Ethics statement

This study used data from the ‘Social, Environmental, and Economic Drivers of Zoonotic disease’ (SEEDZ) project. Approval for the SEEDZ study was provided by the ethics review committees of the Kilimanjaro Christian Medical Centre (KCMC/832) and National Institute of Medical Research (NIMR/2028) in Tanzania, and by the ethics review committee of the College of Medical, Veterinary and Life Sciences at the University of Glasgow (39a/15) in the UK. Formal written consent was obtained from all SEEDZ participants.

Framework overview and study area

A framework was developed to: i) optimise the use of currently available data, capturing uncertainty in livestock production and brucellosis prevalence estimates; ii) account for heterogeneity in livestock production systems across the study area; and, iii) estimate production losses attributable to brucellosis in each system.

The framework was divided into three parts: i) Part I – retrospective data gathering and harmonisation (the process of systematically improving comparability of data collected by multiple studies) [31]; ii) Part II – harmonised household classification models; and, iii) Part III – estimation of production losses due to livestock brucellosis by production system.

This study focused on some of the northern and central regions of Tanzania: Arusha, Manyara, Kilimanjaro, Tanga, Dodoma, Singida, Simiyu, and Mara Regions. These regions cover an area equivalent to approximately 29% of mainland Tanzania but include around half of the livestock population in the country [32,33]. In October 2017, there were 13.0, 10.6, and 3.2 million cattle, goats, and sheep in these regions, respectively, which correspond to 42.6%, 55.8%, and 57.9% of the total population of each species in mainland Tanzania [33].

This large study area has highly diverse socio-economic and geo-climatic settings. The regional average human population density is close to the national average (51 people/km²), except in Kilimanjaro, where there are approximately 124 people/km² [34]. Kilimanjaro is also the region with the smallest average household size (4.3 people per household) in the study area, contrasting with Simiyu, where the average household size is the largest nationally (6.9 people per household) [34]. The area is dominated by savannas, grassland and, to a lower extent, small-scale crop cultivation [35]. The mean annual temperature ranges from 20.2 ^oC in Arusha to 23.2 ^oC in Tanga [36]. The mean annual precipitation ranges from 616 mm in Dodoma to 1,060 in Mara [36]. The elevation ranges from the sea-level values in the Tanga region to the highest point of the African continent: Mount Kilimanjaro (5,895 m) [37]. Arusha and Manyara are dominated by livestock-only, grassland-based systems, whilst the other regions include predominantly rainfed mixed crop/livestock systems [38].

Part I – Retrospective data gathering and harmonisation

A protocol based on the Maelstrom Research guidelines for retrospective data harmonisation was followed [31]. An informal review of the literature on household surveys in northern and central Tanzania and classification of livestock production systems was carried out. Researchers with experience in this context and research area were also contacted for feedback on methodological issues. These preliminary steps enabled the identification of candidate surveys and the definition of survey selection criteria.

The SEEDZ project delivered a large-scale, cross-sectional survey, dedicated to the analysis of risk factors for several zoonoses in people and animals, between February and November 2016, in two regions of northern Tanzania: Arusha and Manyara Regions. In the SEEDZ project, households were classified by production system, and livestock were serologically tested for exposure to Brucella spp. No previous administration of vaccines against Brucella spp. in livestock was reported by participating households. The study design and the methods used have been described in detail elsewhere [15,16]. In the SEEDZ project, a household is defined as a separate dwelling within a compound with a husband, his wife(s) and their children. The SEEDZ survey was used as a contextual, spatial, and temporal reference for gathering data on livestock production from additional surveys.

The following criteria were used for inclusion of additional surveys: (i) surveys that were delivered in the SEEDZ study area (Arusha and Manyara) or neighbouring regions (Kilimanjaro, Tanga, Dodoma, Singida, Simiyu, and Mara), (ii) surveys that collected information on the number of livestock kept and their production characteristics (e.g., wholesale price, milk yield, lactation length, and milk price), household demographics, crop agriculture, livestock management practices, food consumption practices, and indicators of household vulnerability (e.g., crop losses, illness in livestock and people), (iii) surveys that took place between 2013 and 2019 (three years before and after SEEDZ), and (iv) had publicly accessible survey data that included approximate geolocation of the households.

