How differently do SAS and STATA deal with missing values when running regressions?

[lizzy28]

I use both SAS and STATA to run a log-linear regression with the same dataset. The coefficient magnitudes were somehow different. One of the variables in my dataset had 18% missing values. I was wondering whether it was because SAS applied imputation when running regression.

Anyone knows the difference between SAS and STATA in running regression for data with missing values?

Thanks a lot.

[ballardw · Accepted solution]

I don't know about STATA but SAS with the most of the regression procedures will remove any record from the analysis that has any of the model variables missing. The diagnostics of the procedure should tell you how many records were actually used.

[lizzy28]

Thanks, Ballard.

But when I run the data with missing values excluded, the result was different from that given by not excluding missing values on purpose. The key difference is that coefficient values were totally different.

I'm attaching my data in my thread. The way I used was:

proc reg data=temp;

  model lcost_all_adj=age_diag_gp1 age_diag_gp2 age_diag_gp4 age_diag_gp5 mhi_grp2 mhi_grp3 mhi_grp4 days_debr_grp2 days_debr_grp3 days_debr_grp4 flap

                  negpthp days_dswi_grp2 days_dswi_grp3 days_dswi_grp4 los_cs_grp2 los_cs_grp3 los_cs_grp4 los_cs_grp5 comorbid_grp2 comorbid_grp3

                  sepsis transf_bleedcomp;

run;

This gave me corrected total of 1198, as shown below:

Number of Observations Read	1198
Number of Observations Used	1198

		Analysis of Variance
Source	DF	Sum of Squares	Mean Square	F Value	Pr>F
Model	23	441.70924	19.20475	24.58	<.0001
Error	1174	917.16903	0.78123
Corrected Total	1197	1358.87827

However, when I particularly excluded the missing values, as below

proc reg data=temp;

  model lcost_all_adj=age_diag_gp1 age_diag_gp2 age_diag_gp4 age_diag_gp5 mhi_grp2 mhi_grp3 mhi_grp4 days_debr_grp2 days_debr_grp3 days_debr_grp4 flap

                  negpthp days_dswi_grp2 days_dswi_grp3 days_dswi_grp4 los_cs_grp2 los_cs_grp3 los_cs_grp4 los_cs_grp5 comorbid_grp2 comorbid_grp3

                  sepsis transf_bleedcomp;

  where mhi_ctg^=.;

run;

I have

Number of Observations Read	992
Number of Observations Used	992

Analysis of Variance
Source	DF	Sum of Squares	Mean Square	F Value	Pr > F
Model	23	381.081	16.56874	21.28	<.0001
Error	968	753.8316	0.77875
Corrected Total	991	1134.913

[ballardw]

Since your variable mhi_ctg does not appear as a model variable in the first code then those records were not filtered out. When you add it in the second then you are excluding records that have non-missing values for all of the model variables, looks like about 200 of them. I would expect to get different results with about one-fifth of the records excluded.

[lizzy28]

Sorry for the confusion. mhi_grp1-4 were derived from mhi_ctg as mutually exclusive dummy variables. As mhi_grp2-4 were included in both regressions, I believe that they were supposed to run the same data set. Thanks.

[lizzy28]

I figured out what the problem was. After I recoded mhi_ctg into four dummy variables mhi_grp1-4, mhi_grp1 was excluded from the regression, and thus the observations with missing values in mhi_grp1 were treated the same way as the ones taking 1 in the variable.

Thank you, Ballard and Paige!

[PaigeMiller]

I was wondering whether it was because SAS applied imputation when running regression.

SAS does not impute missing values in regression. It does not include observations with missing values among the model terms in the regression calculations.