Policy Research Working Paper 10315 Internal Conflicts and Shocks. A Narrative Meta-Analysis Camille Laville Pierre Mandon Macroeconomics, Trade and Investment Global Practice February 2023 Policy Research Working Paper 10315 Abstract Do income shocks locally affect internal conflicts? To in the agriculture sector or wealth-increasing shocks in the address this question, this paper employs a meta-regression extractive sector on the local risk of conflict. The paper analysis of 2,464 infranational estimates from 64 recent also shows that studies that fail to uncover empirical effects empirical studies on conflicts and income-related shocks that conform to researchers’ expectations on the theoretical in developing countries. After accounting for publication mechanisms are less likely to be published. Differences in selection bias, the analysis finds that, on average, wealth-in- the geographical area of study, the choice of control vari- creasing shocks in the agriculture sector are negatively ables, and the way shocks are measured substantially explain associated with the local risk of conflict. Nonetheless, the the heterogeneity among estimates in the literature. analysis finds no average effect of wealth-decreasing shocks This paper is a product of the Macroeconomics, Trade and Investment Global Practice. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may be contacted at pmandon@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Internal Conflicts and Shocks. A Narrative Meta-Analysis* Camille Laville„ Pierre Mandon… JEL Classification Numbers: C19; C83; D74; H12; H56; Q34 Keywords: Conflicts; climate shocks; commodity shocks; natural resources; income-driven con- flicts; meta-regression analysis * Unit: EAWM2, Macroeconomics, Trade & Investment Global Practice. Activity: background research for Eco- nomic Updates and other ASAs. TTL: Pierre Mandon „ Researcher, Chair of Defense Economics, IHEDN. Associate researcher, CERDI-CNRS-IRD. … Country Economist, Macro Trade Investment (MTI) Global Practice, World Bank Group. 1 Introduction In the last ten years, 528,763 violent events have taken place worldwide, resulting in the loss of 904,881 human lives.1 This human toll, which does not account for indirect deaths (including from displacement, deprivation, and disease), surpasses the estimated number of battle-deaths from the Iran–Iraq War (1980–1988) or the Afghan Civil War (1978–2002) (Lacina and Gleditsch, 2005). In 35% of these battles, violence against civilians or riots were reported in low-income economies and only 3% in high-income countries. This gap widens further in terms of human toll, with only 1% of total casualties reported in high-income countries, compared to 58% in low-income economies. By 2030, up to two-thirds of the world’s extremely poor people are expected to live in countries affected by violence and conflict (World Bank, 2020). In other words, contexts combining extreme poverty and exposure to conflict may become more common in the upcoming years. While there is little doubt about the damaging consequences of conflict on development (see for example Abadie and Gardeazabal, 2003; Islam et al., 2016), the effect of local economic conditions (i.e. incomes and economic prospects) on the risk of conflict is subject to academic discussion about the explanatory mechanisms and the appropriate empirical strategies to model them (Bazzi and Blattman, 2014; Laville, 2019). Understanding how incomes and economic prospects locally affect the risk of conflict is therefore important in order to provide adequate policy recommendations to low to intermediate income countries concerned about “conflict traps”2 and exposed to income shocks related to climate change and commodity price disruptions. Several empirical studies find that low levels of national incomes are robustly correlated with higher risks of internal conflicts (Hegre and Sambanis, 2006; Blattman and Miguel, 2010). How- ever, the direction of the causal relationship is difficult to establish, especially at the infra-national scale. First, there is an endogeneity issue; the long-term negative consequences of insecurity on the economies and institutions could largely influence the direction of the relationship (Corral et al., 2020). Second, aggregated measures of poverty are coarse proxies for the economic constraints faced by individuals, especially in conflict affected areas (Laville, 2019). Third (and corollary), there is a lack of micro data, since obtaining micro evidence through field surveys is limited by obvious secu- rity concerns (Axinn et al., 2012). Fourth, other hard to quantify individual considerations might 1 Source: Authors’ compilation from ACLED for the period between November 18, 2012 and November 18, 2022. 2 In essence, the idea that civil war begets more civil war (See Collier, 2003). 2 be at play. For example, the very idea that internal conflicts are linked to rational individualism is sometimes questioned given the importance of other social and historical conditions and constraints (see, e.g., Cramer, 2002). Since the 2010s, the use of conflict location data and satellite imagery has offered new avenues for understanding in which local contexts violence develops. Empirical frameworks based on small spatial units (especially grid-cells) and georeferenced data introduce key sources of heterogeneity at the local scale, like the location of mineral deposits (Maystadt et al., 2014; Berman et al., 2017) or the volume of production/exports in each area (Dube and Vargas, 2013; Berman and Couttenier, 2015; McGuirk and Burke, 2020). These variations can then be exploited using quasi-experimental frameworks to isolate and test the validity of one mechanism among others (Blattman and Miguel, 2010; Couttenier and Soubeyran, 2015; Laville, 2019). For example, using grid-cells and georeferenced data on conflicts in Sub-Saharan Africa, Berman and Couttenier (2015) conclude that positive income shocks - proxied by shocks in global demand for the main good exported by the cell - decrease conflict probability in the cell,3 and von Uexkull (2014) find that negative income shocks - proxied by the exposure of rainfed croplands to drought - significantly and substantially increase the risk of conflict. Although these two studies reach similar results, they present key methodological differences in the sign and proxies used to define the shock. This makes it difficult to establish income as a key local factor of conflict. Figure 1 presents the distribution of the t -students from 1.391 point estimates from 61 studies on income shocks and conflicts.4 Figure 1a shows that the literature associates almost twice as often income changes with adverse rather than positive effects on conflict, suggesting a possible asymmetric research focus on the link between income and violence. In Figure 1b, we adjust the t -students so that the ex-ante assumed effects all work in the same direction.5 It reveals that finding a result in line with ex-ante expectations is twice more frequent (66.8 % of estimates) than finding one with an unexpected signs (33.2 % of counter-intuitive or null effects). This coincidence between the obtained and expected results may suggest the presence of genuine effects or manipulations by researchers, namely, publication selection bias (i.e., to prefer results with signs consistent with their expectations or prior understanding, or results with higher statistical significance). 3 This effect disappears when spatial units are countries, suggesting a local nature of the explanatory mechanism. 4 Figure 1 only presents the sample of direct estimates from the MRA database of this study. 5 We reverse the sign of t -statistics for regressions focusing on theoretically pacifying mechanisms as they are the only channels with an ex-ante expectation of reducing risks of conflicts. 3 (a) t -students (b) Adjusted t -students Notes: Distribution of the t -statistics taken from 1.391 estimates from 61 studies. A positive (negative) statistically significant t-student suggests that the estimate shows a statistically higher (lower) risk of conflict. A positive (negative) adjusted t-student signifies that the estimate statistically present (does not present) the expected sign. We consider a risk error of 10 percent, so t-student ≥ 1.645 (or ≤ -1.645). All other t -statistics are reported in the “Statistically null effects ” category. Source: Authors’ compilation from MRA database. Figure 1: Distribution of the t -students and Adjusted t -students To analyze these methodological concerns and assess the presence of publication selection bias in the literature, we perform a meta-regression analysis (MRA) on 2,464 estimates from 64 studies published between 2010 and 2021. The MRA is a regression on estimates from existing regressions. In contrast to simple meta-analysis, this statistical method aims to highlight one or more study characteristics that may explain heterogeneity among estimates from selected studies. In other words, the MRA objective is to summarize and “make sense” of statistical heterogeneity (i.e., the true effects in each study not being identical) on a given topic in the literature (Thompson and Higgins, 2002; Balima et al., 2020). Also, following the guidelines by Stanley and Doucouliagos (2012), the MRA distinguishes the genuine multidimensional effects of income shocks on internal conflicts from the potential publication bias inherent to most economic fields (Doucouliagos and Stanley, 2013). We develop a methodological approach in three steps to assess the effect of various transmission channels and take into account the heterogeneity derived from different empirical methodologies. First, a representative sample of empirical studies is built (called the meta-sample henceforth). Second, all the estimated coefficients from these selected studies are collected. Third, we assess the presence of publication selection bias and genuine effects from the collected estimates and explore the drivers of heterogeneity among them. We find that the literature suffers from two types of publication bias: researchers tend to prefer studies that (i) find higher risks of conflicts when they focus on negative agricultural shocks 4 (Type I publication bias); and (ii) promote results with higher statistical significance (Type II publication bias). After filtering them out, we report the presence a genuine negative effect of positive agricultural shocks on conflicts. We also provide evidence that the section of the literature studying adverse or unspecified mechanisms favors empirical results showing a higher risk of conflict, independently of the sector of the economy affected by the shock. Using Bayesian model averaging techniques, we show that several methodological choices made by the authors explain diverging results in the literature. We find that the most influential moderators relate to the direction of the explanatory mechanism (stressful or pacifying/undefined), the characteristics of the spatial units (their size and the geographic region studied), the method of measuring the shock (through climatic variables, through price variations, in the area where the resource is produced/located), the inclusion of interactions, and the control variables included (GDP or population size). This study is, to our knowledge, the first MRA covering the literature on the effect of income shocks on the risk of conflicts. Previous meta-analyses have examined how the risk of conflict responds to natural resource endowment (O’Brochta, 2019; Vesco et al., 2020), commodity price shocks (Blair et al., 2021), and climate change (Hsiang et al., 2013, 2014). The present meta- analysis aims to reconcile all these different approaches through the prism of income shocks; as highlighted by recent literature reviews (Couttenier and Soubeyran, 2015; Laville, 2019), natural resource endowment, (commodity) price shocks and climate change could affect the risk of conflicts through income shocks both for active (or potential) warring groups and individuals. Furthermore, scholars have repeatedly expressed concerns about the presence of biases (including the publication selection bias) in the conflict literature due to the great variety of tested mechanisms, outcomes and empirical strategies (Ioannidis, 2005; Dixon, 2009; Bazzi and Blattman, 2014), the selection of control variables according to statistical significance and not theory (Ward et al., 2010), or the fragility of published results in studies conducted at the country/year level (Hegre and Sambanis, 2006; Ward et al., 2010). This work takes up several of these criticisms and determines whether they are justified in the context of work done at the sub-national scale on income shocks and conflicts. The paper is organized as follows. Section 2 presents the different mechanisms linking income shocks to conflicts according to the literature. Section 3 discuss the construction of the meta-sample. Section 4 deals with publication bias. Section 5 explores different sources of heterogeneity in the collected estimates. Section 6 summarizes our findings and concludes. 5 2 Review of the Mechanisms Involved in the Literature Among all the possible channels of transmission between local incomes and conflicts, the literature emphasizes the role of opportunity cost and rapacity (Blattman and Miguel, 2010; Couttenier and Soubeyran, 2015).6 The opportunity cost mechanism posits that lower wages in productive sectors of the economy increase relative gains from violent appropriation, which lowers the opportunity cost of conflict and increases the risk of predatory behaviors (Grossman, 1991; Hirshleifer, 1995; Collier, 1998). Alternatively, the rapacity effect states that a rise in contestable income increases the risk of conflict by raising gains from appropriation. Put differently, rising commodity prices increase the rent from their capture and the taxes collected in production areas, facilitating the financing and recruitment capabilities of armed groups (Reuveny and Maxwell, 2001; Collier et al., 2008) and increasing their capacity to sustain a rebel movement (Berman et al., 2017).7 In sum, a positive wealth-increasing shock, typically a higher selling price for a given produced commodity, could both increase conflict through rapacity effects, or decrease it through the oppor- tunity cost mechanism. This comes from the fact that wealth-increasing shocks can simultaneously o and Dal B´ affect wages and returns to conflict. In a simple general equilibrium model, Dal B´ o (2011) predict that a key consideration in determining whether an income shock will increase or decrease the risk of conflict is the nature of the economic sector where it occurs. They find that positive shocks to labor-intensive industries diminish the risk of conflict, while positive shocks to capital-intensive industries increase the risk. A shock to a capital-intensive sector expands the capital-intensive industry and contracts the labor-intensive one. As a result, labor is relatively less scarce, resulting in lower wages and lower costs of appropriation activities relative to the amount of appropriable resources. Empirical evidence supports their findings. In Colombia, Dube and Vargas (2013) find that a rise in the price of coffee, a labor-intensive commodity, decreases violence in production areas, while a rise in the price of oil, a capital-intensive commodity, increases violence. 6 A third mechanism, the state capacity, posits that adverse income shocks expose the state to negative growth, which constrains investment in national counterinsurgency capacities (Fearon and Laitin, 2003). This channel appears however off the scope of this paper since it takes place at the national level and/or supposes that the state largely finance its operations by collecting tax revenues locally (which is a strong hypothesis for many low to intermediate income countries). 7 Other channels of transmission have been investigated by the literature, like separatist ambitions in resource-rich regions (Morelli and Rohner, 2015), lower incentives to develop sufficient state capacity to discourage or buy off rebellion in rentier and resource-dependent states (Fearon, 2005), and grievances from environmental degradation and lack of mining jobs (Ross, 2004). 6 They explain these results by the variation in wages affecting the opportunity cost of conflict, and by the rapacity of armed groups seeking to capture the higher oil rents. Fjelde (2015) and Berman and Couttenier (2015) also find results in line with the opportunity cost channel for labor-intensive agricultural commodities produced in Africa. Recent empirical evidence suggests that the opportunity cost is not an exclusive transmission channel for labor-intensive goods, nor is rapacity exclusive to capital-intensive goods. Commodities’ lootability and producers’ taxation opportunities can also influence the nature of transmission chan- nels. Rapacity can be a relevant transmission channel for labor-intensive export resources if armed groups can tax producers. In Colombia, Angrist and Kugler (2008) show that the increase in the world price of cocaine (whose production is labor intensive) in the 1990s did increase the quantity produced but had a modest impact on producers’ incomes. Indeed, the wealth created was captured by armed groups and the number of violent events increased in the producing regions. Crost and Felter (2020) also find a higher risk of conflict when the price of bananas (a high-value exported anchez de la Sierra (2020) finds that commodity) increases. In the Democratic Republic of Congo, S´ positive demand shocks on coltan, a labor-intensive mineral, increase violence and suggests that the effect of the opportunity cost channel can be overridden by a taxation-induced rapacity effect. Income shocks can also have different effects on the risk of conflict when they impact producers or consumers. In Sub-Saharan Africa, McGuirk and Burke (2020) find that higher commodity prices reduce conflict over the control of territory (what they call “factor conflict”) in food-producing areas, and increase conflict over the appropriation of surplus (“output conflict”) in food-consuming ı (2021) find that higher prices of consumed areas. Using survey data in Nigeria, Abidoye and Cal` goods increase the risk of conflict by reducing consumers’ incomes (in line with the opportunity cost channel), while higher oil prices increase the risk of conflict in oil-producing areas (in line with the rapacity effect). Concerning less organized and violent forms of collective actions, a large body of empirical research finds that food prices can act as a trigger for urban riots and social unrest (Bush, 2010; Bellemare, 2015; Hendrix and Haggard, 2015). They generally assume that relative deprivation and grievances are the main explanatory channel, although it is difficult to provide empirical evidence of causality (Martin-Shields and Stojetz, 2019). Another explanatory mechanism is the breakdown of state authority and legitimacy when it fails to provide food security (Arezki and Brueckner, 2014; Buhaug et al., 2015). 