Menu
Menu
Menu
Menu
Refugees may be perceived as a burden to their host communities, and nutrition insecurity is a critical area of contention. We explored the relationship between refugee presence and a host community’s resilience in nutrition outcomes in Cameroon. We also tested an analytical framework for evaluating community resilience during shocks. We used data from repeated cross-sectional Demographic and Health Surveys in Cameroon (2004 and 2011), data on refugee movement, and data on extreme climatic events, epidemics, and conflicts from multiple sources. Outcome variables were maternal underweight, maternal anaemia, and child underweight, anaemia, stunting and wasting. The exposure variable was residence within an area in which refugees settled. We used a genetic matching algorithm to select controls from the rest of the country after excluding areas experiencing concurrent shocks. We used a difference-in-differences analysis to compare outcomes between the exposed and control areas. The 2004 survey comprised 10,656 women and 8,125 children, while the 2011 survey comprised 15,426 women and 11,732 children. Apart from anaemia which showed a decreasing trend in both the refugee-hosting community and the rest of the country, all other indicators (wasting, underweight and stunting) showed increasing trends in the refugee-hosting community but decreasing trends in the rest of the country. The matched control group showed a similar trend of decreasing trend for all the indicators. Controlled comparisons showed no evidence of an association between changes in nutrition outcomes and the presence of refugees. These findings contest a common perception that refugees negatively impact hosting communities. The difference-in-differences analysis and an improved matching technique offer a method for exploring the resilience of communities to shocks.
We undertook a secondary data analysis using repeated cross-sectional Demography and Health Surveys (DHS) data (data collected from a new sample of participants at successive time points). We also used data on refugee movement, extreme climatic events, epidemics, and conflicts from multiple sources to examine if residents of a refugee-hosting community in Cameroon experienced changes in nutrition status attributable to refugees. Between 2004 and 2011, Cameroon experienced a rapid inflow of refugees fleeing conflict from the neighbouring Central African Republic. These refugees settled mainly in a geographically circumscribed area in the Eastern region of the country. Using households in the refugee hosting community as the exposure group, we matched households with similar characteristics from the rest of the country as the control group. We used genetic matching algorithms embedded in the Matching R package to select controls with the optimal balance of covariates. Although Cameroon is diverse in terms of culture and climate, we considered our control group appropriate because the central health and development policies apply equally over the national territory. We further undertook an extensive systematic analysis to identify and exclude potential control areas that experienced other shocks, such as political unrest, epidemics, droughts, and floods in the country over the same period (see Supplementary Table S1). We established an exposure and a control group for each study year (2004 and 2011), with 2004 representing the period immediately before the mass migration of refugees and 2011 representing the time after. We conducted a difference-in-differences analysis on nutrition outcomes between the years to estimate the impact of refugees on nutrition outcomes in the refugee-hosting community. We reported this study following the STROBE (STrengthening the Reporting of OBservational studies in Epidemiology) guidelines [14]. Cameroon is a lower-middle-income country in Central Africa, often described as Africa in miniature because of its diverse climate, geography, culture, and politics. The country experiences various shocks, such as intermittent floods, droughts, and epidemic outbreaks, which increase vulnerability to nutrition insecurity, compounding the effects of the low-income status. Until recently, Cameroon has been a peaceful country and has served as a haven for refugees fleeing threats and violence from troubled neighbouring countries. Until 2004, 57000 refugees had settled in three towns in Cameroon (Banyo, Yaoundé and Douala) [15]; the number of settled refugees was relatively stable for a long time before 2004. Nine thousand rural refugees (mainly in Banyo) were from Nigeria. Most of the 49000 urban refugees (mainly in Douala and Yaoundé) were from Chad, the Central African Republic, DRC, Liberia, Rwanda, and Sierra Leone. Since 2006, refugees fleeing the escalating conflict in the Central African Republic have settled in communities on the eastern borders of Cameroon. By 2011, the country’s total number of refugees doubled the 2004 estimate (Figure 1) [16]. The refugee settlement area was largely rural, with few small towns. The small cities rapidly grew as the refugees settled and economic activities increased. While some refugees lived in camps, most lived within the community; they had access to over 2000 ha of arable land offered by the government and local authorities. Support for the refugees was mainly from the Cameroonian government, the United Nations High Commissioner for Refugees (UNHCR) and other NGOs [17]. Based on the 2011 UNHCR report [16], most refugees settled in a well-defined geographical area; their interaction seems to have been with a distinct group. Therefore, we can estimate the impact of their interaction with the host community. Figure 2 shows the UNHCR geolocation of refugee settlement sites in Cameroon in 2011. The trend of refugees in Cameroon (data source: www.macrotrends.net). Refugee settlement in Cameroon (extracted from the United Nations High Commissioner for Refugees report, 2011). Our primary data came from Cameroon’s Demographic and Health Surveys (DHS) of 2004 and 2011 [18]. The DHS is a quinquennial national household survey in over 90 countries sponsored by USAID; it is representative at different subnational levels, thus allowing for varying levels of within-country subgroup analyses. The surveys employ rigorous standardised methodologies, with only minor contextual adaptation, and indicators are comparable across multiple levels. The indicators captured include alcohol consumption, anaemia levels, anthropometry, blood pressure, domestic violence, HIV behaviour, HIV knowledge, HIV status, malaria status, maternal mortality, and tobacco use. Data are accessible on request from the DHS programme. In addition to the DHS data, we used data from the Emergency Events Database (EM-DAT), which maintains an accessible record of emergency events in countries [19], as well as published reports, grey literature and expert knowledge of the context to identify other shocks that occurred in Cameroon around 2004 and between 2004 and 2011. Participants analysed in this study were mothers (15–49 years) and children (5 years old and below) who had participated in the 2004 or 2011 DHS. Participants in the 2011 survey were not the same as in 2004 since the DHS uses a repeated cross-sectional design. We focused on mothers and their children because they are