11-07-2012 06:55 AM
I have a balanced dataset of 1285 observations (257 firms, 5 years). I have been trying Proc Panel using Parks method. The SAS log provides "The number of observations is too small to permit estimation of the regression parameters in model statement". Any ideas how can I solve this problem.
11-07-2012 08:32 AM
Can you share your code? My first guess is that there are more parameters to be estimated in the variance-covariance matrix than you might think. I tried to dig out what would be estimated in general from the documentation, but I am having a bit of a problem without knowing some specifics about your model.
11-07-2012 05:16 PM
Here is the SAS code:
proc sort data=Warranty.Paneldataset;
by ASXCode Year;
proc panel data=Warranty.Paneldataset outest=Warranty.Reg1_Parks;
ID ASXCode Year;
Reg1_Parks: model Leverage = WBtoTA Size AssetCollateralValue ROA ROAVolatility NonDebtTaxShield TobinsQ TotalAccruals HerfindahlIndex RDIntensity GICS1010 GICS1510 GICS2010 GICS2020 GICS2030 GICS2510 GICS2520 GICS2530 GICS2540 GICS2550 GICS3010 GICS3020 GICS3030 GICS3510 GICS3520 GICS4510 GICS4520 GICS4530/ Parks;
Here is what is I get in the log
"NOTE: Estimated autoregression parameter RHO=1.032 did not satisfy |RHO|<1.0 criteria for cross section ASXCode=WEB. It will be set to a value within range.
ERROR: The number of observations is too small to permit estimation of the regression parameters in model statement REG1_PAR."
The NOTE is written 34 times for different firms.
11-07-2012 05:21 PM
Please note that all of the below are dummy variables
GICS1010 GICS1510 GICS2010 GICS2020 GICS2030 GICS2510 GICS2520 GICS2530 GICS2540 GICS2550 GICS3010 GICS3020 GICS3030 GICS3510 GICS3520 GICS4510 GICS4520 GICS4530
11-08-2012 08:20 AM
OK. Take all of this with a large grain of salt, as this is an area I am just now starting to work in.
I see two possibilities for what is going on. It looks like the Parks method may not be valid for your data. However, this may arise one of two ways. First, the data really do not fit an autoregressive process. I doubt this, but it may be the case. I think that because the regression parameters cannot be adequately estimated, the RHO value is similarly mis-estimated.
So, what might be going on? (I stress MIGHT, and really hope someone with more experience with PROC PANEL jumps in soon.) In the details section of PROC PANEL, I dug around and find N=257 (if this is incorrect, then everything after this is flat wrong). This means that the covariance matrix is a 257x257 arrangement of 5x5 symmetric matrices. These look identical (to me) for all N entries, so that is not the source of the problem. I think it might be trying to estimate 257*256 correlations/heteroscedasticity coefficients. Maybe not, as I would think that might be a pretty common thing that would have to be done, and those aren't the regression parameters. There are 28 parameters, some of which are 0-1. For the firms that "blow up", is there a particular pattern of the dummy variables that might result in a singular matrix? Right now, this is the only thing I can identify as a possiblity.
And I would really like to know what is going on! Perhaps posting this in the Forecasting forum would help, as the folks who follow that forum are a lot more familiar with the ETS procedures.
Message was edited by: Steve Denham
11-08-2012 05:13 PM
Thank you. I have tested for autocorrelation using DW and Godfrey and both tests provide evidence of AR(1). I have just posted that in Forecasting forum. I will keep you posted.
11-09-2012 01:04 AM
I have estimated the same model using proc reg and when I tried to test the heteroscedasticity of errors using SPEC, SAS log provides that the average covariance matrix is singular and could not perform the SPEC. Do you think this is related to the inability of SAS to perform PARKS too?
I also tested the normality and autocorrelation of residuals assumptions for OLS (using proc autoreg for the latter) and results show that both assumptions are violated. This means that I should abandon OLS (proc reg).
11-09-2012 08:00 AM
I think this is an "AHA, that's what's happening" discovery. Any possibility that the dummy variables are linearly dependent? If they are generated from a single variable with a lot of levels, try using PROC GLMMOD to get an output dataset of them that might work better. I am still baffled myself.