SAS power analysis Macros for a linear growth curve model with group information
2009-04-16    Zhang, Z. & Wang, L.   
Print from: Zhiyong Zhang \'s Psychometric Website
Address: https://www.psychstat.org/us/article.php/80
SAS Macros for a linear growth curve model with group information
These macros are licensed under the GNL General Public License Version 2.0. You can use/modify/distribute those macros. For questions or comments, please contact Johnny Zhang at zhiyongzhang(@)nd.edu. It would be appreciated if you cite the macros in the following way:

Zhang, Z., & Wang, L. (2009). Power analysis for growth curve models using SAS. Behavior Research Methods, 41(4), 1083-1094. Request a copy

/*Suppress the output and the log */
options nosource nonotes nosource2 nomprint;

/*Power analysis of growth curve models using SAS*/
/*CHANGE THE PARAMETERS HERE*/
*model parameters;
%LET MuL=10;        *mean level/initial status;
%LET MuS=.1;       *mean slope/rate of change;
%LET Sigma_e=1;                *residual standard deviation;
%LET Sigma_L=2;                 *level standard deviation;
%LET Sigma_S=1;                *slope standard deviation;
%LET rho=0;              *correlation between levels and slopes;
%LET miss=0.1;       *missing data rate, 0: no missing data;
%LET treatmentL=1;    *group difference in the level
%LET treatmentS=1;    *group difference in the rate of change
*power parameters;
%LET R=5000;       *number of simulation replications;
%LET T=5;          *number of measurement occasions;
%LET start=100;     *the minimum sample size to consider;
%LET end=100;     *the maximum sample size to consider;
%LET step=50;      *the step between two sample sizes;
%LET df=1;         *the difference in the numbers of parameters;
%LET seed=4321;     *random number generator seed;
 
/*DO NOT CHANGE CODES BELOW UNLESS YOU KNOW WHAT YOU ARE DOING*/
 
/*The first Macro: LL*/
/*Calculate the chi-square difference of two nested growth curve models*/
%MACRO LL(N,T,seed);
DATA Sim_LinGM;
* set statistical parameters;
  N = &N; seed = &seed;
* setup arrays so that we can include multiple variables for repeated measures;
  ARRAY y_score{&T} y1-y&T;
  ARRAY M{&T} m1-m&T;
  m1=1;
* generate raw data with considering the missing data rate;
  DO _N_ = 1 TO N;
    x = RANBIN(seed,1,.5);
    e_L=RANNOR(seed);
    e_S=&rho*e_L+SQRT(1-&rho**2)*RANNOR(seed);
    L_score=&MuL+&treatmentL*x+&Sigma_L*e_L;
    S_score=&MuS+&treatmentS*x+&Sigma_S*e_S;
* include indicator variables to generate missing data;
    DO t = 1 TO &T;
      y_score{t} = L_score + (t-1) * S_score + &Sigma_e*RANNOR(seed);
      END;
 DO t=2 TO &T;
   m{t}=m{t-1};
   IF m{t-1}=1 AND  RANUNI(seed) > (1-&miss * (t-1))/(1-&miss * (t-2)) THEN m{t} = 0;
      IF m{t}=0  THEN y_score{t}=.;
   END;
    KEEP y1-y&T x;
    OUTPUT;
    END;
RUN;
 
DATA LinGM;
  SET Sim_LinGM;
  %DO t = 1 %TO &T;
    id = _N_; time=&t-1; y=y&t; x=x; OUTPUT;
  %END;
  KEEP id time y x;
RUN;
 
 
/*Fit two nested models to the data*/
ODS OUTPUT FitStatistics(persist=proc)=fit;
*Model 1: the true model;
TITLE1 'Linear Growth Model with covariate';
PROC MIXED DATA=LinGM NOCLPRINT COVTEST MAXITER=100 METHOD=ML;
  CLASS id;
  MODEL y = time x x*time /SOLUTION DDFM=BW CHISQ;
  RANDOM INTERCEPT time / SUBJECT=id TYPE=UN GCORR;
RUN;
 
*Model 2: without treatment effects on the rate of change;
PROC MIXED DATA=LinGM NOCLPRINT COVTEST MAXITER=100 METHOD=ML;
  CLASS id;
  MODEL y =  time x  /SOLUTION DDFM=BW CHISQ;
  RANDOM INTERCEPT time / SUBJECT=id TYPE=UN GCORR;
RUN;
 
ODS OUTPUT CLOSE;
%MEND LL;
 
/*The second Macro: POWER*/
/*This Macro calls the first Macro LL for each replication*/
* Calculate power based on R replications;
%MACRO POWER(R,N,T,seed,df);
DATA tempfit;
  DO _N_=1 TO 8;
    tempfit=_N_;
 OUTPUT;
  END;
RUN;
 
%LL(&N,&T,&seed);
DATA fit;
  MERGE fit tempfit;
RUN;

DATA allfit;
  SET fit;
RUN;

%DO I = 2 %TO  &R;
  PROC DATASETS LIBRARY=WORK; DELETE fit; RUN; QUIT;

  %LL(&N,&T,%eval(&seed+&I*1389));
  DATA fit;
    MERGE fit tempfit;
  RUN;

  DATA allfit;
    SET allfit fit;
  RUN;
  DM 'CLEAR LOG';
%END;

DATA allfit;
  SET allfit;
  IF MOD(_N_,4) ~= 1 THEN DELETE;
  KEEP Value;
RUN;

DATA allfit;
  SET allfit;
  id =INT((_N_-.1)/2)+1;
  modelnum = MOD(_N_+1, 2);
RUN;

PROC TRANSPOSE DATA=allfit OUT=allfit prefix=model;
  BY id;
  ID modelnum;
  VAR Value;
RUN;

DATA allfit;
  SET allfit;
  ss = &N;
  diff = model1 - model0;
  ind = 1;
  IF diff=. THEN DELETE;
  IF diff<0 THEN DELETE;
  IF diff < CINV(.95, &df) THEN ind = 0;
  DROP id _NAME_ model0 model1;
RUN;


PROC MEANS DATA = allfit;
  VAR ss ind;
  OUTPUT OUT=power mean(ss ind)=ss power;
RUN;

%MEND POWER;

 
/*The third Macro: POWERCURVE*/
/* This Macro calls the second Macro for each sample size*/
%MACRO powercurve(R, seed, st, end, step, T,df);
%POWER(&R, &st, &T, %eval(&seed+&st), &df);
DATA allpower;
  SET power;
  RUN;
 
%LET st = %eval(&st + &step);
%DO %WHILE (&st <=  &end);
  %POWER(&R, &st, &T, %eval(&seed+&st), &df);
  %LET st = %eval(&st + &step);
  DATA allpower;
    SET allpower power;
  RUN;
  DM 'CLEAR LOG';
%END;
 
* Save the results for possible future use;
DATA allpower;
  SET allpower;
  FILE "power.txt";
  PUT ss power;
RUN;
 
* Plot the power curve;
ODS PDF FILE='power.pdf' NOTOC;
PROC GPLOT DATA = allpower;
  SYMBOL I=JOIN;
  PLOT power*ss;
RUN;
QUIT;
ODS PDF CLOSE;
%MEND powercurve;
 
ODS RESULTS OFF;
ODS LISTING CLOSE;
%powercurve(&R,&seed,&start,&end,&step,&T,&df);
ODS RESULTS ON;
ODS LISTING;

Editor: johnny