Here is the code to do your rolling window estimation but with the added enhancement that your initial starting values are based upon the previously estimated model.  I would think that this would speed up estimation considerably. SAS Technical support was instrumental in creating this code and should be given full credit.  It is a shining example of the skill of our technical support. -Ken /************************************* * generate some data; data a ; do i = 1 to 100 ; x = normal(1) ; y = 2+3*x + rannor(23); output; end; run; %rolling(data=a , outest=est1 ,regn=20 ,totn=45  ); proc print data= est1; run; **************************************/ %macro rolling(data= , outest= ,regn= ,totn=  ); * this macro will perform the rolling regressions; * each rolling regression uses estimates from previous rolling window as initial values ; * first rolling window uses estimates using all observations as initial values ; *     data=     data set name; *   outest=     OUTEST= data set; *     regn=     number of obs in each regression; *     totn=     number of obs in the data set; * clear out the OUTEST= data set; proc datasets lib=work; delete &outest; run; /*use outest from all observations as initial values for the first rolling window*/ proc autoreg data = a outest = _temp_ noprint; model y =x /nlag= 2 garch=(p=1,q=1); run; %do i= &regn %to &totn %by 1; * what is the first obs?  ; data _null_; x=&i - &regn +1; call symput('start',trim(left(x))); data _null_; lab='r'||trim(left(&start))||'_'||trim(left(&i)); call symput('label',lab); run; data null ; set _temp_; call symput('int',Intercept); call symput('beta',x); call symput('ar1',_A_1); call symput('ar2',_A_2); call symput('arch0',_AH_0); call symput('arch1',_AH_1); call symput('garch1',_GH_1); run; * run the regression; proc autoreg data=&data(firstobs=&start obs=&i) outest=_temp_ noprint plots = none; &label: model y=x /nlag= 2 garch=(p=1,q=1) initial = (&int &beta &ar1 &ar2 &arch0 &arch1 &garch1); run; * append the OUTEST= data set; proc append base=&outest data=_temp_; %end; %mend; * generate some data; data a ; do i = 1 to 100 ; x = normal(1) ; y = 2+3*x + rannor(23); output; end; run; %rolling(data=a , outest=est1 ,regn=20 ,totn=45  ); proc print data= est1; run; ... View more

Hi Wouter, So far I think you have made good use of MACRO and there really isn't anything you could do to speed that part up (the passes).  The developer and I have talked about this issue and one idea is to set the most recent estimates as INITIAL values on the next window of parameters.  This could really speed things up.  We are working some code that would let you do this.  If you have some ideas on how to do that, you can specify the values with this code SAS/ETS(R) 13.1 User's Guide Stay tuned. -Ken ... View more

Thank you for your question. Here is an answer from technical support detailing what happens when NOINT is chosen. You can see the creative use of MACRO at the bottom which would let you define your R^2 in whatever form you would like.  -Ken " With NOINT option, PROC AUTOREG does the same as PROC REG which uses uncorrected total sum of squares instead of corrected total sum of squares. Here is a usage note discussing this redefined rsquare: http://support.sas.com/kb/22/632.html Here is an example reproducing the rsquare with NOINT option: data a ;   do i = 1 to 100 ; x = normal(1) ; y = 3*x + rannor(213); output;   end; run; proc means data= a; var y ; output out = outm mean = my css = ssc uss = ssu; run; ods output fitsummary =fits(where =(Label1 = 'SSE') keep = Label1 cValue1); proc autoreg data =a ; model y = x/noint ; run; proc print data = fits ; run; data null ; set fits; set outm; call symput('sse',cValue1); call symput('ssc',ssc); call symput('ssu',ssu); run; data comp ; rsq_noint= 1-&sse/&ssu ; run; proc print data= comp; run; I hope that helps." ... View more

Hi Lilly, A couple places to look.  If you are doing VCE corrections, many of those are available in PANEL using these HCCME options SAS/ETS(R) 13.1 User's Guide If you are looking to use GLS then I would suggest PROC MIXED as the likely place as it will have longitudinal model options that you would likely want. Good luck-Ken ... View more

Hello Madelaine, I agree with Reeza, especially if you primary interest in in the qualitative marginal response, as a linear panel model opens up the possibility of controlling for fixed effects (unobserved heterogeneity).  Check out this link in the documentation to a two-way fixed effect linear model. At minimum, it is a good place to start. SAS/ETS(R) 13.1 User's Guide Ken ... View more

While I don't quite follow exactly what you would like to do, you need a "parameters ; " statement your PROC MODEL code for it to estimate anything. That is, you need an unknown parameter.  See this SAS/ETS User's Guide Example Programs If your problem is to find an optimal solution, some example code is here. SAS/ETS User's Guide Example Programs Best-Ken ... View more

Hello, Thank you for your question.  It actually seems that your problem is more closely related to reporting and processing longitudinal/panel data than it is about estimating a panel model. There are couple of ways to do this in SAS with the most common to be using DATA step concepts or with PROC MEANS and some BY groups. I would go a different direction and use PROC TIMEDATA to do this.  since you have repeat observations on a student at different moments in time, your 'by group' would be a student ID.  You would then use TIMEDATA like a datastep but instead of thinking about the PDV being a row, think about the PDV as a column of data.  If you provide some data I wouldn't mind taking a crack at solving your problem. -Ken ... View more

