topic Re: proc mianalyze in Statistical Procedures

proc mianalyze

FanRu — Tue, 07 Dec 2021 04:49:19 GMT

Hello statisticians,

I have been using proc mianalyze for sometimes and i found that some variables showed statistical significance in each dataset after multiple imputation, but they didn't show statistical significance in proc mianalyse. I tried to change the number of imputation but unfortunately found that it did not work. I would be very grateful if anyone could explain this phenomenon.

Re: proc mianalyze

Ksharp — Tue, 07 Dec 2021 12:21:49 GMT

Calling @Rick_SAS

Re: proc mianalyze

Rick_SAS — Tue, 07 Dec 2021 13:17:00 GMT

Please post your code so that we can have a better chance of answering your questions.

Re: proc mianalyze

FanRu — Tue, 07 Dec 2021 14:26:46 GMT

I am using proc traj to build a trajectory model and explore whether the covariates affect each trajectory group. Since proc traj does not have a by statement, I put together the parameter estimates for building the model using each impuation.

trajectory_imputation(data):

ID TIME BMI pulse HDLC LDLC TC TG glucose UA

A 2002 23.0 87.0 1.16 2.82 5.16 1.93 4.82 357

A 2003 22.8 . 1.18 . . 1.80 . .

A 2004 . 86 . 0 . 2.30 . . 4.54 359

B 2003 24.0 86 .0 1.19 2.30 5.17 1.75 4.54 358

.....

/*mi*/

proc mi data=trajectory_imputation out=imputed
seed=2021 nimpute=20;
var BMI pulse HDLC LDLC TC TG glucose UA;
mcmc;
run;

/*traj*/

data ParameterEstimates;
set oe; /*include PARMS STDERR COV*/
if _TYPE_="PARMS" or _TYPE_="STDERR";
run;

/*mianalyze for one trajectory group*/

proc transpose data=ParameterEstimates out=ParameterEstimates (rename=(_NAME_=Parameter PARMS=Estimate STDERR=StdErr));
var INTERC01 LINEAR01 QUADRA01 CUBIC01 BMI001 PULSE001 HDLC001 LDLC001 TC001 TG001 GLUCOSE001 UA001;
by _imputation_;
id _TYPE_;
run;

ODS OUTPUT ParameterEstimates=RESULT;
proc mianalyze parms=ParameterEstimates;
modeleffects INTERC01 LINEAR01 QUADRA01 CUBIC01 BMI001 PULSE001 HDLC001 LDLC001 TC001 TG001 GLUCOSE001 UA001;
run;

I also found that this situation may occur if the parameter estimates and standard errors of the variables between each imputation are large.Unfortunately, I can't figure out the reason for the large difference between each imputed datasets.Thank you very much for your help!

Re: proc mianalyze

SAS_Rob — Tue, 07 Dec 2021 14:27:59 GMT

Without seeing your code I would say that this is likely due to a large fraction of missing information (FMI). You should expect an increase in the variance (and thus a reduction in significance), specifically the between imputation variance, when the FMI is high. This section of the documentation will be helpful in that regard.

SAS Help Center: Multiple Imputation Efficiency

The other possible cause is that you have a bad imputation model (in the Proc MI step) or there is non-convergence in the MI models.

SAS Help Center: Checking Convergence in MCMC

If you can post your code and LOG (including the MI, modeling and MIANALYZE steps) then there might be something more concrete I can suggest.

Re: proc mianalyze

FanRu — Tue, 07 Dec 2021 15:36:43 GMT

Thank you for your help! I also found that there is a large gap between the coefficient estimates and standard errors of some covariates between different datasets after imputation.Here below my code and LOG:
/*MI for longitudinal data*/
proc mi data=trajectory_imputation out=imputed seed=2021 nimpute=20;
var BMI pulse HDLC LDLC TC TG glucose UA;
mcmc timeplot(mean(BMI) mean(pulse) mean(HDLC) mean(LDLC) mean(TC) mean(TG) mean(glucose) mean(UA));
run;
WARNING: The TIMEPLOT option is ignored when ODS Graphics is enabled.
NOTE: The EM algorithm (MLE) converges in 12 iterations.
NOTE: The EM algorithm (posterior mode) converges in 1 iterations.
/*Trajectory modeling*/
proc traj data=imputed out=of. outplot=op outstat=os outest=oe;
id ID;
var target_variable0-target_variable12;
indep time0-time12;
model cnorm;
max 240;
ngroups 4;
order 3 5 4 5;
risk age sex;
tcov BMI0-BMI12 pulse0-pulse12 HDLC0-HDLC12 LDLC0-LDLC12 TC0-TC12 TG0-TG12 glucose0-glucose12 UA0-UA12;
run;

data ParameterEstimates;
set oe;/*include PARMS,STDERR and COV*/
if _TYPE_="PARMS" or _TYPE_="STDERR";
run;
/*mianalyze for one trajectory group*/
proc transpose data=ParameterEstimates out=ParameterEstimates(rename=(_NAME_=Parameter PARMS=Estimate STDERR=StdErr));
var INTERC01 LINEAR01 QUADRA01 CUBIC01 BMI001 PULSE001 HDLC001 LDLC001 TC001 TG001 GLUCOSE001 UA001;
by _imputation_;
id _TYPE_;
run;

ODS OUTPUT ParameterEstimates=RESULT;
proc mianalyze parms=ParameterEstimates.;
modeleffects INTERC01 LINEAR01 QUADRA01 CUBIC01 BMI001 PULSE001 HDLC001 LDLC001 TC001 TG001 GLUCOSE001 UA001;
run;

Re: proc mianalyze

FanRu — Tue, 07 Dec 2021 15:39:24 GMT

Thanks a lot!