Violent conflict and the demand for healthcare: How armed conflict reduces trust, instills fear, and increases child mortality

Schaub, Max Article — Published Version Violent conflict and the demand for healthcare: How armed conflict reduces trust, instills fear, and increases child mortality Social Science & Medicine Provided in Cooperation with: WZB Berlin Social Science Center Suggested Citation: Schaub, Max (2024) : Violent conflict and the demand for healthcare: How armed conflict reduces trust, instills fear, and increases child mortality, Social Science & Medicine, ISSN 1873-5347, Elsevier, Amsterdam, Vol. 359, pp. 1-10, https://doi.org/10.1016/j.socscimed.2024.117252 This Version is available at: https://hdl.handle.net/10419/312201 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. http://creativecommons.org/licenses/by/4.0/ Contents lists available at ScienceDirect Social Science & Medicine journal homepage: www.elsevier.com/locate/ssm Violent conflict and the demand for healthcare: How armed conflict reduces trust, instills fear, and increases child mortality✩ Max Schaub University of Hamburg and WZB Berlin Social Science Center, Germany ARTICLE INFO Dataset link:https://doi.org/10.7910/DVN/0B UKZK Keywords: Armed conflict Child health Healthcare-seeking Vaccination Fear Trust Africa ABSTRACT What are the health effects of violent conflict? It is well known that wars kill civilians away from the battlefield and long after the fighting has stopped. Yet why this happens remains only partially understood. While we have good evidence that factors such as the destruction of infrastructure, political neglect, and the out-migration of health workers – what may be called supply-side factors – negatively affect health outcomes, we know much less about how violence shapes the attitudes and behavior towards healthcare use among civilians exposed to violent conflict – what may be called demand-side factors. Here, I theorize that exposure to violence suppresses civilian demand for healthcare through two mediating channels – mistrust of government institutions and fear of future violence – with adverse consequences for health outcomes, particularly child health. To test this theory empirically, I combine information from over 80,000 interviews conducted in 22 conflict-affected countries in Africa with individualand context-level measures of exposure to violent conflict. Exposure to violence is associated with significantly lower levels of political trust and increased fear of future violence, which in turn predict lower healthcare utilization, lower immunization rates, and higher infant and child mortality. To fully address the health consequences of armed conflict, it is essential that we better understand the attitudinal and behavioral correlates of exposure to violence. Introduction What are the indirect health effects of violent conflict? It is well known that wars kill and maim civilians years after the violence has ceased and in areas far from the battlefield (Ghobarah et al.,2003, 2004;Iqbal,2006;Gates et al.,2012;Kesternich et al.,2014;AkbulutYuksel,2017;Wagner et al.,2018). However, why exactly this is the case remains only partially understood. Scholars so far have tended to focus on structural factors and the physical challenges of providing healthcare in the midst of conflict: Destroyed health infrastructure means that populations are poorly served by essential medical services, hampering disease control and perinatal care (Iqbal and Zorn,2010; Kirschner and Finaret,2021). Landmines and unexploded ordnance continue to pose a threat, especially to children and agricultural workers, years after active fighting has ceased (Lin,2022). The exodus of health workers leaves conflict areas undersupplied with essential health services (McKay,1998;Chi et al.,2015;Chukwuma and EkhatorMobayode,2019). What these – clearly important – factors have in common is that they focus on the supply-side of health. In this paper, I ✩Max Schaub is Assistant Professor of Political Science, esp. International Relations and Global Health at the University of Hamburg, Allende-Platz 1, 20146 Hamburg, Germany. For comments and feedback, I would like to thank Carl M "uller-Crepon and seminar participants at EPSA 2023, the Hamburg Center for Health Economics, and the University of Hamburg. I am also grateful for the very constructive comments by two anonymous reviewers. Lennart Kasserra provided outstanding research assistance. This study received financial support under the University of Hamburg’s Close-the-Gap program, grant number 23-CG-2021. E-mail address: [email protected]. instead focus on the attitudes and behavior of the civilians affected by violence, and how these shape their demand for healthcare services. I argue that by inducing fear of future victimization and distrust of government institutions, exposure to violence leads to behavioral changes. People become reluctant to seek out medical services for themselves or their children, and to have their children vaccinated, with detrimental consequences for child health. Scholars have shown that the experience of violence systematically goes along with lowered levels of interpersonal trust and trust in state institutions (Rohner et al.,2013;Hager et al.,2019;Conzo and Salustri, 2019). At the same time, trust is an essential predictor for medical outcomes. Distrustful individuals are less likely to seek and follow medical advice, visit hospitals, and get their children vaccinated (Vinck et al.,2019;Stoop et al.,2021;Lin,2022). The argument put forward in this paper follows from the concatenation of these arguments: Violent conflict undermines trust and instills fear of future violence, leading to reduced use of health services and lower vaccination rates, which are associated with higher mortality, especially among the most vulnerable https://doi.org/10.1016/j.socscimed.2024.117252 Received 31 March 2024; Received in revised form 4 August 2024; Accepted 13 August 2024 Social Science & Medicine 359 (2024) 117252 Available online 2 September 2024 0277-9536/© 2024 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license ( http://creativecommons.org/licenses/by/4.0/ ). M. Schaub – young children. Exposure to violence is also known to induce trauma, often in the form of post-traumatic stress syndrome (PTSD), which has been linked with avoidance behavior (e.g., Obilom and Thacher, 2008;Neugebauer et al.,2009). Those exposed often shy away from socializing and public places, which, I argue, may make them reluctant to seek healthcare for their children, again resulting in elevated rates of infant and child mortality. To