You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I believe currently the intercept is initialized as zero (at least for logistic regression, only model I checked). Given we are often dealing with rare outcomes I think it would be better to use the logit of the outcome rate ( log(p / (1 - p)) ). In this case the model converges faster (as measured in fever iterations, see below) and seems to give the same results. I do think this could end up giving ~ 20-30% speed up for outcome rates of 1-3% which are common in our data. The example below for Eunomia gives a speedup of over 2x with outcome rate of 2%, but it's rather small data (n=1800, p=372). There's also a pure Cyclops reprex at the bottom giving about 30% fewer iterations but I had to change the simulateCyclopsData to give 1-3% outcome rates.
Of course I could supply startingCoefficients myself from the PLP to add this. But if there are no downsides maybe this should be the default in Cyclops ? What do you think ?
Hi @msuchard ,
I believe currently the intercept is initialized as zero (at least for logistic regression, only model I checked). Given we are often dealing with rare outcomes I think it would be better to use the
logit
of the outcome rate (log(p / (1 - p))
). In this case the model converges faster (as measured in fever iterations, see below) and seems to give the same results. I do think this could end up giving ~ 20-30% speed up for outcome rates of 1-3% which are common in our data. The example below for Eunomia gives a speedup of over 2x with outcome rate of 2%, but it's rather small data (n=1800, p=372). There's also a pureCyclops
reprex at the bottom giving about 30% fewer iterations but I had to change thesimulateCyclopsData
to give 1-3% outcome rates.Of course I could supply
startingCoefficients
myself from the PLP to add this. But if there are no downsides maybe this should be the default inCyclops
? What do you think ?reprex:
Created on 2025-01-29 with reprex v2.1.1
Pure
Cyclops
reprex:Created on 2025-01-29 with reprex v2.1.1
The text was updated successfully, but these errors were encountered: