Hello everyone, I need your advice regarding a little simulation I am trying to do. Recently I had a data set which I analyzed using a mixed model. Now I have simulated some data to test if my model was the right one. Ignoring what was in the past, now I have a data I simulated, which is a bivariate gamma distribution with a correlation of 0.8 (see attached photos). The data suppose to simulate a time variable with gamma distribution, and two measurements per each subject. I have 1000 subjects in this particular set. Each subject "gave" 2 samples with a correlation of 0.8 between them. I chose a shape parameter of 0.7 and scale of 1. I wanted a mean of 0.7. The s.d is 0.85, I am not so why, because to my calculation that yields a scale of 1.1, but doesn't matter, maybe it's due to the inaccuracies of the simulation. Now I wanted to estimate the mean and s.d of my data, taking into account the correlation between measurements within a subject, I tried two codes: proc mixed data = Long;   model Value = D / s;   random intercept / subject = ID; run; and proc mixed data = Long;   model Value = D / s;   random ID; run; where D is a vector of 1's. The first code gave me a mean value of 0.697 with s.e of 0.025 while the first code gave me 0.647 with s.e of 0.03447. The first code was closer to the truth, and I wanted to ask why. There is no explanatory variable in this model, so how come a random intercept is the right model ? In addition, shouldn't the correlation affect only the variance, and not the mean value ? I am quite confused here, I thought that using the intercept command, I make like a separate regression line per subject, but here I do not have an "X" variable, only "Y". Thank you ! ... View more

Hello all, I have some basic questions regarding PROC GLIMMIX and PROC MIXED, and probably other mixed models related PROC's. Assume that I have an outcome variable Y, continuous, and a treatment variable X with 2 levels. Let's also say that every subject in the analysis contributes 3 data points (can be 3 treatments in one person, or students within schools). If I write this code: proc mixed data = ....;      class SubjID;      model Y = X / solution cl;      [random statement]     run; what is the difference between: random SubjID; to random intercept / subject = SubjID; and let's say (if it's a legal statement): random Treatment / subject = SubjID; I tried it on a small data set and the first two options gave me identical results. This question of course is relevant also for glimmix with a dist = binary. And one more question, if I have this scenario, and I run a PROC MIXED once with random statement and once with REPEATED, and the results are very similar and leading to the same conclusions, how can I choose if to use R side or G side covariance ? Thank you ! ... View more

Hello all, I have a variable which theoretically should have got discrete values between 0 and 6. This variable is not a Likert scale based, but I think we can think about it as such. Practically in my sample, I had 22 independent subjects with 37 observations, i.e., some of them gave two observations while some only one. All observations apart from one were either 0 or 1. Among those subjects with two observations, apart from one case, all values matched (i.e., 0 twice or 1 twice). I need to calculate the mean of the variable, the standard deviation (or variance) and the 95% CI. I did several things, first out of curiosity, I have calculated the descriptive statistics as if I had 37 independent observations. I got a mean of 0.513 with SD of 1.044. I did it like this (and also manually): proc means data = data;   var Y; run; In the next step I averaged two observations within a subject, what gave me a sample of 22 independent values. I have calculated a weighted mean and weighted SD using weights of either 1 or 2, depending on the number of observations per subject. I got a mean of 0.5135 with SD of 1.358. I did it like this: proc means data = Avged_data;   var Mean_Y;   weight N_Y; run; As a last attempt, I did this: proc surveymeans data = data;   var Y;   cluster Subject_ID; run; And I got a mean of 0.5135 with SD of 1.13 What bothers me, is that I don't understand, or more precisely know, how SAS calculated the variance in the last case. I tried looking at the help documents, and saw some complicated variance formulas (Taylor and others), however I did not see any formula specific for the cluster statement case. I wanted your advice, first for the correctness of calculating a mean and SD when a vast majority of an ordinal variable values are either 0 or 1, and second, how would you do it ? Would you choose the weighted approach, or the PROC SURVEYMEANS with the ready cluster statement that to my understanding is exactly what I need. Usually I do not like using a method I do not understand, and my fear is that at the current moment PROC SURVEYMEANS with cluster statement is a black box, does anyone knows what is the rational behind the calculation. If I'll know some details maybe I can find the article in the literature which is the source of the calculations... Thank you in advance ! Edit: I just tried one more option, I ran a GEE model with an explanatory variable of 1's, i.e. a vector (1,1,1,1,1,...,1) taking an idea Steve gave me with similar problem some time ago with a binary outcome. I did proc genmod data = data;   class SubjectID;   model Y = D1;   repeated subject=SubjectID / corr=cs corrw; run; it gave me a mean of 0.5271 with SD of 1.23 (it gave me S.E of 0.2029 and I have extracted the SD from there). I tried the same with PROC MIXED with no success (didn't give me an intercept "line" at all in the fixed effect table). However, I see no additional value in having another pair of (mean,SD) to choose from, it is confusing enough now. The bottom line, this doesn't change my problem, how do you choose the correct answer, now that I have one more candidate of possible values....? ... View more

Hello all, I would like to simulate data of a bivariate "table" distribution. I want to generate a vector of values (0,1,2,3,4,5) with probabilities (p(0), p(1), p(2), p(3), p(4), p(5)). Then I want to generate another such vector, with a correlation of let's say 0.8 to the first vector. How do I do that ? Thank you in advance. ... View more