Making Pastoralists Count: Geospatial Methods for the Health Surveillance of Nomadic Populations

INTRODUCTION

Nomadic pastoralists defy many basic premises of household demographic surveys, including the assumption that individuals are attached to a geographically stable household, and that this household represents a fixed domestic unit.¹ By contrast, pastoralist settlements are often highly mobile, moving over large areas of remote terrain with the herds of livestock on which they subsist.² Their domestic arrangements are similarly fluid, as family members reside in different geographic locations to manage these livestock. Combined with the dispersed distribution of their encampments and cultural barriers, these mobility patterns make surveying nomadic pastoralists notoriously difficult.

Despite being among the most underserved populations in sub-Saharan Africa,³ nomadic pastoralists are also underrepresented in the population data used to plan health interventions.⁴ Large-scale household surveys such as the Demographic and Health Surveys (DHS) Program typically use census-based sampling frames, which magnify and institutionalize the issue of mobile pastoralists’ under-enumeration in the original census. Such strategies result in the “statistical invisibility” of nomadic populations.¹ Lack of accurate data prevents anything but speculative estimates of the global population of mobile pastoralists, but estimates range from 50 to 217.5 million.^4,5 Pastoralists globally face threats to their health and livelihood, including ecologic disruptions, large-scale development projects, conflict, and protracted humanitarian crises. With mounting concerns about emerging zoonotic pandemic disease, it is critical to find ways of including nomadic populations in household surveys and health surveillance systems.

Numerous methodological approaches to achieving representative samples of mobile populations have been tested, including a “waterpoint approach” in which data are collected at waterholes⁶; a capture–recapture transect approach similar to that used to monitor wildlife⁷; random geographic cluster sampling⁸; and the use of mobile phones.⁹ Many of the approaches explored in these studies have been limited by logistical obstacles or do not lend themselves to being scaled and integrated into data collection exercises such as the DHS.

Geospatial methods present new opportunities for sampling mobile pastoralists and have been used with sedentary populations^10,11 as well as at the single-village level.¹² However, geospatial methods have not been used to develop and implement sampling frames at a scale that could be integrated with existing national health surveys. In this study, we develop and assess the utility and feasibility of a geospatial sampling frame to conduct household surveys among a population of mobile pastoralists in a remote region of southwest Ethiopia.

MATERIALS AND METHODS


SAMPLING DESIGN

We used 0.5-m resolution satellite imagery of the study area in visible and infrared imagery bands, collected within 4 months of survey administration (WorldView-3, April 26, 2017). The imagery was first prepared using the Geographic Data Abstraction Library for pan-sharpening, band-reduction, and export of false-color infrared images in the RGB format, which we used to assist in assessing settlement inhabitance status. The resulting imagery was then converted into an Esri Tile Package and deployed to the ArcGIS Online platform (Redlands, CA) to identify settlements. Using the Hanna Immersive Visualization Environment at Stanford’s Institute for Computational and Mathematical Engineering consisting of 35 high definition screens with an effective resolution of 13,440 × 5,400, we tagged actively inhabited settlements and recorded WGS84 geospatial coordinates for use in clustering. We defined settlements as the area within 500 m of an enclosed camp. Encampments have a distinctive visual profile consisting of an outer ring of kraals and a central collection of huts as well as shade structures, storage platforms, and meeting areas (See Figure 1). Among other populations, the composition of settlements will vary, and different features should be used for identification. Because nomadic dwellings are typically constructed from natural materials such as saplings and grass, inhabited versus abandoned settlements can be difficult to distinguish from satellite imagery. Because of the presence of grazing livestock, inhabited settlements contain little interior vegetation, whereas uninhabited settlements tend to be overgrown. We used the presence of significant vegetation in the interior of a settlement as an indicator that the settlement was likely to be uninhabited.

View Full Size

Nyangatom settlement (diameter approx. 100 m) with huts, livestock kraals, and perimeter fence viewed using infrared and visible multispectral imagery (red indicates vegetation). Captured from imagery covering total land area of 5000 km² at a resolution of 0.53 m (imagery provided by DigitalGlobe Foundation, Westminster, CO).

Citation: The American Journal of Tropical Medicine and Hygiene 101, 3; 10.4269/ajtmh.18-1009

• Download Figure
• Download figure as PowerPoint slide

As the only available sampling frame of primary sampling units (PSUs) was based on census data and, therefore, not representative of the target population, we relied on spatial data to divide the entire area into smaller adjacent subregions as homogeneous as possible in terms of population size. We visually identified 225 settlements in the study area, comprising 3,099 huts. With ArcMap’s Grouping Analysis tool, we used settlement elevation and K-nearest neighbors for initial grouping analysis, which produced 15 seed points (the maximum number allowed by the program). We then grouped settlements into 15 clusters with an average of 206 huts per cluster around these points based on elevation and spatial proximity. To develop PSUs, we generated Thiessen polygons from these clusters using the spatial median of settlements within each cluster, thereby minimizing the total distance to all points. This process was reiterated until no further change to cluster membership occurred, producing 15 PSUs covering the entire survey area. Although other approaches such as treating individual settlements as the PSU were considered, we favored clustering because the sparse distribution of settlements would have otherwise made conducting the field survey infeasible. Our approach ensured that the survey would be operationalizable, minimizing travel time within clusters and maximizing field team safety. We then selected three PSUs to survey using probability proportional to size sampling based on the number of huts per settlement as identified in the geospatial data. Although a greater number of sampling units could have been selected to provide an even more representative sample, our decision to sample three clusters balanced reducing sampling variability with the financial and logistical constraints of the field survey (see the Sample Size Calculation and Sampling Weights section in Supplementary Information).

Last, we carried out a field survey in the selected PSUs. We used individual settlements as identified in satellite imagery as our secondary sampling unit. We generated a randomized list of all settlements within each cluster, visiting a total of 25 settlements to achieve a sufficient sample size (see the Sample Size Calculation section in Supplementary Information). Within each settlement, we conducted in-person interviews with all women of reproductive age who reported having a birth at any time during the five years preceding the survey. We also collected anthropometric data for the three youngest children aged 0–15 years for each woman in our sample. Our survey instruments were developed using questionnaires from the most recent Ethiopian DHS (EDHS 2016).¹⁵

RESULTS

We interviewed 342 women reporting having given birth during the five years preceding the survey. Maternal age was not recorded because the Nyangatom do not keep track of their own age; however, we assumed a reproductive range of 15–49 years.

Table 1 presents background demographic information on the Nyangatom women of reproductive age included in our analysis, including marriage status, wife ranking, and indicators of fertility and mortality among our sample.

Problems accessing health care.

Women reported that the primary impediments to receiving health care were not wanting to travel alone (44.0%, 95% CI: 25.6–64.2%) and distance to the health facility (43.3%, 95% CI: 17.0–74.0%). Compared with regional data from the 2016 EDHS in which 76.7% of women reported being responsible for decisions about her own health care, only 54% of Nyangatom women in our sample reported being involved in decision-making, whereas 45% cited their husbands as being the sole party responsible for making decisions regarding the health care for the woman herself (Figure 2). However, only a small proportion (2.9%, 95% CI: 0.4–20.0, 11/347) of women reported that obtaining permission from her husband to visit a health facility was a barrier to accessing care. Nearly all women requested permission from their husbands before traveling to a health facility (91.5%, 95% CI: 78.0–97.1%), but of these women, 97.7% (95% CI: 91.6–99.4%) reported that their husbands granted them permission all of the time (Table 2).

View Full Size

Women’s participation in decision-making about their own health¹. Source: Nyangatom sample and EDHS (2016). ¹Excluding the 1% of respondents who refused.

Citation: The American Journal of Tropical Medicine and Hygiene 101, 3; 10.4269/ajtmh.18-1009

• Download Figure
• Download figure as PowerPoint slide

Table 2

Problems accessing health care (Nyangatom, 2017)

	n	Weighted %	SE*
Number of respondents (women)	347
Asked permission (yes)	307	91.3	2.2
Permission granted (yes)	300	97.7	0.7
Barriers to care identified by respondents
Obtaining permission to go to the doctor	11	2.9	1.4
Lack of money required for advice or treatment	100	25.5	9.6
Distance to health facility	147	43.3	7.5
Not wanting to go alone	148	44.0	4.7
Other†	41	11.8	0.5

Source: Nyangatom sample.

