CONTRAST statement in proc NLMIXED

[issac]

Hi everyone;

I am working with 9.3 version. My model did work and now want to do some tests for its estimates. The following is the code

proc nlmixed data=x EBOPT MAXFU=800 maxit=600 ;

where id ~=. ;

lambda=exp(b0+badmsource*admsourceHOSPITAL+bseqadmsource*admsourceHOSPITALseq+bseqsloglos*seqsloglos+b1maruser*marMARRIEDuserenrolleeYES+b2maruser*marPREVIOUSLY_MARRIEDuserenrolle

    +b1marseq*marMARRIEDseq+b2marseq*marPREVIOUSLY_MARRIEDseq+bsloglos*sloglos+e);

ll=-lambda*nrtime/***(alpha+1)*/+rstatus*(/*LOG(alpha+1)+alpha*LOG(nrtime)+*/LOG(lambda));

MODEL nrtime~GENERAL(ll);

RANDOM e~NORMAL(0,sd*sd) SUBJECT=id;

PARMS  b0=0 badmsource=0 bseqadmsource=0 bseqsloglos=0  b1maruser=0  b2maruser=0 b1marseq=0 b2marseq=0 bsloglos=0

    sd=1 /*alpha=0*/;

CONTRAST 'maruser' mar*userenrollee 1 -1 ;

run;

I want to test that "b1maruser=b2maruser" or not, but I cannot process it in CONTRAST. These tow estimates are related to crossed effect "mar*userenrollee" that I previously built by output design matrix from proc logistic.

The error is this:

88   CONTRAST 'maruser' mar*userenrollee 1 -1 ;

                                         -

                                         22

                                         76

ERROR 22-322: Syntax error, expecting one of the following: (, *, **, +, ',', -, /, <>, ><, [, {.

ERROR 76-322: Syntax error, statement will be ignored.

Any helpful comments would be highly grateful.

Thanks!

Issac

[lvm · Accepted solution]

You are trying to use the syntax of MIXED or GLIMMIX (etc.) with NLMIXED. They are not the same. In NLMIXED, there are no "levels" of a factor, because factors are not used (explicitly). If you just want to know if b1maruser is different from b2maruser (both parameters in your model), then I suggest:

ESTIMATE 'label' b1maruser - b2maruser;

A slightly harder way, but explicitly using +1 and -1,  would be:

ESTIMATE 'label' 1*b1maruser -1*b2maruser;

In NLMIXED, you use ESTIMATE for linear or nonlinear functions of parameters (no variables listed). The CONTRAST statement is primarily to have multiple ESTIMATEs in the same statement.

[issac]

Thanks so much Ivy!  It is really helpful and constructive note.

Issac

[issac]

Ivy!

What if I want to test that both whether b1maruser and b2maruser are set equal to zero or not? Because when I did what you recommend, it gives

Label Est.      StError DF   t Value  Pr.   Alpha  Lower   Upper

label	-1.1804	0.3313	2443	-3.56	0.0004	0.05	-1.8300	-0.5307

means b1maruser could not be the same as b2maruser, but it does not test whether both could be zero or not?

Really appreciate!

Issac 

[SteveDenham]

Hi Issac,

Check your parameter estimates table.  It should have a t-test (based on the delta method to estimate the standard error) for every parameter, testing the null hypotheis Ho: Parameter=0.

If you want a joint test that the two parameters are simultaneously zero, then a contrast statement such as:

contrast 'Simultaneous test' b1manuser, b2manuser;

ought to give a 2 df test. Note the comma between the two SAS expressions.

Steve Denham

[issac]

Morning Steve;

Exactly what I was looking for. The 2 df are related to the Numerator and for Denominator, it used the same DF as provided in the last parameter estimate table.

Thanks so much!

Issac