7 3 Construction of the Meta-Sample 3.1 Studies’ Collection Strategy and Inclusion Criteria The first step in meta-analysis is to identify a sample of relevant studies. We use three research methods. First, we use reference snowballing techniques on two seminal reviews of the conflict literature: Couttenier and Soubeyran (2015), and Theisen (2012). Second, we collect relevant references from three meta-analysis on related topics: Blair et al. (2021) (on price shocks and conflicts), Vesco et al. (2020) (on natural resources and conflicts), and Hsiang et al. (2013, 2014) (on climate changes and conflicts). Third, we run keyword searches on Google Scholar and extract the results by web scraping using the R package ‘rvest’ (Wickham, 2022). We search for keyword associations in the format ‘”keyword1 shocks and keyword2” ’ (in whole documents) and ‘”keyword1 shocks” keyword2 ’ (in reference titles), where keyword1 specifies the type of shock ( ‘income’, ‘price’, ‘natural resource’, ‘climate’, ‘climatic’ or ‘environmental’) and keyword2 indicates the type of conflict outcome (‘conflict’, ‘war’, ‘violence’, ‘unrest’). To be as comprehensive as possible, we also looked for plural or singular forms of keywords. A total of 112 different keyword associations were searched on Google Scholar in August and September 2022. Our searches identify 565 potentially relevant studies (22 from literature reviews, 138 from meta-analysis, and 405 from keyword searches). To ensure the coherence of our sample, we choose seven inclusion criteria. We only select papers (i) published in peer-reviewed academic journals,8 (ii) between 2010 and 2021, and (iii) reporting exploitable empirical results (including regression coefficients, sample size, standard errors and/or t -statistics). We focus on studies (iv) where the outcome variable is the onset (i.e. conflict start), the incidence (i.e. number of conflicts), or the duration of a form of internal conflict,9 and (v) where units of analysis are sub-national (e.g., grid-cells, municipalities, states, prefectures, or districts). We only select studies (vi) analyzing at least one income-related channel of transmission and (vii) in one or several non-OECD high income countries.10 Sixty-four studies meeting our selection criteria compose our final sample. Figure A3 8 We therefore exclude working papers that may be of high quality in terms of content but have not been subjected to the peer review process. 9 We leave aside their intensity (i.e., number of fatalities) to limit causal heterogeneity between studies since these analysis refer to quite different explanatory mechanisms (see, for example, Lacina, 2006). 10 Source: World Bank Country and Lending Groups, https://datahelpdesk.worldbank.org/knowledgebase/ articles/906519-world-bank-country-and-lending-groups, accessed in November, 2022. 8 depicts the study collection procedure and maps out the number of records identified, included and excluded. 3.2 Estimates Collection For each selected study, we collect estimates of the effect of income-related shocks, as well as information on the shock itself, the conflict variable used, the size and composition of the sample, the estimation techniques, the level of aggregation of spatial units, the covariates used, the publication year and formats, and other relevant information for the MRA. From the 64 studies selected, we collect 1,391 direct estimates and 1,073 conditional (or interactive) estimates. Notes: The solid line shows the mean of reported t -statistics; the dashed line denotes the mean of the median estimates of the study. To ensure the figure’s readability, absolute t -statistics higher than 10 (0.6% of the observations) and below -10 (0.3% of the observations) are not presented. Source: Authors’ compilation from MRA database. Figure 2: Distribution of T -Statistics Figure 2 plots the distribution of the values of t -statistics used in the collected estimates. As we selected studies analyzing either positive or negative shocks, we observe two spikes in the distri- bution, one around -2 and one around 2 (i.e. the statistical significance threshold or 5%). Positive t -statistics are slightly more frequent (59% of the observations), which explains why the mean and 9 median values are 0.6 and 0.9. Of the t -statistics, 22% are below -2 and 36% are above 2. Values exceeding 10 or inferior to -10 only represent 0.9% of the sample, with minimum and maximum values of -27.7 and 46.8. To prevent potential distortions caused by the presence of outliers, we winsorize t -statistics and degrees of freedom at the top and bottom of 5% level (Lipsey and Wilson, 2001; Viechtbauer and Cheung, 2010).11 3.3 Grouping the Collected Estimates According to Their Main Transmission Channels The collected estimates test the relationship between the risk of internal conflict and different types of positive or negative shocks. As a result, our meta-sample is heterogeneous in terms of the variables of interest and the transmission channels implicitly tested by the authors. Analyzing this raw sample would complicate the interpretation of the MRA’s results and limit our contributions to two central debates in conflict economics, namely what economic mechanisms are at play and how to test them empirically. We therefore split the collected estimates into four meta-regression subgroups that differ in the direction of the shock (wealth increasing or decreasing) and the sector of activity that is affected (agriculture, extractive or other sectors). Indeed, the literature suggests that income shocks affecting the extractive or the agriculture sector do not refer to the same main o and Dal B´ channel of transmission (Dal B´ o, 2011). The first group, Negative Agricultural Shock (AS-), contains all estimates of negative tran- sitory agricultural shocks (e.g., droughts, floods, rain deficiencies, etc.). The second one, Positive Agricultural Shock (AS+), includes all estimates of positive transitory agricultural shocks (e.g., increased demand and international prices for the cultivated good, environmental conditions par- ticularly suitable to its production, etc.). The third group, Positive Hydrocarbon/Mineral Shock (HS+), contains all estimates of transitory shocks on extractive goods (i.e., hydrocarbon and minerals), including increases of the international price of the resource and subsidies to min- ing concessions. The other estimates fall into the heterogeneous category Other Shocks, which includes estimates of (positive or negative) pure climatic shocks that are not explicitly related to agriculture, labor market shocks, financial crisis, or shocks to the drug sector. 11 Winsorisation corrects biases linked to extreme values without losing observations. It consists in replacing the outliers by the highest values in given percentiles. 10 Following the literature, we may expect shocks affecting the agriculture sector to mainly test poverty related mechanisms (notably, the opportunity cost channel) as they affect more labor- intensive commodities and as rapacity effects concern a specific subset of high-value agricultural exports. Alternatively, we may expect shocks affecting the extractive sector to mainly test rapacity channels as they focus on more capital-intensive commodities. As a matter of fact, Table 1 confirms our expectations. Of the 808 estimates focusing on positive agricultural shocks, 648 (80%) test poverty-related mechanism and 151 (19%) test rapacity effects. Of the 951 estimates focusing on negative agricultural shocks, 798 (84%) test poverty-related mechanism and only 5 (0.5%) test rapacity effects, and 139 (15%) test other channels of transmission (including 107 cases of regional destruction). Table 1: Main Implicitly Tested Mechanism in Estimates According to their Meta-Regression Sub- group Agricultural shock (+) Agricultural shock (-) Hydrocarbon/Mineral shock (+) Other shock (+ or -) Total Mechanism Poverty-related 648 798 82 46 1574 Rapacity-related 151 5 423 4 583 Other 4 139 38 0 181 Not specified 5 9 0 112 126 Notes: Number of estimates in the total sample of 2,464 estimates. Other shock (+ or -): estimates of negative or positive pure climatic shocks, estimates of labor market shocks, financial crisis, and estimates of shocks to the drug sector. Poverty-related mechanisms include the dynamism of local economy, food scarcity, budget constrains and water insecurity. Rapacity-related mechanisms include rent capture and the funding of insurgents. Other mechanisms include exodus, grievance, imperfect information, regional destruction (incl. agricultural soils) and state capacity. Source: Authors’ compilation from MRA database. Figure 3 presents the proportion of collected estimates in each meta-regression subgroup. Our three main subgroups (AS-, AS+, and HS+) account for 90.2% of the baseline sample of 1,391 re- gressions and 66.8% of the baseline estimates concern agricultural shocks (71.4% of the total sample of 2,464 regressions). Other shocks only account for 6.5% of the total sample. Interestingly, while selected studies on the agriculture sector investigate equally positive and negative shocks, studies focusing on the hydrocarbon/mineral sector only investigate positive shocks. Detailed information on the types of shocks in each meta-regression subgroup is provided in Figure A1, Table A1, and Tables A5 to A8 in the Appendix. 11 Notes: Shares in the total sample of 2,464 estimates are represented by solid bars. Shares in the baseline sample of 1,391 estimates are represented by hatched bars. OTHER 1: estimates of negative pure climatic shocks. OTHER 2: estimates of labor market shocks, financial crisis, and shocks to the drug sector. OTHER 3: estimates of positive pure climatic shocks. Source: Authors’ compilation from MRA database. Figure 3: Distribution of Estimates According to their Meta-Regression Subgroup 4 Publication Selection Bias and Genuine Effects Publication selection is a common phenomenon in empirical studies that can be broadly defined as the process of selecting research papers or estimates for their statistical significance (Stanley and Doucouliagos, 2012). When this bias is substantial, it can distort statistical inference and any resulting understanding of research as more significant effects are overrepresented in the published literature (Stanley and Doucouliagos, 2012). There are two types of publication selection bias. Type I bias is the tendency to prefer results whose signs are consistent with expectations or prior understanding, and Type II bias is the tendency to prefer results with higher statistical significance, regardless of their sign. Publication selection bias can arise due to several patterns intrinsic to em- pirical research, including editors’ predisposition to accept papers consistent with the conventional view and/or presenting highly significant results, as well as researchers’ self-censoring attitudes and 12 tendency to select their models based on conventionally accepted results (Card and Krueger, 1995; Stanley and Doucouliagos, 2012). By filtering these publication biases, we can determine, if any, the ”true effect” (or ”genuine effect”) of income shocks on the risk of internal conflict. 4.1 At First Glance (Funnel Plots) To illustrate the distribution of observations, we produce a funnel chart by plotting the partial correlation against precision (the inverse of its standard error).12 By construction, estimates with a larger standard error (less precision) are spread at the bottom of the graph while those that are more precise form the top of the funnel. In the absence of publication selection bias, the funnel plot should be symmetric, with observations randomly distributed around the “true” effect (Egger et al., 1997; Stanley, 2007; Stanley and Doucouliagos, 2012). We have ex-ante expectations about the selective report of scholars and researchers about higher or lower risks of conflicts, depending on the transmission channels. We may suspect lower risks of conflicts in case of positive agricultural shocks (AS+) (e.g., increase in demand and international producers’ prices, exceptional rainfalls). On the contrary, we may suspect higher risks of conflicts in case of negative agricultural shocks (AS-) (e.g., droughts, floods, rain deficiencies), and positive hydrocarbon/mineral shocks (HS+) (e.g., increase in mineral or oil international prices, subsidies to mining concessions). As AS- represent more than one third (38.6 percent, see Figure 3) of the 2,464 observations, we would expect either a funnel plot centered close to zero and/or an asymmetry towards higher risks of conflicts. Figure 4 highlights a mix of both intuitions; the upper-left chart suggests potentially null average genuine effects as the more precise estimates (at the top of the funnel) are closely distributed around zero. Moreover, the funnel appears slightly right-skewed (i.e., an asymmetry towards higher risks of conflict) for less precise estimates (at the bottom of the funnel), indicating the likelihood of publication selection bias, altogether with potential genuine effects. Switching to subgroups analysis, we find a right-skewed distribution (i.e., an asymmetry towards higher risks of conflict) when focusing on the of impact of AS-, a surprisingly slightly left-skewed distribution (i.e., an 12 t Partial correlation is computed as r = √ , where t is the t-statistic of the regression coefficient and df denotes t2 +df the degrees of freedom. Partial correlation coefficients measure the strength and direction of the association between potential determinants of conflicts and conflicts’ outcomes, holding all other factors constant. The standard error of the partial correlation is computed as (1 − r2 )/df in line with Stanley and Doucouliagos (2012). 13 asymmetry towards lower risks of conflict) when focusing on the impact of HS+, and a slightly right-skewed distribution when focusing on the impact of AS+. This comforts us on the importance of controlling for mechanisms encompassed by authors (e.g., poverty related, rent capture) to assess the impact of local shocks on conflicts. For AS- and possibly HS+, the skewness is more pronounced for less precise estimates (at the bottom of the funnel), indicating the likelihood of publication selection bias. For AS+, a (left-side) skewness exists also for more precise estimates, indicating potentially the presence of genuine lower risks of conflicts despite the overall slightly right-skewed asymmetry. Notes: The dashed vertical line shows the weighted average partial correlation (0.008, see Figure from the Appendices section), using inverse variance weights. Precision is measured as the inverse of the estimated standard error of the partial correlations. Upper-left panel: baseline sample (1,391 obs.). Upper-right panel: baseline subsample for the subgroup of positive agricultural shocks (448 obs.). Lower-left panel: baseline subsample for the subgroup of negative agricultural shocks (481 obs.). Lower-right panel: baseline subsample for the subgroup of positive hydrocarbon/mineral shocks (325 obs.). Source: Authors’ compilation from MRA database. Figure 4: Funnel Plots: Partial Correlations Between Potential Channel of Transmission and Risks of Conflicts 4.2 Method The unit of observation in this MRA is the regression, given that it presents notable difference with other regressions. As a result, estimates within the same study are likely to be interdependent 14 (Balima et al., 2020). To capture the between-study heterogeneity while controlling for within- study dependence, a multilevel model13 is most appropriate (see, e.g., Doucouliagos and Laroche, 2009; Doucouliagos and Stanley, 2009). The multilevel model accounts for within-study dependence through the inclusion of a random individual effect for each study, hence the reference to a multilevel random effect model. More specifically, the following Equation 1 is considered: ef f ectij = β1 + β0 SEij + λj + ϵij (1) where ef f ectij stands for the ith estimate from the j th study on the effect of income shocks on conflicts; SEij for the standard error (or accuracy of the estimate) of ith estimate from the j th study. λj is the study level random effect and ϵij is a disturbance term. β1 stands for the ”true value” of the effect of income shocks on conflicts. β0 SEij captures the ”noise” or very tendency from researchers to prefer results that are statistically significant, and thus to make use of alternative estimation techniques and/or model specifications to get high significance levels, notably in the face of a small sample. If the number of observations increases indefinitely, the standard error SEij will tend towards zero (i.e., the smaller the standard error, the more accurate and reliable the estimate), and the estimated effect of income shocks will get closer to β1 . In other words, if there is no publication selection bias, the collected estimates will vary randomly around the ”true value” β1 , regardless of the standard error. We correct for heteroscedasticity owing to differences across studies in the sample size and model specifications used by dividing Equation 1 by SEij , which becomes Equation 2: 1 λj tij = β0 + β1 + + εij (2) SEij SEij where tij represents the collected t-values. We then assess the existence of Type I publication selection bias by testing the null hypothesis that the intercept in Equation (2) is equal to zero (β0 = 0), which is also known as the ’Funnel Asymmetry Test’ - ’Precision Effect Test’ (FAT-PET). Basically, if the intercept is statistically significant, it means that the collected estimates do not vary symmetrically and randomly around the ”genuine effect” (β0 SEij in Equation 1 has an influence). In this case, it is as if part of the evidence (estimates in the literature) is missing, resulting in a 13 In the present case, a model allowing for heterogeneity both at the study (or paper) level and the estimate level. 15 truncated plot.14 Our MRA focuses on the link between positive and negative income shocks, hence on estimates with likely opposite signs. To ensure comparability of estimates, we replace the left-hand side of Equation 2 with the absolute t-student value and consider alternatively adjusted t-student values: we reverse the signs of t -statistics for regressions focusing on pacifying mechanisms as they are the only channels with an ex-ante expectation of reducing risks of conflicts. This gives us Equation 3: 1 λj |tij | = β0 + β1 + + εij (3) SEij SEij Testing the null hypothesis (β0 = 0) in Equation 3 assesses the presence of Type II publication selection bias. To check the presence of a ”genuine” effect after filtering out potential publication bias, we follow Stanley and Doucouliagos (2012) and carry out the so-called Precision Effect Estimate with Standard Error (PEESE) test. Concretely, we test the null hypothesis that the parameter associated with the inverse standard error (β1 ) in Equation 2 equals zero. In other terms, we test whether in Equation 1, the intercept (β1 or genuine effect) has a statistically significant role regardless of the outcome of the publication selection bias (influence of β0 SEij ). Rejecting the null hypothesis would thus signal that a genuine effect remains after filtering out the publication bias. In addition to the FAT-PET-PEESE procedure, we estimate the bias-adjusted genuine effect using the Weighted Average of Adequately Powered (WAAP) estimator (Ioannidis et al., 2017; 2 ) on the Stanley et al., 2017). The WAAP is a weighted average that uses optimal weights (1/SEij only ’adequately powered’ estimates, which are usually defined as having standard errors smaller than the multilevel mixed-effects model estimates divided by 2.8 (Stanley et al., 2017). 4.3 Results Table 2 reports the associated results for the whole sample of 2,464 estimates (Panel A) and our baseline sample of 1,391 estimates (Panel B ). Columns [1] and [2] present results for Type II publication bias, using respectively the adjusted and the absolute t -statistics. Columns [3] to [7] depict results for Type I publication bias considering continuous t -statistics for each type of shock (i.e., each meta-regression subgroup). 