demonstrate this argument empirically, I use information from over 80,000 interviews from 22 countries in Africa that I analyze by a variety of analytical approaches, including structural equation modeling, fixed effect models, and mediation analysis. To establish that the relationship is plausibly causal, I estimate two-stage differences in differences models (Gardner et al.,2023). Individual-level data on experiences of violence, trust, fear, and healthcare-seeking behavior come from Afrobarometer (2018) surveys; context-level data on violence, vaccine uptake, and infant mortality from the Uppsala Conflict Data Program (UCDP) and the Institute for Health Metrics and Evaluation (IHME), respectively. The results show that exposure to violence – both at the individual and context-level – is associated with significantly reduced levels of trust and increased fear of violence, which then predicts lower rates of DPT (diphtheria-pertussis-tetanus) vaccination. Lower vaccine uptake, in turn, directly translates into higher rates of infant mortality. To the best of my knowledge, this is one of the first studies to focus specifically on the demand for healthcare among populations affected by violent conflict – and to directly link the experience of violence to child health through the channels of trust and fear of future victimization. The findings suggest that despite progress made in accounting for indirect effects of violent conflicts, we may still be underestimating the true cost of war. The findings also suggest that midand post-conflict interventions should include individual-level components that explicitly aim to restore trust and address trauma and fear. As will be demonstrated, these measures are not only important in their own right but will also help improve child welfare. 1. Literature and theory The detrimental consequences of violent conflict on the health of civilians not directly involved in the fighting have long been noted by scholars (McKay,1998;Ghobarah et al.,2003,2004;Plümper and Neumayer,2006;Iqbal,2006;Kesternich et al.,2014;Akbulut-Yuksel, 2017;Wagner et al.,2018,2019;Bendavid et al.,2021). A particular focus has been on the health of young children and women. Wagner et al. (2018) provide numerical estimates. Their study shows that between 1995 and 2015, the deaths of 5 million children under the age of 5 were directly or indirectly related to armed conflict. This was five times the total number of direct deaths from armed conflict during the same period. Most scholarship to-date has linked these staggering figures to the destruction of critical infrastructure and other physical legacies of violent conflict. The destruction or non-maintenance of sanitary infrastructure facilitates the spread of disease, which may then be carried to other areas by refugees (Iqbal,2006;Grundy and Biggs,2018; Hirschfeld et al.,2020;Kirschner and Finaret,2021). Destroyed roads complicate access to vital health services and the transport of food and medicine (Ghobarah et al.,2003;Plümper and Neumayer,2006; Kirschner and Finaret,2021). Other long-term dangers result from unexploded ordnance, including cluster sub-ammunition and landmines that pose a threat particularly to children and people working in agriculture (Lin,2022). Another well-studied phenomenon is the outmigration of healthcare workers (McKay,1998;Sharara and Kanj, 2014;Price and Bohara,2013;Chi et al.,2015). Worried about their own safety (and often in high demand in more peaceful regions) health workers may be the first to leave conflict-stricken areas. Direct attacks on health workers, which have become more common in recent years (Mahase,2022;Daniel,2023;Sauter,2024), only spur this process on. As shown by Chi et al. (2015), the absence of healthcare workers directly translates into worse supply of perinatal care and, plausibly, infant mortality. A characteristic shared by these important studies is that they focus on supply-side factors – factors related to the health-related infrastructure and the health services available to those who seek them. Violent conflict compromises these supply-side factors, often causing lasting damage to the health of civilians. A dimension that has seen much less attention is the effect of violent conflict on health-seeking behavior (see Adeyanju et al.,2024, for a review on this topic). In other words, we know very little about the extent to which the demand for healthcare is affected by violent conflict. This problem is exemplified by a study by Meierrieks and Schaub (2024), who estimate that each year around 40,000 children under 5 die as a consequence of terrorism in Africa. This is despite the fact that terrorist attacks rarely target children and do not cause the largescale physical destruction associated with other types of armed conflict. The authors argue that the negative health effects must therefore be due to behavioral changes. Related findings demonstrate changes in healthcare-seeking behavior in response to conflict but do not typically explore mechanisms such as reduced trust or fear (Price and Bohara 2013,Chi et al. 2015,Chukwuma and Ekhator-Mobayode 2019,Druetz et al. 2020; see Adeyanju et al. 2024 for a review). For example, Chukwuma and Ekhator-Mobayode (2019) show that women exposed to the Boko Haram insurgency in Nigeria access fewer perinatal services, arguably contributing to exceptionally high neonatal and infant mortality rates in the northeast of Nigeria, the epicenter of the insurgency. However, the reasons for this change in behavior remain unclear, leading the authors to call for an exploration of mechanisms linking exposure to violence to reduced healthcare-seeking. An important advance on this question is made by Tapsoba (2023), who suggests that what people respond to is the threat of conflict and the fear of it. Similarly, I here propose two mechanisms: a loss of trust in state institutions and fear of future violence. Mistrust in state institutions. While a rich literature shows how patterns of interpersonal trust change among victims of violence, the nature of the change remains contested (Bellows and Miguel,2009;Voors et al.,2012;Cassar et al.,2013;Rohner et al.,2013;Bauer et al., 2016;Hager et al.,2019). While one strand of the literature tends to describe a positive association between experiences of violence and interpersonal trust – a phenomenon explained with reference to so-called posttraumatic growth (Tedeschi and Calhoun,2004) – most recent works tend to find a negative association. Trauma, most scholarship indicates, undermines the belief that other people will not harm you if vulnerable. Alongside interpersonal