Policy Research Working Paper 10238 Legacies of Conflict Experiences, Self-efficacy and the Formation of Conditional Trust Niklas Buehren Markus Goldstein Imran Rasul Andrea Smurra Africa Region Gender Innovation Lab November 2022 Policy Research Working Paper 10238 Abstract An established literature finds that those exposed to conflict measure of exposure to conflict, self-reported memories/ are more pro-social later in life. This paper builds on this recall of victimization, and ages of exposure to conflict. work in two directions using a sample of 4,200 women born This distinguishes individuals who are traumatized, those during the Sierra Leonean civil war and surveyed 14 years with direct first-hand accounts of conflict, and those with later. First, the paper introduces the notion of conditional second-hand narratives. Empirically, the analysis shows trust, whereby individuals neither outright distrust nor out- that exposure to conflict—either by being in the vicinity right trust others, but can use their perceived self-efficacy to of conflict or through specific experiences of conflict—leads raise the cooperativeness of others. This takes ideas from the respondents to be significantly more likely to conditionally psychology literature documenting survivors of trauma can trust others. The findings show that perceived self-efficacy go through a process of posttraumatic growth generating is higher among those exposed to conflict and this mediates perceived self-efficacy. The paper develops a framework to the impact of conflict on trust preferences. By considering make precise how conditional trust depends on beliefs over the role of memories, narratives/socialization in shaping others, gains from cooperation, risk aversion, and the key experiences of conflict, generating self-efficacy and thus mediating role of self-efficacy in linking conflict and trust. driving trust preferences, the paper provides new avenues Second, the paper constructs a granular typology of expe- for research on how psychological legacies of trauma early in riences of conflict combining information on a geo-coded life shape the long run formation of economic preferences. This paper is a product of the Gender Innovation Lab, Africa Region. 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 nbuehren@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 Legacies of Con‡ict: Experiences, Self-e¢cacy and the Formation of Conditional Trust¤ Niklas Buehren Markus Goldstein Imran Rasul Andrea Smurra JEL: N9, O1 . ¤ We thank all those at BRAC Sierra Leone and IPA Sierra Leone, Gieltje Adriaans, Abdulai Bah, Fernando Fernandez and Leslie Alex for excellent research assistance. We have bene…ted from comments from Nava Ashraf, Oriana Bandiera, Michael Callen, Jean-Paul Carvalho, Martin Cripps, Francisco Ferreira, Michael Frese, Alex Hartman, Moses Shayo, Philipp Strack, Nico Voigtländer and numerous seminar participants. We thank JPAL, IGC, UNICEF and the World Bank Group’s Umbrella Facility for Gender Equality for …nancial support. The views presented are the authors’ and do not represent those of the World Bank, its member countries or DfID. This is an output of the Africa Gender Innovation Lab. Human subjects approval was obtained from the IRB at IPA (#:13October-001). All errors remain our own. Buehren: World Bank, nbuehren@worldbank.org; Goldstein: World Bank, mgoldstein@worldbank.org; Rasul: UCL, i.rasul@ucl.ac.uk; Smurra: UCL, a.smurra.11@ucl.ac.uk. 1 Introduction The recent history of many countries has been shaped by armed con‡ict. Since 1989, 123 countries have experienced such violence, resulting in 10 to 14 million deaths. Much of humanity remains at risk of con‡ict, with two billion residing in fragile states. The scourge of con‡ict is especially acute in Africa, where 39 out of 54 countries have experienced con‡ict since 1989 [UCDP 2018]. Con‡ict determines economic outcomes. It adversely a¤ects the accumulation of human and physical capital – through their destruction, displacement, and by dampening investment incen- tives due to the threat of future con‡ict. All of this can lead to economy wide underdevelopment traps [Collier et al. 2003, Besley and Reynal-Querol 2014]. At the individual level, a growing literature has described psychological legacies of con‡ict that impact economic decision making. Whether such legacies help or hinder post-con‡ict recovery is open to debate: a body of evidence in economics and political science has highlighted that exposure to con‡ict tends to foster pro- social behavior [Bauer et al. 2016, Hartman and Morse 2020], while others have argued exposure to con‡ict can strengthen parochial altruism, whereby in-group cooperation is reinforced at the expense of across-group cooperation [Heinrich and Boyd 2001, Bowles 2008]. Bauer et al. [2016] meta-analyze the literature linking con‡ict and pro-sociality. While they …nd that on many dimensions, those exposed to con‡ict are more pro-social later in life, somewhat puzzlingly, trust is one dimension of pro-sociality where null impacts are found (this work is summarized in Table A1a). This points to a need to better understand the subtleties of the relationship between con‡ict and trust. The importance of doing so is …rst order given trust is an outcome of intrinsic interest: it is the foundation for social and economic interactions and acts as ‘an important lubricant of the social system’ [Arrow 1974]. It underpins anonymized market exchange, and is key to cooperation in non-market exchange such as risk sharing. In this paper we provide novel theory and evidence to study the relationship between con‡ict and trust. We provide a road map for organizing the wider literature, to reconcile why heteroge- neous e¤ects might be found across studies, and to provide avenues for research on psychological legacies of traumatic events for economic preferences. We build on the multi-disciplinary literature on con‡ict and trust in two fundamental directions. First, we introduce the notion of conditional trust in one o¤ anonymized exchange with others, whereby individuals neither outright distrust or outright trust others, but rather can choose to use skills related to their self-e¢cacy to raise the cooperativeness of others. This incorporates ideas from psychology on post-traumatic growth, whereby survivors of serious trauma can go through a process leading to greater personal hardiness and resilience [Tedeschi et al. 1998, Tedeschi and Calhoun 2004]. We develop a framework linking con‡ict and trust to make precise what respondents can have in mind when reporting conditional trust in others. The framework captures factors that have been stressed in the earlier literature such as beliefs over others, the gains from cooperation and risk aversion, and highlights the key mediating role that perceived 2 self-e¢cacy plays in linking con‡ict and trust. By encompassing other proposed mechanisms linking con‡ict and trust such as beliefs about others, the framework o¤ers testable predictions to take to data to distinguish between explana- tions for why and how con‡ict shapes the long run formation of trust preferences. In so doing we tackle head on a critique of Bauer et al. [2016] that, ‘the research to date has done a far better job of establishing the e¤ect of war violence on later cooperation than of explaining it.’ Second, we construct a granular typology of experiences of con‡ict by combining information on: (i) a novel geo-coded measure of exposure to con‡ict; (ii) self-reported memories/recall of victimization during the civil war; (iii) ages at which the individual is exposed to con‡ict. Age of exposure is critical because memory formation begins largely at age three onwards, shaping later life beliefs and attitudes [Fehr et al. 2008, Malmendier and Nagel 2011]. It is thus useful to consider if an individual was aged 0-2 when in the vicinity of con‡ict (and so prior to signi…cant memory formation), and/or whether the individual was aged 3 and above when in the vicinity of con‡ict, and hence more capable of retaining …rst-hand memories of con‡ict. Our objective is to distinguish between individuals who are more likely to be traumatized or suppressing memories of con‡ict, those who are more likely to have direct …rst-hand accounts of con‡ict, and those whose narrative of con‡ict is more likely to be passed onto them second-hand. In examining the mediating role of perceived self-e¢cacy between these distinct experiences of con‡ict and trust, we provide novel insights on how trauma, memory and narratives link to the generation of self-e¢cacy, and how this matters for the long run formation of trust preferences. Our context is Sierra Leone. We survey 4 200 young girls and women in 2014, all of whom were born into the civil war that took place between 1991 and 2001. This was one of most brutal con‡icts experienced by a civilian population in recent times. As has been well documented, the violence largely did not operate along ethnic, religious or political lines [Conibere et al. 2004, Humphreys and Weinstein 2006, Bellows and Miguel 2009]. Instead, the con‡ict is widely recognized as being characterized by indiscriminate episodes of violence and abuse of civilians from all armed sides. As a result, the decade long con‡ict spread to nearly all Chiefdoms in Sierra Leone, with the majority of households potentially exposed to violence and acts of victimization. We measure generalized trust using a question with similar wording to the World Values Survey (WVS): “In general do you think people can be trusted, or that they cannot be trusted?” In contrast to the WVS, we allow respondents to answer yes, no, or it depends (and allow for don’t know or refusals).1 11% of our respondents say ‘no’ (outright distrust of others), 37% say ‘yes’, and 51% say ‘it depends’. Hence the modal answer highlights the prevalence of conditional trust among those born into con‡ict. This response does not re‡ect respondent uncertainty: none answered the trust question with ‘don’t know’ or refused to answer. 1 In the WVS the exact wording of the trust question is, ‘Generally speaking, would you say that most people can be trusted or that you need to be very careful in dealing with people? ’ Permitted answers are, ‘most people can be trusted ’, ‘need to be very careful ’, ‘don’t know ’, ‘no answer ’, or ‘other missing ’. 3 We use three approaches to measure exposure to con‡ict. First, we collect …rst-person accounts of direct exposure to violence during the civil war, as in much of the existing literature. Second and more innovatively, we use information from individual migration journals to construct respondent’s entire life history of Chiefdoms of residence, from their place of birth through to their current Chiefdom of residence. We then geo-reference information from migration journals to geo-coded con‡ict data from the Uppsala Con‡ict Data Program that codes con‡ict-related events at the Chiefdom-year level. This matching allows us to construct: (i) on the extensive margin, whether the individual was ever in the vicinity of con‡ict (so in the same Chiefdom-year as a con‡ict event); (ii) on the intensive margin, the cumulative exposure to con‡ict at any given age. Third, we construct a granular typology of experiences of con‡ict by combining information on: (i) the geo-coded measure of exposure to con‡ict; (ii) self-reported memories/recall of victimization during the civil war; (iii) ages at which the individual is exposed to con‡ict. The constructed typology distinguishes …ve mutually exclusive experiences: (i) those exposed to a ‘background narrative’ of the war because they were never in the vicinity of con‡ict and nor do they recall any victimization; (ii) those with direct …rst-hand experience of con‡ict because they were in vicinity of con‡ict from age 3 and recall being victimized; (iii) those who are traumatized or suppressing memories of con‡ict because they were in the vicinity of con‡ict from age 3 but they do not recall any form of victimization, including …ghting in their area; (iv) those socialized, because they were never in vicinity of con‡ict from age 3, yet they recall being victimized – their memories/recall stem from second-hand narratives passed onto them rather than direct experience; (v) respondents similar to those with a background narrative except they were exposed to violent events early in life, yet narratives of these events have not been passed onto them by their parents/guardians, perhaps because these caregivers are themselves traumatized. To formalize the link between exposure to con‡ict and trust, we develop a parsimonious frame- work of anonymized one-o¤ exchange with a partner. The model views trusting others as a gamble with potential outcomes depending on the probability the partner cooperates. To allow for condi- tional trust, we assume an individual can undertake costly actions to increase the probability of cooperation of the partner. Our approach views the cooperation of others as open to change, and not a …xed type. We use the term ‘self-e¢cacy’ to capture the perceived ability of an individual to increase the probability their partner cooperates. Self-e¢cacy can relate to the ability to persuade or negotiate with others to make them more cooperative [Bandura 1977]. The model generates a solution whereby individuals optimally choose to trust, not trust, or conditionally trust their partner depending on parameter values. The framework makes precise that conditional trust depends on factors: (i) that have been emphasized in the earlier literature linking con‡ict and trust, such as beliefs over others, the gains from cooperation and risk aversion; (ii) have been more discussed in the psychology literature and referred to as post-traumatic growth, captured in our framework through the notion of self-e¢cacy. We derive comparative statics on how key mediators of trust – self-e¢cacy and beliefs over the 4 trustworthiness of others – are impacted by con‡ict, how this subsequently impacts the likelihood of reporting conditional trust, and can have asymmetric e¤ects on the likelihood of expressing outright trust and outright distrust of others. This provides precise and testable microfoundations for whether and how con‡ict impacts trust preferences and allows us to potentially distinguish between alternative explanations. Our main results are as follows. We …rst examine the link between exposure to con‡ict and trust preferences, using our geo- coded measure of ever being in the vicinity of con‡ict. Those ever in the vicinity of con‡ict early in life are signi…cantly more likely to report conditional trust (‘it depends’) rather than outright trust or distrust in others, as measured 14 years later. The marginal e¤ect of ever being in the vicinity of con‡ict is to increase conditional trust by 77pp, corresponding to a 16% increase over the level of conditional trust among those never exposed to con‡ict. This increase in conditional trust arises because of signi…cant reductions in those reporting outright distrust of others (which falls by 29pp), and in outright trust of others (which falls by 5pp). We present a battery of checks to establish this result is replicated when: (i) using alternative econometric models; (ii) accounting for enumerator e¤ects; (iii) accounting for individuals being a¤ected by con‡ict in neighboring Chiefdoms (that might otherwise lead to measurement error of being in the vicinity of con‡ict). We also show the results do not re‡ect indecisiveness more generally among those exposed to con‡ict, and nor do they re‡ect those exposed to con‡ict having smaller social networks and so being more uncertain of the trustworthiness of others for anonymized exchange. To underpin a causal interpretation of the …nding, we: (i) demonstrate how the results vary in alternative subsamples related to the variation in geo-coded exposure to con‡ict; (ii) use a placebo check to strengthen the argument that actual exposure to violence matters, not other geographic factors correlated with con‡ict including pre-war confounders; (iii) use an IV approach, predicting exposure to con‡ict by exploiting geographic and temporal patterns of con‡ict. Second, we examine the link between experiences of con‡ict and trust preferences. Relative to those with the background narrative, those socialized have identical trust preferences over a decade later. This implies narratives passed onto children about the civil war are similar between those families for which violent events occurred outside their Chiefdom of residence, and those families for whom violent events occurred in their Chiefdom of residence but prior to their children forming direct memories of those events. Relative to those with the background narrative, those traumatized are signi…cantly more likely to report conditional trust by 108pp. This increase in conditional trust comes mostly from a fall in outright distrust (by 49pp). A similar pattern of marginal impacts on trust are found for those with direct experience of con‡ict. Hence relative to those with the background narrative, we …nd a comparable association between trust preferences for those traumatized or with direct experience of con‡ict. This implies that while the process of memory formation is important, an individual’s own recall 5 of con‡ict – the extent to which memories of violence are suppressed or not – is not associated with the long formation of trust preferences. This o¤ers mixed news for the literature using self-reported victimization to measure exposure to con‡ict (Table A1): on the one hand, among those with no recall that might otherwise be analyzed similarly, respondents that are traumatized have signi…cantly di¤erent trust preferences than those subject to the background narrative. Indeed, the formation of trust preferences does not di¤er between those with and without direct recall but with experiences of trauma or direct exposure to violence. For those whose parents are traumatized, relative to those with the background narrative, this experience also leads individuals to increase the likelihood of conditional trust by 89pp. They di¤er from the traumatized and direct experience groups in that this increase comes from a signi…cantly decline in outright trust in others (by 75pp) rather than a decline in distrust of others. The …ndings suggest the relationship between experiences of con‡ict and trust do not depend strongly on whether an individual or their parents are traumatized. In short, the …ve di¤erent experiences essentially boil down into two types, at least for the study of the long run relationship between exposure to con‡ict early in life and generalized trust: (i) those with the background narrative combined with those socialized; (ii) those with direct experience, combined with those traumatized and those whose parents are traumatized. Third, having established a robust link between con‡ict and trust we use the framework to guide our analysis of mechanisms underlying the relationship. We measure perceived self-e¢cacy using two approaches: the …rst relates to con…dence respondents express in being able to conduct entrepreneurial activities, in which one-o¤ anonymized exchange with others plays a major role. The second relates to personal traits, including those related to interpersonal interaction. We then establish that those ever in the vicinity of con‡ict have signi…cantly higher perceived self-e¢cacy than those never in the vicinity. Similarly, those with trauma/direct experience have signi…cantly higher perceived self-e¢cacy than those with the background narrative/socialized. What these experiences have in common is that respondents are subject to violence after age 3 when memories are being formed, and these shape later life attitudes [Fehr et al. 2008, Malmendier and Nagel 2011, Bauer et al. 2014, Giuliano and Spilimbergo 2014, Malmendier 2021]. Fourth, we re-examine the link between exposure to con‡ict and trust additionally controlling for self-e¢cacy. We …nd that conditional on ever being in the vicinity of con‡ict, self-e¢cacy correlates to trust preferences. A one standard deviation increase in self-e¢cacy is associated with a signi…cantly higher likelihood of reporting conditional trust: the magnitude of the e¤ect is 56pp, and signi…cant at the 1% level. This increase in conditional trust related to self-e¢cacy arises because of a near equal and opposite reduction in those reporting outright trust of others. Higher perceived self-e¢cacy is asymmetrically associated with the extremes of trust preferences: we can reject equality of shifts from yes/no towards conditional trust ( = 000). This asymmetry is precisely in line with the predictions of the framework. 6 Finally, we use the framework to consider other explanations. A prominent alternative is that con‡ict changes beliefs about others. To examine this hypothesis directly, we asked respondents, “How many people in [group] can you trust?” with potential responses being none, some, most, and all. We do so for ingroups of respondents, as well as for various outgroups. We …nd a remarkably consistent pattern of …ndings: for each and every group, respondents report being signi…cantly less likely to trust no one, being signi…cantly less likely to report trusting all, abut being signi…cantly more likely to trust some/most. That this applies equally to ingroups and outgroups is especially notable because it runs counter to the notion that exposure to con‡ict leads to parochial altruism. This interpretation is in line with narratives of the civil war in Sierra Leone being a con‡ict not rooted along ethnic, religious or political party divides. As described earlier, our work builds on an established literature on con‡ict and trust by introducing the notion of conditional trust, enriching the notion of exposure to con‡ict through the construction of speci…c experiences of con‡ict, and by developing a framework linking con‡ict and trust that highlights the key mediating role played by self-e¢cacy. By highlighting the subtleties in the link between con‡ict and trust, the analysis provides a road map for reconciling existing results in the literature and developing a future agenda. We also add to the literature on the formation and stability of economic preferences. As eco- nomic preferences emerge in childhood and remain stable from late adolescence, many studies have focused on their intergenerational transmission [Dohmen et al. 2012, Doepke and Zilibotti 2017, Kosse et al. 2020, Falk et al. 2021, Chowdhury et al. 2022]. We extend this to under- stand how experiences of con‡ict relating to parent-child interactions such as parental trauma and socialization, distinctively shape self-e¢cacy and the link between exposure to con‡ict and trust. Finally, our granular analysis of experiences of con‡ict contributes to a nascent literature on the role of memory and narratives in economic outcomes. On memory, the focus has largely been on formalizing processes of recall, or understanding limitations that recall biases or motivated beliefs have for choices [Mullainathan 2002, Gennaioli and Shleifer 2010, Bordalo et al. 2020, Fudenberg et al. 2022]. On narratives, economists increasingly recognize they shape expectations and macroeconomic outcomes [Akerlof and Snower 2016, Shiller 2017, Bénabou et al. 2018]. Nonetheless, empirical evidence on both the role of memory and narratives for individual decision making remains scarce. Our work provides new evidence that memory and narratives/socialization shape the link between con‡ict and trust through the generation of self-e¢cacy early in life. Section 2 develops a framework to make precise the notion of conditional trust, and microfound the link between con‡ict and trust. Section 3 details the Sierra Leonian civil war. Section 4 describes our data and key measures. Section 5 presents results on the relationship between exposure to con‡ict and trust. Section 6 uses the framework to guide the analysis of mechanisms. Section 7 discusses external validity, reconciles our data with a literature documenting negative impacts of con‡ict on psychological traits, and outlines an agenda for future work. Section 8 concludes. Additional results are in the Appendix. 