On Fri, 29 Jul 2011, Sam Sam wrote: > I input monthly data from 1980M1 to 1990M12. I set the sample > range from 1980M6 to 1990M12. When I estimate AR(3) model, data > used to estimate the model is 1980M1-1990M12, 1980M6-1980M12, > 1980M9-1990M12, or so on?
Exactly what happens here depends on the estimator. When you use exact ML the Kalman filter can handle pre-sample missing values, but conditional ML (the --conditional option) cannot. You can see the effect by running this: <hansl> nulldata 132 set seed 2908743 setobs 12 1980:01 series y = normal() smpl 1980:06 1990:12 arima 3 0 0 ; y # exact ML arima 3 0 0 ; y --conditional y[1980:03] = NA arima 3 0 0 ; y # exact ML arima 3 0 0 ; y --conditional </hansl> Despite making the y observation for 1980:03 into NA, the estimation range for exact ML still starts in 1980:06. But the start of estimation in conditional mode is pushed to 1980:07 by the missing data. Allin Cottrell