2. Theoretical framework

We consider endogenous leadership selection in local community organizations in which leaders' ability, but not their effort, is observable. Our theoretical framework is based on Grossman and Hanlon (Reference Grossman and Hanlon2014), but has two new features. First, we allow the level of prosocial motivation of the leader to vary. Second, reflecting the nature of our case study (elaborated below) in which natural resource abundance matters, we assume that the returns to providing effort by the leader varies with the abundance of the local resource (the standing stock, R).

The augmented Grossman-Hanlon model is as follows. Consider a group of N members who differ in terms of their innate ability, ${A_i}$, and can be ranked from low- to high-ability individuals: ${A_i} \in ({0,\hat{A}} )$. Ability levels are public knowledge, or there are credible signals that all members can ‘read’ and that can be used to learn something about unobservable dimensions of quality (in the sense of Spence (Reference Spence1973)). Following Grossman and Hanlon (Reference Grossman and Hanlon2014), we assume that one group member will be selected as the leader, and is henceforth responsible for producing a public good valued equally by all group members (but who does not receive a salary – leaders are ‘volunteers’). Each group member has one unit of time. Ordinary group members allocate all their time to an income-earning activity, but leaders divide their time between the income-earning activity and public good provision.Footnote ⁴ The productivity in both activities depends on the leader's ability and effort. In the context of our case study, below, public good provision amounts to (overseeing the) management of a communal resource, which provides a flow of benefits for all group members.

Specifically, the value of public good P produced by leaders depends on their ability $({A_l})$, the share of their effort (time) devoted to public good production $({e_l})$, the resource stock (R), which is assumed to be exogeneous, and a random noise term with mean value of zero $({\eta _p})$.Footnote ⁵ Assume complementarity between effort and ability in producing public goods $(\partial {P_l}/\partial {A_l}\partial {e_l} > 0)$, and between leader's effort and the forest stock, as well as between the leader's ability and the forest stock in producing the public good $(\partial {P_l}/\partial {A_l}\partial R > 0\ \textrm{and}\ \partial {P_l}/\partial {e_l}\partial R > 0)$. The value of the local public good is given by $P = p({{A_l},{e_l},R, {\eta_p}} ).$Footnote ⁶ The leader's private income depends on ability $({A_l})$, the share of effort devoted to income generation $(1 - {e_l})$, as well as a random noise term $({\eta _I})$, and is given by $I({{A_l},1 - {e_l},{\eta_p}} )$. We again assume complementarity in production $(\partial {I_l}/\partial {A_l}\partial {e_l} > 0)$. Ordinary group member j earns a private income $I({{A_j},1,{\eta_I}} )$, where ${\eta _I}$ is a random noise term with mean value of zero. Assume the Inada conditions hold.

The leader also faces additional costs (or rewards), defined as $C({m, {e_l}} )$. These costs depend on the level of monitoring by others, m, and effort. Monitoring enters the problem because effort levels, unlike ability, are not perfectly observed and cannot be inferred from P (because of the random error term η_p). Negative costs imply rewards, which may reflect (effort-dependent) compensation by group members, and graft or corruption. In our context, diversion of group resources is relevant, resulting in a flow of private benefits. The scope for engaging in corrupt behavior is a decreasing function of monitoring. Monitoring reduces opportunities for graft and corruption, and can also be costly for the leader because of direct costs associated with complying with certain behavioral imperatives or rules.Footnote ⁷ Assume the cost (reward) function is twice differentiable, increasing and convex in monitoring intensity m, and decreasing in effort allocated to public good provision e (i.e., leaders receive a lower penalty, or greater reward, if they work harder). We also assume that the marginal benefit of supplying effort is increasing in monitoring intensity: ${\partial ^2}C/\partial m\partial e < 0.$

In addition to behavior-specific costs and rewards there is a fixed candidacy cost $(\phi )$ associated with running for the leadership. This may include opportunity cost of time or social obligations that materialize as a consequence of efforts to mobilize sufficient support. In line with recent literature (e.g., Fedele and Naticchioni, Reference Fedele and Naticchioni2016), payoffs of leaders depend on their prosocial motivation $\beta$. Leaders who intrinsically care about the well-being of their group members derive additional utility from supplying the public good as this makes others better off. Summarizing, the leader's payoff $({Y_l})$ can be written as follows:

