Establishing cutoffs for each indicator

To establish cutoffs for each indicator, there was a need to develop criteria for flagging values that would indicate a likely challenge associated with each indicator. In many applications, cutoffs are used to interpret continuous indicators, where a value on one side of the cutoff is diagnosed as problematic, while a value on the other side is diagnosed as acceptable. Because the severity of a condition is rarely tied to an exact value, but rather to a position of greater or lesser risk within a continuous range of values, setting cutoffs for diagnosis requires careful consideration. Each diagnostic indicator was categorized into three categories: green (unlikely challenge area), yellow (potential challenge area), or red (likely challenge area). Since different levels of evidence exist for each indicator, thresholds were established using four different methods, as follows. First, when possible, pre-defined cutoff values representative of global consensus on public health significance (such as pre-defined low to high categories for the prevalence of stunting in young children) were used (S1 Table). However, for most indicators, such pre-defined cutoff values do not currently exist. Second, where normative recommendations exist, these were used to establish cutoffs (S2 Table). For example, thresholds for fruit supply adequacy were based on globally recommended per capita intakes of fruit, with countries in the green category having a supply of fruit at or above the recommended intake and countries in the red category having a supply of less than half of the recommended amount. Third, where no cutoffs have been published and no normative values exist, the relative values of country data points can be compared as relatively higher or lower. For each indicator, density plots, a variation of histograms, were used to examine the distribution of data, using the data assembled on the FSD (S3 Table). A density plot was chosen over a histogram to view a smoothed distribution of the data using kernel density estimation. Most indicators had an approximately normal distribution and were divided into tertiles, rounded to interpretable values. We prioritized retaining meaningful or more easily interpretable cutoffs over exact tertiles. Fourth, some indicators had a bimodal or highly skewed distribution; in these cases, the peaks were bifurcated by the two cutoff points (low/medium; medium/high). An example of each of these is shown in Fig 2. The cutoffs for each indicator, as well as the method used to set them, are shown in Table 1.

Density plots of the data distribution for A) vegetable losses (% of domestic supply) and B) prevalence of adult obesity with colors indicating cutoffs for diagnostic thresholds: green indicates an unlikely challenge area, yellow indicates a potential challenge area, and red indicates a likely challenge area. Density plots are similar to histograms but use kernel density smoothing, rather than bins, to present a continuous distribution of the data. The peaks of each plot represent where the highest number of observations exists in the data.

Four example indicators are explained to demonstrate the methodology for determining the cut-offs. As mentioned above, the prevalence of stunting is an example of an indicator where cutoffs are based on published consensus on cutoffs [50]. An example of an indicator where cutoffs are based on normative recommendations is vegetable supply. This indicator is included as vegetable supply is a precursor of vegetable consumption; thus, the cutoffs are set based on the World Health Organization’s recommendation for vegetable consumption as part of a healthy diet. Vegetable losses, on the other hand, is an example of an indicator where no normative cutoffs or recommendations exist. Because the data for this indicator are normally distributed across countries, the cutoffs are set using rounded tertiles. The prevalence of adult obesity similarly has no published or accepted cutoffs for public health significance, but the distribution shows two large peaks, so bimodal curve-based binning is used to set cutoffs.