The impact of armed conflict on cancer among civilian populations in low- and middle-income countries: a systematic review

Background Armed conflicts are increasingly impacting countries with a high burden of cancer. The aim of this study is to systematically review the literature on the impact of armed conflict on cancer in low- and middle-income countries (LMICs). Methods In November 2019, we searched five medical databases (Embase, Medline, Global Health, PsychINFO and the Web of Science) without date, language or study design restrictions. We included studies assessing the association between armed conflict and any cancer among civilian populations in LMICs. We systematically re-analysed the data from original studies and assessed quality using the Newcastle-Ottawa Scale. Data were analysed descriptively by cancer site. Results Of 1,543 citations screened, we included 20 studies assessing 8 armed conflicts and 13 site-specific cancers (total study population: 70,172). Two-thirds of the studies were of low methodological quality (score <5) and their findings were often conflicting. However, among outcomes assessed by three or more studies, we found some evidence that armed conflict was associated with increases in the incidence and mortality of non-specific cancers, breast cancer and cervical cancer. Single studies reported a positive association between armed conflict and the incidence of stomach and testicular cancers, some as early as 3 years after the onset of conflict. Some studies reported a post-conflict impact on time to diagnosis. Conclusion Our findings support the need for more rigorous longitudinal and cohort studies of populations in and immediately post-conflict to inform the development of basic packages of cancer services, and post-conflict cancer control planning and development.


Introduction
(e.g., hospital admission data pre-, during-and post-conflict), we excluded studies whose first post-conflict data point was greater than 3 years after the end of the conflict.
We excluded studies reporting on military veterans, combatants and studies from high-income countries (including where refugees had migrated to high-income countries). We also excluded studies whose exposure was weapons (often, nuclear) testing rather than armed conflict. Studies that mentioned armed conflict but did not attempt to measure it were further excluded.

Data analysis
Two reviewers performed all citation screening and data abstraction in duplicate and independently using pilot-tested forms. Disagreements were resolved by discussion, and when needed with the help of a third reviewer. We retrieved full texts of citations considered eligible by at least one reviewer. Data extracted from eligible studies included study provenance (funding source, ethics approval and conflicts of interest), study features (design, timing, conflict, country and level of jurisdiction), population (sample size, mean age/age range and percentage of males) and results (outcome measure definition, outcome measure effect size and precision). We calculated the maximum number of years from the onset or end of conflict to the time of data collection, to give an indication of the length of armed conflict exposure. We used the Newcastle-Ottawa Scale (NOS) [22][23][24] to assess the quality of each study. The NOS has been recommended for use for non-randomised studies by the Cochrane Collaboration [25]. Although the NOS has no established threshold of quality, in line with previous reviews [26,27], we defined studies as low quality (score <5), moderate quality (score 5-6) and high quality (score >6) to simplify the main analysis. Quality scores by NOS domains (selection, comparability and outcome) for each study are detailed in Table S2.
Meta-analysis was not feasible given the degree of between-study heterogeneity in design, armed conflict, population and outcome. We, therefore, analysed data descriptively. To standardise our analytical approach and to reduce bias, we systematically re-analysed reported data and presented a single effect estimate per outcome per study where possible. This included constructing 95% confidence intervals around all effect estimates and considering confidence intervals that did not overlap as statistically significant at an alpha level of 0.05. This also meant we combined outcomes stratified by population subgroups (e.g., by age and sex), and used the overall outcome in our analysis. We did not reanalyse data already presented as odds ratios, beta-coefficients or hazard ratios. Where data were available pre-during-and postconflict, we used a single estimate for the differences between the pre-versus during-conflict data for each study. Furthermore, an analysis of post-conflict data was undertaken separately to understand better changes in trends throughout the conflict cycle. Each outcome from each study was assigned a qualitative effect direction (increase, decrease or no change) following exposure to armed conflict based on the statistical significance of effects. We stratified our analysis by cancer incidence and mortality, and outcomes with greater than three studies were described in more detail and displayed graphically using Harvest plots. Harvest plots take aspects of a forest plot to display data on a matrix of effect direction weighted by several variables [28]. Finally, we visually assessed publication bias by constructing an adapted funnel plot, using the sample size and the qualitative effect direction in place of the standard error and effect size, respectively.

Study characteristics
Of 1,543 records identified through database searching, 38 were potentially eligible and 20 were included in the final analysis ( Figure 1). The total study population was 70,172. Three-quarters of studies used an ecological design (75.0%) and over one-third analysed the Croatian War of Independence (1991-1995) (35.0%). Over half were conducted in cities (55.0%) and 70.0% utilised hospital-derived data. The average follow-up time was 16.8 years (range 3-64 years) and study quality was mostly rated as low (65.0%). Only four outcomes were assessed by three or more studies: the incidence of any, breast and cervical cancer, and mortality from any cancer. of time. Two hospital-based studies from the Balkans showed no change in cancer incidence during the conflict compared to the pre-conflict baseline [30,31]. Another cancer registry study assessed the Lebanese Civil War and showed no change in cancer incidence during the conflict period (1983)(1984)(1985)(1986)(1987)(1988)(1989)(1990)(1991), mean 786 cases/year) compared to the post-conflict period (1992 to 1994, mean 802. 3

