Leaderboard

I was intrigued by the original question on this post: “What is the Statistical Probability that all BF Sightings are False?â€, and explored it further using a different approach. You don’t have to calculate the probability that all sightings are false in all of North America; instead you can focus at the probability results from a sample region (say PNW). Probability analysis on the data from the PNW area should give a good indication of the reality of the phenomenon overall. Also, the way to look at results is not simply bi-modal (True or False); there is always the 3rd choice of Inconclusive. To explore this problem and its key assumptions, I built a simple Bayesian Belief Network model (see attached file with description and results). The model could represent a particular region like PNW. Instead of hundreds of reports, I only modeled 6 BF reports and I added them incrementally. Only reports with “Excellent†Investigation were considered. “Excellent†Investigation defined as the best practices used in BF research (field investigations with good resources). Another difference in this approach is that a BBN model looks at all the Hypotheses being proposed (I only modeled 3) and asks if Hypothesis “X†is true, then what kind of reports (evidence) we would expect. Then, we look at the reports and ask, given this type of report conclusion, what is the most likely Hypothesis. I tried what Parnassus suggested, and added “False†reports incrementally to see how the probability of the BF hypothesis changes. The results agreed with what Parnassus concluded – that adding incremental “False†(negative on BF) reports reduces the probability that BF is real continuously. However, not all conclusions on BF reports are “Falseâ€. A “True†conclusion on a BF report does not mean a body in bag; it just means that after an “Excellent†investigation of a report, the experts concluded that the report indicated a real BF. The model also suggests that if all the report conclusions are “Inconclusive†or a 50/50 mixture of “True†and “False†conclusions, then the most likely outcome is inconclusive. We end up with the same a priori probabilities where we started. I believe others in this Forum also had a similar conclusion. The key assumptions in this model were the Conditional Probability Table and the distribution of “Trueâ€, “False†and “Inconclusive†reports. I am sure each of you would have a different set of assumptions. In academia/industry, we would formally interview Subject Matter Experts (SME) to arrive at theses assessments. I am not sure who in BF research would be able to provide non-biased assessments. Nonetheless, I believe a BBN model like this would be helpful if we had good unbiased assessments associated with a small region of North America. BF Probability using Bayesian Network Model.pdf