Emax model using and comapring FOCE ELS/FOCE LB/QRPEM/IT2S_EM

Hi all,
 
I have a data set to fit the simple Emax model (time vs AUEC0-24, sort by ID). No missing data.  
model E=E0+EMAX*C/(EC50+C) E0 fixed to 0.
 
I used four method to fit the data and turned out the AICs for four method are all quite close (delta <1). The problem is the EC50 from FOCE ELS/FOCE LB/QRPEM are around 0.8, while the EC50 from IT2S_EM  is around 0.6.  Emax are all similar (around -19). Since the EC50 value will be critical for my next step, I wonder if there are any criteria to rationalize my choice between 0.6 and 0.8.

Thank you

Jo

Edited by joybaker, 22 September 2020 - 04:19 AM.

Hi Joy- I have a few concerns with your approach.

1) you say your end point is AUEC0_24 but you have samples for this for 0 to 12h.  maybe rather than an area under the effect curve you should use a raw value if you haven't already as the transfomation may be masking something.

2) you have fixed tvE0 to 0 however your data is much more variable than that for your baseline.

3) it's not really appropriate to compare algorithms by AIC. FOCE-LB is generally less precise than FOCE-ELS so would not bother testing both.

I think you need to refine your model itself in order to make a decison, althoguh you have described the general trend well, your individual fits tend to the population mean and don't respresent the observed data well. 

- did they all take the same dose?

- Are there any covariates you could possibly use? 

-  what about the PK data  ?

if you look at how much your baseline varies then worrying about whether the Ec50 is 0.6 or 0.8 is moot I think. look at the CV% on your EC50s., for QRPEM it is almost 120%

  Simon.

Edited by Simon Davis, 22 September 2020 - 04:01 PM.

Dear Simon,

Thank you for your reply.

Please regard "time" (0-12 h) as C and AUEC0-24 as E for the Emax model. AUEC0-24 has been calculated so can just take as it is, there won't be any 0-24 h in the column. The 0-12 h is for the x-axis as various duration. This is a slight modification of the original meaning of the model,but principle is the same.

The model has been assigned  in the guideline (no covariate or PK need to be considered),  and at time 0, the AUEC0-24 supposed to be zero, so E0 = 0 or it can just be waived as without baseline. Therefore,  even the data is noise, the frame has been fixed. So my question is more like "based on all the bad cards  already and have to be in my hand, how can I identify which is less worse"?

Based on your answer, how can I compare algorithm for the same model, as parameters do  vary. And could you elaborate more on " recose"?

The guideline is attached if you feel confused.

Thank you,

Jo

Edited by Simon Davis, 22 September 2020 - 03:55 PM.

Currently your Emax models aren't really showing anything becuase Eta shrinkage ishuge even when E0 is frozen, you might be better off trying an indirect model.

I still think there is an issue with your data transformation beforehand since so much of the data at around baseline is above 0.

  Simon.