A reader messaged me the following concern
hi ken, what would you do if there’s a massive contradiction between most of the suggestions generated by the general consensus and the suggestions generated by your new “from special studies” biomesight algorithm
The first thing that I want to point out is the warning on that page.
The main issue you may be seeing is in the selection of bacteria. With the regular selection, you focus on extreme values, i.e. top 10%, outside of standard lab ranges, outside of reference ranges from Jason Hawrelak and others. The amount outside of the reference range is used to give a weight to each bacteria for the importance of shifting. Different algorithms are used with different approaches (we do not know what the ideal one is).
With the special studies, we up-ended the algorithm. We picked the bacteria based on a simple “if the amount is above or below the reference norm+/- twice the standard deviation of the mean for the reference population and then use the z-score as the weight (the statistical significance for this bacteria)”.
This change means that a bacteria that is at the 70%ile may be included in the selection (which is very unlikely with the the first methods), and this bacteria could have a very high weight (which is based on statistical significance and NOT the difference from a mile post). A Bacteria at the 99%ile will be totally ignored if it is not statistically significant for the condition.
Statistically, I prefer the special studies approach because we are using the statistical significance of the bacteria for the significance/weight for suggestions instead of the naïve assuming that being high or low is the cause.
- We pick bacteria based on statistical significance for a specific condition and not whether they are high or low in general
- We give the bacteria a weight based on statistical significance for a specific condition and not the difference from a bound.
In theory, with the identical same bacteria and counts selected for two different conditions, you will get different suggestions because the weight assigned will be different since the weight is based on the statistical significance for the condition.
I well understand the confusion of some, the model being used is getting more advanced and handling more complexities.