Does SAS have function to get "Skewness" directly

[ZZB]

Hi all,

I know that SAS has functions to get 'mean' and 'std' directly, as the following codes:

/*Rolling 24-month window*/

proc sql;

create table roll as

select

     h2.cusip,

     h2.monthDate as periodEndDate format=yymmd7.,

     h2.ret,

     mean(h1.ret) as meanRet,

     std(h1.ret) as stdRet

from

     have as h1 inner join

     have as h2 on h1.cusip=h2.cusip and

          intck("MONTH", h1.monthDate, h2.monthDate) between 1 and 24

group by h2.cusip, h2.monthDate, h2.ret

having count(h2.cusip)=24;

quit;

How about the function for 'skewness' or 'idiosyncractic  skewness'?

Thanks

[Murray_Court]

I believe proc Univariate gives Skewness statistics for a given variable.

[jakarman]

This wat the SAS SQL supports: SAS(R) 9.4 SQL Procedure User's Guide (summary-function) this is standard SQL not some dialect.

The SAS procedures have there origin in Stats see: Base SAS(R) 9.4 Procedures Guide, Second Edition (Keywords and Formulas)

When you are coming form some RDBMS that have implemented statistics by grouping it are native dialects.

In those cases you could benefit that SAS will do a translation (as that shoudl word) for in-database processing.

[Reeza]

As indicated by Jaap, there isn't a SKEW function for the SAS SQL implementation.

Also, it seems weird to include your analysis variable (ret) in your Group BY statement.

You can use proc means or proc univariate to get the skewness.

proc sort data=sashelp.class out=class; by sex; run;

proc means data=class noprint;

by sex;

output out=summary mean(weight)=avg_ret std(weight)=std_ret skew(weight)=skew_ret;

run;

EDIT:  to supply working example

[ZZB]

I want to get the rolling skewness. Since there is no SKEW function in SQL, I have to think about other ways.

                              ret         meanRet  stdRet         SkewRet

36110	1985-01	0.16865	0.034400	0.083940
36110	1985-02	-0.02326	0.041494	0.087867
36110	1985-03	-0.10714	0.040525	0.088471
36110	1985-04	-0.10027	0.029867	0.090268
36110	1985-05	0.11940	0.019866	0.090858
36110	1985-06	0.01333	0.015292	0.082216
36110	1985-07	0.15868	0.014435	0.082121
36110	1985-08	0.03429	0.022072	0.086726

[PGStats]

No it doesn't. Back to basics: compute your own :

data test;

do x = 2,4,6,3,1;

  output;

  end;

run;

/* Expected : skewness(2,4,6,3,1) = 0.5901286564 */

proc sql;

select count(x) * (sum(x**3) -3*mean(x)*uss(x)+3*mean(x)**2*sum(x)-count(x)*mean(x)**3) /

  ((count(x)-1)*(count(x)-2)*STD(x)**3) as skewness format=14.10

from test;

quit;

PG

[Tom]

Use PROC MEANS or PROC UNIVARIATE if you want to calculate statistics.  You can define your current 24 month merge as a view if you want avoid creating a large intermediate table.

proc sql ;

create view roll24 as

  select h2.cusip

       , h2.monthDate as periodEndDate format=yymmd7.

       , h2.ret

    from have as h1 inner join have as h2

      on h1.cusip=h2.cusip

     and intck("MONTH", h1.monthDate, h2.monthDate) between 1 and 24

    group by h2.cusip, h2.monthDate, h2.ret

    having count(h2.cusip)=24

;

quit;

proc means data=roll24 ;

  by cuspid periodenddate ;

  var ret ;

  output out=roll mean=meanRet std=stdRet skew=skewRet ;

run;

[ZZB]

I tried your code. It does not seem to work.

[Tom]

The MEANS will calculate the statistics properly, you just need to fix the SQL so that it includes the values that you want to include in your calculations.  It seems dubious to me to group by RET and also include RET in the returned values?  Perhaps you want to group by H2.RET (so you can still see the actual value for that date) and also include H1.RET as another variable so that it cab be fed into PROC MEANS.  A small example of what you want would help (perhaps with 3 or 5 instead of 24 months so that the number of records is smaller).

data have ;

  do cusip=1,2 ;

    rec+1;

    baseDate = '01JAN2000'd ;

       do month=0 to 26 ;

        monthDate = intnx('month',baseDate,month);

        ret = ranuni(0);

               output;

       end;

  end;

run;

proc sql ;

create view roll24 as

  select h2.cusip

       , h2.monthDate as periodEndDate format=yymmd7.

       , h2.ret

       , h1.ret as lagret

    from have as h1

     , have as h2

    where h1.cusip=h2.cusip

      and intck("MONTH", h1.monthDate, h2.monthDate) between 1 and 24

    group by h2.cusip, h2.monthDate, h2.ret

  having count(h2.cusip)=24

;

quit;

proc means data=roll24 noprint ;

  by cusip periodenddate ret ;

  var lagret ;

  output out=roll mean=meanRet std=stdRet skew=skewRet ;

run;

proc print data=roll ;

run;

                period

Obs    cusip    EndDate      ret      _TYPE_    _FREQ_    meanRet     stdRet     skewRet

1       1      2002-01    0.22396       0        24      0.50622    0.29944    -0.26692

2       1      2002-02    0.23832       0        24      0.51477    0.28762    -0.21262

3       1      2002-03    0.64083       0        24      0.49023    0.28492    -0.04201

4       2      2002-01    0.59923       0        24      0.52899    0.28713     0.13578

5       2      2002-02    0.91479       0        24      0.52256    0.28359     0.19230

6       2      2002-03    0.41203       0        24      0.52664    0.28867     0.20478

[jakarman · Accepted solution]

ZZb is there a reason that you must code in SQL? Do you have a third party DBMS? When you have a third-party DBMS wich one? SQL aggregated functions are as standard as babylonian languages (old tst) for the advanced ones.
I your question the same as: https://communities.sas.com/message/192004