proc genmod

[desireatem]

I do not know what is wrong with my code:

Below is the distribution of deposi:

deposi	Frequency	Percent	Cumulative Frequency	Cumulative Percent
0	498	59.50	498	59.50
1	239	28.55	737	88.05
2	100	11.95	837	100.00

I fitted the SAS procedure below and had the error

proc genmod data=stroke;
class gender(ref="0") race_ethnic(ref="0") deposi(ref="0") pay_source(ref="0") cmg_tier(ref="a") ;
model mrs_discharge= age gender race_ethnic deposi pay_source cmg_tier report_cmg weight_cmg expected_los/dist=gamma link=log type3;
run;

ERROR: Invalid reference value for deposi.
ERROR: No valid observations due to invalid or missing values in the response, explanatory, offset,
frequency, or weight variable.
NOTE: The SAS System stopped processing this step because of errors.
NOTE: PROCEDURE GENMOD used (Total process time):
real time 0.01 seconds
cpu time 0.01 seconds

[PaigeMiller]

Whenever you get an error in the log, SHOW US THE LOG. That is, show us the entire log, 100% of it, with nothing chopped out.

Do not show us error messages disconnected from the code. Do not pick and choose parts  of the log to show us and then not show us other parts.

[desireatem]

1204 proc genmod data=stroke;
1205 class gender(ref="0") ert(ref="0") race_ethnic(ref="0") deposi(ref="0") pay_source(ref="0")
1205! cmg_tier(ref="a") ;
1206 model mrs_discharge= age gender race_ethnic deposi ert*deposi pay_source cmg_tier report_cmg
1206! weight_cmg expected_los/dist=gamma link=log type3;
1207 run;

ERROR: Invalid reference value for deposi.
ERROR: No valid observations due to invalid or missing values in the response, explanatory, offset,
frequency, or weight variable.
NOTE: The SAS System stopped processing this step because of errors.
NOTE: PROCEDURE GENMOD used (Total process time):
real time 0.00 seconds
cpu time 0.01 seconds

[Reeza]

One of the variables in the model has too many missing values.
See how you can figure that out here:
https://blogs.sas.com/content/iml/2016/04/18/patterns-of-missing-data-in-sas.html

[desireatem]

Thank you but there is less than 10 % missing.

[ballardw]

Any chance you typed a capitol "o" instead of zero (or vice versa).

What is the format assigned to that variable?

If that variable is character you may also need to check to see if the values have leading blanks. Most of the output tables will justify the text hiding the leading space but the reference value wants the actual level.

Or perhaps just use the keyword First instead of a level .

To see what I am talking about run the code below:

data example;
   x="    0";
run;

proc freq data=example;
   tables x;
run;

The single value of X actually has several leading blanks. In the output for Proc Freq the blanks have been "justified" away and not visible.

[FreelanceReinh]

Hello @desireatem,

To check your data for all three potential issues mentioned by ballardw and Reeza, i.e.,

• unformatted or inappropriately formatted reference level for formatted variable deposi
• hidden character(s) in the reference value of variable deposi or its format label (leading or trailing blanks would not matter, though)
• missing values of one or more other predictor variables in the observations with deposi=reference value (missing values in the dependent variable are largely tolerated),

you could run this DATA step and post the log, in particular the values of _c1, _c2, ..., if any:

data _null_;
array _c[5] $16 (5*' ');
set stroke end=last;
where cmiss(age, gender, race_ethnic, pay_source, cmg_tier, report_cmg, weight_cmg, expected_los)=0;
if vvalue(deposi) ~in: _c & i<dim(_c) then do;
  i+1;
  _c[i]=vvalue(deposi);
end;
if last;
put (_c[*])(=$hex32. /);
run;

[desireatem]

366 data _null_;
367 array _c[5] $16 (5*' ');
368 set stroke end=last;
369 where cmiss(age, gender, race_ethnic, pay_source, cmg_tier, report_cmg, weight_cmg,
369! expected_los)=0;
370 if vvalue(deposi) ~in: _c & i<dim(_c) then do;
371 i+1;
372 _c[i]=vvalue(deposi);
373 end;
374 if last;
375 put (_c[*])(=$hex32. /);
376 run;

_c1=20202020202020202020203120202020
_c2=20202020202020202020202020202020
_c3=20202020202020202020202020202020
_c4=20202020202020202020202020202020
_c5=20202020202020202020202020202020
NOTE: There were 230 observations read from the data set WORK.STROKE.
WHERE CMISS(age, gender, race_ethnic, pay_source, cmg_tier, report_cmg, weight_cmg,
expected_los)=0;
NOTE: DATA statement used (Total process time):
real time 0.01 seconds
cpu time 0.01 seconds

[FreelanceReinh · Accepted solution]

---

@desireatem wrote:

_c1=20202020202020202020203120202020
_c2=20202020202020202020202020202020
_c3=20202020202020202020202020202020
_c4=20202020202020202020202020202020
_c5=20202020202020202020202020202020
NOTE: There were 230 observations read from the data set WORK.STROKE.
WHERE CMISS(age, gender, race_ethnic, pay_source, cmg_tier, report_cmg, weight_cmg,
expected_los)=0;

---

Thanks for this helpful information. So, the third bullet point of my list of potential issues described what happened: Missing values of one or more other predictor variables in the observations with deposi=0 have made ref="0" an invalid reference level specification for deposi. The values of _c1, ..., _c5 indicate that all observations used in the analysis have deposi=1. Therefore, ref="1" is the only possible reference value for deposi. But even this is actually useless because the parameter estimate of the constant variable deposi will be zero, i.e., you can remove deposi from the CLASS and MODEL statements without losing more information.

Given that 72.5% of the 837 observations in work.stroke (with non-missing deposi) have been excluded because of missing values in age, gender, race_ethnic, pay_source, cmg_tier, report_cmg, weight_cmg and/or expected_los, you should check if

1. that many values are correctly missing
2. most of the missing values are from a single variable (or from only two, ...) -- Reeza's suggestion will be helpful for this
3. you can remove those largely missing variables from the model or impute their missing values.

Otherwise, you won't get a very useful model from PROC GENMOD as it would be based on a relatively small, not representative subset of your analysis dataset, disregarding information (e.g., about the relationship between deposi and mrs_discharge) which you could gain from the rest of the dataset.

[Ksharp]

As Reeza pointed out , you have too many variables be missing. which conduct to "deposi" doesn't have value 0 .