Inquiry Regarding Contradictory Results in Bayesian Network Analysis
I have encountered some puzzling results in my analysis. In my data, I observed a significant negative Pearson correlation between depression and several dimensions of mental health literacy. Yet, when I constructed a Bayesian network using JASP or the easybgm package in R with the package = "BDgraph" option, the edge between depression and one specific dimension—coping behaviors for mental illness—appeared as a positive association. This contradicts both theoretical expectations and the initial correlation analysis.😲
Interestingly, when I switched to the package = "BGGM" option within easybgm, the relationship reverted to a negative association, which is more interpretable in the context of my research.
I am uncertain which result to trust and wonder whether this discrepancy stems from differences in the underlying algorithms, prior specifications, or model assumptions between BDgraph and BGGM. Could anyone kindly offer any guidance on how to interpret such conflicting outcomes? Are there specific settings or diagnostic steps you would recommend to resolve this?
Comments
I'll ask some of our network experts.
From the team (with whom you are already in contact I believe):
"It is an issue in BDgraph, which can provide wrong results if the user does not supply mean-centered data (because BDgraph does not estimate the mean vector). This is now fixed in the new update of easybgm (which has just been submitted to CRAN) and will also be fixed in the next JASP update."