Uncertainty analysesΒΆ

In order to account for uncertainties in the input data, assumptions about the business-as-usual scenario, the effects of interventions, etc, we can run many model simulations and vary the input data. In each simulation we randomly draw values for each input parameter from some probability distribution (e.g., a normal distribution, where we set the mean to the input value used in the BAU, and define the standard deviation). Accordingly, each simulation will generate different intervention effects. We then define the 95% uncertainty interval for each output (LYs, HALYs, LE, HALE) as the 2.5% and 97.5% percentiles of the values obtained over all of these simulations.

The basic process is:

  1. Identify rate(s) and/or value(s) for which uncertainties exist;

  2. Define a probability distribution to characterise the uncertainty for each rate/value.

  3. Identify whether the samples drawn from each distribution should be independent, or correlated in some way. For example, you may wish to correlate the samples for each rate across all cohorts (e.g., by age, sex, and ethnicity).

  4. Draw \(N\) samples for each of the rate(s) and/or value(s).

  5. Store these samples according to the same table structure as per the original data, with each sample represented as a separate row, and with one additional column ("draw") that identifies the draw number (\(1 \dots N\)).

This will result in a single, larger data artifact that contains all of the draws. In a model specification, you can then identify both the data artifact and the draw number, and when the simulation is run it will automatically select the correct values from all data tables that contain multiple draw.

See Uncertainty analyses for an example of running such an analysis.