trust, trust in state institutions is undermined. Research consistently shows that individuals who have been victimized have less trust in the state (Grosjean,2014;Voors and Bulte,2014;De Juan and Pierskalla,2016;Gates and Justesen,2020) – an effect that can be extremely lasting, sometimes enduring over several generations (Nunn and Wantchekon,2011;Conzo and Salustri,2019). The (implicit) explanation is that victimization informs individuals that the state failed to fulfill its role as protector, leading to a loss of trust. Importantly, the undermined trust seems to apply not only to security forces, the police, and the courts of law – i.e., institutions directly tasked with citizens’ security – but also extends to state institutions such as the national president and parliament, which bear no direct responsibility for individual episodes of violence. It is therefore at least plausible that mistrust in state institutions also extends to health services, such as government hospitals. Mistrust matters for health outcomes because it prevents both direct contact and belief updating. Distrustful individuals avoid contact with medical professionals, which means that they may not seek medical care for themselves or their children, or they may do so at a late stage when a disease has already progressed and the chances of successful treatment are diminished (Alsan and Wanamaker,2018). Distrust of Social Science & Medicine 359 (2024) 117252 2 M. Schaub government institutions may also mean that official messages become less effective because recipients stop trusting the messenger. Important for children’s health, distrust in government institutions can reduce preventive care, especially vaccination. Distrust of government is one of the strongest predictors of compliance with standard childhood vaccination recommendations (Stoop et al.,2021). Vaccination provides one of the best protections against many serious diseases that contribute to infant and child mortality. Therefore, the speed with which children are presented to health workers is important (Breiman et al.,2004). Vaccination is particularly important in resource-poor settings where other forms of healthcare are not always readily available (Jones et al., 2003). If people mistrust government institutions and officials, they may shy away from contacting health professionals to have their children vaccinated and/or stop listening to official government messages about the importance of vaccination. Mistrust can be persistent because it represents a stable equilibrium. As individuals do not seek help, they also have little opportunity to positively update their (negative) priors with regard to state-individual interactions – simply because they do not interact with the medical system at all. This is exemplified by research on the long-term health effects of the Tuskegee Syphilis Study, where Black sufferers of syphilis were left untreated despite the existence of a cure (Alsan and Wanamaker,2018). The authors show that Black men who learned about this abuse consequently avoided using the medical system and suffered worse medical outcomes than those whose trust was not compromised. Interestingly, the same effect was not found among women. The authors explain that most women come into contact with the medical system when giving birth. This typically positive experience gives them an opportunity to revise their negative prior about the medical system. As a result, their healthcare utilization is no different from women not exposed to the Tuskegee study. This example shows how external events can harm trust, and how this trust is only rebuilt if individuals are ‘forced’ into a situation where they can update their priors. In Nigeria, qualitative work has singled out trust in state institutions as an important factor for explaining vaccinationseeking behavior. In the early 2000s, a polio vaccination campaign came to an abrupt halt because rumors spread that the Nigerian health service and its international partners were trying to sterilize women by means of the injection. The rumors only stopped after religious authorities intervened (Obadare,2005;Jegede,2007). Similarly, during Ebola pandemics in central and western Africa, compliance with containment measures was precipitated by trust (Petherick,2015;Vinck et al.,2019). Medical anthropologists have pointed out that in some areas, compliance was particularly low because communities feared the medical personnel. In their minds, their behavior resembled that of armed groups during the recent civil wars in the region (Wilkinson and Fairhead,2017). Trauma and fear. A second channel potentially linking exposure to violence to healthcare-seeking behavior is trauma and fear. As demonstrated by numerous studies, victims of violence are regularly left traumatized by their experience (Jong et al.,2001;Bleich et al.,2003; Pham et al.,2004;Bayer et al.,2007;Neugebauer et al.,2009). Typical symptoms of post-traumatic stress syndrome include nightmares, flashbacks, difficulty concentrating, depression, and nervousness. The fact that traumatized people also regularly exhibit avoidance behavior is particularly relevant here. Victims usually avoid places linked to the traumatic event, or, more generally, avoid public places and social mingling. Authors have argued that, as a consequence, victims will under-utilize healthcare services (Nworah et al.,2014;Splinter et al.,2021). Depression and generalized anxiety, two other common correlates of experiences of violence, have also been shown to lead to unfavorable patterns of healthcare-seeking for children, such as delayed checkups and missed routine vaccinations (Minkovitz et al.,2005). Trauma due to violence can thus be expected to directly and negatively impact child welfare. Apart from these clinical effects, exposure to violence may also change perceptions of future threat. In work predicting conflict events, past violence is consistently one of the best predictors of future violence (Weidmann and Ward,2010;Hegre et al.,2021). This holds true down to the local level, meaning that areas that have experienced violence in the past face a higher risk of experiencing violence again, whether due to structural factors or social dynamics (Bazzi et al.,2022). In other words, individuals exposed to violence objectively face an elevated risk of becoming victims of future violence, and this objective shift is likely reflected in their subjective threat perception. Note that we would expect such changes in perception even without any psychological condition caused by exposure to traumatic events. Individuals will consequently tend to adjust their behavior to the newly perceived risk – avoiding larger groups of people, going out less, and limiting errands to only those strictly necessary (Malik et al.,2018;Sloan et al., 2021).1It is highly plausible that in this calculus, people will attach lower importance to preventive measures such as having their children vaccinated or presenting them for routine checkups – with potentially fatal outcomes in cases where children contract vaccine-preventable diseases or are presented too late to cure otherwise treatable conditions. My theoretical expectations can be summed up in the following hypotheses, which I will test using various data sources in the remainder of this paper: H1: Exposure to violence leads to reduced trust in state institutions and increased fear of future violence. H2: Reduced trust and increased fear lead to lower healthcare use, lower vaccination coverage, and higher infant and child mortality. 