Peer Networks and Entrepreneurship: a Pan-African RCT

Here we report the results of a large RCT conducted at the pan-African level that wants to shed light on the impact of peer effects on innovation and entrepreneurship. The experiment involved around 5000 entrepreneurs (some established, other just aspiring) from 49 African countries. All of those entrepreneurs completed an online business course, while only the treated ones had the additional possibility of interacting with peers, within groups of sixty, and in one of three different setups: (a) face-to-face, (b) virtually "within" (where interaction was conducted through an Internet platform in groups of entrepreneurs of the same country), (c) virtually "across" (where the virtually connected groups displayed a balanced heterogeneity across countries). After two and a half months, all participants were asked to submit business proposals. The ones submitted were then evaluated in a two-stage procedure. First, they were graded by a panel of African professionals; subsequently, the pool of highest-graded proposals were again assessed and graded by senior investors, who selected some for possible funding. Two outcome variables follow from this evaluation exercise: the (optional) decision of whether to submit a proposal, and the grades (1 to 5) obtained by the proposals that were submitted. Next, we outline our main results concerning the effect of the treatment on the two aforementioned outcomes - submission and quality (measured in the intensive margin) - as well as the combination of both of them that we call, for short, extensive quality. (1) Virtual-within interaction has a positive and significant treatment effect on the three dimensions: submission, intensive quality, and extensive quality. Instead, when interaction is face-to-face (thus also “within") only submission and the extensive quality margin are affected (positively so). (2)Virtual-across interaction yields no significant effect on any of the former three dimensions. (3)When effective on quality (cf. (1)), the treatment operates by shifting up, on average the evaluation grade of business proposals from low levels (grades 1 or 2) to high ones (grades 4 or 5). (4) The baseline quality of entrepreneurs has a positive effect on performance. However, the average such quality of the peers in one's own group has a negative composition effect on intensive quality. In fact, a similarly negative effect is also induced by peers' average experience level. (5) As a robustness test, the core treatment effects described in (1)-(2) are confirmed to remain essentially unchanged under a full range of control (baseline) variables, while the composition effects identified in (4) are found to survive a standard placebo test. As a second step in the analysis, we construct a social network in each group by defining the weigh of a directed link between two entrepreneurs as the amount of information (overall size of messages) written by one of them for which there is evidence that the other has been exposed to, then writing a subsequent message. Then, on the basis of the network structure so defined, we estimate the induced peer effects and arrive at the following conclusions. (6) In large countries (the only ones for which a sufficient number homogeneous groups can be formed), virtual-within interaction leads to positive and significant peer effects on submission and extensive quality, but not intensive quality. Instead, when entrepreneurs of large countries are exposed to virtual-across interaction, no significant peer effects arise in any of the three outcomes. (7) In the set of small countries, where only virtual-across interaction is possible, there are positive and significant peer effects on both extensive and intensive quality but not on submission. (8)Composition effects on network peers are weak, largely captured by (outcome-based) peer effects. (9) Results (6)-(7) are structurally robust to redefining the network links in the following two ways: (a) they are limited to involve less than a maximum communication lag, suitably parametrized; (b) they are two-sided, their weight tailored to the flow information channeled in both directions. A combined consideration of (1)-(9) reveals an interesting contrast between treatment and peer effects. For example, in view of (1)-(3), we may conclude that whereas some group homogeneity - or face-to-face contact - bring about positive treatment effects, the group heterogeneity induced by virtual-across interaction fails to deliver significant such effects on all three dimensions. Instead, (6)-(7) indicate that network-based peer effects deliver an intriguingly different pattern. For, under virtual-within interaction, we find that entrepreneurs' peers exert a significantly positive influence on submission (and the extensive margin) but not so on quality per se (in the intensive margin, while a some what polar behavior arises in small countries who undergo virtual-across interaction. This suggests that whereas homogeneity leads to peer interaction that is rather independent of peer performance, heterogeneity has peer performance play an important role (both in positive or negative terms, depending on the quality of that performance). Overall, this induces an effect of the treatment that is significantly positive under homogeneity (virtual-within interaction for large countries) but not strong enough to be significant under full-fledged heterogeneity (virtual-across interaction for small countries). The aforementioned contrast between the nature and implications of the treatment effects stated in (1)-(4) and the network peer effects in (6)-(8) is interesting and deserves further investigation. A possible explanation for it might hinge upon the positive role that homogeneity/familiarity may play as a source of encouragement (and hence participation), as opposed to the negative impact it could have in reducing the novelty of ideas and/or highlighting the fear of competition (thus dis-incentivizing information sharing and thus a genuine effect induced by peer performance). To gain a good understanding of these issues, however, one needs the help of theory as well as a detailed investigation of how communication actually unfolds in our context. Both lines of work are part of our ongoing research. Here, we provide a preliminary account of the latter, which is included in the final part of the paper. Our approach to semantic analysis relies on the machine-learning tools developed by the modern field of Natural Language Processing (NLP). This methodology is applied to the vast flow of information exchanged by entrepreneurs (over 140,000 messages) in order to identify, first, what have been the modes/categories of peer communication more prevalent in our context, e.g. business focus, sentiment/encouragement, target audience, etc. Then we use this information to understand what are the different patterns of communication most prevalent in our context, as captured by a corresponding set of conditional and unconditional distributions that show and how communication is associated to: (a)endogenous variables such as behavior or performance; (b) exogenous variables, such as treatment type or individual baseline characteristics. The main conclusions obtained so far can be summarized as follows. Messages are quite polarized in either the business or sentiment dimension, showing an inverse dependence in the (strong) FOSD sense between the respective distributions. Applying the same comparison criterion, we also find that highly performing agents use more business-focused messages, which are not only neutral in sentiment but also targeted to specific peers (rather than being general messages). Interestingly, however, the treatment arm (virtual-within or -across) has no significant effect on the type of communication, while baseline quality and a measure of †motivation" do have an effect analogous to that described before for performance. Finally, we also rely on the message categorization induced by the NLP analysis to construct semantically weighted networks on two specific features/categories: business relevance and sentiment. Quite remarkably, the corresponding peer effects are found to be unaffected by either of these “semantic projections" of the social network. This suggests that, even though entrepreneurs' messages focus heavily on business issues, their communication displays a feature that is often observed in ordinary (non-virtual) interaction: there is a balance between business focus and a comparable amount of sentiment-laden talk. Keywords: Social networks, peer effects, peer networks, entrepreneurship, semantic NLP analysis.