(1)\begin{equation}{Y_l} = \alpha I({{A_l}, 1 - {e_l},{\eta_I}} )+ ({1 - \alpha } )({1 + \beta } )P({{A_l},{e_l},R, {\eta_p}} )- C({m, {e_l}} )- \phi, \end{equation}

where $\alpha$ is a group-level indicator of the availability of private income opportunities relative to the value of the public good. We assume all group members face the same market conditions for their non-forest activities, but payoffs per unit of time vary with individual ability.

The ‘voters’ side of the model is simple, even if the voter label is a bit of a misnomer in the context we study (see section 3). We abstract from social capital and prosocial preferences among voters and assume away the second-order social dilemma of who will invest in monitoring. We also assume a simple leader selection process, where voters support the leader who delivers the highest expected income.

Timing of the model is as follows. Group members first decide whether they want to be candidates. Then there is an election and, finally, the elected leader decides how much effort to exert. The model can be solved using backward induction: (i) potential leaders determine their optimal effort level $e_i^\ast$; (ii) group members decide whom to select; and (iii) the expected outcome of the election determines who runs for office (candidates compare the payoff from running and not running). See the online appendix for details. This simple model produces the following testable predictions:

Increasing monitoring intensity reduces returns of leaders, but the impact of monitoring on the quality of group leaders varies with the opportunity cost of their time. In the presence of outside income opportunities, the subsample of candidates for whom ‘running for office’ is privately optimal is reduced. In particular, high-ability individuals with high opportunity costs will opt out of the race. The extensive margin shifts to an individual with lower ability and lower opportunity costs. However, this effect is attenuated by prosocial motivation. Prosocial individuals are less likely to be motivated by private profits when taking on a leadership role, and are less likely to display corrupt behavior.

Leaders have to determine how to divide their time between private activities and effort for the public good. As outside opportunities become more attractive, leaders re-allocate part of their time to private activities – raising the marginal benefit of time allocated to public good provision until the marginal benefit across activities is equalized. Since the benefits of increasing effort are greater for greater levels of monitoring, ${\partial ^2}C/\partial m\partial e < 0$, it follows directly that more monitoring will raise effort.Footnote ⁸ A stronger prosocial motivation increases effort levels but dampens the accentuating effect of intensified monitoring. Increasing the leader's valuation of the public good implies a re-allocation of labor towards the public good, raising the marginal value product of effort allocated to ‘other activities’.

A greater resource stock raises the marginal value product of effort allocated to public good provision $({\partial ^2}P/\partial e\partial R > 0)$. This generates greater benefits for all group members, including the leader. Running for the leadership position will become optimal for a larger set of group members – especially more able ones as ${\partial ^2}P/\partial A\partial R > 0$.

3. Background and context

Ethiopian forests suffer from widespread degradation and deforestation – forest cover declined from 40 per cent at the end of the 19th century to 4 per cent in recent years (Dessie and Christiansson, Reference Dessie and Christiansson2008).Footnote ⁹ In response to this trend, the Ethiopian government initiated a large Participatory Forest Management (PFM) program – decentralizing forest management and giving local communities use rights over demarcated forest blocks. One of the earliest such programs was piloted in the mid-1990s in the Adaba and Dodola districts of West Arsi zone, Oromia regional state. In collaboration with the German Society for International Cooperation, so-called forest user groups or FUGs were established and given exclusive use rights.

FUGs have a maximum group size of 30 individuals, so villages have multiple FUGs (minimum 4 and maximum 20 FUGs per village). With the help of experts from Oromia Forest and Wildlife Enterprise (OFWE), FUGs drafted their own bylaws. This included details regarding the organizational structure, specification of the rights and responsibilities of group members and leaders, and establishment of specific committees (including the monitoring committee featuring in our empirical analysis). While OFWE provided sample bylaws, groups were free to specify their own rules. A ‘typical FUG,’ consists of a General Assembly (all members), an executive committee of five members (chairperson, vice chairperson, secretary, cashier, and one member), and perhaps some technical committees (forest development committee, forest rent determining committee, forest products marketing committeeFootnote ¹⁰).

