Low cost and real-time surveillance of enteric infection and diarrhoeal disease using rapid diagnostic tests in Cox’s Bazar, Bangladesh | Conflict and Health

We aimed to include 400 households with a child between the age of 18 and 48 months in the study, with good dispersion across the area of the camp. We lacked a complete census for the camp that could identify eligible households. We therefore drew a sample from all households in the camp. Each sampled household was then visited and if they had a child under 5 years of age, we proceeded with the consent and interviewing process. The camp is divided into “blocks” and each residential location in each block is assigned a sequential “door number”, starting at one. We obtained the total number of households in each block and then sampled from the door numbers proportional to the block size. Based on previous work in the area we estimated that approximately one third of households would have a child under the age of five and the response rate would be close to 100%. We conducted two rounds of the survey in January–March 2021 (Round 1) and September–October 2021 (Round 2). We therefore sampled 1200 households to obtain a sample of 400 in round 1. We aimed to revisit the same participating households in round 2.

Survey and stool sampling

Household survey

At each participating household, we sought consent from the primary caregiver of the children under five. We then conducted a short survey capturing basic demographic and socioeconomic background data using the Open Data Kit software on tablet devices, including age, sex, level of education, time in Bangladesh, and water and sanitation facilities. We also captured the GPS location of the household in the camp.

Stool testing

A random child under five was sampled and the caregiver was provided with a plastic container with a barcode identifier for the child’s stool. The fieldworkers then returned the next day to collect the stool. A sample of the drinking water from the household was also collected from the container they normally used for testing the concentration of chlorine. The stool was taken to a field office in the camp. We used ProFlow tests produced by Pro-Lab Diagnostics, Wirral, United Kingdom for a set of eight pathogens listed in Table 1. Results from the tests were recorded in a survey form and linked to the household ID. The field worker also recorded whether the stool was diarrhoea or not.

Data and analysis

Sensitivity and specificity

As stated, we planned to account for the uncertainty due to the performance of the diagnostic tests in our statistical analysis. Table 1 provides estimates of test performance in terms of the sensitivity and specificity for each test based on values reported in the literature. All of the tests had high reported sensitivity, however specificity was more variable.

Covariates

We included two spatially-referenced covariates in our statistical model. First, we used the density of structures on the map (Fig. 1) as a proxy for population density. Second, we used the estimated level of water chlorination in parts per million (ppm). From each surveyed household’s water sample we tested the chlorine levels, these data were then smoothed over the area for each survey round using kernel density smoothing. Figure 2 shows the covariates within the boundary of the camp.

Inclusion of covariates can improve predictions and reduce uncertainty [3, 23]. We note that the parameters in geospatial statistical models may be biased and difficult to interpret [24, 25], and so we do not aim to provide inference on the “effects” of either of the included covariates beyond their relative comparisons of their magnitudes in predicting the outcome.

Spatially-referenced covariates in Camp 24: building density (top) and spatially-smoothed levels of chlorine in the drinking water supply in both survey rounds (bottom)

Statistical model

A technical description of the methods is provided in the Supplementary Information. In brief, we specified a binomial geospatial statistical model. For a location \(\:s\) in our area of interest at time \(\:t=\text{1,2}\) we observe the outcome of the test for person \(\:i=1,\dots\:,N\) as \(\:y\left({s}_{i},t\right)\) where:

For each location and time we define the linear predictor:

where \(\:x\left(s,t\right)\) are the spatially and temporally referenced covariates (Fig. 2) and \(\:Z\left(s,t\right)\) is a smooth latent process over the area of interest, which we describe below. If we ignore the diagnostic performance of the tests then the model would have \(\:p\left({s}_{i},t\right)={h}^{-1}\left(\mu\:\left({s}_{i},t\right)\right)\) where \(\:{h}^{-1}(.)\) is the inverse-logit function. We refer to this as the “uncorrected model”.

To take into account the sensitivity and specificity, the probability in the model should reflect the probability of testing positive, rather than the probability of having the disease. The test positive probability is:

where \(\:Sens\) is the sensitivity and \(\:Spec\) is the specificity. We refer to this as the “corrected model”.

Prior distributions

The standard geospatial statistical model formulation specifies a Gaussian process prior for the term \(\:Z(s,t)\). We use an accurate approximation to a Gaussian process prior to improve computational time and stability, given the number and complexity of models, with an exponential covariance function [26, 27]. We use an auto-regressive specification with a single autoregressive parameter \(\:\rho\:\) to allow for temporal correlation.

For the model parameters we specify weakly informative prior distributions, which provide a degree of regularisation and computational stability by limiting the parameters to a plausible range while not being informative within this range.

For the sensitivity and specificity we reviewed previous studies on the performance of RDTs for the different pathogens and specified Beta prior distributions on this basis. The distributions we used are reported in Table 1.