7 2 A Framework for Understanding Conditional Trust We develop a parsimonious framework to characterize the concept of conditional trust, and micro- found the link between con‡ict and trust preferences. We do so in the context of a static model of random pairwise interactions. This is appropriate for anonymized market exchange, the e¤ec- tive functioning of which is critical to post-con‡ict recovery and economic growth more broadly [Knack and Keefer 1997]. As detailed later, such one-o¤ anonymized interaction closely matches the wording of the survey question we use to measure generalized trust. The framework and data are less appropriate for understanding repeated (non-market) exchange, such as in informal risk sharing. There is of course a vast theoretical literature exploring how trust/cooperation can be fostered in repeated interactions, where patience and punishment strategies, signaling reputations and third party enforcement are all more relevant for fostering trust. Despite its parsimony, the framework provides intuitive ways to encompass other prominent explanations on why con‡ict a¤ects trust. We derive comparative statics on how mediators of trust are impacted by con‡ict, how this impacts conditional trust, and can have asymmetric e¤ects on outright trust and outright distrust of others. This provides testable predictions to distinguish between some alternative hypotheses for the link between con‡ict and trust. 2.1 Set up To begin with, consider a simple scenario where individuals can only either choose to trust (T) or not trust (NT) another anonymous agent upon meeting them. We later introduce the notion of conditional trust (CT). We view trusting others as a gamble with two potential outcomes that depend on whether the partner cooperates or not. If the partner cooperates, the individual obtains _ a high payo¤  . This occurs with probability , which captures the prior belief an individual holds that their partner cooperates. With probability 1 ¡  the partner does not cooperate and the individual obtains the low payo¤ , which we normalize to zero. The alternative (safe) option is _  where  = 0  e for the individual not to trust their partner, and obtain a guaranteed payo¤ e   . Individuals have a concave utility function () with (0) = 0. If the individual trusts her _ _ partner, her expected payo¤ is (  ) + (1 ¡ )() = (  ), and if she does not trust then her payo¤ is (e ). Hence, if the agent’s choice is restricted between trusting or not trusting, the individual prefers to trust if: » (  ) T º NT :  ¸ _ =  (1) (  ) This captures the standard intuition that trust is more likely to occur if the individual’s belief over the trustworthiness of their partner (), the gains from cooperation (  1 ) are su¢ciently high, or the agent is less risk averse. We now introduce a third option: to conditionally trust the partner. We operationalize this 8 by assuming an individual can undertake actions at cost  2 [0  ], to increase the probability of the partner cooperating from  to () where () ¸  for all  2 (0 1), (0) ¸ 0 and (1) = 1. Our approach thus views the cooperation of others as open to change, and not immutable: trustworthiness is not a …xed ‘type’. Figure 1A shows an example () function (in bold). () need not be increasing everywhere, it is only assumed () ¸  for all  2 (0 1). Hence the dashed () function is also permissible. We use the term ‘self-e¢cacy’ to capture the ability of an individual, or their belief in them- selves, to increase the cooperativeness of their partner (() ¡ ). Self-e¢cacy intuitively relates to individual actions such as the ability to persuade or negotiate with others to make them more cooperative, and such actions come at cost  . This is narrower than the original formulation of self-e¢cacy in the psychology literature, where self-e¢cacy describes one’s ability to execute actions required to reach a particular goal [Bandura 1977]. In Bandura’s original formulation, perceptions of self-e¢cacy are derived from four sources: accomplishments, vicarious experiences, verbal persuasion and physiological states. 2.2 Solution _ The expected payo¤ from conditionally trusting the agent is ()(  ) + (1 ¡ ())(  ) ¡  = ¡ _ ()(  ) ¡  . With this notion of conditional trust, individuals di¤er along two dimensions: their beliefs about others, and their self-e¢cacy. The following conditions then identify when an individual would prefer to conditionally trust their partner rather than trusting them outright, or not trusting them altogether: _ CT º T : [() ¡ ](  ) ¸  (2) _ CT º NT : [() ¡ ](  ) ¸  The solution is in Figure 1B. There are three regions corresponding to when the individual would prefer to trust, not trust or conditionally trust the partner. In the bottom-left corner of {   }-space, when belief over the trustworthiness of the partner () is very low, then even if the cost of making use of self-e¢cacy ( ) is low, the individual would prefer to not trust because the marginal gains from trusting (proportional to () ¡ ) are too small. In the bottom-right corner, when  is very high then the individual does not need to use their self-e¢cacy, even if the cost of doing so is low, in order to trust. For intermediate values of , the highest levels of costs of self-e¢cacy can be borne to conditionally trust the partner.2 2 If  = 0 the model predicts the partner would never be trusted as CT º T when  = 0, and only the lowest  beliefs sustain not trusting as preferred to conditional trust. Given in our data, a large share of respondents report trusting others, the evidence does not support self-e¢cacy being costless. 9 2.3 Self-E¢cacy as a Mechanism Linking Con‡ict and Trust Conditional trust depends on factors that (i) have been emphasized in the earlier literature in economics and political science on trust and con‡ict such as beliefs over others (), the gains from cooperation (  1 ) and risk aversion; (ii) have been more discussed in the psychology literature. In this literature, largely based on case studies with survivors of serious trauma (accidents, rape, near-death experiences etc.), it is documented that exposure to extreme trauma can lead to greater self-re‡ection and greater personal hardiness [Tedeschi et al. 1998, Tedeschi and Calhoun 2004]. Anthropological studies have also noted individuals exposed to con‡ict display surprising resilience [Betancourt and Khan 2008]. This body of work thus stresses positive changes to individuals that can arise from exposure to con‡ict – collectively labelled as ‘post-traumatic growth’. This notion is captured in our framework through the idea of self-e¢cacy.3 Viewing this as a key mediator of con‡ict onto trust preferences, Figure 1C shows the com- parative statics on trust if belief in oneself rises from () to ¤ () where ¤ ()  () for all  2 (0 1). The …gure shows that for any given underlying distribution of {   }, an increase in self-e¢cacy from () to ¤ () increases the likelihood of conditional trust, and causes asymmetric reductions in the likelihood of not trusting and of trusting others. While we have emphasized positive changes or post-traumatic growth arising from exposure to con‡ict, we recognize that a separate literature examines how exposure to violence can have negative psychological impacts, such as increased anxiety. If so, the comparative statics in Figure 1C are reversed. We return to discuss this more fully later. Finally, we reiterate that our focus is on microfounding the link between con‡ict and trust – with self-e¢cacy being a key mediator. How traumatic events generate self-e¢cacy as a coping strategy has been much discussed in psychology [Bandura 1997]. Benight and Bandura [2004] describe a body of evidence showing how di¤erent types of traumatic experiences – including military combat, sexual and criminal assaults – can all generate self-e¢cacy as a coping mechanism. We are able to dig into this empirically using a re…ned typology of experiences of con‡ict, to see how direct exposure, memory and narratives/socialization each shape self-e¢cacy. 2.4 Alternative Mechanisms The framework captures other explanations for why con‡ict can impact trust, and so can poten- tially provide testable ways to distinguish between them. A prominent class of explanation is that exposure to con‡ict changes beliefs about the trustwor- thiness of others. One suggested mechanism is via parochial norms, whereby con‡ict strengthens 3 By making precise what respondents might have in mind when expressing conditional trust, our work relates to the long-standing debate around the standard World Values Survey measure of generalized trust, in that it captures beliefs over the trustworthiness of others rather than trust in others [Glaeser et al. 2000, Sapienza et al. 2013]. Our framework makes clear how the two relate and richens up the set of determining factors for generalized trust – beyond just preferences (risk aversion, inequality aversion, altruism) that are all captured in (). 10 social preferences towards ingroups [Heinrich and Boyd 2001, Choi and Bowles 2007]. In our framework this can be interpreted as  being a population-share weighted average of high beliefs of cooperation if matched to an ingroup member, and low beliefs of cooperation if matched to an outgroup member. How this impacts generalized trust preferences depends on the likelihood of matching with an ingroup member, and the increased strength of ingroup cooperation caused by con‡ict. We provide direct evidence on this below. Another variant is that con‡ict causes individuals to hold less extreme beliefs over others – say because both acts of violence and acts of protection are experienced. As a result, the distribution of  moves away from the tails. As Figure 1D shows, this direct e¤ect of con‡ict on trust through beliefs over others, , can then also also imply a reduction in the likelihoods of not trusting or of trusting, and an increase in the likelihood of conditional trust. The impact of exposure to con‡ict on trust preferences is then similar to when con‡ict raises the self-e¢cacy of individuals. A related explanation is that through con‡ict, individuals learn about the trustworthiness of others. Interpreting  as the prior belief held about the partner cooperating, and if () represents the posterior belief after some signal of the partner’s type is observed, then the model can be recast as one in which con‡ict allows individuals to learn about others. In the current formulation ()   so individuals only ever receive good news about the cooperativeness of others. With learning, individuals update their beliefs either positively or negatively. The latter is more likely to occur the higher the prior belief is to begin with. In other words, () lies above  for low  and lies below  for high . If exposure to con‡ict a¤ects trust preferences through this channel, the predicted e¤ects on trust preferences are the opposite to what occurs if con‡ict increases increases self-e¢cacy, potentially causing conditional trust to fall, and outright trust to rise.4 A second class of explanation more rooted in neoclassical approaches are that con‡ict increases the return to investing in social capital because of a loss of formal institutions during con‡ict, or as a way to build personal safety and security. These channels are captured in the gains from cooperation, 1 . As these rise, the model predicts outright trust rises, outright distrust falls, and conditional trust might rise or fall depending on (). Similarly if exposure to con‡ict increases risk aversion, the utility function becomes more concave so increasing the threshold in  for the individual to prefer trusting the agent to not trusting the agent. This has more pronounced e¤ects on shifting the tails of trust preferences from outright trust towards outright distrust of others, while conditional trust might rise or fall.5 4 Some have argued that a ‘security dilemma’ leads individuals to update their beliefs that others are not trustworthy in general [Walter and Snyder 1999], while arguments related to ‘collective coping’ suggest improved cohesion with neighbors in order to defend themselves [Pennebaker and Harber 1993]. » 5 If exposure to con‡ict increases risk aversion then the utility function becomes more concave increasing  () _ (  ) and the threshold in  for the individual to prefer trusting the partner. There is mixed evidence on link between exposure to con‡ict and risk aversion. Some studies …nd evidence of higher risk taking using lab-in-the-…eld experiments [Voors et al. 2012], while others …nd evidence of lower risk taking or a stronger preference for certainty [Callen et al. 2014, Moya 2018, Jakiela and Ozier 2019]. 11 Finally, the framework encompasses an alternative interpretation of what respondents mean when they report conditional trust. If individuals provide this reply whenever they are close to indi¤erence to trust or not trust others, the region of conditional trust just spans the threshold for  described in (1) and changes in self-e¢cacy have no impact on conditional trust – something that is testable (and indeed ruled out by the evidence we provide below). 3 Study Context: The Sierra Leonian Civil War The framework developed is general and can be potentially applied to many con‡ict scenarios. To narrow our focus, we discuss the speci…cs of the 1991-2002 civil war in Sierra Leone, two features of which have important implications for our analysis. First, the violence largely did not operate along ethnic, religious or political party divides [Conibere et al. 2004, Humphreys and Weinstein 2006, Bellows and Miguel 2009]. As such, it might not be expected to strengthen ingroup ties or create increased animosity towards outgroups (as we later show). Second, it is widely recognized as being characterized by indiscriminate episodes of violence and abuse of civilians from all armed sides in the con‡ict. The con‡ict spread in waves, rather than being concentrated in speci…c areas/cities. As a result, the decade long con‡ict spread to nearly all Chiefdoms in Sierra Leone, leaving the majority of households – and individuals of both genders and all ages – potentially exposed to violence and acts of victimization. On the eve of the con‡ict, Sierra Leone was ranked 153rd out of 181 countries in terms of GDP per capita. Three decades of one-party predatory rule by Siaka Stevens …rst and General Momoh later had left the country on the brink of economic collapse. The con‡ict began in March 1991 when the Revolutionary United Front (RUF), with support from the special forces of Charles Taylor’s National Patriotic Front of Liberia (NPFL), intervened to overthrow the Momoh government with the stated aim of ending its period of corruption and ineptitude [Reno 1995, Richards 1996]. The RUF were the main perpetrators of violence in the con‡ict, but had no political, ethnic or religious a¢liation [Bellows and Miguel 2009]. They conscripted disenfranchised youths from across Sierra Leone as well as Liberian refugees. They also used child soldiers from Sierra Leone, many of whom were forced to abuse their own communities. Chiefdoms bordering Liberia experienced the most intense and prolonged exposure to violence because the RUF and NPFL entered Sierra Leone via Liberia, their headquarters were based there, and these Chiefdoms were also rich in diamonds – control of which helped …nance the con‡ict. By 1992 the Sierra Leone Army (SLA) had taken over the government. Like the RUF, they were drawn from across ethnic groups, and also employed child soldiers – a quarter of the SLA was aged below 18. As is well documented, there was a degree of cooperation between the SLA and the RUF – they coordinated movements to avoid direct battles, and sometimes made pro…t sharing arrangements in diamond areas. The extent of cooperation was such that some soldiers 12 apparently fought for the SLA by day and the RUF by night. As a result, the SLA engaged in looting to such an extent that they became known as ‘rebels by night’ and over time became largely indistinguishable from rebels in terms of their violence towards civilians [Bellows and Miguel 2009]. In 1993 a government o¤ensive supported by ECOMOG pushed the RUF back towards the Liberian border. Notwithstanding this e¤ort, the RUF regained territories, approaching the capital Freetown in 1995. This period saw the con‡ict spread to many more Chiefdoms throughout Southern and Western Sierra Leone. In 1997, the Armed Forces Revolutionary Council (AFRC) split from the SLA, staged a coup, united with the RUF and took over Freetown [Keen 2005]. Looting, rape, murder followed, alongside a complete collapse of most state institutions. To protect themselves from attack by the RUF and SLA, many communities organized Civil Defense Forces (CDF). While many acts of heroism were conduced by CDF …ghters, over time some CDF members also began to abuse civilians, enter the illicit diamonds trade, and utilize child soldiers [Keen 2005]. After a further brutal rebel attack on Freetown in 1999 and failed peace talks, in 2000 the UK, UN and Guinea intervened to conduct a disarmament campaign and secure a peace treaty – involving the demobilization of the RUF and the reintegration into society of its members. In January 2002 the civil war was declared to be over. On the eve of the con‡ict in 1991, the population of Sierra Leone stood at 45 million. By its end, 25 million individuals had been displaced from their homes, 50 000 were casualties, and approximately 18 000 had died. Despite the collective trauma, the country has seen a period of post-con‡ict political stability and economic recovery. National elections were held in 2002, closely followed by the …rst local elections in decades in 2004, and the post-con‡ict decade was one of largely sustained economic growth. We use the Uppsala Con‡ict Data Program (UCDP) as our key data source on con‡ict events during the civil war. This data codes con‡ict-related events at the Chiefdom-year level, based on o¢cial, NGO or journalistic records. A con‡ict event is de…ned as the use of armed forced by an organized actor against another, or against civilians, that resulted in at least one death. The data records the exact location and time of each event, the parties involved, the number of civilian casualties (and whether they occurred at the hands of rebels or state forces), and fatalities on each side. We retain only those events for which the Chiefdom can be retrieved with con…dence: this results in the selection of 1 297 events (out of 1 495 available in the UCDP data for the Sierra Leone civil war). We aggregate events within a Chiefdom and year to construct measures of violent con‡ict by Chiefdom-year.6 6 This UCDP covers con‡icts that cause at least 25 deaths in a given year. The alternative source to the UCDP used in con‡ict research is from the Armed Con‡ict Location and Event Data Project (ACLED). We use the UCDP because: (i) it spans the entire period of the civil war in Sierra Leone (while ACLED starts only in 1997); (ii) ACLED has looser de…nitions of events and actors, and thus includes protests and troop movements for example; (iii) UCDP focuses on episodes of explicit violence against civilians – that are most relevant for our 13 Figure 2 shows the geographic spread of the con‡ict across the 149 Chiefdoms in Sierra Leone. Each panel shows the cumulative con‡ict in each Chiefdom, at four key dates: its initiation in 1991, when the RUF began to be pushed back by the SLA in 1993, the initial taking of Freetown by the RUF in 1997 and the con‡ict end in 2001. This shows how the con‡ict spread from Chiefdoms close to the Liberian border, and this spread occurred in waves driven by political events exogenous to factors speci…c to any given Chiefdom. In all regions, we observe wide variation in cumulative con‡ict across Chiefdoms, and even among neighboring Chiefdoms. Figure 3 shows the time series variation in con‡ict by Chiefdom, grouping the 14 districts into three regions. Within Chiefdom, a grey cell indicates a year of no con‡ict, and a red cell indicates a year with con‡ict. The Figure highlights the intermittent nature of con‡ict within Chiefdoms: (i) while 98% of Chiefdoms experienced at least one con‡ict spell, more than 90% of Chiefdoms experienced gaps of at least a year between con‡ict episodes; (ii) the average length of a con‡ict episode within a Chiefdom is 16 years; (iii) the autocorrelation in casualty numbers within a con‡ict spell quickly falls, so as with the extensive margin, the intensity of con‡ict rises and falls sporadically within Chiefdoms over time. This evidence all suggests that during the con‡ict it would be di¢cult to migrate across Chiefdoms to avoid violence. Indeed, below we document that few respondents migrated during the con‡ict, with the majority of cross-Chiefdom migration occurring post-con‡ict. Moreover, given the lack of contraceptive availability in Sierra Leone at the time, it would be equally di¢cult for parents to strategically delay fertility during the con‡ict in order to ensure their child was less exposed to violence. The foot of Figure 3 shows a histogram of year of birth in our sample: we observe a slight baby boom in the last year of the con‡ict, when it would be most apparent that the con‡ict was over. We later examine in more detail the robustness of our …ndings to concerns related to strategic migration or delay in fertility timing to avoid exposure to violence.7 4 Data, Key Variables and Empirical Method 4.1 Data Sources Our data was collected as part of an evaluation of a life skills intervention targeting young girls and women in 200 villages in four districts of Sierra Leone: Kambia, Moyamba, Pujehun and Port Loko [Bandiera et al. 2020]. Our baseline data was collected in 2014, from a representative sample of young girls and women aged 12 to 24 resident in these villages. Hence all respondents were research question on the formation of trust preferences; (iv) the UCDP data has been argued to have higher quality for the kinds of research question that we seek to address [Eck 2012]. 