Mortality from any cancer
Four studies assessed mortality from any cancer (Figure 2, bottom left panel). One moderate-to-high quality study assessed the 2003 US-led invasion of Iraq and reported an average 50% increase in the number of households reporting cancer deaths from the pre-conflict period (mean 9.9 cases/year in [2001][2002] to the conflict period (mean 14.8 cases/year in 2003-2010) [33]. We calculated this difference to be statistically significant (4.9 cases/year, 95% CI 0. 4-9.4). Two survivor cohort studies from the Siege of Leningrad (1941Leningrad ( -1944 reported no change in cancer mortality 41 to 64 years after the siege although both adjusted hazard ratios showed positive effect estimates (1.12 (95% CI 0.95 -1.31) and 1.11 (95% CI 0.97 -1.27)) [34,35]. One modelling study (1973 to 1994) used data from the Federal Institute of Statistics to assess the impact of the breakup of Yugoslavia, and found that cancer mortality decreased during periods of war and sanctions [36].

Breast cancer incidence
Six studies, all assessing wars in the Balkans during the 1990s, reported on breast cancer incidence ( Figure 2, top right panel). Both moderateto-high quality studies showed an increase in breast cancer incidence [37,38]. One of these was ecological in design, monitored trends 13 years before the 1999 NATO bombing of Yugoslavia, and reported an increase from an average of 67.2 cases/year before the conflict to 80.2 cases/ year during the conflict [38]. We calculated this difference to be statistically significant (13.0 cases/year, 95% CI 4.1-21.9). The other study used a case-control design and reported increased odds of breast cancer among those with greater exposure to war-related events in Bosnia (pooled odds across all events: 1.55, 95% CI 1.37-1.73) [37]. The remaining four studies, all low quality and ecological in design, showed no change [39,40] or a decrease [31,41] in breast cancer incidence. The study with the shortest follow-up in this review (3 years) was one study that showed a decrease in breast cancer diagnosis during the Croatian War of Independence (32 cases in 2 years) compared to the pre-conflict baseline (86 cases in 2 years) [31]. We considered this decrease statistically significant (−54.0 cases/2 years, 95% CI-75.3 to −32.7).

Cervical cancer incidence
Three studies assessed cervical cancer incidence (Figure 2, bottom right panel). One moderate-to-high quality case-control study of the Vietnam War showed that women with a husband in the army had higher odds of cervical cancer compared to those without (adjusted odds ratio (AOR) 1.32, 95% CI: 1.00-1.75) [42]. One low-quality ecological study in Greece assessed over 35,000 smear tests from hospitals with different proximity to the Yugoslav border, but showed no difference in either cervical cancer or cervical intraepithelial neoplasia incidence between the sites following the NATO bombing of Yugoslavia in 1999 [43]. Another low-quality hospital-based ecological study found a decrease in cervical cancer incidence, from 214 cases in 6 years before the Croatian war, to 142 in 6 years of the war [44]. We found this to be a statistically significant decrease (−72.0, 95% CI: −109.0 to −35.0). Figure 2. The impact of armed conflict on cancer incidence and mortality. Interpretation: Height refers to study quality, colour refers to armed conflict, number refers to length of follow-up between conflict exposure and outcome, bars grouped as showing either an increase, decrease, or no change following exposure to armed conflict.

Other cancers
Eight studies examined other site-specific cancers, but they were too few to display graphically and describe collectively. One hospital-based study from Croatia reported a rise in the incidence of malignant stomach and testicular cancers when comparing 2 years of conflict to 2 years prior [31]. Other studies of various study design and quality found no association between armed conflict and mortality from breast cancer

Post-conflict trends
All seven studies that assessed the conflict cycle (i.e., pre-conflict, conflict and post-conflict) were ecological, hospital-based studies analysing either the Croatian or Bosnian wars of the 1990s [30,39,41,[44][45][46][47]. The three studies that reported no change between the times before and during the conflict then showed an increase in incidence in the post-conflict period [30,39,44]. The one study that reported an increase in incidence between the pre-and during-conflict periods found that this increase was sustained into the post-conflict period [47]).
In the three studies that reported a decrease in incidence between the pre-and during-conflict periods found that this either plateaued [41,46] or returned to pre-conflict levels [44] during the post-conflict period. One ecological study showed mixed findings in the incidence of haematological cancers depending on the type of conflict exposure used (areas affected by depleted uranium, chemical damage or population mixing) and outcome (Hodgkin's lymphoma, non-Hodgkin's lymphoma, lymphatic leukaemia and myeloid leukaemia), but generally found either no change or a decrease in incidence through the post-conflict period [45]. Figure 3 presents an adapted funnel plot to assess publication bias, which includes all 55 outcomes from the 20 included studies. While the absence of actual effect estimates limits interpretation, the plot does not present convincing evidence of asymmetry or the absence of small studies showing no effect, which are indicative of publication bias.