Care to share some data with us? Based on your description, you seem to care slightly more about prediction than consistent estimation.  Is this right?  In that case, you might try all of the methods. The X's you have described are likely jointly determined with the Y's you are using for your analysis.  Look at this paper for an explanation of how to properly lag these values. http://www.clevelandfed.org/research/workpaper/2006/wp0606.pdf to estimate marginal effects in a dynamic panel model. ... View more

This is a great question. Thank you. You are correct that the FE estimator with lagged dependent values suffers from a bias on the AR coefficient inversely proportional to the number of time series observations in the data.  As you noted, the typical solution to the problem is GMM, which you can implement in SAS here. SAS/ETS(R) 13.1 User's Guide The GMM estimator may have its own set of problems which leads you to your point. Which is better if T is big?  Two expositions on the subject are linked below. Fixed effects, lagged dependent variables, or what? http://fmwww.bc.edu/EC-C/S2013/823/EC823.S2013.nn05.slides.pdf The Park's Method might be an attractive option under certain circumstances, however not in your case. (as far as I can tell). As you correctly mentioned, Park's is an AR method and utilizes RE for efficiency purposes.  As with all RE methods, there is an assumption that the unobserved heterogeneity is uncorrelated with the included regressors. With observational data, this assumption is hard to swallow. So if you used Park's you might very well be introducing Omitted Variable Bias (OVB). The consequence of OVB is inconsistency with the added problem of having no clue what the sign of the bias is. And this bias will exist as T grows. Now, you could theoretically test this OVB issue on your misspecified models (note that both models are misspecified) by looking at the Hausman test after using RE estimation in SAS, but I would use this as a loose guide rather than definitive evidence. SAS/ETS(R) 13.1 User's Guide My suggestion if you have a lot of data and want to avoid GMM is to live with the bias. May I ask what kind of data you have? Thanks-Ken ... View more

As Steve mentioned, PROC SYSLIN or PROC MODEL are the easiest ways to estimate a regression using two-stage least squares.  Here is an example: There are also more examples at SAS/ETS 13.1 User's Guide Example Programs data in;   label q = "Quantity"   p = "Price"   s = "Price of Substitutes"   y = "Income"   u = "Unit Cost";   drop i e1 e2;   p = 0; q = 0;   do i = 1 to 60;   y = 1 + .05*i + .15*rannor(123);   u = 2 + .05*rannor(123) + .05*rannor(123);   s = 4 - .001*(i-10)*(i-110) + .5*rannor(123);   e1 = .15 * rannor(123);   e2 = .15 * rannor(123);   demandx = 1 + .3 * y + .35 * s + e1;   supplyx = -1 - 1 * u + e2 - .4*e1;   q = 1.4/2.15 * demandx + .75/2.15 * supplyx;   p = ( - q + supplyx ) / -1.4;   output;   end; run; /*-- OLS Estimation --*/ proc syslin data=in;   demand: model q = p y s;   supply: model q = p u; run; /*-- Two-Stage Least Squares Estimation --*/ proc syslin data=in 2sls;   endogenous p;   instruments y u s;   demand: model q = p y s;   supply: model q = p u; run; ... View more

I have consulted with the author of procedure and he has informed me that currently there does not exist the capability to estimate this model with an inequality constraint.  A product improvement suggestion has been initiated. Thank you. the correct form for an equality constraint in AUTOREG on 12.3 is . data MyData;     do t=1 to 1000; Log_Yield = normal(111); temp = normal(111); if (temp>0) then extravar1 = 1; else extravar1 = 0; temp = normal(111); if (temp>0) then extravar2 = 1; else extravar2 = 0; output;     end;     drop temp t;     run; proc autoreg data=MyData outest=o1;     model Log_Yield = / garch=(p=1,q=1 , type = exp) initial=(1.6 0.01 0.07 0.1 0.01 0.01 5.47) maxiter=1000 converge=0.01 ;     hetero extravar1 extravar2 / coef=nonneg;     restrict _GH_1=0.0099;   run; ... View more

Thanks for using the community. Let's see if I can help. First thing is that since PANEL and TSCSREG are SAS/ETS procedures, you might have more luck posting on Forecasting and Econometrics community page. Feel free to do that in the future. Second is that you are correct in that RANONE will produce the Hausman test for the validity of using Random Effects. BTW, this tends to be only popular in social sciences as they are the ones that tend to work most often with observational data. The syntax for PANEL (and most SAS procedures) does not allow for continuous*continuous interaction effects in most cases.  If your data happen to be categorical, you could pull this off with a CLASS statement such as, proc panel; ID gid pid; class Same3; model YardsMinAvgP3 = PIgP Same3 PIncP*Same3 FracPassGameIncP   YPRGameIgP YPPGameIgP FracSuccGame Down YTG FieldPos PointDiff DiffSpreadvsActual Cold Wind Weather Grass OffHome / ranone ; run; if they are continuous, multiply the columns in a data step. The second part of your question concerns a nonlinear model and the Hausman test.  This test does not exist in the same way as with a linear model. My suggestion would be to simply keep going on the path you currently are and use a linear probability model.  It it is a consistent estimator of the average effect and you could certainly test the validity of the RE estimator in this environment. The closest thing to a FE estimator in the nonlinear world is something called the Correlated Random Effects model. You can read up here. http://www.iza.org/conference_files/SUMS_2013/slides_1_linear_iza.pdf Also, you will find sample code for PANEL here. http://support.sas.com/documentation/onlinedoc/ets/ex_code/121/index.html Copy and paste in your code window. Hope this helps. -Ken ... View more