extended-add21-add35-v1

New data allowed us to include the LNLs V and VII into the model. Consequently, we ran some experiments to compare different graphs, connecting the tumor and the LNLs among each other.

To be able to compare them, we computed the BIC, the maximum likelihood, and the mean likelihood. Using thermodynamic integration, we reassured ourselves that - at least for the extended-base-v1 model - the negative log-evidence and the BIC/2 agreed well.

The investigated graph is show below:

Graph: Network of tumor and lymph node levels. Red arcs depict spread from tumor to the individual LNLs, which were considered in all models. Blue arcs represent spread probabilities from LNL to LNL that were also present in all investigated graphs. The green edges, however, were the ones we added one by one to compare the resulting model's performance.

Results

From the model comparison values tabulated below, we inferred that the model with additional arcs from LNL II to LNL I and from LNL III to V (extended-add21-add35-v1) was the best one for our purposes.

Model	- log-evidence	BIC/2	- max. llh	- mean llh
extended-base-v1	1212.3 ± 3.7	1212.0	1180.4	1184.7
extended-add12-v1		1211.7	1177.0	1180.8
extended-add21-v1		1208.2	1173.5	1177.9
extended-add35-v1		1212.4	1177.7	1182.1
extended-add45-v1		1213.6	1178.9	1183.1
extended-add27-v1		1214.9	1180.2	1184.4
extended-add21-add35-v1		1208.4	1170.5	1174.7
extended-add21-add45-v1		1209.7	1171.8	1176.0
extended-add21-add27-v1		1210.7	1172.8	1177.5