Create simulated data

[trosendal]

We want to test the performance of the model on data where we know the attribution answer. Therefore we will first generate simulated type distributions within sources and sample the humans from those.

1. Use the known type distributions within source to generate a probability vector for the source. All types must have a small non-zero probability of being sampled.

2. Sample from these to generate 'source data'

3. Assign sampling frequency to each source for the human population, Sample the humans from the sources

4. Run model. Determine if estimated sample attribution fractions agrees with actual sampling fraction.