Estimation issues, data, and estimation strategy

Our data are such that we have two-stage sampling where groups were sampled first and then group members were sampled in the second stage. As groups are small, we must consider that data from group members are not independent and standard errors should be corrected for clustering at group level when analyzing individual-level data.

To a large extent, there was also self-selection of members into groups, and this could contribute to stronger ingroup social relations than outgroup social relations. We are to a limited degree able to separate this selection effect from the ingroup social relation formation effects after group formation. Many of the group members knew each other before they formed the group and they typically came from the same neighborhood (got) within the larger village (kushet) and municipality (tabia). The other selection criteria relate to the eligibility for joining a youth business group, which is related to being landless or very land-poor and being a resident of the tabia as well as aiming to establish a livelihood in the community and thereby demanding to join such a group. After joining, there is attrition that varies across groups as an additional selection mechanism, which could be influenced by many individual, group, community, and exogenous factors. We lack detailed data on dropped out members and cannot, therefore, assess the effect of this attrition.

As a potential indication of endogeneity of social preferences, we regressed the ingroup and outgroup preference variables on observable individual and parent characteristics while controlling for group fixed effects, see Table B in S2 Appendix. These characteristics explained very little of the variation in the data and few of the variables were significant even with such a big sample. This may indicate that most of the variation in the social preference variables can be regarded as exogenous. We therefore proceed based on that assumption.

Our empirical strategy is to assess how generalized trust and trustworthiness are related to basic social preferences and norms of reciprocity based on experimental measures of these where the youth group members played the games with unknown youth in other groups in their district. We assume that individual ingroup trust and trustworthiness also depend on these individual outgroup characteristics, complicating the analysis of ingroup trust and trustworthiness. We first do simple correlation analysis for the key variables and stepwise add variables as we move from one dependent variable to the next in the (recursive) conceptual model. We assess whether and to what extent adding variables increases the part of the variance that can be “explained” and how the within-group and between-group variance is affected by the RHS variables in each model. We also assess the stability and statistical significance of the coefficients to get a first impression of their direct and potential indirect effects through added endogenous variables.

To deal with endogeneity, we run structural equation models in a recursive system based on the conceptual model. The identification strategy is as follows. We assume social norms are influenced at the community level and therefore use community (tabia) fixed effects in the ordered probit model for the social norm to reciprocate which has three levels (strong, weak and no obligation to reciprocate). The next level is outgroup trustworthiness, which was elicited with the strategy method by our experimental enumerators. This may have resulted in some enumerator bias in the data and we use enumerator dummies as instruments for identification. We had 12 enumerators to interview one group member each in each youth group. This was done both to ensure no communication among group members during experiments and interviews and to avoid spurious correlations within groups due to enumerator bias. For outgroup trust, we included additional economic preferences in the form of risk tolerance and outgroup expected returns in the game (a categorical variable). Trusting people is risky and more risk tolerant people are therefore expected to invest more, ceteris paribus, but more optimistic subjects with higher expected returns would also invest more.

For ingroup trustworthiness, we assume it is a function of outgroup trust and trustworthiness and use predicted values of these. In addition, we assume that ingroup social preferences and the ingroup norm of reciprocity affect ingroup trustworthiness. Ingroup norm of reciprocity is modeled on the ingroup social preferences with an ordered probit model, like the case of outgroup social norm was modeled on the outgroup social preferences. Ingroup trustworthiness is modeled on ingroup social preferences and the predicted ingroup norm of reciprocity. Finally, ingroup trust is modeled on the predicted ingroup trustworthiness, predicted ingroup social norm, predicted outgroup trustworthiness, predicted outgroup trust, ingroup social preferences, risk tolerance, and ingroup expected returns in the trust game.

To get further insights about within-group and between group variation, we ran single-equation models with random group effects that did not control for endogeneity or error correlations. These model results are included in S2 Appendix. We found substantial heterogeneity across groups making it interesting to study this further. We constructed the average group-level variables to dig deeper into the assessment of group effects.

The main advantage of this is that we can assess the group composition effects for social preferences. We, therefore, run models where the shares of altruistic, egalitarian, spiteful and selfish group members are included as additional variables that may influence individual norms, trustworthiness, and trust in the outgroup and ingroup contexts.

We included the shares of each social preference type in an alternative structural equation model. This implies a re-specification of model Eq 1 as follows:

and, likewise for Eqs 2–6. We hypothesize that these shares have a separate group effect on the dependent variables beyond the individual direct effects on their own norm, trustworthiness, and trust. We can then assess whether the outgroup and ingroup variation in norm, trustworthiness, and trust only are affected by the individual level social preferences or whether there is an additional effect of the frequency of or distribution of these norms in their group. We hypothesize that there are such group effects in the ingroup context but not in the outgroup context. More specifically, in groups with more altruists (egalitarians), we hypothesize that this has an additional positive effect on the ingroup norm to reciprocate, trustworthiness and trust. Likewise, we hypothesize that ingroups with a larger share of spiteful and selfish ingroup members demonstrate a significantly lower level of the average ingroup norm to reciprocate, trustworthiness and trust. We have provided the datasets and a do file in STATA file formats as supporting information (see S1 and S2 Datasets, and S1 File). We acknowledge that there are alternative ways to estimate the system of equations and encourage others to inspect and scrutinize the robustness of our findings.