Special Studies are a conceptual thought experiment. The logic is simple, identify the bacteria that are have major statistical significance from the reference. Then use these bacteria with the weight that each has being the z-score to generate suggestions.
The first reviews that I did using them had good results and agreement with my preferred trio to build consensus, namely:
This last week I have gotten several emails from people who got counter-indicated suggestions. I have verified that for their samples, it produces contrary suggestions.
Digging into the mathematics and fuzzy logic being used, I see several possible failure points that I want to slowly investigate. The top failure points are:
- Using the z-score for the weight to give for each desired shifts. A different formula may resolve it. Two candidate formula are:
- z-score * incidence of bacteria being seen
- z-score * function(bacteria count) — with many possible functions
- The z-score cut off is too low, I am using 5.0 at present, it may need to be raised to a higher value.
- The criteria for picking a bacteria to include may not be specific enough, so a lot of bacteria that are fine are included. This can result in excessive noise in the suggestions
- The data available for suggestions at the species levels that we are working are insufficient (and in some cases, may not exist). A lot of the species flagged are rarely seen in studies showing changes.
Use the suggestions generated with great caution. If they compliment the suggestions from the three preferred consensus methods listed above — good. If they contradict, keep to the original consensus method — I have been getting consistent report that they work. The special studies suggestions are getting inconsistent results.
A special study with a z-score below 6.6 is very suspect and should be ignored.