The application of survey selection criteria led to the inclusion of the following surveys, in addition to SEEDZ: i) the ‘Living Standards Measurement Study’ (LSMS), and ii) the ‘Rural Household Multiple Indicator Survey’ (RHoMIS). General information about these surveys, including research design, data collection procedures, questionnaires, and data processing and evaluation methods, was documented. These surveys gathered data through face-to-face interviews with either household heads or other household members.

The LSMS is a multi-country, nationally-representative household survey programme led by the World Bank since the 1980s [39]. This programme has a longitudinal design, which allows researchers to monitor the living conditions of a sample of households over time in ‘waves’. In Tanzania, four waves of the programme have been delivered until 2019 [40]. Data from waves three (LSMS3, October 2012 to November 2013) and four (LSMS4, October 2014 to November 2015) were used in this study due to their temporal proximity to the SEEDZ survey. In the LSMS project, a ‘household’ is defined as ‘people who live together and share income and also basic needs’ [41,42].

RHoMIS is a standard survey tool focused on household socioeconomic characteristics, which has been applied in more than 30 countries since 2015 [43]. In northern and central Tanzania, based on GPS location data, two applications of the RHoMIS tool were available by June 2020: i) RHoMIS1 - a survey of households randomly selected from community animal health service records in 2017; and, ii) RHoMIS2 - a survey of households randomly selected from village lists obtained from village elders in 2018. In the RHoMIS project, a ‘household’ is defined as people ‘who usually live, sleep and eat’ at a house or compound and includes ‘anyone who is temporarily away for less than 3 months, or anyone who is staying for more than 3 months’ [43].

The definition and availability of variables across survey datasets were compared and analysed to enable an initial evaluation of the potential for running harmonised household classification models, i.e., clustering models based on de Glanville et al. (2020) [16] applied to households from candidate surveys and those in SEEDZ. The clustering analyses of the SEEDZ dataset by de Glanville et al. (2020) [16] served as a basis for data harmonisation and identification of livestock production systems in candidate datasets. The timing, spatial extent, and number of households included in each dataset are available in Table A in Supporting information (S1 File). The spatial distribution of households in each dataset is shown in Fig A in Supporting information (S1 File).

For all datasets, the common variables used for clustering were divided into eight domains (as per de Glanville et al., 2020) [16]: 1) local household environment; 2) household demographics; 3) crop agriculture; 4) numbers of cattle, sheep, and goats owned; 5) other livestock owned; 6) livestock management practices; 7) household food consumption practices; and, 8) indicators of household vulnerability. Domains 1 (local environment) and 4 (number of livestock owned) include only numerical variables. The other domains include only categorical variables (binary variables that indicate the presence or absence of an attribute).

Data on variables in domain 1 (local household environment) were sourced from third-party datasets. A summary description of each environmental variable (and data sources) is available in Table B in Supporting information (S1 File). The methods used to gather these variables were adapted from de Glanville et al. (2020) [16] for SEEDZ. However, some methods were not applicable due to the relatively low precision of GPS locations of households in the candidate surveys. For such cases, buffers were created around the approximated GPS locations of households using the ‘sf’ package [44,45] in R [46]. The buffer radii were set at 5 km for LSMS households in rural areas, 2 km for LSMS households in urban areas, and 1.1 km for all RHoMIS households. These radii were chosen based on the coordinate modification strategy in each survey, which was applied to preserve confidentiality. The mean data value for environmental variables within each buffer was extracted using the ‘exactextractr’ package [47] in R. The proportion of the buffer area classified as each landcover type was extracted using the ‘sf’ package in R. The village area to which the households belonged was calculated as the weighted mean of the areas of the villages covered by the buffer.