* SE = linearized standard error.

† Of the 11.7% (41/347) of women responding that “other” factors prevented them from seeking health care, the majority (69.7%, 28/41) reported that this problem was due to the requirements of caring for livestock and domestic chores.

Women reported having more agency in decision-making for the health care of their children than in their own health care: 61.4% of women reported being involved in decision-making for the health care of her children. Numerous women reported that they consulted their husbands only to obtain money when they themselves did not possess money for transportation to the clinic or services at the clinic. A lack of alternative arrangements to care for livestock and domestic labor was cited by respondents as a major barrier to care (Table 2).

Anthropometric data.

Estimates of HAZ/stunting, WAZ/underweight, and WHZ/wasting were based on samples comprising 339, 190, and 182 children younger than 5 years, respectively (see Supplemental Methods). Mid-upper arm circumference measurements were recorded for 339 children aged 6–59 months. Nutritional status was favorable among Nyangatom children in comparison with rural data from the 2016 EDHS. This finding was most significant with respect to WHZ/wasting. In our sample, 13.3% of children younger than 5 years met criteria for wasting compared with 24.8% from the 2016 rural EDHS, and only 1.1% met criteria for severe wasting compared with 7.3% of children in the 2016 rural EDHS (Figure 3). The lower prevalence of stunting and wasting observed among our sample as compared with the 2016 DHS, calculated using height and weight measurements, was consistent with our finding that very few children (2.1%) met the criteria for SAM, calculated using MUAC (see Table 4 for complete results on stunting, wasting, and other anthropometric data).

View Full Size

Severe acute malnutrition (SAM) and wasting by gender based on weight-for-height of Nyangatom children younger than 5 years in comparison with WHO height-for-weight reference chart values. The lightly shaded region indicates the WHO median to 3 SDs, whereas the darker shaded region is the 95% CI of the Nyangatom sample. Children falling below the dashed line (< − 2 SD) meet the WHO definition for wasting, whereas those with < −3 SD meet criteria for SAM. Source: Nyangatom sample and data from WHO-UNICEF 20091⁶. Note: Sample restricted to children with valid measures for weight-for-height (using the original sample, n = 73). Observation symbol size represents the probability proportional to size sample weighting (15.1, 21.9, and 23.5) used in the Loess smoothed regression line and 95% CI. Where month of birth was not known, we used multiple imputation to estimate date of birth (month/year; see the Methods section in Supplementary Information).

Citation: The American Journal of Tropical Medicine and Hygiene 101, 3; 10.4269/ajtmh.18-1009

• Download Figure
• Download figure as PowerPoint slide

Table 4

Nutrition status children younger than 5 years, Nyangatom 2017

Anthropometric failure	Ethiopia DHS 2016 (weighted Mean/%)		Month and year of birth known			(1) vs. (2) P-value*	Month of birth estimated by multiple imputation
	Rural	Southern Nations, Nationalities, and People’s Region (1)	n	Weighted (mean/%) (2)	95% CI		n	Weighted (mean/%) (3)	95% CI†
HAZ/stunting
Sample size (n)			164				337‡
Mean HAZ (95% CI)	−1.5	−1.5		−0.88	(−1.77, 0.01)	0.10		−0.83	(−1.09, −0.56)
Stunted n (%)	39.9	38.6	59	35.3	(23.9, 48.8)	0.38		33.7	(28.1, 39.3)
Severely stunted	18.4	20.2	36	21.2	(11.5, 35.8)	0.76		19.4	(14.7, 24.1)
WAZ/underweight
Sample size (n)			78				190
Mean WAZ (95% CI)	−0.5	−0.2		−0.89	(−1.76, −0.03)	0.07		−0.77	(−0.96, −0.58)
Underweight n (%)	10.1	6.0	21	26.9	(17.4, 39.3)	0.01		18.5	(12.4, 24.6)
Severely underweight	3.0	1.7	6	7.1	(1.0, 36.4)	0.22		4.5	(1.4, 7.6)
WHZ/Wasting
Sample size (n)			73				182
Mean WHZ (95% CI)	−1.2	−1.0		−1.08	(−1.52, −0.65)	0.50		−1.09	(−1.21, −0.96)
Wasting n (%)	24.8	21.1	14	18.3	(5.2, 47.9)	0.63		13.3	(8.3, 18.4)
Severe wasting	7.3	6.4	2	2.7	(0.4, 16.6)	0.09		1.1	(−0.4, 2.5)
MUAC, cm
Sample size (Number of children younger than 5 years> 6 m)			157				336
Mean (cm)				13.5	(12.8, 14.2)			13.6‡	(13.5, 13.7)
Severe acute malnutrition: MUAC < 11.5 cm)			4	1.1	(0.9, 1.3)			2.2	(0.6, 3.8)

