High exposure food substance combinations calculation
A full Monte Carlo analysis can be unwieldy for large cumulative assessment groups (CAGs) and/or large number of foods or concentration data. An algorithmic approach was developed to handle large CAGs. Two unique features of MCRA are:
contributions to the exposure results can be seen both in terms of foods as eaten (e.g. white bread) and foods as measured (e.g. wheat), and
a drill-down can be made into the exact foods and substances contributing for simulated individuals or individual-days in the upper tail.
The number of combinations of simulation, substance, modelled food and food as eaten can be very large. To avoid memory problems with very large datasets, an additional optional modelling step, named screening, was added to MCRA. Screening should be used if the data dimensions are too large for a direct analysis. Screening identifies risk drivers. A full analysis based on screened risk drivers will still retain all food/substance combinations in the exposure calculation, and will therefore produce exactly the same cumulative exposure distribution, and allow to see contributions of all substances and all modelled foods. Details with respect to foods as eaten are however restricted to the risk drivers selected in the screening step. For more details see screening calculation for large Cumulative Assessment Groups.
The two-step approach consists of:
- Step 1: Data screening and selection of risk drivers
Run a simple analysis for each potential source/substance combination (SCC). Here source means the combination of food as eaten and modelled food, for example apple in apple pie. The screening is based on this combination, and not just foods as measured, to avoid problems with potentially multi-modal consumption distributions as much as possible (see van der Voet et al. 2014). SCCs are also referred to as risk driver components. The screening step in MCRA implements a simple model that is applied to each SCC. The model calculates a percentile of interest in a distribution, consisting of a spike of zeroes (non-consumptions), and a mixture of two lognormal distributions for the exposure related to non-detects and positive concentrations, respectively. SCCs (risk driver components) can be combined to measured source/substance combinations (MSCCs, risk drivers). For example APPLE/apple juice/captan and APPLE/apple pie/captan combine to APPLE/captan. MCRA has an interface which identifies the Top-\(N\) SCCs (based on a chosen exposure percentile, e.g. p95) with an option to select \(N\) based on cumulative importance according to some criterion. Remark: Screening is performed before concentration modelling. Therefore there is no correction for processing at the screening stage. Note, originally SCC stands for Source Compound Combination, MSCC for Measured Source Compound Combination.
- Step 2: Full MC analysis
Perform the standard MC to all combinations of substances and foods, but restrict the stored information regarding foods as eaten to the SCCs selected in step 1.
The screening method requires to specify:
Which percentile to consider for each single source/substance combination (SCC, potential risk driver component) (default p95)
Which percentage of all exposures (according to the screening model) should be covered by the selected set of SCCs (default 95%)
How to impute non-detect concentrations (c < LOR) in the screening step. The choice of a factor 0 (default) represents optimistic imputation, the choice of a factor 1 represents a pessimistic imputation. It may be noted that a factor 1 (pessimistic imputation) may select many SCCs (risk driver components) with relatively high LORs and high RPFs, but with only nondetect measurements. Choosing a lower fraction, e.g. 0.25 can be useful if a more realistic method is sought.
Based on limited experience with the EFSA test data, useful settings of these three screening parameters were found to be (95, 95, 0) in preparation for an EFSA optimistic run, and (50, 95, 0.25) in preparation for an EFSA pessimistic run. See also screening calculation acute exposure and chronic exposure .