comp.soft-sys.sas - The SAS statistics package.
Dale: I know that I speak for many here when I say thank you for your prompt and detailed explanations of all things statistical--you're a treasure! Now that I am done brown-nosing, I have a couple of follow-up/follow-on questions I'd like to put forth for comment. You mention that fixed effect model assume that all off-diagonal covariance terms are zero--but isn't this the same as saying the observations are independent and have an identity matrix (i.e. no correlation)? I had always envisioned the inclusion of fixed terms for each cluster of observations as having a covariance matrix more akin to an exchangeable/compound symmetry than identity--a fixed model (in effect) gives each subject (i.e. cluster) its own intercept, and thus accounts for the marginal effect of membership in this cluster. You continue that if there is some off-diagonal correlation among observations then using fixed effects are inefficient and the SEs can be biased--but, again, I thought by including fixed subject effects you are attempting to account for this correlation. Staying with fixed effects for the moment, I recall that a key factor in deciding whether to go with a fixed or random effects model is the presence of correlation between the subject effect and other explanatory variables in the model--if there is correlation, but you use a random effects model, then the estimated correlation coefficients are biased and inconsistent. Using a fixed effects model protects against this type of specification error, but at a cost of reduced efficiency due to the increased number of parameters. Now from what you have said and what I have previously gathered, the decision regarding fixed vs. random puts one on the horns of a dilemma--if you believe there is correlation within each cluster (and why else would you try to include it in the model?) you should use a model that accounts for off-diagonal correlation (marginal or random effects); however, if there is correlation with other predictors in your model (highly likely in non or quasi-experimental designs) you're better off with a fixed effects model. Now is this the point where you say "using a mixed effects model gets you the best of both worlds..." or am I just fantasizing about a neat, tidy solution ;-) Thanks again for your time and attention, and looking forward to your comments... Pete
I would like to seek your advice. The data have three waves. I am analyzing the data with Fixed Effect Method (FEM) using each individual as a cluster. In the FEM, do I have to be concerned about multicollinearity among independent variables? Could you give me your advice about this issue? Thanks in advance. Best, Sunhee
Hello All, I am testing analyst forecast error given certain characteristics of earnings (let's say Earnings and SURPRISE). My data range from 1980 to 2005, about 100,000 firm-years observations. Possibly many firms are duplicated. I want to run fixed effect model, Proc Mixed Data=ANALYST; Class FirmID Year Quarter; Model ForecastError = FirmID Year Quarter; Random Earnings SURPRISE; Run; Am I correct to run firm and year fixed effect? Thank you for your comment and help. I always appreciate your help. Minsup