Proc genmod - estimate at each visit

[hetuvio]

Dear Steve,

Thank you very much for your help. It looks better... See below the iteration history:

As you can see below, Objective function = 37408 from iteration 6 to iteration 199.

...

I have seen that there are a lot of option regarding the convergence (PCONV, ABSCONV, ABSFCONV, ABSGCONV, ABSXCONV, FCONV, GCONV). How criterion would you use to relax the convergence criterion ?

Violaine

[SteveDenham]

First step, I would increase maxiter to 1000.

PCONV is the only one that directly affects the outer loop process.  It looks at the parameter set convergence.  I generally try to "sneak up" on it.  The default is 1e-8, so I generally try PCONV=1e-7 as my first attempt.  All of the other options refer to the inner loop process and are put in the NLOPTIONS statement.  Unless there is a warning in the log that something is failing in the inner loop, I would recommend staying away from these.

Steve Denham

[hetuvio]

I have increased maxiter to 1000. Proc glimmix took about 20 minutes to run, and it didn't converge. See below the last part of the iteration history:

Again, the Objective function stays close to 37408.

Thanks for your thoughts,

Violaine

[SteveDenham]

Time to use PCONV.  You are close to convergence.  Start with PCONV=1e-7, and if that doesn't work, choose slightly larger values until you get convergence.

Steve Denham

[hetuvio]

The following code works and converges:

proc glimmix data = efftab pconv=1e-5 ;

weekcont=week;

class usubjid trt01pn week country ;

model respfn = week trt01pn week*trt01pn / noint dist=bin link=logit ;

random week / subject=usubjid*country type=sp(pow)(weekcont) ;

*nloptions maxiter=200 ;

run ;

With the following results:

Then, I used the following statement to compute the OR at Week 1 for example :

estimate OR at WEEK 1" trt01pn 1 week 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
                                                  trt01pn*week   1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 / exp ;

But it leads to the following result:

Did I make a mistake? I think that an OR cannot be negative, so I'm probably missing something. Also, in addition to the OR, I'd like to get the CI + p-value of the diff between trt group.

Thanks for any idea,

Violaine

[SteveDenham]

I think to get an odds ratio, you have to have a difference between two groups as a starting point, so try using the ODDSRATIO in the MODEL statement.

However, there is still a problem with the model in that there are no denominator degrees of freedom for the treatment and treatment interaction. Try adding the between-within option to the model statement:

model respfn = week trt01pn week*trt01pn / noint dist=bin link=logit ddfm=bw;

Steve Denham

Message was edited by: Steve Denham

[hetuvio]

Dear Steve,

Thank you very much for answer answer. Indeed, it works better with the "ddfm=bw" option.

I have tried to use the "ODDSRATIO in the MODEL statement.", however I'm not sure to get what I really want (at each visit: OR of the proportion of responder in Active group versus PBO group).

I have to look further in the output result and SAS documentation.

Thank you!

Violaine

[SteveDenham]

I think you will want the ODDSRATIO option in the model statement, and again in an LSMEANS statement.  I believe you will be interested in the comparison at each week, so the slicediff= option comes into play.  Try adding:

lsmeans week*trt01pn /slicediff=week oddsratio;

Steve Denham

[hetuvio]

Dear Steve,

This is my last updated code :

(Fyi, I have removed the country - as discussed and agreed in our team)

  [1] proc glimmix data = efftab pconv=1e-5;

[1]  weekcont=week;

[1] class usubjid trt01pn(ref='0') week ;

[1] model respfn = trt01pn week week*trt01pn / dist=bin link=logit ddfm=bw oddsratio ;

[1] random week / subject=usubjid type=sp(pow)(weekcont) ;

[2] lsmeans week*trt01pn /slicediff=week oddsratio cl ;

[3] estimate"OR of Active vs Pbo at Week 1" trt01pn 1 week 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

                                              week*trt01pn 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 / exp ;

[1]  run ;

[1] The "oddsratio" option from the MODEL statement :

It displays ORs for:

- TRT01PN=Active vs. PBO

- AVISITN=WEEK 1 vs WEEK 52

...

- AVISITN=WEEK 48 vs WEEK 52.

>> But it does not display the OR of Active vs. PBO at each Week. So, I think that I cannot use the information from the "oddsratio" option from the MODEL statement.

.

[3] The "estimate" statement :

The result from the SAS Output is "Non-est". Did I make an issue in the specification of this statement? (As a reminder, TRT01PN is coded 1 for Active vs 0 for Pbo). I don't see why it doesn't work.

[2] The "lsmeans" statement :

I guess that I can retrieve the Odds ratio, CI and p-value from the "Simple Effect Comparisons of TRT01PN*WEEK Least Squares Means By WEEK" table (ODS Table Name = SliceDiffs). Do I understand well ? These OR will be computed considering the main effect + the interaction too. Is it correct ? I was also wondering why we wouldn't use the slice option instead of slicediff ?

Thanks a million in advance for your support,

Best regards,

Violaine

[hetuvio]

Dear Steve,

I have done the 2 following prog glimmix and it has the following "Fit Statistics" :

Glimmix A (considering the real time with a spacial covariance structure):

proc glimmix data = efftab pconv=1e-5 ;

weekcont=week;

class trt01pn(ref='0') usubjid week ;

model respfn = trt01pn week trt01pn*week / dist=bin link=logit ddfm=bw ;

random week / subject=usubjid type=sp(pow)(weekcont) ;

lsmeans week*trt01pn /slicediff=week oddsratio cl ; 

run ;

Glimmix B (considering the original visit variable avisitn with no special type of the covariance structure):

proc glimmix data = efftab pconv=1e-5 ;

class trt01pn(ref='0') usubjid avisitn ;

model respfn = trt01pn avisitn trt01pn*avisitn / dist=bin link=logit ddfm=bw ;

random avisitn / subject=usubjid ;

lsmeans avisitn*trt01pn /slicediff=avisitn oddsratio cl ;

run ;

Looking into the ratio of the generalized chi-square statistic and its degrees of freedom : it is close to 1 in Glimmix B. I understand that it indicates that the variability in these data has been properly modeled, and that there is no residual overdispersion.

Then, considering the Fit Statistics, I would consider the 2nd glimmix modeling (Glimmix B) instead. Could you please let me know your thoughts ?

Thanks so much,

Best regards,

Violaine

[SteveDenham]

I am out of the office until next week, so this will be short.

The estimate statement is non-est because it doesn't compare anything at this point.  To get a difference you have to code both parts of interest, one with a positive 1 the other with a negative 1.  I think I gae an example.

I hope you can wait until next week when I get back to work.

Steve Denham