7 Data from the 2004 census shows that pre-con‡ict, there is a steady upward trend in cohort size – starting from the 1970s – and this does not break with the con‡ict starting in 1991, nor does this trend change in the con‡ict years. Overall, census data do not suggest aggregate falls in fertility away from their long run trend over the con‡ict period. The census data re-con…rms the …nding from our sample shown in Figure 3, that there is an increase in cohort size in the …nal year of the con‡ict. 14 born between 1990 and 2002 and we trim the sample to focus on those born during the civil war. Our working sample comprises 4 188 respondents all born at some point of the civil war between 1991 and 2001 inclusive. This corresponds to the daughters of the generation of con‡ict survivors surveyed in Bellows and Miguel [2009].8 Respondents are on average aged 18, 63% are in a relationship, 53% have ever been pregnant and 30% are married. In terms of family background, 16% belong to a ruling family, and 10% (3%) report their father (mother) to have attained higher levels of education.9 4.2 Measuring Trust Preferences We measure generalized trust at baseline using a question with similar wording to the World Values Survey or Afrobarometer: “In general do you think people can be trusted, or that they cannot be trusted?” However, in contrast to these cross-country surveys, we allow respondents to answer yes, no, or it depends (as well as allowing for don’t know or refusals).10 11% of respondents say no (so report outright distrust of others), 37% say yes, and 51% say it depends. Hence the modal answer is not one that is routinely permitted: it highlights the prevalence of conditional trust among those born into con‡ict. This response does not re‡ect respondent uncertainty: no respondents answered the trust question with ‘don’t know’ or refused to answer. We later show that trust preferences do not correlate with indecisiveness as measured in other survey modules.11 4.3 Measuring Exposure to Con‡ict Migration Journals We measure individual exposure to con‡ict using information from mi- gration journals that we collected from each respondent as part of our midline survey in 2016. We asked each respondent their entire history of Chiefdoms of residence, from their place of birth through to their current Chiefdom. These migration journals thus build migration histories span- ning con‡ict and post-con‡ict years. We record each Chiefdom resided in for at least six months over this period, for each respondent. For each migration spell we record the Chiefdom, year of 8 The survey modules covered a sequence of sensitive subjects – not just those related to con‡ict, but also topics relevant for the underlying evaluation such as risky behaviors, and experience of intimate partner violence. As a result, all surveys were conducted by a trained group of women enumerators. 9 A respondent belongs to a ruling family if household members are eligible for the role of Paramount Chief (district -level) and sub-Chief (chiefdom level). This pre-colonial institution remained the form of sub-national governance until 2004 in Sierra Leone. Eligibility for ruling positions is hereditary, and several families can be eligible within a Chiefdom. Having higher education is completing at least one year of Senior Secondary School (SSS1/form 4 and above), including Teacher’s College. Respondents identify as belonging to one of ten ethnic groups/tribes. In our analysis we include dummies for the three largest ethnic groups: Mende, Temne and Limba. The other 7 (corresponding to 3% of the sample) and are the excluded category throughout. 10 In the WVS the exact wording of the trust question is, ‘Generally speaking, would you say that most people can be trusted or that you need to be very careful in dealing with people? ’ Permitted answers are, ‘most people can be trusted ’, ‘need to be very careful ’, ‘don’t know ’, ‘no answer ’, or ‘other missing ’. 11 Our trust measure was collected prior to the Ebola outbreak, that started after our baseline data collection. 15 the start/end of the spell.12 Table A2 shows that among the sample for which we have migration journals for, trust pref- erences do not correlate with attrition (Column 1), and this remains true if we condition on individual and village controls (Column 2).13 The 4 188 respondents collectively report 4 795 migration spells. 31% still reside in their Chiefdom of birth. Only 14% migrated during the con‡ict – the majority of cross-Chiefdom migration occurs post-con‡ict (and only 5% of all migration spells occur post-Ebola). While our baseline sample was drawn in 2014 from those resident in 23 Chiefdoms across 4 districts for evaluation purposes, given migration histories, respondents are drawn from 101 Chiefdoms of birth across 14 districts. Panels A and B of Figure A1 shows cumulative con‡ict in these Chiefdoms of birth, and Chief- doms of residence when migration journals were collected. For those respondents who migrated during the con‡ict, the lower panels provide evidence on how origin and destination Chiefdoms compare in terms of con‡ict intensity. We calculate: (i) the di¤erence in the total number of episodes of violence at the time of migration between their Chiefdom of origin and destination; (ii) the di¤erence in the total number of episodes of violence that took place between their origin and destination Chiefdom post-migration. Panels C and D show histograms of each of these di¤erences. For both measures there is a mass point in the di¤erence of con‡ict intensity at zero episodes. Hence for the 14% of respondents who migrated during the con‡ict, we do not see evidence of systematic migration towards Chiefdoms with lower con‡ict, either overall or post-migration. To construct individual measures of exposure to con‡ict we then geo-reference the information from migration journals to UCDP geo-coded con‡ict data. This matching by Chiefdom-year allows us to construct: (i) on the extensive margin, whether the individual was ever in the vicinity of con‡ict (in the same Chiefdom-year as a con‡ict event); (ii) on the intensive margin, the cumulative exposure to con‡ict at any given age. To get a sense of what it means to be in the vicinity of con‡ict, we note that on average, Chiefdoms have area of 480km2 , equivalent to one third the size of London, or the same as Nashville. 12 Two points are of note. First, displacements of residence within Chiefdom are not recorded, nor are those occurring across Chiefdoms but for less than six months. This di¤ers from the measures of internally displaced individuals recorded in UCDP data (that measure displacements to any di¤erent location, including those within Chiefdom) and also might di¤er from what individuals provide in self-reported victimization data. For example, there are accounts from the con‡ict of entire communities temporarily displacing themselves to the outskirts of their villages in advance of troop arrivals or battles. Second, we asked respondents to provide the reason for each change of residence: in most cases this is missing, re‡ecting that girls were in childhood during the con‡ict and often do not know the reason for household migration. 13 Reliable data on age is determined through age and year of birth being consistent across surveys, the place of birth is known and the migration journal does not have missing spells. Alternatively, there might be minor inconsistencies in reported age across surveys, but the respondent is either exposed or not exposed to con‡ict regardless of the implied year of birth used. 98% of the tracked sample follow one of these routes. 16 Descriptives Table 1 describes the exposure to con‡ict measure. For the full sample in Column 1 we see that 62% of respondents have ever lived in a Chiefdom where con‡ict occurred. 44% were exposed at birth (so born directly into con‡ict). The vast majority (97%) were exposed in their Chiefdom of birth – very few individuals were able to migrate prior to being exposed in their Chiefdom of birth. This re‡ects the low levels of migration during the con‡ict, and di¢culty of using migration as a strategy to avoid con‡ict. The average respondent was …rst exposed to con‡ict at age 15, and then cumulatively exposed to around six violent events during the war. For individuals who migrated during the con‡ict, they did so once, while 68% of respondents migrated post con‡ict. Columns 2 and 3 compare those who migrated during the civil war to those that did not. Migrants are more likely to be in the vicinity of con‡ict at some point, cumulatively exposed to more con‡ict events, and also more likely to migrate post-con‡ict. Columns 4 and 5 compare exposure to con‡ict between those from ruling families to those from non-ruling families. Many accounts of the civil war highlight that the one group the RUF targeted for attacks were members of traditional authority households, akin to local elites [Richards 1996, Keen 2005]. However, we do not …nd stark di¤erences in their patterns of exposure to con‡ict on either extensive or intensive margins, or their propensity to have migrated during the con‡ict. Table A3 shows correlates of exposure to con‡ict on the extensive and intensive margins (Col- umn 1 and 2). The main correlate is age (equivalent to year of birth). Although the dummies for tribe are collectively signi…cant this likely re‡ects their geographic distribution. In line with narratives of the war, other factors such as family background, being part of a ruling family and religion are not predictive of exposure to con‡ict. 4.4 Constructing Experiences of Con‡ict To provide granular evidence on how exposure to con‡ict impacts trust preferences, we exploit the full richness of our data to construct a typology of experiences of con‡ict. This combines informa- tion on: (i) the geo-coded measure of exposure to con‡ict; (ii) self-reported recall of victimization during the civil war; (iii) ages at which the individual is exposed to con‡ict. This detailed typology provides insights into how memories and narratives shape experiences of con‡ict, and can lead to the generation of self-e¢cacy. 4.4.1 Self-reported Victimization At baseline, we collected …rst-person accounts of direct exposure to …ghting and violence during the civil war. The forms of victimization we asked respondents about are comparable to those in Bellows and Miguel [2009]: whether there was …ghting in their area, they were personally harmed, their family was harmed, they were a refugee overseas, or they were internally displaced. For each victimization type, we ask individuals whether they recall being victimized in this way during the 17 con‡ict. They can respond with yes, no, or state they were too young/do not recall. For those that report they were too young/do not recall no recollections on all dimensions of victimization, we refer to them as not recalling the war. 54% of individuals recall the civil war or are willing to discuss victimization during it. Of these, 80% experienced victimization …rst-hand, while the other 20% recall the war but report not being victimized during it. Figure 4A details the victimization data. The left hand panel shows how self-reported victimization relates to the geo-coded measure of exposure to con‡ict. Among those in the vicinity of con‡ict, 64% recall the war, and conditional on recall, 87% recall being victimized. Of the sample geo-coded never in the vicinity of con‡ict, 39% recall the war, and conditional on recall, 63% recall being victimized. The right hand panel of Figure 4A shows that among those self-reporting victimization, 78% observed …ghting in their area, 62% were internally displayed, 12% were a refugee overseas, 41% had their family harmed, and 6% were personally harmed. This ranking …ts accounts of the civil war in Sierra Leone, with con‡ict occurring in nearly every Chiefdom, the majority of the population being displaced, and violence being in‡icted on civilians by all parties. The fact that respondents were born into the con‡ict also helps explain the low levels of personal harm reported relative to reports of family members being harmed.14  15 As expected, there is an imperfect mapping between the geo-coded and self-reported measures of exposure to con‡ict. Of those geo-coded to be in the vicinity of con‡ict, 56% recall the war and being victimized. Of those geo-coded not ever be in the vicinity of con‡ict, 25% recall being victimized during the con‡ict. These discrepancies are informative. Our sample were all born during the con‡ict, and hence might be too young to themselves recall events – but rather report narratives of the con‡ict handed down to them from parents and others. Digging into this further, Figure 4B details how recall of the civil war, reporting any form of victimization, and the geo-coded measure of exposure to con‡ict, all vary with age. All three indicators increase with age, as expected. For cohorts born later in the con‡ict (and so with a lower age in 2001), the gap between victimization rates and actual vicinity to con‡ict are greatest. These series narrow with age (as we move to the right hand side of the …gure). For those born early in the con‡ict, and so aged 6 and above by the end of the civil war, there is near convergence in rates of victimization and the geo-coded measure of residing in the vicinity of con‡ict. This further 14 Our sample of respondents born into the civil war corresponds to the daughters of the generation of con‡ict survivors surveyed in Bellows and Miguel [2009]. They …nd that 44% of respondents reported a family member being killed during the con‡ict; 35% had a family member being injured (close to the …gure in our sample); 38% report being refugees during the con‡ict (that is far higher than in our sample). 15 Among the sample reporting any victimization, the correlates of being in the vicinity of con‡ict remain largely unchanged (Table A3, Columns 3 and 4). Columns 5 to 11 in Table A3 shows correlates of victimization: again age is the predominant predictor of recalling victimization, and is positively correlated to reports of family being harmed, being internally displaced or there being …ghting in their area. In line with accounts of the civil war, we …nd those from ruling families are more likely to report experiencing some form of victimization, and speci…cally, being a refugee overseas. 18 hints that for girls born nearer the end of the con‡ict, this gap between self-reported victimization and actual exposure to con‡ict is partially explained by a process of socialization, where they have heard narratives about the civil war, and this in‡uences self-reported victimization a decade later. 4.4.2 A Typology of Experiences We combine the geo-coded measure of exposure to con‡ict, self-reported recall and victimization data, together with information on ages of exposure to construct a typology of experiences. Our objective is to distinguish between respondents who are more likely to be suppressing memories of con‡ict, those that are more likely to have direct …rst hand accounts of con‡ict, and those whose narrative of con‡ict is more likely to be second-hand. In this construction, age of exposure is critical because memory formation begins largely after age three. It is thus useful to consider if an individual was aged 0-2 when in the vicinity of con‡ict (and so prior to signi…cant memory formation), and/or whether the individual was aged 3 and above when in the vicinity of con‡ict, and hence more capable of retaining …rst-hand memories of con‡ict.16 For each period in a respondent’s life – when aged 0-2 and when aged 3 onwards, we consider whether the individual: (i) was in the vicinity of con‡ict using the geo-coded measure; (ii) recalls any form of victimization. This provides an 8-way mutually exclusive classi…cation of respondents, as shown in the left hand side of Figure 5. We use this 8-way classi…cation to naturally group respondents into …ve distinct types of experience of civil war. To begin with, consider those with no recall of victimization during the civil war, corresponding to rows 1, 3, 5 and 7 in Figure 5. While these might typically be grouped together in earlier work using self-reported recall, our data allows a …ner gradation among them. Respondents in row 1 are never in the vicinity of con‡ict at any stage of childhood, nor do they recall victimization. The civil war took place in their country, but they did not experience it …rst hand either because they were born after the violence took place (or far more rarely, they moved Chiefdom prior to violence occurring). Nor do they recall any victimization. We refer to this group as having been exposed to a ‘background narrative’ experience of the civil war. They form a natural benchmark group from which to consider other experiences: 27% of respondents are of this type.17 Consider next those in rows 3 and 7 in Figure 5: these individuals were all in the vicinity of con‡ict at age 3 or later – so have a higher likelihood of having memory-based recall of the war. Yet they do not recall any form of victimization, including …ghting in their area. We thus refer 16 Pro-social motivations develop from age 3 and into adolescence, although much of this evidence is based on WEIRD samples [Fehr et al. 2008]. As children rely on parents for priors that shape their responses to new information, socialization during childhood can also a¤ect attitudes [Dohmen et al. 2012, Doepke and Zilibotti 2017, Kosse et al. 2020, Falk et al. 2021, Malmendier 2021, Chowdhury et al. 2022]. 17 This benchmark group comprises those that survived the civil war but did not experience it …rst hand. This is not the same as a control group that never experienced civil war altogether, or for whom civil war was a distance event. Such counterfactuals simply do not exist in the context of the civil war in Sierra Leone. 19 to these groups as ‘traumatized’ in that their experience of the con‡ict leads them to suppress memories – they are unwilling to talk about the con‡ict or any form of victimization. 9% of respondents fall into this type. Those in row 5 were not in the vicinity of con‡ict from age 3 onwards, nor do they recall any victimization. These respondents are similar to those with a background narrative except they were exposed to violent events when aged 0 to 2 – yet narratives of these events have not been passed onto them by their parents or others. Their experience of con‡ict is …ltered through their parents, who might either be deliberately trying to protect their children, or who themselves might be traumatized and so unable to pass on such narratives. We refer to this group of respondents as those with parental trauma: 14% fall into this type. We next consider those with recall or victimization (rows 2, 4, 6 and 8 in Figure 5). Consider those in rows 2 and 6: neither group were in the vicinity of con‡ict from age 3 onwards, when their direct memories would have formed. Yet they recall being victimized. We thus refer to these individuals as having the experience of being ‘socialized’, in that their victimization recall stems from narratives told to them rather than from their own direct experience. This includes those in row 4: like those with parental suppression (row 5), this group did have violent events occur in their Chiefdom of residence when aged 0 to 2, but the key distinction from those in row 5 is that they are able to recall such victimization – suggesting the narratives passed on di¤er between the socialized and parental trauma experiences. 16% of respondents fall into this type. Finally, those in rows 4 and 8 were all in the vicinity of con‡ict from age 3 onwards, when direct memories are more likely to form, and they recall being victimized. We refer to these individuals as having experienced con‡ict ‘directly’. 25% of respondents fall into this type. We thus construct a mutually exclusive …ve-way typology of experiences of civil war: those with a background narrative, those socialized, those with direct experience, those traumatized, and those whose parents are traumatized. Each group is well represented in our sample (with the smallest group still having 9% of respondents in it). Although there are clearly issues related to measurement error, alternative groupings and interpretations, given such experiences have not previously been considered in work linking con‡ict and trust, and they potentially provide insights into how con‡ict generates self-e¢cacy, we proceed with this speci…c formulation. 4.4.3 Descriptive Evidence on Experiences of Con‡ict To underpin the usefulness of the typology, Table 2 provides descriptive evidence by experiences of con‡ict. Panel A focuses on family background. As expected, those traumatized or with direct experience of con‡ict are older (and so born earlier into the con‡ict). Those with direct experience or trauma are also most likely to have migrated during the civil war. Panel B focuses on the geo-coded measure of exposure to con‡ict, the key advantage of which is that it can be measured for each experience. By de…nition, those subject to the background 20 narrative are never in the vicinity of con‡ict. Comparing those traumatized to those with direct experience, these groups are exposed to a similar number of violent events with similar numbers of total and civilian casualties. Where they di¤er is the former group are signi…cantly more likely to have been exposed to events where civilians were targeted (especially by government forces) ( = 011 063), and they have been exposed to signi…cantly more years of con‡ict overall ( = 000 008). The …nal column shows this pattern remains largely robust to conditioning on a polynomial in age and district of birth …xed e¤ects. Panel C focuses on self-reported victimization, that by de…nition is reported only for those with direct experience and those socialized. By construction, those with the background narrative recall the war but report no victimization. Comparing victimization accounts of those with direct experience to those socialized, we see the former report signi…cantly higher rates of each form of victimization. The …nal column shows that once we condition on age and district of birth …xed e¤ects, the signi…cant di¤erences are that those with direct experience are more likely to be internally displaced ( = 017), and those socialized are signi…cantly more likely to be a refugee overseas ( = 048). 