Discussion
The literature on the impact of armed conflict on cancer incidence and mortality is very sparse, methodologically poor, and often contradictory. This is despite the fact that some have extensive follow-up periods, which averaged 18 years. The main limitations to many studies were their design, namely, ecological, and thus subject to ecological fallacies; nearly all failed to acknowledge this, in addition to failing to account for sudden population demographic changes following forced migration. There was also limited adjustment for confounding variables in risk factor exposure and behaviour changes. The lack of data on factors, which may mediate the impact of armed conflict on cancer, is an additional serious limitation in the extant literature.
The one cancer (breast) that did have several studies showing an increase in incidence following armed conflict did not have, however, sufficient data to advance understanding of plausible aetiological factors. Armed conflict has been shown to change reproductive strategies in populations affected with greater parity and lower maternal age, both of which are protective of breast cancer [49]. Thus it is unclear, whether the increased incidence of breast cancer is real or an artefact.
The factors that affect cancer incidence and mortality in armed conflict are multifactorial and multilevel; these includes changes to risk factor exposure, behavioural changes, delays to presentation, the availability of timely and affordable complex care (depending on the site-specific cancer), the ability to access care, etc. Furthermore, the ability to collect reliable data from registries, hospitals or camps can be substantially hampered during periods of conflict. In some cases, this is because systems are destroyed, data are not collected (too costly or to protect patients identities) or because care data are fragmented across multiple disconnected places of care [50,51]. Reported data may be inaccurate due to limited diagnostic facilities and available pathologists, so any statistical inference should provide a contextual interrogation to the quality of the data. Reduced case ascertainment featured prominently as a serious lacunae in data collected during the Lebanese Civil War (1975)(1976)(1977)(1978)(1979)(1980)(1981)(1982)(1983)(1984)(1985)(1986)(1987)(1988)(1989)(1990)(1991), when the American University of Beirut Medical Center (AUBMC) was the only functioning cancer referral site in the entire county and it was estimated at least two-thirds of the cancer burden during this period went either undiagnosed or unreported [32]. AUBMC and other cancer centres only become accessible after the end of the conflict [32], so any increase in incidence during the post-conflict period may simply reflect a return of the status quo. A similar conclusion was reached in analysing the cancer incidence data collected during the Croatian War of Independence; road blockades across the country and the removal of free care services such as breast cancer checkups radically reduced health service accessibility [40]. In another analysis of the same conflict, an observed post-conflict increase in cancer incidence was also attributed to the introduction of a new cancer screening programme, better organisation of cancer care services and the introduction of more accurate and up-to-date diagnostic equipment in hospitals [39].
In armed conflict, there is an expected rise in cancer-related mortality due to the loss of skilled personnel, the shift of such personnel into acute care, shortage or failure of key equipment-diagnostic imaging, surgical instruments, radiotherapy and cancer drugs, for example-and the inability of patients to access what care remains due to security or affordability barriers, all factors that led to the rise in cancer mortality during the armed conflict in Serbia in 1999 [38]. Yet it is possible that the same factors that worsen cancer mortality are the same that inhibit the timely and accurate reporting of such mortality, which may explain why many of the studies included in this review reported no change in the incidence or mortality of cancer during or after armed conflict.
Better quality research to study cancer in armed conflict is essential, and our review findings have several research implications. Although resources are often scarce in conflict settings, making use of hospital-based registries or other sources of routinely collected data have excellent potential for robust inquiry. In instances where control groups are not feasible, data could be subject to interrupted time series or difference-in-difference analyses with adjustment for confounders or with age-/sex-standardised rates of cancer incidence. Importantly, researchers should outline the status of screening programmes and other mediators in the relationship between armed conflict and cancer, so that these can be appropriately accounted for in the study design. This will make a more informative contribution to the current literature which is lacking in methodological rigour and often reports crude numbers over time. One notable absence from the literature was studies from humanitarian organisations. Although often unable to collect pre-conflict data, they are in a strong position to assess the degree of conflict exposure among their patients using tools such as the Harvard Trauma Questionnaire [52]. Future research could assess the impact of armed conflict on stage of diagnosis, in addition to inequalities by socioeconomic groups (e.g. age, sex, residence and deprivation). Most studies with very long follow-up times (>30 years) hypothesised that in utero, infant or adolescent exposure to armed conflict would have a greater impact on cancer risk to those exposed at older ages [34,35,53]. However, the failure to properly control for the many confounders has seriously hampered research to examine the link between toxic contamination of the environment due to armed conflict and long-term health impacts such as cancer.
Our findings also have important policy implications. Despite a number of guidance documents on cancer care in complex emergencies and post disaster, e.g., post typhoon Haiyan issued by WHO [54, 55] the literature is silent on what might constitute basic packages of cancer care, for UN and international NGOs for example and on approaches to post-conflict cancer systems reconstruction, or in supporting host countries absorb and provide care to refugees in both formal and informal (sans papier) settings. Although, it is to be recognised that the latter is intimately linked to post-conflict health systems reconstruction per se. More research is needed to urgently inform cancer policies and planning in the context of armed conflicts, particularly now that so many are occurring in high-burden countries with populations that have gone through the demographic and epidemiological transitions.