Part II – Harmonised household classification models

de Glanville et al. (2020) [16] applied multiple factor and hierarchical clustering analyses to classify SEEDZ households as pastoral, agro-pastoral, or smallholder. Multiple factor analysis (MFA) enabled the detection of underlying patterns among variables organised in domains [48]. Hierarchical clustering analysis used the outputs of MFA to group SEEDZ households into production systems based on similarities between households. The goal in the current study was to apply the same methodology to classify LSMS and RHoMIS households. However, many of the variables in SEEDZ were not available in LSMS and RHoMIS. The SEEDZ dataset included in de Glanville et al. (2020) [16] had 76 variables. Out of these 76 variables, 12 (15.8%) were absent in LSMS, and 28 (36.8%) were absent in RHoMIS. All the variables that were absent in LSMS were also absent in RHoMIS.

Given differences in availability of variables between datasets, several household classification models were run with different combinations of datasets and variables. Three main integration options were tested: i) adding one non-SEEDZ household at a time to the original model (M0), which initially contained SEEDZ data only, as [16]; ii) adding one non-SEEDZ dataset at a time to M0; or, iii) adding all data together.

The performance of each model was evaluated with reference to the classification of SEEDZ households derived with M0. Given that the purpose of this study was the classification of households into two predefined production systems (pastoral and non-pastoral), agro-pastoral (cluster two) and smallholder (cluster three) households were merged and characterised together as non-pastoral households. The performance metric used was the probability of a model classifying a SEEDZ household correctly as pastoral or non-pastoral. The best-performing model was selected for classifying LSMS and RHoMIS households by production system. The household classification algorithm is described in detail in Supporting information (S2 File).

Part III – Estimation of production losses due to livestock brucellosis by production system

Part III of this study had two main objectives: i) to gather system-specific demographic, brucellosis prevalence, and production parameter values from the three survey datasets; and, ii) to run herd production simulations in the presence and in the absence of brucellosis, using literature-based information on disease impacts and additional production characteristics. The losses attributable to brucellosis were calculated as the difference between the estimated value of livestock production (including all revenue-generating elements captured in market value and other benefits, such as own consumption of animal products) simulated in the absence and in the presence of the disease. The same workflow, i.e. a series of steps and processes for data analysis, was used to estimate losses in cattle, sheep, and goats. However, whilst only offtake (wholesale, home slaughter, or gifts given away) and milk contributed to production or income from sheep and goats, draught power and dung (e.g., to be used as manure, cooking fuel, and in construction) were also considered in cattle. Indirect consequences of production losses on value-added products such as yoghurt or cheese were not captured in this study.

The workflow was divided into two main layers (global and system-specific parameters) and seven interrelated steps (from A to G), as shown in Fig 1. Steps A to F consisted of processes that gathered and selected parameter values through several mechanisms: i) probabilistic distributions (system-specific distributions fitted to survey data - steps A and C; or non-system-specific, literature-based, uniform distributions - step D); and, ii) fixed values (system-specific demography - step B). Step A applied a Bayesian latent-class analysis (LCA) model to derive individual animal-level brucellosis prevalence estimates for pastoral and non-pastoral households in the SEEDZ dataset, as described by Bodenham et al. (2021) [15]. In this process, a simplifying assumption was made: an animal with serological evidence of exposure to Brucella spp. was considered ‘an infected animal’. Steps B and C gathered production-system-specific demographic and production values in the presence of brucellosis. Step D gathered from the literature global parameter values (same values across production systems) on direct brucellosis impacts and some production characteristics. The parameter values gathered and selected in steps A to D were used to estimate demography and production characteristics that would be expected in the absence of brucellosis (steps E and F). These steps were then followed by the modelling of herd production (simulations) in the presence and in the absence of brucellosis to estimate the losses attributable to the disease (step G). Losses were estimated per infected animal, per household, and per region. Losses attributable to each abortion were also estimated by linear regression using model outputs, i.e., losses per household, and abortions per herd, as regression variables. All monetary values were discounted over time at a rate of 3%. This means that the income from livestock is valued over time as a (discounted) present value. The discount rate of 3% aligns with the value used in a similar application in the study area [49]. The whole sequence of steps was repeated 50,000 times to fully explore the multidimensional parameter space. Each of these steps is described in detail in Supporting information (S3 File).

