Interpreting Evidence

[andersrm]

I have a Bayesian network containing a continuous variable x discretized into a number of possible states.  I obtain evidence/an observation that gives the average value of x. No other information is available about the evidence.  Is it reasonable and defensible to model the observation as likelihood (or soft) evidence?  I realize that without additional information about the evidence, there's still an infinite number of ways to interpret it.  Any thoughts and/or references to handling this issue (in Hugin or in general) would be appreciated.

andersm

[Anders L Madsen]

In principle, you could argue that the average the of X is a function of X (could some how be modeled as a child of X in your model). Observing a child node produces a likelihood on the parent(s). From this point of view it makes sense to enter the evidence as likelihood.

Notice that the probability of the evidence can be difficult to interpret in the case of likelihood evidence.