Closed thenineteen closed 3 years ago
in branch Bayesian-correction
pull #222 this allows inverse variance weighting by modelling the GIF parcellations as binomial random variables (and the marginals p(semiology) from SS and p(GIF) from TS data), as well as an equal weightings (mean) method.
variance of binomial proportions = p(1-p)/n
I discussed these methods with Chibueze, Tim Stone, Normal WIliams from UCL, and they all agreed should prefer the above (cf both marginals from all-data, or using the data-query results as the estimate of the marginals).
NB this is an approximation as the GIFs are not independent (e.g.if it involves amygdala, likely will involve hippocampus)
original issue on combining semiologies #8 used MinMaxScaler and softmax. MinMaxScaler is sensitive to range and doesn't account for variance. Softmax is only great for the top prediction.
After proportions were supported akin to probabilities #118, we can now combine semiologies in a statistically more sound manner by using proportions and inverse variance weighting as alluded to in #8