The authors identified four options for dealing with bimodality: (i) use all data to fit the SSD using a unimodal model (i.e. do not account for bimodality); (ii) use only the data from the most sensitive species; (iii) use all the data but assign greater weight to those values in the lefttail region (or alternatively, down-weight or censor more extreme values on the right); and (iv) use all data to fit the SSD using a statistical mixture model. The authors conclude “given enough data, bimodal distributions should be modelled using statistical mixture models.” The key phrase here is “given enough data.”

4 Responses

  • The ammonia example (see issue 18) shows clearly a high weight for the mixture even with only 8 data-points. This suggests that where the bi-modality is quite strong, the mixture model may prove useful even when data-sets are relatively small.

  • Yes, we agree and with time more and more data will become available and so more complex models will be required.

Comments are closed.