The sample bylaws also propose that FUGs have a monitoring committee, and about 30 per cent of the FUGs in our sample indeed have such a committee – introduced when the group was established (see below). Monitoring committees are charged with the responsibility of monitoring the behavior of the group's chair (or leader). We will exploit the presence or absence of such monitoring committees in our analysis of observational data, below.

After drafting the group bylaws, leaders and committee members are selected by the group members in a participatory process. The details of this process deviate slightly from group to group, but the main elements are consistent across groups and reflect existing (informal) institutions. The selection process is not based on anonymous voting. Instead, as written by Deserranno et al. (Reference Deserranno, Stryjan and Sulaiman2019: 240) who study leader selection in microfinance service groups in rural Uganda, ‘groups are embedded in local power structures that can enter their governance and create a bias in favor of influential community members’. During a meeting of the General Assembly, members engage in a public discussion about who are suitable candidates and whom should be selected for the various leadership positions. Any member may volunteer to take on such responsibilities, but not all members have an equal chance of being supported by their peers. Deserranno et al. (Reference Deserranno, Stryjan and Sulaiman2019: 241) write ‘when decision-making is public, less powerful members can be coerced or intimidated into supporting certain [candidates], tilting selection decisions towards outcomes that are less representative of the low-income group members’ preferences’. This is consistent with our understanding of how group leaders are selected in Ethiopia – if local elites believe that assuming a leadership position is privately profitable, they may ‘muscle their way’ into such positions (see also Gugerty and Kremer, Reference Gugerty and Kremer2008). Kahsay and Bulte (Reference Kahsay and Bulte2021) document that leaders ‘have higher income and consumption levels, are more educated and wealthy, and often have a role as religious or clan leader’.

OFWE was concerned about elite capture and corruption in our study region. Kahsay and Bulte (Reference Kahsay and Bulte2021) study elite capture and corruption by FUG leaders and discuss several approaches through which elites can ‘grab’ more than their fair share of forest rents: ‘In [some] cases, the poor lack complementary assets to benefit fully from forest resources. For example, forage from forests is especially valuable for livestock-owning households – the elites. If forests are managed to maximize the flow of forage, poor community members may benefit little. There is also anecdotal evidence on cases of blatant abuse, where proceeds from the collective sale of forest products are appropriated or divided based on family and clan ties as well as bribes from (non) members to facilitate illegal extraction of forest resources.’ (p. 114).

The group bylaws stipulate the role and responsibilities of group leaders. These responsibilities are varied, and include patrolling the forest and organizing and overseeing the patrolling of fellow group members. Illegal extraction of forest resources should be minimized, and individuals engaged in illegal extraction – whether they are group members or others – should be punished. In addition to patrolling, leaders play a role in legal extraction, perhaps negotiating sales with outside buyers of tradable forest resources (e.g., timber), and liaise with the government when necessary. The bulk of the time that group leaders allocate to group-based activities, however, goes to patrolling. As shown below, on average leaders spend some 40 days per year patrolling the forest. While being educated and having business experience are relevant characteristics for leaders, it is imperative that leaders are willing and able to allocate enough time (effort) to this key component of their job description. As mentioned, leaders do not receive financial compensation for their time. Instead, they may expect indirect financial and non-financial benefits, including benefits from the public good and legal forest harvesting; participation in workshops and trainings, which are sometimes organized by OFWE and involve per diems; free labor assistance from group members during agricultural activities; and non-pecuniary benefits in the form of status. In addition, there may be opportunities for covert diversion of group income or side-selling of forest assets.

This setting is ideal for testing the predictions of the augmented Grossman and Hanlon model. Leaders of the FUG are voluntary group members who are ‘selected’ by their peers, who know them well as they live in close proximity (e.g., Hussam et al., Reference Hussam, Rigol and Roth2022). It is reasonable to assume that leader ability is known, or at least signals of ability. Ability is relevant as leaders have to organize forest management, keep records, and negotiate with traders, the government and NGOs. On the other hand, leader effort is also important and likely imperfectly observable – it is at least partly private information. Group size is fixed, and there is no entry in response to rent grabbing opportunities.Footnote ¹¹ Leadership is a part-time activity and leaders must decide how to allocate their time between public good production and income generation.

