Our study is the first to use geospatial modelling to predict SAM prevalence at a 1-km spatial resolution from sparse survey data, and our work presents some critical findings. We estimate that approximately 6.3% (95% CI 4.2–10.9%) of all Papuan children under 2 years of age were experiencing SAM in late 2018. Based on the geostatistical analysis, there are areas within Papua that very likely experienced even higher levels of SAM. Importantly, we used a Bayesian framework to estimate our models which allowed for uncertainty in the predictions to be quantified. Producing estimates on a gridded surface also allowed the results to be easily visualized, providing the flexibility to aggregate the gridded estimates into any geographically defined unit, which might be useful from a policy or programmatic perspective. We demonstrated this step by aggregating the gridded predictions to the level of districts in Papua. The results of the geostatistical model predicted the proportion of Papuan children under 2 years of age experiencing SAM. By combining these estimates with gridded estimates of the population at risk, we were also able to predict the total number of children who experienced SAM. The use of proportions should be compared with predicted estimates (counts) as they may have different implications in terms of policy responses. For example, although the prevalence of SAM in Asmat district (14%) is higher than that in Mimika district (8%), the total number of children affected by SAM in Mimika (2007) was much higher than that in Asmat (568 children).
The use of exceedance probabilities to express uncertainty in predictions exceeding thresholds of SAM can be particularly useful from a policy-making standpoint. For instance, certain areas of the Papua province were likely to be in a critical situation, with well over 15% of Papuan children under the age of two being severely acutely malnourished. This has important implications for malnutrition programming in Papua to target those most in need. Our analyses highlight that significant advances in addressing malnutrition are required in the province if it is to meet the WHO Global Nutrition Target (GNT) to reduce wasting prevalence to less than 5% or the UN SDG to end all forms of hunger and malnutrition by 2030 [22].
Limitations
With reference to this particular study, the analysis has some limitations related to its source datasets. First, the distribution of the sample locations is not ideal for geostatistical modelling methods. Geostatistical models draw strength from spatial distribution of the sample sites and the assumption that areas near to observed samples are more similar. However, in this case study, the primary sampling units are located in a small number of districts, leading to a low spread of observations across the study area. That leaves large parts of the study area to be predicted from distant data points, which may lead to higher uncertainty in the predictions and it limits our ability to validate the outputs in these areas. Moreover, sites in close proximity (< 1 km) to one another may be coded as experiencing the same or very similar geospatial covariate values (due to the spatial resolution of the covariates). Therefore, it is difficult for the covariates to explain the observed differences in SAM among these clusters. The pattern of cluster locations may have affected the parameter estimates for the covariates and result is more uncertainty in the gridded predictions. Finer resolution covariates could be explored to account for this, though the predictions become more computationally challenging.
In addition, the source data are not representative in the same way that a national survey, such as the DHS or National Nutrition Surveys, would be. The baseline survey data were sampled from households in Papua where the caregiver self-identified as being of indigenous Papuan ethnicity [8]. We used this sample to examine the geographic variation in SAM, therefore our predicted SAM risk is most representative of that population of children in Papua. In the absence of estimates of the indigenous Papuan population, we used total population estimates to approximate the population distribution. If children of different ethnic groups in Papua experience higher (or lower) rates of SAM, then our estimates of the absolute number of children who were SAM—which rely on an estimate of the total population—could be under- (or over-) estimated. Future studies are needed to understand the distribution of different population groups in Papua and their risk of malnutrition.
Additionally, treatment districts for the Child Grant (BANGGA Papua) were specifically targeted and selected from the poorest districts in the province. We did not explicitly model this characteristic of the sample, but these factors were taken into account by controlling for accessibility and local context so that predictions in un-sampled areas were as accurate as possible. With only one set of source data for the Papua analysis, however, validation options for modelling were also limited. Cross-validation was employed to evaluate out-of-sample precision.
It should also be noted that SAM is a relatively fast-moving indicator of malnutrition. In this regard, SAM or wasting reflects acute or short-term malnutrition, while stunting reflects chronic or long-term malnutrition [23] and while we can predict SAM at one point in time, the prevalence of this indicator might have changed soon after measurement. Cross-sectional surveys, as used in this study, may not fully capture the fast-moving and changing risk of SAM. More waves of data at shorter time intervals (e.g. multiple times per season) could help identify “hot spots” with a consistently higher risk of SAM.
Policy relevance and benefits
The use of modelling methods to combine geospatial data with sparse geolocated survey data to predict health outcomes at high resolution or into unsampled areas offers many potential benefits in planning programs and monitoring progress toward government targets and the SDGs.
As noted earlier, many areas within Papua are very remote and cannot be accessed securely due to outbreaks of violent conflict, making ground-level data collection expensive or impossible [8]. Our findings suggest that this approach could provide real benefits in similar contexts where data collection is not possible or traditional surveys might experience gaps in coverage, such as remote areas, conflict-affected states, or areas with security concerns.
While the baseline data for Papua covered only six districts, these modelling techniques enabled us to predict SAM prevalence for the entire province – including districts that were not initially included in the baseline survey. Some of these districts also had high predicted levels of SAM, illustrating how this approach enabled us to identify SAM hot-spots that could be targeted with interventions that are known to work when tackling child undernutrition, such as for example the WHO-recommended approach of community management of acute malnutrition (CMAM) and ready-to-use therapeutic foods (RUTF) in community settings [24, 25].