2. Data and methods The challenge with testing these hypotheses is that no single dataset exists that combines all required variables in one place. I therefore combine data from various sources, measured at both the individualand the context-level.2A disadvantage of this strategy is that for certain concepts, imperfect proxies have to be used. At the same time, a clear advantage is scale. The data sources used here rely on long-standing data collection efforts that cover a large geographic area – representing all major regions of Africa – and a long time span – from the early 2000s to around 2017. I use data from all 22 countries that were affected by violence according to UCDP/PRIO’s Armed Conflict Dataset (Gleditsch et al.,2002;Davies et al.,2023) – the most widely used resource for studying armed conflict around the world – while data collection for the Afrobarometer was ongoing. These countries include Algeria, Burkina Faso, Burundi, Cameroon, Cote d’Ivoire, Egypt, Guinea, Kenya, Liberia, Mali, Morocco, Mozambique, Niger, Nigeria, Senegal, Sierra Leone, South Africa, Sudan, Togo, Tunisia, Uganda, and Zimbabwe (see Fig. 1). While the degree of violence suffered by the different countries varies widely, I opted for this rather broad selection to make the findings more representative. This selection is also warranted because other conflict datasets, notably the Armed Conflict Location and Event Data Project (ACLED), another commonly used resource (which is used here for robustness checks), record rather high levels of exposure to violence in the selected countries, even when they were not in outright war. To further address concerns about country selection, I replicate core results for a more restrictive sample of countries that have seen 200 battle-related deaths per year – another fairly widely used threshold introduced by Weart (1998). This selection includes Algeria, Burundi, Cameroon, Egypt, Mali, Niger, Nigeria, Sudan, and Uganda. As can be seen in Table A9 in the Appendix, all of the core results hold for this sub-sample as well. 1Whether the calculus is actually correct, i.e., whether the behavioral adjustments are rational in the sense that they prevent more harm than they cause, is a question of debate (see, for example, the discussions in Becker and Rubinstein,2004;Meierrieks and Schaub,2024). 2Figure A1 in the Appendix provides a graphical overview of the data structure. Social Science & Medicine 359 (2024) 117252 3 M. Schaub Fig. 1. Map of conflict-affected countries included in the study. Individual-level data on demographics, trust, and fear of future violence comes from the Afrobarometer (2018), a large-scale data collection project that has been conducting nationally representative public attitude surveys on democracy, governance, the economy, and society in many African countries since 1999. I use data from rounds 3 to 7 of the geocoded version of the Afrobarometer, with geocoding provided by BenYishay et al. (2017). The main independent variable of interest is exposure to violence. Both a context-level and an individual-level indicator are used to capture this concept. The main measure for contextual exposure to violent conflict is derived from the UCDP’s Georeferenced Event Dataset (GED) (Sundberg and Melander,2013;Davies et al.,2023), which provides conflict-event data, including precise coordinates and the number of casualties. As with all context-level information, I aggregate the data at the 0.5◦×0.5◦(∼55 ×55 km at the equator) grid-year level using the PRIO-grid (Tollefsen et al.,2012).3Context-level exposure is then measured by a binary indicator that takes the value of one if an event was recorded in the UCDP GED dataset in a given year and grid cell, and zero if not. The dataset only records violent events belonging to armed conflicts resulting in at least 25 deaths per year, meaning that only fairly serious violent events tend to be recorded. As a robustness check, I replicate all analyses with the ACLED (Raleigh et al.,2010,2023) dataset. Similar to the UCDP GED, ACLED provides information on violent conflict events, their location and severity in terms of individuals killed, but does not require a threshold of severity for an event to be recorded. ACLED therefore tends to record a far higher number of events (as can be seen in the summary statistics in Table A1 in the Appendix), making it a more inclusive, but arguably less precise (cp. Eck,2012) measure of exposure to violence. Respondents are coded exposed to violence if ACLED recorded a violent event in the grid cell and year that the interview took place. Replicating all results using this indicator instead of the UCDP GED-based one gives very similar results, as shown in Section D of the Appendix. Individual-level exposure to violence is measured with the Afrobarometer survey item ‘‘During the past year, have you or anyone in your family been physically attacked?’’, to which respondents could respond ‘‘No’’, ‘‘Once’’, ‘‘Twice’’, or ‘‘Three or more times’’ (as with all items, they could also choose ‘‘Don’t know/Refuse’’). All non-missing values for this variable are used. It is important to note that we do not know whether or not the violence reported by a respondent is 3The combined dataset includes around 1700 distinct grid cells, and around 4000 grid cell years. related to the broader conflict captured by the UCDP GED (or ACLED) data; instead, the individual-level measure should be understood as an alternative measure of violent victimization. Of particular interest for this study are the hypothesized mediators ‘‘trust in the state’’ and ‘‘fear of future violence’’, and the intermediate outcome ‘‘health-seeking behavior’’. Measures for these variables are