14 Hence the reference to a ”funnel asymmetry” test. 16 For both Panel A and Panel B, the intercepts (FAT) in columns [1] and [2] are positive15 and highly significant, pointing to the existence of Type II publication bias. This suggests that researchers have incentives to promote results with higher statistical significance, in line with most MRA findings (Doucouliagos and Stanley, 2013). Table 2: Publication Selection Bias and Genuine Effect Tests [Baseline Results] [1] [2] [3] [4] [5] All local shocks Main channels of transmissison Absolute t-student Adjusted t-student Agr. shock (+) Agr. shock (-) Hydr./Min. shock (+) Panel A: whole sample [i] FAT-PET Mean beyond bias (PET) Precision (1/SE) -0.000 * 0.000 -0.003 *** -0.001 0.000 (0.000) (0.000) (0.001) (0.001) (0.001) Publication bias (FAT) Constant 2.383 *** 1.381 *** -0.312 1.641 *** 0.566 (0.119) (0.184) (0.397) (0.296) (0.460) [ii] WAAP Mean beyond bias Constant - - - - - - - - - - #studies 64 64 25 34 14 Observations 2 464 2 464 808 951 543 %Observations 100% 100% 33% 39% 22% Panel B: baseline coefficients only [i] FAT-PET Mean beyond bias (PET) Precision (1/SE) 0.000 0.001 -0.005 *** 0.000 0.001 (0.000) (0.000) (0.001) (0.001) (0.001) Publication bias (FAT) Constant 2.502 *** 1.745 *** -0.084 2.165 *** 0.708 (0.129) (0.223) (0.493) (0.324) (0.601) [ii] WAAP Mean beyond bias Constant - - 0.000 0.002 *** 0.001 - - (0.001) (0.000) (0.000) #studies 61 61 23 32 13 Observations 1 391 1 391 448 481 325 %Observations 100% 100% 32% 35% 23% Notes: FAT-PET models are estimated with a multilevel mixed-effects model and the weighted average of the adequately powered (WAAP) is derived from Ioannidis et al. (2017) and Stanley et al. (2017). Panel A refers to the whole sample of 2,464 observations and 64 studies and Panel B refers to the sample of baseline coefficients, including 1,411 observations and 63 studies. The dependent variable is the t -statistic of the estimate of interest on conflicts as dependent variable. Standard errors, derived from observed information matrix, are reported in parentheses. Columns [1] and [2] report the results for all local shocks, using respectively the absolute value of the t -statistic of the estimate of interest and the adjusted t -statistic of the estimate of interest as dependent variables. Columns [3] to [5] focus on the main channels of transmission for conflict, namely positive agricultural shocks (col [3]), negative agricultural shocks (col [4]) and positive hydrocarbon/mineral shocks (for commodity exporters, col [5]). Results on pure (positive and negative) climatic shocks (respectively 39 and 85 observations over the whole sample of 2,464 observations) and other potential shocks (50 observations over the whole sample of 2,464 observations) are presented in Table A2 from the Appendices section. A detailed description of all variables is available in Table A3. ∗p < 0.10, ∗ ∗ p < 0.05, ∗ ∗ ∗p < 0.01. Source: MRA database. 15 The sign of the intercept is only meaningful for the absolute measure of the t -Student, its absolute value being necessarily positive. 17 To refine the assessment of publication bias (and derive the genuine risks of conflicts, if any), we then focus on our three main meta-regression subgroups: AS+, AS-, and HS+ in columns [3] to [5] of Table 2 (the results for the other types of shocks are presented in the Appendix). The dependent variables are the t -statistics value of the collected estimate, as opposed to the adjusted and absolute value used when considering the whole sample. Results associated with these more homogeneous sets of estimates show a significant intercept in column [4] only. This suggests the presence of a Type I publication bias in studies analyzing the effect of AS-. In other words, researchers tend to prefer studies in line with conventional views (Card and Krueger, 1995) when they study negative agricultural shocks, and in the present case, higher risks of conflicts. Our results also point to a rather “substantial” selectivity in that sub-literature, as supported by the FAT values.16 Interestingly, as already suggested by the funnel plots (see Figure 4 above), we do not find any evidence of Type I publication bias for studies focusing on AS+ (column [3]) and HS+ (column [5]). Since the majority of the estimates collected for AS+ and HS+ are price shocks (see table A1), these results are consistent with those of Blair et al. (2021), who find no evidence of publication selection bias in the literature on price shocks and conflicts. One potential but not definitive explanation is the nature of shocks: while AS+, and HS+ largely rely on market and price shocks, AS- largely rely on natural events (see table A1). It is possible that the analysis of climate events allows for a more arbitrary selection of results. Climate shocks can be captured by a range of variables (precipitation, drought, temperature, etc.) and parameters (different indicators, measures of variance over different periods, etc.). These methodological choices have theoretical implications and can lead authors into various empirical pitfalls (Auffhammer et al., 2013).17 The results for other types of shocks (see Table A2) also point toward this explanation. Indeed, they suggest the presence of selection bias only in the regressions of pure positive and negative climate shocks, which are also measured with a heterogeneous set of climate variables (see Figure A1 and Table A1 in the Appendix). These elements will be discussed further in the heterogeneity analysis section. 16 A FAT absolute value smaller than 1 is synonymous of ”little to modest” selection bias, while a FAT test absolute value ranging between 1 and 2 rather signals “substantial” selectivity (Doucouliagos and Stanley, 2013). 17 Additionally, the WAAP by Ioannidis et al. (2017) and Stanley et al. (2017) shows few adequately powered significant estimates for the subsample of baseline AS- estimates only. In other words, the greater variety of shocks associated with AS- is associated both with overall opportunities of (type I) publication selection bias and genuinely more unconditional risks of conflicts for the few subsample (150 out of 481 observations) of adequately powered estimates. 18 When the adjusted and absolute t -statistics are used as dependent variable, it is hard to interpret the coefficients associated with 1/SE (precision parameter) as genuine effects, as the meta-group consists of a synthesis of studies that do not rest on a single transmission channel. Columns [3] to [5] of Table 2 allow going beyond the publication bias and testing for the existence of genuine effects of income shocks on conflicts for each meta-subgroup. After filtering out the publication bias (slope coefficients reported in columns [3] and [5]), we find a negative effect of AS+ and no evidence of genuine effects for AS- and HS+. Put differently, after filtering out the publication bias, we find that only positive agricultural shocks genuinely influence (here, reduce) the risk of conflict. This does not mean that AS- and HS+ have no effect on conflicts, but that their effects depend on several factors that will be discussed in the heterogeneity analysis section. To sum up, these results show that the literature on the effects of income shocks on conflicts suffers from two types of publication bias: researchers tend to prefer studies that (i) find higher risks of conflicts when they focus on AS- (Type I publication bias); and (ii) generally promote results with higher statistical significance (Type II publication bias). We also find evidence that AS+ are associated with lower risks of conflicts. In the next section, we look at how key interactive factors might affect the baseline results. 5 Heterogeneity Analysis As we already pointed out, individual studies on the link between income shocks and conflict vary greatly in terms of data and method used. The purpose of this section is to investigate whether authors’ methodological choices systematically influence the estimated partial correlation coefficients and whether the estimated coefficients of publication bias from Section 4 survive the addition of moderator variables. Our approach to heterogeneity analysis in divided into two steps. We start by examining if sections of the literature that focus specifically on pacifying or destabilizing shocks or use interaction models are affected differently by publication bias. We then use Bayesian methods to test whether systematic differences in the data and methodological choices made by the authors explain the variations in the results that they obtain. 19 5.1 Publication Bias and Authors’ Initial Assumptions on the Sign and Strength of the Mechanisms at Play Figure 5 illustrates how our sample of selected regressions breaks down according to (i) the authors’ initial assumptions about the effect of the mechanism they are testing, and (ii) the interaction format they apply to test its robustness. We identify regressions testing a theoretically pacifying mechanism by selecting the ones where the authors analyze the effect of AS+ through any other mechanism than rapacity,18 the effect of AS- through rapacity only, or the effect of HS+ through the dynamism of the local economy or the state’s capacity to impede rebellion. Alternatively, re- gressions testing a theoretically stressful mechanism are all regressions where the tested mechanism does not have a pacifying effect.19 Additionally, we identify three categories of key interactive terms: (i) ex-ante risk enhancers, (ii) ex-ante risk absorbers, and (iii) interaction terms with no specific ex-ante expectations (undetermined moderator). The ”risk enhancers” moderator contains all in- teractive terms that the authors assume will increase the risk of conflict ceteris paribus (e.g., not cultivating sweet potatoes, not living in a dam downstream district, living far from key infrastruc- ture). Conversely, the ”risk absorbers” moderator contains estimates where the variable interacted with the shock is expected to mitigate the risk of conflict (e.g., cultivating sweet potatoes, living in a dam downstream district). The ”undetermined” moderator contains all other conditional charac- teristics. It is not assumed to have a homogenous impact on conflicts and authors are more agnostic on average. Table A11 in the Appendix summarizes the different key interactive terms and the rationale underpinning their introduction. Figure 5 shows that 65% of the selected regressions test a stressful mechanism, 30% test a pacifying mechanism, and 5% test a mechanism with undetermined effects. This indicates that a majority of studies on income shocks and conflicts focus on the detrimental impact of income shocks, which narrows the relative amount of empirical evidence available on local income shocks as a way out of conflicts. The figure also shows that regressions where the income shock is interacted with one or more unit-specific variables account for 40.6% of the collected estimates (1,073 estimates), 18 As noted in the literature review section, some studies find that a positive price shock in the agriculture sector can increase the risk of conflict if it concerns a good with a high export value (Crost and Felter, 2020; S´ anchez de la Sierra, 2020), or if it affects the purchasing power of consumers (McGuirk and Burke, 2020; Abidoye and Cal` ı, 2021). 19 For this identification process, we only take into account regressions where the authors clearly state ex-ante what is the mechanism they are testing. We therefore leave aside 121 regressions for which the authors do not specify which mechanism they are testing. 20 revealing a recurrent practice in empirical models of conflicts. For example, in a study on weather shocks and peasant revolts in historical China, Jia (2014) interacts indicators of droughts and floods with sweet potatoes as food production (known to be more resilient to bad weather than rice or wheat production) as a ”risk absorber” moderator. The inclusion of interaction terms can help researchers identify the mechanisms at play by highlighting heterogeneous effects depending on the characteristics of the geographical areas studied. These moderator variables are model-specific and rarely similar across studies. Because interaction models can help identify or mitigate the action of specific mechanisms, result searching bias is likely to be captured, if existent, in the choice of interacted variables. Indirect (+) Risk Enhancer (258) Theoretically Mechanisms Risk Absorber (99) Stressfull Undetermined (364) (878) Higher risk of conflict Direct (+) (1599) SHOCK Direct (-) Lower risk of conflict (400) (744) Theoretically Mechanisms Pacifying Indirect (-) Risk Enhancer (159) Risk Absorber (60) Undetermined (125) Notes: Expected signs of direct and indirect (interaction) estimates. The number of observations concerned is given in parenthesis. 121 observations (not shown) refer to mechanisms with uncertain sign (including 113 direct estimates, 5 indirect estimates with a risk enhancer, 3 indirect estimates with a risk absorber). Source: Authors compilation from MRA database. Figure 5: Expected Signs of Direct and Indirect Channels of Transmission To test if these methodological choices are a source of publication bias per se, we augment the baseline model with dummy variables indicating the type of mechanism or the type of interac- tion that is tested. When the multilevel random effect model includes covariates (or moderators) accounting for heterogeneity between studies, the model becomes best described as a ”multilevel mixed-effect model”. More specifically, Equations 2 and 3 become: 21 ′ 1 xij λj tij = β0 + β1 + βk + + εij (4) SEij SEij SEij or ′ 1 xij λj |tij | = β0 + β1 + βk + + εij (5) SEij SEij SEij ′ where xij stands for a set of meta-independent variables capturing empirical study characteristics from the meta-sample that explain the differences in estimates between studies. β0 measures the severity of publication bias conditional on the inclusion of controls, and β1 is the mean β estimate corrected for publication bias but also conditional on the variables included. Several multilevel random-effect methods have been proposed in the MRA literature to estimate the between-study variance in meta-regressions. The most used method computes the unknown variance of the random-effect model through an iterative residual (restricted) maximum likelihood process (REML), with normal distributions assumed for both the within and between-study ef- fects. We rely on the multilevel mixed-effect REML, given its properties, namely avoiding not only downward-biased estimates of the between-study variance, but also underestimated standard errors and anti-conservative inference.20 In addition, the restricted maximum likelihood (REML) is preferred to the maximum likelihood (ML) because the latter does not account for the degrees of freedom employed when estimating the fixed-effect portion of the model, which is a key short- coming, especially given the small number of studies included in the MRA (Thompson and Sharp, 1999). The regressions are performed on the extended sample of 2,464 observations (our 1,391 base- line estimates plus 1,073 interactive terms) to study the heterogeneity in the estimated effects of income shocks on conflicts due to interactions. But to study the heterogeneity due to the (ex-ante ) direction of tested mechanisms by the authors, we do the regressions only on the baseline sample of 1,391 estimates, to exclude the potential “noisy” effects of interactive terms in the analysis. Table 3 presents the results we obtain for AS+, AS-, and HS+ when we control for the type of mechanism tested. Coefficients associated with the analysis of a supposedly pacifying mechanism are negative and statistically significant in Columns [2], [6], and [10]. Coefficients associated with the analysis of a theoretically stressful mechanism are positive and statistically significant in Columns 20 Anti-conservative inference happens when scholars and researchers do not update their prior and beliefs when they face incomplete information, noise, fallible data or counter-intuitive results for instance. 22 [3], [7], and [11]. This indicates that studies focusing on pacifying mechanisms in each meta- subgroup report a lower risk of conflict following an income shock, while studies focusing on stressful mechanisms report a higher risk of conflict. The presence of publication bias appears conditional to the type of mechanism tested. We find a positive and statistically significant intercept when we control for pacifying mechanisms in columns [2], [6] and [10]. When we control for stressful mechanisms, the intercept in column [11] is also statistically significant (at the 0.10 level) but negative. These results indicate the presence of a substantial Type I publication selection bias in studies that do not focus on pacifying mechanisms, and in studies encompassing HS+ that do not test stressful mechanisms. In other words, the current literature favors the publication of results showing: (i) the detrimental impacts of income shocks on the local risk of conflict through theoretically stressful/undefined mechanisms, (ii) that pacifying mechanisms in the extractive sector reduce the local risk of conflict. Table 3: Explaining Heterogeneity in the Estimated Effects of Income Shocks on Conflicts [Direction of the Tested Mechanism] [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] Main channels of transmission AS+ AS- HS+ Whole sample Precision (1/SE) -0.005 *** -0.005 *** -0.005 *** -0.005 *** 0.000 0.000 0.000 0.000 0.001 0.001 0.001 0.001 (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) Constant -0.084 1.695 *** -0.360 -0.084 2.165 *** 2.475 *** 0.575 2.172 *** 0.708 1.242 ** -2.108 * 0.708 (0.493) (0.533) (0.451) (0.493) (0.324) (0.261) (0.712) (0.331) (0.601) (0.556) (1.237) (0.601) Type of mechanism Mecha. pacifying=1 -2.055 *** -4.885 *** -3.350 *** (0.333) (1.043) (1.353) Mecha. stressful=1 2.055 *** 1.814 *** 3.350 ** (0.333) (0.736) (1.353) Mecha. undefined=1 - -0.123 - - (0.930) - #studies 23 23 23 23 32 32 32 32 13 13 13 13 Observations 448 448 448 448 481 481 481 481 325 325 325 325 %Observations 32% 32% 32% 32% 35% 35% 35% 35% 23% 23% 23% 23% Notes: All models are estimated with a multilevel mixed-effects model on the whole sample of 2,464 observations and 64 studies. The dependent variable is the t-statistic of the estimate of interest on conflicts as dependent variable. Standard errors are reported in parentheses. Columns [1] to [4] focus on positive agricultural shocks, columns [5] to [8] focus on negative agricultural shocks, and columns [9] to [12] focus on positive hydrocarbon/mineral shocks. Results on pure (positive and negative) climatic shocks and other potential shocks are presented in Table A2 from the Appendices section. A detailed description of all variables is available in Table A3. The detailed composition of mechanisms is available in Tables A5, A6, and A7. ∗p < 0.10, ∗ ∗ p < 0.05, ∗ ∗ ∗p < 0.01. Source: Authors’ compilation from MRA database. Table 4 presents our results when we control for the type of interactive term in the regression. Coefficients associated with the inclusion of a risk absorber interaction term are negative and statis- tically significant in Columns [2], [6] and [10]. Conversely, coefficients associated with the inclusion of a risk enhancer interaction term appear positive and statistically significant in Columns [3], [7] and [11]. This indicates that studies implementing interactions conclude a higher (or lower) local risk of conflict in line with the authors’ expectations. Columns [4], [8] and [12] show that coeffi- 23 cients associated with the inclusion of an interaction term with uncertain effects are statistically significant for shocks to the agriculture sector, but not for shocks to the extractive sector. The coefficient is positive for AS+ and negative for AS-. Therefore, these uncertain moderators tend to behave like risk enhancers of AS+, and risk absorbers for AS-. Overall, our results suggest that the authors prefer to highlight contrasted results that support the mechanisms being tested. For instance, Jia (2014) provides evidence about the detrimental impact of droughts and other negative weather shocks (AS-) on conflicts in historical Chinese prefectures (from 1470-1900) but she finds that cultivating sweet potatoes (drought resistant, risk absorber) mitigates this impact. These find- ings support her hypothesis that weather shocks increase the risk of revolt due to their negative impact on food supply. Columns [2], [6], and [10] indicate that the presence of publication bias depends on the type of interaction tested. Studies focusing on AS+ prefer a result for the intercept that shows a lower risk of conflict when they include an interaction with a risk enhancer. Similarly, studies focusing on HS+ and AS- tend to prefer an intercept showing a higher risk of conflict when there is an interaction with a risk absorber. This indicates a small or modest Type I publication selection bias: a tendency to select interaction terms of opposite direction as long as they support the tested mechanism.21 Table 4: Explaining Heterogeneity in the Estimated Effects of Income Shocks on Conflicts [Inclusion of Interactions] [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] Main channels of transmission AS+ AS- HS+ Whole sample Precision (1/SE) -0.003 *** -0.003 ** -0.003 ** -0.003 *** -0.001 0.000 -0.001 -0.001 0.000 0.000 0.000 0.000 (0.001) (0.001) (0.001) (0.001) (0.001) (0.000) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) Constant -0.312 -0.124 -0.714 * -0.332 1.641 *** 2.121 *** 1.463 *** 1.665 *** 0.566 0.899 * 0.287 0.549 (0.397) (0.406) (0.390) (0.400) (0.296) (0.270) (0.321) (0.294) (0.460) (0.474) (0.467) (0.461) Type of interaction Risk absorber=1 -2.175 *** -3.435 *** -2.726 *** (0.241) (0.107) (0.235) Risk enhancer=1 2.923 *** 1.405 *** 2.385 *** (0.179) (0.167) (0.287) Risks uncertain=1 0.743 ** -0.912 ** -0.298 (0.318) (0.442) (0.304) #studies 25 25 25 25 34 34 34 34 14 14 14 14 Observations 808 808 808 808 951 951 951 951 543 543 543 543 %Observations 33% 33% 33% 33% 39% 39% 39% 39% 22% 22% 22% 22% Notes: All models are estimated with a multilevel mixed-effects model on the whole sample of 2,464 observations and 64 studies. The dependent variable is the t -statistic of the estimate of interest on conflicts as dependent variable. Standard errors are reported in parentheses. Columns [1] to [4] focus on positive agricultural shocks, columns [5] to [8] focus on negative agricultural shocks, and columns [9] to [12] focus on positive hydrocarbon/mineral shocks. Results on pure (positive and negative) climatic shocks and other potential shocks are presented in Table A2 from the Appendices section. A detailed description of all variables is available in Table A3. The detailed composition of risk absorbers, risk enhancers and risk uncertain is available in Tables A9 and A10. ∗p < 0.10, ∗ ∗ p < 0.05, ∗ ∗ ∗p < 0.01. Source: Authors’ compilation from MRA database. 