To get residuals in PROC MIXED for a large data set, you need to add the RESIDUAL option to the model statement. This requests plots of both conditional and marginal residuals. However, ODS supresses plots with more than 5000 points by default. To remove this restriction, add plot(maxpoints=none) to the PROC MIXED statement. Here's an example:   proc mixed data=mydata plots(maxpoints=none); class id; model y = x / ddfm=kr2 residual; random int / subject=id; run;   On a side note, I recommend that you also plot the predicted values for the random effects (the EBLUPS). Like the residuals, the random effects are assumed to be normally distributed and unrelated to the predictors. You can request EBLUPS with the SOLUTION option on the RANDOM statement, and then output them with an ODS OUTPUT statement. Here's an example:   proc mixed data=mydata plots(maxpoints=none); class id; model y = x / ddfm=kr2 residual; random int / subject=id solution; ods output solutionr=randomeffects; run;   The data set of EBLUPS can then be combined with your original data for graphing.    ... View more

Sorry about the website confusion. I took down the website when I stopped offering consulting services. As for the macro, I no longer recommend using MMI_IMPUTE. MMI_IMPUTE uses an imputation algorithm called PAN, which was developed a while back by Joseph Schafer. Unfortunately, PAN is a bit outdated, and is less flexible than newer algorithms (e.g., the algorithm can't incorporate random effects between incomplete variables; all incomplete variables are required to be normally distributed). I recommend that you instead use a standalone software package called Blimp. Blimp was written by Brian Keller and Craig Enders at UCLA. Unlike MMI_IMPUTE, Blimp can handle random effects between incomplete variables and can also handle some non-normal incomplete variables (e.g., binary variables). The software and associated documentation are available at http://www.appliedmissingdata.com/multilevel-imputation.html. Blimp is free, and the website contains scripts for using it from SAS. ... View more

Hi,   The problem lies with the data. Run the following syntax:   proc freq data = laptops_dataset_raw;    tables Memory_Technology; run;   It shows that MEMORY_TECHNOLOGY has no missing values, but it does have values are coded as '?'. A question mark is not recognized as a missing value in SAS. Recode the question marks to SAS missing values. Note that this should also be done for other variables that use question marks (e.g., MAX_HORIZONTAL_RESOLUTION). After the data are cleaned up, try running PROC MI again.   Best, Steve   ... View more

The key feature that distinguishes FCS imputation from MCMC imputation is that FCS imputes variables one at a time. In PROC MI with the FCS statement, the variables are imputed sequentially in the order specified in the VAR statement. Because the underlying algorithm relies on random sampling, changing the order of the variables changes which part of the random number stream is used to draw each new parameter and imputed value. So, changing the order of the variables on the VAR statement should always result in some change to the imputed data.   As for the ordering of the variables, the order should not have much impact on your pooled parameter estimates if your number of iterations is large (van Buuren & Groothuis-Oudshoorn, 2011). Increasing the number of imputed data sets will further reduce the impact of the ordering of the variables on the results. However, the ordering of the variables will still effect the efficiency of the imputation process. It is generally recommended that you list the variables according to the number of missing observations for each (from most complete to least complete). ... View more

You essentially want to score the data set. You can use the OUT= option to output the original data set plus the factor scores.   Proc IRT Documentation   Best, Steve ... View more

Hi Steve, I'm glad you like it! If we want to get really fancy, we can simplify the code further with Kronecker products. All of this:      l1_random = j(n,2,0);      l2_id = j(n,1,0);     do i = 1 to (n_l2);         *Generate cluster ID variable;         l2_id[n_l1*(i-1)+1:n_l1*(i-1)+n_l1,1]=i;         *Copy random values from level-2 matrix to level-1 matrix;         do k = 1 to n_l1;             l1_random[n_l1*(i-1)+k,]=l2_random[i,];         end;     end; can be replaced with this:    *Convert l2_random to a level-1 vector (same values repeat within cluster);    l1_random = l2_random @ j(n_l1,1,1);    *Generate level-2 ID variable;    l2_id = (1:n_l2)` @ j(n_l1,1,1); Because the new code has no DO loops, it will run even faster. Best, Steve ... View more