4.5 Measuring Self-E¢cacy Self-e¢cacy describes one’s ability to execute actions required to reach a particular goal [Bandura 1977]. In Bandura’s original formulation, perceptions of self-e¢cacy are derived from four sources: accomplishments, vicarious experiences, verbal persuasion and physiological states. We focus on one manifestation of this: the belief in oneself to increase the cooperativeness of others (()). We use two baseline survey modules to proxy this belief. The …rst relates to con…dence respondents express in being able to conduct various tasks relevant for entrepreneurial activities, in which one-o¤ anonymized exchange with others plays a major role. Although in the context of entre- preneurship, some of these questions link directly to interpersonal skills, powers of persuasion and cooperation with others.18 The second module relates to personal traits, including those related to interpersonal interaction. These questions related more to underlying notions of self-e¢cacy, grit, determination, locus of control, self-image and some elements of interpersonal interaction.19 18 On entrepreneurial skills we asked a sequence of questions to elicit self-reported ability to perform the following tasks: (i) run your own business; (ii) identify opportunities to start a new business; (iii) obtain credit to start a new business/expand an existing business; (iv) save in order to invest in future business opportunities; (v) make sure your employees get work done properly; (vi) manage …nancial accounts; (vii) bargain to obtain cheap prices when you are buying inputs; (viii) bargain to obtain high prices when you are selling outputs; (ix) protect your business assets from harm by others; (x) collect money someone owes you; (xi) …nd information about paid work opportunities in your community. Responses to each were recorded on a 1-10 point Likert scale ranging from ‘de…nitely cannot do this’, to ‘can de…nitely do this’. Some of these dimensions pick up persuasion or communication skills, thus linking to a evidence from laboratory settings showing that pre-play communication, even if cheap talk, can enhance subsequent cooperation and trust [Charness and Dufwenberg 2006, Vanberg 2008, Serra-Garcia et al. 2013, Ederer and Schneider 2022]. 19 Specifally, we asked respondents’ agreement with the following statements about themselves: (i) if I start working on a task, I de…nitely see the end of it no matter how di¢cult it is; (ii) while doing any task, it is 21 Rather than pick speci…c questions, we use all of them to construct separate indices for entre- preneurial skills and personal traits, using the approach of Anderson [2008]. We also construct an overall index of self-e¢cacy that uses all questions across measures. We focus on this as our baseline measure, although also show impacts of each index separately, as well as the impacts of each and every component. Each index is standardized with respect to the control group (those not in the vicinity of con‡ict), so the impacts on the index can be interpreted as e¤ect sizes.20 4.6 Empirical Method To analyze the relationship between exposure to con‡ict during the civil war and generalized trust measured more than a decade later, we treat generalized trust preferences – yes, no, it depends – as a sequence of unordered alternatives. The framework developed makes clear that allowing the ‡exibility for exposure to con‡ict to have non-monotonic impacts on trust preferences is key for understanding their link. We use the following multinomial logit model: µ ¶ P [trust=Yes j   X ]  =    + X    (3) P [trust=It Depends j   X ] µ ¶ P [trust=No j   X ]  =    + X    P [trust=It Depends j   X ] In our baseline speci…cation we consider the geo-coded measure of exposure, so  = 1 for those ever in the vicinity of con‡ict, namely if any con‡ict episode took place in the respondent’s chiefdom of residence while they resided there. We then examine the relationship between experiences of con‡ict and trust preferences. This sheds light on the roles that the suppression of memories of con‡ict, direct accounts of con‡ict, and second-hand narratives of con‡ict all play relative to those with a background narrative. Standard errors are clustered by village of residence at baseline.21 In each batch of three columns we show the average marginal e¤ects from (4) for each trust response  2 f     g:   = E (P [ =  j X   = 1] ¡ P [ =  j X   = 0])  (4) important for me to do it better than others; (iii) if I have the chance, I would make a good leader; (iv) I want to be a respectful person in my village; (v) I do not care what others think about my success or failure; (vi) I am in control of what happens in my life; (vii) I save regularly; (viii) a person can get rich by taking risks; (ix) I often make plans for the future; (x) I believe that my future is determined by luck no matter how hard I work. Responses to each were again recorded on a 1-10 point Likert scale ranging from ‘de…nitely false’, to ‘de…nitely true’. 20 An alternatively widely used scale to measure self-e¢cacy is that developed by Schwarzer and Jerusalem [1995], although other researchers have used variants of this to capture the same basic concept [McKelway 2021]. While self-e¢cacy is distinct from traits such as self-esteem or locus of control, they are highly related and self-e¢cacy scales are indeed often validated through their correlations with these other constructs. 21 The controls in X include age (where we allow for a cubic polynomial in age at baseline), father’s education, ruling family status, religion and tribe dummies, village size and the average poverty score of households in the village of residence at baseline, and district of birth dummies. 22 The coe¢cient of interest   is the average marginal e¤ect on trust response  between those not ever in the vicinity of con‡ict ( = 0) to those in the vicinity of con‡ict ( = 1), holding constant all other covariates X . These estimates are akin to intent-to-treat e¤ects. The later results on experiences of con‡ict and trust are more akin to treatment-on-the-treated e¤ects. 5 Con‡ict and Trust 5.1 Vicinity to Con‡ict Table 3 presents our baseline results linking the geo-coded measure of being in the vicinity of con‡ict, when respondents were aged 0 to 11, to trust preferences expressed 14 years later. We …nd that, irrespective of the set of controls, those who were in the vicinity of con‡ict are signi…cantly more likely to report conditional trust (‘it depends’) rather than outright trust or distrust in others. When all covariates are conditioned on in Columns 2a-2c, the marginal e¤ect of ever being in the vicinity of con‡ict is to increase conditional trust by 77pp, corresponding to a 16% increase over the level of conditional trust among those never exposed to con‡ict ( = 0). The estimated e¤ect is precisely estimated, being statistically signi…cant at the 1% level. This increase in conditional trust arises because of signi…cant reductions in those reporting outright distrust of others (the marginal e¤ect on not trusting others reduces by 28pp or 23% relative to those not in the vicinity of con‡ict), and in outright trust of others (the marginal e¤ect on trusting others falls by 5pp or 13% relative to those not in the vicinity of con‡ict). We cannot reject equality of shifts from yes/no towards conditional trust ( = 534). Three points are of note. First, the fact that exposure to con‡ict has persistent e¤ects on trust preferences means there is no strong general equilibrium e¤ect, where all individuals – even those never exposed – converge to conditional trust preferences as a best response to those that were exposed to con‡ict or through some process of collective coping [Pennebaker and Harber 1993, Lyons et al. 1998]. This already suggests that beliefs about others is unlikely to fully explain how con‡ict shifts trust preferences. Second, that trust preferences non-monotonically move in from both extremes towards condi- tional trust is revealing. Had we used the simpler categorization of yes/no answers and modelled those as ordered choices, then those answering it depends would be shifted to these tails. This makes it more likely that no change in trust preferences is detected, which would be in line with the meta-analysis of Bauer et al. [2016], that …nds on many dimensions of pro-sociality, those exposed to con‡ict tend to be more pro-social later in life, but imprecise null impacts are found for trust. Hence one way to reconcile varying results in the literature linking con‡ict and trust is to recognize the value in allowing respondents to express conditional trust – not doing so might lead to the false conclusion that trust preferences are una¤ected by con‡ict.22 22 The framework makes precise exactly how those expressing conditional trust would be reallocated to yes/no 23 Third, we can also use the geo-coded measure to examine the relationship between the intensive margin of exposure to con‡ict and trust preferences. We do so by estimating a speci…cation analogous to (3) but where we control for the number of episodes exposed to, or the number of casualties in the episodes exposed to (in each case we take an asinh transformation to take account of a mass point of zero because 39% of respondents are never in the vicinity of con‡ict). The results are in Table A4 and show a positive dose response: being exposed to more con‡ict is associated with a higher likelihood of reporting conditional trust using either measure. 5.2 Robustness Checks and Underpinning a Causal Interpretation We extend our baseline result in two directions. First, we present a battery of robustness checks on the …nding to establish it is replicated when: (i) using alternative econometric models; (ii) accounting for enumerator e¤ects; (iii) accounting for individuals being a¤ected by con‡ict in neighboring Chiefdoms (that might otherwise lead to measurement error of being in the vicinity of con‡ict). We also show the results do not re‡ect indecisiveness more generally among those exposed to con‡ict, and nor do they re‡ect those exposed to con‡ict having smaller social networks and so being more uncertain of the trustworthiness of others for anonymized exchange.23 Second, to help underpin a causal interpretation of the …nding, we …rst note that for our geo-coded measure of exposure to con‡ict, its construction is entirely independent of respondents being able and willing to recall victimization post-con‡ict. Nevertheless, we probe the data to present: (i) how the results vary in alternative subsamples related to the fact that variation in geo-coded exposure to con‡ict  arises from three sources – year of birth, Chiefdom of birth, migration during or post-con‡ict (this also relates to concerns that respondents might sort into villages based on their trust preferences [Gilligan et al. 2013]); (ii) a placebo check to strengthen the argument that actual exposure to violence matters, not other geographic factors correlated with con‡ict including pre-war confounders; (iii) an IV approach, where we predict exposure to con‡ict exploiting the geographic and timing patterns of the con‡ict described in Section 3. The bottom line from these extensions is the same: those ever in the vicinity of con‡ict are responses if only those were allowed. Those to the right of the threshold in (1) would report outright trust (rather than conditional trust), and those to the left of the threshold would report outright distrust. Given the asymmetry in the model, these reallocations would not be of equal magnitude, and depend on the () and (). In our data, among those never in vicinity of con‡ict (38% of respondents), 12% say no, 39% say yes, and 49% say it depends. While the WVS is not conducted in Sierra Leone, we can draw a comparison with the 2012 Afrobarometer survey in Sierra Leone (close to the time of our baseline), that is based on a sample of 600. This asks respondents the general trust question in a format very close to that used in the WVS: around 20% of respondents report outright distrust of others, and 78% report outright trust of others (with 2% reporting don’t know). The breakdown is similar if we focus on those aged 18-25 in 2012. Comparing these distributions of trust preferences across our baseline and Afrobarometer surveys, suggests 16% of those that report conditional trust would shift to reporting outright distrust of others, while 84% of them would shift to reporting outright trust in others. 23 The concern relates to Nunn and Wantchekon [2011] who show that a history of violence going to the slave trade in Africa impacts contemporary trust negatively – they attribute this to destruction of social ties by inter-ethnic slave raiding. 24 signi…cantly more likely to report conditional trust, and are less likely to express outright distrust or outright trust in others. This holds across samples, de…nitions, and empirical approaches. We thus press on to understand how experiences of con‡ict relate to trust preferences. 5.3 Experiences of Con‡ict To shed light on the link between experiences of con‡ict and the formation of trust preferences we re-estimate (3), with the estimates being more akin to treatment-on-the-treated e¤ects. The results are in Table 4, where we report average marginal e¤ects throughout and the omitted category are those with the background narrative. To begin with, we contrast those two groups never in the vicinity of con‡ict from age 3: those with the background narrative (no exposure ever, no recall) and those socialized (i.e. not exposed at age 3 or later and with recall). Relative to those with the background narrative, those socialized have identical trust preferences over a decade later – all estimated marginal e¤ects on each trust preference are below 008 in absolute value. This implies narratives passed onto children about the civil war might well be similar between those families for which violent events occurred in di¤erent locations to their own Chiefdom of residence, and those families for whom violent events occurred in their Chiefdom of residence but prior to their children forming direct memories of those events. This underpins the use of the geo-coded measure of exposure to con‡ict: both groups coded as never in the vicinity of con‡ict end up with identical trust preferences more than a decade later, irrespective of their exact experience of con‡ict (background narrative vs. socialized). Second, relative to those with the background narrative, those traumatized are signi…cantly more likely to report conditional trust (‘it depends’). The magnitude of the e¤ect is to marginally increase the likelihood of conditional trust by 108pp (or by 21% over those with a background narrative). This increase in conditional trust comes from both falls in outright distrust (by 49pp) and a fall in outright trust in others (by 6pp) although only the former e¤ect is statistically di¤erent from zero. A very similar pattern of marginal impacts on trust are found for those with direct experience of con‡ict: both experiences signi…cantly increase conditional trust, and both do so through signif- icant falls in outright distrust of others. Relative to those with the background narrative, we thus …nd a comparable association between trust preferences and: (i) those traumatized (i.e. exposed at age 3 or later but not recalling the war or any victimization); (ii) those with direct experience (i.e. exposed at age 3 or later and with recall). This implies that while the process of memory formation is important, an individual’s own recall of the war – the extent to which memories of actual con‡ict are suppressed or not – is not associated with the formation of conditional trust. For the remainder of the analysis we thus group together these experiences (traumatized, direct). This o¤ers mixed news for the literature using self-reported victimization to measure expo- sure to con‡ict (Table A1): on the one hand, among those with no recall that might otherwise 25 be analyzed similarly, we …nd that respondents that are traumatized have signi…cantly di¤erent trust preferences than those subject to the background narrative. Indeed, the formation of trust preferences does not di¤er between those with and without direct recall but with experiences of trauma or direct exposure to violence. Third, for those whose parents are traumatized, we also see a somewhat similar relation with trust preferences over a decade later. Relative to those with the background narrative, this set of experiences also leads individuals to signi…cantly increase the likelihood of conditional trust by 89pp. They di¤er from the traumatized and direct groups in the sense that this increase in con- ditional trust comes from a signi…cantly decline in outright trust in others (by 75pp) rather than a decline in distrust of others. We therefore consider them separately to the traumatized/direct grouping, although in terms of conditional trust, the …ndings suggest the relationship between ex- periences of civil war and trust do not depend strongly on whether an individual or their parents are traumatized. In short, the …ve di¤erent experiences essentially boil down into two types, at least for the study of the long run relationship between exposure to con‡ict early in life on long run trust preferences: (i) those with the background narrative and those socialized; (ii) those with direct experience, those traumatized, and those whose parents are traumatized, although the last group only subtly di¤ers from the …rst two in terms of whether the increase in conditional trust is driven by falls in outright trust or outright distrust. The second batch of estimates in Table 4 show the results from pooling experiences of con‡ict, where the omitted experiences combine those with the background narrative and those socialized. These replicate the earlier …ndings in terms of direct/trauma/parental trauma experiences of con- ‡ict increasing conditional trust. This restricted speci…cation also allows us to make precise that there are signi…cant di¤erences in the fall in outright distrust between those with direct/trauma experiences and those with parental trauma ( = 066).24 6 Mechanisms The framework makes precise what respondents can have in mind when they report conditional trust in others. This depends on factors: (i) that have been emphasized in the existing literature in economics and political science on trust and con‡ict such as beliefs over others (), gains from cooperation (  1 ) and risk aversion; (ii) have been more discussed in the psychology literature related to post-traumatic growth and captured through self-e¢cacy. We now take these predictions to the data. 24 In comparing the marginal e¤ects on trust preferences of the geo-coded exposure to con‡ict (Table 3) to those from experiences of con‡ict (Table 4), we do not …nd stark di¤erences – so that the intent-to-treat e¤ects captured by vicinity to con‡ict are only slightly smaller than the e¤ects of experiences of con‡ict, that account for recall and direct victimization. 26 6.1 Con‡ict and Self-E¢cacy As a preliminary step in using the framework to understand the link between con‡ict and trust, we establish whether exposure to con‡ict correlates to perceived self-e¢cacy. Table 5 shows OLS estimates of the partial correlation between residing in the vicinity of con‡ict and the overall index of self-e¢cacy, where we condition on the same set of background and geographic controls as in (3). Column 1 shows those ever in the vicinity of con‡ict have higher self-e¢cacy than those never in the vicinity. The magnitude of the di¤erence is 089 and is statistically signi…cant ( = 053).25 Figure A2 summarizes the results if we repeat the exercise for each sub-component of the self- e¢cacy index. Components of both the business and personal traits measures signi…cantly increase self-e¢cacy. For the business measure, vicinity to con‡ict is statistically signi…cantly correlated to the ability to run a business, bargain on purchases, obtain credit, manage accounts and ensure employees work. For the personal traits measures, the strongest and most precise correlations are found with respect to a drive to do better than others, making a good leader, making plans for the future, seeing tasks to completion, and saving regularly. In Column 2 of Table 5 we repeat the analysis but examine how experiences of con‡ict partially correlate to self-e¢cacy. Following the earlier results, we group those traumatized and with direct experience, and the omitted category groups together those with the background narrative and those socialized. Experiences matter: those with trauma/direct experience have signi…cantly higher self-e¢cacy than those with the background narrative/socialized. The magnitude of the e¤ect is 121 and is statistically di¤erent to zero at the 5% level. What these experiences have in common is that respondents are subject to violence after age 3 when memories are being formed, and these shape later life beliefs and attitudes [Fehr et al. 2008, Malmendier and Nagel 2011, Bauer et al. 2014, Giuliano and Spilimbergo 2014, Malmendier 2021]. Column 2 also shows that those with parental trauma also have signi…cantly higher self-e¢cacy than the omitted group – the magnitude of the e¤ect is 143 and is precisely measured, being statistically di¤erent to zero at the 1% level. The partial correlations with self-e¢cacy are not di¤erent between the traumatized/direct group and the parental trauma group ( = 713). Our results build on a long-standing evidence base in psychology, showing the development of self-e¢cacy is an important coping strategy in the face of extreme trauma [Bandura 1997, Benight and Bandura 2004]. Our …ndings suggest this process starts early in life – as soon as memories start forming. The fact that perceived self-e¢cacy is not higher among those with traumatized parents is also informative. Recall these respondents are similar to those with a background narrative except they were exposed to violent events when aged 0 to 2 – yet narratives of these events have not been passed onto them by their parents or others. We refer to them as with parental trauma, but an alternative interpretation is that parents might be deliberately trying to protect 25 To reiterate, the …nding that self-e¢cacy increases for those exposed to con‡ict is relative to those not exposed but still in Sierra Leone. We do not have a counterfactual to compare self-e¢cacy to those never exposed to con‡ict. Hence it could still be the case that self-e¢cacy is lower than if con‡ict had never occurred. 