals to be highly correlated with membership of Rotating Saving and Credit Associations. There is also a significant correlation between the degree of trust exhibited in the game and the degree of trust declared in response to survey questions. However, survey responses do not capture all of the systematic variation in experimental behaviour, and understate the importance of ROSCA membership in predicting trust. Keywords – Africa, Cameroon, social capital, trust, economic experiments, ROSCAs 1 1. Introduction Much of the existing empirical literature on social capital relies on survey-based measures of trust, and in particular on responses to the question, “Generally speaking, do you think most people can be trusted?” The fraction of the population responding in the affirmative to this generalised trust question is used to proxy a country’s overall level of social capital. Such a measure has been used to explain international differences in income levels and rates of economic growth (Knack & Keefer, 1997; Zak & Knack, 2001; La Porta et al., 1997), environmental quality (Grafton & Knowles, 2004), health and education (La Porta et al., 1997) and levels of financial development (Guiso et al., 2004). Average levels of trust are found to be lower in LDCs, and particularly low in Africa. For example, in the 1999-2002 wave of the World Values Survey (www.worldvaluessurvey.org), 35.4% of participants in North America and Western Europe responded in the affirmative to the generalized trust question; the figure for the small number of African countries surveyed was 19.3%, and for the rest of the world 26.0%. One advantage of the generalised trust question is that responses are reported for a wide variety of countries on a regular basis, facilitating international comparisons. However, it is unclear how well these responses capture an underlying propensity for trusting, co-operative behaviour. A more direct way to measure the propensity for trusting behaviour is through controlled experiments such as the Trust Game. Following Glaeser et al. (2000), several studies have examined the relationship between responses to trust questions and behaviour in the Trust Game. We have conducted a Trust Game experiment and a trust survey in a village in rural Cameroon. As we will see, overall levels of trust – both experimental and survey trust – are higher in our village than has typically been found elsewhere in Africa. (Ours is the first study of trust in Cameroon, and the first study of experimental trust in West Africa.) One salient feature of our village, and of many other parts of rural Cameroon, is the high level of participation in Rotating Saving and Credit Associations (ROSCAs). Members of a ROSCA each contribute an agreed sum to a common fund at regular intervals, the fund being allocated to each member in turn, in order to facilitate lumpy expenditures. The successful operation of a ROSCA requires that those who are allocated the fund early in the round continue to make contributions later on. If those considering forming a ROSCA do not have much trust in other potential members, then ROSCAs are less likely to be formed in the first place and are more likely to fail. 2 In this paper we investigate the relationship between experimental trust, survey trust and ROSCA membership, conditional on other social and economic characteristics. We show that while levels of survey trust and experimental trust are positively correlated, the distribution of experimental trust (both the unconditional distribution and the distribution conditional on survey trust) is a function of ROSCA membership. However, not all of the survey questions reveal a significantly higher level of trust among ROSCA members, and in this sense the survey understates the extent to which ROSCA membership is associated with more trusting behaviour. 2. Literature Review (a) Experimental trust In the Trust Game (Berg et al., 1995), participants are divided into two groups: Senders and Recipients. Each Sender is paired with a Recipient, sometimes anonymously. A Sender is given a sum of money and must decide how much of this money, if any, to transfer to the Recipient. The amount of money transferred is tripled by the experimenter. The Recipient must then decide how much of the money, if any, to return to the Sender. The amount of money sent by Sender is interpreted as a measure of her trust in the Recipient (who may be anonymous), while the amount returned by the Recipient is interpreted as a measure of trustworthiness or reciprocity. To what extent does the Trust Game measure trust and trustworthiness? Ashraf et al. (2006, p.193) define trust as a situation in which “one party, the trustor, makes himself vulnerable by taking a course of action that creates incentives for the other party, the trustee, to exploit him.” Camerer (2003, p.85) argues that “[t]rust must be risky. Trustworthiness must also go against the Trustee’s self-interest, to test whether people are willing to sacrifice moral obligation.” A decision by the Sender to transfer money gives the Recipient an incentive to exploit her by keeping it all. Returning any money is contrary to the Recipient’s self-interest. Therefore, the