drawn from the Afrobarometer. With respect to trust in the state, scholars working on sub-Saharan Africa tend to emphasize the importance of trust in the president because most political regimes in the region are characterized by strong presidential governments (e.g. van Cranenburgh,2008;Azevedo-Harman,2011;Gates and Justesen, 2020). However, scholars working on political trust more generally typically opt for some form of index that combines measures of trust in different government institutions and offices (e.g. Mattes and Moreno, 2018). Often, these scholars further distinguish between political trust– trust in elected leaders such as the president or the ruling political partyand institutional trust–trust in nonpartisan ‘order’ institutions such as the police or the courts (e.g. Rothstein and Stolle,2008;Marien, 2011). Because of the importance of the executive in the African context, I here focus on political trust. In the Afrobarometer, respondents are asked, ‘‘How much do you trust the following?’’ followed by a list of state institutions, including the president, the ruling party, the courts, and the police. Respondents can choose ‘‘not at all’’, ‘‘just a little’’, ‘‘somewhat’’, or ‘‘trust a lot’’. Political trust is measured by averaging trust in the president and trust in the ruling party.4In Table A11 of the Appendix, I replicate the main analyses with an indicator for institutional trust, coded as the average of trust in the police and trust in the courts of law.5Unfortunately, the Afrobarometer does not include questions on trust in the medical system, which would be of great interest in the context of this paper. Fear of future violence is measured with an item asking respondents how often they or their family fear crime in their own home, to which they could respond ‘‘Never’’, ‘‘Just once or twice’’, ‘‘Several times’’, ‘‘Many times’’, or ‘‘Always’’. Here, the assumption is that the fear of crime encompasses the fear of being violently attacked, particularly in conflict-prone contexts, on which this study focuses. While not exactly the same as fear of victimization, this is the only measure available in the Afrobarometer. The imprecision introduced by this less-thanoptimal measure may explain the relatively weak explanatory power of the measure in some of the analyses below. Health-seeking behavior is captured by an Afrobarometer item inquiring whether respondents or their family members have ‘‘gone without medicines or medical treatment’’, to which they again could respond ‘‘Never’’, ‘‘Just once or twice’’, ‘‘Several times’’, ‘‘Many times’’, or ‘‘Always’’. I use this variable in its reversed form so that higher values indicate more frequent healthcare use. Apart from these variables of interest, all models also include a set of control variables taken from the Afrobarometer, notably a respondent’s age, sex, education level, and employment status. Ultimate outcomes of interest are infant mortality and vaccination rates. Infant mortality rates are from Burstein et al. (2019), who provide high-resolution estimates covering the whole of mainland Africa for the 2000–2017 period. Infant mortality is measured as the probability of a child dying before reaching the age of one. These estimates are based on national censuses and survey data, including the Demographic and Health Survey (DHS) and UNICEF Multiple Indicator Cluster Survey (MICS) household survey series, and are here used aggregated at the PRIO grid-cell-year level. As a robustness check, in Appendix Table A10 I replicate all core results using data on child mortality (also from Burstein et al. 2019), defined as the probability of a child dying before the age of five. 4These two dimensions of political trust have also been found to increase with improvements in health service delivery, see Chukwuma et al. (2019). 5The models show that while most relationships between the different variables are similar for both political and institutional trust, institutional trust does not predict vaccination rates, arguably due to the important messaging role of political leaders. Social Science & Medicine 359 (2024) 117252 4 M. Schaub Fig. 2. Trends in core variables over time, conditional on context-level exposure and individual victimization. Note: The figure shows trends over time for the indicated health-related outcomes and attitudinal and behavioral measures (a) in conflict-affected and non-affected areas, and (b) among victimized and non-victimized individuals. Context-level exposure to violence is measured as grid cells that experienced one or more conflict events in a given year, as recorded in the UCDP GED. Individual victimization is measured as respondents interviewed in a given year indicating that they or their family members were physically attacked. Infant mortality is measured as the probability that a child will not survive its first year of life. DPT coverage measures the proportion of under-5 year old who have received the recommended three doses of the diphtheria-pertussis-tetanus vaccine. Fear of violence measures respondents’ self-reported concern that they will experience crime of some form in their home. Political trust is an index that combines respondents’ self-reported trust in the president and trust in the ruling party. Frequency of healthcare use records the ease and frequency with which respondents could access healthcare services in the past year. Lines are polynomial smooths with a bandwidth of three; shaded areas are 95% confidence intervals. Social Science & Medicine 359 (2024) 117252 5 M. Schaub As the indicator for vaccination coverage, diphtheria-pertussis-tetanus (DPT) vaccine coverage is used, measured as the share of children under five who are fully vaccinated (i.e., who have received three shots or more). These estimates were produced by Mosser et al. (2019) using data from household surveys that included dose-specific information on DPT coverage.6As a different type of outcome and robustness check, I examine whether similar results as for vaccination can be reproduced for oral rehydration solution (ORS) use. ORS use is interesting because it represents a type of health-seeking behavior that is different from vaccination. However, similar to vaccination, it is a behavior that requires contact with and trust in health professionals. This is because although in principle ORS can be easily administered at home by caregivers, in practice, ‘‘the information regarding its adequate usage is restricted within the healthcare centers and professionals’’ (Aghsaeifard et al.,2022). The indicator comes from Wiens et al. (2020) and measures the share of children under 5 