27 their children in suppressing accounts of violent events in their early years. If so, we …nd they are not able to engender signi…cantly higher levels of perceived self-e¢cacy among their children, relative to those with experiences of direct exposure or trauma.26 6.2 Con‡ict, Self-E¢cacy and Trust We can now re-examine the link between con‡ict and trust by estimating (3) and additionally controlling for self-e¢cacy. The results are in Table 6 where we report average marginal e¤ects throughout. Columns 1a-1c show that all else equal, self-e¢cacy correlates to trust preferences. A one standard deviation increase in self-e¢cacy is associated with a signi…cantly higher likelihood of conditionally trust: the magnitude of the e¤ect is 56pp, corresponding to a 11% increase over the level of conditional trust among those never exposed to con‡ict ( = 0). The e¤ect is statistically signi…cant at the 1% level. This increase in conditional trust related to self-e¢cacy arises because of a near equal and opposite reduction in those reporting outright trust of others (the marginal e¤ect of trusting others reduces by 55pp). Higher self-e¢cacy is asymmetrically associated with trust preferences: we can reject equality of shifts from yes/no towards conditional trust ( = 000). This is in line with the asymmetric predictions of the framework, as described in Figure 6C. We also continue to …nd a signi…cant correlation between vicinity to con‡ict and conditional trust. This suggests there remain residual channels through which exposure to con‡ict still directly a¤ects trust preferences through. These other channels include that con‡ict increases the return to investing in social capital because of a loss of formal institutions during con‡ict, or as a way to build personal safety and security (  1 ). We cannot empirically capture these directly. Column 2a-2c repeat the analysis but for experiences of con‡ict. We again see that, all else equal: (i) self-e¢cacy relates to trust preferences: those with great self-e¢cacy are signi…cantly more likely to report conditional trust, and this increase comes from a near equal and oppo- site likelihood of reporting outright trust in others; (ii) experiences of con‡ict correlate to trust preferences in a very similar way as documented earlier in Table 4. 6.3 Con‡ict and Beliefs Over Others The framework captures alternative hypotheses for why con‡ict can impact trust preferences: a prominent class of explanation for why con‡ict a¤ects generalized trust in one-o¤ interactions is because it changes beliefs over the trustworthiness of others. One suggested mechanism is via parochial norms, whereby con‡ict strengthens social preferences towards ingroups [Heinrich and Boyd 2001, Choi and Bowles 2007]. In the framework this can be interpreted as  being a 26 Our …ndings complement recent work showing it is possible to design interventions to increase self-e¢cacy [McKelway 2021], and to design interventions to improve imagery-based decision making, especially among trauma victims [Ashraf .et al 2022]. 28 population-share weighted average of high beliefs of cooperation if matched to an ingroup member, and low beliefs of cooperation if matched to an outgroup member. Distinct from the standard question on generalized trust in others, to proxy for beliefs over speci…c groups we asked respondents at baseline, “How many people in [group] can you trust?” with potential responses being none, some, most, and all. We do so for groups considered in- groups of respondents – such as residents of their village, other women, those attending their same church/mosque, and those of the same religion. We also do so for various outgroups – men, those of a di¤erent religion, and whites. We also asked the same question about groups related to state and market institutions – police o¢cers, politicians, those in national government, those in local government, banks and money lenders. We use this data to estimate a speci…cation analogous to (3) where we examine the relationship between being in the vicinity of con‡ict ( ) and responses to the question of how many in each group can be trusted. Figure 6 summarizes the results: it shows the average marginal e¤ect of being in the vicinity of con‡ict on each response (none, some/most, and all). We see a remarkably consistent pattern of …ndings: for each and every group, respondents report being signi…cantly less likely to trust no one in each group, they are also signi…cantly less likely to report trusting all in nearly each group. They are signi…cantly more likely to report trusting some/most in each group. This represents a shift in the distribution of beliefs over others, so  becomes concentrated towards central values and away from extremes. The fact that this applies equally to ingroups and outgroups is especially notable. This result runs counter to the notion that exposure to con‡ict leads to parochial altruism or greater trust in the ingroup at the expense of less trust towards outgroups. This interpretation is in line with narratives of the civil war in Sierra Leone being a con‡ict not rooted along ethnic, religious or political party divides. Finally, the fact that self-e¢cacy a¤ects trust preferences helps rule out an alternative ex- planation for why individuals reply ‘it depends’ when asked our trust question: that they do so whenever they are close to indi¤erence to trusting or not. If that were true, then the region of conditional trust spans the threshold for  described in (1) and changes in self-e¢cacy should have no impact on conditional trust – something emphatically ruled out by the evidence. 7 Discussion 7.1 Negative Psychological Legacies of Con‡ict We have demonstrated the mediating role of self-e¢cacy in linking exposure to con‡ict and trust preferences. This builds on case studies in psychology highlighting positive changes to individuals that arise through post-traumatic growth. At the same time, it is important to understand whether our data can be reconciled with an equally important literature documenting negative 29 impacts of con‡ict on psychological traits, such as depression and anxiety [Ehlers and Clark 2000, Yehuda 2002, Vinck et al. 2007]. This is especially relevant given our sample, if women are more susceptible to negative impacts of trauma [Baranov et al. 2020]. We shed light on this in two ways: (i) by examining the relation between exposure to con‡ict, anxiety and life satisfaction; (ii) probing how speci…c types of victimization correlate to trust. 7.1.1 Anxiety and Satisfaction On anxiety, we asked respondents about whether they worry about various dimensions of life: …nding a job, …nding a husband, money and violence/theft. The results in Table 7 show that: (i) relative to those never in the vicinity of con‡ict, those in the vicinity of con‡ict are signi…cantly more likely to report worrying about violence/theft, but not other dimensions (Columns 1-4); (ii) in terms of experiences, relative to those with the background narrative/socialized, those traumatized or who experienced violence directly, also are signi…cantly more likely to report worrying about violence/theft, but not these other dimensions (Columns 5-8). Those with the experience of parental trauma report no more anxiety than those with the background narrative. This provides added credence to our gradations of experiences of con‡ict. Given the evidence suggests those exposed to con‡ict have both higher anxiety and self-e¢cacy in the long run, one way to establish the net e¤ect on welfare is to examine correlations between these measures and life satisfaction. To be clear, these comparisons are relative to those never in the vicinity of con‡ict or those with the background narrative. The results cannot be interpreted as relative to a counterfactual absent civil war altogether. We are able to construct indices across three dimensions: social, economic and life in general. These results are in the remaining Columns of Table 7. We see overall higher satisfaction along all three dimensions for those ever in the vicinity of con‡ict (Columns 9-11), or with various experiences of con‡ict relative to those with the background narrative (Columns 12-14). This is consistent with any positive e¤ects on self-e¢cacy induced by exposure to con‡ict more than compensating for negative impacts through anxiety about future violence.27 7.1.2 Victimization, Self-E¢cacy and Trust Our second approach probes further the link between speci…c types of victimization and generalized trust. We have so far only used the victimization data in terms of memory recall to construct experiences of con‡ict. We start by examining how speci…c forms of victimization correlate to self-e¢cacy. Column 1 of Table 8 shows the results: this highlights that some, but not all, forms of victimization increase self-e¢cacy. Those that report their family being harmed or being a 27 The satisfaction indices are inverse covariance weighted indices constructed from questions asking respondents to rate their satisfaction on a Likert scale. The dimensions over which this questions are asked are: (i) Social: friends, family, community, house; (ii) Economic: own education, job, earnings, household …nances, access to credit; (ii) Life: life as a whole, future prospects. The indices are each standardized relative to the omitted group. 30 refugee overseas have signi…cantly higher self-e¢cacy. Notably however, reporting being personally harmed signi…cantly reduces self-e¢cacy. Hence types of victimization shift self-e¢cacy in di¤erent directions. This result was masked in the earlier analysis because the most common forms of victimization (…ghting in the area, internally displaced or family harmed) all positively correlate to self-e¢cacy, while only 6% of respondents report being personally harmed. The results point to the need for targeted post-con‡ict policy interventions towards those personally harmed, the returns to which might be very high given the costs of worse mental health over the life cycle [Currie 2009, Adhvaryu et al. 2019, Ridley et al. 2020]. The remaining Columns of Table 8 then link forms of victimization to trust preferences. We see that the two most common forms of victimization – …ghting in the area or being displaced – are signi…cantly related to conditional trust, and pull in trust preferences from the tails. However, for other forms of victimization such as family being harmed or being personally harmed, we see no association with trust preferences. To the extent con‡icts vary in terms of the speci…c types of victimization civilians su¤er from, the …nding suggest that the impact of con‡ict on trust preferences will be tempered. Most notably in con‡icts where many are personally harmed, their families harmed, or they are refugees, the impact on conditional trust might be far weaker. This helps reconcile the range of estimates found in the literature, that lead to the meta-analysis of Bauer et al. [2016] …nding no overall e¤ect of con‡ict on trust (counter to other forms of pro-social behavior).28 7.2 Survivor Bias As in all studies linking con‡ict and trust, we can only base our analysis on the select sample of those that survive the con‡ict. As Table A1 highlights, our sample is uniquely based on a group born into con‡ict and that survive 14 years post-con‡ict. While this helps establish the long run e¤ects of con‡ict and the formation of trust preferences, this leads to a particular form of survivor bias, relevant for external validity. Earlier studies have demonstrated that children who live through con‡ict are less healthy, less educated, and have worse labor market outcomes as adults [Bundervoet et al. 2009, Blattman and Annan 2010]. While all these focus on socio- economic outcomes, our …ndings additionally highlight that those born into and exposed to con‡ict can develop psychological resistance in the form of higher self-e¢cacy.29 28 Relative to other con‡icts, the civil war in Sierra Leone might be characterized as one in which although casualty rates were relatively low (estimated to be 15% of the population), there were high rates of displacement of the civilian population (estimated to be 58% of the population), and as emphasized throughout, the con‡ict did not operate along parochial lines. In contrast, during the ethnically-based Rwandan civil war of 1990-4, casualty rates were 9%, while 20% of the population were displaced. Our results suggest to the extent that forms of victimization di¤er across con‡icts, these can translate into self-e¢cacy being increased or decreased, and consequent di¤erential impacts on the long run formation of trust preferences. 29 The other dimension of selection relevant for external validity is that our sample is based on girls born into con‡ict. As such, our focus is largely on victims of con‡ict, and our results do not speak to the formation of trust preferences among perpetrators of violence. However, Bauer et al. [2016] report similar …ndings on exposure to 31 7.3 Future Agenda Our analysis points to three important directions for future research on psychological legacies of exposure to con‡ict and violence early in life. First, the framework developed emphasizes the mediating role that self-e¢cacy has, and our evidence hints to the processes generating self-e¢cacy. The results on experiences of con‡ict and trust inform some of the modelling choices. Given the similarity of …ndings between those traumatized by con‡ict and those with direct experience of con‡ict, the framework does not focus on the role that being able and willing to recall victimization plays in explaining trust preferences. Similarly, the …ndings suggested a similarity between those with the background narrative and those socialized. Hence the framework does not focus on the endogenous choice of narratives passed down by parents to children born into con‡ict [Doepke and Zilibotti 2017, Bénabou et al. 2018, Akerlof and Rayo 2020]. We leave such extensions for future work, as such channels might well be important in other con‡ict scenarios. Second, there can be valuable gains from more closely integrating work on post-con‡ict psy- chological legacies for those born into con‡ict with the established literature studying the process of human capital accumulation in early childhood, a recent overview of which is provided by At- tanasio et al. [2022]. These strands of research have developed largely in isolation from each other. Work in early childhood development has focused on the accumulation of cognitive skills, non-cognitive skills and health, in response to inputs such as resources, information, nutrition and parenting. Less attention has been paid to the development of traits such as self-e¢cacy, or the importance of aggregate and traumatic shocks in shaping parent-child interactions and the production function for human capital early in life.30 Third, when constructing experiences of con‡ict, age of exposure is critical. We considered if an individual was aged 0-2 when in the vicinity of con‡ict (and so prior to signi…cant memory formation), and/or whether the individual was aged 3 and above when in the vicinity of con‡ict, and hence more capable of retaining …rst-hand memories of con‡ict. While this age split is useful, there might be other critical periods of childhood development when exposure to con‡ict leaves psychological legacies [Kim and Lee 2014]. The brain develops in cyclical spurts from birth through to the end of teenage years [Lampl et al. 1992, Heckman 2007]. We are unable to study other sensitive periods of development because of the long lasting nature of the con‡ict in Sierra Leone, so most respondents are exposed to con‡ict early in life. Understanding how con‡ict shapes which traits are formed across phases of development, and their dynamic interlinkages are all rich areas for future study. con‡ict and pro-social behavior for victims and perpetrators of violence, and for both genders. 30 Other example of collective traumas and aggregate shocks impacting traits include slavery, natural disasters and economic downturns [Malmendier and Nagel 2011, Nunn and Wantchekon 2011, Callen 2015]. 32 8 Conclusion Exposure to armed con‡ict in early life is an extremely traumatic experience, and yet it is common, a¤ecting over 400 million children worldwide, and with the number of countries experiencing armed con‡ict in 2021 being among the highest in 30 years [Ostby et al. 2020]. We build on the literature documenting persistent psychological legacies of con‡ict. We shed new light on the role that post traumatic growth and building self-e¢cacy play in the long run formation of trust preferences, that is the foundation for social and economic interactions and acts as ‘an important lubricant of the social system’ [Arrow 1974]. Taken together our analysis provides a framework in which to understand and reconcile disparate …ndings in the multi-disciplinary literature linking con‡ict and trust. The macro evidence suggests on average, post-war periods are often characterized by rapid economic recovery [Cerra and Saxena 2008, Miguel and Roland 2011]. Whether psychological legacies of violence help or hinder post-con‡ict recovery has been subject to debate [Knack and Keefer 1997, Guiso et al. 2009]. Despite the collective trauma, Sierra Leone has seen a remarkable period of post-con‡ict recovery and stability– at least until the Ebola outbreak of 2014. National elections were held in 2002, closely followed by the …rst local elections in decades in 2004, and the post-con‡ict decade was one of largely sustained economic growth. Linking psychological legacies of con‡ict to understanding whether such changes in self-e¢cacy and trust also help foster public goods provision or allow new forms of social organization to emerge [e.g. Grief 1993, Bowles 2008, Blattman and Miguel 2010] would be a natural next step in understanding more general patterns of post-war economic recovery in an increasingly violent world. A Appendix A.1 Robustness We present a series of robustness checks on the main …nding from Table 3 linking exposure to con‡ict to trust. Table A5 shows our main result is robust to: (i) alternative clustering of standard errors than by village of residence, such as clustering by district of birth, by age, or allowing for robust standard errors (Columns 1a-1c); (ii) alternative empirical models, such as probit or linear probability speci…cations (where we group the yes and no trust answers together). In both alternative models we …nd comparable increases in the likelihood to report conditional trust among those ever in the vicinity of con‡ict (Columns 2, 3). Given the sensitive nature of survey questions related to con‡ict and trust, we address the concern that results are driven by enumerator style by conditioning on 53 enumerator …xed e¤ects in the linear probability model speci…cation (from across the 200 villages in our sample). Column 4 shows that the coe¢cient of interest reduces in magnitude (from 77 to 44) but remains precisely 33 estimated and statistically di¤erent from zero at the 5% signi…cance level. Given enumerators typically survey across a handful of villages, this speci…cation is the closest we can get (without losing precision altogether) to documenting within village variation in conditional trust being related to individuals ever being in the vicinity of con‡ict. We next account for individuals being a¤ected by con‡ict in neighboring Chiefdoms (that might otherwise lead to measurement error of being in the vicinity of con‡ict). To do so, we construct a measure of con‡ict exposure that takes into account all episodes of violence that took place within a …xed radius from each Chiefdom’s centroid, and weight episodes by the inverse of their distance. We thus rede…ne exposure to con‡ict as: Ã ! XX ¡1  =  (1 +  ) (5)  2 where  includes all years between birth of respondent  and 2001,  is the subset of episodes of violence that took place in year , and  is the distance in kilometers between the centroid of the chiefdom of residence of respondent  in year  and the location of con‡ict episode . Table A6 shows results from our benchmark speci…cation using this measure of being in the vicinity of con‡ict, for various distance cuto¤s, where we note the average Chiefdom has an approximate radius of 12km. We see that using larger areas rather than the Chiefdom boundaries gradually weakens the results. This is as expected and reinforces the idea that it is exposure to localized violence that matters for the long run formation of trust preferences. We next examine whether the results re‡ect indecisiveness more generally among those exposed to con‡ict. To do so we consider responses to an altogether di¤erent survey module, on various dimensions of life satisfaction. These questions asked, “How do you feel about [category]?” where the categories were own education, family, friends, job, income, own dwelling, own school, own community, future prospects, family’s …nancial situation, access to credit, and life as a whole. Responses were coded on a 7-point Likert scale, ranging from very happy/optimistic to very sad/pessimistic. The middle of the scale is labelled as neutral. We combine responses across these dimensions and code whether: (i) the respondent gives a neutral reply to any of these dimensions of satisfaction; (ii) the share of dimensions that they give a neutral response to. Table A7 shows OLS estimates of these outcomes regressed against the geo-coded exposure to con‡ict measure, controlling for the usual set of covariates included in (3), and allowing standard errors to be clustered by village of residence. We …nd those exposed to con‡ict are less likely to report being neutral on any dimension of satisfaction (Column 1) and their share of neutral answers across all dimensions is uncorrelated to exposure to con‡ict (Column 2). Finally, to check whether our main …nding might re‡ect those exposed to con‡ict having smaller social networks and so are more uncertain of the trustworthiness of others for anonymized exchange, we use social networks data collected in our midline survey from a random subset of 34 respondents. We report Tobit estimates of the relationship between con‡ict exposure and the degree of various types of social networks of respondents. These networks include friendship ties, others they speak to about intimate topics, work issues and opportunities, or issues related to …nances and credit. As reported in Columns 3 to 6 of Table A7, on each type of social network, we see no relationship between the size of networks reported and having been ever in the vicinity of con‡ict. A.2 Causality We use a variety of approaches to underpin a case for the link between exposure to con‡ict early in life and trust preferences expressed a decade later as being causal. First, we note that variation in the geo-coded exposure to con‡ict  arises from three sources: (i) year of birth; (ii) Chiefdom of birth; (iii) migration during or post-con‡ict. In Table A8 we explore how the relationship between con‡ict and trust varies along these dimensions. We do so using the linear probability speci…cation, in which heterogeneous e¤ects of con‡ict on trust are most transparently estimated. In Column 1 we drop those born in the …nal year of con‡ict. This cohort are born to parents who would be more certain of the con‡ict’s end than those born earlier, and they might also be born to parents who strategically delayed fertility during the con‡ict. Either form of selection could correlate with trust preferences. We …nd the relationship between exposure to con‡ict and conditional trust remain unchanged when dropping this cohort from the sample: despite the 15% reduction in sample size, the coe¢cient of interest remains identical (077) and its standard error increases only marginally. Columns 2 and 3 explore how the relationship between con‡ict and trust varies with location of birth. We …nd those born into districts neighboring Liberia – that experienced the earliest and severest con‡ict (Figure 2) – have no di¤erential relationship between exposure to con‡ict and conditional trust. Column 3 shows this to also be the case examining heterogeneous relationships across regions. If migration is endogenous to con‡ict, the results capture a total e¤ect accounting for migra- tory insurance responses of households. While policy relevant, the concern is that this picks up factors driving both the decision to migrate during con‡ict and the formation of trust preferences. To assess endogenous migration choices driving exposure to con‡ict, we examine if there is a het- erogeneous relationship between vicinity to con‡ict and conditional trust between those that did and did not migrate across Chiefdoms during the civil war (Column 4). We …nd no evidence for this. The same holds if we also allow the relationship between con‡ict and trust to vary with whether the respondent migrated post-con‡ict. Taken together, the evidence suggests if migrating parents di¤er in attributes, those di¤erences appear unrelated to the formation of conditional trust preferences of their children. This also helps address concerns that respondents might sort into villages based on their trust preferences [Gilligan et al. 2013]. 