more vulnerable to undernutrition during shocks that lead to food and nutrition insecurity. The consequence of undernutrition is most severe in these groups. Refugees were not part of the DHS dataset due to not being Cameroonian nationals. We analysed data from secondary sources and did not initially engage with the public during the study design and planning phases. The output from this analysis will be available to the population of Cameroon, especially the refugee-hosting communities, through the Ministry of Public Health. In addition, publications from the Demographic and Health Surveys are available on the DHS website for dissemination to the surveyed and broader communities. Because this analysis was an ex-post evaluation, the initial sample size constrained the number in the exposed group. We included all individuals surveyed in the refugee area. We increased the total sample size by performing one-exposed to two-control matches. With an average of 800 individuals in the exposed group for each outcome variable, we estimated the average sample of 2400 individuals per outcome. We used a simplistic formula suggested by Hu and Hoover [20] to determine that the sample size had up to 80% power to detect effect sizes as small as 5%. The outcome variables for children were weight-for-height (for wasting, and defined as wasted or not), height-for-age (for stunting, and defined as stunted or not), weight-for-age (for underweight, and defined as underweight or not) [21] and anaemia status (defined as anaemic or not). Because of their clinical and public health relevance, we treated all outcome variables as binary. We considered measurements of weight-for-height, height-for-age or weight-for-age that were less than -2 standard deviations from the reference medians to be malnourished. For mothers, we used the body mass index – BMI (and defined as underweight or not) and anaemia status (defined as anaemic or not) as outcome variables. In mothers, BMI ≤18.5 was considered underweight. Despite the known limitations of using BMI as an indicator of nutritional status in the adolescent population, owing to their rapid growth during this time [22], we judged that it was still suitable for comparing group outcomes [23, 24]. All participants who had any form of anaemia (mild, moderate, or severe) were considered anaemic. Our exposure variable was the place of residence, i.e., in either the refugee-hosting community or elsewhere in the country, with no other identified shock. We delineated the refugee-hosting community based on the UNHCR 2011 report on Cameroon, which described the refugee areas (Figure 2). Participants were assigned to the refugee hosting community using the quantum geographical information system (QGIS) version 2.14. First, we redistributed sample clusters in the datasets over the national territory using cluster GPS coordinates. Then we considered as exposed all clusters falling within the area mapped out as the refugee area in the UNHCR 2011 country report [16]. Similarly, control participants were selected by matching after mapping and subsequently excluding areas that experienced shocks during the time of interest [25]. Our covariates for matching—household characteristics that could predict the outcome—initially included the following: mother’s educational level (no education, primary, secondary, or tertiary education), residence (urban or rural), wealth index (poorest, poorer, middle, richer, or richest), household size, sex of household head (male or female), age of household head, use of bed nets in the household, aridity index and proximity to a water source. These covariates were selected based on literature and a checklist for selecting covariates for difference-in-differences analysis [26, 27]. All selected covariates were initially included for matching, and we subsequently deleted variables that did not help in achieving a balance between the groups in a stepwise manner. The deleted variables were likely to generate a difference in our study groups, but they correlated with other variables. For example, a covariate such as wealth index is from several other potential covariates (including the household’s ownership of selected assets, such as televisions and bicycles; materials used for housing construction; and types of water access and sanitation facilities). These could easily correlate with other variables like household size and the sex of the household head. The final set of covariates ensured that the exposure and control groups had similar background characteristics at baseline and follow-up. These were wealth index, rural/urban residence, household size, mother education, sex of household head, age of household head, aridity index and proximity to a water source. We extracted relevant variables for children and mothers from the respective files of the 2004 and 2011 DHS datasets. We dichotomised outcome variables using public health-relevant cut-off values while maintaining covariates in their original forms. Our exploratory analysis started by evaluating distributions for each numeric variable and frequencies and percentages for each categorical variable. Specifically, we examined numeric variables for normality in distributions and categorical variables for near-zero variance (very few observations in any class) [28]. We explored missing data using a combination of graphical displays and univariate, bivariate and multivariate methods. We used the GenMatch function in the Matching package in the R statistical software to match controls to exposure [29]. This function uses a genetic search algorithm to determine the weight of each covariate and finds optimal balance using multivariate matching. The genetic algorithm thus allowed for optimised one-to-two nearest neighbour matching. We evaluated the balance of covariates between groups for each outcome using a combination of plots and statistical tests (t-test and Kolmogorov-Smirnov (KS) bootstrap). Once we found balanced controls, we calculated the difference in the proportion of each outcome between groups using the generalised linear model. We calculated the difference-in-differences of outcomes between 2004 and 2011 in per cent point difference. This was the coefficient of the model’s interaction term between year and exposure (exposure∗year). We used a significance level of p-value less than 0.05 and a 95% confidence interval to interpret statistical tests. For sensitivity analyses, we performed similar calculations in the following instances (i) controls sampled randomly from the rest of the country without matching. Here, we aimed to create comparable groups to evaluate the added value of our matching (ii) by removing the regions with the highest number of disasters from the control pool and (iii) by a stepwise reduction in the number of covariates in the matching. Ethics did not apply to this study since this was a secondary data analysis of anonymised datasets. The DHS programme only provided anonymised datasets on individuals and households. To ensure that we further protected potentially identifiable personal information, we respected the following guidelines during analyses (i) adhered to the DHS data use policy, including sharing data only among registered co-authors and (ii) made no attempt to identify any participant in the anonymised datasets.