Hi Tim, You specified G to be a diagonal matrix (no correlation between the random intercept and random slope). This is fine, but make sure that it is what you intended. Also, you drew a single sample for the random effects (rather than 200 samples). This means that everyone in the data set has the same intercept and slope, hence the zero variances. Try the code below. I replicated what you were trying to do, but without a compound symmetric structure for R. I left that off because you should generally avoid specifying a compound symmetric structure for R while also including random effects in the model (the random intercept and the residual within-cluster covariance end up being redundant). An AR(1) structure might be preferable if you want both R-side random effects and G-side random effects in the same model. Also, I switched the sampling of the random effects and the error to be performed in blocks. This decreases the memory requirements for the simulation, which will be helpful if you want to test the simulation for a large number of subjects. Best, Steve proc iml;    /***************************************************************************    Specify model parameters    ***************************************************************************/    n_l2 = 200;   *Number of people;    n_l1 = 4;     *Number of time points per person;    n = n_l1 * n_l2;    seed = 314159;   *Random value seed;    call randseed(seed); *Specifies a seed for all random number generation;    *Specify rixed intercept and slope;    fixed = {16.7611 0.6602};    *Specify block of G matrix;    *Treating G as block-diagonal reduces memory requirements;    g_block = {1.8430 0,               0      0.02228};    *Specify block of R matrix;    *Treating R as block-diagonal reduces memory requirements;    *R is currently set to have a block-diagonal structure;    r_block = 1.8787 # i(n_l1);    /***************************************************************************    Set up matrices and draw random values    ***************************************************************************/    *Draw intercepts and slopes from a multivariate normal distribution;    l2_random = randnormal(n_l2, fixed, g_block);    l1_random = j(n,2,0);    l2_id = j(n,1,0);   do i = 1 to (n_l2);       *Generate cluster ID variable;   l2_id[n_l1*(i-1)+1:n_l1*(i-1)+n_l1,1]=i;       *Copy random values from level-2 matrix to level-1 matrix;   do k = 1 to n_l1;   l1_random[n_l1*(i-1)+k,]=l2_random[i,];   end;   end;    *Generate level-1 ID variable;    l1_id = (1:n)`;    *Generate age variable;    age = repeat(do(8,14,2)`,n_l2,1);    *Draw level-1 errors from a multivariate normal distribution;    errors = colvec(randnormal(n_l2, j(n_l1,1,0), r_block));       *Generate dependant variable;    dv = l1_random[,1] + l1_random[,2] # age + errors;    output = l2_id || l1_id || age || dv;   names = {l2_id l1_id age dv};   create multilevelData from output [colname = names];   append from output;    close multilevelData; quit; proc mixed data=multilevelData;    model dv = age / s;    random int age / subject=l2_id type=un; run; ... View more

Yes, exactly. Or you can just copy the code that I posted previously. ... View more

I misunderstood. You are correct that the code suggested by PG would also work, as long as the variable names are changed to be the same as those used by mesdz (e.g., mn_x1, not nm_x1). ... View more

The do loop is not necessary after using the TRANSPOSE procedure. The data set would appear as follows: S 1 5 7 If you want to keep the variable label for each observation, just remove "drop=_name_" from the code I previously posted and the result is as follows. The _name_ variable is not necessary for generating the histogram, but you can keep it if you like. The code for generating the histogram is: proc univariate data=histo;    var s;    histogram; run; ... View more

The easiest way to do that is to use the TRANSPOSE procedure. After you transpose the data, you'll need to drop the _name_ variable and rename the col1 variable to s. The code below shows how to do this. data proc_means;    input mn_x1 mn_x2 mn_x3;    datalines; 1 5 7 ; run; proc transpose data=proc_means out=histo;    var mn:; run; data histo;    set histo (drop=_NAME_ rename=col1=s); run; ... View more

PGStats used nm_x as a prefix for the variable names, rather than mn_x. If you change the code to use mn_x as a prefix, the code works perfectly. The code below demonstrates that with three means. data pop;    input mn_x1 mn_x2 mn_x3;    datalines; 1 5 7 ; run; data popList; length id $12; set pop; array mn{*} mn_x:; do i = 1 to dim(mn);     x = mn{i};     id = catx("_", _n_, i);     output;     end; keep id x; run; proc univariate data=popList; var x; id id; histogram; run; ... View more

Though you could impute gender using a multiple imputation method that assumes normality, better approaches are available. The MI procedure includes fully conditional specification (FCS) multiple imputation using the FCS statement. The beauty of FCS multiple imputation is that it allows the user to employ logistic regression for imputing some variables and OLS regression for imputing others. The link below provides an example of using logistic regression to impute a nominal variable (species) and OLS regression to impute the other variables in the data set (length and width). SAS/STAT(R) 13.1 User's Guide I believe that the FCS statement became production in SAS/STAT 13.1. So, you'll need a recent version of SAS to perform FCS multiple imputation. Best, Steve ... View more