libname mydata "/folders/myfolders/";

11   ods listing close;ods html5 file=stdout options(bitmap_mode='inline') device=png; ods graphics on / outputfmt=png;
NOTE: Writing HTML5 Body file: STDOUT
12   
13   libname mydata "/folders/myfolders/";
NOTE: Libref MYDATA was successfully assigned as follows: 
      Engine:        V9 
      Physical Name: /folders/myfolders
14   ods html5 close;ods listing;

15   

proc print data=mydata.claim_history (obs=5);
run;

proc contents data=mydata.claim_history varnum;
run;

*-------------------------------------------------------------------------*;
*STEPWISE SELECTON MACRO FOR PROC GENMOD;
*-------------------------------------------------------------------------*;

*
The first part of the macro is the construction of all single factor
models and saving their results in a dataset (aggregated), where those
with p-values greater than 0.2 could be easily filtered out
;

************************** MACRO PARAMETERS *******************************;

*Put the name of the modeling dataset:;

%let dataset = mydata.claim_history;

*Put here the name of the name of the target varible;

%let target = CLM_AMT;

*Put here the name of all the potential explanatory variables, the
categorical and the coninous:;

%let explanatory =
GENDER
CAR_USE
EDUCATION
CAR_TYPE
MSTATUS
OCCUPATION
RED_CAR
URBANICITY
REVOKED
PARENT1
AGE
BIRTH
BLUEBOOK
CAR_AGE
HOMEKIDS
HOME_VAL
INCOME
KIDSDRIV
MVR_PTS
TIF
TRAVTIME
YOJ
;

*Put the name of the character variables with commas:;

%let char=
"GENDER"       ,
"CAR_USE"      ,
"EDUCATION"    ,
"CAR_TYPE"     ,
"MSTATUS"      ,
"OCCUPATION"   ,
"RED_CAR"      ,
"URBANICITY"   ,
"REVOKED"      ,
"PARENT1"
;

*Put the name of the character variables with commas:;

%let char_2=
GENDER       ,
CAR_USE      ,
EDUCATION    ,
CAR_TYPE     ,
MSTATUS      ,
OCCUPATION   ,
RED_CAR      ,
URBANICITY   ,
REVOKED      ,
PARENT1
;

*Put the name of the continous variables. Include the target variable also:;

%let interval =
CLM_AMT
AGE
BIRTH
BLUEBOOK
CAR_AGE
HOMEKIDS
HOME_VAL
INCOME
KIDSDRIV
MVR_PTS
TIF
TRAVTIME
YOJ
;

*In order to minimize the number of models that needed to be run, a stepwise
selection model was created, considering susceptible for entry all those
variables with a p-value in a single model less than 0.2, and with a p-value
in the aggregated model of less than 0.25;

%let single_model=0.2;
%let aggregated_model=0.05;

*Set the power Tweedie distribution parameter;

%let power = 1.5;

************************** END MACRO PARAMETERS ****************************;


/* Creation of the Aggregated Data Set*/


data explanatory;
    set /*tmp1*/
    &dataset
        (keep= &explanatory);
run;

proc contents data = explanatory varnum nodetails noprint
    out=explanatory_names (keep=name);
run;

data explanatory_names;
    set explanatory_names;
    j = _n_;
run;

* Determine the number of observations;
data _NULL_;
    if 0 then
        set explanatory_names nobs=n;
    call symputx('nrows',n);
    stop;
run;

proc datasets lib = work nolist;
    delete pva;
run;

options minoperator mlogic mprint;
%macro agrega / mindelimiter=',';
    %do obs = 1 %to &nrows;

data _null_;
    set explanatory_names;

    if j = &obs then
        call symputx("var", put(name, 30.));
run;