The model herd size (N) in each system, i.e., the initial number of animals of each species per household, was given by the median number of animals observed per household in each system in the SEEDZ survey. The expected consequences of the hypothetical, sudden elimination of brucellosis from the system on the herd size after ten years (simulation time horizon, t =  10 years) were modelled through two different model scenarios: i) ‘no increase in herd size’: all excess stock due to brucellosis elimination was valued as offtake (sold, slaughtered, or given away as gifts); and, ii) ‘unrestricted increase in herd size’. Unless otherwise stated, results refer to the first model scenario (‘no increase in herd size’; reference scenario) as this is likely the most appropriate to represent how households would manage herd size in the event of brucellosis elimination.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Workflow used to estimate losses attributable to brucellosis in livestock in northern and central Tanzania.

Fixed parameter values are represented with a vertical red line. Parameter values obtained from distributions fitted to data are represented with a vertical red line in a grey normal distribution. Parameter values obtained from uniform distributions are represented with a vertical red line in a grey horizontal rectangle. Boxes A to G represent the step-wise processes. In the panel for step G (the model), the red shaded areas indicate the losses estimated in one iteration of the simulation (∑), under the two model scenarios considered: no increase (left) and unrestricted increase (right). ^* Parameters that entered the model as fixed values. ^† The values of milk price (p) were obtained from LSMS and RHoMIS and entered the model as distributions fitted to data.

https://doi.org/10.1371/journal.pntd.0012814.g001

Statistical analyses

General descriptive statistics were calculated for each of the main model outputs: i) losses per infected animal; ii) losses per household; and, iii) losses per region. Pearson’s Chi-squared test was used to assess differences between production systems in (i) demographic attributes of animal populations (e.g., proportion of animals in each age-sex group) and (ii) the proportion of households harvesting milk [50]. The herd sizes, the productivity of livestock (e.g., milk yield) as well as the estimated production losses per infected animal and per household were compared between production systems using the Wilcoxon rank-sum test [50]. The significance level was set at 5%. These comparative analyses were also carried out visually, through the inspection of violin and bar plots, which were made with the ‘ggpubr’ and ‘ggplot2’ packages, in R [51,52]. The influence of the model parameter values on the estimated production losses was assessed through elasticity analyses [53]. Elasticity analyses estimated the percentage change in production losses (per infected animal and household) that resulted from a percentage change in parameter values. These analyses were performed separately for the two model scenarios: ‘no increase in herd size’ and ‘unrestricted increase in herd size’.

Elasticity was calculated using a log-log function, as follows [54]:

(1)

where y is the production loss per infected animal in each species, I is the regression intercept, E is the elasticity (percentage change in production loss that results from a percentage change in the parameter value), and x is the input parameter.

Household classification by production system

The percentage of households classified as pastoral with the best-performing model was 42.8% (173 out of 404), 12.9% (96 out of 746), and 29.4% (115 out of 391) in SEEDZ, LSMS, and RHoMIS, respectively. The individual scores of LSMS and RHoMIS households on the first and second factors from the multiple factor analysis followed broadly the distribution of SEEDZ households scores of M0. This result is shown in Fig 2, where three distinct clusters are visible in LSMS and RHoMIS, overlapping those in SEEDZ: i) cluster one, predominately negative scores on dimension 1, consistent with pastoral households; ii) cluster two, predominately positive scores on dimension 1, consistent with agro-pastoral households; and, iii) cluster three, predominately positive scores on dimensions 1 and 2, consistent with smallholder households [16].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Individual scores of LSMS and RHoMIS households in shaped and coloured dots, as compared to SEEDZ’s (grey squares, as per de Glanville et al. (2020) [16]), on the first and second factors of the multiple factor analysis, and membership of each household to a cluster.

