topic Re: Regression models with restricted cubic splines in Statistical Procedures

Regression models with restricted cubic splines

BayzidurRahman — Wed, 17 Jul 2024 04:31:12 GMT

I need to fit a poisson model with offset including restricted cubic spline of a continuous covariate. I want to use the effect statement to automatically generate the spline variables at my preferred values. Can you please let me know what procedures can do it? I can see that proc genmod doesn’t allow select statement.

Re: Regression models with restricted cubic splines

sbxkoenk — Wed, 17 Jul 2024 11:07:54 GMT

Splines are available in many procedures when you use the EFFECT statement.

The following procedures provide an EFFECT statement: GEE, GENMOD, GLIMMIX, GLMSELECT, HPMIXED, LOGISTIC, ORTHOREG, PHREG, PLM, PLS, PROBIT, QUANTLIFE, QUANTREG, QUANTSELECT, ROBUSTREG, SURVEYLOGISTIC, and SURVEYREG.

GENMOD is in the list!

Go to the Google search-bar and submit :

effect statement spline site:https://blogs.sas.com/content/iml

BR, Koen

Re: Regression models with restricted cubic splines

sbxkoenk — Wed, 17 Jul 2024 11:10:44 GMT

Also, the TRANSREG procedure in SAS/STAT supports spline effects but has its own syntax.

You can enrich your dataset with new covariates (spline effects) using PROC TRANSREG and subsequently use the procedure of your choice.

Koen

Re: Regression models with restricted cubic splines

Rick_SAS — Fri, 19 Jul 2024 12:44:12 GMT

I don't know where that sentence came from, but it is not correct. GENMOD does not support the EFFECT statement so the OP must use GLMMIX.

DOC for EFFECT stmt: SAS Help Center: EFFECT Statement says "This section applies to the following procedures: GLIMMIX, GLMSELECT, HPMIXED, LOGISTIC, ORTHOREG, PHREG, PLS, QUANTLIFE, QUANTREG, QUANTSELECT, ROBUSTREG, SURVEYLOGISTIC, and SURVEYREG."

Re: Regression models with restricted cubic splines

Rick_SAS — Wed, 17 Jul 2024 15:29:59 GMT

> You can enrich your dataset with new covariates (spline effects) using PROC TRANSREG and subsequently use the procedure of your choice.

An easier choice is PROC GLMSELECT, which supports an OUTDESIGN= option. See the example at https://blogs.sas.com/content/iml/2019/10/16/visualize-regression-splines.html

Re: Regression models with restricted cubic splines

sbxkoenk — Wed, 17 Jul 2024 21:37:07 GMT

@Rick_SAS wrote:

I don't know where that sentence came from, but it is not correct. GENMOD does not support the EFFECT statement so the OP must use GLIMMIX.

Oh. 🤐
That sentence came from SAS/STAT® 15.2 User's Guide.
Statistical Graphics Using ODS
The EFFECT and EFFECTPLOT Statements
SAS Help Center: The EFFECT and EFFECTPLOT Statements

Koen

Re: Regression models with restricted cubic splines

Rick_SAS — Thu, 18 Jul 2024 10:17:22 GMT

Thanks. I have alerted the manager for that documentation. For the record, the GEE and GENMOD procedures support the EFFECTPLOT statement, but not the EFFECT statement.

Re: Regression models with restricted cubic splines

jiltao — Thu, 18 Jul 2024 13:41:22 GMT

Use PROC GLIMMIX with the EFFECT statement.

Re: Regression models with restricted cubic splines

BayzidurRahman — Thu, 18 Jul 2024 23:23:15 GMT

Thanks for all of your support. the glimmix works.

My objective is to estimate incidence rate ratio (IRR) from the poisson model with splines for given values of the continuous variable. A manual process of doing that is described in the following presentation using Stata. Is there a simpler way to achieve the same in SAS?
https://www.stata.com/meeting/4nasug/RCsplines.pdf

Re: Regression models with restricted cubic splines

jiltao — Fri, 19 Jul 2024 12:04:19 GMT

You might use the ESTIMATE statement in PROC GLIMMIX to get what you want. The exact syntax would depend on your data and model. What is your PROC GLIMMIX program and exactly what do you want to estimate?

