proc logistic vs. proc surveylogistic

[sodonnell]

I ran a model using proc logisitic and proc surveylogistic and upon comparing the results I found the OR estimates to be same (as one would expect) however, the parameter estimates were different.

Upon taking the EXP of the parameter estimates from the proc logistic output I was able to reproduce the ORs but I was not able to do the same with the parameter estimates from the proc survey logistic output.  As a result, I believe the parameter estimates from proc logistic to be true.

Does anyone know why the parameter estimates would differ?

[Reeza]

Code for each proc?

[sodonnell]

proc surveylogistic  data=Mergedset_rev /*VARMETHOD=BRR;*/

class sex age2 marital edu_r Inc_Quin immigt aborig employ_4 cc_alt area smoke alco;

model sub2 (event='2')=sex age2 marital edu_r Inc_Quin immigt aborig employ_4 cc_alt area smoke alco/EXPB clodds alpha=.05;

ods output CLOddsWald=OR ParameterEstimates=parm;

weight wts_s;

/*REPWEIGHTS bsw1-bsw500;*/

run;

proc logistic descending data=Mergedset_rev;

model sub2= female _15_29 _30_49 w_s_d single le_sec_r sec_r so_pos_r inc_q1 inc_q2 inc_q3 inc_q4 imm abor no_work students onetw_ch three_ch daily occfor alcoh rural;

estimate 'female' female 1/ exp;

estimate '_15_29' _15_29 1/exp;

estimate '_30_49' _30_49 1/exp;

estimate 'w_s_d' w_s_d 1/ exp;

estimate 'single' single 1/exp;

estimate 'le_sec_r' le_sec_r 1/ exp;

estimate 'sec_r' sec_r 1/ exp;

estimate 'so_pos_r' so_pos_r 1/ exp;

estimate 'inc_q1' inc_q1 1/ exp;

estimate 'inc_q2' inc_q2 1/ exp;

estimate 'inc_q3' inc_q3 1/ exp;

estimate 'inc_q4' inc_q4 1/ exp;

estimate 'imm' imm 1/ exp;

estimate 'abor' abor 1/ exp;

estimate 'no_work' no_work 1/ exp;

estimate 'students' students 1/ exp;

estimate 'onetw_ch' onetw_ch 1/ exp;

estimate 'three_ch' three_ch 1/ exp;

estimate 'daily' daily 1/ exp;

estimate 'occfor' occfor 1/ exp;

estimate 'alcoh' alcoh 1/ exp;

estimate 'rural' rural 1/ exp;

weight wts_s;

run;

[ballardw]

Since there is significant difference in the two model statements I wouldn't expect much to be the same...

[Reeza]

Assuming you've done everything else right, it might be the common mistake of not adding param=ref to the class statement in proc surveylogistic. This creates parameter estimates that via effect coding not referential coding which is probably what you're after.

Check the docs for further information. Also, if you're using weights you should be using surveylogistic, again check the docs for the reasons.