21 Empirically, a lower (or higher) value of the intercept (i.e. the effect of the shock when the interaction term is null) implies a seemingly smaller (or larger) risk of conflict when the conditions of interaction term are not met. 24 5.2 Do the Scientific Publication Process, Data, and Methodological Choices Affect the Results? 5.2.1 Model Estimated and Description of the Moderators In order to investigate systematic differences among the reported estimates, we select and present some key study characteristics (also called moderators ) that are likely to drive the results. We code 41 variables according to the following categories: transmission channels (see Section 3.3), mecha- nisms and interactive models (see Section 5.1), publication outlet, geography, model characteristics, conflict characteristics, measures of shocks, and model specifications. Descriptive statistics of the moderators are detailed in Table A3 in the Appendix. For the model specification, we follow Stanley and Doucouliagos (2012): 41 pcc ′ P CCij = β0 + β1 SEij + βk xij + εij (6) k=2 where P CC is the partial correlation coefficient between local income shocks and conflicts’ out- come of the ith estimate from the j th study, SE pcc denotes the standard error of the PCC, and εij ′ is the error term. The vector xij stands for a set of meta-independent variables capturing study- and regression-specific characteristics associated with the j th estimate as discussed in Section 5.1, with potential bearing on risks of conflicts. In other words, heterogeneity introduced and detailed below can be identified and quantified by the coefficients βk . Publication Outlet To some extent, differences in the publication process to which the studies were subjected may explain the heterogeneity of their estimates. We verify the validity of this concern through several controls. First, a careful peer review process can limit the degree of imprecision in the estimates of published studies. We consider two measures that can reasonably reflect a stricter publishing process: the impact score of academic journals (proxied by their SJR22 score), and publications in the top five journal in economics or political science. Second, academic journals specializing in conflict publish relatively more conflict-related work, including field studies with detailed contextual information. They may be more willing to publish insignificant and/or counter-intuitive estimates 22 Scimago Journal & Country Rank (SJR): https://www.scimagojr.com/journalrank.php. 25 if they are justified by a careful analysis of local contexts. Additionally, these unusual results may be reviewed by highly specialized peers, which limits the degree of imprecision in the estimates of published studies. Third, additional materials consisting of careful sensitivity and robustness checks can help identify significant relationships or, alternatively, mitigate them. Fourth, as the availability of disaggregated data and statistical methodologies have improved over the 2010s (Laville, 2019), we expect older studies to present, ceteris paribus, a lower degree of precision in their estimates. Geography A commonly held limitation of the conflict literature is that phenomena applying to one country or region hardly apply to the rest of the world since the formation of armed groups is often rooted in the long history of nation-building (de Mesquita, 1985; Cramer, 2002; Michalopoulos and Pa- paioannou, 2020). This point seems likely given our sample of studies: 55% of the selected estimates focus at least partially on a panel of African countries, and half of the selected articles focus on a single country (see Table A3 of the Appendices section). We introduce several geography covariates to control for this potential source of heterogeneity. Specifically, we include dummy variables to assess whether estimates from country case studies or works on specific geographic regions explain heterogeneity among results. Model Characteristics Some choices made in model construction may also explain heterogeneity. First, we test if estimator choices affect the results. Maximum-likelihood (MLE) and Ordinary Least Squares (OLS) are the two most common classes of estimators used by the literature. We create a dummy variable taking the value one if the estimator is from the MLE class (i.e., Logit, Negative binomial or Poisson pseudo maximum-likelihood), and zero otherwise. Second, inclusion of specific effects (including fixed and random effects) can increase the statistical power of the estimates by reducing unobserved heterogeneity. We create a dummy taking the value one when specific (fixed, random or mixed) effects are introduced in estimates, and zero otherwise. Third, clustering has implications on sta- tistical inference. We include a dummy variable indicating if estimates are clustered on the unit of reference. Fourth, we include a variable indicating whether the period of interest in the regression is before or after the end of the Cold War (i.e., after 1991). Finally, we include information on 26 the level of disaggregation of the estimates (regions, 0.5°x0.5° grid cells, 0.25°x0.25° grid cells, etc.). Several reviews of the literature suggest using small geographic units and quasi-experimental design to avoid biases arising from unobserved heterogeneity or competing mechanisms (Collier and Hoef- fler, 2007; Blattman and Miguel, 2010; Couttenier and Soubeyran, 2015). However, these gains for the identification strategy may also depend on the existence of subnational root causes of conflict and the degree of precision of the data at that scale (for a discussion, see Laville, 2019). To test for the influence of the size of the spatial unit on the results, we include a dummy variable indicating whether the estimates apply to grid cells of size less than or equal to 0.5°x0.5° at the equator (ap- prox. 55km Ö 55km), and the natural logarithm of the sample size. Conflict Characteristics We use several variables to control for heterogeneity in the estimates resulting from differences in the empirical definition of the conflict outcome. First, we control for the origin of the conflict data. Conflict studies often rely on data from the Uppsala Conflict Data Program (UCDP/PRIO) (Sundberg and Melander, 2013) or Armed Conflict Location and Event Data (ACLED) (Raleigh et al., 2010). The institutions behind these datasets have the capacity to collect daily data on the timing and location of violent events, while maintaining a consistent definition of violent phe- nomena (which supports the comparability of estimates across studies). However, their respective definition of violent phenomena and the potential omission, inflation or misrepresentation of events could systematically influence the literature due to their widespread use (Miller et al., 2022). We therefore include a dummy variable indicating if the regression relies on UCDP/PRIO or ACLED data. Second, we control for the type of conflict outcome by including three dummy variables in- dicating estimates of conflict incidence (if the outcome variable is the total number of conflict per day/month/year, or if it takes the value one when at least one conflict is observed), onset (if the out- come variable only takes the value one at the starting day/month/year of the conflict), or duration (if the outcome variable is the number of days, months or years of active conflict). Third, we include a dummy variable indicating if the spatial dimension of the dependent variable is local (below Ad- min1 level or at small grids level), or larger (above Admin2 level or at large grids level). Fourth, we include a dummy variable indicating if the conflict implied human casualties, indicating a relatively more intense form of conflict. Finally, we control for the type of conflict using dummy variables in- 27 dicating if the authors study armed conflicts, crimes, social unrest, and/or violence against civilians. Measures of Shocks How the authors measure income shocks may influence their results. First, we include a dummy variable indicating if the estimate uses a quasi-experimental design, namely if a measure of shock (e.g., prices variations, rainfall variation, floods, etc.) is interacted with a measure of resource endowment in the cell (e.g., production areas of coal, coffee intensity, oil and gas reserves, etc.). o and Dal B´ Second, following Dal B´ o (2011) and Dube and Vargas (2013), we include a dummy variable indicating if the shock specifically concerns a labor-intensive agricultural commodity.23 Agricultural shocks that do not concern the production of a labor-intensive good include general climatic shocks that could influence labor-intensive and capital-intensive productions simultane- ously, shocks that concern the livestock sector,24 shocks on consumer commodity prices, and shocks related to resource scarcity. Finally, we separate climatic shocks, price shocks and other types of shocks using dummy variables. Model Specifications Estimates can be sensitive to the choice of explanatory variables. In a sensitivity analysis of the conflict literature, Hegre and Sambanis (2006) identify a robust correlation between conflict onset and several variables including low per capita income, slow income growth, recent political insta- bility, large population size, and war-prone neighbors. As noted by Blattman and Miguel (2010), the inclusion of such correlates in conflict models may bias other estimates in unknown directions due to endogeneity or insufficient knowledge about the mechanisms involved. We control for the inclusion of these correlates (population size, GDP, past conflicts, and neighborhood conflicts) as 23 Unfortunately, extending this variable to the case of mineral resources proved to be more complicated. A first issue is that a resource can be labor-intensive in one country, but not in another one if the latter has the necessary infrastructure for a larger scale production. An adverse selection problem also arises since we depend on the author’s knowledge of the labor-intensity in the production of the commodity. For example, Corvalan and Pazzona (2019) analyze the links between copper market shocks and crime in Chile. Although they focus on a transmission channel passing through the local labor market, they do not specify the extent to which this production is labor intensive. Therefore, we chose to limit ourselves to the case of agricultural goods that involves less subjective choices in our coding strategy. Nevertheless, the intensity of labor in the production of mineral resources is central to understand how artisanal mining influences the risk of conflict and future MRA should address these issues. 24 Shocks in the livestock sector concern only a few regressions. We separate them so that the variable is homoge- neous. Indeed, although livestock farming requires labor force, it requires relatively little compared to agriculture. Also, its effects on the labor market are less direct because the creation or abandonment of an entire herd cannot be immediate. 28 control variables using dummy variables. 5.2.2 Estimation Technique and Results Ideally, we would include all our 41 moderators in Equation 6. However, it is likely that some ex- planatory variables will prove redundant, and including too many covariates relative to the number of studies increases the risk of false positive conclusions (Thompson and Higgins, 2002). Addi- tionally, including the wrong variables in the equation leads to misspecification bias and invalid inference. Therefore, a fundamental problem in estimating Equation 6 is model uncertainty asso- ciated with the variables to include. Two popular strategies employed in the literature to address model uncertainty are model selection and model averaging (Steel, 2020). Stepwise regression is the easiest and most commonly used approach for model selection. However, this method may erro- neously exclude important variables in sequential t -tests, and it does not account for the selection process when presenting the results of the final equation. To avoid results dependent on the se- lected model, we employ a model averaging technique, which takes into account all possible models. Following the recent literature on meta-analysis, we apply the Bayesian Model Averaging (BMA) approach to deal with model uncertainty (Havranek et al., 2017, 2018; Zigraiova et al., 2021). The goal of BMA is to find the best possible approximation of the distribution of regression parameters by running regressions based on different subsets of moderators. In our case, since we consider 41 variables, this yields 241 possible models to estimate. We therefore apply the Markov chain Monte Carlo algorithm, which approximates the model space and walks the part that contains the models with the highest posterior model probabilities (PMP), which measures the ‘goodness of fit’ of each model with the data. For each variable in the model, BMA reports three parameters: posterior mean, posterior standard deviation and posterior inclusion probability (PIP). PIP aggregates the PMPs of all the models in which the variable is included. A PIP above 0.5 is usually regarded as the threshold to include variables in the model (Jeffreys, 1961; Eicher et al., 2011). Figure 6 is a graphical representation of the BMA results for the whole sample of 2,464 regres- sions. The vertical axis lists the explanatory variables sorted by PIP in descending order. The horizontal axis is the PMP of each model sorted in ascending order. The blue (dark) color indicates the positive sign of the variable in the model, and the red color (light) denotes the negative sign of the variable. The blank cell suggests that the variable is excluded from the regression model. Figure 29 6 shows that nearly one-third of the variables are included in the best model and that their signs are robustly consistent across different models. Figure 7 is a graphical representation of our BMA results for AS+ (a), AS- (b), and HS+ (c). Approximately one-third of the moderators are included in the best models for agricultural shocks, compared to half of them for shocks in the extractive sector. However, fewer moderators are tested for HS+, which is most likely due to the smaller number of observations in this sub-sample. Overall, their effect’s sign also appears consistent across different models. We note that they are mainly negative for AS- studies, positive for HS+ studies, and nuanced for AS+ studies. Notes: The figure depicts the results of Bayesian Model Averaging. The explanatory variables are ranked according to their posterior inclusion probabilities from the highest on the top to the lowest at the bottom. The horizontal axis shows the values of cumulative posterior probability. Blue and red colours denote the positive and negative sign of the estimated parameter of explanatory variable, respectively. No colour means the corresponding explanatory variable is not included in the model. Numerical results are reported in Table 5. All variables are described in Table A3. Source: Authors’ compilation from MRA database. Figure 6: Model Inclusion in Bayesian Model Averaging [Total Sample] Table 5 presents the empirical results of BMA for the whole sample of 2,464 regressions. We also report the OLS results using the variables from BMA with PIP higher than 0.5. When interpreting the BMA results, we follows Jeffreys (1961), who considers that the value of PIP indicates a decisive effect if it exceeds 0.99, a strong effect if it is between 0.95 and 0.99, a positive effect if it is between 0.75 and 0.95, and a weak effect if it is between 0.5 and 0.75. The publication bias term in Table 5 is positive and statistically significant at the 0.10 level after controlling for our set of moderators. It confirms that the result obtained in Table 2 is not a spurious outcome caused by omitted variables 30 and confirms that the literature suffers from publication selection bias. Thirteen variables present a PIP higher than 0.5, indicating that they are relevant for explaining the differences in the estimates. We focus on the variables for which we have the most robust evidence across the two specifications: at least a strong PIP in BMA, and a significance level of at least 10% in OLS. As could be expected due to the meta-sample’s heterogeneity, studies testing theoretically detrimental mechanism or introducing interactions with a risk enhancer find a higher local risk of conflict subsequent to income shocks. Studies controlling for GDP find a higher risk of conflict and, conversely, controlling for the size of the population, using a quasi-experimental framework or studying small grid cells is associated with a smaller risk of conflict. Table 6 reports the results of the BMA and OLS for the sub-samples of AS+ (808 regressions), AS- (951 regressions), and HS+ (543 regressions) studies. The results we obtain for publication bias after controlling for our set of moderators confirms the presence of a substantial positive publication bias in the literature focusing on AS-, and the absence of evidence for such biases in the literature on AS+. However, Table 6 shows a negative and statistically significant coefficient of publication bias for HS+, which differs from the statistically non-significant constant found in Table 2. This indicates that HS+ studies not published recently in a top 5 journal suffer from a substantial negative publication selection bias when they focus on a small sample and countries from different regions, do not control spatially for conflicts or for specific effects, and do not study armed conflicts and price shocks. We find that AS+ studies specifically testing for stressful mechanisms25 or introducing interactions with a risk enhancer find a higher risk of conflict. They also find a higher risk of conflict when they go through a stricter publishing process or when they examine the post-Cold War period. Studies focusing on AS+ in Latin American and African countries find a generally lower risk of conflict. Studies encompassing negative agricultural shocks (AS-) approaching income shocks through climatic events or, to a lesser extent, price variations find a higher risk of conflict. They also find a stronger association between income shocks and local criminality (e.g., robberies, property trespassing, assaults). Finally, studies published in journals specialized in conflict analysis, controlling for the size of the population, or focusing on Latin American countries find a smaller risk of conflict. Studies focusing on positive hydrocarbon/mineral shocks (HS+) find a lower risk of conflict when they encompass larger samples or test for price shocks. 25 A few studies of AS+ test agitating mechanisms through rapacity effects. 