5 replies to this topic

[pdender]

when doing VLE / LLE/ SLE data regression with PRODE PROPERTIES utility

the procedure shows with results the relative errors (to input data),

these data fitting errors can be high for small fractions (say 1.0E-3 or below) ,

which impact do have these small fractions (and relative errors)  on data fitting procedure ?

Should I filter (i.e. remove) the small fractions (say < 1.0E-3) before to run the data regression utility ?

[PaoloPemi]

in PRODE PROPERTIES data regression utility

the errors are calculated as

 

Y(specified)-Y(calculated)

-------------------------------------

       Y(specified)

 

relative errors can be high for small fractions,

however I do not see problems to include even small fractions in the data set,

as the procedure attempts to fit all data points to find the best (weighted) fit.

 

However by removing small fractions you can improve data fitting accuracy

on other fractions so filtering may be useful in some cases.

[mrbabu]

thanks Paolo,

is that the way the errors reported in databank are calculated ? 

[pdender]

thanks,

as I see data regression procedure takes in account any data point

including small fractions where relative errors may become very large,

I presume one should evaluate all data points before to do data regression.

[staffel]

pdender wrote

"I presume one should evaluate all data points before to do data regression."

 

that is the usual way to proceed with data regression procedures,

data filtering is a important step to obtain a good model.

[chemdoc]

I suggest to follow this procedure

 

1) data filtering

2) data regression

3) binary txy or pxy plot to verify results

 

with models based on liquid activity you may be surprised to notice

how  calculated parameters show large variations depending

from a single point,

the step 3 (binary txy or pxy plot) is important to

verify results, I utilize tpxy.xls page in Excel

included with Prode

 tpxy.jpg   68.25KB   1 downloads