DHS = Demographic and Health Surveys; HAZ = height-for-age; MUAC = mid-upper arm circumference; WAZ = weight-for-age; WHZ = weight-for-height.

Source: Nyangatom sample

* P-value of the adjusted Wald test (that accounts for the survey design) for the null hypothesis H₀: Prop_DHS2016 = Prop_{Nyangatom2017}, with a level of significance of 5% [“Prop” stands for “mean” or “proportion”].

† Accounts for the variation in the imputed samples for age in months.

‡ Estimation sample varies across imputations; sample sizes vary between 337 and 343 for HAZ/stunting, and between 336 and 343 for MUAC.

Values and percentages in (OS) are based on reported month and year of birth. Percentages and 95% CI in (MI) were calculated using 120 imputed values on month of birth

DISCUSSION

Our results demonstrate that geospatial sampling frameworks can be used to construct representative samples of nomadic populations in a logistically feasible, cost-sensitive manner. Our approach was designed to overcome the limitations of census-based sampling frames such as those used by household surveys, including the DHS, the World Bank’s Living Standards Measurement Study, and UNICEF’s Multiple Indicator Cluster Surveys. Conventional methods fail to account for population mobility, contributing to the exclusion of nomadic groups in survey data.

Although enumerations using geospatial imagery have been reported at the single settlement level,¹² this is to our knowledge the first time a geospatial sampling frame was implemented to conduct a population-level assessment. Previous studies found that dispatching survey personnel without first determining the inhabitance status of randomized geographic regions was logistically infeasible.⁸ By enumerating currently inhabited settlements in advance, we were able to assess active population distributions to generate density estimates in a remote, difficult-to-access area. We then demonstrated the logistical feasibility of this method through our field validation.

The use of recent satellite imagery played a crucial role by providing nearly real-time information about the distribution of the target population. Of the 25 settlements selected for survey, two exhibited signs of very recent abandonment (both were active when scouted in the week before the intended date of survey, but abandoned by the time of survey), two showed signs of abandonment at a more distant time point (demonstrating the limitations of identifying settlement status from space), and one site was not located because of a suspected faulty coordinate pair. Refining our methods in the future, we expect that we will be able to compress the timeline from satellite imagery capture to survey deployment, reducing the number of settlements abandoned by the time of the field team’s arrival. For subsequent surveys, we recommend visual confirmation and coordinate checks for all selected settlements, once the sampling frame has been generated.

Once the satellite imagery for our study was prepared, manual reconnaissance required ∼6 hours to visually enumerate settlements in an area of over 5,000 km². This approach is minimally resource intensive over large areas of terrain characteristic of pastoralist environments. Our method depends on environmental features such as sparse vegetation and the distinctive visual signature of settlements. Although implementation may be more difficult in heavily forested settings, nomadic populations tend to inhabit areas of sparse vegetation similar to the conditions of our study area. For this reason, we believe that our methods should be generalizable to nomadic pastoralists in general, and in sub-Saharan Africa in particular. In addition, we believe that machine automation of parts of the reconnaissance process will be possible, reducing human intervention to verification.¹⁷ The increasing availability of high-resolution, high-cadence satellite imagery and software pipelines for automated pre-processing and feature extraction will reduce human reconnaissance time and increase the capacity for greater geographic coverage.