31 (a) Positive Agricultural Shocks (AS+) (b) Negative Agricultural Shocks (AS-) (c) Positive Hydrocarbon and Mineral Shocks (HS+) Notes: The figure depicts the results of Bayesian Model Averaging. The explanatory variables are ranked according to their posterior inclusion probabilities from the highest on the top to the lowest at the bottom. The horizontal axis shows the values of cumulative posterior probability. Blue and red colours denote the positive and negative sign of the estimated parameter of explanatory variable, respectively. No colour means the corresponding explanatory variable is not included in the model. Numerical results are reported in Table 6. All variables are described in Table A3. Source: Authors’ compilation from MRA database. Figure 7: Model Inclusion in Bayesian Model Averaging [Sub-samples] 32 Table 5: Explaining Heterogeneity in the Estimated Effects of Income Shocks on Conflicts [BMA - Total Sample] Full sample BMA OLS Post Mean Post SD PIP Coef. SE p-Value (i) Baseline Publication bias (β 1) 0.789 NA 1.000 0.737 0.393 0.065 Precision (β 0) 0.000 0.002 0.052 -0.000 0.003 0.893 (ii) Transmission channels Positive agricultural shock (Ref.) Negative agricultural shock 0.001 0.002 0.127 - - - Positive hydrocarbon shock -0.002 0.002 0.492 - - - Other shock 0.000 0.001 0.096 - - - (iii) Mechanisms and interactive models Other mechanisms (Ref.) Mechanism stressful 0.007 0.001 0.994 0.006 0.002 0.008 Others (Ref.) Risk enhancer 0.004 0.000 1.000 0.004 0.002 0.093 (iv) Publication outlet No top 5 reviews (Ref.) Top 5 0.003 0.003 0.627 0.005 0.002 0.011 No conflict reviews (Ref.) Conflict review 0.000 0.000 0.046 - - - Core results (Ref.) Appendix results 0.000 0.000 0.024 - - - Continuous variables Age of study 0.000 0.000 0.079 - - - SJR score 0.000 0.000 0.418 - - - (v) Geography No country focus (Ref.) Country focus -0.009 0.002 0.992 -0.004 0.003 0.228 Worldwide (Ref.) Region: Africa -0.001 0.002 0.127 - - - Region: East Asia Pacific (EAP) 0.002 0.003 0.457 - - - Region: Latin America & the Caribbean (LAC) -0.003 0.003 0.548 -0.005 0.002 0.025 Region: South Asia -0.019 0.005 0.995 -0.020 0.012 0.089 (vi) Model characteristics No MLE (Ref.) Maximum likelihood estimator 0.000 0.001 0.060 - - - No specific effects (Ref.) Specific effects 0.000 0.000 0.016 - - - No cluster unit reference (Ref.) Cluster unit reference 0.000 0.001 0.221 - - - Pre cold war (Ref.) Post cold war 0.004 0.001 0.994 0.002 0.002 0.268 No small grids (Ref.) Small grids -0.008 0.002 1.000 -0.006 0.003 0.043 Continuous variable Sample size 0.000 0.000 0.140 - - - (vii) Conflict characteristics Usual conflict data (Ref.) Unusual conflict data 0.000 0.001 0.042 - - - Conflict incidence (Ref.) Conflict duration 0.000 0.000 0.103 - - - Conflict onset 0.000 0.000 0.028 - - - No local conflicts (Ref.) Conflict local 0.001 0.002 0.140 - - - Non lethal (Ref.) Lethal 0.000 0.001 0.087 - - - Others (Ref.) Armed conflicts 0.000 0.000 0.018 - - - Crimes 0.000 0.001 0.089 - - - Social unrests 0.000 0.001 0.062 - - - Violence against citizens 0.000 0.000 0.020 - - - (viii) Measures of shocks No quasi experimental (Ref.) Quasi experimental -0.006 0.001 1.000 -0.005 0.002 0.028 No labor intensive (Ref.) Labor intensive resources 0.000 0.000 0.014 - - - Others (Ref.) Climatic shock 0.003 0.001 0.809 0.003 0.002 0.096 Price shock 0.000 0.001 0.133 - - - (ix) Model specification No past conflict (Ref.) Control for past conflicts 0.000 0.001 0.139 - - - No spatial factor (Ref.) Control for conflict spatial factor -0.006 0.002 0.929 - - - No population (Ref.) Control for population -0.003 0.000 1.000 -0.003 0.002 0.036 No GDP (Ref.) Control for GDP 0.007 0.001 1.000 0.004 0.002 0.042 Notes: All models are estimated on the whole sample of 2,464 observations and 64 studies focusing on positive agricultural shocks, with the weight being the inverse of SE pcc . The dependent variable is the partial correlation between shocks and conflicts. SD = standard deviation. SE = standard error. PIP = posterior inclusion probability. Bayesian model averaging (BMA) employs the unit information and the dilution prior (UIP g-prior; uniform model prior on the BMS package by Zeugner and Feldkircher (2015)) suggested by George (2010). The frequentist check (OLS) includes only explanatory variables with a PIP above 50 percent (≥ 0.50) in the BMA and is estimated using clustered standard errors (cluster at the study level). In bold, explanatory variable with a PIP above 50 percent (≥ 0.50) (BMA) and/or statistically significant at 10 percent or less (OLS). Source: Authors’ compilation from MRA database. 33 Table 6: Explaining Heterogeneity in the Estimated Effects of Income Shocks on Conflicts [BMA - Sub-Samples] Positive Agricultural Shocks (AS+) Negative Agricultural Shocks (AS-) Positive Hydrocarbon and Mineral Shocks (HS+) BMA OLS BMA OLS BMA OLS Post Mean Post SD PIP Coef. SE p-Value Post Mean Post SD PIP Coef. SE p-Value Post Mean Post SD PIP Coef. SE p-Value (i) Baseline Publication bias (β 1) -0.482 NA 1.000 -0.521 0.312 0.108 1.629 NA 1.000 1.693 0.262 0.000 -3.283 NA 1.000 -2.870 0.718 0.002 Precision (β 0) 0.009 0.013 0.374 0.022 0.007 0.004 0.000 0.002 0.047 -0.001 0.003 0.815 0.133 0.078 0.846 0.109 0.039 0.014 (ii) Mechanisms and interactive models Other mechanisms (Ref.) Mechanism stressful 0.045 0.005 1.000 0.045 0.005 0.000 0.000 0.001 0.037 - - - 0.001 0.006 0.071 - - - Others (Ref.) Risk enhancer 0.014 0.001 1.000 0.014 0.005 0.008 0.002 0.000 1.000 0.002 0.002 0.155 0.002 0.001 0.798 0.002 0.002 0.315 (iii) Publication outlet No top 5 reviews (Ref.) Top 5 -0.001 0.004 0.158 - - - 0.000 0.006 0.091 - - - -0.032 0.019 0.923 -0.027 0.005 0.000 No conflict reviews (Ref.) Conflict review 0.002 0.005 0.150 - - - -0.017 0.004 0.998 -0.016 0.003 0.000 0.001 0.004 0.158 - - - Core results (Ref.) Appendix results 0.000 0.000 0.019 - - - 0.000 0.000 0.026 - - - 0.001 0.005 0.081 - - - Continuous variables Age of study 0.000 0.000 0.407 - - - 0.000 0.000 0.082 - - - 0.008 0.005 0.820 0.008 0.001 0.000 SJR score 0.001 0.000 0.972 0.001 0.000 0.000 -0.001 0.000 0.931 -0.001 0.000 0.003 0.000 0.001 0.302 (iv) Geography No country focus (Ref.) Country focus 0.000 0.002 0.038 - - - 0.000 0.001 0.058 - - - -0.057 0.032 0.866 -0.061 0.012 0.000 Worldwide (Ref.) Region: Africa -0.028 0.009 0.993 -0.034 0.006 0.000 -0.002 0.002 0.755 -0.003 0.001 0.001 0.008 0.043 0.368 - - - Region: EAP 0.002 0.007 0.078 - - - 0.000 0.002 0.082 - - - 0.059 0.058 0.708 0.056 0.014 0.001 Region: LAC -0.033 0.010 0.988 -0.036 0.005 0.000 -0.018 0.005 0.974 -0.020 0.005 0.001 0.035 0.050 0.628 0.035 0.004 0.000 Region: South Asia -0.021 0.012 0.859 -0.025 0.007 0.003 -0.041 0.016 0.938 -0.046 0.012 0.000 0.081 0.058 0.753 0.101 0.015 0.000 (v) Model characteristics No MLE (Ref.) Maximum likelihood estimator 0.000 0.000 0.021 - - - 0.000 0.000 0.035 - - - 0.000 0.002 0.058 - - - No specific effects (Ref.) Specific effects 0.003 0.005 0.260 - - - 0.000 0.000 0.034 - - - 0.047 0.055 0.520 0.048 0.021 0.042 Others (Ref.) Cluster unit reference 0.000 0.000 0.017 - - - 0.000 0.000 0.031 - - - 0.000 0.013 0.155 - - - Pre cold war (Ref.) Post cold war 0.006 0.001 1.000 0.006 0.003 0.095 0.000 0.001 0.072 - - - - - - - - - No small grids (Ref.) Small grids 0.000 0.001 0.052 - - - 0.000 0.001 0.041 - - - - - - - - - Continuous variable Sample size 0.000 0.000 0.064 - - - 0.000 0.000 0.043 - - - -0.011 0.003 1.000 -0.009 0.002 0.001 (vi) Conflict characteristics Usual conflict data (Ref.) Unusual conflict data 0.000 0.002 0.079 - - - 0.000 0.000 0.041 - - - - - - - - - Conflict incidence (Ref.) Conflict duration - - - - - - 0.000 0.000 0.319 - - - - - - - - - Conflict onset 0.000 0.001 0.073 - - - 0.000 0.000 0.097 - - - - - - - - - No local conflicts (Ref.) Conflict local 0.000 0.001 0.044 - - - 0.000 0.001 0.096 - - - 0.003 0.011 0.242 - - - Non lethal (Ref.) Lethal 0.000 0.001 0.058 - - - 0.001 0.002 0.210 - - - -0.001 0.002 0.134 - - - Others (Ref.) Armed conflicts 0.005 0.002 0.930 0.004 0.001 0.002 0.001 0.002 0.307 - - - -0.001 0.001 0.595 -0.002 0.001 0.007 Crimes 0.000 0.001 0.047 - - - 0.036 0.004 1.000 0.035 0.005 0.000 0.000 0.002 0.034 - - - Social unrests -0.006 0.007 0.522 -0.010 0.006 0.088 0.000 0.001 0.103 - - - 0.000 0.001 0.057 - - - Violence against citizens 0.000 0.001 0.023 - - - 0.000 0.000 0.026 - - - 0.000 0.002 0.065 - - - (vii) Measures of shocks No quasi experimental (Ref.) Quasi experimental -0.003 0.003 0.545 -0.005 0.002 0.034 -0.004 0.001 1.000 -0.004 0.002 0.105 0.009 0.014 0.377 - - - No labor intensive (Ref.) Labor intensive resources 0.000 0.002 0.042 - - - -0.005 0.001 1.000 -0.005 0.003 0.104 - - - - - - Others (Ref.) Climatic shock 0.000 0.001 0.030 - - - 0.006 0.001 1.000 0.007 0.002 0.000 - - - - - - Price shock 0.000 0.001 0.038 - - - 0.016 0.005 0.968 0.018 0.004 0.000 -0.081 0.037 0.969 -0.063 0.017 0.003 (viii) Model specification No past conflict (Ref.) Control for past conflicts 0.000 0.000 0.024 - - - 0.000 0.000 0.025 - - - -0.001 0.005 0.100 - - - No spatial factor (Ref.) Control for conflict spatial factor 0.000 0.001 0.030 - - - 0.000 0.001 0.063 - - - 0.017 0.005 0.995 0.016 0.010 0.118 No population (Ref.) Control for population 0.001 0.002 0.270 - - - -0.002 0.000 1.000 -0.002 0.001 0.088 0.000 0.007 0.072 - - - No GDP (Ref.) Control for GDP 0.000 0.001 0.100 - - - 0.000 0.000 0.055 - - - -0.004 0.014 0.090 - - - Notes: AS+ sample: 808 observations and 25 studies focusing on positive agricultural shocks; AS- sample: 951 observations and 34 studies focusing on negative agricultural shocks; HS+ sample: 543 observations and 14 studies focusing on positive hydrocarbon/mineral shocks. The weights are the inverse of SE pcc . The dependent variable is the partial correlation between the type of shock and conflicts. SD = standard deviation. SE = standard error. PIP = posterior inclusion probability. Bayesian model averaging (BMA) employs the unit information and the dilution prior (UIP g-prior; uniform model prior on the BMS package by Zeugner and Feldkircher (2015)) suggested by George (2010). The frequentist check (OLS) includes only explanatory variables with a PIP above 50 percent (≥ 0.50) in the BMA and is estimated using clustered standard errors (cluster at the study level). In bold, explanatory variable with a PIP above 50 percent (≥ 0.50) (BMA) and/or statistically significant at 10 percent or less (OLS). Source: Authors’ compilation from MRA database. 34 6 Discussion and Conclusion Do income shocks locally affect the probability of internal conflicts? This paper uses meta-regression analysis (MRA) to review the recent empirical literature on local income shocks and conflicts in developing countries. Using a sample of 2,464 point estimates from 64 studies, we evaluate the pres- ence of publication selection bias and discuss the relevance of a range of methodological limitations and advances reported in previous reviews of the conflict literature. We find no evidence of an average and unconditional genuine effect of income-related shocks on the risk of conflict at the infra-national scale. Therefore, the effect of climate changes, commodity price fluctuations and natural resources endowment on the risk of conflict does not seem to sys- tematically transit through variations in local incomes and economic prospects. Nonetheless, the fluctuation of local incomes may affect the risk of conflict under certain conditions. To account for heterogeneity in the authors’ narratives, we divide the meta-sample into coherent sets based on the type (agricultural, extractive, or other) and direction (wealth-increasing or -decreasing) of the shock. For these sub-samples, we find a positive and statistically significant effect of wealth- increasing shock in the agriculture sector on the risk of conflict, but no statistically significant effects of wealth-decreasing shocks or shocks in the extractive sector. Moreover, we find that the literature suffers substantially from two types of publication selection bias: it favors the publication of results with higher statistical significance (Type II), and showing the detrimental effects of income shocks (Type I). Studies of negative agricultural shocks (droughts, floods) are particularly affected by the latter type of bias, probably because 99.5% of the estimates in this subgroup test theoretically stressful mechanisms. We also suspect the influence of result searching bias in the choice of climate- related variables, since we find that overestimation is more likely for studies focusing on climate shocks. Indeed, climatic variables are intricate (Auffhammer et al., 2013) and recent micro-evidence suggests that their effect on local poverty is heterogeneous (Azzarri and Signorelli, 2020). Several reviews of the literature advocate for the use of highly disaggregated spatial units in conflict models based on methodological considerations (Collier and Hoeffler, 2007; Blattman and Miguel, 2010; Couttenier and Soubeyran, 2015) or study-specific robustness checks (Berman and Couttenier, 2015). The results of this MRA, and more precisely the heterogeneities captured with the BMA, suggest that quasi-experimental models and small grid cells may indeed palliate the 35 overestimation tendency we find in the literature on income shocks as a whole. However, these moderators never strongly influence the estimates when we divide the literature into coherent sets. When decomposing for the nature of main income shocks (AS+, AS-, HS+), the BMA technique enables us to precisely detect how complex heterogeneities — related to (i) the publication outlet, (ii) geographical area encompassed, (iii) model characteristics, (iv) the nature of conflicts, (v) measurements of shocks, and (vi) covariates introduced — affect the results from the literature on the risks of conflicts. This MRA complements the existing body of meta-analysis on the economic origins of conflict in developing countries. O’Brochta (2019) finds no aggregate relationship between conflicts and natural resources, and Blair et al. (2021) obtain similar conclusions for commodity prices. Our results complement theirs by showing that there is no aggregated effect of income shocks on the risk of conflict at the infra-national scale. Vesco et al. (2020) find that both extractive resource wealth and renewable resource scarcity (e.g., forest, water, vegetation) increase the risk of conflict. When decomposing for the nature of main income shocks (AS+, AS-, HS+), our results suggest that wealth-increasing and -decreasing shocks do not have a symmetric effect on conflict, backing supporters of non-wealth-related factors (including grievances and state capacity) as motives of conflicts (de Mesquita, 1985; Cramer, 2002; Michalopoulos and Papaioannou, 2020). Hsiang et al. (2013, 2014) find that deviations from normal precipitation and mild temperatures systematically increase the risk of conflict. In line with their conclusions and several scholars (Theisen et al., 2012; Sarsons, 2015), our results suggest that the effect of climatic events is unlikely to be operating solely through changes in local agricultural income. Finally, our results show that researchers’ expectations on the theoretical mechanisms at play (above all, opportunity cost and rapacity) distort statistical inference and the resulting understanding of research on the local causes of conflict. Overall, this MRA advocates for caution in generalizing the results of conflict studies, especially for recommending policies to countries outside the study area. The genuine effect of positive agricultural shocks and the asymmetries we find in the literature call for more analysis of the peacemaking mechanisms linking income shocks to local risk of conflict. 36 References Abadie, Alberto and Javier Gardeazabal. 2003. “The Economic Costs of Conflict: A Case Study of the Basque Country”. American Economic Review 93 (1), 113–132. ı. 2021. “Income Shocks and Conflict: Evidence from Abidoye, Babatunde and Massimiliano Cal` Nigeria”. Journal of African Economies 30 (5), 478–507. Angrist, Joshua D. and Adriana D. Kugler. 2008. “Rural Windfall or a New Resource Curse? Coca, Income, and Civil Conflict in Colombia”. Review of Economics and Statistics 90 (2), 191–215. Arezki, Rabah and Markus Brueckner. 2014. “Effects of International Food Price Shocks on Political Institutions in Low-Income Countries: Evidence from an International Food Net-Export Price Index”. World Development 61, 142–153. Auffhammer, Maximilian, Solomon M. Hsiang, Wolfram Schlenker, and Adam Sobel. 2013. “Using Weather Data and Climate Model Output in Economic Analyses of Climate Change”. Review of Environmental Economics and Policy 7 (2), 181–198. Axinn, William G., Dirgha Ghimire, and Nathalie E. Williams. 2012. “Collecting Survey Data during Armed Conflict”. Journal of Official Statistics 28 (2), 153–171. Azzarri, Carlo and Sara Signorelli. 2020. “Climate and poverty in Africa South of the Sahara”. World Development 125, 104691. e Tapsoba. 2020. “Inflation Targeting: Genuine Balima, Hippolyte W., Eric G. Kilama, and Ren´ Effects or Publication Selection Bias?”. European Economic Review 128, 103520. Bazzi, Samuel and Christopher Blattman. 2014. “Economic Shocks and Conflict: Evidence from Commodity Prices”. American Economic Journal: Macroeconomics 6 (4), 1–38. Bellemare, Marc F.. 2015. “Rising Food Prices, Food Price Volatility, and Social Unrest”. American Journal of Agricultural Economics 97 (1), 1–21. Berman, Nicolas and Mathieu Couttenier. 2015. “External Shocks, Internal Shots: The Geography of Civil Conflicts”. Review of Economics and Statistics 97 (4), 758–776. 37 Berman, Nicolas, Mathieu Couttenier, Dominic Rohner, and Mathias Thoenig. 2017. “This Mine is Mine! How Minerals Fuel Conflicts in Africa”. American Economic Review 107 (6), 1564–1610. Blair, Graeme, Darin Christensen, and Aaron Rudkin. 2021. “Do Commodity Price Shocks Cause Armed Conflict? A Meta-Analysis of Natural Experiments”. American Political Science Re- view 115 (2), 709–716. Blattman, Christopher and Edward Miguel. 2010. “Civil War”. Journal of Economic Litera- ture 48 (1), 3–57. Buhaug, Halvard, Tor A Benjaminsen, Espen Sjaastad, and Ole Magnus Theisen. 2015. “Climate variability, food production shocks, and violent conflict in Sub-Saharan Africa”. Environmental Research Letters 10 (12), 125015. Bush, Ray. 2010. “Food Riots: Poverty, Power and Protest”. Journal of Agrarian Change 10 (1), 119–129. Card, David and Alan B. Krueger. 1995. “Time-Series Minimum-Wage Studies: A Meta-analysis”. The American Economic Review 85 (2), 238–243. Collier, P. 1998. “On Economic Causes of Civil War”. Oxford Economic Papers 50 (4), 563–573. Collier, Paul (Ed.). 2003. Breaking the conflict trap: civil war and development policy. A World Bank policy research report. Washington, DC : [New York]: World Bank ; Oxford University Press. Collier, Paul and Anke Hoeffler. 2007. “Chapter 23 Civil War”. In Handbook of Defense Economics, Volume 2, pp. 711–739. Elsevier. Collier, P., A. Hoeffler, and D. Rohner. 2008. “Beyond Greed and Grievance: Feasibility and Civil War”. Oxford Economic Papers 61 (1), 1–27. Corral, Paul, Alexander Irwin, Nandini Krishnan, Daniel Gerszon Mahler, and Tara Vishwanath. 2020. Fragility and Conflict: On the Front Lines of the Fight against Poverty. The World Bank. Corvalan, Alejandro and Matteo Pazzona. 2019. “Persistent Commodity Shocks and Transitory Crime Effects”. Journal of Economic Behavior & Organization 158, 110–127. 38 Couttenier, Mathieu and Raphael Soubeyran. 2015. “A Survey of the Causes of Civil Conflicts: Natural Factors and Economic Conditions”. Revue d’economie politique 125 (6), 787–810. Cramer, C. 2002. “Homo Economicus Goes to War: Methodological Individualism, Rational Choice and the Political Economy of War”. World Development 30 (11), 1845–1864. Crost, Benjamin and Joseph H Felter. 2020. “Export Crops and Civil Conflict”. Journal of the European Economic Association 18 (3), 1484–1520. o, Ernesto and Pedro Dal B´ Dal B´ o. 2011. “Workers, Warriors, and Criminals: Social Conflict in General Equilibrium”. Journal of the European Economic Association 9 (4), 646–677. de Mesquita, Bruce Bueno. 1985. “Toward a Scientific Understanding of International Conflict: A Personal View”. International Studies Quarterly 29 (2), 121. Dixon, Jeffrey. 