Trust Game does measure both trust and trustworthiness. Note also that the Trust Game is typically played as a one-shot game, not a repeated game. In a repeated game reputations can be built up, so the Sender is not necessarily making herself vulnerable to exploitation, and a repeated game would not measure trust so well (Barr, 2003, pp.617-8). Table 1 summarises Trust Game results from LDCs, along with results from two seminal US studies, Berg et al. (1995) and Glaeser et al. (2000); this summary updates Cardenas & Carpenter (2008). The table shows the mean proportion of money sent and returned in each game, indicating whether the experiment was carried out on students or non-students and 3 whether the participants were anonymous to their partners. It also notes cases in which the authors tested the significance of the correlation between experimental and survey trust or between experimental trust and group membership. There is substantial variation in the mean proportion sent (30-83%) and returned (18-82%); the reasons for such variation are not well understood. However, the overall average proportion sent in the African studies in the table is 45%, compared with 57% in the other LDCs, which is consistent with Africa’s relatively low levels of survey trust. 4 Table 1. Trust Game studies in developing countries and in the USA Author (s) Country Survey Students Proportion Proportion Anonymous Correlation with Survey Questions Correlation with Group Sent (%) Returned Memberships (%) Berg et al. USA No Yes 52 46 Yes Not Examined Not Examined (1995) Ashraf et al. Russia Yes Yes 49 29 Yes The question about trust in strangers was Memberships of any (2006) South Yes Yes 43 27 Yes positively correlated with the proportion returned organizations were Africa but not correlated with the proportion sent. insignificant in explaining either the proportion sent or the proportion returned. Barr (2003) Zimbabwe No No 43 43 Yes Not Examined Not Examined Bouma et al. India Yes No 49 33 Yes Not Examined Not Examined (2005) Buchan & China Yes Yes 66 35 Yes Not Examined Not Examined Croson (2004) Buchan et al. China Yes Yes 73 50 Yes Not Examined Not Examined (2006) South Yes Yes 64 49 Yes Korea Burns (2004) South Yes Yes 33 23 No Not Examined Not Examined Africa Carter & Castillo South No No 53 38 Yes Not Examined Memberships of any social (2003) Africa groups were insignificant in explaining either the proportion sent or the proportion returned. Castillo & Carter Honduras No No 49 42 Yes Not Examined Not Examined (2004) Danielson & Tanzania Yes No 56 46 Yes Questions regarding attitudes to trust (including Not Examined Holm (2007) generalized trust) and past trusting behaviour (as measured by an index of three survey questions) do not predict the amount sent. Generalized trust and self-reported trustworthiness fail to predict the proportion returned. The trust index is significantly positively correlated with the proportion returned. Ensminger Kenya No No 44 18 Yes Not Examined Not Examined (2000) 5 Table 1 (continued) Glaeser et al. USA Yes Yes` 83 45 No Questions about past trusting behaviour were Not Examined (2000) positively correlated with the proportion sent. Many attitudinal trust questions (including the generalized trust question) were positively correlated with the proportion returned. Greig & Bohnet Kenya Yes No 30 82 Yes Not Examined Not Examined (2005) Haile et al. South Yes Yes 55 28 Yes Positively correlated with the proportion returned Not Examined (2004) Africa 45 29 No but not correlated with the proportion sent. Holm & Tanzania Yes Yes 53 37 Yes Attitudinal trust questions were insignificant in Not Examined Danielson (2005) explaining either the proportion sent or the proportion returned. The question about past trusting behaviour was found to be negatively correlated with the proportion returned. Johansson- Bangladesh Yes No 46 46 Yes Generalized trust was positively correlated with Memberships of any Stenman et al. the proportion returned but uncorrelated with the voluntary groups were (2006) proportion sent. Questions about past trusting insignificant in explaining the behaviour and past experience (victim of crime) proportion sent and the were insignificant in explaining either the proportion returned. proportion sent or the proportion returned. Karlan (2005) Peru Yes No 46 43 No Survey trust questions were positively correlated Whether the pair comes with the proportion returned but not with the from the same lending group proportion sent. had an insignificant effect on either the proportion sent or the proportion returned. If the pair attends the same church was positively correlated with the proportion sent but not correlated with the proportion returned. Lazzarini et al. Brazil Yes Yes 56 34 Yes Trust questions (including WVS) were positively Not Examined (2004) 86 49 No correlated with both the proportion sent and the proportion returned (anonymous sample). Mosley & Uganda Yes No 49 33 Yes Not Examined Memberships of groups Verschoor (2005) within the village (index of bonding social capital) were significantly positively correlated with both the 6 proportion sent and the proportion returned. Schechter (2007) Paraguay Yes No 