In our setting, the hypothesized link between monitoring and leadership quality is as follows. We hypothesize that more intensive monitoring will reduce the scope for engaging in ‘cases of blatant abuse’ – reducing the private profitability of taking on a leadership role for individuals who are prone to corrupt behavior (individuals with weak prosocial preferences). Some individuals therefore opt out of the race for leadership, and the quality of the remaining top candidate (weakly) goes down. This will be particularly true if relatively well-educated local elites opt out. Incentives to opt out will be strongest in settings where leaders have alternative income opportunities.

4. Experiment and data

This paper is based on observational data, and data from an RCT. We have access to two waves of data collection in Adaba and Dodola districts, involving all 132 FUGs from the 13 villages (kebelles) in the study area. Baseline household data and OFWE registration data were collected in the Spring of 2017 (see Kahsay and Bulte, Reference Kahsay and Bulte2019). We explore the correlation between the presence of monitoring committees and leader ability and effort using these baseline data (cross-section analysis). A randomized monitoring intervention started in May 2017, introduced below, and end-line data were collected in autumn of 2018 (see Kahsay and Bulte, Reference Kahsay and Bulte2021). We leverage exogenous variation in monitoring induced by the experiment, and estimate a diff-in-diff model combining baseline and end-line data. The experiment was based on informed consent of all FUG members and there was no attrition or non-compliance.

The RCT approach is superior from an identification perspective (due to exogenous variation in monitoring) but our experimental results are driven by a relatively small number of cases where leaders were replaced by new ones during the study period of 16 months due to the monitoring (leader turnover during the intervention increased from less than 3 per cent for the control group to more than 17 per cent for the treatment group). Instead, the results of the observational analysis are based on a cross-section of the RCT panel – using all 132 leaders in our sample. We, therefore, believe there is merit in presenting both the observational and RCT results. Together, these complementary approaches provide a rich picture of how monitoring affects leader ability and effort.

5. Econometric method and identification strategy

To identify the effect of monitoring on leaders' ability and effort, we estimate OLS and diff-in-diff models. Consider the former first. Using the baseline data, we regress measures of ability and effort of the leader of group i, Y_i, on a dummy variable indicating the presence of a monitoring committee, Monitoring _i, a measure of Outside Income Opportunities (OIO _i), the interaction between these terms, and additional controls, ${X_i}$:Footnote ¹⁸

(2)\begin{equation}{Y_i} = {\beta _0} + {\beta _1}Monitorin{g_i} + {\beta _2}OI{O_i} + {\beta _3}Monitorin{g_i} \times OI{O_i} + {\beta _4}{X_i} + {\alpha _z} + {\varepsilon _i}.\end{equation}

In (2), α _z captures kebelle (village) fixed effects to capture time invariant factors like geophysical conditions and local governance (one kebelle hosts multiple FUGs), and ε is a random error term. We are mainly interested in coefficient ${\beta _3}$ that captures the conditional association between ability or quality and the presence of a monitoring committee (when outside income opportunities are available). In additional models we include (interaction terms involving) our measure of the value of the baseline forest stock. This enables exploration of whether high-ability leaders are attracted by the prospect of greater resource wealth.

To interpret the association captured by ${\beta _3}$ as the ‘causal effect’ of monitoring institutions (in the presence of outside income opportunities) on leader ability or effort requires additional assumptions. Specifically, we need to assume that the presence of the monitoring committee is an exogenous variable in (2). However, it is possible that monitoring institutions are established in response to (expectations about) weak leadership, or that certain factors for which we cannot control drive both leadership and monitoring. In other words, the presence of monitoring institutions may be endogenous in models explaining dimensions of leadership.