2009. “What Causes Civil Wars? Integrating Quantitative Research Findings”. International Studies Review 11 (4), 707–735. Doucouliagos, Chris and T.d. Stanley. 2013. “Are All Economic Facts Greatly Exaggerated? Theory Competition and Selectivity”. Journal of Economic Surveys 27 (2), 316–339. Doucouliagos, Hristos and Patrice Laroche. 2009. “Unions and Profits: A Meta-Regression Analy- sis”. Industrial Relations: A Journal of Economy and Society 48 (1), 146–184. Doucouliagos, Hristos and T. D. Stanley. 2009. “Publication Selection Bias in Minimum-Wage Research? A Meta-Regression Analysis”. British Journal of Industrial Relations 47 (2), 406–428. Dube, Oeindrila and Juan F. Vargas. 2013. “Commodity Price Shocks and Civil Conflict: Evidence from Colombia”. The Review of Economic Studies 80 (4), 1384–1421. Egger, M., G. D. Smith, M. Schneider, and C. Minder. 1997. “Bias in Meta-Analysis Detected by a Simple Graphical Test”. BMJ 315 (7109), 629–634. Eicher, Theo S., Chris Papageorgiou, and Adrian E. Raftery. 2011. “Default priors and predictive performance in Bayesian model averaging, with application to growth determinants”. Journal of Applied Econometrics 26 (1), 30–55. 39 Fearon, James D.. 2005. “Primary Commodity Exports and Civil War”. Journal of Conflict Resolution 49 (4), 483–507. Fearon, James D. and David D. Laitin. 2003. “Ethnicity, Insurgency, and Civil War”. American Political Science Review 97 (01), 75–90. Fjelde, Hanne. 2015. “Farming or Fighting? Agricultural Price Shocks and Civil War in Africa”. World Development 67, 525–534. George, Edward I.. 2010. “Dilution priors: Compensating for model space redundancy”. In In- stitute of Mathematical Statistics Collections, pp. 158–165. Beachwood, Ohio, USA: Institute of Mathematical Statistics. Grossman, Herschell I.. 1991. “A General Equilibrium Model of Insurrections”. The American Economic Review 81 (4), 912–921. Havranek, Tomas, Zuzana Irsova, and Olesia Zeynalova. 2018. “Tuition Fees and University Enrol- ment: A Meta-Regression Analysis”. Oxford Bulletin of Economics and Statistics 80 (6), 1145– 1184. Havranek, Tomas, Marek Rusnak, and Anna Sokolova. 2017. “Habit formation in consumption: A meta-analysis”. European Economic Review 95, 142–167. avard and Nicholas Sambanis. 2006. “Sensitivity Analysis of Empirical Results on Civil Hegre, H˚ War Onset”. Journal of Conflict Resolution 50 (4), 508–535. Hendrix, Cullen S and Stephan Haggard. 2015. “Global food prices, regime type, and urban unrest in the developing world”. Journal of Peace Research 52 (2), 143–157. Hirshleifer, Jack. 1995. “Anarchy and its Breakdown”. Journal of Political Economy 103 (1), 26–52. Hsiang, Solomon M., Marshall Burke, and Edward Miguel. 2013. “Quantifying the Influence of Climate on Human Conflict”. Science . Hsiang, Solomon M., Marshall Burke, and Edward Miguel. 2014. “Reconciling Climate-Conflict Meta-Analyses: reply to Buhaug et al.”. Climatic Change 127 (3-4), 399–405. 40 Ioannidis, John P. A.. 2005. “Why Most Published Research Findings Are False”. PLoS Medicine 2 (8), e124. Ioannidis, John P. A., T. D. Stanley, and Hristos Doucouliagos. 2017. “The Power of Bias in Economics Research”. The Economic Journal 127 (605), F236–F265. Islam, Asadul, Chandarany Ouch, Russell Smyth, and Liang Choon Wang. 2016. “The Long- Term Effects of Civil Conflicts on Education, Earnings, and Fertility: Evidence from Cambodia”. Journal of Comparative Economics 44 (3), 800–820. Jeffreys, Harold. 1961. Theory of probability. Oxford classic texts in the physical sciences. Oxford [Oxfordshire] : New York: Oxford University Press. Jia, Ruixue. 2014. “Weather Shocks, Sweet Potatoes and Peasant Revolts in Historical China”. The Economic Journal 124 (575), 92–118. Lacina, Bethany. 2006. “Explaining the Severity of Civil Wars”. Journal of Conflict Resolu- tion 50 (2), 276–289. Lacina, Bethany and Nils Petter Gleditsch. 2005. “Monitoring Trends in Global Combat: A New Dataset of Battle Deaths”. eenne de European Journal of Population / Revue europ´ emographie 21 (2-3), 145–166. D´ ees R´ Laville, Camille. 2019. “Les Causes des Conflits Civils : Avanc´ a l’Aide de Donn´ ecentes ` ees esagr´ D´ ees”. Revue fran¸ eg´ economie XXXIV (3), 123–165. caise d’´ Lipsey, Mark W. and David B. Wilson. 2001. Practical meta-analysis. Practical meta-analysis. Thousand Oaks, CA, US: Sage Publications, Inc. Pages: ix, 247. Martin-Shields, Charles P. and Wolfgang Stojetz. 2019. “Food security and conflict: Empirical challenges and future opportunities for research and policy making on food security and conflict”. World Development 119, 150–164. cois, Giacomo De Luca, Petros G. Sekeris, and John Ulimwengu. 2014. “Mineral Maystadt, Jean-Fran¸ Resources and Conflicts in DRC: a Case of Ecological Fallacy?”. Oxford Economic Papers 66 (3), 721–749. 41 McGuirk, Eoin and Marshall Burke. 2020. “The Economic Origins of Conflict in Africa”. Journal of Political Economy 128 (10), 3940–3997. Michalopoulos, Stelios and Elias Papaioannou. 2020. “Historical Legacies and African Develop- ment”. Journal of Economic Literature 58 (1), 53–128. Miller, Erin, Roudabeh Kishi, Clionadh Raleigh, and Caitriona Dowd. 2022. “An agenda for ad- dressing bias in conflict data”. Scientific Data 9 (1), 593. Morelli, Massimo and Dominic Rohner. 2015. “Resource Concentration and Civil Wars”. Journal of Development Economics 117, 32–47. O’Brochta, William. 2019. “A Meta-Analysis of Natural Resources and Conflict”. Research & Politics 6 (1), 205316801881823. Page, Matthew J, Joanne E McKenzie, Patrick M Bossuyt, Isabelle Boutron, Tammy C Hoffmann, Cynthia D Mulrow, Larissa Shamseer, Jennifer M Tetzlaff, Elie A Akl, Sue E Brennan, Roger objartsson, Manoj M Lalu, Tianjing Li, Chou, Julie Glanville, Jeremy M Grimshaw, Asbjørn Hr´ Elizabeth W Loder, Evan Mayo-Wilson, Steve McDonald, Luke A McGuinness, Lesley A Stewart, James Thomas, Andrea C Tricco, Vivian A Welch, Penny Whiting, and David Moher. 2021. “The PRISMA 2020 statement: an updated guideline for reporting systematic reviews”. BMJ , n71. avard Hegre, and Joakim Karlsen. 2010. “Introducing ACLED: Raleigh, Clionadh, Andrew Linke, H˚ An Armed Conflict Location and Event Dataset: Special Data Feature”. Journal of Peace Re- search 47 (5), 651–660. Reuveny, Rafael and John W. Maxwell. 2001. “Conflict and Renewable Resources”. Journal of Conflict Resolution 45 (6), 719–742. Ross, Michael L.. 2004. “What Do We Know about Natural Resources and Civil War?”. Journal of Peace Research 41 (3), 337–356. Sarsons, Heather. 2015. “Rainfall and Conflict: a Cautionary Tale”. Journal of Development Economics 115, 62–72. 42 Stanley, T.D. and Hristos Doucouliagos. 2012. Meta-Regression Analysis in Economics and Business (0 ed.). Routledge. Stanley, T. D.. 2007. “Meta-Regression Methods for Detecting and Estimating Empirical Effects in the Presence of Publication Selection”. Oxford Bulletin of Economics and Statistics 0 (0). Stanley, T. D., Hristos Doucouliagos, and John P. A. Ioannidis. 2017. “Finding the Power to Reduce Publication Bias”. Statistics in Medicine . Steel, Mark F. J.. 2020. “Model Averaging and Its Use in Economics”. Journal of Economic Literature 58 (3), 644–719. Sundberg, Ralph and Erik Melander. 2013. “Introducing the UCDP Georeferenced Event Dataset”. Journal of Peace Research 50 (4), 523–532. anchez de la Sierra, Ra´ S´ ul. 2020. “On the Origins of the State: Stationary Bandits and Taxation in Eastern Congo”. Journal of Political Economy 128 (1), 32–74. Theisen, Ole Magnus. 2012. “Climate clashes? Weather variability, land pressure, and organized violence in Kenya, 1989–2004”. Journal of Peace Research 49 (1), 81–96. Theisen, Ole Magnus, Helge Holtermann, and Halvard Buhaug. 2012. “Climate Wars? Assessing the Claim That Drought Breeds Conflict”. International Security 36 (3), 79–106. Thompson, Simon G. and Julian P. T. Higgins. 2002. “How Should Meta-Regression Analyses be Undertaken and Interpreted?”. Statistics in Medicine 21 (11), 1559–1573. Thompson, Simon G. and Stephen J. Sharp. 1999. “Explaining Heterogeneity in Meta-Analysis: a Comparison of Methods”. Statistics in Medicine 18 (20), 2693–2708. Vesco, Paola, Shouro Dasgupta, Enrica De Cian, and Carlo Carraro. 2020. “Natural Resources and Conflict: A Meta-Analysis of the Empirical Literature”. Ecological Economics 172, 106633. Viechtbauer, Wolfgang and Mike W.-L. Cheung. 2010. “Outlier and influence diagnostics for meta- analysis”. Research Synthesis Methods 1 (2), 112–125. von Uexkull, Nina. 2014. “Sustained Drought, Vulnerability and Civil Conflict in Sub-Saharan Africa”. Political Geography 43, 16–26. 43 Ward, Michael D, Brian D Greenhill, and Kristin M Bakke. 2010. “The Perils of Policy by P-value: Predicting Civil Conflicts”. Journal of Peace Research 47 (4), 363–375. Wickham, Hadley. 2022. rvest: Easily Harvest (Scrape) Web Pages. World Bank. 2020. Poverty and Shared Prosperity 2020 : Reversals of Fortune. Washington, DC: World Bank. Zeugner, Stefan and Martin Feldkircher. 2015. “Bayesian Model Averaging Employing Fixed and Flexible Priors: The BMS Package for R”. Journal of Statistical Software 68 (4). Zigraiova, Diana, Tomas Havranek, Zuzana Irsova, and Jiri Novak. 2021. “How puzzling is the forward premium puzzle? A meta-analysis”. European Economic Review 134, 103714. Meta Sample ı. 2021. “Income Shocks and Conflict: Evidence from Abidoye, Babatunde and Massimiliano Cal` Nigeria”. Journal of African Economies 30 (5), 478–507. Acharya, Avidit, Robin Harding, and J. Andrew Harris. 2020. “Security in the Absence of a State: Traditional Authority, Livestock Trading, and Maritime Piracy in Northern Somalia”. Journal of Theoretical Politics 32 (4), 497–537. Ahrens, Achim. 2015. “Civil Conflicts, Economic Shocks and Night-time Lights”. Peace Economics, Peace Science and Public Policy 21 (4), 433–444. er´ Almer, Christian, J´ emy Laurent-Lucchetti, and Manuel Oechslin. 2017. “Water Scarcity and Ri- oting: Disaggregated Evidence from Sub-Saharan Africa”. Journal of Environmental Economics and Management 86, 193–209. Bagozzi, Benjamin E., Ore Koren, and Bumba Mukherjee. 2017. “Droughts, Land Appropriation, and Rebel Violence in the Developing World”. The Journal of Politics 79 (3), 1057–1072. Bai, Ying and James Kai-sing Kung. 2011. “Climate Shocks and Sino-nomadic Conflict”. Review of Economics and Statistics 93 (3), 970–981. 44 Berman, Nicolas and Mathieu Couttenier. 2015. “External Shocks, Internal Shots: The Geography of Civil Conflicts”. Review of Economics and Statistics 97 (4), 758–776. Berman, Nicolas, Mathieu Couttenier, Dominic Rohner, and Mathias Thoenig. 2017. “This Mine is Mine! How Minerals Fuel Conflicts in Africa”. American Economic Review 107 (6), 1564–1610. Bhavnani, Rikhil R. and Bethany Lacina. 2015. “The Effects of Weather-Induced Migration on Sons of the Soil Riots in India”. World Politics 67 (4), 760–794. Bohlken, Anjali Thomas and Ernest John Sergenti. 2010. “Economic Growth and Ethnic Violence: an Empirical Investigation of Hindu—Muslim Riots in India”. Journal of Peace Research 47 (5), 589–600. Bollfrass, Alexander and Andrew Shaver. 2015. “The Effects of Temperature on Political Violence: Global Evidence at the Subnational Level”. PLOS ONE 10 (5), e0123505. Buhaug, Halvard, Mihai Croicu, Hanne Fjelde, and Nina von Uexkull. 2021. “A Conditional Model of Local Income Shock and Civil Conflict”. The Journal of Politics 83 (1), 354–366. Carreri, Maria and Oeindrila Dube. 2017. “Do Natural Resources Influence Who Comes to Power, and How?”. The Journal of Politics 79 (2), 502–518. Caruso, Raul, Ilaria Petrarca, and Roberto Ricciuti. 2016. “Climate Change, Rice Crops, and Violence: Evidence from Indonesia”. Journal of Peace Research 53 (1), 66–83. Christensen, Darin. 2019. “Concession Stands: How Mining Investments Incite Protest in Africa”. International Organization 73 (1), 65–101. Christensen, Darin, Mai Nguyen, and Renard Sexton. 2019. “Strategic Violence during Democrati- zation: Evidence from Myanmar”. World Politics 71 (2), 332–366. Corvalan, Alejandro and Matteo Pazzona. 2019. “Persistent Commodity Shocks and Transitory Crime Effects”. Journal of Economic Behavior & Organization 158, 110–127. Crost, Benjamin and Joseph H Felter. 2020. “Export Crops and Civil Conflict”. Journal of the European Economic Association 18 (3), 1484–1520. 45 cois Maystadt. 2018. “Violence, Selection Dagnelie, Olivier, Giacomo Davide De Luca, and Jean-Fran¸ and Infant Mortality in Congo”. Journal of Health Economics 59, 153–177. De Juan, Alexander. 2015. “Long-Term Environmental Change and Geographical Patterns of Vio- lence in Darfur, 2003–2005”. Political Geography 45, 22–33. Detges, Adrien. 2016. “Local Conditions of Drought-Related Violence in Sub-Saharan Africa: the Role of Road and Water Infrastructures”. Journal of Peace Research 53 (5), 696–710. ıa-Ponce, and Kevin Thom. 2016. “From Maize to Haze: Agricul- Dube, Oeindrila, Omar Garc´ tural Shocks and the Growth of the Mexican Drug Sector”. Journal of the European Economic Association 14 (5), 1181–1224. Dube, Oeindrila and Juan F. Vargas. 2013. “Commodity Price Shocks and Civil Conflict: Evidence from Colombia”. The Review of Economic Studies 80 (4), 1384–1421. oring, Stefan. 2020. “Come Rain, or come Wells: How Access to Groundwater Affects Communal D¨ Violence”. Political Geography 76, 102073. Eastin, Joshua. 2018. “Hell and High Water: Precipitation Shocks and Conflict Violence in the Philippines”. Political Geography 63, 116–134. Fetzer, Thiemo. 2020. “Can Workfare Programs Moderate Conflict? Evidence from India”. Journal of the European Economic Association 18 (6), 3337–3375. Fjelde, Hanne. 2015. “Farming or Fighting? Agricultural Price Shocks and Civil War in Africa”. World Development 67, 525–534. ee Nilsson. 2012. “Rebels against Rebels: Explaining Violence between Fjelde, Hanne and Desir´ Rebel Groups”. Journal of Conflict Resolution 56 (4), 604–628. Fjelde, Hanne and Nina von Uexkull. 2012. “Climate triggers: Rainfall anomalies, vulnerability and communal conflict in Sub-Saharan Africa”. Political Geography 31 (7), 444–453. Gong, Erick and Katherine A. Sullivan. 2017. “Conflict and Coffee: Are Higher Coffee Prices Fuelling Rebellion in Uganda?”. Journal of African Economies 26 (3), 322–341. 46 Guardado, Jenny. 2018. “Land Tenure, Price Shocks, and Insurgency: Evidence from Peru and Colombia”. World Development 111, 256–269. Harari, Mariaflavia and Eliana La Ferrara. 2018. “Conflict, Climate, and Cells: A Disaggregated Analysis”. The Review of Economics and Statistics 100 (4), 594–608. Hidalgo, F. Daniel, Suresh Naidu, Simeon Nichter, and Neal Richardson. 2010. “Economic Deter- minants of Land Invasions”. Review of Economics and Statistics 92 (3), 505–523. Hong, Ji Yeon and Wenhui Yang. 2020. “Oilfields, Mosques and Violence: Is There a Resource Curse in Xinjiang?”. British Journal of Political Science 50 (1), 45–78. Jia, Ruixue. 2014. “Weather Shocks, Sweet Potatoes and Peasant Revolts in Historical China”. The Economic Journal 124 (575), 92–118. Kung, James Kai-sing and Chicheng Ma. 2014. “Can Cultural Norms Reduce Conflicts? Confucian- ism and Peasant Rebellions in Qing China”. Journal of Development Economics 111, 132–149. Landis, Steven T., Babak Rezaeedaryakenari, Yifan Zhang, Cameron G. Thies, and Ross Maciejew- ski. 2017. “Fording Differences? Conditions Mitigating Water Insecurity in the Niger River Basin”. Political Geography 56, 77–90. Lessmann, Christian and Arne Steinkraus. 2019. “The Geography of Natural Resources, Ethnic Inequality and Civil Conflicts”. European Journal of Political Economy 59, 33–51. Linke, Andrew M., John O’Loughlin, J. Terrence McCabe, Jaroslav Tir, and Frank D.W. Witmer. 2015. “Rainfall Variability and Violence in Rural Kenya: Investigating the Effects of Drought and the Role of Local Institutions with Survey Data”. Global Environmental Change 34, 35–47. Linke, Andrew M., Frank D. W. Witmer, John O’Loughlin, J. Terrence McCabe, and Jaroslav Tir. 2018. “Drought, Local Institutional Contexts, and Support for Violence in Kenya”. Journal of Conflict Resolution 62 (7), 1544–1578. aivi. 2010. “The Spoils of Nature: Armed Civil Conflict and Rebel Access to Natural Lujala, P¨ Resources”. Journal of Peace Research 47 (1), 15–28. 47 Maystadt, J.-F., M. Calderone, and L. You. 2015. “Local warming and violent conflict in North and South Sudan”. Journal of Economic Geography 15 (3), 649–671. cois, Giacomo De Luca, Petros G. Sekeris, and John Ulimwengu. 2014. “Mineral Maystadt, Jean-Fran¸ Resources and Conflicts in DRC: a Case of Ecological Fallacy?”. Oxford Economic Papers 66 (3), 721–749. cois and Olivier Ecker. 2014. “Extreme Weather and Civil War: Does Drought Maystadt, Jean-Fran¸ Fuel Conflict in Somalia through Livestock Price Shocks?”. American Journal of Agricultural Economics 96 (4), 1157–1182. McGuirk, Eoin and Marshall Burke. 2020. “The Economic Origins of Conflict in Africa”. Journal of Political Economy 128 (10), 3940–3997. Nordkvelle, Jonas, Siri Aas Rustad, and Monika Salmivalli. 2017. “Identifying the Effect of Climate Variability on Communal Conflict Through Randomization”. Climatic Change 141 (4), 627–639. O’Loughlin, John, Andrew M. Linke, and Frank D. W. Witmer. 2014. “Effects of Temperature and Precipitation Variability on the Risk of Violence in Sub-Saharan Africa, 1980–2012”. Proceedings of the National Academy of Sciences 111 (47), 16712–16717. O’Loughlin, John, Frank D. W. Witmer, Andrew M. Linke, Arlene Laing, Andrew Gettelman, and Jimy Dudhia. 2012. “Climate Variability and Conflict Risk in East Africa, 1990–2009”. Proceedings of the National Academy of Sciences 109 (45), 18344–18349. Papaioannou, Kostadis J.. 2016. “Climate Shocks and Conflict: Evidence from Colonial Nigeria”. Political Geography 50, 33–47. Papaioannou, Kostadis J.. 2017. ““Hunger Makes a Thief of Any Man”: Poverty and Crime in British Colonial Asia”. European Review of Economic History , ereh;hew019v1. Papaioannou, Kostadis J. and Michiel de Haas. 2017. “Weather Shocks and Agricultural Com- mercialization in Colonial Tropical Africa: Did Cash Crops Alleviate Social Distress?”. World Development 94, 346–365. 48 Raleigh, Clionadh, Hyun Jin Choi, and Dominic Kniveton. 2015. “The Devil is in the Details: an Investigation of the Relationships between Conflict, Food Price and Climate across Africa”. Global Environmental Change 32, 187–199. Raleigh, Clionadh and Dominic Kniveton. 2012. “Come Rain or Shine: An Analysis of Conflict and Climate Variability in East Africa”. Journal of Peace Research 49 (1), 51–64. Rigterink, Anouk S.. 2020. “Diamonds, Rebel’s and Farmer’s Best Friend: Impact of Variation in the Price of a Lootable, Labor-Intensive Natural Resource on the Intensity of Violent Conflict”. Journal of Conflict Resolution 64 (1), 90–126. Rowhani, Pedram, Olivier Degomme, Debarati Guha-Sapir, and Eric F. Lambin. 2011. “Malnu- trition and Conflict in East Africa: the Impacts of Resource Variability on Human Security”. Climatic Change 105 (1-2), 207–222. Sarsons, Heather. 2015. “Rainfall and Conflict: a Cautionary Tale”. Journal of Development Economics 115, 62–72. Shapiro, Jacob N. and Oliver Vanden Eynde. 2023. “Fiscal Incentives for Conflict: Evidence from India’s Red Corridor”. The Review of Economics and Statistics 105 (1), 217–225. First published online, March, 2021. Theisen, Ole Magnus. 2012. “Climate clashes? Weather variability, land pressure, and organized violence in Kenya, 1989–2004”. Journal of Peace Research 49 (1), 81–96. Theisen, Ole Magnus, Helge Holtermann, and Halvard Buhaug. 2012. “Climate Wars? Assessing the Claim That Drought Breeds Conflict”. International Security 36 (3), 79–106. Vanden Eynde, Oliver. 2018. “Targets of Violence: Evidence From India’s Naxalite Conflict”. The Economic Journal 128 (609), 887–916. von Uexkull, Nina. 2014. “Sustained Drought, Vulnerability and Civil Conflict in Sub-Saharan Africa”. Political Geography 43, 16–26. Wischnath, Gerdis and Halvard Buhaug. 2014. “On Climate Variability and Civil War in Asia”. Climatic Change 122 (4), 709–721. 49 Witmer, Frank DW, Andrew M Linke, John O’Loughlin, Andrew Gettelman, and Arlene Laing. 2017. “Subnational Violent Conflict Forecasts for Sub-Saharan Africa, 2015–65, Using Climate- sensitive Models”. Journal of Peace Research 54 (2), 175–192. Yeeles, Adam. 2015. “Weathering Unrest: the Ecology of Urban Social Disturbances in Africa and Asia”. Journal of Peace Research 52 (2), 158–170. 