Results obtained with the best-performing model: a model with all SEEDZ variables, except those that were missing in LSMS, integrated ‘one by one’ to the whole SEEDZ dataset, and with partition of the hierarchical tree constrained to three clusters. SEEDZ refers to the Social, Environmental, and Economic Drivers of Zoonotic disease project. LSMS refers to the Living Standards Measurement Study. RHoMIS refers to the Rural Household Multiple Indicator Survey.

https://doi.org/10.1371/journal.pntd.0012814.g002

Livestock brucellosis prevalence

The estimates of individual-level brucellosis prevalence by species and production system, obtained with the latent-class analysis model, are shown in Fig 3. The highest prevalence estimates were in pastoral cattle (median of 5.1%), which were over seven times more likely to be infected than non-pastoral cattle (median of 0.7%). In small ruminants, the prevalence estimates were similar across production systems, and slightly higher in goats (median of 2.5% in both systems) compared to sheep (median of 1.3 and 1.6% in pastoral and non-pastoral systems, respectively).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Latent-class analysis-derived brucellosis prevalence estimates by production system (pastoral and non-pastoral) and species (cattle, sheep, and goats).

White dots indicate the median, black diamonds the mean, black boxes the 50% credible interval, and black lines the 95% credible interval. ‘n’ indicates the number of animals sampled and tested for brucellosis with both the Rose Bengal test (RBT) and competitive enzyme-linked immunosorbent assay (cELISA).

https://doi.org/10.1371/journal.pntd.0012814.g003

Production losses attributable to livestock brucellosis

The estimated production losses per infected animal, by species, production system, and model scenario, are shown in Fig 4. Under the model scenario of ‘no increase in herd size’ (left panel in Fig 4), the median (95% uncertainty interval, UI) annual losses attributable to brucellosis in cattle, sheep, and goats, per infected animal, were 74.4 (26.2-211.7), 9.7 (3.4-23.1), and 10.6 (3.7-25.0) international dollars (int. $) in the pastoral system, and 62.3 (16.8-228.6), 6.3 (1.8-17.1), and 7.0 (2.2-17.9) int. $ in the non-pastoral system, respectively. In cattle, the estimated production losses were approximately seven to ten times as high as those estimated in sheep and goats. For all species, greater losses were estimated in the pastoral system, compared to the non-pastoral system. Additionally, within the same system, greater losses per infected animal were estimated in goats, compared to sheep.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Estimated production losses attributable to brucellosis per infected animal per year, by production system (pastoral and non-pastoral), species (cattle, sheep, and goats), and model scenario (‘no increase in herd size’ and ‘unrestricted increase in herd size’).

White dots indicate the median, black diamonds the mean, black boxes the inter-quartile range (50% UI), and black lines the 2.5 and 97.5 percentiles (95% UI). Uncertainty intervals were obtained from Monte Carlo simulation (50,000 iterations). The values in the y-axis are on the logarithmic scale. Int. $: International dollars.

https://doi.org/10.1371/journal.pntd.0012814.g004

Under the model scenario of ‘unrestricted increase in herd size’ (right panel in Fig 4), the median (95% UI) annual losses attributable to brucellosis in cattle, sheep, and goats, per infected animal, were 66.0 (21.1–329.2), 3.6 (0.5–29.2), and 6.9 (1.9–33.2) int. $ in the pastoral system, and 65.9 (22.3–445.0), 4.2 (1.3–10.8), and 4.0 (1.3–9.8) int. $ in the non-pastoral system, respectively. The difference in losses between cattle and the other species was slightly greater in the model scenario of ‘unrestricted increase in herd size’, compared to the scenario of ‘no increase in herd size’.