%if %upcase(&var) in(&char_2) %then
    %do;
        ods output Type3=pva&obs (rename=source=parm keep=source ProbChiSq);

        proc genmod data=&dataset NAMELEN=50;
            if _resp_ > 0 then
                d = 2*(_resp_*(_resp_**(1-&power)-_mean_**(1-&power))/
                      (1-&power)-(_resp_**(2-&power)-_mean_**(2-&power))/
                      (2-&power));
            else d = 2* _mean_**(2-&power)/(2-&power);
            variance var = _mean_**&power;
            deviance dev = d;
            class &var;
            model &target =  &var / link=log type3 scale=pearson;

            *scwgt expos;
            title "&var";
        run;

    %end;
%else
    %do;
        ods output Type3=pva&obs (rename=source=parm keep=source ProbChiSq);

        proc genmod data=&dataset NAMELEN=50;
            if _resp_ > 0 then
                d = 2*(_resp_*(_resp_**(1-&power)-_mean_**(1-&power))/
                      (1-&power)-(_resp_**(2-&power)-_mean_**(2-&power))/
                      (2-&power));
            else d = 2* _mean_**(2-&power)/(2-&power);
            variance var = _mean_**&power;
            deviance dev = d;
            model &target =  &var / link=log type3 scale=pearson;

            *scwgt expos;
            title "&var";
        run;

    %end;

data pva&obs;
    length parm $30;
    set pva&obs;
run;

proc append base=pva data=pva&obs force;
run;

    %end;
%mend agrega;

%agrega;

ods select all;

proc sort data=pva out=aggregate;
    by ProbChiSq;
run;

data elegibles;
    set aggregate;
    where probchisq le &single_model;
run;

proc sort data = elegibles;
    by probchisq;
run;

data elegibles;
    set elegibles;
    rename parm = source;
run;

title "Aggregate (all potential explanatory variables)";

proc print data=aggregate;
run;

title "Elegibles (individual model threshold = &single_model)";

proc print data=elegibles;
run;

/* STEPWISE PROCESS */

%let i = 1;

proc sql noprint;
    select count(source) into: totalfactor from elegibles;
quit;
* Also, the necessary set of macro variables needed to be initialized
as null, as the variables had to exist in order to be used in the rest
of the code;

%let catvar=;
%let invars=;
%let tempvar=;
%let tempcat=;

*options NOSYNTAXCHECK;/*Use this oiption before the last macro called
ONE only if you are sure that the failure of the macro is due to a lack
of convergence problem*/
options mlogic;

%macro one;
    %do %while (&i le &totalfactor);

data elegibles;
    set elegibles;

    if _n_ = 1 then
        do;
            call symput('testvar',source);

            if upcase(source) in (&char) then
                do;
                    call symput('catvar_t',source);
                end;
            else
                do;
                    call symput('catvar_t','');
                end;

            delete;
        end;
run;

title "Step &i";
title2 "Elegibles (aggregated model threshold = &aggregated_model)";
ods output parameterestimates = parms type3 = type3;

proc genmod data=&dataset NAMELEN=50;
    if _resp_ > 0 then
        d = 2*(_resp_*(_resp_**(1-&power)-_mean_**(1-&power))/
              (1-&power)-(_resp_**(2-&power)-_mean_**(2-&power))/
              (2-&power));
    else d = 2* _mean_**(2-&power)/(2-&power);
    variance var = _mean_**&power;
    deviance dev = d;
    class &catvar &catvar_t;
    model &target = &invars &testvar / link=log type3 scale=pearson;

    *scwgt expos;
run;

data stay;
    set type3;

    If source = "&testvar" and probchisq le &aggregated_model then
        do;
            call symput ('invars',"&tempvar"||" "||strip("&testvar"));
            call symput ('catvar',"&tempcat"||" "||strip("&catvar_t"));
        end;
    else if source = "&testvar" then
        delete;
run;

%let tempvar = &invars;
%let tempcat = &catvar;
%let i = %eval(&i+1);
    %end;
%mend one;

%one;

title "Selected Variables";
proc print data=stay;
run;


proc datasets lib=work kill nolist;
run;