serial correlation

[appleyard]

hi there i really need help for my regression. I have found serial correlation via the BG LM test in my regression and want to fix it, but I am just unsure about the process. Do I just ad AR(1) to my regressors- and this is it? That fixes all the serial correlation problems or? I am a bit confused, please help me out, I'm a bit new to this.

NOTE when i entered AR(1) in my DW statistic was above the dU limit, so the test is conclusive right and there is no more serial correlation?


Here are my results:

Dependent Variable: FORRATE
Method: Least Squares
Date: 03/11/12 Time: 19:05
Sample (adjusted): 2 3142
Included observations: 3129 after adjustments
Convergence achieved after 7 iterations

Variable Coefficient Std. Error t-Statistic Prob.

C 0.072224 0.013922 5.187769 0.0000
ASIAN 0.116053 0.015257 7.606478 0.0000
BLACK 0.027659 0.004039 6.848391 0.0000
HISP 0.030511 0.002733 11.16292 0.0000
WHITE 0.012150 0.003812 3.187408 0.0014
MIDWEST 0.006834 0.000932 7.329219 0.0000
PACIFIC -0.028791 0.004338 -6.637492 0.0000
WEST -0.004759 0.001318 -3.610978 0.0003
BACH -0.084305 0.005275 -15.98341 0.0000
HS 0.018763 0.006859 2.735330 0.0063
POVERTY 0.026004 0.007442 3.494140 0.0005
UNEMP 0.393981 0.017074 23.07528 0.0000
LNMEDVALHUNIT -0.005808 0.000969 -5.995244 0.0000
AR(1) 0.364028 0.016902 21.53787 0.0000

R-squared 0.562896 Mean dependent var 0.049220
Adjusted R-squared 0.561072 S.D. dependent var 0.020924
S.E. of regression 0.013863 Akaike info criterion -5.714750
Sum squared resid 0.598633 Schwarz criterion -5.687687
Log likelihood 8954.726 Hannan-Quinn criter. -5.705036
F-statistic 308.5729 Durbin-Watson stat 2.212066
Prob(F-statistic) 0.000000

Inverted AR Roots .36

[appleyard]

please help... i am confused about the whole ar(1) when reading it in eviews help section. does this remove serial correlation?

should i just use a lagged dependent variable ??

[JimForest]

You want to make sure you don't have nonstationarity in the dependent variable. Use the ADF test, I suggest with auto selection and don't worry about how many lags. YOu could try KPSS but keep in mind the null hypothesis is reversed. If your dependent is nonstationarity, you may want to transform it to a first difference.

An AR(1) term is a lag of the dependent variable. Use the ACF and PACF to see if there is significant autocorrelation at different lags. If you need to you can add ar terms or seasonal ar terms. But it looks like you have a factor model as opposed to a forecasting model? So maybe you want to use lagged values of the regressors as it could take time for a change in an independent variable to affect the dependent.

I can't really say I can tell if your results are from a spurious regression but I would be more concerned if the R-squared was really - like 99%.

If you just want a good forecast, then by all means consider AR and/or MA terms. If you are just testing hypotheses on coefficients or groups of coefficients, then I would pay more attention to modeling lags in the predictors.

You just need to brush up on the following concepts: stationarity, unit-root nonstationarity, trend stationarity and the concept of a spurious regression.

Hope this helps.