Search code examples
sascorrelationsas-iml

Generate correlated random variables that follow beta distributions


I need to generate random values for two beta-distributed variables that are correlated using SAS. The two variables of interest are characterized as follows:


X1 has mean = 0.896 and variance = 0.001.

X2 has mean = 0.206 and variance = 0.004.

For X1 and X2, p = 0.5, where p is the correlation coefficient.


Using SAS, I understand how to generate a random number specifying a beta distribution using the function X = RAND('BETA', a, b), where a and b are the two shape parameters for a variable X that can be calculated from the mean and variance. However, I want to generate values for both X1 and X2 simultaneously while specifying that they are correlated at p = 0.5.


Solution

  • This solution is based on modified methods used from Chapter 9 of Simulating Data with SAS by Rick Wicklin.

    In this particular example, I first have to define variable means, variances, and shape-parameters (alpha, beta) that are associated with the beta distribution:

    data beta_corr_vars;
        input x1 var1 x2 var2;  *mean1, variance1, mean2, variance2;
        *calculate shape parameters alpha and beta from means and variances;
        alpha1 = ((1 - x1) / var1 - 1/ x1) * x1**2;   
        alpha2 = ((1 - x2) / var2 - 1/ x2) * x2**2; 
        beta1 = alpha1 * (1 / x1 - 1);
        beta2 = alpha2 * (1 / x2 - 1);
        *here are the means and variances referred to in the original question;
        datalines; 
    0.896 0.001 0.206 0.004
    ;
    run;
    proc print data = beta_corr_vars;
    run;
    

    Once these variables are defined:

    proc iml;
      use beta_corr_vars; read all; 
      call randseed(12345);
          N = 10000;                  *number of random variable sets to generate;
          *simulate bivariate normal data with a specified correlation (here, rho = 0.5);
          Z = RandNormal(N, {0, 0}, {1 0.5, 0.5 1});   *RandNormal(N, Mean, Cov);
          *transform the normal variates into uniform variates;
          U = cdf("Normal", Z);      
    
          *From here, we can obtain beta variates for each column of U by; 
          *applying the inverse beta CDF;
          x1_beta = quantile("Beta", U[,1], alpha1, beta1);        
          x2_beta = quantile("Beta", U[,2], alpha2, beta2); 
          X = x1_beta || x2_beta; 
    
      *check adequacy of rho values--they approach the desired values with more sims (N);
      rhoZ = corr(Z)[1,2];                
      rhoX = corr(X)[1,2];
    
    print X;
    print rhoZ rhoX;
    

    Thank you to all users who contributed to this answer.