This concern is mitigated, but not eliminated, by the following considerations. First, we have access to a rich set of controls, attenuating concerns of omitted variables. Second, as mentioned, much of the variation at the group level is caused by expectations and beliefs of the implementing agent of OFWE. Local OFWE officers worked with villagers to modify the sample bylaws in accordance with local needs. While this process could be informed by local conditions, our data suggest this is not the case.Footnote ¹⁹ Third, our institutional measures are quite persistent. Personal communication with villagers and OFWE representatives revealed that groups that have a monitoring committee established the committees at the beginning of the FUG-creating process – when local bylaws were drafted. The election of the first cohort of leaders happened afterwards, and since these days nearly all groups have changed leadership (typically more than once) (Kahsay and Medhin, Reference Kahsay and Medhin2020). As a robustness analysis, we consider the subsample of FUGs with their leaders elected during the past six years – with the institutional setting firmly in place. Finally, we examine whether our results are sensitive to omitted variable bias by the coefficient stability approach (Altonji et al., Reference Altonji, Elder and Taber2005; González and Miguel, Reference González and Miguel2015; Oster, Reference Oster2019).

Notwithstanding these considerations, we understand that not all concerns about endogeneity of the monitoring variable can be eliminated with our cross section data. We therefore also combine our baseline and end-line measures of ability and effort, and the random assignment to the intensified monitoring treatment, and estimate the following ability model:

(3)\begin{align}\begin{aligned} {Y_{it}} & ={\beta _0} + {\beta _1}{T_{\boldsymbol{i}}} + {\beta _2}OI{O_i} + {\beta _3}Year + {\beta _{4j}}OI{O_i} \times {T_i} + {\beta _5}OI{O_i} \times Year\\ & \quad + {\beta _6}{T_i} \times Year + {\beta _7}{T_i} \times OI{O_i} \times Year + {\beta _8}{X_{it}} + {\alpha _v} + {\varepsilon _{it}}. \end{aligned}. \end{align}

In (3), ${T_i}$ is the treatment dummy, $OI{O_i}$ is the measure of outside income opportunity, and $Year$ is a year dummy (1 if end-line period, 0 otherwise). Our coefficient of interest is ${\beta _7}$, associated with the triple interaction of outside income opportunities, the treatment, and the end-line dummy. In the diff-in-diff model, the monitoring committee variable measured at baseline enters as one of the controls. For the effort model we estimate a simple diff-in-diff model in which we interact the treatment and end-line dummies (and include outside income opportunities as a control variable).

Finally, to probe the net welfare effects of monitoring institutions we estimate a model explaining variation in average forest-based income, F, across FUGs. Since monitoring is expected to increase effort but lower ability, the net effect on welfare is ambiguous. We estimate a simple OLS model with a dummy variable for the presence (or absence) of a monitoring committee as the main explanatory variable which enables us to probe the net effect.

(4)\begin{equation}{F_i} = {\beta _0} + {\beta _1}Monitorin{g_i} + {\beta _2}{\boldsymbol{X}_i} + {\alpha _z} + {\varepsilon _i}.\end{equation}

Again, X _i is a vector of controls and α_Z are kebelle fixed effects. We also estimate a diff-in-diff model with assignment to the monitoring treatment, ${T_i}$, as the main explanatory variable (interacted with the end-line dummy):

(5)\begin{equation}{F_i} = {\beta _0} + {\beta _2}{T_i} + {\beta _2}Year + {\beta _3}{T_i} \times Year + {\beta _2}{\boldsymbol{X}_i} + {\alpha _z} + {\varepsilon _i}.\end{equation}

We estimate the OLS and diff-in-diff models using two different samples. First, we include many of the key control variables also used by Grossman and Hanlon (Reference Grossman and Hanlon2014), and use the largest sub-sample for which these variables are available. Including controls is especially important for the OLS analysis, but also helps to increase precision of the diff-in-diff estimations.Footnote ²⁰ Second, we re-estimate all models for larger samples without controls (but including kebelle fixed effects). The results and statistical significance of the two samples are very similar, and the same is true for models based on a restricted vector of controls (selected using LASSO). To economize on space we only report the regression results of the former sample.Footnote ²¹ Third, we estimate models where we imputed missing values by the mean of the variable, and included dummy variables to indicate imputation. Fourth, we estimate a series of ANCOVA models. Finally, we estimated models with the three monitoring interventions entering separately, rather than pooled into one monitoring dummy. In all models, we cluster standard errors at the kebelle level. However, results without clustered standard errors (not reported here, but available upon request) remain the same.