Logistic regression

[malena]

Good afternoon,

I recently took over running some logistic code from a previous programmer and I am note sure why a "freq" line is included before the class statement in the logistic procedure. Does anyone know what it does ? the probabilities I get are much higher when this line of code is included than when it is not.

proc logistic data=sponly DESCEND PLOTS(MAXPOINTS=NONE) ;
freq pending;

class

ageg4 (ref='5') /*ref=15-34*/
foreign (ref='1')
marital_status (ref='1') /*ref=Single*/
race (ref='2') /*ref=WNH*/
edu (ref='1') /*ref=High School or Less*/
gender (ref='2') /*ref=male*/
weekend (ref='0') /*ref=weekday*/
veteran_deaths(ref='2') /*ref=no veteran*/
;

model opioidtotal = ageg4 foreign marital_status race edu gender weekend veteran_deaths

/details rsquare ExpEst lackfit hierarchy=none CLPARM=Both CLODDS=Both /*CORRB CTABLE*/; roc;

** Model only for all knowns **;

where gender <> 9 and ageg4 <> 99 and race <>9 ;

** Create output data set for predictions **;

output OUT=PRED_SP p=_prob lower=_lower upper=_upper / /*alpha=.01 this will be for 99% CI*/;

run;

[Reeza]

FREQ is used when you have a single line of data representing multiple observations. 

So, it will definitely change your results.

The FREQ statement identifies a variable that contains the frequency of occurrence of each observation.

http://documentation.sas.com/?docsetId=statug&docsetTarget=statug_logistic_syntax18.htm&docsetVersio...

Two other things: 

1. Verify that <> is working as Not Equals, it actually means MAX in a SAS data step, but I'm not sure how it evaluates in a WHERE clause. Pretty sure it will be Not Equals, but it's worth a quick check.

2. On the CLASS statement I would expect PARAM=REF to be somewhere, but I don't see that, unless you do want GLM parameterization.

[malena]

thank you- then we should remove that line since the dataset only has one observation per person

[Reeza]

---

@malena wrote:

thank you- then we should remove that line since the dataset only has one observation per person

---

Then why do the results change if FREQ PENDING is included? The variable is obviously not all 1's. Is it 1/0 so it's a method to exclude pending observations or only considering pending cases?

[PaigeMiller]

---

@malena wrote:

thank you- then we should remove that line since the dataset only has one observation per person

---

Not necessarily, if the intent is to weight some observations more than others, or to exclude some observations which have a zero weight.

[WarrenKuhfeld]

To understand how a FREQ variable works, work through this example.  Compare the data sets X and X2 (the freq variable COUNT in X2 contains a count for each unique value).  Compare the two analyses.  They are identical.  Yes, you can use zero weights and freqs to exclude observations.

data x;
   do i = 1 to 200;
      x = ceil(uniform(7) * 10);
      y = ceil(uniform(7) * 10);
      output;
      end;
   run;
   
proc reg data=x;
   model y = x;
   quit;
   
proc freq data=x noprint;
   tables x * y / out=x2;
   run;
   
proc reg data=x2;
   freq count;
   model y = x;
   quit;