Integration with existing DHS methodology is crucial. This could be carried out by working with Ministries of Health and other national partners to identify pastoralist regions in which a geospatial approach should be implemented in favor of traditional census-based enumerations. Survey weighting and population estimates remain a challenge. However, we feel that our approach provides the possibility of more accurate population estimates than existing methods.

Compared with the 2016 EDHS sample regionally and in rural areas nationally, Nyangatom women demonstrate a relatively high uptake of ANC services. Recent NGO- and government-supported initiatives, including the provision of free MCH services at the nearest health clinic as well as recent road construction increasing access to transportation may explain this trend. Our finding that, under certain conditions, pastoralist women seek treatment at health facilities at rates comparable with regional non-pastoralist populations counters narratives that pastoralists tend to underutilize health services because of cultural beliefs or preferences for traditional healers.¹⁸ At a minimum, our findings suggest that such reports cannot be extrapolated to characterize the health-seeking behavior of pastoralist populations in general.

Despite higher ANC-seeking rates, skilled birth attendance (SBA) was significantly lower (6.4%) among the Nyangatom compared not only with the average from rural areas nationally (21.2%), but also lower than the lowest regional figure reported by the 2016 Ethiopia DHS (16.4% in the Afar region). The national trend in the reduction of at-home births has not reached this population, and the vast majority of births occur at home (91.2%) compared with the national rural average (79.0). Only 6.8% of women in our sample delivered at a health facility versus 19.6% in rural areas nationally. The determinates of this discrepancy would benefit from further research. Antenatal care services may be more accessible for pastoralist women because they can time their visits when their encampments are settled in a location favoring travel to a health facility. By contrast, pastoralist women may be unable to time their deliveries to give birth at a health clinic, as they frequently continue to migrate with mobile livestock encampments and perform domestic labor in locations distant from health facilities throughout their pregnancy. Although the generalizability of this finding to other pastoralist populations should be investigated, this finding highlights opportunities to target programming, outreach, and research toward SBA and delivery practices among pastoralists.

Although limited by lack of reliable documentation, Nyangatom children in our sample had low vaccination coverage compared with the EDHS, with only 62.8% of Nyangatom children aged 12–35 months having ever received a vaccination versus 84.1% of children regionally in SNNPR. The majority of our respondents did not retain possession of their vaccination cards even when these documents had been provided to them, or produced documents that were unreadable because of weathering from dirt and mold. As a result, mothers were not able to furnish definitive documentation of full immunization coverage for any of the children in our sample. By mothers’ recall, these findings are consistent with other reports of low rates of vaccination coverage among other pastoralist populations.^19–21

Ethiopia’s National Expanded Programme on Immunization has identified pastoralists as the predominant population at risk for missing immunization services.²² Many respondents in our sample reported that their children missed vaccination events while residing in a different area than where coverage was provided. These accounts are similar to those of Fulani pastoralists in Nigeria, most of who reported they were missed by supplementary immunization activities during an oral polio vaccination campaign because of absence of the child at the time of health workers’ visit.²³ Such studies provide support for the argument that sampling bias may lead to substantial inaccuracies in estimates of vaccination coverage,²⁴ and suggest that even mobile vaccination programs may be insufficient if they are not informed by knowledge of seasonal transhumance patterns.

Our findings suggest that conventional narratives of factors preventing pastoralist women from seeking health care may be inaccurate in some settings. Among our study population, participants reported that money, women’s empowerment, and distance to a health facility were not as significant an impediment to seeking care as insecurity and the issue of domestic labor. Subsistence livelihoods rely on daily labor from female household members such as milking and watering livestock, and collecting firewood and water. Whereas for urban populations lack of access to money may be a major impediment, among pastoralists, the requirements of household labor may be more significant. Likewise, living in a patriarchal society such as that of the Nyangatom cannot be interpreted to mean that obtaining permission from husbands is a significant barrier to care. This finding differs from previous studies reporting that men’s control over health-care decisions is a primary barrier to care for pastoralist women.^21,25 Although most women in our sample needed to obtain their husband’s permission to seek care, the vast majority (97%) reported they were granted this permission all of the time and that this was not, therefore, a barrier to care.