35 We next provide a placebo check on our main result: we do so by adding …ve years to the actual year of birth of each respondent (where we restrict the sample to those that would still then be born sometime during the con‡ict – this avoids a mass of respondents being shifted to never being exposed). Columns 1a-1c in Table A9 show the result of doing so when estimating (3): we see that under his placebo we …nd no evidence that measured exposure to con‡ict relates to trust preferences. This placebo strengthens the argument that actual exposure to violence matters, not other geographic factors correlated with con‡ict. Given the few respondents that migrate during the civil war, this check is reassuring that trust preferences do not just pick up factors related to place of birth. Finally, we address causality through an IV approach where we predict exposure to con‡ict. We exploit the geographic patterns of the con‡ict described in Section 2. More precisely, we consider as instruments an interaction between: (i) the distance from the respondents’ place of birth and the …rst recorded episode of violence in the con‡ict in 1991 (  ); (ii) the year of birth of the respondent.31 The …rst stage of the IV is then given by: 2000 X  =     I[ ¸   ] +    +   (6) =1991 where  is the measure of vicinity to con‡ict,  includes our standard set of controls and …xed e¤ects (including district of birth and district of residence),  = ( 1991      2000 )0 and   are column vectors of parameters. The exclusion restriction is that conditional on these controls, distance to the …rst violent event of the con‡ict predicts con‡ict but has no direct e¤ect on the long run formation of trust preferences. As a second set of instruments we also include interactions of respondent’s year of birth with their distance from Freetown (  ) so the …rst stage is: 2000 X 2000 X  =     I[ ¸   ] +     I[ ¸   ] +    +   (7) =1991 =1991 Figure A3 graphs the …rst stage estimates of the instruments on exposure to con‡ict. In the left hand panel we see the …rst stage from (6). Residing further from the …rst violent event decreases the likelihood of ever being in the vicinity of con‡ict for those born up until 1997, and the e¤ect of the instruments reverses later in time as the con‡ict advanced to regions away from the border with Liberia. The right hand panel shows the predictability of the both instrument sets from the …rst stage described in (7). Again we see the instruments vary in their predictability with year of birth of ever being in the vicinity of con‡ict. Columns 2 and 3 in Table A9 report the second stage 2SLS estimates, where we use the LPM speci…cation for the second stage. For both sets of instruments, we …nd a strong …rst stage (F- 31 Alternative approaches to predicting con‡ict have included using machine learning methods using multiple data sources or newspaper archives [Bazzi et al. 2022, Mueller and Rauh 2022]. 36 statistics above 30) and the second stage remains precisely estimated. The magnitude of the IV e¤ects of con‡ict on conditional trust are stronger than in the baseline LPM speci…cation, being around double in size. References [1] adhvaryu.a, j.fenske and a.nyshadham (2019) “Early Life Circumstance and Adult Mental Health,” Journal of Political Economy 127: 1516-49. [2] akerlof.g.a and s.j.dennis (2016) “Bread and Bullets,” Journal of Economic Behavior & Organization 126: 58-71. [3] akerlof.r and l.rayo (2020) Narratives and the Economics of the Family, mimeo Warwick. [4] anderson.m.l (2008) “Multiple Inference and Gender Di¤erences in the E¤ects of Early Intervention: A Reevaluation of the Abecedarian, Perry Preschool, and Early Training Projects,” Journal of the American Statistical Association 103: 1481-95. [5] arrow.k.j (1974) The Limits of Organization. WW Norton & Company. [6] ashraf.n, g.bryan, a.delfino, e.holmes, l.iacovone and a.pople (2022) Learning to see the World’s Ppportunities: The Impact of Imagery on Entrepreneurial Success, mimeo LSE. [7] attanasio.o, c.meghir and s.cattan (2022) “Early Childhood Development, Human Capital and Poverty ” Annual Review of Economics 14: 853-92. [8] bandiera.o, n.nbuehren, m.goldstein, i.rasul and a.smurra (2020) The Economic Lives of Young Women in the Time of Ebola, mimeo LSE. [9] bandura.a (1977) “Self-E¢cacy: Toward a Unifying Theory of Behavioral Change,” Psy- chological Review 84: 191-215. [10] bandura.a (1997) Self-E¢cacy: The Exercise of Control, New York: Freeman. [11] baranov.v, s.bhalotra, p.biroli and j.maselko (2020) “Maternal Depression, Women’s Empowerment, and Parental Investment: Evidence from a Randomized Controlled Trial ” American Economic Review 110: 824-59. [12] bauer.m, a.cassar, j.chytilova and j.henrich (2014) “War’s Enduring E¤ects on the Development of Egalitarian Motivations and In-group Biases,” Psychological Science 25: 47- 57. 37 [13] bauer.m, c.blattman, c.chytilova, j.henrich, e.miguel and t.mitts (2016) “Can War Foster Cooperation?” Journal of Economic Perspectives 30: 249-74. [14] bazzi.s, r.a.blair, c.blattman, o.dube, m.gudgeon and r.peck (2022) “The Promise and Pitfalls of Con‡ict Prediction: Evidence from Colombia and Indonesia,” The Review of Economics and Statistics 104: 764-79. [15] bellows.j and e.miguel (2009) “War and Local Collective Action in Sierra Leone,” Journal of Public Economics 93: 1144-57. [16] benabou.r, a.falk and j.tirole (2018) Narratives, Imperatives, and Moral Reasoning, NBER WP 24798. [17] benight.c.c and a.bandura (2004) “Social Cognitive Theory of Posttraumatic Recovery: The Role of Perceived Self-E¢cacy,” Behavior Research and Therapy 42: 1129-48. [18] besley.t and m.reynal-querol (2014) “The Legacy of Historical Con‡ict: Evidence from Africa,” American Political Science Review 108: 319-36. [19] betancourt.t.s and k.t.khan (2008) “The Mental Health of Children A¤ected by Armed Con‡ict: Protective Processes and Pathways to Resilience,” International Review of Psychi- atry 20: 317-28 [20] blattman.c and j.annan (2010) “The Consequences of Child Soldiering,” Review of Eco- nomics and Statistics 92: 882-98. [21] blattman.c and e.miguel (2010) “Civil War,” Journal of Economic Literature 48: 3-57. [22] bordalo.p, n.gennaioli and a.shleifer (2020) “Memory, Attention and Choice,” Quar- terly Journal of Economics 135: 1399-442. [23] bowles.s (2008) “Con‡ict: Altruism’s Midwife,” Nature 456: 326-7. [24] bundervoet.t and p.r.verwimp (2009) “Health and Civil War in Rural Burundi ,” Journal of Human Resources 44: 536-63. [25] callen.m, m.isaqzadeh, j.d.long and c.sprenger (2014) “Violent Trauma and Risk Preference: Experimental Evidence from Afghanistan,” American Economic Rev. 104: 123- 48. [26] callen.m (2015) “Catastrophes and Time Preference: Evidence from the Indian Ocean Earthquake,” Journal of Economic Behavior and Organization 118: 199-214. [27] cerra.v and s.c.saxena (2008) “Growth Dynamics: The Myth of Economic Recovery,” American Economic Review 98: 439-57. 38 [28] charness.g and m.dufwenberg (2006) “Promises and Partnership,” Econometrica 74: 1579-1601. [29] choi.j-kyoo and a.bowles (2007) “The Coevolution of Parochial Altruism and War,” Science 318: 636-40. [30] chowdhury.s, m.sutter and k.f.zimmermann (2022) “Economic Preferences Across Generations and Family Clusters: A Large-scale Experiment in a Developing Country,” Jour- nal of Political Economy, forthcoming. [31] collier.p, v.l.elliott, h.hegre, a.hoeffler, m.reynal-querol and n.sambanis (2003) Breaking the Con‡ict Trap: Civil War and Development Policy, World Bank and Oxford University Press. [32] conibere.r et al. (2004) “Statistical Appendix to the Report of Truth and Reconciliation Commission: Report of Sierra Leone,” Human Rights Data Analysis Group, The Benetech Initiative. [33] currie.j (2009) “Healthy, Wealthy and Wise: Is there a Causal Relationship Between Child Health and Human Capital Development?” Journal of Economic Literature 47: 87-122. [34] doepke.m and f.zilibotti (2017) “Parenting with Style: Altruism and Paternalism in Intergenerational Preference Transmission,” Econometrica 85: 1331-71. [35] dohmen.t, a.falk, d.huffman and u.sunde (2012) “The Intergenerational Transmission of Risk and Trust Attitudes,” Review of Economic Studies 79: 645-77. [36] eck.k (2012) “In Data We Trust? A Comparison of UCDP GED and ACLED Con‡ict Events Datasets,” Cooperation and Con‡ict 47: 124-41. [37] ederer.f and f.schneider (2022) “Trust and Promises over Time,” American Economic Journal: Microeconomics 14: 304-20. [38] ehlers.anke, and d.m.clark (2000) “A Cognitive Model of Post-traumatic Stress Disor- der,” Behaviour Research and Therapy 38: 319-45. [39] falk.a, f.kosse, p.pinger, h.schildberg-horisch and t.deckers (2021) “Socio- Economic Status and Inequalities in Children’s IQ and Economic Preferences,” Journal of Political Economy 129: 2504-45. [40] fehr.e, h.bernhard and b.rockenbach (2008) “Egalitarianism in Young Children,” Na- ture 454: 1079-83. 39 [41] fudenberg.d, g.lanzani and p.strack (2022) Selective Memory Equilibrium, mimeo Yale. [42] gennaioli.n and a.shleifer (2010) “What Comes to Mind,” The Quarterly Journal of Economics 125: 1399-433. [43] gilligan.m.j, b.j.pasquale and c.samii (2014) “Civil War and Social Cohesion: Lab-in- the-Field Evidence from Nepal,” American Journal of Political Science 58: 604-19. [44] giuliano.p and a.spilimbergo (2014) “Growing up in a Recession,” Review of Economic Studies 81: 787-814. [45] glaerser.e, d.laibson, j.a.scheinkman and c.l.soutter (2000) “Measuring Trust,” Quarterly Journal of Economics 115: 811-46. [46] grief.a (1993) “Contract Enforcement and Economic Instituions in Early Trade: The Maghribi Traders’ Coalition,” American Economic Review 83: 525-48. [47] guiso.l, p.sapienza and l.zingales (2009) “Cultural Biases in Economic Exchange?,” Quarterly Journal of Economics 124: 1095-131. [48] hartman.a.c and b.s.morse (2020) “Violence, Empathy and Altruism: Evidence from the Ivorian Refugee Crisis in Liberia,” British Journal of Political Science 50: 731-55. [49] heckman.j.j (2007) “The Economics, Technology and Neuroscience of Human Capability Formation,” Proceedings of the National Academy of Science 104: 13250-5. [50] henrich.j and r.boyd (2001) “Why People Punish Defectors: Weak Conformist Trans- mission Can Stabilize Costly Enforcement of Norms in Cooperative Dilemmas,” Journal of Theoretical Biology 208: 79-89. [51] humphreys.m and j.m.weinstein (2006) “Handling and Manhandling Civilians in Civil War,” American Political Science Review 100: 429-47. [52] jakiela.p and o.ozier (2019) “The Impact of Violence on Individual Risk Preferences: Evidence from a Natural Experiment,” Review of Economics and Statistics 101: 547-59. [53] keen.d (2005) Con‡ict and Collusion in Sierra Leone, James Currey. [54] kim.y and j.lee (2014) “The Long-run Impact of a Traumatic Experience on Risk Aversion,” Journal of Economic Behavior and Organization 108: 174-86. [55] knack.s and p.keefer (1997) “Does Social Capital Have an Economic Payo¤? A Cross- Country Investigation,” Quarterly Journal of Economics 112: 1251-88. 40 [56] kosse.f, t.deckers, h.schildberg-horisch and a.falk (2020) “The Formation of Prosociality: Causal Evidence on the Role of the Social Environment,” Journal of Politi- cal Economy 128: 434-67. [57] lampl.m.j, d.veldhuis and m.l.johnson (1992) “Saltation and Stasis: A Model of Human Growth,” Science 258(5083): 801-3. [58] malmendier.u (2021) “Exposure, Experience, and Expertise: Why Personal Histories Mat- ter in Economics,” Journal of the European Economic Association 19: 2857-94. [59] malmendier.u and s.nagel (2011) “Depression Babies: Do Macroeconomic Experiences A¤ect Risk Taking?” Quarterly Journal of Economics 126: 373-416. [60] mckelway.m (2021) Women’s Employment in India: Intra-Household and Intra-Personal Constraints, mimeo Dartmouth. [61] miguel.e and g.roland (2011) “The Long-run Impact of Bombing Vietnam” Journal of Development Economics 96: 1-15. [62] moya.a (2018) “Violence, Psychological Trauma, and Risk Attitudes: Evidence from Victims of Violence in Colombia,” Journal of Development Economics 131: 15-27. [63] mueller.h and c.rauh (2002) “The Hard Problem of Prediction for Con‡ict Prevention,” Journal of the European Economic Association forthcoming. [64] mullainathan.s (2002) “A Memory-Based Model of Bounded Rationality ” Quarterly Jour- nal of Economics 117: 735-74. [65] nunn.n and l.wantchekon (2011) “The Slave Trade and the Origins of Mistrust in Africa,” American Economic Review 101: 3221-52. [66] ostby.g, s.a.rustad and a.f.tollefsen (2020) Children A¤ected by Armed Con‡ict, 1990-2019, Peace Research Institute Oslo: Oslo. [67] pennebaker.j.w and k.d.harber (1993) “A Social Stage Model of Collective Coping: The Loma Prieta Earthquake and the Persian Gulf War,” Journal of Social Issues 49: 125-45. [68] reno.w (1995) Corruption and State Politics in Sierra Leone, CUP: Cambridge. [69] richards.p (1996) Fighting for the Rainforest: War, Youth and Resources in Sierra Leone, London: James Currey. [70] ridley.m, g.rao, f.schilbach and v.patel (2020) “Poverty, Depression and Anxiety: Causal Evidence and Mechanisms,” Science 370(6522). 41 [71] sapienza.p, a.toldra-simats and l.zingales (2013) “Understanding Trust,” Economic Journal 123: 1313-32. [72] schwarzer.r and m.jerusalem (1995) “Generalized Self-E¢cacy Scale,” in Measures in Health Psychology: A User’s Portfolio. Causal and Control Beliefs, J.Weinman, S.C.Wright and M.Johnston (eds.) Windsor, England: NFER-Nelson. [73] serra-garcia.m, e.van damme and j.potters (2013) “Lying About What You Know or About What You Do?” Journal of the European Economic Association 11: 1204-29. [74] shiller.j.r (2017) “Narrative Economics,” American Economic Review 107: 967-1004. [75] tedeschi.r.g, c.l.park and l.g.calhoun (1998) Posttraumatic Growth: Positive Changes in the Aftermath of Crisis, Routledge. [76] tedeschi.r.g and l.g.calhoun (2004) “Posttraumatic Growth: Conceptual Foundations and Empirical Evidence,” Psychological Inquiry 15: 1-18. [77] ucdp (2018) Uppsala Con‡ict Data Program, date of retrieval: 2018/07/10, UCDP Con‡ict Encyclopedia: www.ucdp.uu.se, Uppsala University. [78] vanberg.c (2008) “Why Do People Keep Their Promises? An Experimental Test of Two Explanations,” Econometrica 76: 1467-80. [79] vinck.p, p.n.pham, e.stover and h.m.weinstein (2007) “Exposure to War Crimes and Implications for Peace Building in Northern Uganda,” Journal of the American Medical As- sociation 298: 543-54. [80] voors.m.j, e.m.nillesen, p.verwimp, e.h.bulte, r.lensink, and d.p.van soest (2012) “Violent Con‡ict and Behavior: A Field Experiment in Burundi,” American Economic Review 102: 941-64. [81] walter.f.b and j.snyder (1999) Civil Wars, Insecurity, and Intervention, Columbia Uni- versity Press. [82] yehuda.r (2002) “Post-traumatic Stress Disorder,” New England Journal of Medicine 346: 108-14. 42 Table 1: Geo-Coded Exposure to Conflict Means (standard deviation in parentheses) (2) Migrated (3) Stable Residence (4) Ruling (5) Non-Ruling (1) Full Sample During Civil War During Civil War Family Family [1991,2001] [1991,2001] Number of respondents 4,188 577 3,611 658 3,424 In vicinity of any conflict .617 .849 .580 .653 .614 In vicinity of conflict at age zero .439 .386 .452 .444 .434 In vicinity of conflict in Chiefdom of birth .974 .888 .995 .985 .971 Youngest age when in vicinity 1.46 1.50 1.44 1.54 1.45 (1.89) (1.86) (1.90) (2.03) (1.86) Cumulative number of conflict episodes in 6.36 11.2 5.57 6.83 6.34 vicinity of (9.77) (14.0) (8.62) (11.3) (9.50) Migration during civil war Ever migrated .138 1 0 .160 .135 Number of migrations | migrated 1.16 1.16 - 1.18 1.16 (.423) (.423) - (.434) (.422) Migration post civil war Ever migrated .677 .967 .630 .696 .671 Number of migrations | migrated 1.75 1.41 1.83 1.76 1.75 (1.01) (1.19) (.947) (1.06) (1.01) Notes: Migration is recorded within the survey's migration journal, where respondents are asked to list all of the location where they resided for at least 6 months. Respondents where also asked about episodes of internal displacement during the war in the victimization module. These includes migration episodes that were either shorter than 6 months, seen as temporary, or towards location in the proximity of the village of residence. As such some of these episodes might have not been listed in the migration journal and do note count as migration spells. Table 2: Experiences of Conflict Means, p-values in braces (5) Parental (1) Background (2) Trauma (3) Direct (4) Socialized Trauma Number of observations (%) 1114 (29%) 380 (10%) 1033 (27%) 656 (17%) 594 (16%) Panel A. Family Background Age 14.5 19.9 21.4 17.6 16.5 Ruling family .141 .146 .170 .193 .185 Born in district close to Liberia .488 .142 .251 .419 .219 Migrated during civil war .038 .197 .238 .108 .121 Conditional on a polynomial Unconditional in age and district of birth Panel B. Vicinity of Conflict Trauma = Direct Trauma = Direct Total number of events 0 13.4 12.8 2.40 5.43 {.236} {.320} Civilian targets 0 6.74 6.14 1.17 2.24 {.011} {.063} Civilians targeted by government 0 1.60 1.19 .241 .545 {.000} {.186} Civilians targeted by rebels 0 5.07 4.91 .905 1.64 {.438} {.145} Total number of casualties 0 206 197 28.1 82.5 {.593} {.559} Total number of civilian casualties 0 65.6 60.4 11.6 20.5 {.277} {.140} Years of exposure 0 3.68 3.48 .733 1.66 {.000} {.008} Age at first exposure - 1.49 2.48 .338 .209 {.000} {.000} Panel C. Victimization Recall Direct = Socialized Direct = Socialized Fighting in the area 0 .925 .892 {.017} {.752} Personally harmed 0 .089 .058 {.018} {.909} Family harmed 0 .500 .489 {.603} {.697} Internally displaced 0 .785 .673 {.000} {.017} Refugee overseas 0 .134 .155 {.221} {.048} Notes: Respondents assigned to groups based on their wartime experience. Panel A and C contain information from the respondents' survey collected in 2014. The information in Panel B is constructed by matching respondents' migration journals, collected as part of the survey in 2014, with conflict data from Uppsala Conflict Data Program. The table refers to all episodes of violence that took place in each respondent's chiefdom(s) of residence. Features are then averaged across respondents. The last two columns report simple test of equality of means across groups. The first of set of tests - unconditional - is performed by regressing the variable of interest on dummies for each of the experience groups, and then testing for equality of the relevant coefficients. The second set of tests is analogous, but regressions additionally control for a cubic polynomial in age and district of birth fixed effects. Table 3: Vicinity to Conflict and Trust Multinomial logit, average marginal effects reported Dependent variable: In general do you think people can be trusted? Standard errors in parentheses, clustered by village of residence (1a) NO (1b) IT DEPENDS (1c) YES (2a) NO (2b) IT DEPENDS (2c) YES Vicinity to Conflict -.015 .071*** -.056** -.028* .077*** -.050* (.016) (.027) (.027) (.017) (.029) (.028) p-value (yes = no) {.250} {.534} Age ✓ ✓ ✓ ✓ ✓ ✓ Indiv. Controls ✓ ✓ ✓ Tribe & Religion ✓ ✓ ✓ Village Controls ✓ ✓ ✓ District of birth FE ✓ ✓ ✓ ✓ ✓ ✓ District 2014 FE ✓ ✓ ✓ ✓ ✓ ✓ Sample Avg | Not in the vicinity of conflict .119 .496 .385 .121 .489 .390 Observations 4,187 3,784 Note: ***, ** and * denote significance at the 1%, 5%, and 10% levels. The table reports average marginal effects for the estimation sample from a multinomial Logit specification. All specifications include a cubic polynomial in age, district of birth and district of residence fixed effects. Controls include dummy variables for: respondents belonging to each of the three largest tribes in our sample (Mende, Temne and Limba); respondents identifying as Muslim; respondents' father has completed junior secondary education; respondent belongs to a ruling family. Village controls include the number of households residing in the community and the average PPI score, the latter being an index capturing household wealth. Standard errors are clustered at the community level. Table 4: Experiences of Conflict and Trust Multinomial logit, average marginal effects reported Dependent variable: In general do you think people can be trusted? Standard errors in parentheses, clustered by village of residence p-value p-value (1a) NO (1b) IT DEPENDS (1c) YES (2a) NO (2b) IT DEPENDS (2c) YES (yes = no) (yes = no) Socialized -.004 .008 -.003 {.981} (.020) (.034) (.033) Traumatized -.049* .108** -.060 {.850} (.028) (.045) (.044) Direct -.053** .071* -.018 {.532} (.023) (.044) (.045) Parental Trauma -.014 .089** -.075** {.210} (.024) (.036) (.036) Trauma, Direct -.049*** .078*** -.029 {.608} (.017) (.029) (.029) Parental Trauma -.011 .084*** -.073** {.122} (.021) (.031) (.029) Background [OMITTED] ✓ ✓ ✓ Background, Socialized [OMITTED] ✓ ✓ ✓ Traumatized = Direct (P-VALUE) {.830} {.361} {.260} Trauma, Direct = Parental Trauma (P-VALUE) {.066} {.816} {.125} Controls for age, individual ✓ ✓ ✓ ✓ ✓ ✓ characteristics, tribe and religion Village Controls ✓ ✓ ✓ ✓ ✓ ✓ District of birth, district of ✓ ✓ ✓ ✓ ✓ ✓ residence fixed effects Sample Avg | Omitted Group .116 .512 .372 .119 .502 .380 Observations 3,411 3,412 Notes: ***, ** and * denote significance at the 1%, 5%, and 10% levels. The table reports average marginal effects for the estimation sample from a multinomial Logit specification. Respondents are grouped based on their experience of conflict, further details can be found in figure 4. All specifications include a cubic polynomial in age, district of birth and district of residence fixed effects. Controls include dummy variables for: respondents belonging to each of the three largest tribes in our sample (Mende, Temne and Limba); respondents identifying as Muslim; respondents' father has completed junior secondary education; respondent belongs to a ruling family. Village controls include the number of households residing in the community and the average PPI score, the latter being an index capturing household wealth. Standard errors are clustered at the community level. Table 5: Conflict and Self-Efficacy OLS regression estimates Dependent variable: self-efficacy measure Standard errors in parentheses, clustered by village of residence (1) Self-Efficacy (2) Self-Efficacy Exposure to Conflict .089* (.046) Trauma, Direct .121** (.053) Parental Trauma .143*** (.054) Sample Average | Omitted group 0 0 Trauma and Direct = Parental Trauma (P-VALUE) {.713} Background + Socialized [OMITTED] ✓ Controls for age, individual characteristics, ✓ ✓ tribe and religion Village Controls ✓ ✓ District of birth, district of residence fixed ✓ ✓ effects Observations 3,784 3,411 Note: ***, ** and * denote significance at the 1%, 5%, and 10% levels. The table reports OLS estimates. All specifications include a cubic polynomial in age, district of birth and district of residence fixed effects. Controls include dummy variables for: respondents belonging to each of the three largest tribes in our sample (Mende, Temne and Limba); respondents identifying as Muslim; respondents' father has completed junior secondary education; respondent belongs to a ruling family. Village controls include the number of households residing in the community and the average PPI score, the latter being an index capturing household wealth. Standard errors are clustered at the community level. The outcome variable is a summary index of several measures of efficacy. We measure efficacy across two dimensions. Business Efficacy is self-reported ability to perform the following tasks on a 10-points Likert scale: Run your own business; Identify business opportunities to start up a new business; Obtain credit to start up a new business or expand an existing business; Save in order to invest in future business opportunities; Make sure that your employees get the work done properly; Manage financial accounts; Bargain to obtain cheap prices when you are buying anything for business (inputs); Bargain to obtain high prices when you are selling anything for business (outputs); Protect your business assets from harm by others; Collecting the money someone owes you; Find information about paid work opportunities in your community. Self-Efficacy is captured by questions recording respondents' agreement with the following statements on a 10-points Likert scale: If I start working on a task, I definitely see the end of it no matter how difficult it is; While doing any task, it is important for me to do it better than others; If I have the chance, I would make a good leader; I want to be a respectful person in my village; I do not care what others think about my success or failure; I am in control of what happens in my life; I save regularly; A person can get rich by taking risks; I often make plans for the future; I believe that my future is determined by luck no matter how hard I work (reverse scale). Components are then aggregated in a Inverse Covariance Weighted Index (Anderson, 2008). In each column, the index is standardised with respect to the control group, i.e. effects can be interpreted as standard deviations from the control group's average. Table 6: Conflict, Self-Efficacy and Trust Multinomial logit, average marginal effects reported Dependent variable: In general do you think people can be trusted? Standard errors clustered by village of residence (1a) NO (1b) IT DEPENDS (1c) YES (2a) NO (2b) IT DEPENDS (2c) YES λ: Self-Efficacy -.001 .056*** -.055*** .007 .052*** -.058*** (.006) (.010) (.009) (.006) (.011) (.010) Exposure to Conflict -.029 .073** -.044* (.017) (.029) (.027) Traumatized and Direct -.050*** .072** -.022 (.017) (.029) (.029) Parental Trauma -.013 .078** -.064** (.021) (.031) (.028) p-value (λ yes = no) {.000} {.000} Controls for age, individual ✓ ✓ ✓ ✓ ✓ ✓ characteristics, tribe and religion Village Controls ✓ ✓ ✓ ✓ ✓ ✓ District of birth, district of residence ✓ ✓ ✓ ✓ ✓ ✓ fixed effects Sample Avg | Not Exp. .121 .489 .390 .118 .502 .380 Observations 3,784 3,411 Notes: ***, ** and * denote significance at the 1%, 5%, and 10% levels. The table reports average marginal effects for the estimation sample from a multinomial Logit specification. All specifications include a cubic polynomial in age, district of birth and district of residence fixed effects. Controls include dummy variables for: respondents belonging to each of the three largest tribes in our sample (Mende, Temne and Limba); respondents identifying as Muslim; respondents' father has completed junior secondary education; respondent belongs to a ruling family. Village controls include the number of households residing in the community and the average PPI score, the latter being an index capturing household wealth. Standard errors are clustered at the community level. The components of the Business Efficacy Index are self-reported ability to perform the following tasks on a 10-points Likert scale: Run your own business; Identify business opportunities to start up a new business; Obtain credit to start up a new business or expand an existing business; Save in order to invest in future business opportunities; Make sure that your employees get the work done properly; Manage financial accounts; Bargain to obtain cheap prices when you are buying anything for business (inputs); Bargain to obtain high prices when you are selling anything for business (outputs); Protect your business assets from harm by others; Collecting the money someone owes you; Find information about paid work opportunities in your community. The components of the Self-Efficacy Index are questions recording respondents' agreement with the following statements on a 10-points Likert scale: If I start working on a task, I definitely see the end of it no matter how difficult it is; While doing any task, it is important for me to do it better than others; If I have the chance, I would make a good leader; I want to be a respectful person in my village; I do not care what others think about my success or failure; I am in control of what happens in my life; I save regularly; A person can get rich by taking risks; I often make plans for the future; I believe that my future is determined by luck no matter how hard I work (reverse scale). Components are aggregated in a Inverse Covariance Weighted Index (Anderson, 2008). In each column, the index is standardised with respect to the control group, i.e. effects can be interpreted as standard deviations from the control group's average. Table 7: Conflict, Anxiety and Satisfaction Dependent variable Columns 1-8: Ever worry about…? Dependent variable Column 9-14: Satisfaction indices Standard errors in parentheses, clustered by village of residence Anxiety Satisfaction (2) Finding (4) Violence/ (6) Finding (8) Violence/ (1) Job (3) Money (5) Job (7) Money (9) Social (10) Economic (11) Life (12) Social (13) Economic (14) Life Husband Theft Husband Theft Exposure to Conflict .034 .005 .029 .062** .118** .101* .158** (.023) (.024) (.022) (.025) (.055) (.056) (.061) Traumatized and Direct .017 -.019 .040 .095*** .115* .154** .218*** (.026) (.027) (.027) (.029) (.063) (.060) (.068) Parental Trauma .005 .019 -.003 .046 .158*** .141** .275*** (.026) (.028) (.025) (.029) (.056) (.058) (.060) Background and Socialized [OMITTED] ✓ ✓ ✓ ✓ ✓ ✓ ✓ Controls for age, individual characteristics, tribe and ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ religion Village Controls ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ District of birth, district of ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ residence fixed effects Sample Avg | Omitted group .572 .490 .751 .462 .603 .507 .774 .471 0 0 0 0 0 0 Observations 3,784 3,412 3,784 3,784 3,784 3,412 3,412 3,412 Notes: The table reports OLS estimates. In Columns 1 to 8, the outcome variables are dummies indicating whether the respondent ever worries about a specific topic or issue. In Columns 9 to 14 the outcome variables are indices of satisfaction computed over three spheres: social, economic, and life in general. The indices are inverse covariance weighted indices [Anderson 2008] constructed from questions asking respondents to rate their satisfaction on a Likert scale. The dimensions over which this questions are asked are: (i) Social: friends, family, community, house; (ii) Economic: own education, job, earnings, household finances, access to credit; (ii) Life: life as a whole, future prospects. The indices are standardised relative to the omitted group in each Column. All specifications include a cubic polynomial in age, district of birth and district of residence fixed effects. Controls include dummy variables for: respondents belonging to each of the three largest tribes in our sample (Mende, Temne and Limba); respondents identifying as Muslim; respondents' father has completed junior secondary education; respondent belongs to a ruling family. Village controls include the number of households residing in the community and the average PPI score, the latter being an index capturing household wealth. Standard errors are clustered at the community level. Table 8: Victimization and Trust Column 1: OLS regression estimates, dependent variable = self-efficacy measure Dependent variable Columns 2a-2c: In general do you think people can be trusted? Standard errors in parentheses, clustered by village of residence (1) Self-efficacy (2a) NO (2b) IT DEPENDS (2c) YES Victimized: Fighting in the area .012 .002 .107** -.108** (.090) (.030) (.047) (.044) Victimized: Personally harmed -.144 .072 -.054 -.019 (.090) (.049) (.054) (.053) Victimized: Family harmed .130** -.002 -.022 .024 (.062) (.024) (.034) (.035) Victimized: Internally displaced .100 -.086*** .157*** -.071 (.079) (.025) (.041) (.045) Victimized: Refugee overseas .179* -.096*** .009 .087 (.102) (.015) (.058) (.058) Sample Avg | VICT==0 0 .116 .512 .372 Controls for age, individual characteristics, tribe and religion ✓ ✓ ✓ ✓ Village Controls ✓ ✓ ✓ ✓ District of birth, district of residence fixed effects ✓ ✓ ✓ ✓ Observations 1,879 1,879 Note: ***, ** and * denote significance at the 1%, 5%, and 10% levels. Column 1 reports OLS estimates, where the dependent variable is our index of self-efficacy. Column 2a through 2c report AMEs from a multinomial Logit specification, where the dependent variable is our key measure of trust in others. All specifications include a cubic polynomial in age, district of birth and district of residence fixed effects. Controls include dummy variables for: respondents belonging to each of the three largest tribes in our sample (Mende, Temne and Limba); respondents identifying as Muslim; respondents' father has completed junior secondary education; respondent belongs to a ruling family. Village controls include the number of households residing in the community and the average PPI score, the latter being an index capturing household wealth. Standard errors are clustered at the community level. In Column 1, the outcome variable is a summary index of several measures of efficacy. We measure efficacy across two dimensions. Business Efficacy is self-reported ability to perform the following tasks on a 10-points Likert scale: Run your own business; Identify business opportunities to start up a new business; Obtain credit to start up a new business or expand an existing business; Save in order to invest in future business opportunities; Make sure that your employees get the work done properly; Manage financial accounts; Bargain to obtain cheap prices when you are buying anything for business (inputs); Bargain to obtain high prices when you are selling anything for business (outputs); Protect your business assets from harm by others; Collecting the money someone owes you; Find information about paid work opportunities in your community. Self-Efficacy is captured by questions recording respondents' agreement with the following statements on a 10-points Likert scale: If I start working on a task, I definitely see the end of it no matter how difficult it is; While doing any task, it is important for me to do it better than others; If I have the chance, I would make a good leader; I want to be a respectful person in my village; I do not care what others think about my success or failure; I am in control of what happens in my life; I save regularly; A person can get rich by taking risks; I often make plans for the future; I believe that my future is determined by luck no matter how hard I work (reverse scale). Components are then aggregated in a Inverse Covariance Weighted Index (Anderson, 2008). In each column, the index is standardised with respect to the control group, i.e. effects can be interpreted as standard deviations from the group that experienced no victimization. All victimization measures are self-reported and measured in 2014. Figure 1: Framework for Understanding Trust Preferences A. Persuasion B. Equilibrium C. Comparative Static: Increase in Self Efficacy D. Comparative Static: Less Extreme Beliefs over Others Figure 2: Cumulative Fighting Intensity, by Chiefdom 1991 (Start, Close to Border with Liberia) 1993 (RUF pushed back) LIBERIA 1997 (coup and taking of Freetown) 2001 (End) Notes: Data from the Uppsala Conflict Data Program (UCDP). The maps portray the cumulative number of episodes of conflict, at the chiefdom level, at the end of four key years of the Sierra Leonean civil war. In this context, an episode of conflict is defined as the use of armed forced by an organized actor against another, or against civilians, that resulted in at least 1 death. The conflict started in 1991 when the Revolutionary United Front (RUF) began taking control of the country, with limited opposition from the Sierra Leonean Army (SLA). In 1993 a government offensive supported by ECOMOG pushed the RUF back towards the Liberian border. Notwithstanding this effort, RUF regained territories, approaching the capital Freetown in 1995. In 1997, dissidents within the SLA formed the Armed Forces Revolutionary council (AFRC) and staged a successful coup with RUF's blessing, leading to rebels conquering the capital. A UN peace keeping mission was deployed in 1999 and, joined by British forces in 2000, they quickly regained control of country. The last episode of conflict was recorded in 2001, and the war was declared to be over in January 2002. Figure 3: Time Series Variation in Conflict Panel A. Fighting by year and Chiefdom Panel B. Sample year of birth Notes: The central panel reports, for each year, which chiefdoms experienced any episode of fighting (source: UCDP). The bottom panel reports year of birth of respondents in our sample. Region 1: Pujehun, Kenema, Kailahun. Region 2: Bonthe, Moyamba, Bo, Tonkolili, Kono, Koinadugu. Region 3: Western Urban, Western Rural, Port Loko, Kambia, Bombali. Figure 4: Victimization A. Victimization B. Victimization, by Age at End of Civil War Notes: Panel A (left) reports sample averages for being in the vicinity of conflict, recalling wartime experience and victimization. Vicinity to conflict is computed by matching self-reported migration journals with conflict data from UCDP. The match is performed at the chiefdom-year level. Respondents are identified as recalling their civil war experience if they answered at least one of the questions on victimization, as opposed to answering "don't know/don't remember" to all five questions. The sample shares shown are conditional on vicinity status. The third set of bars shows the share of the sample reporting any form of victimization, conditional on vicinity status and on recalling war time. Panel A (right) shows details on the five form of victimization for which data was collected. Sample share are conditional on recalling war time. Panel (B) shows sample shares for the three variables - vicinity, recall, and victimization - conditional on respondent's age in 2001, the year when the last episode of violence took place. Figure 5: Constructing Experiences of Conflict : not in vicinity of conflict : in vicinity of conflict Sample Share Background narrative .266 .266 Pre-memory (0-2) Memory (recall) (3+) Recall Victimization? 1 No Socialized .085 .157 2 Yes Socialized .072 3 No 4 Yes Trauma .020 .091 5 No Trauma .071 6 Yes 7 No Direct .107 .247 8 Yes Direct .140 Parental trauma .142 .142 Notes: Vicinity to conflict is measured using data from migration journals, which record every Chiefdom respondents lived in for at least six months, since birth. A respondent is in the vicinity of conflict if at least one episode of violence took place in her Chiefdom of residence. Victimization is measured over five dimensions: personal harm, family members harmed, fighting in the area of residence, internal displacement, refugee abroad. We asked respondents whether they experienced each form of victimization, and the possible answers were: (i) "did not live in Sierra Leone", (ii) "too young/don't remember", (iii) "no", and (iv) "yes". Respondents are classified as recalling victimization if they answered "yes" to at least one type of victimization. They are then split into Socialised and Direct Experience based on their vicinity to conflict. Respondents who answer "too young/don't remember" to all five types of victimization are classified as not recalling any victimization. They form the Background Narrative, Trauma and Parental Trauma groups based on their vicinity to conflict. We also consider those respondents that answered "no" to each victimization question, i.e. those respondents that recall not being victimized at all. These respondents enter the Background Narrative and Parental Trauma groups depending on the vicinity status. For these two experiences, both not recalling victimization or recalling not being victimized fit within the assigned classification given that they were either not in the vicinity of conflict or too young to remember. The exception is the Trauma group, where we do not include those respondents recalling explicitly not being victimized (answering "no " to every question) even if they were in the vicinity of conflict from age 3 onwards, as this specific case does not fit within the type of experience that characterizes this group. Figure 6: Vicinity to Conflict and Trust in Groups Notes: The figure reports average marginal effects from a multinomial Logit speciation, with 90% confidence intervals. The outcome variable is the answer to the following question: "How many people in [group ] can you trust?". This is coded to take three categorical value. All specifications include a cubic polynomial in age, district of birth and district of residence fixed effects. Controls include dummy variables for: respondents belonging to each of the three largest tribes in our sample (Mende, Temne and Limba); respondents identifying as Muslim; respondents' father has completed junior secondary education; respondent belongs to a ruling family. Village controls include the number of households residing in the community and the average PPI score, the latter being an index capturing household wealth. Standard errors are clustered at the community level. Table A1: Literature Review (studies covered in Bauer et al . 2016 meta-analysis) Time since Age of Paper Country Conflict Sample war Measure of Exposure Exposure exposure 1) Self-reported This paper Sierra Leone Civil war (1991–2002) ~4,000 young women 12 years 0-9 2) Geo-reference via migration and conflict data Representative sample adolescence, early Annan, Blattman, Mazurana, and Lord’s Resistance Army (LRA) of youth, some of whom adulthood (abduction Self-reported abduction by LRA; questions about 17 specific acts Uganda ~7 years Carlson (2011) insurgency (1986–2006) were conscripted by for conscription of violence, experienced or perpetrated, on self and family LRA; N = 613 purpose) Georgia: war with Russia over Georgia: children; N = Georgia: 6 Bauer, Cassar, Chytilová, and Georgia and South Ossetia (2008) 565 months Georgia: 3-12 Self-reported victimization and displacement Henrich (2014) Sierra Leone Sierra Leone: civil war Sierra Leone: adult Sierra Leone: 8 Sierra Leone: 0-75 (1991–2002) population; N = 586 years Young men, some of Abducted from Lord’s Resistance Army Self-reported abduction by LRA; questions about 17 specific acts Bauer, Fiala, and Levely (2014) Uganda whom were conscripted 5 years childhood to later insurgency (1986–2006) of violence, experienced or perpetrated, on self and family by LRA; N = 337 years. 