47 44 Yes Not Examined Not Examined 7 (b) Survey trust Questions about trust normally take two forms: questions about attitude (asking whether people in general or certain groups of people can be trusted) and questions about past trusting behaviour (asking, for example, whether subjects have ever lent money to others, or intentionally left their doors unlocked). Some authors, for example Glaeser et al. (2000), who sample Harvard economics students, find that the proportion sent is significantly correlated with responses to behavioural questions but not with responses to attitudinal questions. Studies investigating such a correlation among developing country university students include Holm & Danielson (2005, Tanzania), Lazzarini et al. (2004, Brazil), Ashraf et al. (2006, South Africa) and Haile et al. (2004, South Africa). Only Lazzarini et al. find a positive correlation between the proportion sent and generalised trust, and then only when the players are anonymous to each other. University students may not be representative of the wider population. However, results from the few studies using non-student samples in developing countries are similarly mixed. Johansson- Stenman et al. (2006) find a positive correlation of both the amount sent and the amount returned with responses to the generalised trust question among Bangladeshi household heads, but not with responses to behavioural trust questions. Karlan (2005), using subjects from a group-based lending scheme in Peru, finds no correlation of survey-based trust with the proportion sent, although there is a significant positive correlation with the proportion returned. Danielson & Holm (2007) find no correlation between the amount sent and responses to the generalised trust question, or other questions about attitudinal trust, for a sample drawn from a Tanzanian church congregation. However, they do find a positive correlation between an index of survey trust questions and the proportion returned. One possible reason for the absence of a correlation between the generalised trust question and the amount sent is that many surveys focus on questions about “most people”, whereas the Trust Game is typically played between a much narrower, homogeneous group of people, such as household heads in a particular village. A few studies have examined whether there is a correlation between membership of some kind of co-operative group and the amount sent or returned in the Trust Game. Ashraf et al. (2006), Carter & Castillo (2003) and Johansson-Stenman et al. (2006) find no evidence of any such correlation. However, Mosley and Verschoor (2005) do find a positive correlation among group members living in the same village. Only Karlan (2003) asks whether membership of specific 8 types of group affects experimental trust. Here, membership of the same church is correlated with experimental trust, but membership of the same group credit scheme is not. Our paper adds to the literature by presenting survey and experimental results from Cameroonian villagers among whom there is a substantial amount of economic and demographic heterogeneity, some dimensions of which impact on experimental trust. Unlike most previous studies, our investigation into the relationship between experimental and survey trust pays attention to the radius of trust, with questions about trust in people with whom respondents interact regularly (for example, people in the same ROSCA) as well as questions about trust in people in general. Also, we are the first to explore whether ROSCA members are more trusting or trustworthy than non-members, and whether the amount of trust or trustworthiness increases with the duration of ROSCA membership. 3. Experimental Design (a) The village The Trust Game and the questionnaire were administered in a village in the South West Province of Cameroon. The village has a population of approximately 1000 inhabitants, and is ethnically homogenous: everyone belongs to the same ethnic group and speaks the same dialect. The main economic activity is agriculture; coffee and cocoa are the most important crops. Most people are illiterate. The closest neighbouring villages are about 5km away, and the nearest large town 40km away. The village has no post office or bank, and a high proportion of villagers belong to a ROSCA. The field work was carried out in January 2007 by one of the authors and three research assistants, all fluent in the local dialect but having no previous contact with the village. 17 ROSCAs operate in the village with a total of 426 members, representing more than half of the adult population. Around 60% of ROSCA members are female. Groups may consist of men only, women only or both men and women. ROSCA size ranges from 11 to 45 members, and the average age of a ROSCA is 8 years. The median contribution to a ROSCA is 1000 CFA francs per meeting, with meetings typically held every month or every two weeks. In addition to the rotating fund, some ROSCAs also run a savings and loan fund or an insurance fund as an optional extra. ROSCA membership in the village has been increasing slowly over time. ROSCA transactions do not require any written contract: all agreements are monitored and enforced 9