topic Re: MMRM and Pooled Cohort in Statistical Procedures

MMRM and Pooled Cohort

yw2757 — Sun, 22 Nov 2020 02:08:00 GMT

Suppose have 3 Treatment Group, Placebo (1), Group A with Dose 10mg, Group B with Dose 20mg.

Study duration: Baseline (Day 1), Week 4, Week 8, Week 12.

Goal: 1) compare the change from baseline across visits among treatments separately 2) compare the change from baseline across different visits among (placebo) VS. (combined Group A & B)

When using PROC MIXED model, "lsmeans" can provide result of Goal (1), "estimate" can provide (2); however, I am not very sure how to use "estimate" or "lsmestimate" or "contrast" to get Goal (2)

proc mixed data= long;
  class treatment visit;
  model CHG =  Treatment*visit;
  repeated visit / subject=id type=UN ;
  estimate 'placebo vs pool' treatment 1 -0.5 -0.5; 
  lsmeans treatment*visit/ slice=visit; /*differences in exertype for each time point*/
run;

Re: MMRM and Pooled Cohort

SteveDenham — Mon, 23 Nov 2020 14:03:51 GMT

We'll need some clarification. It seems you have an ESTIMATE statement that gives you an answer to your Goal(2), so I suspect you want something more. Is that a comparison of placebo vs. pooled treatments at each time point? That would require knowing how many timepoints, but otherwise is a perfect case for using an LSMESTIMATE statement.

SteveDenham

Re: MMRM and Pooled Cohort

StatsMan — Mon, 23 Nov 2020 14:49:36 GMT

Adding the /e option to your LSMEANS statement is always a good idea when you are unsure of the hypothesis you are testing with your ESTIMATE (or CONTRAST) statement. The hypothesis tested by LSMEANS is a test of a linear combination of the paramter estimates in the model, achieved by the vector multiplication of two vectors L*B. B is the vector of the paramter estimates (which you can see by adding the /s option to the MODEL statement). L is the vector of coefficients as specified on the ESTIMATE statement (seen by adding that /e option).

This https://support.sas.com/kb/24/447.html gives lots of examples of writing contrasts and estimates.

Many estimates are more easily constructed if you take a linear combination of the LSMEANS, instead of the model parameters. As @SteveDenham pointed out, that can be achieved through the LSMESTIMATE statement.

Re: MMRM and Pooled Cohort

yw2757 — Mon, 23 Nov 2020 20:51:16 GMT

" Is that a comparison of placebo vs. pooled treatments at each time point?" is the questions I wish to ask how to use "lsmestimate"

Confused how to set the coefficients

Re: MMRM and Pooled Cohort

StatsMan — Tue, 24 Nov 2020 12:54:27 GMT

In the absence of real data, simulated data can come in handy. The code below simulates data that looks like the situation you have and runs a mixed model similar to yours.

data test;
call streaminit(56340);
do id=1 to 100;
trt=ceil(rand("uniform")*3);
base=rand("uniform");
do time=1 to 3;
y=trt*time*rand("uniform")-base;
output;
end; end;
run;

proc mixed data=test;
class trt time;
model y=trt time trt*time;
repeated time / subject=id type=un;
lsmeans trt*time / slice=time;
lsmestimate trt*time 1 -.5 -.5 / e;
run;

The LSMESTIMATE statement above is our first attempt at comparing the result of the placebo (treatment 1) to the pooled result of the two active treatments (treatments 2 and 3) at time point 1. The /e option will help us see if we got the coefficients right:

Looking at the results of the /e option, it appears that we did not set up the coefficients correctly. The comparison here is within the first level of treatment, comparing the result at time 1 to the pooled result at times 2 and 3. That is not what we had in mind. Let's change the coefficients on the LSMESTIMATE statement:

proc mixed data=test;
class trt time;
model y=trt time trt*time;
repeated time / subject=id type=un;
lsmeans trt*time / slice=time;
lsmestimate trt*time 1 0 0 -.5 0 0 -.5 0 0 / e;
run;

Now the /e option shows which paramters we are combining:

Now we can see that we have a 1 next to the lsmean for trt 1 and time 1, and -.5 next to the lsmeans for trt 2, time 1 and trt 3, time 1. These coefficients line up with the comparison we want, comparing trt 1 to the pooled result of trt 2 and 3 at time point 1.

You can verify the result algebraically too. If you look at the output of the LSMEANS table:

We wish to compare the lsmean for trt 1 at time 1 to the pooled result of trt 2 and trt 3 at time 1. That would be .04008 - (.7140+1.3651)/2 = -.9995, the result we get from the LSMESTIMATE statement.

Re: MMRM and Pooled Cohort

yw2757 — Tue, 24 Nov 2020 13:21:08 GMT

This is really helpful and exactly I am looking for.

Really appreciate!