50 Appendices (a) Total (2,464) Agricultural shock (+) Agricultural shock (-) Hydrocarbon/Mineral shock (+) Other shock (+ or -) 0 20 40 60 80 100 0 20 40 60 80 100 percent Price Climate (b) Baseline (1,391) Financial Crisis Resource endowment Agricultural shock (+) Subsidies to Mine Agricultural shock (-) Unemployment Other Hydrocarbon/Mineral shock (+) Other shock (+ or -) 0 20 40 60 80 100 0 20 40 60 80 100 percent Price Climate Financial Crisis Resource endowment Subsidies to Mine Unemployment Other Notes: Categories are not totally exclusive, so percentages do not necessarily sum to 100 for each meta-subgroup. For example, estimates that instrument an output measure with a climate variable will be classified both as climate and production shocks. The shares of each subgroup and their components are detailed in Table A1. Source: Authors’ compilation from MRA database. Figure A1: Channels of Transmission [Share of Estimates in Meta-Subgroups] 51 Table A1: Channels of Transmission [Share of Estimates for Each Category of Shock in Each Meta- Subgroup] Total Baseline AS+ AS- HS+ Other AS+ AS- HS+ Other (%) (%) (%) (%) (%) (%) (%) (%) Price shocks 65.7 10.7 89.5 0.0 75.2 8.5 83.7 0.0 Incl. production*prices 63.9 6.1 13.4 0.0 72.8 1.9 20.6 0.0 Incl. location*prices 0.0 0.0 68.9 0.0 0.0 0.0 60.3 0.0 Incl. consumption*prices 0.0 4.6 0.0 0.0 0.0 6.7 0.0 0.0 Incl. prices 1.9 0.0 0.0 0.0 2.5 0.0 0.0 0.0 Incl. rents 0.0 0.0 7.2 0.0 0.0 0.0 2.8 0.0 Climate shocks 30.4 79.3 0.0 69.8 19.2 78.4 0.0 76.6 Incl. precipitations 27.6 28.6 0.0 38.3 15.2 29.3 0.0 45.3 Incl. droughts 2.8 37.9 0.0 3.1 4.0 35.6 0.0 3.6 Incl. temperatures 0.5 5.2 0.0 27.2 0.9 7.3 0.0 26.3 Incl. floods 0.0 10.9 0.0 1.9 0.0 12.3 0.0 2.2 Ressource endowment shocks 3.1 12.4 6.8 2.5 5.6 15.4 10.2 2.9 Incl. water access 1.6 0.6 0.0 0.0 2.9 1.2 0.0 0.0 Incl. production/endowment 1.5 11.8 6.8 2.5 2.7 14.1 10.2 2.9 Other shocks 0.9 1.4 3.7 28.4 0.2 2.7 6.2 21.2 Incl. financial crisis 0.0 0.0 0.0 27.8 0.0 0.0 0.0 20.4 Incl. subsidies to Mine 0.0 0.0 3.7 0.0 0.0 0.0 6.2 0.0 Incl. unemplyment 0.0 0.0 0.0 0.6 0.0 0.0 0.0 0.7 Incl other 0.9 1.4 0.0 0.0 0.2 2.7 0.0 0.0 Observations 808 951 543 162 448 481 325 137 Notes: Categories are not totally exclusive, so percentages do not necessarily sum to 100 for each meta-subgroup. For example, estimates that instrument an output measure with a climate variable will be classified both as climate and production shocks. AS+: positive agricultural shock. AS-: negative agricultural shock. HS+: positive hydrocarbon and mineral shock. Source: Authors’ compilation from MRA database. Table A2: Table of Results [Other Shocks] Pure climatic shock (+) Pure climatic shock (-) Other potential detrimental shocks Precision (1/SE) Constant Precision (1/SE) Constant Precision (1/SE) Constant Baseline regression 0.006 -2.387* -0.001 2.563 *** -0.002 1.262 Non-lineartities Control for risk absorber=1 0.006 -2.318 *** -0.001 2.563 *** 0.001 1.577 ** Control for risk enhancer=1 0.006 -2.416 *** 0.000 2.465 *** -0.002 1.262 Control for risk uncertain=1 0.006 -2.387* -0.001 2.563 *** -0.002 1.262 Baseline regression 0.004 -2.130 *** 0.000 2.414 *** -0.001 1.726 *** Mechanisms Mecha. pacifying=1 0.004 -2.130 *** 0.000 2.414 *** -0.001 1.726 *** Mecha. stressful=1 0.004 -2.130 *** 0.000 2.414 *** -0.001 1.726 *** Mecha. undefined=1 0.004 -2.130 *** 0.000 2.414 *** -0.001 1.726 *** Notes: All models are estimated with a multilevel mixed-effects model. The dependent variable is the t -statistic of the estimate of interest on conflicts as dependent variable. Standard errors are not reported to save space. The table reports the mean beyond bias (Precision (1/SE)) and the publication bias (Constant) for each specification A detailed description of all variables is available in Table A3. The detailed composition of mechanisms is available in Table A8. The detailed composition of risk absorbers, risk enhancers and risk uncertain is available in Tables A8 and A9. The detailed composition of mechanisms is available in Table xx. ∗p < 0.10, ∗ ∗ p < 0.05, ∗ ∗ ∗p < 0.01. Source: Authors’ compilation from MRA database. 52 Notes: Dots denote outliers. Most PCC values fall between -0.1 and 0.1, implying that income shocks may have a small impact on the risk of conflict locally. Source: Authors’ compilation from MRA database. Figure A2: The Box Plot of PCCs within Studies. 53 Table A3: Meta-Regression Variables Definition and Descriptive Statistics (N= 2,464) Variable name Variable description Mean S.D. t -student Student’s t-test of estimates between income shocks and conflicts 0.561 2.612 Absolute t -student Absolute value of t-student 2.324 1.316 Adjusted t -student Adjusted t-student (reversed sign for Pos. Agr. Shocks) 1.129 2.421 Partial Partial correlation of estimates between income shocks and conflicts 0.010 0.056 Precision 1/Standard error of the partial correlation 219.292 245.048 Transmission channels Pos. agr. shocks* BV = 1: income shock is a positive agricultural shock 0.328 0.470 Neg. agr. shocks BV = 1: income shock is a negative agricultural shock 0.386 0.487 Pos. hydr./min. shocks BV = 1: income shock is a positive hydrocarbon or mineral shock 0.220 0.415 Other shocks BV = 1: other income shock (e.g., financial shock, drug shock) 0.066 0.248 Mechanisms and interactive models Mecha. uncertain* BV = 1: if mechanism is uncertain 0.049 0.216 Mecha. peace* BV = 1: if pacifying mechanism encompassed 0.302 0.459 Mecha. stress BV = 1: if stressfull mechanism encompassed 0.649 0.477 Risk uncertain* BV = 1: if interactive term with no clear ex-ante effect 0.065 0.246 Risk absorber* BV = 1: if interactive term is theoretically an abridgment for conflict 0.200 0.400 Risk enhancer BV = 1: if interactive term is theoretically an amplifier for conflict 0.171 0.377 Publication outlet No top 5* BV = 1: the paper is not in top 5 Economics or Political science journal 0.827 0.378 Top 5 BV = 1: the paper is in top 5 Economics or Political science journal 0.173 0.378 No conflict review* BV = 1: the paper is not in conflict specialized journal 0.929 0.256 Conflict review BV = 1: the paper is published in a conflict specialized journal 0.071 0.256 Core results* BV = 1: estimates taken from the core paper 0.761 0.427 Result from appendix BV = 1: estimates taken from appendix 0.239 0.427 Age of study Age of study, in year, in 2023 6.725 2.507 SJR score SCImago Journal score of the paper (the year of publication) 5.508 4.303 Geography No country focus* BV = 1: estimates is not focusing on a given country 0.500 0.500 Country focus BV = 1: estimates focus on a given country 0.500 0.500 Worldwide* BV = 1: focus on two or several regions 0.066 0.248 Africa BV = 1: focus on Africa as a whole or Sub-Saharan Africa (SSA) 0.551 0.497 EAP BV = 1: focus on East Asia Pacific (World Bank Group definition) 0.198 0.399 LAC BV = 1: focus on Latin America and the Caribbean (WBG def.) 0.111 0.314 South Asia BV = 1: focus on Souht Asia (WBG def.) 0.074 0.262 Model characteristics No MLE* BV = 1: if linear estimator used (e.g., GLM, OLS) 0.778 0.415 Maximum-likelihood est. BV = 1: if maximum-likelihood estimator used (e.g., Logit, PPML) 0.222 0.415 No specific effects* BV = 1: if no specific effect used 0.072 0.259 Specific effects BV = 1: if fixed, mixed or random effects used 0.928 0.259 No cluster unit reference* BV = 1: if no clustering on the unit of reference 0.531 0.499 Cluster unit reference BV = 1: if estimates clustered on the unit of reference 0.469 0.499 Pre Cold War end* BV = 1: if pre Cold War end (until 1991) 0.535 0.499 Post Cold War end BV = 1: if post Cold War end (after 1991) 0.465 0.499 No small grids* BV = 1: if no small grids 0.603 0.489 Small grids BV = 1: if small grids (≤0.5°*0.5°) are used to capture conflicts 0.397 0.489 Sample size Natural logarithm of sample size 9.594 2.274 Notes: BV means binary variable, with a value of 1 if condition is fulfilled and zero otherwise. DV: dependent variable. *: used as reference category in BMA. #: 92 observations out of 2,464 use a mix of price and climatic shocks. Source: Authors’ compilation from MRA database. 54 Table A3 continued: Meta-Regression Variables Definition and Descriptive Statistics (N= 2,464) Variable name Variable description Mean S.D. Conflict characteristics Usual conflict data* BV = 1: if usual conflict data used (i.e., ACLED, UCDP, UCDP/PRIO) 0.459 0.498 Unusual conflict data BV = 1: if non-usual conflict data used 0.541 0.498 Conflict incidence* BV = 1: if DV is conflict incidence 0.903 0.296 Conflict duration BV = 1: if DV is conflict duration 0.028 0.164 Conflict onset BV = 1: if DV is conflict onset 0.069 0.254 No local conflict* BV = 1: if DV is not a local conflict (above Admin2 level or at large grids level) 0.181 0.385 Conflict local BV = 1: if DV is a local conflict (below Admin1 level or at small grids level) 0.819 0.385 Non lethal* BV = 1: if DV is not a lethal conflict 0.235 0.424 Lethal BV = 1: if DV is a lethal conflict by definition 0.765 0.424 Other conflict* BV = 1: if DV refers to other/undefined type of conflicts 0.337 0.473 Armed conflicts BV = 1: if DV refers to armed conflicts and battles 0.278 0.448 Crimes BV = 1: if DV refers to crimes 0.126 0.332 Social unrests BV = 1: if DV refers to social unrest 0.235 0.424 Violence against citizens BV = 1: if DV refers to violence against citizens 0.024 0.153 Measures of shocks No quasi experimental* BV = 1: if not endowment*shock used 0.496 0.500 Quasi experimental BV = 1: if endowment*shock used 0.504 0.500 No labor intensive* BV =1: if shock not based on labor intensive resource 0.665 0.472 Labor intensive resources BV =1: if shock based on labor intensive resource 0.335 0.472 Others* BV =1: if shock refers to other type of shocks 0.165 0.371 Climatic shock# BV =1: if climatic shock used 0.507 0.500 Price shock# BV =1: if price shock used 0.365 0.481 Model specification No past conflict* BV = 1: if the estimate doesn’t control for past conflicts 0.892 0.310 Control for past conflicts BV = 1: if the estimate controls for past conflicts 0.108 0.310 No spatial factor* BV = 1: if the estimate doesn’t control for neighboring conflict 0.916 0.277 Control conflict spatial BV = 1: if the estimate controls for neighboring conflict 0.084 0.277 No population* BV = 1: if the estimate doesn’t control for population density/pop. size 0.691 0.462 Control population BV = 1: if the estimate controls for population density/pop. size 0.309 0.462 No GDP* BV = 1: if the estimate doesn’t control for GDP 0.876 0.329 Control GDP BV = 1: if the estimate controls for GDP or equivalent 0.124 0.329 Notes: BV means binary variable, with a value of 1 if condition is fulfilled and zero otherwise. DV: dependent variable. *: used as reference category in BMA. #: 92 observations out of 2,464 use a mix of price and climatic shocks. Source: Authors’ compilation from MRA database. 55 Table A4: Characteristics of Individual Studies # Study Review SJR Regressions Interaction Subgroup Mechanism(s) Time Region Country focus score # % total # % period(s) 1 Abidoye and Cal` ı Journal of African 0.520 22 0.9% 0 0.0% AS+; AS- Pacifying; 2004-2011 SSA Nigeria (2021) Economies stressful 2 Acharya et al. (2020) Journal of Theoretical 0.954 10 0.4% 10 0.9% AS+ Pacifying 2000-2012 SSA Somalia Politics (Somaliland, Puntland) 3 Ahrens (2015) Peace Economics, 0.186 4 0.2% 0 0.0% AS+ Pacifying 1992-2010 SSA No Peace Science and Public Policy# 4 Almer et al. (2017) Journal of 2.198 162 6.6% 84 7.8% AS- Stressful 1990-2011 SSA No Environmental Economics and Management 5 Bagozzi et al. (2017) The Journal of Politics 4.220 4 0.2% 0 0.0% AS- Stressful 1995-2008 World No 6 Bai and Kung (2011) The Review of 6.765 26 1.1% 0 0.0% AS+; AS- Pacifying; -220-1839 EAP China Economics and stressful Statistics 7 Berman and Couttenier The Review of 5.133 246 10.0% 120 11.2% AS+; other Pacifying; 1989-2005; SSA No (2015) Economics and stressful 1989-2006; Statistics 1997-2006 8 Berman et al. (2017) American Economic 10.472 227 9.2% 130 12.1% HS+ Stressful 1997-2010 SSA No Review° 9 Bhavnani and Lacina World Politics 3.646 5 0.2% 4 0.4% AS- Stressful 1982-2000 South Asia India (2015) 10 Bohlken and Sergenti Journal of Peace 2.272 2 0.1% 0 0.0% AS+ Pacifying 1982-1995 South Asia India (2010) Research# 11 Bollfrass and Shaver PLoS ONE 1.427 30 1.2% 0 0.0% AS+; AS-; Pacifying; 1989-2008 World No (2015) CS+; CS- uncertain 12 Buhaug et al. (2021) The Journal of Politics 3.027 15 0.6% 12 1.1% AS- Stressful 1971-2013 World No 13 Carreri and Dube The Journal of Politics 4.220 4 0.2% 0 0.0% HS+ Stressful 1997-2005 LAC Colombia (2017) 14 Caruso et al. (2016) Journal of Peace 3.586 20 0.8% 0 0.0% AS- Stressful 1993-2003 EAP Indonesia Research# 15 Christensen (2019) International 7.363 21 0.9% 14 1.3% HS+ Stressful 1997-2013 Africa No Organization 16 Christensen et al. World Politics 2.861 5 0.2% 2 0.2% HS+ Stressful 2006-2010; EAP Myanmar (2019) 2006-2015; 2011-2015 17 Corvalan and Pazzona Journal of Economic 1.482 34 1.4% 0 0.0% HS+ Pacifying 2003-2008; LAC Chile (2019) Behavior and 2003-2013 Organization 18 Crost and Felter (2020) Journal of the 7.791 118 4.8% 46 4.3% AS+ Pacifying; 2001-2009; EAP Philippines European Economic stressful 2003-2009 Association 19 Dagnelie et al. (2018) Journal of Health 3.106 80 3.2% 0 0.0% HS+ Stressful 1997-2004 SSA Congo, Dem. Economics Rep. 20 De Juan (2015) Political Geography 2.025 10 0.4% 0 0.0% AS+; AS- Pacifying; 2003-2005 SSA Southern stressful Sudan (Darfur) 21 Detges (2016) Journal of Peace 3.586 10 0.4% 8 0.7% AS- Stressful 1990-2010 SSA No Research# 22 D¨oring (2020) Political Geography 1.527 6 0.2% 0 0.0% AS- Stressful 1990-2014 World; SSA No 23 Dube and Vargas Review of Economic 12.200 55 2.2% 6 0.6% AS+; HS+ Pacifying; 1988-2004; LAC Colombia (2013) Studies° stressful 1988-2005 24 Dube et al. (2016) Journal of the 8.113 54 2.2% 0 0.0% AS+ Pacifying 1990-2005; LAC Mexico European Economic 1990-2010 Association 25 Eastin (2018) Political Geography 1.659 12 0.5% 0 0.0% AS- Stressful 2001-2007 EAP Philippines 26 Fetzer (2020) Journal of the 7.792 43 1.7% 28 2.6% AS+ Pacifying 2000-2010; South Asia India European Economic 2000-2014 Association 27 Fjelde (2015) World Development 2.253 12 0.5% 0 0.0% AS+; AS- Pacifying; 1990-2010 Africa No stressful 28 Fjelde and Nilsson Journal of Conflict 3.448 12 0.5% 0 0.0% HS+; other Stressful 1987-2007 World No (2012) Resolution# 29 Fjelde and von Uexkull Political Geography 2.137 15 0.6% 5 0.5% AS+; AS- Pacifying; 1990-2008 SSA No (2012) stressful 30 Gong and Sullivan Journal of African 0.533 47 1.9% 0 0.0% AS+; AS- Pacifying; 2002-2014 SSA Uganda (2017) Economies stressful 31 Guardado (2018) World Development 2.254 54 2.2% 49 4.6% AS- Stressful 1980-2000; LAC Colombia; 1988-2005 Peru 32 Harari and Ferrara The Review of 8.363 19 0.8% 5 0.5% AS+ Pacifying 1997-2011 SSA No (2018) Economics and Statistics 56 Table A4 continued: Characteristics of Individual Studies # Study Review SJR Regressions Interaction Channel(s) Mechanism(s) Time Region Country focus score # % total # % of transmission period(s) 33 Hidalgo et al. (2010) The Review of 7.882 73 3.0% 56 5.2% AS+; AS-; Pacifying; 1988-2004; LAC Brazil Economics and other stressful 1991; 2000 Statistics 34 Hong and Yang (2020) British Journal of 4.116 48 1.9% 30 2.8% HS+ Pacifying 1998-2005 EAP China Political Science° (Xinjiang) 35 Jia (2014) The Economic Journal 5.264 182 7.4% 92 8.6% AS- Stressful 1470-1900 EAP China 36 Kung and Ma (2014)) Journal of Development 4.712 77 3.1% 32 3.0% AS- Stressful 1651-1910 EAP China Economics (Shandong) 37 Landis et al. (2017) Political Geography 1.770 72 2.9% 72 6.7% AS+; AS- Pacifying; 1997-2012 SSA No stressful 38 Lessmann and European Journal of 1.107 9 0.4% 4 0.4% HS+ Stressful 2000-2012 World No Steinkraus (2019) Political Economy 39 Linke et al. (2015) Global Environmental 3.504 10 0.4% 8 0.7% AS+; AS- Pacifying; 2013 SSA Kenya Change stressful 40 Linke et al. (2018) Journal of Conflict 4.341 19 0.8% 16 1.5% AS- Stressful 2014 SSA Kenya Resolution# 41 Lujala (2010) Journal of Peace 2.272 14 0.6% 0 0.0% HS+ Stressful 1946-2001 World No Research# 42 Maystadt and Ecker American Journal of 1.521 2 0.1% 0 0.0% AS- Stressful 1997-2009 SSA Somalia (2014) Agricultural Economics 43 Maystadt et al. (2014) Oxford Economic 0.687 20 0.8% 0 0.0% HS+ Stressful 1997-2007 SSA Congo, Dem. Papers Rep. 44 Maystadt et al. (2015) Journal of Economic 2.957 49 2.0% 22 2.1% AS-; CS+; Uncertain; 1997-2009 SSA Southern Geography CS- Stressful Sudan (incl. future South Sudan) 45 McGuirk and Burke Journal of Political 21.034 96 3.9% 53 4.9% AS+; AS- Pacifying; 1989-2010; Africa No (2020) Economy° stressful 1997-2013; 1999-2009 46 Nordkvelle et al. (2017) Climatic Change 2.035 7 0.3% 0 0.0% CS- Uncertain 1989-2013 World No 47 O’Loughlin et al. Proceedings of the 6.866 10 0.4% 0 0.0% AS+; AS- Pacifying; 1991-2009 SSA No (2012) National Academy of stressful Sciences (PNAS) 48 O’Loughlin et al. Proceedings of the 6.092 2 0.1% 0 0.0% CS- Uncertain 1980-2012 SSA No (2014) National Academy of Sciences (PNAS) 49 Papaioannou (2016) Political Geography 2.098 2 0.1% 0 0.0% AS- Stressful 1912-1945 SSA Nigeria 50 Papaioannou (2017) European Review of 0.702 33 1.3% 12 1.1% AS- Stressful 1910-1939 World No Economic History 51 Papaioannou and World Development 2.122 93 3.8% 18 1.7% AS- Stressful 1920-1939 SSA No de Haas (2017) (former British colonial area) 52 Raleigh and Kniveton Journal of Peace 2.985 16 0.6% 0 0.0% CS+: CS- Uncertain 1997-2009 SSA No (2012) Research# 53 Raleigh et al. (2015) Global Environmental 3.504 8 0.3% 0 0.0% AS+; AS- Pacifying; 1997-2010 SSA No Change stressful 54 Rigterink (2020) Journal of Conflict 2.671 32 1.3% 32 3.0% HS+ Stressful 2004-2015 Africa No Resolution# 55 Rowhani et al. (2011) Climatic Change 1.532 2 0.1% 0 0.0% AS+ Pacifying 2005-2010 SSA No 56 Sarsons (2015) Journal of Development 3.100 63 2.6% 35 3.3% AS+ Pacifying 1970-1995 South Asia India Economics 57 Shapiro and The Review of 8.245 6 0.2% 0 0.0% HS+ Stressful 2007-2011 South Asia India (Red Vanden Eynde (2023)* Economics and corridor) Statistics 58 Theisen (2012) Journal of Peace 2.985 16 0.6% 0 0.0% AS+; AS- Pacifying; 1989-2004 SSA Kenya Research# stressful 59 Theisen et al. (2012) International Security# 4.318 3 0.1% 2 0.2% AS- Stressful 1960-2004 Africa No 60 Vanden Eynde (2018) Economic Journal 5.009 63 2.6% 54 5.0% AS- Stressful 2005-2011 South Asia India 61 Wischnath and Buhaug Climatic Change 2.440 11 0.4% 0 0.0% CS+; CS- Uncertain 1951-2008 World No (2014) 62 Witmer et al. (2017) Journal of Peace 3.888 2 0.1% 0 0.0% CS- Uncertain 1980-2012 SSA No Research# 63 Yeeles (2015) Journal of Peace 3.892 24 1.0% 0 0.0% CS+; CS- Uncertain 1960-2006 World No Research# 64 von Uexkull (2014) Political Geography 2.815 16 0.6% 12 1.1% AS- Stressful 1989-2008 SSA No Notes: *: published online in 2021 on peer-review journal website, but attributed to a journal issue in 2023. #: conflict specialized peer- review journal. °: top five peer-review journal in Economics or Political Science. SJR scores are established for the year of the publication. AS+: positive agricultural shock. AS-: negative agricultural shock. HS+: positive hydrocarbon shock. CS+: positive pure climatic shock with no explicit impact through agriculture. CS-: negative pure climatic shock with no explicit impact through agriculture. Other: other potential detrimental shocks (positive drug shock; negative financial shock; negative labor market shock). EAP: East Asia & Pacific (World Bank definition). LAC: Latin America & the Caribbean (World Bank definition). SSA: Sub-Saharan Africa (World Bank definition). Except for Africa (Sub-Saharan Africa and North Africa), we consider a worldwide sample if two or more developing regions included in a given regression. For a detailed description of transmission channels, see Tables A5, A6, A7, and A8. For a detailed description of mechanisms, see Tables A9 and A10. For a detailed description of interactive terms, see Table A11. Source: Authors’ compilation from MRA database. 