missing values

[mintbit]

Hi everyone. 

I am wondering if I have understood a problem correctly.

I have missing values in my predictor and outcome. I have even more missing in control variables. 

SAS only excludes variables included in step so in the first step (predictor/outcome) the ones with missing values on control variables are included. Then in next step (Predictor+Control/Outcome) SAS excludes more variables. 

I suspect this is wrong procedure, and I have to exclude/impute all observations with missing values prior to any regression steps?

[ballardw]

Code? or better log entry to show what you are doing. Copy the log of all the steps you are concerned with including the code and all messages, notes or warnings and paste into a code box opened on the forum to maintain legibility and preserve formatting of any diagnostic information.

Not enough detail to tell what you are attempting or which tools may help.

Many of the modeling procedures that use categorical variables have a CLASS statement and you can usually specify that missing is a valid level for those.

[mintbit]

Sorry for beeing imprecise. I'll try to explaing more in detail.
Step1: My main analysis including predictors and outcome have 150 missing cases wich I have deleted and then run regression.
Step 2: The next step wich includes a confounder has additionally 200 missing cases. Now, I want to exclude these as well, and the program does so automatically.

My question is it wrong to do step 1 with more cases than step 2, or should I exclude cases with missing confounder before doing Step1?

[PaigeMiller]

There was a very specific and clear request for you to show us your code.

[mintbit]

 
 1          OPTIONS NONOTES NOSTIMER NOSOURCE NOSYNTAXCHECK;
 70         
 71         
 72         title "Step 1";
 73         proc logistic data=u.mf plots (only)= (effect oddsratio);
 74         class ses (ref='1') kjønn klasse /param=ref;
 75         model &dep(event='1')=&v/ clodds=pl;
 76         run;
 
 NOTE: PROC LOGISTIC is modeling the probability that akt_med=1.
 NOTE: Convergence criterion (GCONV=1E-8) satisfied.
 NOTE: There were 3463 observations read from the data set U.MF.
 NOTE: PROCEDURE LOGISTIC used (Total process time):
       real time           0.43 seconds
       user cpu time       0.28 seconds
       system cpu time     0.03 seconds
       memory              22273.12k
       OS Memory           50544.00k
       Timestamp           19.09.2020 10:43:41 PM
       Step Count                        91  Switch Count  6
       Page Faults                       0
       Page Reclaims                     4889
       Page Swaps                        0
       Voluntary Context Switches        540
       Involuntary Context Switches      1
       Block Input Operations            0
       Block Output Operations           1064
       
 
 77         
 78         
 79         title "Step 2";
 80         proc logistic data=u.mf plots (only)= (effect oddsratio);
 81         class ses (ref='1') kjønn klasse
 82         nærtilb2 nærtilb3  /param=ref;
 83         model &dep(event='1')=&v nærtilb2 nærtilb3/ clodds=pl ;
 84         run;
 
 NOTE: PROC LOGISTIC is modeling the probability that akt_med=1.
 NOTE: Convergence criterion (GCONV=1E-8) satisfied.
 NOTE: There were 3463 observations read from the data set U.MF.
 NOTE: PROCEDURE LOGISTIC used (Total process time):
       real time           0.49 seconds
       user cpu time       0.35 seconds
       system cpu time     0.02 seconds
       memory              8719.37k
       OS Memory           52268.00k
       Timestamp           19.09.2020 10:43:41 PM
       Step Count                        92  Switch Count  6
       Page Faults                       0
       Page Reclaims                     1406
       Page Swaps                        0
       Voluntary Context Switches        571
       Involuntary Context Switches      1
       Block Input Operations            0
       Block Output Operations           768
       
 
 85         
 86         
 87         OPTIONS NONOTES NOSTIMER NOSOURCE NOSYNTAXCHECK;
 98         

 
title "Step 1"; 
proc logistic data=u.mf plots (only)= (effect oddsratio); 
	class ses (ref='1') kjønn klasse /param=ref; 
	model &dep(event='1')=&v/ clodds=pl; 
run; 
 
 
title "Step 2"; 
proc logistic data=u.mf plots (only)= (effect oddsratio); 
	class ses (ref='1') kjønn klasse  
	nærtilb2 nærtilb3  /param=ref; 
	model &dep(event='1')=&v nærtilb2 nærtilb3/ clodds=pl ; 
run; 
 

[PaigeMiller]

---

@mintbit wrote:

Hi everyone. 

I am wondering if I have understood a problem correctly.

I have missing values in my predictor and outcome. I have even more missing in control variables. 

SAS only excludes variables included in step so in the first step (predictor/outcome) the ones with missing values on control variables are included. Then in next step (Predictor+Control/Outcome) SAS excludes more variables. 

I suspect this is wrong procedure, and I have to exclude/impute all observations with missing values prior to any regression steps?

---

I do not understand why this is being done via two sequential logistic regressions. What do you gain from doing this as two separate regressions, instead of one combined regression? What does the first one tell you?

[mintbit]

I wanted to look at the effect with and without the confounder. Also for model selection.

[PaigeMiller · Accepted solution]

I think any comparison of models or comparison of coefficients has to be done on the same data, so it would be the data with all missings (predictors and confounders) removed.

[mintbit]

Thank you for helping me.