Proc surveylogistic and proc logistic

[Sas_User_TWu]

Hi

I wonder why running proc logistic without weight statement and proc surveylogistic without weight statement will have different SE/variance as both weight would be treated as default 1? E.G. code A vs code B below.

A. Proc logistic data=xyz;  /*ignore sample survey weight*/

    model Y=x1-x8/link=glogit;

  run;

B. Proc surveylogistic data=xyz; /*ignore sample survey weight*/

    model Y=x1-x8/link=glogit;

   run;

Also, what is the harm/negative consequences that I only apply proc logistic without weight to build a model against sample survey with known designed weight(variable) instead of apply proc surveylogistic together with weight statement, which is SAS document/procedure recommended way. E.G. code A vs code C below.

A. Proc logistic data=xyz;  /*ignore sample survey weight*/

    model Y=x1-x8/link=glogit;

  run;

C. Proc surveylogistic data=xyz;

    model Y=x1-x8/link=glogit;

    weight w;

run;

Thank you

TWU

[SAS_Rob]

This has to do with the manner in which the SE themselves are calculated.  SURVEYLOGISTIC uses (by default) a Taylor Series approximation which readily accomodates weights, strata and clusters.  SURVEYLOGISTIC is built for complex survey data so when you don't have any of the design effects adjusted for, there is no reason to use it.  The documentation has more details on the variance estimation.

https://go.documentation.sas.com/doc/en/pgmsascdc/9.4_3.4/statug/statug_surveylogistic_details27.htm...

LOGISTIC on the other hand uses maximum likelihood for the variance estimation.

https://go.documentation.sas.com/doc/en/pgmsascdc/9.4_3.4/statug/statug_logistic_details05.htm 

[Sas_User_TWu]

Thanks Rob,

Am I understand correctly that without weight statement, Taylor series approximation will end up with different variance/SE from fisher score method.

TWU

[SAS_Rob]

Yes, they can be different even when there is no weight variable.

[SAS_Rob · Accepted solution]

To answer your question regarding the harm, you will be biasing your results by ignoring any of the design effects.  To what extent they are bias can be somewhat quantified by calculating the design effect.

https://go.documentation.sas.com/doc/en/pgmsascdc/9.4_3.4/statug/statug_surveylogistic_details40.htm 

[Sas_User_TWu]

Great, thank you very much!

Regards

TWu