Column quantiles or probabilites by ...

[#1]

I need jsl code that will save the probabilities of a given column distribution, by other columns.

I know how to do this un-scripted by taking a distribution of the column by the other columns and then under the red action button while holding control, and selecting Save Prob Scores.  But I don't know how to script this.

I looked through other posts but didn't find an answer to how to script the above or doing a more direct calculation.

The only thing I could think of is to do multiple iterations over each "by": taking the subset, sorting those tables, error checking for a null table, calculate estimated quantiles using row number (not sure how to get this), and updating the original table with each.  It seems like a lot of coding that maybe can be done with one line with the right function.

I'm sure someone has a beter idea.

[#2 · Accepted solution]

Hi, pcarroll1!

I think there are many ways to calculate and script what you wish.

Using a method from Blom G., Statistical estimates and transformed beta-variables, Wiley; New York: 1958, you could calculate the quantiles of, for instance, height by sex in the Big Class data table using the formula 

(Col Rank( :height, :sex ) - 3 / 8) / (Col Number( :height, :sex ) + 1 / 4)

I can think of several more sophisticated ways to do what I think you want, but making a new column with a formula (or Set Values) in it is pretty easy in JSL...

Good luck!

[#3]

Kevin,

    Thanks.  This works, though I don't understand the 3/8 and 1/4.  I've seen the equation as:

(Col Rank( :height, :sex ) ) / (Col Number( :height, :sex ) + 1)

Is this wrong?

The other issue is for me is constructing the equation for given column name inputs. 

The response I get from JMP are driving me crazy.  Forcing me several times to have to kill JMP with the Task Manager.

Pat

[#4 · Accepted solution]

Hi, Pat!

"Wrong" may be an overstatement!

The reference from Blom is specifically regarding Inverse Normal Transformations, which are of the form

The parenthetical argument is the quantile.  Blom recommends that c=3/8;  Van der Waerden recommends c=0; Bliss and others recommend different c's.

No less than John W. Tukey suggests in Tukey JW.; "The future of data analysis"; Ann Math Stat. 1962;33:pp1–67 that there is a trivial difference between any of the methods and the Expected Normal Scores, so it really doesn't matter which one you pick.  I chose Blom because I've studied it and I'm confident that it closely matches ENS even with small sample sizes.

Crazy is a short drive for me, but would you like some assistance with constructing the equation for given column name inputs? 

[#5]

Thanks for the offer. 

 I took a pill and then finally decided to construct my code as a string and then use eval(parse()) to convert it into an expression and run it.  In the end it worked!

Pat