1# yakilili

/**********************************************************
Purpose: trim or winsorize SAS dataset to remove the impact from extreme values

Input
dsetin  : dataset to winsorize/trim
byvar   : define subset to winsorize/trim,e.g. 'date'. type 'none' for the whole dataset
type    : 'delete' or 'winsor' (delete will trim, winsor will winsorize
vars    : subsetting variables to winsorize/trim on; type 'none for no byvar
pctl    : the percenagte of left and right tails to trim/winsorize

Output
dsetout : dataset to output with winsorized/trimmed values
************************************************************/

%macro winsor(dsetin=, dsetout=, byvar=none, vars=, type=winsor, pctl=1 99);

%if &dsetout = %then %let dsetout = &dsetin;
   
%let varL=;
%let varH=;
%let xn=1;

%do %until ( %scan(&vars,&xn)= );
    %let token = %scan(&vars,&xn);
    %let varL = &varL &token.L;
    %let varH = &varH &token.H;
    %let xn=%EVAL(&xn + 1);
%end;

%let xn=%eval(&xn-1);

data xtemp;
    set &dsetin;
    run;

%if &byvar = none %then %do;

    data xtemp;
        set xtemp;
        xbyvar = 1;
        run;

    %let byvar = xbyvar;

%end;

proc sort data = xtemp;
    by &byvar;
    run;

proc univariate data = xtemp noprint;
    by &byvar;
    var &vars;
    output out = xtemp_pctl PCTLPTS = &pctl PCTLPRE = &vars PCTLNAME = L H;
    run;

data &dsetout;
    merge xtemp xtemp_pctl;
    by &byvar;
    array trimvars{&xn} &vars;
    array trimvarl{&xn} &varL;
    array trimvarh{&xn} &varH;

    do xi = 1 to dim(trimvars);

        %if &type = winsor %then %do;
            if not missing(trimvars{xi}) then do;
              if (trimvars{xi} < trimvarl{xi}) then trimvars{xi} = trimvarl{xi};
              if (trimvars{xi} > trimvarh{xi}) then trimvars{xi} = trimvarh{xi};
            end;
        %end;

        %else %do;
            if not missing(trimvars{xi}) then do;
              if (trimvars{xi} < trimvarl{xi}) then delete;
              if (trimvars{xi} > trimvarh{xi}) then delete;
            end;
        %end;

    end;
    drop &varL &varH xbyvar xi;
    run;

%mend winsor;

什么是winsorise 原理是啥？

Winsorize的原理：如果一个样本某变量的值大于该变量的99分位数，则该样本的值被强制指定为99分位数的值；类似的，如果一个样本某变量的值小于该变量的1分位数，则该样本该变量的值被强制指定为1分位数。

受教了！