04-23-2015 09:47 PM
I'm doing a replication of an empirical paper examining the effects of bus transit on various economic measures. I have panel data for about 80 counties over 16 years, and for this particular model I'm regressing the unemployment rate against operating expenses of the bus systems. There are many missing values, which makes this data unbalanced, and I am aware that this is problematic for some measures of autocorrelation. What I'm more concerned about is just making sure I understand the process, rather than necessarily getting correct values.
Working with non-panel time series data, I know how to use AUTOREG to find a Durbin-Watson score for the original model, and then set a lag time to correct, then get a new DW score reflecting the corrected model. So maybe the uncorrected model gives DW=.3, then corrected shows DW=2 (approximately).
Now, I'm not sure exactly how to implement this in PROC PANEL. I have run the following code:
proc panel data=full plots=none;
id county_id year;
model unemployment_rate = real_oe_percap/ dw;
model unemployment_rate = real_oe_percap/ FIXONE dw;
model unemployment_rate = real_oe_percap/ FIXONE hccme=4 dw;
model unemployment_rate = real_oe_percap/ FIXONE hac dw;
A note: the DW score in PANEL is experimental, and didn't even seem to be available in SAS Studio until after I spoke to customer service today. However, I've done this with non-experimental tests and gotten the same results, so I think I still have some kind of a problem.
The first model defaults to random effects, which I don't want, but am including for debugging purposes. The next three use a fixed effects model, with what I intend to be corrections for autocorrelation in the last two. So here's the thing: the three FixOne models give identical results for fit statistics and parameter estimates, with the last two having a slightly different standard errors from the second. However, all four of them have identical DW scores of .44, which suggests positive autocorrelation. Am I implementing these incorrectly? Shouldn't I be able to perform a correction as in AUTOREG that will result in a DW score closer to 2?
Also, I have performed a test for heteroskedasticity as follows:
proc model data=full plots=none;
parms a1 b1;
unemployment_rate = a1 + b1 * real_oe_percap;
fit unemployment_rate / white;
title 'Test for Heteroskedasticity';
This gives me a score for White's Test, indicating heteroskedasticity in the basic model, which I am attempting to correct for here. How can I tell if I have successfully corrected for that? PROC PANEL doesn't take white as an option.
04-30-2015 10:55 AM
04-27-2015 02:36 PM
04-30-2015 10:55 AM
Need further help from the community? Please ask a new question.