Nyangatom children in our sample had a lower prevalence of SAM and stunting compared with national averages. Although we believe that these results represent adequate nutritional status of the pastoralist children in our sample, the issue of survivorship remains a possibility if undernourished children do not survive to appear in our sample. For children older than approximately 6 months, our results are consistent with findings from previous research demonstrating that settled children may have poorer nutritional status. This has been attributed to decreased access to milk and animal products associated with a sedentarized diet, even when total caloric intake may be greater.^26,27 There may be variability between populations, as other studies have found children of nomadic families to have unfavorable nutritional status compared with sedentary counterparts.^20,28 These findings underscore that arguments for sedentarization premised on improvements in childhood nutritional status (historically put forward by development agencies, national governments, and nongovernmental organizations) must have an empiric basis.²⁷ An unexpected finding of our study concerned the misuse of veterinary medicine for self-treatment. Numerous mothers reported administering livestock medicine to their children, including the use of injectable cattle ivermectin applied directly to the eyes for pediatric ophthalmic illness, and veterinary oxytetracycline as an orally administered antitussive. This potentially concerning practice merits further research.

The limitations of this study highlight several challenges inherent in surveying pastoralist populations. First, most Nyangatom women in our study population were unable to accurately date events and ages. As a result, it was necessary to impute ages for a number of children in our sample, with a potential impact on indicators depending on children’s precise age. Second, the domestic labor schedules of our participants constrained both the hours during which it was feasible to collect data as well as participants’ susceptibility to survey fatigue, limiting the length of our survey. During data collection, women were occupied chasing birds away from their fields and often left their dwellings before dawn. For similar reasons, we were unable to interview household heads, who herded cattle during the day and were rarely present in the same location as women participants.

Our findings suggest several ways in which MCH interventions among pastoralists can be improved. First, vaccination outreach campaigns should coordinate their timing to occur when settlements are at their largest size (among the Nyangatom, this is during the wet season), to avoid missing inhabitants of livestock encampments dispersed across remote grazing territory in times of scarcity.²⁶ Second, alternative record-keeping systems for immunizations should be developed and implemented with pastoralists. Nomadic households are unlikely to retain these documents, and paper cards will not weather exposure to the elements incurred during frequent migrations.

In addition, our findings suggest that men’s control over health-care decisions may not be a major barrier to pastoralist women’s ability to seek care even in cases where a husband’s permission is required. By contrast, domestic labor and responsibilities such as caring for the family’s livestock, homestead, and other children appear to be significant and under-recognized barriers to pastoralist women seeking care. Efforts targeting these domains may be more effective in improving MCH among pastoralists, and empirically informed studies investigating the determinants of health-seeking behavior should be used to guide programming and decision-making. Interventions focusing on men’s awareness, for example, may not be as effective or impactful as providing secure, expedient transportation to a clinic to minimize time away from the household.

Despite pastoralist groups’ long history of invisibility in population data, our sampling design and successful field deployment indicate that geospatially derived sampling frames hold potential for the systematic inclusion of mobile populations in household surveys. This method is a viable alternative for use in pastoralist regions and has the potential to increase the representation of mobile populations in survey data. The disparities we found between the study population and DHS-derived country estimates for certain MCH indicators suggest that without the implementation of sampling strategies capable of accounting for mobility, key indicators in data used to monitor progress on Sustainable Development Goals and other targets may present an inaccurate picture of pastoralists’ health status.

Above all, we propose that policymakers and national governments implement the use of geospatial sampling frames in pastoralist regions to reduce under-coverage and prevent bias in national estimates. This is a crucial step toward designing health systems and implementing longitudinal health surveillance for underserved mobile groups. Such an approach can aid in more effectively targeting the limited resources of national Ministries of Health and serve as an overdue foundation to include mobile pastoralists in frameworks to achieve universal health coverage.

Supplemental materials