57 Table A5: Details of Transmission Channels [Positive Agricultural Shocks, AS+] Study Sub-component Type # Regressions % transmission channel Abidoye and Cali (2021) Prices of produced commodities Price 11 1.4% Acharya et al. (2020) Exports (log) instrumented by Hajj months; international lamb price change; local sheep and goat price change Price; other 10 1.2% Ahrens (2015) Growth instrumented by temperature Climate 4 0.5% Bai and Kai-sing (2011) Share of years with records of levee breaches of Yellow River in a given decade Climate 13 1.6% Berman and Couttenier (2015) (Positive) agricultural demand shock Other 201 24.9% Bohlken and Sergenti (2010) Growth instrumented by rainfalls Climate 2 0.2% Bollfrass and Shaver (2015) Precipitation Climate 5 0.6% Crost and Felter (2020) Several crops*price (mainly cavendish bananas) Climate; price 118 14.6% De Juan (2015) Normalized Difference Vegetation Index (NDVI) 1998-2002 Climate 8 1.0% Dube and Vargas (2013) Coffee intensity*price Price 20 2.5% Dube et al. (2016) Agro-climatically attainable yield for maize *national maize price in year (instrumented by lagged weather conditions) Climate; price 54 6.7% Fetzer (2020) Log(Monsoon t-1) Climate 43 5.3% Fjelde (2015) Spatial data on crop production*international prices Price 11 1.4% Fjelde and Uexkull (2012) Inter-annual positive rainfall anomaly Climate 1 0.1% Gong and Sullivan (2017) Several crops*price (mainly coffee) Price 44 5.4% Harari and La Ferrara (2018) Standardized Precipitation-Evapotranspiration Index (SPEI)*Growing Season, t-1 Climate 19 2.4% Hidalgo et al. (2010) Agricultural income instrumented by rain deviation; Log GDP per capita instrumented by rain deviation Climate 64 7.9% Landis et al. (2017) Precipitation trend Climate 36 4.5% Linke et al. (2015) Changes in Vegetation conditions (VCI) Climate 2 0.2% McGuirk and Burke (2020) Producer price index: price*crop share land Price 68 8.4% O’Loughlin et al. (2012) Precipitation (SPI6); Precipitation (wet) Climate 4 0.5% Raleigh et al. (2015) Positive rainfalls lagged 1yr Climate 1 0.1% Rowhani et al. (2011) iEVI (ecosystem productivity: total annual vegetation activity) Climate 2 0.2% Sarsons (2015) Rain growth; rain shock Climate 63 7.8% Theisen (2012) Distance to Drought (SPI6) Climate 4 0.5% Total 808 100% Notes: For a data visualization of subcomponents of transmission channels, see Figure A1 and Table A1. Source: Authors’ compilation from MRA database. Table A6: Details of Transmission Channels [Negative Agricultural Shocks, AS-] Study Sub-component Type # Regressions % transmission channel Abidoye and Cali (2021) Prices of consumed commodities Price 11 1.2% Almer et al. (2017) Standardized Precipitation-Evapotranspiration Index (SPEI) Climate 162 17.0% Bagozzi et al. (2017) Drought Climate 4 0.4% Bai and Kai-sing (2011) Share of years with records of drought disasters on the central plains in a given decade Climate 13 1.4% Bhavnani and Lacina (2015) Internal migration instrumented by abnormal rainfall (Monsoon) Climate 5 0.5% Bollfrass and Shaver (2015) Temperature Climate 5 0.5% Buhaug et al. (2021) Drought (SPEI) Climate 15 1.6% Caruso et al. (2016) Paddy rice production instrumented by temperature deviation Climate 20 2.1% De Juan (2015) Normalized Difference Vegetation Index (NDVI) 1998-2002 Climate 2 0.2% Detges (2016) Extreme drought Climate 10 1.1% Doring (2020) Several measures on groundwater scarcity Climate 6 0.6% Eastin (2018) Typhoon last year Climate 12 1.3% Fjelde (2015) Spatial data on crop production*international prices Price 1 0.1% Fjelde and Uexkull (2012) Inter-annual negative rainfall anomaly Climate 14 1.5% Gong and Sullivan (2017) Coffee intensity*price (-1 standard deviation) Price 3 0.3% Guardao (2018) Coffee intensity*price; agro-climatic attainable yield for coffee*price Climate; price 54 5.7% Hidalgo et al. (2010) Agricultural income instrumented by rain deviation; rainfall deviation Climate 8 0.8% Jia (2014) Several measures on droughts and floods Climate 182 19.1% Kun and Ma (2014) Crop failure; Drought; Flood Climate; other 77 8.1% Landis et al. (2017) Negative precipitation variability Climate 36 3.8% Linke et al. (2015) Drought (SAT and TAMSAT) Climate 8 0.8% Linke et al. (2018) Drought (SAT and TAMSAT) Climate 19 2.0% Maystadt and Ecker (2014) Drought length (in months); cattle price (log) instrumented by drought length (in month) Climate 2 0.2% Maystadt et al. (2015) Precipitation anomaly; Temperature anomaly Climate 19 2.0% McGuirk and Burke (2020) Consumer price index: price*crop share land Price 28 2.9% O’Loughlin et al. (2012) Precipitation (dry); several measures of temperature Climate 6 0.6% Papaioannou (2016) Rainfall deviation square Climate 2 0.2% Papaioannou (2017) Several measure on rainfall deviations/shocks Climate 33 3.5% Papaioannou and De Haas (2017) Several measure on rainfall deviations/shocks Climate 93 9.8% Raleigh et al. (2015) Several measures of commodity prices instrumented by negative shocks; Negative rainfalls Climate; price 7 0.7% Theisen (2012) Drought (SPI6); Rainfall defficiency (SPI6); Temperature (SPI6) Climate 12 1.5% Theisen et al. (2012) Drougt (SPI) Climate 3 0.3% Vanden Eynde (2018) Rain deficiency t-1 Climate 63 6.6% von Uexkull (2014) Several measure of drought Climate 16 1.7% Total 951 100% Notes: For a data visualization of subcomponents of transmission channels, see Figure A1 and Table A1. Source: Authors’ compilation from MRA database. 58 Table A7: Details of Transmission Channels [Positive Hydrocarbon and Mineral Shocks, HS+] Study Sub-component Type # Regressions % transmission channel Berman et al. (2017) Mines*prices Price 227 41.8% Carreri and Dube (2017) Municipality that produced oil in 1993*oil price Price 4 0.7% Christensen (2019) Active mine*price Price 21 3.9% Christensen et al. (2019) Active mine*price Price 5 0.9% Corvalan and Pazzona (2019) Production copper in 2000*prices Price 34 6.3% Dagnelie et al. (2018) Weighted minerals endowment*price for each mineral (several minerals) Price 80 14.7% Dube and Vargas (2013) Coal prod*coal price; gold prod*gold price; oil prod*oil price Price 35 6.4% Fjelde and Nilsson (2012) Presence of gemstones in conflict area; presence of oil/gas in conflict area Other 8 1.5% Hong and Yang (2018) Several measures based on gas and oil prices and gas and oil revenue Price; other 48 8.8% Lessmann and Steinkraus (2019) Mineral Gini as concentration of mine light across ethnicities Other 9 1.7% Lujala (2010) Mineral and hydrocarbon production/reserves in conflict zone Other 14 2.6% Maystadt et al. (2014) Subsidies to mining concessions Price 20 3.7% Rigterink (2020) Diamond propensity*price; diamond propensity*price (price instrumented with Russian prod. volume) Price 32 5.9% Shapiro and Vanden Eynde (2021) Iron deposits*post new royalty regime Other 6 1.1% Total 543 100% Notes: For a data visualization of subcomponents of transmission channels, see Figure A1 and Table A1. Source: Authors’ compilation from MRA database. Table A8: Details of Transmission Channels [Other Shocks] Study Sub-component Type # Regressions % transmission channel Pure negative climatic shocks (Other 1) Bollfrass and Shaver (2015) Precipitation; Temperature Climate 18 23.4% Maystadt et al. (2015) Precipitation anomaly; Temperature anomaly Climate 21 27.3% Nordkvelle et al. (2017) Absolute SPI; Drought; Flood Climate 7 9.1% O’Loughlin et al. (2014) Precipitation (SPI6); Temperature (SPI6) Climate 2 2.6% Raleigh and Kniveton (2012) Rainfall variation Climate 8 10.4% Wischnath and Buhaug (2014) Drought; precipitation deviation, growth and level Climate 6 7.8% Witmer et al. (2017) Precipitation (SPI6); Temperature (SPI6) Climate 2 2.6% Yeeles (2015) Precipitation; Temperature Climate 13 16.9% Total 77 100% Other potential detrimental shocks (Other 2) Berman and Couttenier (2015) (Negative) financial crises shock Other 45 90.0% Fjelde and Nilsson (2012) Presence of drugs in conflict area Other 4 8.0% Hidalgo et al. (2010) Log rural unemployment instrumented by rain deviation Climate 1 2.0% Total 50 100% Pure positive climatic shocks (Other 3) Bollfrass and Shaver (2015) Precipitation Climate 2 5.7% Maystadt et al. (2015) Precipitation anomaly; Temperature anomaly Climate 9 25.7% Raleigh and Kniveton (2012) Rainfall variation Climate 8 22.9% Wischnath and Buhaug (2014) Temperature deviation, growth and level Climate 5 14.3% Yeeles (2015) Precipitation; Temperature Climate 11 31.4% Total 35 100% Notes: For a data visualization of subcomponents of transmission channels, see Figure A1 and Table A1. Source: Authors’ compilation from MRA database. 59 Table A9: Details of Stressful Mechanisms Study Sub-component # Regressions % transmission channel Abidoye and Cali (2021) Enhanced poverty (budget constraint) 11 0.70% Almer et al. (2017) Enhanced poverty (water insecurity) 162 10.10% Bagozzi et al. (2017) Enhanced poverty (food scarcity) 4 0.30% Bai and Kai-sing (2011) Enhanced poverty (food scarcity) 13 0.80% Berman and Couttenier (2015) Enhanced poverty (budget constraint) 45 2.80% Berman et al. (2017) Feasability for insurgents to fund their activity 227 14.20% Bhavnani and Lacina (2015) Exodus 5 0.30% Buhaug et al. (2021) Enhanced poverty (lack dynamism of local economy) 15 0.90% Carreri and Dube (2017) Rent capture 4 0.30% Caruso et al. (2016) Enhanced poverty (food scarcity) 20 1.30% Christensen (2019) Imperfect information (lack transparency extractive dividend) 21 1.30% Christensen et al. (2019) Imperfect information (lack transparency extractive dividend) 5 0.30% Crost and Felter (2020) Rent capture 102 6.40% Dagnelie et al. (2018) Rent capture 80 5.00% De Juan (2015) Rent capture (livestock and vegetation for pastoral population) 8 0.50% Detges (2016) Enhanced poverty (food scarcity) 10 0.60% Doring (2020) Enhanced poverty (water insecurity) 6 0.40% Dube and Vargas (2013) Rent capture 35 2.20% Eastin (2018) Regional destruction (including agricultural soil) 12 0.80% Fjelde (2015) Enhanced poverty (budget constraint) 1 0.10% Fjelde and Nilsson (2012) Rent capture 12 0.80% Fjelde and Uexkull (2012) Enhanced poverty (water insecurity) 14 0.90% Gong and Sullivan (2017) Rent capture 41 2.60% Guardao (2018) Enhanced poverty (budget constraint) 54 3.40% Hidalgo et al. (2010) Enhanced poverty (budget constraint) 9 0.60% Jia (2014) Enhanced poverty (food scarcity) 182 11.40% Kun and Ma (2014) Enhanced poverty (food scarcity) 77 4.80% Landis et al. (2017) Enhanced poverty (water insecurity) 36 2.30% Lessmann and Steinkraus (2019) Grievance 9 0.60% Linke et al. (2015) Grievance 8 0.50% Linke et al. (2018) Grievance 19 1.20% Lujala (2010) Rent capture 14 0.90% Maystadt and Ecker (2014) Enhanced poverty (budget constraint) 2 0.10% Maystadt et al. (2014) Rent capture 20 1.30% Maystadt et al. (2015) Enhanced poverty (food scarcity; water insecurity) 15 0.90% McGuirk and Burke (2020) Enhanced poverty (budget constraint) 28 1.80% O’Loughlin et al. (2012) Enhanced poverty (food scarcity) 6 0.40% Papaioannou (2016) Regional destruction (including agricultural soil) 2 0.10% Papaioannou (2017) Enhanced poverty (food scarcity) 33 2.10% Papaioannou and De Haas (2017) Regional destruction (including agricultural soil) 93 5.80% Raleigh et al. (2015) Enhanced poverty (budget constraint) 7 0.40% Rigterink (2020) Rent capture 32 2.00% Shapiro and Vanden Eynde (2021) Rent capture; state capacity (security operations) 6 0.40% Theisen (2012) Enhanced poverty (food scarcity) 12 0.80% Theisen et al. (2012) Enhanced poverty (food scarcity) 3 0.20% Vanden Eynde (2018) Enhanced poverty (budget constraint) 63 3.90% von Uexkull (2014) Enhanced poverty (food scarcity) 16 1.00% Total 1 599 100% Source: Authors’ compilation from MRA database. 60 Table A10: Details of Pacifying and Uncertain Mechanisms Study Sub-component # Regressions % transmission channel Pacifying mechanisms Abidoye and Cali (2021) Reduced poverty (budget constraint) 11 1.50% Acharya et al. (2020) Reduced poverty (budget constraint) 10 1.30% Ahrens (2015) Reduced poverty (dyn. of local economy) 4 0.50% Bai and Kai-sing (2011) Reduced poverty (food scarcity) 13 1.70% Berman and Couttenier (2015) Reduced poverty (budget constraint) 201 27.00% Bohlken and Sergenti (2010) Reduced poverty (dyn. of local economy) 2 0.30% Bollfrass and Shaver (2015) None specifically 5 0.70% Corvalan and Pazzona (2019) Reduced poverty (dyn. of local economy) 34 4.60% Crost and Felter (2020) Reduced poverty (budget constraint; food scarcity) 16 2.20% De Juan (2015) Lower incentive for rent capture (livestock and vegetation pastoral pop.) 2 0.30% Dube and Vargas (2013) Reduced poverty (budget constraint) 20 2.70% Dube et al. (2016) Reduced poverty (budget constraint) 54 7.30% Fetzer (2020) Reduced poverty (budget constraint; food scarcity) 43 5.80% Fjelde (2015) Reduced poverty (budget constraint) 11 1.50% Fjelde and Uexkull (2012) Reduced poverty (water insecurity) 1 0.10% Gong and Sullivan (2017) Reduced poverty (food scarcity); Lower rent capture 6 0.80% Harari and La Ferrara (2018) Reduced poverty (food scarcity) 19 2.60% Hidalgo et al. (2010) Reduced poverty (budget constraint) 64 8.60% Hong and Yang (2018) Reduced poverty (budget constraint; water scarcity) 48 6.50% Landis et al. (2017) Reduced poverty (water insecurity) 36 4.80% Linke et al. (2015) Lower grievance 2 0.30% McGuirk and Burke (2020) Reduced poverty (budget constraint) 68 9.10% O’Loughlin et al. (2012) Reduced poverty (food scarcity) 4 0.50% Raleigh et al. (2015) Reduced poverty (budget constraint) 1 0.10% Rowhani et al. (2011) Lower grievance or reverse causality 2 0.30% Sarsons (2015) Reduced poverty (budget constraint) 63 8.50% Theisen (2012) Reduced poverty (food scarcity) 4 0.50% Total 744 100% Uncertain mechanisms Bollfrass and Shaver (2015) None specifically 25 20.70% Maystadt et al. (2015) None specifically 34 28.10% Nordkvelle et al. (2017) None specifically 7 5.80% O’Loughlin et al. (2014) None specifically 2 1.70% Raleigh and Kniveton (2012) None specifically 16 13.20% Wischnath and Buhaug (2014) None specifically 11 9.10% Witmer et al. (2017) None specifically 2 1.70% Yeeles (2015) None specifically 24 19.80% Total 121 100% Source: Authors’ compilation from MRA database. 61 Table A11: Details of interaction terms # Study Subgroup Risk absorber Risk enhancer Risk undetermined 1 Abidoye and Cali (2021) AS+/- - - - 2 Acharya et al. (2020) AS+ Institutional framework in Somaliland Institutional framework in Puntland - 3 Ahrens (2015) AS+ - - 4 Almer et al. (2017) AS- Several including high blue water per capita; Several including low blue water per capita; - high distance to urban center; no ethnic low distance to urban center; ethnic diversity diversity 5 Bagozzi et al. (2017) AS+ - - - 6 Bai and Kung (2011) AS+/- - - - 7 Berman and Couttenier (2015) AS+, Low distance to seaport; local revenue Distance to seaport; distance to natural Various measures other mobilization resources 8 Berman et al. (2017) HS+ Several including relatively higher governance Several including ethnic and religious Mines characteristics indicators; proxy of lack rent incentives fract./polarization; Gini index; mineral rents 9 Bhavnani and Lacina (2015) AS- Low unemployment in host state; political High unemployment in host state; political - match between migrant and host state mismatch between migrant and host state 10 Bohlken and Sergenti (2010) AS+ - - - 11 Bollfrass and Shaver (2015) AS+/-, - - - other 12 Buhaug et al. (2021) AS- - Several terms based on discriminated and - downgraded groups 13 Carreri and Dube (2017) HS+ - - - 14 Caruso et al. (2016) AS- - - - 15 Christensen (2019) HS+ WGI control of corruption; EITI* candidate; Mineral price - EITI* compliant member 16 Christensen et al. (2019) HS+ Post-2000 (democratization) Mineral price - 17 Corvalan and Pazzona (2019) HS+ - - - 18 Crost and Felter (2020) AS+ - Several terms based on resource endowment, - muslim pop., territories control by rebel groups 19 Dagnelie et al. (2018) HS+ - - - 20 De Juan (2015) AS+/- - - - 21 Detges (2016) AS- Relatively good access to alternative water Low density paved roads - sources; relatively high density paved roads 22 Doring (2020) AS- - - - 23 Dube and Vargas (2013) AS+, HS+ - Years with pro-para majority local councils - 24 Dube et al. (2016) AS+ - - - 25 Eastin (2018) AS- - - - 26 Fetzer (2020) AS+ Several measures based on NREGA** social Inside red corridor - program; outside red corridor 27 Fjelde (2015) AS+/- - - - 28 Fjelde and Nilsson (2012) HS+, - - - other 29 Fjelde and Uexkull (2012) AS/- - - various measures 30 Gong and Sullivan (2017) AS+/- - - - 31 Guardao (2018) AS- Several measures based on shared arrangement Agricultural price - on lands 32 Harari and La Ferrara (2018) AS+ - - Various measures based on groups and infrastructures 33 Hidalgo et al. (2010) AS+/-, Several measures based on low land Gini Several measures based on land Gini Mainly contracts characteristics other 34 Hong and Yang (2018) HS+ Several measures based on gas and oil prices Mosque density in Xinjiang - and gas and oil revenue 35 Jia (2014) AS- Mainly the practice of culture of sweet potatoes Mainly the lack of culture of sweet potatoes Time trend 36 Kung and Ma (2014) AS- Several proxies of Confucean culture (#chaste - - women per area; #temples; #sages) 37 Landis et al. (2017) AS+/- Shared ethnicity; higher road density The distance to Niger river - 38 Lessmann and Steinkraus (2019) HS+ Higher institutional quality; low inequality in Higher inequality in mineral distribution - mineral districts 39 Linke et al. (2015) AS+/- Presence of rules to manage resources and No rule to manage resources - conflicts on resources 40 Linke et al. (2018) AS- Presence of rules to manage resources and No rule to manage resources - conflicts on resources 41 Lujala (2010) HS+ - - - 42 Maystadt and Ecker (2014) AS- - - - 43 Maystadt et al. (2014) HS+ - - - 44 Maystadt et al. (2015) AS-, other Several measures based on favorable Several measures based on unfavorable - agricultural potential agricultural potential 45 McGuirk and Burke (2020) AS+/- Cash crops; high luminosity as proxy of urban Food crops; low luminosity as proxy of urban - activity activity 46 Nordkvelle et al. (2017) other - - - 47 O’Loughlin et al. (2012) AS+/- - - - 48 O’Loughlin et al. (2014) other - - - 49 Papaioannou (2016) AS- - Several measures based on unfavorable - agricultural potential 50 Papaioannou (2017) AS- Several measures based on favorable Exceptional drought; exceptional flood - agricultural potential 51 Papaioannou and De Haas (2017) AS- Export crop production and suitability - - 52 Raleigh and Kniveton (2012) other - - - 53 Raleigh et al. (2015) AS+/- - - - 54 Rigterink (2020) HS+ Off-archon kimberlite; primary diamond Secondary diamond; Upstream rivers diamond - 55 Rowhani et al. (2011) AS+ - - - 56 Sarsons (2015) AS+ Presence of district dam downstream No district dam downstream Several measures based on slopes and soil elevation 57 Shapiro and Vanden Eynde (2021) HS+ - - - 58 Theisen (2012) AS+/- - - - 59 Theisen et al. (2012) AS- No of marginalized ethnic groups Presence of marginalized ethnic groups - 60 Vanden Eynde (2018) AS- - Agricultural share for neighbours; external Agricultural share for locals; various source of revenues for locals measures based on the lack of mining neighbours 61 Wischnath and Buhaug (2014) other - - - 62 Witmer et al. (2017) other - - - 63 Yeeles (2015) other - - - 64 von Uexkull (2014) AS- No rainfed croplands Rainfed croplands - Notes: *: EITI stands for Extractives Industries Transparency Initiative.**: National Rural Employment Guarantee Act. Source: Authors’ compilation from MRA database. 62 Notes: This Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) diagram shows the data collection flow of the meta-analysis. *: Excluded records do not meet the Inclusion Criteria (IC) 1 (to be published in a peer-reviewed journal) and/or the IC 2 (to be published between 2010 and 2021). The inclusion criteria (IC) 3 is to present exploitable empirical results (including standard errors or Student’s t); IC 4 is to have a conflict output; IC 5 is to use a sub-national scale of analysis; IC 6 is to examine an income-related channel of transmission; IC 7 is to analyze non OECD’s High incomes countries. Records are marked as ineligible by automation tools if they are citations or duplicates. Source: Authors’ compilation based on Page et al. (2021). For more information, visit: http: // www. prisma-statement. org/ . Figure A3: PRISMA 2020 Flow Diagram 63