comp.soft-sys.sas - The SAS statistics package.
Can anyone explain which occasions it would be appropriate to use Random effects as opposed to fixed effects (in a regression). In an experimental design sense a fixed effect is when we are only concerned about making inferences about the particular treatments such as 10mg, 20 mg, 30mg. A random effect seeks to make inferences about all treatments (population) based on a random sample of 10mg, 20mg 30mg. This concept is understandable, but a random effect's use in a regression model isn't. Problem. If we are looking to model the amount of revenue expected to be collected given a bill amount (y being the amount paid on that bill) or the probability of paying a bill, or to build an attrition model using many other variables as predictors (some predictors are repeated measures such as monthly usage while other aren't such as rate or salary. The bill amounts are taken for a 3 year period per account. Note that all accounts may not have 3 years worth of data. This appears to be a repeated measure problem. What models or procs would be needed to model these two models or models like it? Would you treat this as a random or a fixed effect? If one chooses random wouldn't it only be because the bill amounts were chosen from a population of bill amounts per account? If one uses all possible amounts per customer would it then be a fixed effect? What models are procs should be used for eithr case? Is a repeated measure model the only way to model this? Thanks DW
Hi everyone, I'm fairly new to SAS (and repeated measures analysis and mixed models), so bear with me here: The dataset I'm working with consists of counts of birds flying into n = 27 study sites, on multiple days, in multiple years (note: no more than one observation per day at a given site). The tricky part is that the data are unbalanced: Not all sites were surveyed in all years, with some sites surveyed in more years than others. Likewise, not all sites were surveyed on the same day, with some sites surveyed on more days than others in a given year. What I like to know, is whether there are significant year and day effects on the number of birds observed flying into each site (i.e., I'd like to examine both daily and yearly variation). The model would therefore look something like this: IBijk = Yeari + Day(Yeari )j + Eijk where IBijk = the number of incoming birds in year i on day j at site k. With missing data and unequal time intervals, my options for analysis are either PROC mixed or PROC Glimmix (if I don't want entire observations dropped if one value is missing, which I don't). For PROC mixed, I stuck on how to specify a day (i.e., Julian date) and year time period in the repeated statement (i.e., the epeated-effect. Right now I have: PROC mixed DATA = dataset; CLASS SiteName Year JulDay; MODEL IB= Year JulDay(Year); REPEATED<
/subject= SiteName type=un; run; However, I don know what to put in the< , or if the code above is even appropriate. I've been all through the literature and forums, but can't seem to find an example where there are two within-subject effects (i.e., where measurements on the subject are repeated, and being examined, over both day and year, as above, or something similar), and where one of these effects is nested in the other. Any help would be greatly appreciated. Please let me know if you need more details. Cheers, Jenn
Hi all, I wonder what kind of analysis technique I should use in the following situation: We had baseline data + repeated mesures on covariates and also on the outcome. It turned out that outcome measures were highly variables (there also was a change in the reading procedure). So we decided to have a clinician judge the overall evolution of each patient's outcome and decide if its status regress or not. I still have the baseline info. My question is regarding the repeated covariates measures. Each patient had a different number of visits (it goes from 5 to 15 visits), each time, some measurements were done (some variables were also measured more often than some other). We would like to use that info. What kind of model I should go with? It's the repeated covariates but single outcome thing that messes me up. Thanks for you input (Here and private mail, if possible). Thanks, JP
Hi, Random effect model and random coefficient model are two different models. However, I am confused by SAS code how to specify them. Anyone has simple codes to distinguish them? Thanks. Specifically, I am interested in the following specification Y=X*beta + Z*gamma + e, where X,Z are covariates, beta is a fixed coefficient, gamma is a random coefficient, which equals T*alpha + u, both u and e are error term By simplification, Y=X*beta+Z*T*alpha+Z*u+e This is a random coefficient model, isn't it? Can I get estimates for both gamma, and alpha? Thanks very much.