with diarrhea who received ORS. The results of this robustness check are reported in Appendix Table A12 and confirm the results presented below for vaccination coverage. Context-level control variables include factors that may influence both the onset of violent events and health outcomes. Along with longitude and latitude, these are: (i) population size, as provided by LandScan (Bright et al.,2018); (ii) nightlight intensity, a measure of economic development collected by the U.S. Air Force Defense Meteorological Satellite Program and made available for research by Elvidge et al. (2021), Ghosh et al. (2021); (iii) market access and state reach, measured as the travel time to the nearest large city and to regional and national capitals, respectively (Müller-Crepon,2021); (iv) female education, measured as the average years of education of 20-24-yearold women, from Graetz et al. (2018); and (v) climate extremes, as measured by the standardized precipitation evapotranspiration index (SPEI), from Peng et al. (2019). Apart from the geographic coordinates, context-level control variables are again aggregated at the grid-year level using the PRIO-grid. Since some of the variables are only available up to 2015, I extrapolated missing values for 2016 in order to not lose observations due to missingness in the control variables. Table A1 in the Appendix provides summary statistics for all individual-level and context-level variables, and Fig. 2 shows trends over time for the core variables of interest. Trends are displayed separately for areas exposed to conflict (that have seen a conflict event as defined by UCDP) vs. those that have not – Panel (a); and for individuals that have been victimized (physically attacked) vs. those that have not – Panel (b). We can see that in areas not directly exposed to violence, infant mortality rates have seen a steep decline over the years under observation. This trend does not hold in conflict-exposed areas, however. While these areas started out with initially lower rates of infant mortality (likely because conflict tends to concentrate in more urban areas, which typically also have lower infant mortality rates), infant mortality rates then stagnated. Given these trends, negative effects due to conflict exposure are best interpreted as the absence of improvement in infant mortality rather than absolute increases. However, this is still concerning because absolute figures remain extremely high (as shown in Table A1 in the Appendix): on average, 5.9 percent of children did not complete their first year of life, with figures as high as 12.9 percent in some grid cells. Even more disparate trends are visible with regard to vaccination rates. While these have tended to increase in non-conflict-affected grid cells, the trend is clearly negative in areas exposed to conflict, likely contributing to the stagnating infant mortality rates. Similar, albeit more complex, trends are evident in the attitudinal and behavioral measures. In terms of fear of future violence, while overall trends are downward sloping – arguably reflecting declines in 6These and other health-related, sub-national indicators can be downloaded from the website of the Institute for Health Metrics and Evaluation (IHME), see: https://www.healthdata.org. violent crime unrelated to armed conflict (van Dijk et al.,2022) – context-level exposure to violence is consistently associated with higher levels of fear. No particular trends are visible for individual victimization, though, which is consistently associated with higher levels of fear. As for political trust, again we see that trust is clearly lower among respondents exposed to violence or victimized. The gap in trust between those exposed to violence at the context level and those not exposed seems to be narrowing over time. It is possible that countries in Africa and the international community have become better at dealing with some of the consequences of violent conflict, thereby reducing the negative impact of exposure to violence on political trust. However, no such trend is evident when looking at individual victimization, where differences in trust levels are large and constant over time. A somewhat paradoxical trend is observed with respect to access to healthcare. In areas exposed to violence, the proportion of people not using healthcare services decreased over time and is lower than in areas not affected by violence. I can only speculate as to why this is the case. One possibility is that access has improved. As pointed out in a review article on health-seeking behavior in the midst of conflict (Adeyanju et al.,2024), access to healthcare sometimes improves in conflict settings due to the humanitarian response of national and international NGOs. However, there is no such positive trend when we look at individual victimization. In all years, victims of violence are significantly less likely to have interacted with medical professionals than non-victims, and the discrepancy remains highly constant over time. It is possible that access to and use of healthcare is improving for some in conflict-affected areas, but not for direct victims of violence. More importantly, the trends may be confounded by third variables and thus may not truly capture the effect of violence on the reported outcome. For this reason, we now turn to more rigorous analyses. 3. Analysis Structural equation models. I begin the analysis by investigating how the different variables just introduced relate to each other using structural equation models. The models, graphically presented in Fig. 3, show all relationships that are both theoretically interesting and statically significant. Again, I show results for both contextand individuallevel exposure to violence. Effects manifest themselves at the individual level (‘fear of violence’, ‘political trust ’, and ‘healthcare use’) and the context level (‘DPT-3 coverage’, and ‘infant mortality’). Several observations stand out: As hypothesized, there are statistically significant correlations between exposure to violence, fear of violence, and lower trust in state institutions. Higher political trust positively predicts healthcare use, while increased fear of victimization has the opposite effect. Contrary to what was hypothesized, there is no significant relationship between fear of violence and vaccination rates (arrow omitted). However, again in line with the theory, vaccination rates are strongly related to political mistrust: lower trust in state leaders predicts lower vaccination rates. Healthcare use and vaccination rates, in turn, are strongly negatively associated with higher infant