10-50 Self-reported questions used to create a victimization index and Nationally representative All adults/no age number of reported attacks and battles within each chiefdom as Bellows and Miguel (2006, 2009) Sierra Leone Civil war (1991–2002) 3–5 years sample; N = 10,496 range information another violence measure. Abducted from Young men, some of Lord’s Resistance Army childhood to later Blattman (2009) Uganda whom were conscripted ~5 years Self-reported violence related questions insurgency (1986–2006) years. Mean age of by LRA; N = 741 abduction=15 yrs Cassar, Grosjean, and Whitt (2013) Tajikistan Civil war (1992–1997) Adult population; N = 426 13 years Ages 0-64 Self-reported violence related questions Cecchi, Leuveld, Voors, and van Youth male street Sierra Leone Civil war (1991–2002) 8 years Ages 1-23 Self-reported questions used to create a victimization index der Wal (2015) football players; N = 162 Lord’s Resistance Army Nationally representative All adults/no age De Luca and Verpoorten (2015a) Uganda 12 years External violent acts reports - ACLED insurgency (1986–2006) sample; N = 4,671 range information Lord’s Resistance Army Nationally representative All adults/no age De Luca and Verpoorten (2015b) Uganda 12 years External violent acts reports - ACLED insurgency (1986–2006) sample; N = 4,671 range information Gilligan, Pasquale, and Samii Household heads; N = Nepal Civil war (1996–2006) 3 years 14-76 Village Development Committees-level fatality figures (2014) 252 Senior citizens that Gneezy and Fessler (2012) Israel Israel–Hezbollah war (2006) Senior citizens; N = 50 1 year were not called into Experiment before and after war - no spatial conflict data used military service WWII (1939–45); 35 countries in Yugoslav wars (1991–95); Mostly Europe, the Kosovo war (1998–99); Nationally representative 5 months–65 Self-reported survey questions on if you or parents or Grosjean (2014) parent/grandparent Caucasus, and Tajik civil war (1992–97); samples; N = 38,864 years grandparents were harmed exposure Central Asia Chechen wars (1994–2009); Kyrgyzstan clashes (2010) Former soldiers who Grossman, Manekin, and enlisted between Health rankings assigned in the IDF recruitment process as an Israel Israeli–Palestinian conflict (1967+) 1–12 years 21-33 Miodownik (2015) 1998–2003 and instrument for combat exposure 2004–2009; N = 2,334 Rohner, Thoenig, and Zilibotti Lord’s Resistance Army Nationally representative Uganda 8 years Ages 10-73 Fighting events reported in ACLED (2013) insurgency (1986–2006) sample; N = 2,431 1) Self-reported community level conflict victimization from village Household heads, N = Voors et al. (2012) Burundi Civil war (1993–2005) 4–6 years Ages 2-85 level focus groups 2) self-reported household level conflict 287 questions used for a victimization index Adult population; N = Voors and Bulte (2014) Burundi Civil war (1993–2005) 4 years 14-85 Self-reported questions used to create a victimization index 874 Table A1: Literature Review Continued (other studies) Time since Age of Paper Country Conflict Sample war Measure of Exposure Exposure exposure 1) Self-reported. This paper Sierra Leone Civil war (1991–2002) ~4,000 young women 12 years 0-9 2) Geo-reference via migration and conflict data Civil war, focus on period between 1127 adults near polling All adults/no age Incident records of the International Security Assistance Force, a Callen et al. (2014) Afghanistan 0 years 2002-2010 stations range information multilateral military body present since December 2001 1280 adults across 64 Hartman and Morse (2008) Liberia Civil war (1990-2003) 10 years 15-65 Self-reported violence related questions villages Jakiela and Ozier (2019) Kenya Post election crisis 2008 N=5049 1 year 13-30 Self-reported violence related questions Population data from 1949 census + civilian injuries and casualty Kim and Lee (2014) Kenya Kenya war (1950-1953) N=7047 50 years 0-31 data 0-10 years Modern conflict in Columbia (1985 All adults/no age Moya (2018) Columbia N=284 IDPs (average 2.5 Self-reported violence related questions - ongoing) range information years) Not eligible for met-analysis Number of total killings per VDC. Information is taken from the De Juan and Pierskalla Nationally representative All adults/no age Nepal Civil war (1996–2006) 0–7 years Informal Sector Services Center (INSEC). Use respondent's GPS (2016) sample; N = 8,822 range information info to match with VDCs Adult population; N ~ All adults/no age Hartman and Morse (2020) Liberia Civil war (1989–2003) 10 years Self-reported exposure to violence questions 1,600 range information Indonesia: N = 1,752 Indonesia, Indonesia: insurgency in Aceh Bosnia: nationally Bosnia and (1976–2005) representative sample; N Hercegovina, Shewfelt (2009) B&H: civil war (1992–1995) = 3,580. 2–11 years United States United States: Vietnam war United States: male (Vietnam (1955–1975) Vietnam theater veterans) veterans; N = 1,171 Other forms of violence Latin America: 39,238 United States and Canada: 3,000 All adults/no age Bateson (2012) 70 countries Crime victimization Self-reported crime victimization Africa: 27,713 range information Europe: 17,088 Asia: 16,725 Becchetti, Conzo, and Romeo Kenyan crisis, post-election Nairobi slum-dwellers; N All adults/no age Self-reported crime victimization, divided into 3 categories: i) Kenya (2014) violence (2007–2008) = 404 range information direct or indirect harm; ii)economic losses; iii)forced relocation Hopfensitz and Miquel-Florensa All adults/no age Displacement history. Also linked to average homicide rate over Colombia Colombian conflict (1964+) Coffee farmers; N = 260 (2014) range information the last 10 years Nationally representative Rojo-Mendoza (2014) Mexico Crime victimization sample; N = 7,416 Table A2: Attrition Dependent Variable = 1 if respondent is tracked 2014-16 OLS estimates, standard errors in parentheses (1) (2) (3) Trust: No -.015 -.018 (.017) (.017) Trust: It depends .012 .006 (.011) (.012) Does not recall -.006 (.012) Individual Controls No Yes Yes Tribe and Religion No Yes Yes Village Controls No Yes Yes F-tests: p-values on joint significance: Individual Controls {.055} {.061} Tribe and Religion {.000} {.000} Village Controls {.008} {.009} Mean of outcome variable .838 .838 .838 Observations 5,376 4,979 4,979 Notes: *** denotes significance at 1%, ** at 5%, and * at 10%. Results from a regression where the outcome variable is a dummy equal to 1 if the respondent not tracked at endline (2016). All specifications include a cubic polynomial in age, district of birth and district of residence fixed effects. Controls include dummy variables for: respondents belonging to each of the three largest tribes in our sample (Mende, Temne and Limba); respondents identifying as Muslim; respondents' father has completed junior secondary education; respondent belongs to a ruling family. Village controls include the number of households residing in the community and the average PPI score, the latter being an index capturing household wealth. Standard errors are clustered at the community level. Table A3: Correlates of Exposure to Conflict Standard errors in parentheses, clustered by village of residence Panel A: Geo-Matched Exposure Panel B: Self-Reported Victimization Conflict Exposure Conflict Exposure Any Personally Family Internally Fighting in Outcome: Exposure Intensity Exposure Intensity Recall Refugee Victimisation Harmed Harmed Displaced the area (any) (Tobit) (any) (Tobit) Sample: Full Victimization Victimization (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) Age (AME) .106*** 3.78*** .079*** 2.74*** .045*** .044*** .003 .028*** -.003 .042*** .042*** (.004) (.251) (.004) (.251) (.005) (.005) (.003) (.007) (.003) (.006) (.006) Father's H-Education -.009 -.016 -.010 .081 .003 .027 .027 -.021 .040 -.006 .017 (mother) (.020) (.596) (.030) (.810) (.026) (.028) (.020) (.038) (.025) (.036) (.030) Ruling Family -.006 -.207 -.019 -.957 -.006 .020 .019 .015 .046** -.017 .010 (.016) (.616) (.021) (.722) (.019) (.020) (.016) (.030) (.022) (.029) (.026) F-tests: p-values on joint significance Parental Characteristics {.890} {.817} {.807} {.341} {.759} {.836} {.776} {.465} {.832} {.829} {.871} Tribe {.000} {.000} {.000} {.000} {.021} {.384} {.643} {.348} {.000} {.000} {.062} Religion {.789} {.682} {.908} {.563} {.477} {.424} {.222} {.079} {.168} {.102} {.050} District of birth FE ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ District 2014 FE ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Sample Mean .626 6.47 .729 8.08 .443 .850 .065 .402 .120 .615 .764 Observations 3,784 3,517 2,107 1,954 3,784 2,107 1,975 1,993 1,980 1,953 1,925 Notes: *** denotes significance at 1%, ** at 5%, and * at 10%. All victimization measures are self-reported. Geo-matched exposure intensity is defined as the total number of war-related episodes of violence that took place in the chiefdom where the respondent was residing, as reported by the UCDP. Column (3) and (4) repeat the analysis for the subsample of respondents that completed the victimization module in its entirety. Coefficients estimated via OLS, unless otherwise specified. Age is controlled for via a cubic polynomial, and the table reports average marginal effects. The education control is a dummy equal to one if respondents' father has completed junior secondary. The variable Ruling Family is a dummy equal to 1 if the respondent belongs to one of the families whose members can run for the role of paramount chief. These families were recognised and empowered by the British Colonial Administration in 1896. Paramount chiefs are elected by local notables, their mandate is for life and they represented the only form of local administration until local councils were formed in 2004. Their powers include collecting taxes, overseeing the judicial system and allocating land. Each regression includes a set of dummies for each three major tribes, a dummy equal to 1 if respondents identify as Muslim, district of birth and district of residence fixed effects. Standard errors are clustered at the community level. Table A4: Intensive Margin of the Geo-coded Exposure to Conflict and Trust Multinomial logit, average marginal effects reported Dependent variable: In general do you think people can be trusted? Standard errors in parentheses, clustered by village of residence (1a) NO (1b) IT DEPENDS (1c) YES (2a) NO (2b) IT DEPENDS (2c) YES Exposure to Conflict, intensive -.008 .022** -.014 (Number of Episodes, asinh transf.) (.006) (.010) (.010) Exposure to Conflict, intensive -.008** .011** -.004 (Number of casualties, asinh transf.) (.003) (.006) (.006) p-value (yes = no) {.602} {.538} Age ✓ ✓ ✓ ✓ ✓ ✓ Indiv. Controls ✓ ✓ ✓ ✓ ✓ ✓ Tribe & Religion ✓ ✓ ✓ ✓ ✓ ✓ Village Controls ✓ ✓ ✓ ✓ ✓ ✓ District of birth FE ✓ ✓ ✓ ✓ ✓ ✓ District 2014 FE ✓ ✓ ✓ ✓ ✓ ✓ Sample Avg | Not in the vicinity of conflict .119 .498 .383 .120 .497 .384 Observations 3,517 3,517 Note: ***, ** and * denote significance at the 1%, 5%, and 10% levels. Columns 1a-1c use the total number of episodes of violence experienced during the war, while Columns 2a-2c use the total number of deaths that took place in each respondent's chiefdom(s) of residence. Given the large mass at zero and the long tail for each of the two variables, we employ inverse hyperbolic sine transformations. All specifications include a cubic polynomial in age, district of birth and district of residence fixed effects. Controls include dummy variables for: respondents belonging to each of the three largest tribes in our sample (Mende, Temne and Limba); respondents identifying as Muslim; respondents' father has completed junior secondary education; respondent belongs to a ruling family. Village controls include the number of households residing in the community and the average PPI score, the latter being an index capturing household wealth. Standard errors are clustered at the community level. Table A5: Robustness Dependent variable: In general do you think people can be trusted? Standard errors in parentheses, clustered by village of residence Multinomial Logit, AME Probit, AME LPM (1a) NO (1b) IT DEPENDS (1c) YES (2) IT DEPENDS (3) IT DEPENDS (4) IT DEPENDS Exposure to Conflict -.027* .077*** -.050* .077*** .077** .046*** (.017) (.029) (.028) (.030) (.030) (.017) Alternative clustering of standard errors: district of birth (.017) (.037) (.037) age (.017) (.026) (.023) robust (.016) (.023) (.022) Age ✓ ✓ ✓ ✓ ✓ ✓ Controls for age, individual ✓ ✓ ✓ ✓ ✓ ✓ characteristics, tribe and religion Village Controls ✓ ✓ ✓ ✓ ✓ ✓ District of birth, district of ✓ ✓ ✓ ✓ ✓ ✓ residence fixed effects Enumerator fixed effects ✓ Sample Avg | Not Exp. .121 .489 .390 .489 .489 .489 Observations 3,784 3,784 3,784 3,784 Note: ***, ** and * denote significance at the 1%, 5%, and 10% levels. All specifications include a cubic polynomial in age, district of birth and district of residence fixed effects. Controls include dummy variables for: respondents belonging to each of the three largest tribes in our sample (Mende, Temne and Limba); respondents identifying as Muslim; respondents' father has completed junior secondary education; respondent belongs to a ruling family. Village controls include the number of households residing in the community and the average PPI score, the latter being an index capturing household wealth. Standard errors are clustered at the community level. Table A6: Vicinity to Conflict and Trust Multinomial logit, average marginal effects reported Dependent variable: In general do you think people can be trusted? Standard errors in parentheses, clustered by village of residence Within 50km Within 100km (1a) NO (1b) IT DEPENDS (1c) YES (2a) NO (2b) IT DEPENDS (2c) YES Vicinity to Conflict - -.007 .027* -.020 -.017 .021 -.004 Distance Weighted (.010) (.016) (.015) (.018) (.021) (.014) Age ✓ ✓ ✓ ✓ ✓ ✓ Indiv. Controls ✓ ✓ ✓ ✓ ✓ ✓ Tribe & Religion ✓ ✓ ✓ ✓ ✓ ✓ Village Controls ✓ ✓ ✓ ✓ ✓ ✓ District of birth FE ✓ ✓ ✓ ✓ ✓ ✓ District 2014 FE ✓ ✓ ✓ ✓ ✓ ✓ Observations 2,906 3,293 Notes: ***, ** and * denote significance at the 1%, 5%, and 10% levels. The table reports average marginal effects for the estimation sample from a multinomial Logit specification. The vicinity variables are the sum on the number of episodes of violence experienced within the stated threshold, weighted by the inverse of the distance between the chiefdom's centroid and the location where fighting took place. The logarithm of these measures are then standardised and coefficients can be interpreted as effect sizes. We present results for different cutoffs, with a the average Chiefdom size corresponding to a threshold of 22km. All specifications include a cubic polynomial in age, district of birth and district of residence fixed effects. Controls include dummy variables for: respondents belonging to each of the three largest tribes in our sample (Mende, Temne and Limba); respondents identifying as Muslim; respondents' father has completed junior secondary education; respondent belongs to a ruling family. Village controls include the number of households residing in the community and the average PPI score, the latter being an index capturing household wealth. Standard errors are clustered at the community level. Table A7: Indecisiveness, Social Networks OLS regression estimates Outcomes: Neutral Answers, Life Satisfaction module Standard errors in parentheses, clustered by village of residence Indecisiveness Network Degree (1) Neutral Answer, (2) Neutral (4) Intimate (3) Friends (5) Work (6) Credit Any Answers, Share Topics Exposure to Conflict -.056* -.003 .015 .044 .052 .058 (.029) (008) (.072) (.046) (.060) (.076) Sample Average | Omitted group .608 .112 1.96 .847 .833 .877 Controls for age, individual characteristics, ✓ ✓ ✓ ✓ ✓ ✓ tribe and religion ✓ ✓ ✓ ✓ Village Controls ✓ ✓ ✓ ✓ ✓ ✓ District of birth, district of residence fixed ✓ ✓ ✓ ✓ ✓ ✓ effects Observations 3,784 3,784 2,529 Notes: ***, ** and * denote significance at the 1%, 5%, and 10% levels. The table reports OLS estimates. The dependent variables in columns 1 and 2 are derived from a set of questions measuring respondents' satisfaction along a number of dimensions. The questions are phrased as follows: "How do you feel about [category]?". The available categories are: own education, family, friends, job, income, own dwelling, own school, own community, future prospects, family's financial situation, access to credit, and life as a whole. Available answers belong to a 7-points Likert scale, ranging from very happy/optimistic to very sad/pessimistic. The middle of the scale is labelled as "neutral". Columns 3-6 report Tobit estimates of the impact of conflict exposure on respondents' network degree. This was measured at endline, in 2016, for a random subset of respondents who were asked to list the people they consider friends, they speak about intimate topics with, they can discuss work issues and opportunities with, or they can discuss issues related to finances and credit. All specifications include a cubic polynomial in age, district of birth and district of residence fixed effects. Controls include dummy variables for: respondents belonging to each of the three largest tribes in our sample (Mende, Temne and Limba); respondents identifying as Muslim; respondents' father has completed junior secondary education; respondent belongs to a ruling family. Village controls include the number of households residing in the community and the average PPI score, the latter being an index capturing household wealth. Standard errors are clustered at the community level. Table A8: Sources of Variation in Exposure to Conflict Linear Probability Model Dependent variable: In general do you think people can be trusted? (=1 if it depends) Standard errors clustered by village of residence Age 14-25 Geography of Birth Migration (1) (2) (3) (4) (5) Exposure to Conflict .079** .074** .075** .085*** .109*** (.031) (.036) (.039) (.031) (.040) Exposure to Conflict x Born into district bordering Liberia .007 (.046) Exposure to Conflict x Born into region 2 -.101** (.051) Exposure to Conflict x Born into region 3 .060 (.051) Exposure to Conflict x Migrated during civil war -.074 (.063) Exposure to Conflict x Migrated after civil war -.043 (.038) Controls for age, individual characteristics, ✓ ✓ ✓ ✓ ✓ tribe and religion Village Controls ✓ ✓ ✓ ✓ ✓ District of birth, district of residence fixed ✓ ✓ ✓ ✓ ✓ effects Sample Avg | Not Exp. .513 .490 .498 .489 .498 Observations 3,216 3,784 3,784 3,517 3,784 Note: ***, ** and * denote significance at the 1%, 5%, and 10% levels. The table reports average marginal effects for the estimation sample from a multinomial Logit specification. All specifications include a cubic polynomial in age, district of birth and district of residence fixed effects. Controls include dummy variables for: respondents belonging to each of the three largest tribes in our sample (Mende, Temne and Limba); respondents identifying as Muslim; respondents' father has completed junior secondary education; respondent belongs to a ruling family. Village controls include the number of households residing in the community and the average PPI score, the latter being an index capturing household wealth. Standard errors are clustered at the community level. The geographical regions correspond to those detailed in Figure 2. Standard errors are clustered at the community level. Table A9: Placebo and IV Specifications Multinomial logit, average marginal effects reported Dependent variable: In general do you think people can be trusted? Standard errors in parentheses, clustered by village of residence Placebo: YoB+5 IV IV (1a) NO (1b) IT DEPENDS (1c) YES (2) IT DEPENDS (3) IT DEPENDS Exposure to Conflict -.015 -.066 .080 .136** .165*** (.096) (.121) (.124) (.063) (.063) p-value (yes = no) {.608} Controls for age, individual ✓ ✓ ✓ ✓ ✓ characteristics, tribe and religion ✓ ✓ ✓ ✓ ✓ Village Controls ✓ ✓ ✓ ✓ ✓ District of birth FE ✓ ✓ ✓ District 2014 FE ✓ ✓ ✓ ✓ ✓ IV: Distance from first episode at birth ✓ ✓ IV: Distance from Freetown at birth ✓ IV: F-Stat / {p-value} 32.7 / {.000} 34.5 / {.000} Sample Avg | Not in the vicinity of conflict .116 .481 .404 .509 .509 Observations 2,061 3,705 3,705 Note: ***, ** and * denote significance at the 1%, 5%, and 10% levels. The table reports average marginal effects for the estimation sample from a multinomial Logit specification (Columns 1a-1c), and 2SLS estimates (Columns 2 and 3). Columns 1a-1c report a placebo test whereby year of birth and the date of each migration spell is increased by 5 years before matching this information with UCDP data and compute vicinity to conflict. We restrict the estimation sample to those respondent who would still have been born no later than 2001 according to this placebo year of birth. The last two columns reports 2SLS estimates. We use two sets of instruments for vicinity to conflict: a set of dummy variables for each year of the conflict and equal to 1 from respondents' year of birth onwards, interacted with either distance between place of birth and location of first episode of violence in 1991, or distance between place of birth and Freetown. All specifications include a cubic polynomial in age, and district of residence fixed effects. Controls include dummy variables for: respondents belonging to each of the three largest tribes in our sample (Mende, Temne and Limba); respondents identifying as Muslim; respondents' father has completed junior secondary education; respondent belongs to a ruling family. Village controls include the number of households residing in the community and the average PPI score, the latter being an index capturing household wealth. Standard errors are clustered at the community level. Figure A1: Migration Histories A. Respondents' Chiefdom of Birth B. Respondents' Chiefdom of Residence, 2016 C. Past Conflict Intensity, Origin Minus Destination D. Future Conflict Intensity, Post migration Notes: Panel A shows cumulative conflict intensity for all chiefdoms where at least on respondent was born. Panel B shows the same data for the chiefdoms where sampled respondents currently live. Cumulative conflict intensity is defined as the total number of violent episodes that took place in a Chiefdom between 1991 and 2001. A violent episode is defined as the use of force by an organized group, against another group or civilians, that resulted in at least one casualty. Panel C and D shows data for 921 migration spells that took place between 1991 and 2000. For each spell, we compute (i) difference in total number of episodes of violence between chiefdom of origin and of destination at the time of migration, (ii) difference in total number of episodes of violence taking place in the two chiefdoms after the migration spell. Figure A2: Components of Self-Efficacy OLS coefficients of Vicinity to Conflict Effect sizes, 90% confidence intervals Personal traits Notes: OLS coefficients with 90% confidence intervals. The outcome variables are the self-reported ability to perform a range of business-related tasks (top panel), and agreement with a sequence of statements about oneself, measured on a 10 point scale and later standardized. Estimates obtained by regressing each of these measures on vicinity to conflict. All specifications include a cubic polynomial in age, district of birth and district of residence fixed effects. Controls include dummy variables for: respondents belonging to each of the three largest tribes in our sample (Mende, Temne and Limba); respondents identifying as Muslim; respondents' father has completed junior secondary education; respondent belongs to a ruling family. Village controls include the number of households residing in the community and the average PPI score, the latter being an index capturing household wealth. Standard errors are clustered at the community level. Figure A3: IV, First Stage Estimates Notes: Each picture show first stage coefficient estimates from the IV models reported. For convenience, we report only the coefficients on the instruments, with 90% confidence intervals. These instruments are constructed by multiplying a vector of dummy variables, one for each year between 1991-2000 and equal from the respondent year of birth onward, with two distance measures. In the left panel, the distance used is between place of birth and the first episode of violence recorded in UCDP data, which took place in Pujehun district in 1991. In the right hand panel, we the previous measure as well as the distance between the respondent's place of birth and Freetown, Sierra Leone's capital.