Jill

Re: Regression models with restricted cubic splines

BayzidurRahman — Fri, 19 Jul 2024 12:29:20 GMT

Here is my code.

proc glimmix data=spl;
offset=log(psyc_den_m);
   effect tspl = spline(month_n/ details naturalcubic basis=tpf(noint)
             knotmethod=list(7, 15, 22, 28, 35) ); /* new in SAS/STAT 15.1 (SAS 9.4M6)  */
   model psyc_m_1 = tspl / dist=Poisson link=log offset=offset s;       /* fit model by using spline effects */
        ods select ParameterEstimates;
quit;

I want to estimate the incidence rate ration for month_n 15 vs 7 from the fitted model.

Re: Regression models with restricted cubic splines

jiltao — Fri, 19 Jul 2024 21:33:51 GMT

Try the following ESTIMATE Statement --

estimate 'month 15 vs. 7' tspl [1, 15][-1,7] / exp;

Thanks,

Jill

Re: Regression models with restricted cubic splines

BayzidurRahman — Fri, 19 Jul 2024 21:52:33 GMT

Hi Jill,

Thanks very much. It works. Do you know how can we define scaled variance option in glimmix? SCALE = P option works in GENMOD but glimmix doesn't recognise this option.

Re: Regression models with restricted cubic splines

BayzidurRahman — Fri, 19 Jul 2024 22:17:16 GMT

Hi Jill, how can we calculate risk difference for the same values?

Re: Regression models with restricted cubic splines

jiltao — Fri, 19 Jul 2024 22:51:15 GMT

The SCALE=P option in PROC GENMOD is one way to address overdispersion, as explained in the documentation below --

https://go.documentation.sas.com/doc/en/pgmsascdc/v_021/statug/statug_genmod_syntax22.htm#statug.genmod.genmodscale

PROC GLIMMIX does not have this option, it uses a separate statement to address overdispersion -

random _residual_;

This is explained in the link below --

https://go.documentation.sas.com/doc/en/pgmsascdc/v_021/statug/statug_glimmix_details05.htm

Re: Regression models with restricted cubic splines

jiltao — Fri, 19 Jul 2024 23:06:29 GMT

You might be able to use the STORE statement in PROC GLMMIX then use the %NLESTIMATE macro to compute the risk difference.

%NLESTIMATE macro is described in the details at the link below --

58775 - Estimating nonlinear combinations of model parameters (sas.com)

The Results tab has various examples showing how to use this macro.

Thanks,

Jill

Re: Regression models with restricted cubic splines

Ksharp — Sat, 20 Jul 2024 06:46:57 GMT

You could use PROC LOGISTIC + MODEL 's option SCALE= to get it.
You also could use PROC LOGISTIC +EFFECT to get cubic splines effect.
You also could use PROC LOGISTIC +/param=glm + lsmean trt/ ilink diff + %nlmeans to get risk difference or risk ratio.
Check the second example:
https://support.sas.com/kb/23/003.html
For Risk Ratio or Relative Risk

https://support.sas.com/kb/37/228.html
For Risk Difference

Re: Regression models with restricted cubic splines

BayzidurRahman — Sun, 21 Jul 2024 13:18:31 GMT

Hi, My outcome is count with an offset.

Re: Regression models with restricted cubic splines

Ksharp — Mon, 22 Jul 2024 00:15:41 GMT

You could change it into Binomial Distribution, and feed it into PROC LOGISTIC .


n    c  age
10   2  1
4    1  2

---->
y  age
1  1
1  1
0  1
0  1
0  1
0  1
0  1
0  1
0  1
0  1

1  2
0  2
0  2
0  2

And using:
model Y(event='1') = age;

Or just usa another style of MODEL statement:
model c/n = age;


Note: This assuming Binomial Distribution , Not Poisson Distribution, 
But I think they are almost the same thing, since Poisson is modeling for more discrete data,
as long as your data is NOT so discrete ,Binomial is more than suffice .

Re: Regression models with restricted cubic splines

BayzidurRahman — Mon, 22 Jul 2024 00:39:53 GMT

Thanks for the information. My data is very discrete with overdispersion. I need to either adopt scaled variance with Poisson distribution or negative binomial distribution.