mortality. Two-way fixed effects and multilevel models. While the analyses provide supportive evidence for the hypothesized relationship, the models in Fig. 3 shows mere correlations without taking into account potential confounding factors. I therefore proceed by presenting regression models, all of which include the full set of control variables introduced above. I estimate the individual-context and the context-context links using multilevel models (allowing slopes to vary at the grid-level), and the individual-individual links using yearand grid-cell fixed effects models. The latter specification is particularly stringent because it compares individuals who were interviewed within the same local area, hence limiting the potential confounding effect of context. The inclusion of year fixed effects ensures that time trends cannot influence the estimates. Results are shown in Table 1. Exposure to violence at the grid cell level is associated with a 0.14 point decrease in the political trust scale, and a 0.22 point increase in Social Science & Medicine 359 (2024) 117252 6 M. Schaub Fig. 3. Structural equation models showing relationships between core variables. Note: Structural equation models relating (a) context-level exposure to violence, and (b) individual-level victimization to attitudinal and behavioral measures and structural outcomes. Estimated using maximum likelihood estimation, standard errors clustered at the grid-cell level, ∗𝑝 < 0.1,∗∗ 𝑝 < 0.05,∗∗∗ 𝑝 < 0.01. Results in tabular format are shown in Table A4 in the Appendix. Table 1 Two-way-fixed-effects (TWFE) and multilevel models. TWFE models Multilevel models (1) (2) (3) (4) (5) (6) (7) (8) (9) Pol. trust Fear Pol. trust Fear Healthc. Healthc. DPT-3 cov. Inf. mort. Inf. mort. Context. viol. exp. −0.143∗∗ 0.223∗∗∗ (0.056) (0.055) Ind. victimization −0.091∗∗∗ 0.482∗∗∗ (0.008) (0.011) Fear of violence −0.337∗∗∗ (0.013) Political trust 0.298∗∗∗ 0.124∗∗∗ (0.017) (0.032) Freq of healthcare use −0.062∗∗∗ (0.009) DPT-3 coverage −0.644∗∗ (0.290) N observations 88,562 88,770 87,359 88,627 88,411 88,200 88,563 89,744 90,136 N fixed/random effects 1,785 1,783 1,782 1,783 1,782 1,784 1,786 1,786 1,786 Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Note: Two-way-fixed-effects (TWFE) and multilevel models relating the indicated variables of interest to each other. TWFE models estimated using OLS; multilevel models estimated using maximum likelihood estimations. Observations vary due to missing values in individual dependent or control variables. The abbreviated column and row titles translate as follows: Pol. trust∼Political trust; Fear∼Fear of violence; Healthc.∼Frequency of healthcare use; DPT3-cov.∼DPT-3 coverage; Inf. mort.∼Infant mortality, probability of a child dying before the age of 1; Context. viol. exp.∼Exposure to violence at the context-level; Ind. victimization∼Individual-level victimization of the respondent or their family. Standard errors clustered at the grid-cell level in parentheses, ∗𝑝 < 0.1,∗∗ 𝑝 < 0.05,∗∗∗ 𝑝 < 0.01. Full regression output shown in Table A3 in the Appendix. the measure for fear of violence. Similar effects are seen for individuallevel victimization. Victimization predicts a 0.09 point decrease in political trust and a 0.48 point increase in fear of future violence. Fear, in turn, is associated with a 0.34 point decrease in healthcare use, while a 1 point increase in political trust predicts a 0.3 point increase in healthcare use. The multilevel models (Models 7 through 9) show that the known relationships between trust and vaccination rates, between the use of medical services and infant mortality, and between vaccination rates and infant mortality all hold. Taken together, the results thus confirm the robustness of the correlational findings above. Two-stage differences in differences. Two-way fixed effects have faced criticism for not clearly identifying the proper counterfactual of treated units due to treatment effect heterogeneity (Sun and Abraham,2021; Imai and Kim,2021). To circumvent this problem, I adopt the two-stage differences -in-differences approach suggested by Gardner et al. (2023) and implemented by Butts and Gardner (2022). This approach is only applicable to the regressions probing for the effect of contextual-level exposure to violence (i.e., Models 1 and 2), where a clear time-varying treatment – being affected by a UCDP-recorded conflict event – can be identified. The idea of Gardner et al. (2023) is to first regress the dependent variable of interest on the unit and time fixed effects plus Table 2 Two-stage differences-in-differences model. (1) (2) Pol. trust Fear Context. viol. exp. −0.013∗∗∗ 0.029∗∗∗ (0.004) (0.005) N observations 88,563 88,770 Controls Yes Yes N fixed effects 1,785 1,783 Note: Two-stage differences-in-differences model (Gardner et al.,2023) for political trust and fear on contextual exposure to violence. The abbreviated column and row titles translate as follows: Pol. trust∼Political trust; Fear∼Fear of violence; Context. viol. exp.∼Exposure to violence at the context-level. Standard errors clustered at the grid-cell level in parentheses, ∗𝑝 < 0.1,∗∗ 𝑝 < 0.05,∗∗∗ 𝑝 < 0.01. controls using only not previously treated units. In a second step, the predicted residuals from this first regression are then regressed on the treatment indicator. In this way, the average treatment effect on the treated can be identified in a consistent manner. Because the analysis is performed on the residuals only, the method tends to produce more conservative estimates. The results shown in Table 2 indicate that the findings are robust to this procedure. While the estimated effects of contextual exposure to violence are much reduced in size, they Social Science & Medicine 359 (2024) 117252 7 M. Schaub Table 3 Meditation analysis. A. Context. exp. viol. →Healthcare use Mediator: Fear Pol. trust Natural indirect effect −0.021∗∗∗ −0.012∗∗∗ (0.006) (0.004) Total effect −0.031 −0.033 (0.040) (0.040) Proportion mediated 0.68 0.35 N 86,902 86,902 Controls Yes Yes B. Ind. victimization →Healthcare use Mediator: Fear Pol. trust Natural indirect effect −0.083∗∗∗ −0.012∗∗∗ (0.008) (0.002) Total effect −0.329∗∗∗ −0.327∗∗∗ (0.015) (0.015) Proportion mediated 0.25 0.04 N 86,773 86,773 Controls Yes Yes Note: Mediation analysis investigating the extent to which the relationship between exposure to violence at the context- (A) and individual-level (B) is mediated by fear of violence and institutional (mis-)trust. The abbreviated column and row titles translate as follows: Fear∼Fear of violence; Pol. trust∼Political trust; Context. viol. exp.