I have been using GeNIe for a long time, but relatively new to using it to learn parameters from data. I have a data set of about 15,000 observations for seven variables. The model structure is a grandchild, two parents and two sets of grandparents. I have discovered that I get a different conditional distribution for the child depending on whether I learn with just the two parents or include the grandparents, even though the network is structured so that the child is independent of the grandparents given the parents. Does anyone have insight has to why this would be happening? I have been setting the confidence level at 1 (it didn't seem to work when random was checked). The algorithm for the network is at the default of clustering. They are all multistate variables with as many as ten states. Any suggestions or advise on using GeNIe for parameter learning would be appreciated--especially what could account for the different distributions. I have checked that the states are ordered the same in both cases.
Thanks,
Alan