The estimated production losses per household, by species, production system, and model scenario, are shown in Fig 5. The household-level losses attributable to brucellosis in the pastoral system (101.4 int. $ per year) were equivalent to 4.4% (95% UI: 2.1–8.8) of the median annual, livestock-derived, household income expected in the absence of the disease (approximately 2,265 int. $; model scenario of ‘no increase in herd size’). In the non-pastoral system, the household-level losses attributable to brucellosis (4.3 int. $ per year) were equivalent to 0.6% (95% UI: 0.2–1.6) (near eight-fold lower than in the pastoral system) of the median annual, livestock-derived, household income expected in the absence of the disease (approximately 755 int. $; model scenario of ‘no increase in herd size’). Cattle contributed the most to household-level losses in both production systems and model scenarios. Sheep and goats jointly contributed around one-tenth and less than one-fourth of the household-level losses in pastoral and non-pastoral systems, respectively. In cattle, sheep, and goats, each abortion represented a median loss of 1,054.6 (1,034.6–1,074.6), 142.8 (141.6–144.0), and 134.5 (133.5–135.6) int. $ in the pastoral system, and 1,599.8 (1,574.9–1,624.7), 133.7 (132.5–134.9), and 106.8 (105.9–107.7) int. $ in the non-pastoral system, respectively.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Estimated production losses attributable to brucellosis per household per year, by production system (pastoral and non-pastoral), species (cattle, sheep, and goats), and model scenario (‘no increase in herd size’ and ‘unrestricted increase in herd size’).

The numbers within each bar or connected to it with a line segment indicate the median household-level losses attributable to brucellosis in each species. Estimates of household-level losses assume a herd size of 24 cattle, 20 sheep, and 38 goats in the pastoral system, and seven cattle, four sheep, and seven goats in the non-pastoral system. Int. $: International dollars.

https://doi.org/10.1371/journal.pntd.0012814.g005

The annual, region-level losses attributable to brucellosis in northern and central Tanzania are shown in Fig 6. Manyara was the region with the highest estimated production losses per year (5.7 million (M.) int. $), followed by Arusha (3.8 M. int. $) and Dodoma (3.1 M. int. $). These three regions were those with the largest population of livestock kept by pastoral households, as indicated in Table F in Supporting information (S3 File). Kilimanjaro was the region with the lowest estimated production losses, at 0.6 M. int. $ per year. In this region, livestock were kept predominantly by non-pastoral households.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Choropleth map of the median losses attributable to brucellosis in the study area, northern and central Tanzania, per year, by region.

The estimates of production losses were obtained from the reference model scenario (‘no increase in herd size’). Solid lines delineate regional and national borders. M. int. $: million international dollars. The inset at the upper right corner shows central and southeastern parts of Africa, with the study area highlighted in red. This figure was made in R [46] with packages ‘ggplot2’, ‘raster’, ‘sf’, and ‘rgdal’ [44,45,51,55,56]. The polygon shapefiles were obtained from ‘GADM’, the Database of Global Administrative Areas (https://gadm.org/). The base map was obtained from ‘Natural Earth’: ‘Cross Blended Hypso with Shaded Relief and Water’ raster, version 2.0.0 (www.naturalearthdata.com).

https://doi.org/10.1371/journal.pntd.0012814.g006

Determinants of production losses attributable to brucellosis

The results of the elasticity analyses of production losses per infected animal per year, under the model scenario of ‘no increase in herd size’, are shown in Fig 7. The parameters for which variation affected the estimated production losses per infected animal the most were the wholesale price and the abortion probability. This was consistent for all species (cattle, sheep, and goats) and both production systems (pastoral and non-pastoral). The next most influential variables were related to milk value (milk yield, lactation length, milk price, and milk harvest) in cattle and litter size in sheep and goats. Further results from elasticity analyses are shown in Supporting information (at household level under the model scenario of ‘no increase in herd size’, and at the individual and household levels under the scenario of ‘unrestricted increase in herd size’; Figs A-C; S4 File).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. Tornado plot showing the parameter elasticity of production losses per infected animal (percentage change in production losses per infected animal per year when parameter values change by 1%), by production system (pastoral and non-pastoral) and species (cattle, sheep, and goats) (50,000 model iterations, scenario of ‘no increase in herd size’).

The colours of the bars indicate the significance of the parameter elasticity estimate (dark and light grey indicate p-value  <  0.05 and p-value  ≥  0.05, respectively).

https://doi.org/10.1371/journal.pntd.0012814.g007