∼Exposure to violence at the context-level; Ind. victimization∼Individual-level victimization of the respondent or their family. OLS regression on simulated potential outcomes using Stata’s causal mediation package; standard errors clustered at the grid-cell level in parentheses, ∗𝑝 < 0.1,∗∗ 𝑝 < 0.05,∗∗∗ 𝑝 < 0.01. remain highly statistically significant, increasing our confidence that the relationship we observe is indeed causal. Mediation analysis. How much is the effect of exposure to violence on healthcare-seeking behavior mediated by political mistrust and fear of violence? To answer this question, I conduct a mediation analysis. I adopt the causal mediation framework proposed by Imai et al. (2010), which also provides methods for assessing the sequential ignorability assumption in this type of analysis. In cases such as the present one, where the total effect is hypothesized to run through several mediators, it is recommended to estimate mediation effects in the same model, or to control for other potential mediators (VanderWeele and Vansteelandt,2014). Alongside the standard set of control variables, I therefore include the non-considered mediator as a control variable, plus the interaction term between the exposure variable and the mediator. As can be seen in Table 3, the reduction in healthcare seeking behavior due to the exposure to violence is strongly mediated by the fear of violence. This is particularly true for contextual exposure to violence, where an estimated 68% of the effect is mediated by fear of future violence, compared to 25% for individual victimization. In contrast, the mediating role of political mistrust is less pronounced. For contextual exposure, an estimated 35% of the effect is mediated through lowered political trust, while for individual exposure it is only 4%. The mediation analysis thus provides clear support for the theory that exposure to violence drives down healthcare-seeking behavior through the channels of fear of future violence, and, to a lesser extent, mistrust in the political leadership. However, it can also be seen that only part of the relationship between violence and healthcare-seeking is explained by the proposed mediators. This means that there must be other factors mediating the effect of exposure to violence that remain unaccounted for. In the theory section, I suggested that one such mediator might be psychological trauma, which could not be tested here because there was no appropriate measure in the data. Exploration of this and other mediators will have to be left to future research. Conclusion This paper establishes a relationship between exposure to violence and reduced health-seeking behavior. This relationship is mediated by political mistrust and the fear of future violence. The knock-on effects are lowered vaccination rates, reduced healthcare use, and, ultimately, higher levels of infant and child mortality. The study combines a large number of data sources from 22 conflict-affected countries in Africa, observed over a time period of 12 years. This approach is both a strength and a weakness. On the one hand, the breadth of the study increases our confidence that the observed effects are quite general; their applicability beyond the African continent is however a question for further research. At the same time, several concepts could not be optimally measured. This is certainly true of trust in political and state institutions, which could only be operationalized in a general fashion. Future research should explicitly test for trust in health-related messages, and should also measure trust in the medical system and health workers. Moreover, rather than relying on context-level indicators of infant mortality and vaccine coverage, it would be preferable to measure those outcomes within the same instrument where trust and fear of violence are assessed. Such research should also explore additional channels, such as psychological trauma. Conducting such unified research would be an important area for future exploration. Another possible extension would be to rigorously investigate the longterm effects of the proposed channels. Future studies could explore how lasting the effects of violence on trust and fear are, and whether adverse health effects of these factors remain visible in the long run. This paper contributes to the literature on the health effects of violence by highlighting the largely overlooked demand-side effects of violent conflict on health. Exposure to violence changes the attitudes and behaviors of civilians, with negative consequences for health outcomes, particularly child health. These findings should be seen as complementary to, rather than challenging, well-established pathways, such as the destruction of critical infrastructure or the out-migration of health workers, which previous research has shown to severely and negatively affect health outcomes in conflict-affected regions. Future work could explore interactions between supplyand demand-side factors. For example, it is possible that the decline in the quality of healthcare due to the out-migration of health professionals from conflict-affected regions negatively affects trust in the health system, which in turn affects healthcare-seeking behavior in the long run. While additional research is necessary before definitive conclusions can be drawn, the findings presented here seem clearly relevant for policy. They suggest that in order to mitigate the detrimental health effects of violence, attitudinal and behavioral consequences must be explicitly addressed. Adjusting existing programs could be part of this effort. For example, trust-building measures are already a well-established component of post-conflict reconstruction and peacebuilding programs (Wong,2016;Svensson and Brounéus,2013). These programs could be modified so to consider health as an explicit area of activity. While trust-building measures would likely be one component, the provision of ‘hard’ security also matters. As shown, the fear of victimization is a major driver of detrimental adjustments in health-seeking behavior. Measures that contribute to providing physical security in conflict and post-conflict settings – be they in the realm of security-sector reform, or in the realm of welfare policy – can therefore also likely benefit population health and reduce child mortality. CRediT authorship contribution statement Max Schaub: Writing – original draft, Visualization, Validation, Supervision, Software, Resources, Project administration, Methodology, Investigation, Funding acquisition, Formal analysis, Data curation, Conceptualization. Social Science & Medicine 359 (2024) 117252 8