Banana boxes

• Summary: Evaluate approaches to calculate the total drug weight in a batch
• Developed by: Netherlands Forensics Institute, team Evidence Evaluation & Statistics, 2026
• Publication: submitted for publication (DOI-link will be added when available)

General background

Situation being modelled

• A batch of cardboard banana boxes has been seized. The bottoms of the boxes were impregnated with cocaine.
• The entire batch consists of 1000 boxes, of which 300 have been analysed for cocaine presence.
• From 5 cocaine-containing boxes, the weights of the bottoms have been measured (including cocaine).
• For 10 cardboard bottoms, the cocaine concentrations have been measured using 3 repeats per bottom.
• An estimate is required for the total weight of cocaine in the batch, together with an uncertainty interval.

Modelling approaches

This repository currently supports the evaluation of five modelling approaches:

• An existing frequentist approach, using weight and drug-presence measurements, based on a tool from ENFSI.
• An existing frequentist approach, using weight and drug-presence measurements, based on a tool from SWGDRUG.
• An existing frequentist approach, using weight and drug-presence measurements, based on Alberink et al. (2010).
• An existing frequentist approach, using weight and concentration measurements, based on Alberink et al. (2016).
• A novel Bayesian approach, using weight, drug-presence and concentration measurements, and using MCMC-simulations.

Using the repository

Setting things up

• Python version: 3.12
• Packages used: listed in requirements.txt (including version numbers)
• Configuration: the input_configurations folder contains all files hat define the simulations:
  • default_case.yaml: default values for all simulation settings and parameters
  • local_adjustments.yaml: values in here will override the values of all simulations being done.
    An example local_adjustments.yaml-example is provided, it can be renamed to local_adjustments.yaml.
  • To reproduce specific simulation results, change the random_seed: from empty/none to a 'date/time-stamp' number.
    This date/time-stamp can be found in the folder name of that specific simulation, for example 20260415035613.

Running simulations

There are two ways to run the simulations: via the Python command line, or using a file with a series of OS-commands.

Python command line

• Start in main.py
• By default a single simulation run is done, based on the values in default_case.yaml and local_adjustments.yaml.
• Several optional input arguments are supported:
  • -no_default_run: Do not run the default simulation defined in default_case.yaml.
  • -run_single_yaml, in combination with -yaml_file and path/to/file.yaml:
    Run the simulation specified in file.yaml, located in the folder path/to.
  • -run_all_yaml_files_in_folder, in combination with -folder_yaml_files and path/to/folder:
    Run the simulations specified by all yaml-files located in the folder path/to/folder.
  • -run_method_comparison:
    Run the simulations specified by all yaml-files located in the folder input_configurations/method_comparison.
  • -run_sensitivity_analysis:
    Run the simulations specified by all yaml-files located in the folder input_configurations/sensitivity_analysis.
  • -create_overview_results, in combination with -results_dir and path/to/folder:
    Make an overview table of all simulations results located in the sub-folders of the folder path/to/folder.
• The four -run_... arguments should not be combined; use only one of them at a time.
  • If any of them is used, -results_dir is overwritten with the output folder of that run-argument.
• Some example usages:
  • To do the sensitivity analysis including default run and summarising overview:
    python main.py -run_sensitivity_analysis -create_overview_results
  • To do the method comparison without default run and including summarising overview:
    python main.py -no_default_run -run_method_comparison -create_overview_results
  • To make a summarising overview of a selection previously run simulation results:
    python main.py -no_default_run -create_overview_results -results_dir results/my_result_selection/

Series of OS-commands

• Automatically clearing pytensor cache
  • When running a single yaml-file, the 'simulation time per run' slowly increases from the first to the last run.
  • This behaviour is caused by the growing size of the pytensor cache, related to the MCMC-calculations.
  • By default, this cache is cleared before the start of a simulation (clear_cache: True).
  • Clearing the cache is only possible before Python starts its calculation, not during the calculations.
  • To automatically run many simulation-yamls while clearing the cache in between, a series of OS-commands can be used.
• Examples files are included for two common types of Operating System (OS):
  • Windows: run_all_yamls.bat-example
  • Linux: run_all_yamls.sh-example
• To use such a file, do the following:
  • Copy the relevant file, and remove -example from the file extension
  • Update some settings in these files, for example:
    • Location of the Python interpreter to use
    • Location of folder(s) with simulation yaml-files to run
  • Execute the .bat or .sh file

Description of other folders and files

• tests: Folder with several files that contain multiple tests, to verify the implementations of the modelling approaches.
  The random nature of the MCMC-calculations may cause the tests in test_baysian.py to fail, depending on hardware.
• boundary_methods.py: Functions to calculate the relevant percentiles, based on weights, presences, and concentrations.
  • calculate_percentiles_enfsi: Using ENFSI approach: lower bound, best estimate, and upper bound.
  • calculate_percentile_swgdrug: Using SWGDRUG approach: lower bound.
  • calculate_percentiles_alberink2010: Using approach from Alberink et al. (2010): lower bound, best estimate, and upper bound
  • calculate_percentiles_alberink2016: Using approach from Alberink et al. (2016): lower bound, best estimate, and upper bound
  • calculate_percentiles_bayesian: Using novel Bayesian approach: lower bound, best estimate (median), upper bound.
• simulation_steps: Functions used to execute distinct parts of a simulation.
  • process_cfg: Process a configuration file (file.yaml) into input variables used for a simulation.
  • run_single_simulation: Do a single simulation run: generate ground truth, sample from it, apply methods.
  • aggregate_simulation_results: Summarise the results of a simulation, by aggregating over its runs.
  • save_aggregated_results: Save the aggregated results as csv-file including rounding, do some MCMC-checks.
• utils: Functions to perform various supporting tasks.
  • lower_bound_hyper: Determine lower bound for number of drug-containing units, using a hypergeometric distribution.
  • draw_normal_mcmc_samples: Draw samples from a normal distribution using MCMC-simulations.
  • get_dirs_per_measurement_types: Make lists of sub-folders with results, grouped by measurements types present.
  • create_overview_results: Make an overview table of all simulations results located in a specific folder.